John J. Heim
An Econometric
Model of the
US Economy
Structural Analysis in 56 Equations
John J. Heim
University at Albany-SUNY
Albany, New York, USA
I left academic life in 1972, after getting my Ph.D. At that time large-scale
econometric modeling of the economy was the rage; everyone thought it
would be just a matter of time before we had “done enough science” to
allow economists to discuss economics in the classroom, not in terms of
the alphas and betas of theoretical models, but in terms of the real-world
coefficients they represent. Economics would become the next branch of
engineering, or so many thought.
Much to my surprise, when I returned to academic life 25 years later
things had not much progressed. Most economists were still using alphas
and betas to describe how one variable affects another in economics. For
lack of vigorous, concerted effort over those 25 years to pursue the hard
numbers underlying the theories, and their statistical significance, econom-
ists were still just discussing theories with the best “numbers” we had – the
abstract alphas and betas of pure theoretical discourse. Because we hadn’t
disciplined our presentation of theories to those scientifically proven to
work, even more theories abounded than was the case in 1972. Worse, the
overriding emphasis in economic theory was not on “what works?”, but
on “what’s new?”.
My engineering students knew the difference. When I tried to describe
macroeconomics as real science, and then described the coefficients that
connect one variable to another in alphas and betas, instead of real num-
bers, they just snickered. “Yes, but what is the real relationship?” they
would ask, meaning what are the real numbers? “And if you don’t have
them, why do you call this science?” they would ask. Certainly in their
vii
viii PREFACE
ix
x ACKNOWLEDGMENTS
Nor could the book have been written without the strong support of
my wife Sue. This book required 2 years full-time work, and before that,
considerable part-time work. The problems to be resolved required endless
long hours at work, and endlessly preoccupied my mind, even at home.
Sue was always willing to make the sacrifices necessary to cope with all
that.
Finally, I must acknowledge the secretarial assistance provided by
Annemarie Hebert. She has helped pull together, duplicate, and send out
endless drafts of this work.
C ONTENTS
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 Modern Macroeconomics: Moving from
the Methods of Economic Philosophy to Those of
Economic Science 5
1.2 Summary of Ways in Which This Large-
Scale Econometric Model Improves on Past Work 9
1.3 The 56-Equation Model: 30 Behavioral Equations,
15 Identities (Product Side of National Income
and Product Accounts (NIPA)), and 8 Behavioral
Equations, 3 Identities (Income Side of NIPA) 14
1.4 The 38 Behavioral Equations: Coefficients,
Significance, R2 , and Durbin Watson Tests:
(Summary of Results: Detailed Explanations of
Findings Presented in Chapters 4–20) 16
2 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.1 General Methodological Issues 40
2.2 Choosing Between VAR, DSGE, and
Cowles Commission Models 48
xi
xii CONTENTS
3.2 Otto Eckstein’s (1983) The DRI Model of the U.S. Economy 125
3.3 Ray Fair’s Estimating How the Macroeconomy Works (2004) 131
3.4 Federal Reserve Board/U.S. Model (1996) 140
3.5 Literature Review Summary 144
15 Endogeneity of Government
Spending Levels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 309
15.1 The Model for Total Government Spending for All
Purposes: Goods, Services, and Transfers 309
15.2 The Model for Government Spending on Goods
and Services Only 313
18 Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 363
18.1 Introduction 363
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 445
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 451
L IST OF F IGURES
xvii
xviii LIST OF FIGURES
xix
xx LIST OF TABLES
The book has two parts: Part I contains 45 equations describing in detail
the “product side” of the National Income and Product Accounts (NIPA).
It contains tested models of the GDP and its major components, and the
determinants of their level of production (Chapters 4–19). Part II provides
11 additional equations describing how the value of the product generated
producing the GDP is distributed among the factors of production. For
each factor of production there are two equations. The first describes the
variables that were found to determine each factor’s percentage share of
national income. The second describes the variables found to determine
the total amount (the level) of each factor’s total income. These models
describe the variables whose own changes cause the distribution of income
among factors to shift from one factor to another over time (Chapter 20).
Chapter 19 provides a summary of the substantive findings as to the
determinants of GDP and its components. Chapter 20, Section 20.5,
summarizes the determinants of factor shares and levels of income.
xxvii
xxviii SUMMARY
METHODOLOGY
Good science requires replicability of results. This chapter’s goal was to
provide, to the best extent possible, models whose results meet the rep-
licability standard. Largely, this goal appears to be achieved, though in
some areas more remains to be done. Hopefully, future generations of
researchers will find it worthwhile to take up where this study leaves off.
In particular, in some equations we were not able to fully resolve the “left
out” variables and multicollinearity problems that affects the credibility of
parameter estimates in any economic model.
In most models 85–95% of the variance is explained. However, in some
models, there are definitely some “left out” explanatory variables remain-
ing to be found. Less of the total variance in the model than we would
like is explained by the variables. Models with this problem are identified
in the text.
In addition, the problem of multicollinearity needs to be better
resolved. It is perhaps the most serious impediment to doing good science
in economics today. To mitigate the problem in this study, we use first dif-
ferencing, and careful selection of combinations of explanatory variables
used. In addition, we do extensive robustness testing, by adding and sub-
tracting explanatory variables to a model, to ensure (reasonable) model
changes do not cause marked changes in other parameter estimates. For
most of our parameter estimates we are able to show these techniques
achieved the desired level of stability, but not for all. For some models,
parameter estimates are still sensitive to exactly what other variables are
included in the model (these models are identified in the text). Economists
needs to develop better scientific methods for dealing with this problem.
CHAPTER 1
Introduction
fiscal policy (with or without crowd out effect variables included). This
reconciles a major formulation of the neoclassical model with Keynesian
mechanics. It shows that under the right assumptions, Keynesian results
obtain from this neoclassical model.
This chapter chooses to use Cowles Commission structural mod-
eling techniques, yet today most macroeconomic models use either
Vector autoregression (VAR) or Dynamic, Stochastic, General Equilib-
rium(DSGE) techniques. One can ask why the Cowles technique was
chosen. The answer is Cowles-type models are both good economics
(based on recognizable economic theory) and good science (use the best
econometric methods available for presenting and testing the enormous
detail in such large models). By comparison, VAR is usually thought of
as good science, but bad economics (atheoretical, so results can be hard
to interpret), and DSGE is generally considered good economics, but bad
science (key parameters calibrated, not estimated). These deficiencies have
been widely criticized during the past decade by major economists. A sum-
mary of those criticisms is presented below. They suggest that a better
alternative is needed. Cowles modeling, by combining good economics
and good science, is not only better, it is really the only alternative available
for large-scale modeling of the economy’s multitudinous determinants.
DSGE criticisms
VAR criticisms
Eckstein (1983) (Cowles) structural model forecasts better than VAR models tested
Gale and Orszag (2004) (∼Cowles) structural model forecasts better than VAR models tested
Fair (2004) (Cowles) structural model forecasts better than VAR models tested
4 1 INTRODUCTION
The wide range of variables used also builds on the more limited
number typically used in VAR and DSGE analysis.
10 1 INTRODUCTION
GDP ½ of 1%
Consumption ½ of 1%
Investment 3.2%
CT CT CD CD CM CM CB2 CB2
Equation #: 4.1T 4.1T.TR 4.4 4.4.TR 4.2 4.2.TR 4.6 4.6.TR
DJ0
DJ–1 –1.65∗∗∗ –1.37∗∗
DJ–2 0.43∗∗∗ 0.44∗∗∗ 0.44∗∗∗ 0.48∗∗∗ 0.11∗
DJAV-2, -3
DJAV-0, -1, -2
XRAV 1.44 –0.33 3.13∗∗∗ 2.22∗∗ 14.06∗∗∗ 13.16∗∗∗
POP16 –418.25 –517.17∗∗∗ –515.07∗∗∗ 239.00∗∗ 68.02
POP 0.018∗∗∗ 0.017∗∗∗ 0.020∗∗∗ 0.020∗∗∗ –0.002 –0.012∗∗
ICC–1 0.37 0.41 0.53∗∗ –14 0.52
M1Real-1
M2AV-2-4 46.31∗∗∗ 44.78∗∗∗ 38.16∗∗∗ 38.00∗∗∗ 7.16 –30.60
(M2-M1)Real –0.14
PERSAV–2 –0.07 –0.07∗∗∗(2)
CB 0.12∗∗∗ 0.13∗∗∗ 0.10∗∗∗ 0.09∗∗∗ +0.12∗∗
X 0.47∗∗∗
IHousing
PHousing
Int.%Mortgage
CNONDUR
CDUR
R2 (%) 95.3 94.8 88.7 87.8 86.7 76.7 59.7 55.3
D.W. 1.6 1.6 2.0 2.2 2.2 1.5 2.5 2.2
Table 1.4.1 (continued)
(continued)
1.4 THE 38 BEHAVIORAL EQUATIONS: COEFFICIENTS, SIGNIFICANCE, R2 ,. . .
19
Table 1.4.1 (continued)
IT IT ID ID IM IM IBor IBor
Equation # 5.2 5.2.TR 5.4 5.4.TR 5.6 5.6.TR 5.8 5.8.TR
(continued)
Table 1.4.2 (continued) 22
Significance levels: ∗ 10%; ∗∗ 5%; ∗∗∗ 1%.1 (when residential investment is removed from total investment, this coefficient is 0.10∗ , more consistent with our
IP&E finding (0.13∗ ∗) in the next table. Likewise, the domestic investment coefficient increases from 0.02 to 0.07. Similarly, removing IP&E from ITotal
allows CBOR to have a significant coefficient in this model as it does in the IRES equation in the following table. We conclude the failure of IBOR-1 and
CBor to appear as determinants of IT and IDom is technical, not a sign of inconsistency of results.)
where
IT = real total investment in domestically produced and imported investment goods
ID = real total investment in domestically produced investment goods
IM = real total investment in imported investment goods
IB2 = real business borrowing
IP&E = real investment in plant and equipment
IRES = real investment in residential construction
IInventory = real inventory investment
Disp. Inc. = real disposable income (Y-T)
ACC = real accelerator (YT -YT-1 )
T = real total government revenue
G = real total government spending
DEP = real depreciation allowances
XRAV0-3 = real exchange rate average0-3
CapUtil% = % of productive capacity utilized
Y = real GDP
PRREAL = real prime interest rate
CBOR = consumer borrowing
DJAV = NYSE Composite Index
M1REAL = real M1
PROF = real corporate profits
PHousing = real price of housing
POP= U.S. population size
CT = real total consumption
1.4 THE 38 BEHAVIORAL EQUATIONS: COEFFICIENTS, SIGNIFICANCE, R2 ,. . .
where
TTotal = real total government receipts
G∗ Total = real total government spending (see p. 156 note on coefficient validity)
TDeficit = tax reduction-induced deficit
GDeficit = government spending increase-induced deficit
PR = real prime interest rate
DJAV = NYSE Composite Index
XRAV = real exchange rate average, current and past three years
POP16 = ratio of population 24 and under to population 65 and over population
POP = U.S. population
ICC–1 = Conference Board’s Index of Consumer Confidence
M2AV-2-4 = real M2 money average for period from 2–4 years ago
CBor = real consumer borrowing
ACC = real accelerator (YT -YT–1 )
DEP = real depreciation allowances
CAP–1 = capacity utilization percentage, lagged 1 year
PROF0 = real corporate profits
IBOR(–1) = real business borrowing, lagged 1 year
X = real exports
(X-M) = real net exports
1.4 THE 38 BEHAVIORAL EQUATIONS: COEFFICIENTS, SIGNIFICANCE, R2 ,. . . 25
Table 1.4.4 Is the prime interest rate determined by the Taylor rule?
Initial model
PRREAL = .42INFL – 1.29UNEM – .008 M1REAL + .012 M1REAL(–1)
(t =) (2.6) (–4.8) (–3.4) (2.8)
(Eq. 9.2)
–.002 Taxes – .001 G.Spend + .44AR(1)
2
(–1.1) (–0.5) (2.4) R = .81 D.W. = 2.
Notes: TDef and GDef included because government deficits thought to affect the economy by raising
interest rates. See text discussion.
Robust model
PRREAL = .42INFL – 1.30UNEM + .44AR(1) R2 = .67 D.W. = 1.9
(Eq. 9.2TR)
(t =) (2.8) (–6.6) (2.4)
Table 1.4.5 Is the prime interest rate determined by traditional Keynesian “LM”
theory?
PRREAL = .002 GDP – .0.27M1REAL + .022 M1REAL AV(–1–2) ) R2 = .22, D.W. = 1.6
(t =) (1.9) (–4.3) (4.0)
(Eqs. 10.2 and 10.2TR)
All three variables were significant in all four time periods tested and robust to model changes, hence
Eq. 10.2 is also our robust Keynesian model 10.2.TR.
Tables 1.4.4 and 1.4.5 Significance levels: 1.7 = 10%, 2.0 = 5% and 2.7 = 1%.
where
PRReal = real prime interest rate
INFL = inflation rate (CPI)
UNEM = unemployment rate
M1Real = real M1 money supply
Taxes = real total government revenue (T)
G.Spend = real total government spending (G)
AR(1) = first-order autocorrelation control
GDP = real gross domestic product
Table 1.4.6 Determinants of savings
26
Personal savings Personal savings Corporate savings Corporate savings Deprec. savings Deprec. savings
Equation #: 13.3.1 13.3.1TR 13.1.2 13.1.2TR 13.2.1 13.2.1.TR
PRReal(–2) 2.72
DJAV0 –.25
DJAV0+(–2) 9 –2.29E-25∗∗∗ –2.27E-25∗∗∗
XRAV0–4 1.80 7.92∗∗∗
POPY/Old Ratio 440.60
POP 0.011∗∗
ICC0.1 –713.64∗∗∗ –725.37∗∗∗
M2AV-2-4 –30.86∗∗∗
CBOR –0.18∗∗∗ –0.16∗∗∗
INFL3 –0.02∗∗ –0.03∗∗∗
Tax Increase93 –32.39∗∗∗
BEA DefnChge99 –195.10∗∗∗ –195.94∗∗∗
Katrina Shock05 –160.05∗∗∗ –176.30∗∗∗
ACC –0.03 –0.17∗∗∗
DEP –0.96∗
CAP–1 –0.39
PROF 0.73∗∗
IBOR 0.10
INV0 0.06∗∗∗ 0.06∗∗∗
INV–1 0.10∗∗∗ 0.10∗∗∗
INV–2 0.10∗∗∗ 0.10∗∗∗
INV–3 0.07∗∗∗ 0.07∗∗∗
INV–4 0.03∗∗ 0.03∗∗
INV–5 0.04∗∗∗ 0.04∗∗∗
INV–6–10 0.03∗∗∗ 0.03∗∗∗
INV–11–17 0.04∗∗∗ 0.04∗∗∗
R2 (%) 84.4 78.7 94.0 78.2 96.6 96.6
D.W. 2.6 2.1 2.2 2.1 2.1 2.1
Total government receipts∗ Total government spending Only G&S government spending
Equation #: 14.1 14.1.TR 15.1.1 15.1.1.TR 15.2.1 15.2.1.TR
where
GDP = gross domestic product
INFL%AV-1-2 = inflation % (CPI), average of two past years
UNEM % = unemployment %
TaxCutShock86 = 1986 Reagan tax cut shock
Tax Incr. Shock93 = 1993 tax increase on wealthy
POP–31–21 = growth in population from 21 to 31 years ago (proxy for school and infrastructure costs associated with new family formation period for
cohort)
Vietnam Build Up = years of Vietnam military spending build up
R. Mil. Build Up, Iraq = years of Reagan and Iraq military build ups
Shock08 (Fin.Crisis) = 2008 financial crisis
AR(1) = first-order autocorrelation control
Table 1.4.8 Determinants of unemployment and inflation
(continued)
1.4 THE 38 BEHAVIORAL EQUATIONS: COEFFICIENTS, SIGNIFICANCE, R2 ,. . .
29
30
Table 1.4.10 Determinants of velocity robust models only (where V1or2 = Y(P/M1or2 )
where
(DebtB&C ) = total business and consumer real debt levels
(PRAv -1,-2,-3 ) = average real prime interest rates
(Baa/NI) = real Baa bond interest rate/national income ratio
(EMPL/NI) = employment/NI ratio
(DJAV) = NYSE Composite Index
(T-G)/GDP = the ratio of the government deficit to GDP
(Prof/NI) = profit income/NI ratio
(Labor/NI) = labor income/NI ratio
PART I
Methodology
The methodology section is divided into two parts. The first part
(Section 2.1) discusses how standard methodological issues like station-
arity, endogeneity, heteroskedasticity, and multicollinearity are dealt with.
Treatment is brief because methods for dealing with most of these
problems are commonly agreed upon.
The second part (Section 2.2) deals with what many economists would
consider to be the first and perhaps most important methodological issue
a large – scale modeler has to deal with: whether to model using DSGE,
VAR, or Cowles Commission methodology. This section is 10 times as
long as the previous methodology section. This is because the issue dis-
cussed remains unresolved, unlike the most issues discussed in the first
section (2.1). Not all economists agree which of these three modeling
methods to use. Therefore, we feel the need to extensively empirically and
theoretically show that the modeling method chosen here (Cowles) is by
far the best at modeling how the actual economy performs, and there-
fore the logical methodological choice for those attempting to develop a
large-scale, detailed model of how the macroeconomy works.
Another reason these lengthy comparisons are important is that large-
scale econometric models have lost credibility in recent decades. They are
no longer routinely relied upon by economists and policy makers for reli-
able guidance, as was the case with, say, Eckstein’s DRI model, 30 years
ago (a Cowles model). The change in the type of models to VARs and
For readers already comfortable with Cowles methodology, and not inter-
ested in reading the full 38 pages devoted to discussing differences in
Cowles, DSGE, and VAR methods, the long second section of the meth-
odology can be skipped. The reader can move on to Chapter 3 (literature
review) or Chapters 4–19 (detailed presentation of the findings for each
equation in the model) without loss of ability to understand what is
presented.
All models were tested extensively to ensure the robustness of the findings.
This is particularly important when using nonexperimental techniques
like regression analysis where even moderate levels of multicollinearity or
2.1 GENERAL METHODOLOGICAL ISSUES 43
and subtracting variables from the time–period robust model to see if the
parameter estimates on remaining variables remain reasonably stable. Vari-
ables’ coefficients’ are considered to remain stable if adding/subtracting
variables from the model does not change their values by more than one
third. Some exceptions are made to this rule depending on particular
model characteristics. Reasons for exceptions are noted in the text.
For those who wish to move more quickly from initial to final fully
robust models, we note at the beginning of each equation’s development
process the initial model we are starting with, and where to find the final
robust model we end up with, so the reader can jump from start to finish.
We also note that Section 1.4 presents all initial and robust models results
for each variable used side by side, with each variable used identified.
This leaves “personal” income taxes defined as the total of PIT, ½ FICA,
Sales, and Misc. taxes. These added to the CIT and the other half of
FICA taxes, equal total government revenue. (See Economic Report of the
President, 2012, Table 83 for government revenue, Table 27 for personal
income)
Results are repeated below for the consumption model using the (Y-TT )
definition, and for exactly the same model using the disposable personal
income definition, also calculated using OLS since the standard method
for calculating Hausman endogeneity indicated no statistically significant
level of endogeneity was present among the right hand side variables.
Model 2.1.1
Consumption Function (Using Y-TT ):
CT = 0.50(Y – TT ) + 0.55(TT ) – 0.26(GT& I ) – 11.81PR
(t =) (11.4) (11.4) (–3.7) (–5.1)
+0.42DJ–2 + 3.42XRAV – 336.65POP16 + 0.012POP
(5.3) (2.3) (–1.3) (2.6)
+0.36ICC–1 + 40.86M2AV + 0.12 CB2
(1.3) (3.8) (3.1)
+0.04 GDPReal(–3) R2 = 94.9% D.W. = 1.8 MSE = 25.45
(1.1)
(2.1.1)
48 2 METHODOLOGY
Model 2.1.1.a
Same Consumption Function as 2.1
(Except Disposable Personal Income Used)
Cowles Models
Cowles models are structural. They provide answers grounded in both
good science and good economic theory to very detailed questions, such
as “What variables drive year- to – year consumer spending on durables?
Nondurables? Inventory Investment?” How are they different? It is the
only one of the three methods in which you will typically find upward of
a dozen explanatory variable in each equation, and dozens of explanat-
ory equations, richly describing the complexity of relations in a modern
economy. It is one of two reasons some economists prefer Cowles-type
structural modeling. DSGEs are also structural, but typically much smaller
and do not include the large number of variables and relationships need to
answer such detailed questions. Nor do VARs.
The second reason some economists prefer Cowles models, is that such
models when tested, tend to explain the actual behavior of the US macroe-
conomy far better than do VAR and DSGE models. This is the judgment
of many economists, whose comparisons are discussed in detail below. It
is also this study’s finding, based on detailed comparisons of the three
models’ performance, also presented below.
Cowles-type structural models were the standard way the science of
macroeconomics was applied from the until the 1980s. The first Nobel
prize in economics was awarded, in part, to Jan Tinbergen for developing
the first econometric model of the macroeconomy, and in 1980 another
was awarded to Lawrence Klein for the same type of research. The mod-
els tend to be demand driven, and therefore have a decidedly Keynesian
look. Cowles models take a top down, “macro principles first” modeling
approach. They cut the macro model into more and more micro-sized
pieces, adding relative price variables as they go along. In this way they
identify the effects of demand for other products on the demand for the
one being studied, thereby allowing for integration of micro into macro.
Eckstein (1983) provides numerous examples of this. This “macro prin-
ciples first” approach is based on the notion that most year-to-year changes
in demand for products are likely to be caused by macro factors (like busi-
ness cycles), not micro factors like prices, though relative prices do matter.
The Cowles approach is continued today in Ray Fair’s models of the
US and world economies. Cowles models incorporate the multitudinous
variables and relations necessary to describe the detailed structure of the
economy. Many believe macroeconomics parameters are sufficiently stable
to allow macroeconomic relationships to be evaluated as a science.Those
who agree would argue the goal of modeling should be to correctly
50 2 METHODOLOGY
DSGE Models
DSGE advocates argued that Cowles models lack a firm micro-foundation
in their explanations of economic behavior, i.e., are not built to show how
profit maximization and utility maximization conduct is actualized from
year to year. DSGE models are constructed and tested on the assumption
this “micro principles first” approach best describes what drives year-to-
year fluctuations in macroeconomic variables.
DSGE models also assume rational expectations characterize the actions
or consumers and businesses, i.e., they are predicated on the notion that
consumers accurately foresee future income, at least in the probability
distribution sense, and armed with this foresight, consumers base cur-
rent consumption decisions on expected future lifetime average income,
not just current income (Keynes) or past income (DeLeuuw/Modigliani).
Basing it on (accurate knowledge of) lifetime income allows (accurate)
maximization of lifetime utility, the consumers’ driving objective in a
DSGE model. Businesses are assumed to operate in similar profit max-
imizing fashion. Though the assumed micro foundations are not directly
testable, it is possible to look at, say, the first-order conditions for inter-
temporal utility maximization implied by the model, and test to see if some
of these results occur. For example, DSGE models assume consumers
can accurately deduce what savings level is required to maximize utility
intertemporally. If so, an implication of the theory that the savings rate
is such that consumption should be constant in all periods, for example,
Ct = Ct+1 = Ct+n except for unexpected shocks.
DSGE models are often parameterized rather than estimated. All theor-
etical concepts underlying a Cowles model are estimated. There is no such
thing as a “non-testable” component of the theory underlying Cowles
models.
VAR Models
The a priori use of theory to decide what relationships to test was
criticized in the 1980s by Sims and others and led to the development of
VAR modeling. VAR is based on the premise that everything (potentially)
may be a function of everything else, at least after one or more lags, and
2.2 CHOOSING BETWEEN VAR, DSGE, AND COWLES COMMISSION MODELS 51
The VAR error likely results from combining likely causal and spuri-
ous regression coefficients in models, a problem that is unavoidable when
using the same explanatory variables in every equation, and is discussed
in the VAR section below. The error also likely results from failure to
provide any way of adjusting forecasts for unanticipated future regime
change. Out-of-sample performance of Cowles models is better because
they control for variables which might cause regime change in the future.
This allows forecasts of (say) the effect of interest rate changes today on
investment 2 years in the future to be adjusted in the future in prede-
termined ways as specific types of change in the control variables occurs,
changes which were unforeseen at the time the initial forecast was made. In
this sense, the out-of-sample performance record of Cowles models better
describes how well they explain the causes of what actually happed, rather
than how well they forecast it. This is a key difference with VARs.
The DSGE flaws result from use of incorrect theory. Tests indicate using
current year income in consumption functions explains more variance than
estimates of lifetime income. Failure of the lifetime income assumption
casts doubt on the validity of the rational expectations assumption, a key
postulate of DSGE theory. The DSGE assumption of constancy of con-
sumer spending from year to year, except for technological shocks (Romer
1996) was also tested. The data tested did not support this theory. Wide
variation in consumer spending was found, aside from that caused by tech-
nological shocks. It was better explained by factors commonly found in
more Keynesian – looking, demand driven models, like Cowles models.
2.2.3.1 Tests by Others of How Well DSGE Models Fit the Data
The test of any theory’s usefulness is its ability to explain or forecast accur-
ately beyond the sample period used to estimate the model’s parameters.
For DSGE, perhaps the most popular model today is the Smets–Wouters
“New Keynesian” DSGE model.
Comparisons with Smets–Wouters
In examining the Smets–Wouters model, Edge and Gurkaynak (2011)
found that it explained very little variance in the data, only 13% or less,
depending on the test. Two tables with their findings are presented below
(Tables 2.2.3.1.1 and 2.2.3.1.2).
Some DSGE advocates have argued that, because of rational expecta-
tions, income should be predictable from year to year except for changes
due to unforeseeable technological progress. It is argued that this means
DSGE’s should not be able to successfully explain much of the variance
in year-to-year changes in consumption. This is a random walk explan-
ation of consumption growth over time. However, tests below indicate,
most of the year-to-year changes in consumption are due to factors other
than year-to-year technological change. Most of the change that Smets–
Wouters can’t explain – which is roughly 90% of it according to Edge
and Gurkaynak, stems from sources other than technological progress,
DSGE model
Slope 0.451∗∗ 0.089 0.031 0.209 0.167 0.134
(0.108) (0.149) (0.250) (0.261) (0.216) (0.174)
Intercept 0.261∗∗ 0.421∗∗ 0.446∗∗ 0.363∗∗ 0.386∗∗ 0.398∗∗
(0.051) (0.082) (0.122) (0.128) (0.112) (0.112)
R2 0.13 0.00 0.00 0.02 0.01 0.01
No. of obs 104 104 104 104 104 104
Source: Edge and Gurkaynak (2011), Board of Governors, Federal Reserve System, p. 20
Notes: **/* denotes 1/5% significance levels
54 2 METHODOLOGY
DSGE model
Slope 0.374∗ 0.485 0.477 0.507 0.495 0.553
(0.174) (0.249) (0.321) (0.303) (0.312) (0.279)
Intercept 0.419∗ 0.313 0.331 0.299 0.320 0.284
(0.206) (0.292) (0.362) (0.346) (0.344) (0.311)
R2 0.08 0.09 0.07 0.08 0.07 0.06
No. of obs 104 104 104 104 104 104
Source: Edge and Gurkaynak (2011), Board of Governors, Federal Reserve System, p. 20
Notes: Standard Errors in Parentheses; ∗∗ /∗ denotes 1/5% significance levels
like business cycle effects. Business cycle effects are in Cowles models, and
may account for why they explain year-to-year variation in consumption
so well.
[Comparisons with Fair]
Fair (2007) compared his Cowles model to a DSGE model presented in
Del Negro et al. (2007), and found the Cowles model performed better.
The average root mean square error of fit were as follows:
VAR models do still fit better than DSGE’s when they are applied to real
data (and not to processed data that has had trend removed by filtering or
regression). (p. 53)
(VARs) still do better than DSGEs when they are applied to real data (ie
data that has not been processed by removing the trend, either by filtering
or regression)
However, there are other analysts who have found their DSGE models
perform as well as VARs. For example, Christoffel et al. (2010) note:
Smets and Wouters (2003, p. 1125) found that their DSGE model
performed as well as VARs.
Wickens (2012) notes:
A review of the literature shows that forecasts from DSGE models are not
more accurate than either times series models or official forecasts, but neither
are they any worse (Abstract of paper).
As far as I’m aware, private-sector firms don’t hire anyone to make DSGE
models, implement DSGE models, or even scan the DSGE literature . . . As
I see it, this is currently the most damning critique of the whole DSGE
paradigm
2.2.3.2 This Study’s Tests of How Well DSGE Models Fit the Data
In subsections 1 and 2 below, this study tests the DSGE assumption
of (accurate) intertemporal utility maximization. We test whether con-
sumers can foresee lifetime income accurately, and whether consumption
is constant from year to year, except for unexpected technology shocks.
Subsection 3 tests whether the Lucas critique holds, i.e., whether the past
effects of policy variable changes can predict the effects of similar changes
in the future. Subsections 4 and 5 compare the out-of-sample perform-
ance of the FRB/NY and FDR/US DSGE models with Cowles models
during the 2001–2010 period, based on models whose parameters were
previously estimated using the same estimation periods.
tests. Results were compared to models using only current income. This
old Keynesian formulation explained more variance in consumer spending
than any of the rational or adaptive expectations models.
We replicate this approach here, using a typical “Old” Keynesian
model consumption function. The new model adds some explanatory
variables not used in the Heim model. For example, prior years’ sav-
ings, and consumer borrowing. Testing also covers a slightly longer time
period (1960–2006). Controls include the government deficit, current
year interest rates, a wealth variable, the exchange rate, and other variables.
Variables used are defined as follows:
“lifetime” average income, but still produces results inferior to the “Old”
Keynesian model, as shown in Eq. 2.2.3.2.1.3:
Tests using any one future year’s disposable income hugely reduce the
explanatory power, compared to using current year income, no matter
what future year is used. This again leads us to reject the hypothesis that
average income is a better determinant of consumer behavior than current
income. As an example, we show the results for the same model as tested
above, except that we substitute next year’s actual income for this year’s
as the income variable in the test. Results show that this reduces explained
variance from 94.8% (using current income) to 79.7% (using next year’s
income).
Table 2.2.3.2.1(1) Current and four future year annual changes in income
(real GDP) (Billions of 2005 Dollars)
The second set of tests repeats the simple and sophisticated tests done in
test one, except in per capita spending, rather than total spending, terms.
In the simple model, consumption and disposable income (after the por-
tion attributable to yearly productivity growth is deducted) are run in per
64
2 METHODOLOGY
Table 2.2.3.2.2(1) Yearly variation in consumer spending 1960–2010. Explained by yearly variation in TFP compared to
other determinants of consumption
capita terms. In the sophisticated model, the deficit, money supply, and
consumer borrowing variables are also measured in per capita terms.
In the simple model 81.5% of the total consumption is explained by the
full model, which includes both the disposable income per capita variable
and productivity index as separate variables. Removing the index reduces
R2 to 81.2%, implying the productivity index can uniquely explain only
3/10 of 1% of consumption’s yearly variation. By comparison, remov-
ing the disposable income variable indicates non-TFP disposable income
changes (such as those due to a recession) uniquely explain 58.4% of the
yearly variation in per capita consumption over the 50-year period. The
remaining 22.8% of the explained variance can be equally well explained
by either variable.
Test two’s more sophisticated model of consumer behavior included
not only (non-TFP) changes in disposable income and TFP’s effect on
income but also interest rates, wealth, exchange rates, the mix of young
and old in the population, the overall population size, the consumer con-
fidence index, savings accumulations in the prior 3 years, and access to
borrowing. All variables together, including TFP, explained 94.1% of the
total year-to-year variation in consumption over the 50-year period tested.
Removing the productivity index indicated productivity growth could only
uniquely explain 1.7% of the year-to-year variation in consumption (4.4%
including the variation that could be explained by either the productivity
index or the other variables). The rest of the variables uniquely explained
89.7% of year-to-year consumption variance, 53 times as much as the
productivity index could uniquely explain. This becomes 92.4% if all the
variance that could be explained by either productivity or the group of
variables equally well was assigned to this group of non-productivity index
variables.
The Third Set of Tests A third set of test results are also in per capita terms.
The models are identical to those in test two, except the current productiv-
ity index and four lagged values of it are included. This was done to test
the possibility that a productivity gain in this year’s disposable income
might only slowly lead to changes in consumer spending. For the simple
model, total explained variance was 67.1%, of which the productivity index
explained 23.6% uniquely. The disposable income variable explained 7.1 %
uniquely. Either variable could explain and additional 36.4% of the vari-
ance. For the sophisticated model, total explained variance was 95.4%, of
which only 2.9% could be uniquely explained by current year and four lags
of the productivity index. Forty-eight percent could be explained uniquely
66 2 METHODOLOGY
Table 2.2.3.2.3(2) Robustness over time: (2SLS model 5.2, 1960–2010 data)
So far, the evidence suggests that changes in policy regimes are not among
the principal causes of simulation error, that forecast error is largely created
by other exogenous factors and the stochastic character of the economy.
(Eckstein 1983, pp. 50–51)
Quarter GDP% Qtr Chge Annual Chge DSGE model Forecast vs. actual
Est. forecast
2002:IV 10,095.8
2003: I 10,138.6 0.0042 1.64 1.64 2003 qtr. av.
II 10,230.4 0.0091 3.64 3.64 3.1% DSGE
III 10,410.9 0.0176 7.04 3.6 4.0%Actual
IV 10,502.6 0.0088 3.52 3.55 (9/10 of 1%)
2004: I 10,612.5 0.0105 4.20 3.0 2004 qtr. av.
II 10,704.1 0.0086 3.44 3.0 3.1%DSGE
III 10,808.9 0.0098 3.92 3.0 3.7%Actual
IV 10,897.1 0.0082 3.28 3.3 (6/10 of 1%)
2005: I 10,999.3 0.0094 3.76 3.3
Using these estimated model parameters, the model was then used to
calculate estimated values of (yt ) for the out-of-sample period 2001–
2009. The Sbordone et al. model estimate of average yearly change in
GDP was 1.34%, 36% below the actual average change of 2.1% in the
absolute value of yearly changes). By comparison, for the Cowles model
(described below in this paper’s VAR section), the average out-of-sample
error was less than half of one percent (46/100) compared to Sbordone’s
1.34%. The Sbordone DSGE model out-of-sample error is 2.9 times as
large.
For the Cowles model developed later in this paper (see VAR section
below), the error for 2003 was much smaller: 9/100 of 1% compared
to 90/100 of 1% for the DSGE model above. For 2004, the Cowles error
of fit was 28/100 of 1% compared to 60/100 of 1% for the DSGE model
above.
In this note, I focus on the effects of forward guidance and compare the pre-
dictions of three structural models: FRB/US, EDO and the model of Smets
and Wouters (2007), which is a representative class of dynamic stochastic
general equilibrium (DSGE) models . . . All three models considered here
share a core New Keynesian structure. (p. 13)
• The lagged deviation of the decision variable from its target (y∗t–1 –yt–1 )
• Lagged changes in the decision variable (y)
• A weighted sum of expected future changes in the target variable (y∗ )
One criticism of the FRB/US model is that not all its parameters were
estimated; some were calibrated. Fair (2004, website version) noted that
74 2 METHODOLOGY
The FRB/US model . . . has strong interest rate effects . . . In most of the
expenditure equations real interest rate effects are imposed rather than
estimated. Direct tests of nominal versus real interest rates . . . are not done,
and so there is no way of knowing what the data actually support in the
FRB/US expenditure equations. (p. 120)
(Brayton et al. (April 2014) acknowledge as much.)
The data set used to test the model spanned the quarterly periods:
63:1–95:4.
The following definitions are used in this and other consumption
equations below:
The model explains slightly more than half the variance in total consump-
tion, and even less for subcomponents such as motor vehicles (R2 = 43%),
and other durables (R2 = 34%). For residential construction, R2 = (60%).
For the business investment categories, producer durables (40%) and
inventories (42%) results are also unimpressive. Hence, the model does
2.2 CHOOSING BETWEEN VAR, DSGE, AND COWLES COMMISSION MODELS 75
This original model has been updated over the years. The 2014 ver-
sion (for consumer nondurables and nonhousing services only) looks like
this:
aggregates are the sum the first differences of component year’s incomes,
not their levels.
The variable ct–1 ∗ is the sum of five income (GDP) variables. Each of
the five income variables is comprised of the sum of the first differences
of current and five leads values of income; and the same year’s unem-
ployment rate (a proxy for the output gap) is subtracted from the income
aggregate. A lagged version would contain last year’s value, the current
and four lead values. This is our approximation of the consumption equi-
librium condition used in FRB/US. The ct+1 ∗e variable is estimated using
the average of the first differences of seven leads of the variable c∗e (or
Y∗e in the 2014 version). The unemployment rate in the same period is
subtracted from each of the income components. Because the dependent
variable was not stationarity, nor fully cointegrated with the right hand
side variables, a trend variable was added, resolving the stationarity issue,
increasing the reliability of the regression coefficients. It is not needed in
the out of sample projections using these corrected regression coefficients.
Because of the need to use data for leads, the latest out-of-sample full dec-
ade we could test was 1991–2000. The in sample estimation period was
1960–1990.
Using the 1960–1990 data, the model’s parameters were estimated as
All variables are in logs, all c values are in real terms and taken from the Eco-
nomic Report of the President (ERP). Unlike the FRB/US model, which
assumes “rule of thumb” (i.e., Keynesian) consumers only account for
18% of consumption, this model allow the data to determine the actual
percentage. The “real” percentage is substantially higher, and as a result,
our approximation of the FRB/US model explains far more variance in
consumer spending than the actual FRB/US model. Hence, it is likely to
overstate the actual FRB/US equation’s ability to explain out-of-sample
variance.
The results indicate the current income variable (yt ) plus the trend
variable explain 70% of the 77% of the variance explained by the
model. By comparison, the three DSGE variables (plus trend) when
78 2 METHODOLOGY
used alone, explain only 15%. About 8% of the total variance can
be explained equally well by either Keynesian or rational expectations
variables.
The same model in levels rather than logs was estimated as
Fair (2012) noted the counter intuitive nature of these key DSGE
assumptions:
Fair also argues it is not realistic to think that agents are sophisticated
enough to have rational expectations, and concludes the Smets–Wouters
model seems highly misspecified.
A more philosophical objection to DSGE methodology is that it is not
a modern way of deciding how the economy operates. It is not scientific
and inductive, but rather based on self-evident truths (“consumers max-
imize utility,” etc.) and deductions from them. This is a throwback to the
nonscientific methods used by the natural and moral philosophers of the
seventeenth and eighteenth centuries to determine what was “true” and
“real.” For example, Sims (2006) described the difference noting
aggregate DSGE models are story – telling devices, not hard scientific
theories. (p. 153)
decade as the next. Part of how the modeler would decide they got the
model “right” would derive from the model’s ability to explain as well all
2-year variations in consumption, not just the one of special interest.
The scientist (a.k.a. Cowles or VAR modeler) would let the data
decide what variables are best included in a theory of the determinants
of consumption.
In this study, when we let the data decide, we find there simply is no
empirical support for DSGE’s key assumption that today’s consumption is
a function of (accurate) estimates of lifetime average income by consumers.
Without this, the DSGE assumption of consumers being able to accur-
ately estimate future income and therefore successfully be able to maximize
utility intertemporally simply collapses, as well as any policy change implic-
ations subsequently deduced from the theory, whose validity is dependent
on such assumptions.
Fair (2004, website version) has noted the limited empirical underpin-
nings of DSGE models:
The econometric approach that dominated research from the 1940s until
the 1980s (and perhaps still dominates) takes the observed data as a given
2.2 CHOOSING BETWEEN VAR, DSGE, AND COWLES COMMISSION MODELS 85
Nobel laureate Robert Solow (2010) also raised issues about how realist-
ically DSGE theory reflects economic reality. In congressional testimony
he decried the then-current situation in macroeconomics:
I do not think that the currently popular DSGE models pass the smell test.
They take it for granted that the whole economy can be thought about as if it
were a single, consistent person or dynasty carrying out a rationally designed
long-term plan, occasionally disturbed by unexpected shocks, but adapt-
ing to them in a rational, consistent way . . . . One important consequence
of this “representative agent” assumption is that there are no conflicts of
interest, no incompatible expectations, no deceptions. This cannot be an
adequate description of a national economy . . . An obvious example is that
the DSGE story has no real room for unemployment of the kind we see
most of the time . . . the only way DSGE and related models can cope with
unemployment is to make it somehow voluntary, a choice of current leis-
ure or a desire to retain some kind of flexibility for the future or something
like that. But this is exactly the sort of explanation that does not pass the
smell test.
David Colander (2010) has also criticized the policy usefulness of DSGE
models:
what the empirical evidence and common sense might tell them. Some of
the most outspoken advocates of this approach . . . admit the DSGE model
does not fit the data, but state that a model neither “can nor should fit most
aspects of the data” (Chari et al. 2009, p. 243). Despite their agreement
that their model does not fit the data, they are willing to draw strong policy
implications from it, for example, they write “discretionary policy making
has only costs and no benefits, so that if policy makers can be made to
commit to a rule, society should make them do so.”
(if at all) over time. Hence, they would argue next year’s effect of some
change in income on consumption is likely to be the same as last year’s
effect, ceteris paribus.
Fair notes choice of modeling method depends on one’s assumption
about rational expectations, where rational expectations means that, people
can (correctly) forecast their likely economic futures, and therefore make
decisions today about how to save and spend that will accurately maximize
their lifetime utility.
If rational expectations theory holds, when testing, we should explain
more variance in consumption using DSGE’s claim that current consump-
tion is determined by lifetime income, than is explained by the Keynesian
“current income only” hypothesis. Some research indicates this is not the
case. In testing, the more years – forward and/or backward added into the
average representing lifetime average income, the worse it gets at explain-
ing the effect of income on consumer spending (Heim 2008). Without
rational expectations, DSGE collapses to a Keynesian model (i.e., con-
sumption depends on current income only), or an adaptive expectations
model (e.g., the FRB/MPS), where consumer spending depends on both
current and past income.
We showed above that in typically structured consumption functions
C = f (disposable income, wealth, interest rates, etc.), current income
alone always explains more variance in consumption than long-term
income averages of the Modigliani “Life Cycle” or Friedman “Permanent
Income” type. (See section 2.2.3.2.1.). In short, “old” Keynesian models
explain the economy better than their DSGE and adaptive expectations
counterparts.
If so, one can ask why DSGE theory and its less empirical modeling
method still dominate macroeconomic thinking, and not Cowles or VAR
models? This was addressed by Eckstein (1983):
New classical and new Keynesian research has had little impact on practical
macroeconomists who are charged with the messy task of conducting actual
monetary and fiscal policy . . . From the standpoint of macroeconomic engin-
eering, the work of the past several decades looks like an unfortunate wrong
turn . . . it is clear that the new classical economists promised more than they
could deliver . . . The new Keynesians can be criticized for having taken the
new classicals’ bait, and, as a result, pursuing a research program that turned
out to be too abstract and insufficiently practical. (pp. 39, 44)
The Cowles models used in this paper are consistent with Colander’s
recommendation.
future direction and rate of motion as a linear vector that simply extends
its past direction and rate of motion. For very short future periods, peri-
ods too small for regime change to take place, this is a sound, scientific
approach to forecasting the future. High school physics students use sim-
ilar forms of vector analysis to determine (e.g.) where and when a boat
leaving a dock on one side of a river with a current will reach the other side.
The problem is, if you project far enough into the future, there will
eventually be regime changes (e.g., changes in the water current’s speed).
You must predict those accurately too if your forecasts are to remain accur-
ate. Because we cannot accurately foresee the future, this is not likely. For
this reason, it is not something VAR analysts try to do. Every VAR forecast
is accompanied by a strong ceteris paribus assumption. This limits the util-
ity of the simplest form of VAR to short run forecasting. As an example of
this problem, in structural models, testing may indicate 75 or more differ-
ent variables and lags are necessary to explain movements in the economy.
Changes in many of them are frequent and constitute a “regime change.”
Incorporating each would be important to the accuracy of a VAR forecast,
but would not typically be done in a simple VAR model.
2.2.4.2 Tests by Others if How Well VAR Models Fit the Data
Eckstein (1983) tested a Cowles model (the DRI model) and found it
had smaller forecasting errors than optimal ARIMA forecasting formulas
2.2 CHOOSING BETWEEN VAR, DSGE, AND COWLES COMMISSION MODELS 91
developed for the same period. Eckstein noted that “since ARIMA will not
produce turning points, the method is then guaranteed to fail.”
Comparative forecasting errors for 20 key economic variables indicated
the Cowles model had better 1 quarter ahead forecasts for all 20 variables,
and for 4 quarter ahead forecasts, 17 of the 20 Cowles model forecasts
were better than the ARIMA model.
Gale and Orszag (2004, p. 149), also found VAR projections to be
inferior to others when evaluating the impact of deficits on interest rates,
noting that most studies find a change in interest rates when deficits
increase, but that those that don’t are usually VARs, suggesting VAR-
based may be biased toward showing no effect of deficits on interest rates.
They also noted “VAR–based projections have been shown to be inferior
to those produced by the OMB or by Data Resources, Inc.” (p. 152).
Fair (2004) also found that his structural model performed better than
both a 7-variable VAR model with four lags for itself, and two lags for each
of the other six variables.
In addition, VARs commonly use lagged values of the dependent vari-
able in their models. If there is serial correlation in the error terms, using
lagged dependent variables results in biased and inconsistent parameter
estimates. (Griffiths et al. 2008, 2011), affecting forecast accuracy.
Full Model
Partial Model
Because of the limited size of our data set (40 observations, 1960–
2000) available for estimating the model, we could not run Montford
and Uhlig’s full model with more than three lags (they used six). We
also ran a partial model, with seven variables and five lags. Our version
of Mountford and Uhlig’s model, though very similar, was not identical
to their model in several respects. They used a crude materials price index,
while we used a total materials index when recreating their model. Also,
they used the GDP definition of government spending, which excludes
transfer spending, while we used the full amount in our updated ver-
sion of their model. Results indicated the average predicted change in
the GDP in the decade following model estimation for the Montford and
Uhlig models, was about twice as large as the actual average change in
GDP:
A fourth VAR, labeled the “general” VAR model was also tested, used
five lagged values of six variables for current year GDP determinants:
Lagged GDP, consumption, investment, government spending, exports,
and imports. Since GDP is an identity comprised of all of their current
values, we expected applying the VAR method, with absolutely nothing
that actually does determine GDP missing, would yield the best projec-
tions. Instead, it yielded the worst. The model tested is given below (all
variables are in real terms)
94 2 METHODOLOGY
Here again the estimates from the 1960 to 2000 sample are quite sim-
ilar the full 1960–2010 year sample estimates, This was expected, since
the graph of this equation shows the model explains consumer behavior
equally well in each of the five decades included in the sample.
The parameter estimates for the model of domestically produced goods
for export for the 1960–2000 period gives the model results shown in the
equation below:
used in test to those consistent with theory (to protect against confus-
ing spurious correlations with those consistent with causality), and uses
established econometric techniques (Hausman endogeneity tests, 2SLS)
to ensure identification issues are resolved.
In classic VAR models, each dependent variable is determined by
numerous lagged values of all economic variables in the larger economic
model, including lagged values of the dependent variable itself. This allows
the regression to determine how much each of several lags matters. It
also avoids what Sims considered to be the identification problem with
structural models, i.e., structural models attempt to define a priori, not
necessarily through scientific testing, precisely which variables are determin-
ants of each specific variable (like consumption or investment), and with
precisely what lags.
Results above indicate the Cowles approach led to better performance
results than the VAR approach.
Rt = r∗ + 1.5(πt – π∗ ) – 1.25(ut – u∗ )
+ (lagged values of R, π, u) + ξt(R) (2.2.4.4.1)
π = (lagged values of R, π, u) + ξt(π) (2.2.4.4.2)
U = (lagged values of R, π, u) + ξt(u) (2.2.4.4.3)
where r∗ is the desired real interest rate, πt and ut are the average values
of inflation and unemployment rate over several past periods, π∗t and u∗
are the target inflation and unemployment rates, and ξt is the error term
in the equation.
The theory is that this type of structural VAR avoids the problem of
having to decide what year’s error term in the interest rate equation can
be considered caused by (only) an interest rate shock.
The basic VAR model we tested in the previous section and compared
to a Cowles model was conceptually like the model above, but without the
Taylor Rule structural component included in the interest rate equation.
When separately testing the Structural and VAR pieces of a model nearly
identical to the Stock and Watson model, we found the following using a
1960–2010 data sample:
Structural Only
where “2” in the Taylor rule equation is both the desired inflation and
unemployment rate.
VAR Only
Structural VAR
Note the stability differences between the Structural model variable and
the VAR lagged variables.
1. The only variable significant in all four periods was the Structural
model variable.
2. Only one of the VAR variables was significant at all, and then only in
two test periods.
3. For the structural model variable, the lowest sample estimate was
only 37% below the highest estimate.
4. For the VAR variables, six of the nine variable coefficients were less
stable than the structural variable coefficient
Out-of-Sample Comparisons
How the VAR model performs compared to the structural model in
out-of-sample tests is now examined. Such tests are considered the best
100 2 METHODOLOGY
Model 2.2.4.4.7
Cowles Commission – Type Taylor Rule Structural Model
PRREAL = 0.552(INFL – 2%) – 1.322(UNEM – 2%)
(t =) (5.2) (–6.5) (2.2.4.4.7)
R2 = 0.63; MSE 1.25; DW 1.6
Model 2.2.4.4.8
Structural VAR Model
Structural VAR
1∗ : Assumes different Taylor Rule Shock each year, based on shocks that actually occurred. Equation 2.36
coefficients used, except coefficient for Taylor rule variable set at 1.00; setting it at the value in Eq. 2.36
increased average error to .94. Using coefficients from a traditional VAR, with the Taylor variable added
with coefficient of 1.00 raised the average error to 119%.
2∗ : Assumes 2001 Taylor shock is reversed in 2002; no additional shocks (Taylor Rule changes) thereafter.
(This leaves some internal inconsistency within the model: U, π held constant in Taylor rule variable, but
allowed to vary from year to year in the VAR component in accordance with the actual changes in the
economy.)
structural and SVAR models for the 10 out-of-sample years was calcu-
lated, and is presented below. Two SVAR models errors were calculated:
(1) assuming the Taylor rule shock changed each year to reflect what actu-
ally happened to inflation and unemployment; and (2) assumed the first
year shock lasted 1 year only. Results are presented in Table 2.2.4.4.2.
In the Table 2.2.4.4.2, the statistics indicate that the Cowles-type struc-
tural equation has a smaller average error of fit and smaller MSE than
either structural VAR tested. We conclude that even when adding a struc-
tural model to a standard VAR equation, the Cowles structural model
outperforms the Structural VAR, i.e., provide a better set of estimates of
out-of-sample data.
Note that with even this small model, it would be impractical to make
each equation a structural VAR. As we show in later chapters with inflation
Model 11.1.TR and unemployment Model 12.4.TR, there are at least five
additional variables driving inflation and unemployment in addition to the
three noted above. Each would need its own equation, and its own three
lags added to each equation, and that undoubtedly would add additional
variables to the model, each of which would need its own equation, and
also need its own three lags added to each equation. Each equation could
end up with 25 or more highly multicollinear explanatory variables.
In short, even if the other problems mentioned did not exist, it does
not seem feasible that the SVARs could be used to construct a large-
scale, detailed model of the economy. Hence, they would not seem to
be a suitable alternative to Cowles models for this purpose.
102 2 METHODOLOGY
2.2.4.5 Factors Which May Explain the Poor Performance of VAR Models
Sims’ equation for GDP determination was
Testing these hypotheses on the 1960–2010 data set used in this study, we
obtain the following regression estimates of marginal effects:
2.2 CHOOSING BETWEEN VAR, DSGE, AND COWLES COMMISSION MODELS 105
C = 0.663 Y (2.2.4.5.2)
I = 0.478 Acc + 1.38 Dep (2.2.4.5.3)
Suppose these are the “real” marginal effects of Y, Acc, and Dep. By
adding them together, we could deduce that
By adding C and I, this suggesting that the real total effects of Y, Acc, and
Dep are
Since the consumption and investment equations have exactly the same
determinants (and only because of this), regression of (C+I) on the three
determinants produces exactly the same results we obtained adding the
parameter estimates in equations Eqs. 2.2.4.5.5 and 2.2.4.5.6.
We assume the effect of Y on investment in the investment equation
is spuriously correlational not causal, because that’s what theory tells us
(though that would not be the case for the accelerator: Y). But in a VAR
model, purely correlational and causal effects are added together when
estimating the effects of income on (C+I), biasing the resulting estim-
ate. In addition, including the depreciation and accelerator variables in the
consumption function biases the estimate of income’s marginal effect on
C in the consumption function. By our assumptions, neither the accel-
erator nor depreciation has any effect on consumption. Yet because the
regression coefficient on income in the consumption function is (in part)
a function of the level of multicollinearity between income and these two
variables (Fox 1968), the estimate in the VAR-type consumption equation
is biased by any non-zero correlation the accelerator and depreciation have
with income. In fact, it is lower (.526) than the estimate provided by the
106 2 METHODOLOGY
It is never the case that all the coefficients in a stochastic equation are chosen
ahead of time and thus no estimation done: every stochastic equation is
estimated. In this sense, the data rule. (p. 4)
This method does not calibrate parameter estimates which best allow a
favored predetermined explanation (i.e., theory) to fit a particular set of
data. It defines as the best explanation, the one theory among many which
best explains the empirical behavior of variables over long periods of time,
under varying economic conditions. To do this, it tests hypotheses directly
expressing an underlying economic theory to help improve the likelihood
their estimated effects are not spurious, selecting the one that fits the data
best. To Cowles modelers, demand-driven models of the economy, like
those described in “old” Keynesian models, seem to best fit this bill. They
explain an uncannily large amount of the total variance of key variables
2.2 CHOOSING BETWEEN VAR, DSGE, AND COWLES COMMISSION MODELS 109
little empirical support for key assumptions underlying DSGE models that
are required for them to work, for example, current consumption spending
being based on accurate estimates of lifetime income averages, the Lucas
critique’s accuracy, etc. This makes DSGEs bad science, and therefore of
questionable value to modern economics.
In Cowles models each component of the GDP (consumption, invest-
ment, exports, demand for government goods and services) is tested to
find its determinants. Nothing is taken on faith (i.e., there is no use of
“self-evident” truths unless they have been directly tested and verified).
Each component, such as consumption, is divided into its major sub-
components (e.g., consumer durables, nondurables, services, consumer
saving and borrowing). Each subcomponent is tested separately to assess
its hypothesized determinants.
DSGE’s most innovative feature is its attempt to unify macro and
micro theory using the micro foundations method. The Cowles alternat-
ive is a “macro foundations of micro” method of deriving microeconomic
demand models for each firm and industry from a larger macroeconomic
model. A model with macro foundations, from which micro relations are
derived, makes sense because in any given period, changes in demand
for individual products, like cars, are likely to be principally driven by
macroeconomic phenomena, like booms and recessions, not by microeco-
nomic phenomena like price changes. Eckstein (1983) presents a Cowles
model that includes macro equations subsequently cut into smaller more
micro-sized sector and industry pieces, each with relative price variables to
capture microeconomic effects of price change. To the extent the supply
constraints are of interest, modern updates of Leontief’s (1951) input-
output models of the macro economy provide equivalent or greater level
of supply side micro detail, also scientifically estimated. Together they
provide a comprehensive science-based, rather than calibrated, alternative
to DSGE modeling.
In this sense, Cowles modeling provides a simple and straightforward
way of providing macroeconomic foundations for microeconomics, devoid
of the strong types of hypothesizing necessary to allow the build – up
of macroeconomic theory from microeconomic foundations, for example,
consumers and business have perfect knowledge of the future; they are
capable of doing Lagrangian maximizations in their head to determine
lifetime utility maximizations regarding how much to save and spend this
year, etc. It also avoids all the well-known tractability issues micro founda-
tions modelers face in trying to solve their models. Parameter estimates for
2.2 CHOOSING BETWEEN VAR, DSGE, AND COWLES COMMISSION MODELS 111
all levels of model aggregation can be solved using OLS or 2SLS applied
to mathematically simple hypotheses about how particular determinants
affect some dependent variable.
Past efforts to put a micro foundation under macro have met with very
limited empirical success. As noted earlier, with what some consider the
best of the micro foundations models (Smets–Wouters), a recent Federal
Reserve Board study by Edge and Gurkaynak (2011) indicated its forecasts
only were able to explain 8–13% of the variance in actual movements of
the GDP for periods as small as one to six quarters forward. In Cowles
models, the unexplained variance is more typically like this.
Cowles structural models typically explain about 90% of the variance in
key variables over time, and do so with models that explain what drove
economic behavior the 1960s or any other decade, more or less equally
well. This is true even when the data are tested in 1st differences, where
it is generally harder to explain variance than in levels. Hence, modern
versions of Keynesian structural models are very powerful in explaining
economic behavior. The difference in capacity to explain variance com-
pared to DSGE or VAR models is due to use of a theory that better
explains how consumers and businesses actually behave.
In general in such systems, the level of production is thought to be
demand driven, that human wants are endless, and that as long as people
have the purchasing power to buy more rather than less, they will do so.
We assume profit maximization provides the incentive that drives pro-
ducers to produce more in response to demand, provided there are not
supply constraints. In this sense all the great structural models of the
past including Klein’s, Eckstein’s, Fair’s, etc. are demand-driven models
of what determines the level of production. This and the belief that one
needs to know the essential determinants of demand to explain variation
in production over time accurately, constitute the essential criterion of a
Cowles model (not whether fiscal policy works or not).
Comparing DSGE models with Cowles models is to compare two
different ways of discerning empirical reality: one fundamentally is philo-
sophical and deductive, reaching its conclusions from deductions from
“self-evident” truths about the economic behavior of consumers and busi-
nesses. Deduction from self-evident truths was a methodology used to
determine “truth” and “reality” by natural and moral philosophers before
the scientific method was available. Cowles methods are more modern:
they are inductive and scientific. Consumption and investment functions
used in GDP models are those obtained from testing. Eckstein (1983), an
112 2 METHODOLOGY
investors who merely want to trade in current investments so they can buy
others which currently look more profitable. In doing so, the open market
process causes the market valuation of securities to rise, but aside from the
small mpc out of wealth, or business spending out of proceeds from newly
issued securities, does not affect the real GDP. If this hypothesis is correct,
much of the additional money created by the open market process merely
increases money chasing investments, causing the stock and bond markets
to rise, not in buying goods and services that raise the real GDP. As an
explanation as to why monetary policy fails to stimulate the real economy,
this explanation is considered nearly an article of faith by much of the busi-
ness press. But it is not well researched by economists. It is a topic which
deserves future additional investigation.
Open market purchases of bonds from banks does increase free reserves
held by banks. However, the reserves may just stay in the bank in poor
economic times, failing to result in increased real spending when it is most
needed. This also may be part of the reason we empirically find little rela-
tionship between monetary policy and the real economy. The model tested
in this paper suggest monetary policy principally can affect the real eco-
nomy – consumption and investment – through its effect on the prime
interest rate and the level of consumer borrowing, particularly for housing,
but the measured effects are small.
2.2.6 Conclusions
Our tests indicated the out-of-sample performance of Cowles models was
far better than that of VARs and DSGEs. Out-of-sample performance test-
ing is considered by many to be the gold standard for evaluating the
usefulness of models.
Tests showed Cowles models have greater explanatory power than
VARs or DSGEs. The average error with which the Cowles models fit
the data in the decade after estimation was only half that of DSGE mod-
els, and 1/6 that of VARs. The larger VAR error likely results from two
problems:
For all these reasons we have chosen the Cowles methodology for use in
developing this paper’s large-scale econometric model of the US economy,
presented in Chapters 4–19 below.
CHAPTER 3
Literature Review
and investment spending that the Keynesian models do. Also, because of
its atheoretic VAR approach to some aspects of model building, the eco-
nomic meaning of some equations can be hard to interpret. Nonetheless, it
does reflect dominant current trends in macroeconomic theoretical mod-
eling. Hence, it is important to compare the variables found important in
predicting and explaining consumption and investment in such models, to
those using more traditional methods having greater explanatory power.
The statistical technique used to estimate the equations was 2SLS if there
was evidence of endogeneity. Instruments developed to replace endo-
genous variables were constructed using factor analysis to develop 12
orthogonal (no multicollinearity) principal components for use in the
instrument instead of using individual variables, as is more common today.
Klein and Even’s model uses Almon and Koyck lags in investment and
consumption models to show lags (e.g., for a construction project which
affects several future years’ GDP). Such lags are commonly used as a way of
avoiding the problem of multicollinearity distorting individual year effects,
when several lagged values of an explanatory variable are required to show
its full influence on the current value of the dependent variable.
Key parts of the complete Klein-Evens model include
Three consumption models
3.1 LAWRENCE KLEIN AND MICHAEL EVANS (1968): THE WHARTON . . . 117
• Manufacturing
• Mining and regulated (utility) industries
• Commercial (office buildings, stores, shopping centers)
• Housing
• Inventory (manufacturing)
• Inventory (nonmanufacturing)
• Food imports
• Raw materials and semi-finished goods
• Manufactures and services
All models analyzed below are given in real terms, unless otherwise noted.
where Xwt = index of world trade; Pwt = price of world trade, and
Pe = implicit deflator for exports.
Klein’s export demand is defined as a function of the level of world
trade, average prices in world trade relative to our export prices, and the
average level of our exports the past four years. It is not likely these are
causal factors; causally, we would expect exports to be driven more by
our trading partners income levels, exchange rates, and to the extent that
current trade is needed to give foreigners cash to buy U.S. exports, our
import levels may be a factor.
where Y/N = disposable income per capita, Pif = deflator for food
imports, and Pf = prices received by farmers. Demand for crude and food
imports is positively related to disposable income and negatively the price
of these imports relative to domestic prices for agricultural goods.
. . . The basic rationale for continuing to use structural models in the face
of the rational expectations criticism is this: changes in policy regime seem
to have been among the minor sources of structural change of the eco-
nomy and of forecasting error in the actual historical record. The principal
obstacles to structural consistency and forecast accuracy seem to lie in the
exogenous shocks of wars and OPEC, and in the unpredictability of the
exogenous monetary policy variables. The central assumption of the rational
expectations school, that the forecasts on which businesses and households
make their decisions are free of bias and that markets clear instantaneously,
so far do not seem confirmed by the historical record (Eckstein 1983, pp.
xi–xii) and
. . . the DRI model shows a nervous consumer, subject to the risk
of an unstable environment acting on an exposed financial position,
forming “permanent” expectations rather quickly and acting to a con-
siderable degree in response to short term changes. This is a far
cry. . . from the life-time permanent income theories of stable spending . . . .
(Eckstein 1983, p. 113)
126 3 LITERATURE REVIEW
This model finds per capita demand for furniture and appliances to
be positively related to per capita permanent income, the ratio of cur-
rent disposable income to permanent disposable income, the index of
consumer sentiment, and net wealth. Demand was negatively related
to growth in the price of these goods relative to all consumer
prices.
128 3 LITERATURE REVIEW
This model finds per capita demand for autos to be a positive function
permanent income, the ratio of current income to permanent income, and
the index of consumer sentiment. Demand for autos was negatively related
to the price of autos relative to all consumer goods and the existing per
capita level of car ownership.
housing, and the cost of major home owner utilities – gas and elec-
tric. (It is surprising that income and wealth variables did not prove
significant.)
methodology. Eckstein was also aware of the growing interest in VAR and
DSGE modeling, and argued against them. He offered evidence in favor
of his Cowles-type model with arguments commonly still used today when
comparing model strengths and weaknesses the three types of large-scale
modeling.
That said, Eckstein always considered his models a work in progress,
and there are a few remaining equations (though not many) with explan-
atory variables that may confuse directions of causation, whose estimates
have the wrong signs, or use a Koyck distributed lag to push forward
a past empirical trend for which the theoretical basis is not obvious. In
addition he heavily used formulations of permanent income hypothesis
(averages of several year’s income) as the income variable driving con-
sumption, without indicating if this was the result of an empirical finding,
or just an off-the-shelf theoretical argument assumed to be valid. Our
own tests have indicated current income alone explains more variance in
consumption.
Eckstein’s DRI model has export and import sub-models, but they are
not included in his 1983 book describing the model. Also, no models of
interest rates are presented.
Fair found nominal (vs. real) interest rates explained variation in con-
sumption and investment models better, i.e., interest rate effects were
more statistically significant. His tests were complicated (see p. 71) and
are not repeated here. For comparison, we tested standard consumption
and investment equations in this study’s models two ways; one using the
real prime interest rate, then retesting the exact same model except using
the nominal prime interest rate. Results for OLS first difference models
are presented below:
Real does substantially better for consumption; nominal does better for
investment (more marginally). Our practice will be to use real interest rates
for investment as well as consumption; they are more theoretically sensible
and only give marginally different results in terms of their ability to explain
variance.
the largest segment), the previous year’s spending on services, per cap-
ita disposable income (YD/POP), the after tax three-month treasury bill
rate-negatively (RSA), per capita wealth levels (AA/POP), and a time
trend.(T).
CD = consumer durables
CN = consumer nondurables
CS = consumer services
IHH = residential investment
AG1 = % of 16 + population 26–55 minus % 16–25
AG2 = % of 16+ population 56–65 minus % 16–25
AG3 = % of 16+ population 66+ minus % 16–25
POP = noninstitutional population 16+
3.3 RAY FAIR’S ESTIMATING HOW THE MACROECONOMY WORKS (2004) 135
YD = disposable income
RSA = after tax bill interest rate (% points)
AA = total net wealth
RMA = after tax mortgage interest rate
PH = price deflator for CD, CN, CS, and IHH
KD = stock of durable goods
CDA = peak-to-peak interpolation of CD/POP
KH = stock of housing
IHHA = peak-to-peak interpolation of IHH/POP
RHO = autocorrelation control
DELH = housing depreciation rate
DELD = durables depreciation rate
AA = net wealth
V = 0.76V–1 + 0.30Y–1
where PF = price deflator for (total sales-farm GDP) and PIM = price
deflator for imports.
Demand for imports (per capita) is given as a positive function of last
year’s imports level, total consumption, and investment demand (which
includes demand for imports) and the relative growth in domestic prices
compared to imports. Dummies are used to represent unique events in
1969 and at the end of 1971 and beginning of 1972.
where
Y = real private sector production, derived from Log Y
Fair did not just enter all the variables he found to be determinants
of consumption, investment, and imports into one IS curve regression
3.3 RAY FAIR’S ESTIMATING HOW THE MACROECONOMY WORKS (2004) 139
variables can add to this problem, though they may be useful in forecast-
ing. The lagged dependent variable’s own determinants, lagged one more
year, are the same explanatory variables as are explicitly included in the
current year model along with the lagged dependent variable. Including
the lagged dependent variable as explanatory just indicates the modeler’s
belief that not all changes in an explanatory variable are felt in the year the
change occurs; some are not felt until the following year. This substantive
message may be clearer if these explanatory variable lags were added to the
model, as well as the more current values already included.
Consumption was hypothesized as a determinant of imports; though
imported consumer goods and services are usually counted as part of total
consumption, so while use of the variable explains variance, it does not
explain much about what makes import demand vary. An income variable
might work better since it is generally considered the major determin-
ant of consumer demand. We also noticed the depreciation model did
not include prior years’ investment levels as explanatory variables. Prior
year investment generally determines allowable percentage write offs as
depreciation from year to year.
The model explains slightly more than half the variation in the change in
the log of total consumption. This is considerably less than was the case for
most Cowles structural models examined, where R2 s in the 0.80–0.95 range
are more common not only for total consumption, but for its individual
components including any equations estimated in logs. Estimation in logs
is the standard practice in the FRB/US and other DSGE models, but less
common in Cowles models.
How should we interpret all this? The guide to the FRB/US indicates that
Dynamic adjustment:
cdv = – 0.30(cdv–1 – c∗ dv–1 ) – 0.28lags1 (cdv–i ) + 3.22Leads (c∗e dv+i )
+ 7.46Lags4 (c∗e dv+i ) R2 = 0.43; (No t-stats.provided)
cdo = – 0.10(cdo–1 – c∗ do–1 ) + 0.17lags1 (cdo–i ) + 2.15Leads (c∗e do+i )
+ 1.12Lags4 (c∗e do+i ) R2 = 0.34; (No t-stats.provided)
ih = – 0.09(ih–1 – i h–1 ) + 0.38lags1 (ih–i ) + 6.10Leads (i∗e h+i )
∗
Both producer durables and inventories are a function, first, of the rate of
adjustment of current levels of them to desired levels. In addition current
period changes are found to be linear projections of past and anticipated
rates of change each quarter in these variables. From a substantive per-
spective, projecting future change from past changes amounts to arguing
that there are no substantive determinants of changes in spending on pro-
ducer durables or inventories, only the inertia of past trends in spending
growth. The role of future changes in investment in determining the cur-
rent period change is not so obvious. The theoretical credibility of the first
term, i.e., that the desired change this period depends on prior decisions
about how quickly to try to move from current levels to desired levels
hangs on how reasonable it is to assume that desired levels of invest-
ment = current output minus current price of producers durables plus
depreciation plus 19.5 times the accelerator. This would seem to be an
empirical question not addressed in the model.
Exports, Imports, and Interest Rates
The new FRB/U.S. model of the economy has export, import, and
interest rate equations, but these were not included in the Federal
Reserve’s “A Guide to FRB/US”: A Macroeconomic Model of the
United States (1996)
PR = real prime interest rate defined as the nominal rate minus the average
of the past two completed years inflation
DJ–2 = a measure of wealth (NYSE Composite Index), lagged two years
(Federal Reserve/MIT Large econometric model showed stock mar-
ket activity to be a major determinant of household wealth, which in
turn was a determinant of consumer spending. See Eckstein 1983, p. 5)
XRAV = the real broad U.S. exchange rate average for current and past three
years (foreign currency per dollar)
POP16 = ratio of young (20–24) to old (65+) in population
POP = population size
ICC–1 = Index of Consumer Confidence (Conference Board measure), lagged
one year
M2AV = real M2 money supply; average of second, third and fourth past years
M2-M1 = savings components of M2 for the current period
CB , CB2 = consumer borrowing (annual change in consumer debt). (Federal
Reserve/MIT model indicates credit access a major determinant of
housing demand (Eckstein 1983, p. 5))
Above, as well as in tests below, current period values are typically denoted
without subscript. Other subscripts indicate periods lagged.
Variables in the consumption model such as disposable income or the
deficit variables (T, G) may be endogenous with consumption since con-
sumption is a part of GDP and GDP is a determinant of all the models
tested. Explanatory variables “suspected” a priori as being endogen-
ous with the dependent variable in consumption spending or borrowing
models are given in Table 4.0.1.
Endogeneity of these “suspect” variables was tested using the Hausman
endogeneity test.
All other explanatory variables beside those listed in Table 4.0.1, used
in either the consumption function or in standard investment models in
Chapter 5, which follows, were assumed to be exogenous, or only affect
• Disposable income
• The government deficit (T, G)
• The exchange rate
• The prime interest rate
• Consumer borrowing
• The exchange rate average
THE CONSUMPTION MODELS 149
Model 4.1
OLS Standard Consumer Spending Model, with Borrowing
Included as a Determinant of Consumer Spending (No Variable
Found Endogenous, So No 2SLS Models)
Model 4.1.T
OLS Standard Consumer Spending Model, with Borrowing
Included as a Determinant of Consumer Spending;
Stationarity Issues Resolved
(No Variable Found Endogenous, So No 2SLS Model Needed)
Model 4.1.T(1VarDef)
OLS Standard Consumer Spending Model, with Borrowing
Included as a Determinant of Consumer Spending (No 2SLS
Models; 1Variable Definition of Deficit Used)
400
300
200
100
60 0
40 –100
20 –200
0
–20
–40
–60
1960 1965 1970 1975 1980 1985 1990 1995 2000 2005 2010
Residual Actual Fitted
Fig. 4.1.1 Actual consumption compared to levels calculated from Model 4.1.T
1960–2010
154 4 THE CONSUMPTION MODELS
out does not mean pulling all its influence out. Stepwise is an illumin-
ating, but imperfect, tool. For the same reason, first-in stepwise tends
to overstate contributions. As such, stepwise results can provide some
information on variable importance, but either approach alone is con-
sidered a definitive measurement of a variable’s contribution to explained
variance.
Table 4.1.2 tests Model 4.1.T in four sample periods to determine if
coefficients are stable from sample-to-sample.
Results are quite robust for all periods sampled. The similarity of effect
of fiscal policy variables in different historic time periods strongly contra-
dicts the Lucas critique notion that how the economy reacted to past
policy changes is no guide to how it will react to the same changes in
the future. As the four samples above show, clearly you can use the results
in one sample period as a reliable guide to the likely results in another, for
both fiscal policy and other variables. Eckstein (1983), who also tested the
Lucas critique, also could not find any support in the data for it. Hence,
simulation of policy changes in the future should produce reliable guides
to the actual effects that will occur, at least if the estimate of past effects is
obtained from a Keynesian-style, demand-driven structural model like Eq.
4.1.T above or other similar structural models used in other consumption
models in this study.
Table 4.1.2 Robustness over time – (2SLS detrended model, Eq. 4.1.T)
Model 4.1.T.TR
Time Period Robust Standard Consumption Model
Model 4.1.T.TRa
Time Period Robust Standard Consumption Model
(Two Variables Deleted)
Coefficients and significance levels are robust for all the remaining
variables.
We have already seen that dropping the POP16/64 and exchange
rate variables left the remaining parameter estimates nearly unchanged,
so adding it back in would do the same. Adding the real M1money sup-
ply variable to the sample period robust model again leaves coefficient and
significance level estimates stable.
4.2 SPENDING ON IMPORTED CONSUMER GOODS – OLS ESTIMATES 157
Model 4.1.T.TRb
Time Period Robust Standard Consumption Model
(One Variable Added)
CT = 0.49(Y – TT ) + 0.56(TT ) – 0.36(GT& I ) – 9.74PR
(t =) (11.4) (12.3) (–7.6) (–5.1)
+ .41DJ–2 + 0.017POP + 0.41ICC–1 + 49.30M2AV
(5.8) (4.3) (1.2) (4.9)
+ 0.12 CB2 – 0.15 M1Real R2 = 96.0% D.W. = 1.7 MSE = 22.46
(4.0) (–2.1) (4.1T.TRb)
We conclude that estimates of effects of other variables on consumption
in the original model are robust to specification changes in the model.
Since the full sample period robust model (4.1T.TRb), from a theoret-
ical perspective, is better specified than the truncated alternatives or the
alternative with the added variable, we take the results for 4.1T.TRb as
robust to model specification changes as well as changes in the time period
sampled.
Model 4.2
Consumer Imports Model – Exports Added
CM = 0.09(Y – TT ) + 0.13(TT ) – 0.08(GT& I ) – 2.09PR
(t =) (1.9) (3.3) (–1.8) (–1.3)
+ 0.11DJ–2 + 3.13XRAV + 239.00POP16 – 0.002POP
(1.8) (3.3) (2.1) (–1.2)
(4.2)
– 0.14ICC–1 + 7.16M2AV + 0.12 CB2 + 0.47 X
(–0.6) (1.1) (2.4) (5.7)
R2 = 86.7% D.W. = 2.2 MSE = 17.65
Explained variance First-out stepwise method First-in stepwise method (constant added)
(R2 = 0.87 to start) (R2 = 0.00 to start)
Periods sampled
Variable 1960–2010 1970–2010 1970–2000 1960–2000
those initially found significant at least three of four samples, they are kept.
A new model is run with all these variables and retested in the four sample
periods. If the variables that initially failed the three out of four tests now
are significant in at least three of four, they, along with the variables that
initially passed the three out of four tests, become our final, sample period
robust model (Eq. 4.3.TR).
One variable originally insignificant (the prime rate) was added to the
five “core” variables originally found significant in three out of four tests
(income, tax and spending deficits, the exchange rate, and the stock mar-
ket index variable). With the new variable in, the stock market variable
was no longer significant in at least three of four time periods sampled
and was dropped from the “final” time period robust model given in
Eq. 4.3.TR.
Model 4.2.TR
Time Period Robust Consumer Imports Model
CM = 0.19(Y – TT ) + 0.25(TT ) – 0.18(GT& I ) + 2.22XRAV
(t =) (7.8) (7.4) (–5.4) (2.6)
– 3.06PR0 R2 = 76.7% D.W. = 1.5 MSE = 21.55
(–1.8) (4.2.TR)
4.3 SPENDING ON IMPORTED CONSUMER GOODS – 2SLS ESTIMATES 161
Model 4.2.TR.a
Time Period Robust Consumer Imports Model
(One Variable Deleted)
CM = 0.19(Y – TT ) + 0.23(TT ) – 0.15(GT&I ) + 2.20XRAV
(t =) (7.5) (6.5) (–5.1) (2.2)
R2 = 75.4% D.W. = 1.4 MSE = 21.89 (4.2.Alt.TR.a)
Remaining coefficients are robust as are significance levels.
Adding the population and consumer confidence variables to the sample
period robust model (4.2.TR) and re-estimating, we obtain
Model 4.2.TR.b
Time Period Robust Consumer Imports Model
(Two Variables Added)
CM = 0.22(Y – TT ) + 0.26(TT ) – 0.18(GT& I + 2.05XRAV )
(t =) (6.6) (6.7) (–4.4) (2.3)
– 2.35PR0 . – 002POP – 0.25ICC–1 –
(–0.9) (–1.0) (–0.6)
2 (4.2.TR.b)
R = 77.7% D.W. = 1.6 MSE = 21.53
Results are very robust to these model changes, with the exception of the
interest rate variable, whose magnitude is reasonably stable, but which now
becomes insignificant.
We conclude that estimates of effects of other variables on consumption
of imports in the sample period robust model are also robust to specifica-
tion changes, with the exception of the interest rate variable’s significance
level in the model. Hence, Eq. 4.2.TR becomes our final sample period
and model specification robust model of what affect consumer purchases
of imports, and how much.
Model 4.4
OLS Standard Consumer Spending Model,
with Borrowing Included as a Determinant of Consumer
Spending (No Variable Found Endogenous, So No
2SLS Models)
first-out results indicate disposable income and tax cut deficits and popula-
tion growth were the most important factors affecting consumer spending
in the 1960–2010 period. The first-in results suggest disposable income,
tax deficits, consumer confidence, and consumer borrowing. First-out
stepwise results tend to understate contributions because the omitted
variable typically shares some ability to explain variable with a remaining
variable. For the same reason, first-in stepwise tends to overstate contribu-
tions. As such, stepwise results can provide some information on variable
importance, but are not considered definitive measurements of a variable’s
contribution to explained variance.
Robustness Over Time
To test for replicability of the initial results, the model was tested in four
different, but overlapping time periods. Findings are shown in Table 4.4.2.
Results are quite robust for at least three of the four sample periods
sampled for all variables except two: the exchange rate and consumer
confidence. Hence our model robust to period tested will include all
variables except those two. Since multicollinearity or insufficient degrees
of freedom can cause significant variables to appear insignificant in models
with a large number of explanatory variables (like the models above), we
also add back, one at a time, the variables not found significant in at least
three sample periods, and re-estimate. Variables now found significant
are included in the final time period robust model if they now prove
significant in three of four sample periods when tested in this model.
No additional variables met this criterion. The final sample period robust
model is shown in Eq. 4.4.TR.
Model 4.4.TR
Domestically Produced Consumer Goods – Time Period
Robust Model
Model 4.4.TR.a
Domestically Produced Consumer Goods – Time Period
Robust Model
(Two Variables Deleted)
Model 4.4.TR.b
Domestically Produced Consumer Goods – Time Period
Robust Model
(Two Variables Added)
CD = 0.29(Y – TT ) + 0.31(TT ) – 0.20(GT& I ) – 7.02PR
(t =) (6.1) (5.7) (–3.7) (–2.5)
166 4 THE CONSUMPTION MODELS
Findings for the initial model tested, which included all variables found
significant in at least one previous consumer spending study reviewed,
4.5 DETERMINANTS OF CONSUMER BORROWING – OLS ESTIMATES 167
are presented in Model 4.5 The final model, robust to changes in time
period sampled and the number of other variables included in the model,
is presented in Model 4.6.TR further below.
Model 4.5
OLS Standard Consumer Borrowing Model, Savings Variable Added
(XR Variable Found Endogenous; See Also 2SLS Model Below)
Model 4.6
2SLS Model of Eq. 4.5; Exchange Rate Endogenous and Replaced
with Strong Instrument
Table 4.6.2 Robustness over time – consumer borrowing, 2SLS Model 4.6
Deleting the variables in the initial model that were not significant in
at least three of the four sample periods gives the core components of a
model that is robust to time periods sampled.
Because significant variables can sometimes show as insignificant
because of multicollinearity or inadequate degrees of freedom, we read-
mit the ones deleted, one by one, to the core robust model, and then
re-estimate. Previously deleted variables that are now found significant
are also included in the sample period robust model, provided only that
they now prove to be significant in three of the four sample periods when
retested. No additional variables met this criterion. The final time period
robust model is shown in Eq. 4.6.TR:
Two other adjustments were made. (M2–M1) and the (PerSav) vari-
ables, both of which measure consumer savings, were combined. Doing
so markedly improved the statistical significance. Also, as it left the
population variable insignificant in 3 out of 4 time periods, it was
dropped. Serious multicollinearity of personal savings and population
growth (r = 0.70) seemed the underlying reason for the population vari-
able having the wrong sign and both the personal savings variable and the
savings components of the M2 variable (M2-M1) were found statistically
insignificant.
Model 4.6.TR
2SLS Standard Consumer Borrowing Model – Time Period Robust
Though our sample-period robust findings are strong, they still reliably
explain only 55.3% of all the causes of variation in consumer borrow-
ing. Clearly important factors that affect borrowing are not uncovered
yet. More research is needed.
Robustness to Model Specification Changes (1960–2010 Data Set):
Dropping the last two variables in Eq. 4.6.TR and re-estimating, we have
4.7 MODELING THE MAJOR COMPONENTS OF TOTAL CONSUMPTION 171
Model 4.6.TR.a
2SLS Standard Consumer Borrowing Model – Time Period Robust
(Two Variables Deleted)
Coefficients and statistical significance are robust for all variables left in
except the coefficient on the income variable.
Finally, adding the Population ratio of young to old, and the consumer
confidence variables to the time period robust Model 4.6 gives:
Model 4.6.TR.b
2SLS Standard Consumer Borrowing Model – Time Period Robust
(One Variable Added)
Coefficient change is minor. The time – period robust model is also robust
to the addition of these variables to the model.
Overall, we conclude the coefficients in the consumer borrowing model
given by Eq. 4.6.TR are robust to significant changes in both sample
period tested and changes in model specification.
Model 4.8
OLS Standard Consumer Durables Spending Model, with
Borrowing Included as a Determinant
(Four Variables Found Endogenous, See Also 2SLS Model Below)
Generally results are the same as with the OLS model, except the interest
rate’s coefficient changed considerably, as did the M1 variables’ coefficient
and significance level.
In first differences, the consumer durables variable was stationary. Two
of its nine right-hand-side hypothesized determinants were nonstation-
ary: disposable income, and the wealth variable (stock market average).
Both were cointegrated with the dependent variable, so no detrending
was needed. Standard consumption function variables like POP16, POP,
consumer confidence, M2AV are not included because they were found
nonsignificant in preliminary testing.
Table 4.9.2 Robustness over time – consumer durables, 2SLS Model (Eq. 4.9)
Five of ten variables were significant in all four periods sampled (dis-
posable income, tax and spending deficits, the exchange rate, and the level
of nondurables purchased). Two others significant in two sample periods
including the 2001–2010 decade, when we had a housing boom which
drove up demand for durables like kitchen appliances. They were the level
of residential investment and the level of house prices relative to income.
The large growth in the stock market during the same decade also con-
tributed to its significance then, but not in the two samples ending in
2000. The prime rate and the M1 money supply were not significant in
any period.
The five significant variables explain 82% of the variance and form the
core of our sample period robust model given in Eq. 4.9.TR. However,
in large models, a variable that truly is significant can appear insignificant
because of multicollinearity or too few degrees of freedom. To ensure we
haven’t rejected truly explanatory variables for this reason, we retest each
rejected variable separately with the core model. The additional variables
found significant this way are then added to the core robust model and
again tested in all four time periods. Only if significant in at least three are
they included in the time period robust model, Eq. 4.9.TR. No additional
variables were found significant in at least three of the four tests. The only
one to come close was the Housing Price/GDP ratio variable, which was
significant in samples including the 2001–2010 decade, as before, but not
in the two samples ending in 2000. We interpret this as meaning it takes
major growth in the ratio to have effect on durables purchases distinct
from that caused by other forces affecting durables demand. The final time
period robust model is given in Eq. 4.9.TR.
Model 4.9.TR
Time Period Robust Standard Consumer Durables Model
Model 4.9.TR.a
Time Period Robust Standard Consumer Durables Model
(One Variable Deleted)
Model 4.9.TR.b
Time Period Robust Standard Consumer Durables Model
(One Variable Added)
The time period robust model coefficients and significance levels remained
stable (using current year values of M1 did not make a difference). Hence
Model 4.9.TR appears robust to both sample periods tested as well as
(reasonable) changes to the model specified.
the expectation that the more expensive housing was, the less would be
bought, and the more consumers would shift spending to nondurables),
and the level of spending on durable goods (on the expectation that the
more spent on durables, the less available for spending on nondurables).
Use of both the consumer durables and residential investment variable (I),
the same model tended to leave one insignificant, though both were sig-
nificant if used in the model without the other. As we had seen earlier,
the two are related. Hence, both were picking up much the same vari-
ance. The consumer durables variable was dropped. It seemed the weaker
determinant of nondurable goods demand because the investment variable
has higher levels of statistical significance when used alone. This suggests
that while the demand for durables resulting from residential investment
was substantial, other factors were also at work, namely consumer spend-
ing on nondurables dropping as demand shifts to a desire for new housing.
Results are shown in 4.10 below.
The final model, robust to changes in time period sampled and the
number of other variables included in the model, is presented in Model
4.11.TR further below
Model 4.10
OLS Standard Consumer Spending Model, with Borrowing
Included as a Determinant
(Y-T Variable Found Endogenous, So See 2SLS Model Below)
Model 4.11
2SLS Estimates of Spending on Consumer Nondurables
(Hausman Tests Indicate TT and CDur were Endogenous)
Table 4.11.2 Robustness over time – nondurables, 2SLS Model (Eq. 4.11)
The model was tested in four different but overlapping time periods to
determine if initial results could be replicated in different time periods.
Findings are shown in Table 4.11.2 .
Results are quite robust for all variables except residential investment,
the M2 money supply and the stock market index. Both are significant
for samples including 2001–2010, but not for the two samples ending in
2000.
Re-estimating the model using only the six variables found signi-
ficant in three of the four tests, gives the core sample period robust
model. To avoid having discarded variables not meeting this criteria
for statistical, not substantive reasons, e.g., too few degrees of free-
dom for the number of explanatory variables used in the initial large
model, or multicollinearity problems, we add back the discarded vari-
ables one at a time and re-estimate. All that are now found significant
we add to the core model, provided they are significant in three of the
four sample periods. No additional variables met this condition. The
final sample period robust model resulting from this process is given in
Eq. 4.11.TR
Model 4.11.TR
2SLS Estimates of Spending on Consumer Nondurables – Time
Period Robust
Model 4.11.TR.a
2SLS Estimates of Spending on Consumer Nondurables – Time
Period Robust
(Two Variables Deleted)
CND = 0.23(Y – TT ) + 0.13(TT ) – 0.06(GT& I ) + 9.73MORTRREAL
(t =) (9.7) (5.8) (3.1) (4.1)
R2 = 76.2% D.W. = 1.5 MSE = 16.24 (4.11.TR.a)
Model 4.11.TR.b
2SLS Estimates of Spending on Consumer Nondurables – Time
Period Robust
(One Variable Added)
CND = 0.24(Y – TT ) + 0.20(TT ) – 0.14GT& I + 8.20MORTRREAL
(t =) (11.7) (5.8) (4.2) (2.0)
– 2.73PR + 0.21IRes + 0.09M1Real
(–2.6) (4.9) (0.8)
R2 = 80.3% D.W. = 1.3 MSE = 15.30 (4.11.TR.b)
Stationarity
The dependent variable was stationary. Three RHS variables in Model 4.12
were nonstationary disposable income, government spending, and the
stock market average but all were cointegrated with the dependent
variable.
Model 4.12
OLS Consumer Services Spending Model
Table 4.13.2 Robustness over time – consumer services, 2SLS model (Eq. 4.12)
Only one, the ratio of house prices to GDP), was only significant in only
two samples: those including the 2001–2010 decade.
A sample period robust model is presented below. It includes a core
model with the six variables found significant in all four sample periods and
the two found significant in three periods. The final core model is shown
in Eq. 4.12.TR, with results shown for the 1960–2010 period testing.
Model 4.12.TR
OLS Estimates of Spending on Consumer Services – Time Period
Robust Model
Model 4.12.TR.a
OLS Estimates of Spending on Consumer Services – Time Period
Robust Model
(Two Variables Deleted)
Model 4.12.TR.b
OLS Estimates of Spending on Consumer Services – Time Period
Robust Model
(Two Variables Added)
Eckstein’s findings indicate crowd out more than completely offsets stim-
ulus effects of fiscal policy. The greater than 100% crowd out effect
presumably stems from the discrete nature of housing borrowing: when
government borrowing to finance deficits reduce the amount banks have
to lend to consumers by, say, $10,000, the bank has to tell someone who
needs to borrow, say, $100,000 they can only get $90,000. This may cause
a drop in private spending of $100,000, while only providing a government
stimulus of $10,000. Eckstein (p. 37) also noted that “. . . real crowding out
occurs even when the economy is far from the full employment ceiling . . . ”.
This study’s findings are consistent with his, and underscore the importance
of including the deficit variable, a measure of the magnitude of crowd out).
PR = real prime interest rate defined as the nominal rate minus the average of
the past two completed years inflation.
DJ–2 = a measure of wealth (NYSE Composite Index), lagged 2 years (other lags
also used).
XRAV = the real broad U.S. exchange rate average for current and past three years
(foreign currency per dollar).
POP = population size.
M2AV = real M2 money supply; average of second, third, and fourth past years.
M2-M1 = savings components of M2 for the current period.
IB , = business borrowing (annual change in business debt).
Prof–2 = business profits, lagged 2 years (other lags also used).
DEP = depreciation variable.
CAP–1 = % of productive capacity utilized, lagged 1 year.
Model 5.1
OLS Estimates of Determinants of Total Investment Spending
Four variables in the model were nonstationary, but were all cointeg-
rated with the dependent variable, which was stationary. The nonstation-
ary variables were government spending, depreciation, the Tobin’s q proxy
and the population size variable.
Model 5.2
2SLS Estimates of Determinants of Investment Spending
Hausman tests indicate one variable endogenous with investment spend-
ing (DJAV), so Instrumented: (Wald test F = 9.1, but DJ–2 t = 3.5)
and Sargan test NR2 = 6.12 < X2(95,15) = 25.00. Hence, this is a strong
instrument model.
In the 2SLS model, the lagged prime rate and the stock market index were
significant, unlike the OLS model. But the profits and borrowing variables
which were significant in the OLSS model, were not in the 2SLS model.
Table 5.2.2 Robustness over time – total investment, 2SLS Model 5.2
accelerator and the stock market variable, which serves as our imperfect
proxy for Tobin’s q, and the business borrowing variable. Also important
were the two types of deficit and the level of profits.
Model 5.2.TR
2SLS Estimates of Determinants of Total Investment
Spending – Time Period Robust
Model 5.2.TR.a
2SLS Estimates of Determinants of Total Investment
Spending – Time Period Robust
(Two Variables Deleted)
Model 5.2.TR.b
2SLS Estimates of Determinants of Total Investment
Spending – Time Period Robust
(Two Variables Added)
Model 5.3
OLS Estimates of Determinants of Spending on Domestically
Produced Investment Goods
2SLS results are also presented further below since the average exchange
rate and depreciation variables were found endogenously related to
spending on domestically produced investment goods.
5.4 2SLS ESTIMATES OF THE DETERMINANTS OF DOMESTICALLY PRODUCED. . . 193
Model 5.4
2SLS Estimates of Determinants of Spending on Domestically
Produced Investment Goods
Hausman tests indicate two variables endogenous with investment spend-
ing, so instrumented: XRAV (XRAV(–1) t = 8.0), DEP (DEP–1 t = 7.9), so
this is a strong instrument model. Sargan test NR2 = 5.64 < X2(95,16) =
26.30. Hence, the instrument is not endogenous.
ID = + 0.25(ACC) + 0.29(TT ) – 0.31(GT&I ) + 0.08DEP
(t =) (6.7) (2.7) (–3.7) (0.3)
+ 2.60CAP–1 – 3.01PR–2 – 0.19DJ0 + 0.08PROF–0
(1.5) (–1.9) (–0.8) (1.2) (5.4)
+7.08XRAV + 0.011POP + 0.02(BOR–1 )
(2.2) (3.4) (0.3)
R2 = 84.4% D.W. = 2.0 MSE = 28.77
To allow later use in Chapter 16, the same model is re-estimated using only
1960–2000 data. This will allow out-of-sample testing of its fit during the
2001–2010 period. Re-estimation yields the following.
Model 5.4.TR
OLS Estimates of Determinants of Spending on Domestically
Produced Investment Goods
(Time Period Robust Model)
Model 5.4.TR.a
OLS Estimates of Determinants of Spending on Domestically
Produced Investment Goods
(Time Period Robust Model, with Two Variables Deleted)
Model 5.4.TR.b
OLS Estimates of Determinants of Spending on Domestically
Produced Investment Goods
(Time Period Robust Model, with Two Variables Added)
The time robust model variables’ signs, coefficients, and significance are
robust to these additions.
Hence we conclude the model given by Eq. 5.4.TR is robust to both
time period tested and changes in the exact way the model is specified.
(The business borrowing variable is highly collinear with the deficit is
the likely reason for its lack of significance, since it is very highly posit-
ively correlated with government receipts and highly negatively correlated
with government spending (r = .78) when the two are run as one defi-
cit variable. In short, government tax and spending deficits are strongly
related to declines in business borrowing as we might expect due to
crowd out.)
5.5.1 Nonstationarity
For investment imports, five of the explanatory variables were found
nonstationary, but were found cointegrated with total investment, the
dependent variable. The nonstationary variables were government spend-
ing, depreciation, the stock market index, profits, and population
growth.
Model 5.5
OLS Estimates of Determinants of Spending on Imported
Investment Goods
(Several Variables Endogenous; See Section 5.6 for 2SLS Model)
Model 5.6
2SLS Estimates of Determinants of Spending on Imported
Investment Goods
Hausman tests indicate three variables endogenous with investment spend-
ing on imports: the accelerator, profits, and DJAV, The instrument used
was a strong instrument for the first two, but not for the third. Sargan test
NR2 = 6.24 < X2(95,15) = 25.0. Hence, the instruments are not endogenous.
Results indicate the key factors affect the demand for imported investment
goods positively are the deficit, interest rates, the stock market, profits, the
exchange rate, and access to borrowing.
5.6 2SLS ESTIMATES OF THE DETERMINANTS OF IMPORTED INVESTMENT GOODS 199
small amount for much of the period sampled), or because many of the
initial findings were spurious and the model fails to include important
determinants. More research is needed.
The core of our time period robust model will be the prime rate
and stock market variables. When using our usual procedure of adding
variables individually to the robust model, we noted the accelerator was
significant in all periods. We then added the rest individually to this two
variable “core” robust model and found most – individually added as a
third variable – were significant. However, when all those successful in this
way were added together to the model, only two were significant. Elim-
inating the rest left the Eq. 5.6.TR as our final time period robust model,
with each variable being significant in at least three of the four periods
sampled:
Model 5.6.TR
2SLS Estimates of Determinants of Spending on Imported
Investment Goods
(Time Period Robust Model)
Model 5.6.TR.a
2SLS Estimates of Determinants of Spending on Imported
Investment Goods
(Time Period Robust Model, with One Variable Deleted)
Model 5.6.TR.b
2SLS Estimates of Determinants of Spending on Imported
Investment Goods
(Time Period Robust Model, with One Variable Added)
Model 5.7
OLS Estimates of Determinants of Business Borrowing
Model 5.8
2SLS Estimates of Determinants of Business Borrowing
deficits, the accelerator, and profits. The other variables only contrib-
uted marginally to explained variance. From the first-in perspective, profit
growth accounted for much more variance than the others; the acceler-
ator and spending deficits were also significantly more important than the
other variables.
Model 5.8.TR
2SLS Estimates of Determinants of Business Borrowing – Time
Period Robust Model
Model 5.8.TR.a
2SLS Estimates of Determinants of Business Borrowing – Time
Period Robust Model
(Two Variables Deleted)
Finally, let us add the money supply and depreciation variables to the
full model time period robust model given in Eq. 5.8. The new model
results are
Model 5.8.TR.b
2SLS Estimates of Determinants of Business Borrowing – Time
Period Robust Model
(Two Variables Added)
Model 5.10
2SLS Model of Investment Spending on Plant and Equipment
regression the OLS model given in Eq. 5.10.1 over the 1960–2010 period
(Table 5.10.1).
From a first-out perspective, the four factors which uniquely explain
the most variance in plant and equipment investment are spending defi-
cits, depreciation allowances, borrowing and profits, but only 3–4%. The
other variables unique contribution was marginal at best. From the first-
in perspective, the Tobin’s q proxy stock market index and the level of
borrowing were the most important.
Table 5.10.2 Robustness over time – plant and equipment, 2SLS Model 5.10
Model 5.10.TR
Time period Robust 2SLS Model of Investment Spending on
Plant and Equipment
The tax deficit variable and the borrowing variable are positively correlated
(r = 0.48), indicating the smaller the deficit, the greater private borrow-
ing. This relationship held in all four test periods. Including two highly
correlated variables typically cause the significance level of at least one to
drop, and that is what happened here. Because of the relatively high cor-
relation level, in two of the test periods the deficit variable was significant
and the borrowing variable was not; it the other two tests just the oppos-
ite occurred. We included both in our “final” time period robust model,
convinced that the reason they were sometimes showing as insignificant
was not that their fundamental relationship to the dependent variable was
spurious, but rather because of a multicollinearity problem’s predictable
consequences. (When the other variable was dropped from the model, the
remaining variable was strongly significant in all four periods in the time
period robust model.)
5.10 DETERMINANTS OF SPENDING ON FIXED PLANT AND EQUIPMENT. . . 211
Model 5.10.TR.a
Time period Robust 2SLS Model of Investment Spending on Plant
and Equipment
(Two Variables Deleted)
Model 5.10.TR.b
Time Period Robust 2SLS Model of Investment Spending on Plant
and Equipment
(Two Variables Added)
Model 5.11
OLS Estimates of Determinants of Residential Construction
Spending
(No 2SLS Model Needed – See Below)
Table 5.11.2 Robustness over time – residential investment, OLS Model 5.11
Model 5.11.TR
OLS Estimates of Determinants of Residential Construction
Spending
Model 5.11.TR.a
OLS Estimates of Determinants of Residential Construction
Spending.
(Two Variables Deleted)
All variables held their sign, and statistical significance, indicating the res-
ults are generally robust. Magnitudes are also generally robust, except for
the stock market index variable, whose coefficient drops by more than
half, and the exchange rate variable whose coefficient grows by more
than half.
Finally, let us add the profits and depreciation variables to the full model
(5.11.TR):
5.13 DETERMINANTS OF SPENDING ON INVENTORY INVESTMENT (OLS) 215
Model 5.11.TR.b
OLS Estimates of Determinants of Residential Construction
Spending
(Two Variables Deleted)
Model 5.13
OLS Estimates of Determinants of Inventory Investment
at a much smaller level, were lagged GDP and (negatively) the level of
consumer spending. The other variables only contributed marginally at
best to explained variance. From the first-in perspective, the accelerator
was again by far the most important, with profits and % capacity utilized
also explaining significant variance.
Model 5.13.TR
OLS Estimates of Determinants of Inventory Investment
Model 5.13.TR.a
OLS Estimates of Determinants of Inventory Investment
(Two Variables Deleted)
The remaining variable had the same sign remained statistically significant
as model 5.13.TR. The magnitude of its marginal effect on inventory did
drop from 0.32 to 0.19, we cannot call its magnitude fully robust to this
particular change in specification. We also note that dropping only one of
the last two variables leaves the other insignificant, so our most robust res-
ult is that the Samuelson accelerator effect is by far the main determinant
of inventory change, explaining nearly 80% of the variance.
Finally, if we added back the capacity utilization and exchange rate
variables to the full three variable model 5.13.TR, all three variables stay
significant and coefficients remain reasonably stable. This has is shown in
Model 5.13.b:
Model 5.13.TR.b
OLS Estimates of Determinants of Inventory Investment
(Two Variables Added)
Coefficients and significance levels of the time period robust model vari-
ables fluctuate somewhat, but are not significantly changed. We conclude
the three variables of the time period robust model are robust to the
addition of other variables to the model, and generally robust to variable
deletions, though with deletions, magnitudes of remaining variables may
fluctuate noticeably.
CHAPTER 6
In 2012, 13.5% of the U.S. GDP was produced to meet foreign demand
(EOP, 2013). This is substantial and changes in foreign demand can have
a significant effect on the U.S. economy. The model of export demand
presented below initially theorized that the principal determinants of
demand for U.S. exports were (1) foreign nations’ income and (2) the
exchange rate, i.e., much the same as key determinants driving domestic
demand for the same items. Using Federal Reserve data on the relative
weights of different countries purchases of exports since 1973, and UN
data on our export trading partners GDPs each year during the 1973–
2010 period, we constructed a trade-weighted real average income variable
to represent yearly variation in our trading partners’ real income. The real
exchange rate variable for each year is the average of the current and past
three years exchange rates.
These two variables only explained a small amount of the variance
in spending on exports over the past 40 years, and we had to search
out other factors to explain most of the variance during this period.
Through a process of trial and error, we found two variables exerting a
mild, but systematic relationship with exports: the U.S. real prime interest
rate (positively related, for reasons not clear) and the U.S. inflation rate
(negatively related). Clearly, one would expect inflation to be negat-
ively related, though its marginal levels of statistical significance probably
reflects the fact that we are already controlling for it in the real exchange
rate; but perhaps less than completely. A theory for why the U.S. interest
rate (average for prior 2 years) is empirically positively related to foreign
demand is not so well established; but we hypothesize that since other
evidence in earlier equations indicates domestic U.S. demand drops with
rising rates, that this results in some U.S. production being offered on
the foreign market that normally would be sold domestically. It may also
mean that foreign earnings on U.S. securities may have been high the
past two years because interest rates were high, and that is now translat-
ing into increased foreign demand for U.S. goods. Another reason may
be that the high interest rate implies lower inflation in the U.S. relative
to other countries, stimulating export demand. Finally, it may also be that
foreign interest rates are even higher when U.S. rates are high, limiting the
funds foreigners can afford to borrow, forcing them to search for cheaper
products. Perhaps they find them in the U.S.
Besides the interest rate and inflation rate, a third and by far the most
important other variable found systematically related to foreign demand
for U.S. exports was U.S. demand for imports. This variable explained
more variance in exports over the past 50 years than all the other determ-
inants of export demand combined, even controlling for both U.S. and
foreign income levels.
The economic explanation may be that the single most important
factor affecting foreigners’ ability to purchase American exports is the U.S.
money they received selling their own goods to the U.S. The coefficient
on this variable indicates that almost 0.60 cents of every dollar Americans
spend on imports is returned in the same period to America in the form
of foreign demand for U.S. exports (and this coefficient’s statistical signi-
ficance is very stable regardless of whether U.S. and foreign income are
controlled for).
It is hard to conceive of why foreigners’ MPC U.S. goods out of in-
come from U.S. imports would be as high as 56%. Would, on average, the
rest of the world’s MPC for U.S. goods be 56% out of their total income
from farming or working for other domestic industries? Probably not; the
high MPC out of imports income is more plausibly because to get U.S.
cash with which to buy U.S. goods, foreigners must sell the U.S. their
goods, and that this is the principal reason they are in the export market.
If this is the primary reason foreigners export to the U.S., we would
expect imports to equal or exceed U.S. exports, which is the case.
This relationship may also reflect the tendency of many countries to
import intermediate goods and then export them as finished goods (or to
THE EXPORTS DEMAND EQUATION 223
more than part of the reason why our export levels seem so tied to our
import levels. The larger reason would seem to be that without selling
goods to the U.S., foreigners cannot obtain the U.S. currency needed to
buy goods from the U.S.
Our causality analysis suggests American demand for imports is a
stronger determinant of the demand for U.S. exports than the demand
for U.S. exports is a determinant of the demand by the U.S. for imports.
And from earlier analysis of the demand for U.S. imports, we know U.S.
income is a major determinant of import demand. Taken together, these
two facts suggest the principal reason for the demand by the U.S. for for-
eign imports is to satisfy U.S. consumer and business needs, not to add
value by fabricating and re-exporting. This then implies that the system-
atic relationship of every dollar of imports being related to a 0.56 cent
increase in demand for U.S. exports is probably due to the desire of for-
eigners to meet their demand for U.S. goods through dollars earned from
trading, rather than by borrowing U.S. currency.
As usual in our models, if theory (or past empirical findings) indic-
ates a variable should be included in a model, we include it using the lag
level (or level of linearity) most consistent with theory, i.e., the one that
explains the most variance in the dependent variable. For foreign income,
we found our trading partners income lagged two periods worked best
(controlling for the effects of U.S. imports), the 4-year average exchange
rate, the average interest rate for the past two years, and the average U.S.
inflation rate.
Further, we examined whether perhaps the relationship might work in
reverse: foreign demand for U.S. exports providing needed funds to allow
Americans to buy imports. As shown in Eqs. 6.2 and 6.3, our demand
for our imports is slightly better explained by our export levels than vice
versa. But, if it is the desire to obtain currency to buy U.S. exports that
accounts for the strong sales of foreign goods to the U.S., the coefficient
on exports (1.28) in Eq. 6.3 is irrational. It suggests that for every dol-
lar of imports we buy, we sell more than a dollar’s worth of exports to
foreigners, which we know is not the case. The coefficient in Eq. 6.2
is more reasonable; it suggests that for every dollar of goods the U.S.
imports, $0.65 is returned to the U.S. in the form of purchases of U.S.
exports.
The equation expressing the demand for U.S. exports was found to be
as shown in Eq. 6.1. It was estimated using OLS reflecting the finding of
no endogeneity of the dependent variable with right-hand-side variables.
6.1 OLS MODEL OF EXPORT DEMAND 225
Model 6.1
OLS Model of Export Demand
X = 0.16(WGDPRealTP(–2) ) – 9.47(XRAV0 to –3 ) + 0.56(M)
(t =) (2.9) (–4.1) (18.6)
+ 14.74(PRRealAV–1–2 ) – 11.58INFLAV–1 to –2 ) – 0.49AR(6)
(–3.9) (–2.0) (–1.7)
R2 = 87.9; DW = 1.6 MSE = 24.69 (6.1.TR)
Graph 6.1.1
The export demand model in 6.1 explains variation in exports about
equally well in each of the five decades of the sample period.
200
100
–100
80
60 –200
40
20
0
–20
–40
82 84 86 88 90 92 94 96 98 00 02 04 06 08 10
Residual Actual Fitted
The results indicate that U.S. demand for imports is at least a margin-
ally better explanation of the foreign demand for U.S. exports than is
the opposite hypothesis, i.e., that our import demand is conditioned by
foreign demand for our exports.
Explained Variance and Robustness Tests
Overwhelmingly, the principal determinant of demand for U.S. exports
is U.S. demand for imports (Table 6.1.1). This may be because of U.S.
demand for intermediate goods, which are then repackaged and sold as
final goods back to foreign markets, or it may reflect a need for foreigners
to obtain U.S. currency in order to buy U.S. goods, and that most of
this has to be obtained from trade, with foreigners unwilling or unable to
borrow sufficient dollars to pay for their desired exports from America.
In first differences only one of its determinants (government spending)
was nonstationary, the variable measuring export goods and services, but
it was cointegrated with the dependent variable, so no detrending was
needed (Table 6.1.2).
Since the earliest data available was from the late 1970s only three
samples of reasonable size could be taken. Overall, variables statistically
significant in one sample were significant in all. Coefficients were not quite
as stable, especially for the 1980–2000 small sample, but this may have
more to do with small sample size than with robustness over time. Hence,
our time period robust model is the same as our initial Model 6.1.
Model 6.1.TR
Time Period Robust OLS Model of Export Demand
X = 0.16(WGDPRealTP(–2) ) – 9.47(XRAV0 to –3 ) + 0.56(M)
(t =) (2.9) (–4.1) (18.6)
+ 14.74(PRRealAV–1–2 ) – 11.58INFLAV–1 to –2 ) – 0.49AR(6)
(–3.9) (–2.0) (–1.7)
R2 = 87.9; DW = 1.6 MSE = 24.69 (6.1.TR)
Model 6.1.TR.a
Time Period Robust OLS Model of Export Demand
(Two Variables Deleted)
X = 0.10(WGDPRealTP(0) ) – 8.26(XRAV0 to –3 ) + 0.58(MT ) – 0.39AR(6)
(t =) (1.8) (–4.2) (16.9) (–1.5)
R2 = 0.73; DW = 1.3 MSE = 31.06 (6.1.TR.a)
228 6 THE EXPORTS DEMAND EQUATION
Model 6.1.TR.b
Time Period Robust OLS Model of Export Demand
(One Variable Deleted)
X = – 10.17(XRAV0 to –3 ) + 0.58(MT ) + 14.24(PRRealAV–1–2 )
(t =) (–4.4) (20.0) (–3.4)
– 11.68 INFLAV–1 to –2 ) – 0.10 AR(6)
(–2.0) (–0.5) (6.1.TR.b)
Results are nearly identical to the full model.
Finally, we add the real money supply variable to the full time period
robust Model 6.1.TR and re-estimate:
Model 6.1.TR.c
Time Period Robust OLS Model of Export Demand
X = 0.13(WGDPRealTP(–2) ) – 9.69(XRAV0 to –3 ) + 0.57(M)
(t =) (3.0) (–4.3) (21.8)
+ 15.43(PRRealAV–1–2 ) – 11.20INFLAV–1 to –2 ) + 0.09M1Real
(3.2) (–2.0) ()
2
–0.49 AR(6) R = 88.9; DW = 1.6 MSE = 24.12
(11) (6.1.TR.c)
Coefficients and significance levels remain robust with the addition of
this variable.
Hence, we conclude the full model (6.1.TR) is very robust to specific-
ation changes as well as time period changes.
CHAPTER 7
To calculate the GDP, the GDP identity is used as a basis for developing
a model comprised of the system’s exogenous and lagged variables (the
“reduced form,” or “IS” curve model). The general form of such models
is described in Section 2. The reduced form or IS model is derived from
the GDP identity
GDP = C + I + G + (X – M) (7.1)
GDP = CD + ID + GD + X (7.2)
Model 7.1.1
GDP Determination Model: Single Regression Estimates
YT = .26(TT ) – .17(GT& I ) – 7.94PR – .09DJ–0 – .006DJ–2
(t =) (2.4) (–2.0) (–1.9) (–0.5) (0.0)
+ 7.39XRAV + 107.28POP16 + .049POP + 1.32ICC–1
(–1.1) (0.2) (4.4) (2.9)
+ .001M2AV + .62(ACC) + 3.17DEP + 5.76CAP–1
(0.0) (15.0) (2.4) (2.0)
+ .61PR–2 + .05PROF–0 + .03 (CB2 + (IB(–1) ) – .02X
(0.1) (0.4) (0.4) (0.1)
+1.00AR(2) – .34AR(7)
(+5.3) (–2.0)
R2 = 95.8% D.W. = 1.9 MSE = 41.59
(7.1.1)
Clearly the IS curve has many insignificant variables and many variables
whose coefficients seem larger or smaller than we would expect given their
values in the consumption and investment equations. This can result from
four things: Briefly summarized they are:
Clearly we do not have any generally accepted theory that tells us that while rising interest rates have
negative effects on consumption and investment, rising interest rates have a positive effect on government
spending. The government effect is a spurious correlation at best; an example of the “correlation does not
imply causation.” Further, consumption function studies (Eq. 4.4 only found current interest rates, not
lagged rates, statistically significant. Hence, the lagged effect is presumed spurious This leads good studies
to only test variables that have a sound theoretical foundation. Running a regression of GDP on all the
determinants of GDP simple gives an estimate of the marginal effect of the interest rate variable that is
the sum of its measured effects in all the individual component models of GDP, whether these effects are
causal or merely spuriously correlative).
Section 2.2.4.4 presented an even simpler example of the same problem and how it affects VAR models.
As can be seen in the table above, the sum of the regression coefficients
for the parts equals the regression coefficient for the whole and considerably
misstates the effects of the lagged interest rate considered causal. There-
fore, we conclude that only adding values for coefficients of variables
actually included in the consumption, investment, etc., function models
is more correct than the method employed in 7.3 above: regressing GDP
on all variables found significant in any equation in the system.
The failure to do so is a is a particularly serious potential flaw in VAR –
methodologies, which typically regress each dependent variable on exactly
the same right hand side variables. If one of the VAR model’s depend-
ent variables (e.g., GDP) is the sum of several subcomponents (e.g., C,
I., etc.), each in reality determined by different variables, estimates of the
effect on GDP of any one variable will invariably be biased by the VAR
standard method of inclusion of the same determining variables in each
VAR equation. In this way, spurious components in one equation can be
added to causal components in another to get the total effect of (say) C
and I determinants on GDP. VAR only avoids this problem if all depend-
ent variables in its model are causally determined by exactly the same
right-hand side variables (and lags of them). See Section 2.2.4.4 for an
example.
We will show further below that estimating the GDP not by regres-
sion, but by adding only coefficients in subcomponent regressions where
the variable is actually considered a determinant leads to estimates whose
average error of estimate over time is only 40% of the average error
obtained using the single statistically estimated GDP function, such as
7.1.1 above. A smaller average error is exactly what we would expect if
estimates in the single equation GDP determination model, econometric-
ally estimated, distort the actual causal effects of the GDP’s determinants
on the GDP.
Hence, adding the specific parameter estimate for each variable
obtained in our consumption, investment and export regressions, where
only variables thought causally related are included, rather than estimating
a GDP regression, will be the process used to estimate GDP in Section 8.
(This would not be a problem if the scientific techniques we rely on in
economics were experimental rather than correlational, but they are not.)
Generally, this is a problem which makes statistical estimation of a func-
tion that is the sum of two or more separate sub-functions that have
separate determinants inappropriate. It results in biased coefficients. We
have shown this for the GDP function here.
236 7 STATISTICALLY ESTIMATED REAL GDP DETERMINATION FUNCTIONS. . .
A few other examples from the same equations used to illustrate interest
rate effects in Table 7.1.1 above are provided below:
By the same line of reasoning, is there any theory that says spending defi-
cits increase exports or that wealth and population growth cut them? Yet
this enters into the IS curve-type regression calculation of a change in
these variables on the economy. Here again, the only way to avoid this
problem is to regress each component of the GDP, (C, I, G, and X–M),
on its own determinants, and only add those estimated effect together to
get the total effect on GDP.
Model 7.2.1
GDP Determination from One Regression Equation (X–M Model)
found nonstationary, but all were cointegrated with the dependent variable
(YT ) so no detrending was necessary.
As you can see, the same problem with combining causal and correlat-
ive effects is present, so a better way of estimating the separate effects of,
e.g., current period versus lagged interest rate effects needs to be con-
structed. For this reason, Eqs. 7.1.1 and 7.2.1 are included here only
for reference and as a way of illustrating the problems resulting from
this method of statistical GDP determination. The actual behavioral IS
model of GDP determination used as part of this large scale model of the
American economy is given in Chapter 8.
CHAPTER 8
GDP = CD + ID + GD + X (7.0.2)
Model 5.4
2SLS Estimates of Determinants of Spending on Domestically
Produced Investment Goods
Model 8.1.1.1
GDP Determination by Adding Together Subcomponent Model
Parameter Estimates
YT = – 0.41 (TT ) + 0.69 (GT& I ) + 0.85(TDef ) – 0.72 (GDef )
– 9.67PR – 4.24PR–2 – 0.27DJ–0 + 0.62 DJ–2
+ 9.52 XRAV – 729.21POP16 + 0.044POP
+ 0.75ICC–1 + 53.81M2AV + 0.14 CB2
+ 0.35(ACC) + 0.11DEP + 3.67CAP–1
+ 0.11PROF–0 + 0.03(IB(–1) ) + 1.41 X
(8.1.1.1)
stimulus effect and its crowd out effect. Our empirical estimate of stimu-
lus effects – the coefficient on the government spending variable – will be
compared to the 1.00 coefficient commonly assumed when constructing IS
curves. To check the validity of our estimates, we will then run the same
model, this time deleting the spending deficit crowd out variable – only
the total government spending level variable will be retained. It should
measure the net stimulus effect, after accounting for crowd out, e.g.,
GNet = GStimulus – GCrowd out . It does so, exactly. This provides addi-
tional confidence our estimate of the coefficient on the spending vari-
able is correct. Results for the modified consumption equation are as
follows:
And the equation showing only the net of consumption stimulus and
crowd out effects for government spending is Eq. 4.4.Rev2 below. The
Revision 2 model is exactly the same as the Revision 1 model, except the
stimulus and crowd out effects have been combined, i.e., 0.02GT+I –
0.21(GDef )=02GT+I – 0.21(GT+I – LF) = –0.19(GT+I ) +0.21LF)
where LF is the loanable funds variable. Regression coefficients obtained
are exactly the same as would have been obtained arithmetically from Revi-
sion 1, increasing our confidence in the correctness of our results. Coef-
ficients on all other variables in Revision 2 stay exactly the same as they
were in Revision 1. t-statistics are the same or only marginally different.
And the effects of deleting the (GDEF ) variable so that (GT+I ) only
captures the net of the stimulus and crowd out effects of government
spending deficits on investment is given in Revision 2 below:
for estimating the other variable effects on GDP, and therefore provides
better empirical estimates.
Model 8.1.2.1TR
GDP Determination by Adding Together Subcomponent
Model Parameter Estimates
(Robust Model)
Model 4.1.T
Total Consumption Model (Repeated)
Model 5.2
Total Investment Model (Repeated)
Following the same procedure used in Section 8.1, but with the model
GDP = CT + IT + GT + (X – M) (7.0.1)
Model 8.2.1.1
GDP Determination by Addition of Component
functions of the GDP
YT = – 0.92(TT ) + 1.04(GT& I ) + 1.51 (TDef ) – 1.25(GDef )
– 19.16PR – 13.23PR–2 + 1.02DJ–0 + 0.83 DJ–2
+ 14.07 XRAV – 803.04POP16 + 0.04POP + 0.71ICC
+ 88.92M2AV + 0.23CB2 + 0.58 ACC + 1.31DEP
+ 4.38CAP–1 + 0.06PROF–0 + 0.12(IB(–1) ) + 1.92 (X – M)
(8.2.1.1)
out) variable – only the separate spending level variable will be retained.
It should measure the net stimulus effect, after accounting for crowd
out, e.g., GNet = GStimulus – GCrowd out . It does, exactly. This provides
additional confidence our estimate of the coefficient on the spending vari-
able is correct. Results for the modified consumption equation used in IS
Model 8.2.1.1 below are as follows:
to be estimated capturing just the net of stimulus and crowd out effects
of a change in government spending. The results adding the separate
spending variable to capture stimulus effects are:
Model 4.1.T.TR
Robust Total Consumption Model (Repeated)
CT = 0.49(Y–TT ) + 0.57(TDef ) – 0.38(GDef ) – 9.31PR + 0.44 DJ–2
(t =) (10.8) (11.0) (–7.9) (–4.6) (5.4)
+ 0.017POP + 0.41ICC–1 + 44.78M2AV + 0.13 CB2
(4.3) (1.2) (4.3) (3.6)
R2 = 94.8% D.W. = 1.6 MSE = 24.75
(4.1T.TR)
Repeating the final robust total investment model from Eq. 5.2.TR above
Model 5.2.TR
Robust Total Investment Model (Repeated)
IT = + 0.25(ACC) + 0.30(TDef ) – 0.32(GDef ) – 10.53 PR–2
(t =) (8.2) (2.7) (–4.4) (–4.3)
+ 0.87 DJAV + 3.18XRAV + 0.97 DEP
(3.3) (1.5) (4.1) (5.2.TR)
R2 = 86.6% D.W. = 2.2 MSE = 29.43
Model 8.2.2.1TR
GDP Determination by Addition of Component
Functions of the GDP
Y = – 0.96(TT ) + 1.06 (GT& I ) + 1.71(TDEF ) – 1.37(GDEF )
– 18.25PR – 20.64 PR–2 + 1.71 DJ0 + 0.86 DJ–2 + 0.033POP
+ 0.80ICC–1 + 87.77M2AV + 0.25 CBOR + 0.49ACC
+ 6.23XRAV + 1.90 (DEP) + 1.96 (X – M)
(8.2.2.1TR)
Most of the coefficients are close in value to the initial model, despite the
elimination of several variables as spuriously significant in the first test.
CHAPTER 9
Heim (2008) examined a wide range of interest rates, including the federal
funds rate, the prime interest rate the Aaa and Baa rates, and the mortgage
rate to determine which had the largest and most systematic impact on
the economy. A typical simple Keynesian IS model in which consumption
was driven mainly by income and investment was driven mainly by interest
rates and the accelerator was used to evaluate the results. That study indic-
ated the rate most systematically related to variation in GDP was the prime
interest rate, which is derived from the federal funds rate. This is consist-
ent with data indicating bank lending historically was a greater source of
business credit than the bond market.
Traditionally, the prime interest rate has been set by the business com-
munity rigidly at 3 percentage points above the federal funds rate (though
that spread was reduced for a while during the 2008 fiscal crisis), not by
the interaction of supply and demand. Since the federal funds rate is set
by the Board of Governors of the Federal Reserve, we can consider the
federal funds rate and therefore the prime rate, as exogenously determ-
ined by an outside body. However, to the extent the behavior of the Fed
is determined by economic conditions, the rate can be considered endo-
genously determined by the variables that affect Fed behavior. Modern
practice has been to consider it determined endogenously, with the Fed-
eral Reserve’s “exogenous” behavior determined some by variant of the
Taylor rule and/or other variables. We have been able to explain about
80% of the variation in the prime rate endogenously over the past 50 years
using the following Taylor rule model:
Using the first-out method, the three variables uniquely accounting for
the most variation in the prime rate were the inflation rate, the unemploy-
ment rate and the inflation effect of an increase in the money supply. Using
the first-in method, the three variables accounting for both unique vari-
ance and variance that could be assigned to other variables are inflation,
unemployment and government spending.
Robustness Over Time
Table 9.2.2 shows the results of testing the prime interest rate in four dif-
ferent but overlapping time periods as a means of testing the replicability
of the initial results in other time periods. The ability to replicate initial
findings is a key requirement of good science, a key goal of this study.
Results indicate that the effects of the Taylor rule variables (inflation
and unemployment) are extremely robust to time period sampled, as are
the results for the deficit variables (consistently insignificant).
The effects of changes in the money supply on the prime rate are not
robust: only with the 2001–2010 decade in the sample do money supply
changes appear significant. With that decade in the sample, they seem to
9.2 2SLS ESTIMATES 257
Table 9.2.2 Robustness over time – Taylor rule model: 2SLS model 9.2
have a strong effect in a way consistent with theory. This may be because
variation in the yearly change in the money supply was much greater in the
later part of the 2001–2010 decade, allowing a tighter confidence interval
around its estimated average yearly effect.
In the 1960–2000 and 1970–2000 samples, changes in the money sup-
ply did not seem to affect the prime rate. One theory why (untested here)
is that the Fed makes changes in the federal funds rate when the consensus
in the credit markets is that changes are needed. Hence, just announcing
changes to the federal funds rate just sends a signal to the banking com-
munity that now is the time to make changes to the prime rate, changes
they already thought it was time to do. In this way, by changing the
prime, the banks are simply, concurring with the timeliness of the Fed’s
own change to the federal funds rate. Of course, by this theory, the Fed’s
change only occurred because the Fed sensed the credit markets thought
it was time for a change.
Our time period robust model becomes the two variables found sig-
nificant in all four tests above, plus any found insignificant above which
are now found significant in three of four tests of this time-period robust
model (none were). It is given as Eq. 9.2.TR. This becomes our final time
period robust model:
The model is robust in that the signs on variables remain the same,
and they remain statistically significant. However, the magnitude of the
remaining variable’s coefficient rises, which is to be expected as long as
any correlation between the two exists and both are major explanatory
variables. Hence, we judge the model robust to this change as regards its
sign and statistical significance, but not with respect to coefficient mag-
nitude, particularly for the inflation variable whose coefficient fluctuates
more (0.42 vs. 0.78) than our rough limit for defining consistency (1/3).
Adding the GDP variable to the time period robust model:
(M/P)D = αY – βr
where, (M/P)D represents the demand for real money balances, Y repres-
ents aggregate income, and r represents the interest rate
Implying that when (M/P)S = (M/P)D
r = (α/β)Y – (1/β)(M/P)S
with the first determinant (Y) representing the transactions and precau-
tionary need for money, and the second the speculative.
Overall, the standard Keynesian model of money demand and interest rate
determination explains less than half the variance explained by the Taylor
rule model. A third of the variance explained by the Keynesian model is
only due to adding a lagged money supply variable (average of M1–1 and
M1–2 ) variable to test the inflationary effects as well as the liquidity effects
of expansion of the money supply. This would not have been done in the
early Keynesian models and gives a more modern interpretation to the
classic Keynesian function. Further, as was the case with the Taylor rule
model, the subsequent inflationary effects of increasing the money supply
more than offset the current year liquidity (stimulus) effects, consistent
with the more classical notion that inflation is the only long run effect of
changing the money supply.
Results are similar to the OLS results, but with the money supply variable
having stronger liquidity and inflation effects, this time with the liquidity
effect slightly stronger than the inflation effect. Clearly though the initial
Keynesian LM model of interest rate determination leaves most of the
behavior of the prime interest rate unexplained. The initial Taylor rule
model discussed earlier, by comparison, explained 80% of the variance over
10.2 2SLS MODELS OF THE LM CURVE 261
the same period and therefore provides a much more compelling theory
of what drives interest rates. In Keynes’ defense, his interest rates were
market determined by the supply and demand for money; the prime rate
is not; it is an administered rate, determined in small part by fluctuation
in the M1 money supply, but principally by inflation and unemployment,
which explains most of the variation in the prime rate (and therefore the
federal funds rate) over time.
In first differences, neither the dependent variable nor any one of its
three determinants were found ADF nonstationary, so no cointegration
tests or detrending was needed.
model 10.2.TR. Overall, the results seem relatively robust to time period
sampled. The variable representing the transactions demand effect is only
marginally statistically significant regardless of time period tested. If dis-
posable income is used instead of gross income, results are the same. This is
significant because virtually all theories of money demand historically have
considered transactions demand as the principal cause of the demand for
money. This seems not to be the case, since it was only of marginal signi-
ficance here, and of no significance when added to the Taylor rule interest
rate function earlier. As Baumol (1952) and Tobin (1956) suggested over
half a century ago, even transactions demand for money may be interest
rate driven.
Essentially the same model with inflation and unemployment variables
added to it is the Taylor rule model. When these two are added, the
money variables become insignificant for the first 40 (1960–2000) of the
50 years (1960–2010) tested. This suggests the prime (and federal funds)
rate relationship to how much money the Fed is creating is more likely
correlational than causal. Though the Fed reduces target interest rates in
recessions, the money supply may not have to increase much to do so,
causing the data to show the unemployment rate is a more important
influence than the money supply. A similar argument can be made sug-
gesting that in periods of inflation, a Fed attempt to increase the target
rate may not take much of a reduction in the money supply. If the Fed
announces an increase in the target Fed funds rate, banks are more than
happy (usually the same day) to increase the prime rate before the Fed
has even had a chance to engage in open market operations to achieve its
goal.
UnemAV(0 and –1) = average unemployment rate for the current and past
year
M1 = change in the M1 Money supply, lagged two years
((M-X)/Y)RealAV(0–1) = net exports as a percent of GDP (average of cur-
rent and past year). Results show trade deficits are deflationary, since
deficits (M-X) have a positive sign.
ForBor–1 /Y–1 = real foreign borrowing by the U.S. as a percentage of
real investment
GrossSav(–1) /Y(–1) = prior year net U.S. savings + prior year depreciation
allowances, divided by prior year GDP
OPEC 73, 78 Shocks = OPEC 1973 & 78 Oil Price Increases (dummy
variable: 1 for 74, 75, 79, 80). Shocks were also entered in first
differences.
In first differences, neither the dependent variable nor any one of its five
determinants were found ADF nonstationary, so no cointegration tests or
detrending was needed. No current period variables were found Hausman-
endogenous, so no 2SLS was needed. Further tests below are in OLS.
Hence, though not identical, the functions are closely related: M1–1 and
M1–2 play a key role in all of them.
This can be shown statistically as well. If we add M1–2 in the Taylor
rule model (11.1.1) by averaging it with the lagged M1 variable already in
it, M1–1 , we reduce the coefficient and significance level of the inflation
variable in the Taylor rule equation, suggesting something in the M1–2
variable is now picking up variation formerly picked up by the inflation
variable, as shown in 11.1.4 below
for the inflation variable (and vice versa), as shown in 11.1.5 and 11.1.6
below:
Determinants of Unemployment
In the long run, factors like technological progress, levels of capital, and
changes in the size of the labor force play a role in determining the level
of output. In the short run, since technology and capital are fixed, gen-
erally, to produce more output, more labor from the existing labor force
is required, though varying levels of efficiency over the business cycle can
affect how much. Therefore, we would expect growth in demand for GDP
to be a driving force in levels of demand for labor at the macroeconomic
level.
Four OLS models of unemployment variation are presented below
which explain 70–80% of the amount of variance in unemployment over
the 1960–2010 sample period.
Or, in the equivalent form we will use, which more clearly shows the
demand-driven nature of the relationship:
Model 12.1.1
Okun Unemployment Model (OLS)
Model 12.1.2
Okun Unemployment Model (OLS Modified)
Model 12.1.3
Okun OLS Unemployment Model Augmented to
Include the Effects of Inflation and Shocks
In first differences, the dependent variable was stationary, as were all of the
RHS variables. The Woodridge (2003) method of using dummy variables
in levels with a constant term was used; it explained significantly more
variance than the same model running the dummies in first difference form
without a constant.
The current year inflation rate was found significant. The best res-
ults for the 1973 OPEC oil price shock were found using a dummy
variable with negative shock effects indicated for only 1974 and 1975;
and for the 1978 shock, best estimates were for shock effects for 1979
and 1980. The best dummy for Katrina counted 2005 and 2006 as
impact years, and the dummy for 2008 hypothesized 2009 as the impact
year (small effects for 2008 and 2010 were also seen, but were not
statistically significant). Overall, this augmented Okun model captures
the effect of the GDP on unemployment quite nicely, as shown in
Graph 12.2.1. Interestingly, not all shocks have a negative effect on
the unemployment rate: the 1973, 1978, and 2008 shocks did, but the
2005 natural disaster shock did not presumably because of the imme-
diate and large scale rebuilding effort that followed that shock required
large amounts of labor. Interestingly, the shocks that did increase the
unemployment rate were financial; the shock that decreased it was real
(Katrina).
The basic Okun model explains unemployment trends the past 50 years
quite well, including the sharp rise in unemployment in 2008 and 2009 as
the banks crisis impacted the economy.
276 12 DETERMINANTS OF UNEMPLOYMENT
0
1.2
0.8 –2
0.4
–4
0.0
–0.4
–0.8
1965 1970 1975 1980 1985 1990 1995 2000 2005 2010
Residual Actual Fitted
Graph. 12.2.1 The augmented Okun model (Eq. 12.4) model for explaining
variation in unemployment 1960–2010
The OLS and 2SLS results are similar, differing most notably for the
OPEC shocks, where the estimated effects and statistical significance of
both decline, though the 1973 shock stays significant.
Contribution to Explained Variance
The importance of these variables in explaining variation in unemployment
during the 1960–2010 period can be examined using stepwise regres-
sion the 2SLS model given in Eq. 12.4 over the 1960–2010 period
(Table 12.2.1)
Clearly, changes in the GDP are the most important factor driving
changes in unemployment.
Robustness Over Time
The model was tested in four different but overlapping time periods to
allow examination of robustness of the initial results in different time
periods. Results are given in Table 12.2.2.
Estimates of current year GDP, inflation, the 1973 oil shock, the Kat-
rina and financial crisis shocks were all robust to the time period sampled.
Estimates of the effects of the 1978 oil shock and prior year GDP effects
were not.
We define the core time period robust model as a model containing
the variables above that were found significant in at least three of four
time periods sampled. To this model we add either of the two nonsignific-
ant variables which, when added to this reduced size core robust model,
now are found significant in three of four periods sampled, but not when
both are included in the same model. Hence, they are left out of the
time period robust model. This final time period robust model is given
in Eq. 12.2.1.TR.
Model 12.2.1.TR
Okun 2SLS Unemployment Model Augmented to Include the
Effects of Inflation and Shocks
(Time Period Robust)
Model 12.2.1.TR.a
Okun 2SLS Unemployment Model Augmented to Include the
Effects of Inflation and Shocks
(Time Period Robust Model, with Three Variables Deleted)
Model 12.3.1
OLS Technological Change Model Augmented by Shocks
Model 12.4.1
2SLS Technological Change Model Augmented by Shocks
The same variables are significant as were in the OLS model, except for the
OPEC oil shocks. Estimated marginal effects, again, are very similar to the
OLS model findings, especially for the statistically significant variables. The
73 and 78 shock values changed significantly, with both declining to lack of
statistical significance. The estimate marginal effect of the 2008 crisis also
declined almost half, and its strong significance reduced to marginal signi-
ficance. Just as GDP growth in the Okun model was the most significant
determinant of unemployment, GDP growth dampened by technological
progress is the most important here.
The extent to which this model accurately models the movement of
unemployment over the 1960–2010 period is shown in Graph 12.4.1.
Contributions to Explained Variance
The importance of these variables in explaining variation in unemploy-
ment during the 1960–2010 period can be examined using stepwise
regression the 2SLS model given in Eq. 15.5A over the 1960–2010 period
(Table 12.4.1).
12.4 THE 2SLS TECHNOLOGICAL CHANGE MODEL 281
0
0.8
–2
0.4
–4
0.0
–0.4
–0.8
1960 1965 1970 1975 1980 1985 1990 1995 2000 2005 2010
Residual Actual Fitted
In the first out tests, only current GDP had more than a marginal
effect on employment levels. The first in tests also indicated current
GDP was the most important factor, but also indicated all the other vari-
ables were of more than marginal importance, except the 1978 and 2005
shocks.
282 12 DETERMINANTS OF UNEMPLOYMENT
Model 12.4.1.TR
2SLS Technological Change Model Augmented by Shocks
(Time Period Robust Model)
%Unemployed = 1.55 – 63.53 (GDP0.17 ) – 0.48 Shock05
(12.4) (–12.9) (–5.7) (12.4.1.TR)
R2 = 0.78; DW = 2.2; MSE = 0.49
Model 12.4.1.TR.a
2SLS Technological Change Model Augmented by Shocks
(Time Period Robust Model with One Variable Deleted)
%Unemployed = 1.53 – 63.64 (GDP0.17 )
(12.4) (–13.0)
R2 = 83; DW = 1.8; MSE = 0.43
(12.4.1.TR.a)
Table 12.4.2 Robustness over time: (tech. progress unemployment, 2SLS model)
Retesting 12.4.1.TR by adding back the variables deleted from Eq. 12.4.1
indicates the signs and significance levels of the two variables in the time
period robust model stay the same, and the variation in coefficient is
moderate.
We conclude the time period robust model 12.4.1.TR is also largely
robust to specification changes as well.
CHAPTER 13
The flow of funds accounts divides total U.S. savings into personal savings,
corporate savings, and depreciation. All three components constitute gross
U.S. savings and are taken from Table B.32 (“Gross Savings and Invest-
ment”) of the 2005 (for 1959–1963 data) and the 2013 (for 1964–2010
data) Economic Report of the President. Net savings are gross savings minus
depreciation, and equal net private saving (undistributed corporate profits
and personal saving) and net government saving. Gross saving and foreign
borrowing equal total investment (+ statistical discrepancy) in the flow of
funds accounts and in Table B.32.
Model 13.1.1
OLS Estimates of Determinants of Corporate Saving
(2SLS results presented further below since the average exchange rate and
government spending were found endogenously related to spending on
domestically produced investment goods.)
SC = 0.02(ACC) + 0.82(TT ) – 0.91(GT& I ) – 59DEP
(t =) (0.4) (8.4) (–12.3) (–1.4)
= 0.06CAP–1 + 3.42PR–2 – 0.30DJ–0 + 0.70PROF–0
(0.0) (1.0) (–1.8) (7.9) (13.1.1)
+ 7.65XRAV + 0.007POP – 0.10(BOR–1 )
(3.6) (1.6) (1.3)
2
R = 94.0% D.W. = 1.9 MSE = 36.39
Key variables affecting the level of corporate saving include the gov-
ernment deficit, which affects corporate savings negatively regardless of
whether the deficit is the result of tax cuts or spending increases, profits
and the exchange rate. Marginally significant was the negative relationship
between stock market values and corporate savings. The model explains
virtually all the variance in corporate savings the past 50 years and does so
as well in one decade as another.
The completeness with which the model explains yearly changes
in corporate savings is shown in the graph below of Eq. 13.1.1
(Graph 13.1.1).
400
200
120 –200
80 –400
40 –600
–40
–80
1960 1965 1970 1975 1980 1985 1990 1995 2000 2005 2010
Residual Actual Fitted
Graph. 13.1.1 Fifty years annual variation in corporate saving (calculated from
Eq. 13.1.1, then compared to actual)
Model 13.1.2
2SLS Estimates of Determinants of Corporate Saving
The same variables in the 2SLS model were found significant or insig-
nificant as in the OLS model, with three exceptions. In the 2SLS model
the DJ variable (Tobin’s q proxy) was not significant, and the depreciation
and population variables were, contrary to the OLS findings. Coefficients
288 13 THE SAVINGS FUNCTIONS
Explained variance: First-out stepwise method First-in stepwise method (constant added)
(R2 = 0.95 to start) (R2 = 0.00 to start)
Table 13.1.2 Robustness over time – corporate savings, Eq. 13.1.2 2SLS Model
robust variables. If was found significant in at least three of the four sample
periods, they became the “final” time period robust model. The results are
shown in Eq. 13.1.2.TR.
Model 13.1.2.TR
Time Period Robust 2SLS Estimates
of Determinants of Corporate Saving
Model 13.1.2.TR.a
Time Period Robust 2SLS Estimates of Determinants
of Corporate Saving
(One Variable Deleted)
Model 13.1.2.TR.b
Time Period Robust 2SLS Estimates of Determinants of
Corporate Saving
(One Variable Added)
explain as much variance as the average investment variables for the same
periods. Investment averages older than 17 years did were not statistically
significant.
Initial findings are presented in Model 13.2.1. The final model, robust
to changes in time period sampled, is presented in Model 13.2.1.TR fur-
ther below. Because of multicollinearity issues discussed after that model
is presented, we cannot say the model is robust to the addition or deletion
of variables to the model.
Model 13.2.1
OLS Model of Depreciation Savings
The same model explains depreciation well in any decade from 1980
to 2010, which was the longest we could test for with the data avail-
able. In first differences, the dependent variable (depreciation savings)
was nonstationary (Graph 13.2.1). Two of the eight right-hand-side
variables in model were nonstationary in first differences: average invest-
ment lagged 6–10 years, and average investment lagged 11–17 years, but
both were cointegrated with the dependent variable, hence detrending
was not required. The investment variable was tested, but no vari-
ables were found endogenous with depreciation, so no 2SLS model was
developed.
Contributions to Explained Variance
We can examine contributions to explained variance over the past 50 years
using the stepwise regression models in Table 13.2.1.
Using the first-out method, clearly the highest contributors to
explained variance in depreciation savings is investment taken from 6 to
17 years year earlier, which were significantly affecting total depreciation
allowances taken in the current year. In addition are investment in the cur-
rent and past year were important using the first-out method. The first-in
13.2 THE DEPRECIATION ALLOWANCES SAVINGS FUNCTION 293
80
60
40
12 20
8 0
4 –20
–4
–8
86 88 90 92 94 96 98 00 02 04 06 08 10
Residual Actual Fitted
Graph 13.2.1 Explained and actual depreciation allowance savings the past 50
years
With so many lagged values in the model, and the need for eight additional lags for the AR(8) process, the
model could only be tested on the 1985–2010 data set, or subdivisions of it. Three samples were tested.
Significance Level: * 10%; ∗∗ 5%; ∗∗∗ 1%.
Model 13.2.1.TR
OLS Model of Depreciation Savings
Model 13.2.1.TR.a
Time Period Robust OLS Model of Depreciation Savings
(Lags 6–17 Deleted)
The variables representing lags 6–17 explains nearly half the total explained
variance. Since there is some collinearity between 1 year’s investment levels
and another, it is not surprising that we note the coefficients on the
variables remaining in the equation get larger. We conclude the original
model is robust as regards the continued statistical significance of variables
remaining in the equation, but not as to their coefficient values.
The original model without the current period and lagged one period
investment variable.
Model 13.2.1.TR.b
Time Period Robust OLS Model of Depreciation Savings
(Lags 0, 1 Deleted)
Both coefficients and significance levels of investment lags 2–5 are substan-
tially changed by omitting lags 0 and 1, though not for the coefficients and
significance of the lag 6–10 variable or the lag 11–17 variable.
Again there is nearly a 50% drop in explained variance, suggesting that
most firms take accelerated depreciation, since these two lags account for
nearly as much depreciation variance as lags 6–17 combined.
We conclude that though time period robust, Model 13.2.1.TR is
not fully specification robust. Data on current and lagged values is
too multicollinear. Removing variables that explain a lot of variance
296 13 THE SAVINGS FUNCTIONS
almost invariably will distort coefficients on variables with which they are
collinear. This is perhaps the single most important unresolved issue in
econometrics today.
Model 13.3.1
OLS Estimates of Determinants of Personal Saving.
(No 2SLS results presented since no variables were found endogenously
related to personal saving.)
400
300
200
100
0
–100
100 –200
75
50
25
0
–25
–50
1960 1965 1970 1975 1980 1985 1990 1995 2000 2005 2010
to 82% Graph 13.3.1 shows how well adding these three shock variables
explains the large previously unexplained drops in savings in the three
shock years.
In first differences, the dependent variable (personal savings) was
stationary. Seven of the 13 right-hand-side variables in model were non-
stationary in first differences: government spending, the current year stock
market average, the twice lagged stock market index, young to old age
ratio in the population, population, and the lagged money supply average
and population, but all were cointegrated with the dependent variable so
no detrending was required.
Explained Variance
Using the first-out process, the most important determinants of positive
personal saving are disposable income, wealth, and to lesser extent, rising
interest rates. Negatively affecting personal saving were the growth of (the
non-M1 components) of the M2 money supply, the 99 shock caused by
changing the savings definition, the Katrina 05 shock, and the 1993 tax
increase (Table 13.3.1).
300 13 THE SAVINGS FUNCTIONS
Using the first-in approach, increases in current year wealth have the
most positive effect on personal savings. The most important negative
effects on personal saving result from increases in consumer borrowing,
the Katrina Hurricane 05 shock, the 99 shock, and tax increases.
Model 13.3.1.TR
Time Period Robust OLS Estimates of Determinants of
Personal Saving
SP = – 17.89 + 0.22(Y – TT ) – 0.14(TDef ) + 0.14PR3 – 725.37ICC0.10
(t =) (–1.6) (4.8) (–1.9) (3.6) (3.9)
– 0.03Infl3 – 195.94 Shock99 – 716.30 Shock05
(–3.1) (12.9) (–18.3)
–2.27E-25 (DJ9 + DJ9–2 ) – 0.16CB2
(–5.8) (–4.4)
R2 = 78.7% D.W. = 2.1 MSE = 35.82
(13.3.1.TR)
Model 13.3.1.TR.a
Time Period Robust OLS Estimates of Determinants of
Personal Saving
(Two Variables Deleted)
SP = – 21.11 + 0.25 (Y – TT ) – 0.18(TDef ) + 0.14PR3
(t =) (–1.0) (1.9) (–2.4) (3.2)
– 1179.80ICC0.10 – 0.02 Infl 3
– 202.03 Shock99
(2.8) (–5.5) (5.6)
– 172.58 Shock05 R2 = 78.7% D.W. = 2.1 MSE = 35.82
(–16.7)
(13.3.1.TR.a)
Clearly, the signs, magnitudes and statistical significance of the remaining
variables are robust to the model change, with very little variation generally
occurring, except possibly for the magnitude of the consumer confidence
variable.
Adding the money supply and the 1993 shock to the full time period
robust model
Model 13.3.1.TR.b
Time Period Robust OLS Estimates of Determinants of
Personal Saving
(Two Variables Added)
SP = –16.87 + 0.25(Y – TT ) – 0.13(TDef ) + 0.10PR3
(t =) (–1.7) (4.7) (–2.29) (2.3)
0 – 0.03Infl – 187.99 Shock99
3
– 766.79ICC0.1
(3.7) (–3.1) (9.6)
– 191.23 Shock05 – 2.63E-25 (DJ9 + DJ9–2 ) – 0.16CB2
(–14.0) (–5.4) (–3.9)
– 0.19 M1Real – 18.99DJ
(–1.4) (–1.3)
R2 = 80.1% D.W. = 2.3 MSE = 35.51
(13.3.1.TR.b)
Here again, coefficients, signs, and statistical significance of the model are
robust to these added variables.
We conclude the personal savings model results given in model
13.3.1.TR are robust to variation over time, and reasonable variation in
model specification.
CHAPTER 14
Model 14.1
OLS Government Receipts Model
Table 14.2 Robustness over time – government receipts (assumes 1993 tax
increase repealed by 2001 tax cut)
The GDP, inflation and unemployment variable, and the 93 shock res-
ults are robust to changes in period sampled. The results for the 1986
tax cut are ambiguous: the sign is positive for samples including only
1960–2000 data, indicating the tax cut increased tax revenue. But when
2001–2010 data is added to the sample the effect of the tax cut on gov-
ernment revenue is not significantly different from zero. One (untested)
hypothesis is that this may reflect the difference in the two key aspects
of the 1986 cut: a ∼25% cut in rates, and indexing and when they made
a noticeable difference. The rate cut would be felt quickly; the indexing
effect would manifest itself more slowly, taking until 2001 to offset the
insignificant effect of the rate cut on government revenue. (The observed
non-significant effect of the rate cut on total revenue is consistent with
the notion of a backward bending supply curve, i.e., that tax cuts finance
themselves.)
The 1993 tax rate increase also appears to have had a positive, but
not statistically significant effect on revenue 1994–2000. There was some
uncertainty as to whether to model the 1993 tax increase as effect-
ive only through 2000, or continue it until 2010; the model above
shows the results of modeling its effects only through 2000, on the
assumption the Bush tax cut of 2001 restored the pre-1993 rates.
Below we repeat the same time period robustness tests, using exactly
the same models except for an alternate 1993 shock variable which
contains the assumption the 1993 tax increase on high-income earners
continued in effect through 2010. Doing so gives results consistent with
the notion the tax increase had a negative effect on total government
revenues.
306 14 DETERMINANTS OF GOVERNMENT RECEIPTS
Model 14.1.Alt
Determinants of Government Receipts Assuming 1993
Tax Increase Continued Through 2010
Table 14.3 Alt robustness over time (assumes 1993 tax increase continues
through 2010)
Model 14.1.TR
OLS Government Receipts Model
Model 14.1.TR.a
OLS Government Receipts Model (One Variable Deleted)
Model 14.1.TR.b
OLS Government Receipts Model (Two Variables Added)
Endogeneity of Government
Spending Levels
1/3 of all the variation in government spending during the period. Initial
OLS model results are shown in Eq. 15.1.1. The more limited, but time
period and model change robust model is presented as Model 15.1.1.TR
further below.
Model 15.1.1
OLS Total Government Spending Model
GT& I = 61.69 + 0.03 (GDP–0 ) + 23.85(UNEM)
(t =) (4.6) (0.5) (3.2)
+ 0.028 (Pop. Size–31–21 ) + 64.66(Viet. Build Up)
(4.5) (4.9)
+ 34.85(Reagan, Iraq Build Up) + 74.56 (Fin. Crisis Shock’09)
(2.1) (4.5)
R2 = 0.66; DW = 1.8; MSE = 30.62
(15.1.1)
Attempts to insert Okun’s law for the unemployment variable gave the
expected statistically significant signs for the GDP growth rate, but
explained some, but not as much variance as using the unemployment
rate itself as an explanatory variable. It is also worth noting that GDP
tested with 1–4 lags also was not significant (even if unemployment vari-
able was removed from the model). The sign on the GDP variable is
positive for all lags tested except the first year lag which was negat-
ive, presumably indicating major increases in government spending on
transfer payments via unemployment insurance in the year after a decline
in GDP.
The unemployment rate and GDP variables were tested, but not found
endogenous with government spending, so no 2SLS model was needed.
In first differences, the dependent variable (government spending) was
nonstationary. One of the right-hand side variables in model was non-
stationary in first differences: population changes between 21 and 31 years
earlier periods but it was cointegrated with the dependent variable, so
detrending was not required.
From the first-out perspective, the variables which accounted for the
largest amount of variance which they alone could account for were the
population growth and Vietnam build up variables, with Iraq and unem-
ployment next. From the first-in perspective, unemployment, Iraq, and the
shock in 2009 were the most important.
Model 15.1.1.TR
OLS Total Government Spending Model
Model 15.1.1.TR.a
OLS Total Government Spending Model
(One Variable Eliminated)
Model 15.1.1.TR.b
OLS Total Government Spending Model (One Variable Added)
All variables remain relatively stable except the significance of the Iraq
military spending shock variable, which becomes insignificant. Notice the
M1 variable is not statistically significant. We expect that when a variable
which is significant when included alone in a regression are included with
a highly correlated second variable, significance levels on both drop. This
multicollinearity problem appears to be the culprit here. Initially, adding
it to a model with a (significant) Iraq variable results in two insignific-
ant variables. Remove either one form the model, and the remaining one
becomes significant. Hence, the lack of robustness is a sign of a technical,
not a substantive problem.
We conclude the time period robust model is also generally model spe-
cification robust, though one variable is vulnerable to the multi collinearity
problem if the M1 money supply is added to the model.
Model 15.2.1
OLS Model of Government Spending on Goods and Services Only
Table 15.2.2 Robustness over time – government spending (goods and services
only)
Model 15.2.1.TR.
OLS Government Goods and Services Spending Model
GG& S = 2.76 + 0.10 (GDP–2 ) + 30.88(Viet. Build Up)
(t =) (0.5) (5.8) (3.1)
+ 32.32 (Iraq Build Up) – 17.43 Shock09 + 0.48 AR(1)
(6.3) (–7.0) (4.0)
R2 = 0.68; DW = 1.8; MSE = 19.60
(15.2.1.TR)
316 15 ENDOGENEITY OF GOVERNMENT SPENDING LEVELS
Model 15.2.1.TR.a
OLS Government Goods and Services Spending Model (One
Variable Eliminated)
GG& S = 5.56 + 0.10 (GDP–2 ) + 29.69 (Iraq Build Up) – 17.12 Shock09
(t =) (0.7) (6.2) (6.2) (–7.6)
+ 0.58 AR(1) R2 = 0.65; DW = 1.6; MSE = 20.24
(7.5)
(15.2.1.TR.a)
The remaining variables’ coefficients and significance levels are nearly
identical to the full model, therefore the model is very robust to this
particular change.
Adding the population variable back to the core robust model
15.2.1.TR and re-estimating:
Model 15.2.1.TR.b
OLS Government Goods and Services Spending Model (One
Variable Eliminated)
The GDP variable is very robust to this change; the population size, Viet-
nam build up, and 2009 shock coefficients change 25–33%, though their
significance levels remain fairly stable.
Overall, the model 15.2.1.TR seems reasonably robust to changes in
model specification as well as robust to changes in time period sampled.
CHAPTER 16
We already know the model of the real economy presented in this book is
quite robust in explaining variations in investment and consumer behavior
within the sample period: R2 s are generally 0.90% or higher; coefficients
are generally stable over different sample periods and model specifications.
However, here is nothing scientific about our work unless the models
found to be good at explaining within-sample period variance are as good
at explaining behavior 5–10 years beyond the sample period. We must
show that our results have identified the true underlying, (reasonably) time
invariant, structure of the economy, not just (nonrepeating) peculiarities
of the particular set of data we examined. We now wish to examine how
well the structural models employed in this study explained the economy’s
behavior in years long after the end of the period used to estimate the
model. Since our data covered the period 1960–2010, not a lot of time
has passed that we can use to test the hypothesis that the model is robust
for extended periods of time outside the period sampled – up to 10 years.
To do this we re-estimated our successful models using only data from
1960 up to the year 2000. The theory was that if our “successful” models
really were successful at uncovering the true structure of the economy, the
model would yield similar parameter estimates m the 40-year sample data.
Since the same (true) structure would underlie the economy in the out
of sample period, the 1960–2000 parameter estimates should explain the
out-of-sample data as well.
Obtaining parameter estimates this way for all the explanatory variables
in the model, we then use these parameters and the actual values of the
explanatory variables that occur during the period 2001–2010 to calculate
the model’s estimates for the values of the GDP, C, and I during these
10 years after the end of the period covered by the sample. These estim-
ates will then be compared to the actual data. The basic model used to
determine the GDP is
GDP = CD + ID + G + X
Here again the estimates from the 1960–2000 sample are quite similar
to the full 50-year period model (and for the same reason – the model
explains investment behavior about equally well in any decade).
In Section 6, Eq. 6.1 showed the parameter estimates for the model
of domestically produced goods for export for the full 1960–2010 data
sample. Results are repeated here:
Y = real GDP
(Y-TT ) = disposable Income
TT = total government receipts
PR0 =prime interest rate (calculated from Taylor rule theory of rate
determination)
PR–1 = “ ”
PRAV–1–2 = “ “
PR–2 = “ “
DEP = depreciation allowances retained by business
X = exports
Infl = inflation
16.2 MODEL 2: TREATING C, I, AND X MODEL DETERMINANTS FOR WHICH WE HAVE. . . 323
The models used were presented earlier in the text and are repeated
here for reference. Results are shown in Model 2 below:
(1) GDP
(5) Exports
Table 16.2.1 How well models 1 and 2 fit the data for the 10 periods following
the 1960–2000 estimation period (billions of 2005 dollars)
The extended model appears to fit the data for a decade beyond the
estimation period less well than the simpler model, but still reasonably
well relative to GDP size, since we can say our average estimate of GDP
for the 10 years after the model’s parameters were estimated is only 1.7%
off the actual GDP. This we take as a sign of internal consistency within the
different regression results we obtained for the 45 component equations
of this large scale econometric model. That said, the average yearly error
of estimate is about ¾ of the actual size of the change in GDP, which is
not so good. The better estimator is Model 1. Since each of the Model 1
determinants was replaced in Model 2 inexactly by its own determinants,
this result should not surprise us.
It was noted earlier that in general, adding the coefficients of variables
from the previously tested subcomponents C, I, and X of the GDP gave
more rational estimates of parameter estimates of their relationship to the
subcomponent (and therefore to the GDP when added together) than
just running a regression on the same variables, using CD + ID + G + X
as the dependent variable. We noted parameter estimates in the aggregate
regression can vary considerably, even their sign, compared to what was
obtained in desegregated function (See Section 7.1). The same problem
also affects how well the two approaches (Models 1 and 2) estimate the
GDP in the decade after estimation. Below we compare actual GDP to
calculated figures for 2001–2010 based on the parameters obtained from
regression of GDP on its determinants (16.3.16, repeated below),
Table 16.2.2 How well models 1 and 2 fit the data for the 10 periods follow-
ing the 1960–2000 estimation period (nine additional equations substituted for
variables treated as exogenous in Model 1)
Yearly Single GDP regression model Sum of C & I Eq. coefficients, plus G+X model
change
(GDPACTUAL – GDPPRED ) (GDPACTUAL – GDPPRED )
(as a %
of GDP) GDPACTUAL GDPACTUAL
But 8.1.1.1 was calculated from Eqs. 4.4 and 5.4 estimated using the
whole 1960–2010 data set. Re-estimating those equations (Eqs. 4.4.16
and 5.4.16 presented earlier) using only the 1960–2000 data set is actu-
ally what we did earlier in Model 1, where the average error for the decade
after the estimation period was found to be 0.46 of 1%.
The average error of the GDP was calculated as (Actual GDP-calculated
GDP)/Actual GDP). For the model (7.1.1.16) above which just did one
regression on all the C and I equation determinants, plus G and X, for the
10 years 2001–2010, the average error of fit was 2.1% as shown in the first
column of Table 16.2.2. This was 4.6 times more than the 46/100 of 1%
error for the GDP estimates obtained for our first model which added
330 16 CAPACITY OF THE MODEL TO EXPLAIN BEHAVIOR. . .
C&I model parameter estimates for each variable, plus G and X (all
adjusted by the multiplier).
The reason for discrepancy is that the single regression of GDP on a
variable captures correlational as well as causal relationships the variable
may have with the C and I subcomponents of the GDP. The observed
coefficient is the sum of the causal effect measure in one subcomponent
of GDP and the merely correlational (spurious) effect found in another.
The approach of adding up only measured effects of a variable only from
subcomponent equations in the GDP in which the variable was statistic-
ally significant, or at least left in the equation because it was considered
theoretically important, avoids this. The calculated errors are presented in
Table 16.2.2.
CHAPTER 17
the two theories of what causes the SRAS curve to be up-sloping have
exactly the opposite directions of causation: diminishing returns suggests
decisions to increase production forces the increase in prices just to main-
tain profits, while the increased profits view suggests increased prices leads
to increased production.)
Either the flat or the up-sloping SRAS curves are consistent with Keyne-
sian mechanics; both allow for shifts in aggregate demand to effect the
GDP, the result predicted by traditional Keynesian IS/LM mechanics.
Only a vertical curve is inconsistent with this possibility.
The long-run supply curve is portrayed as a vertical line drawn at the
point on the horizontal axis representing the economy’s maximum level
of (long-term sustainable) output at a particular point in time. It also rep-
resents the more neoclassical point of view that shifts in demand result in
instantaneous adjustment of prices and wages, and therefore have no effect
on real GDP (or employment).
Temporary deviations along the SRAS above this level are allowed for
when for cultural or other reasons, e.g., the Christmas shopping rush,
wartime patriotism), additional supplies of labor make themselves tempor-
arily available, but only for the holiday season or national crisis. Similarly,
machinery (capital) may be used temporarily above its long-term sustain-
able level by temporarily ignoring the need for down time to perform
preventive maintenance such as lubrication, cleaning, etc. This also allows
for temporary increases in output above LRAS levels, but increases that
are not sustainable over time.
(MV)
=Y (17.2.1)
P
(PY) P P
i.e., V = = (Y) = fy (. . .) (17.2.2a)
M M M
Equation 17.2.2a indicates that holding the price level and the money supply
constant, velocity can be increased by some action that increases the demand
for real GDP, e.g., increased government demand for goods (or tax cuts)
financed by borrowing unused funds – idle reserves – from banks. This will
provide the increase in velocity desired. Put another way, any change in the
GDP not caused by an increase in the real money supply must be the result of
some action that increases velocity, and the same effect occurs regardless of
whether we measure in nominal or real values if the price level is constant.
As we will show below, increasing government spending or cutting taxes
does increase velocity, provided they induce or increase the government
deficit.
Financing either tax cuts or government spending by means of a deficit
means financing them out of money borrowed from a bank or other finan-
cial institution where people are storing their savings until needed. Using
these savings to finance spending now (that would otherwise not occur
now if the money were left unused in banks), increases aggregate money
demand, and in Keynesian mechanics this elicits an increase in the GDP.
If M2 stays constant, it increases V. (If we are only using M1 money in
Fisher’s equation, M1 might increase if the loan to the government is fin-
anced out of non-M1 components of M2 (like savings deposits) available
as excess reserves, leaving V constant).
Bank lending to the government out of savings which the bank expects
will remain unused and in the bank until a future period, clearly increases
the current GDP by increasing the number of times the same money is
used to purchase goods in the current period, i.e., by increasing the rate
at which the same money is turned over in successive sales (velocity). This
holds for loans financed out of excess reserves. This is basic Keynesian
stimulus theory as it affects the economy in the short run.
Long-run effects may be different. In later periods, government repay-
ment of debts to banks allows the bank to repay its debt to the original
saver when the original savers wish to withdraw their money. This repay-
ment may reduce the government deficit, as it requires revenues in excess
of those earmarked for spending, thereby reducing the GDP (and velo-
city) in that future period. This would leave the net long-run effect of the
stimulus program zero. Alternatively, the government can just roll over
the debt in perpetuity, allowing repayment of the old debt out of issuance
17.2 THE AGGREGATE DEMAND CURVE AND THE ROLE OF VELOCITY. . . 335
those who will spend it in its entirety on goods and services, not just on
purchases of other existing securities, whose main effect will be to increase
existing security prices, not raise the GDP.
Since the effect of deficits can have different theoretical outcomes when
considering the current period alone (depending on repayment schedules
and crowd out effects), the effect of deficits on velocity for the current
period becomes an empirical question, answered in the sections below.
In the following sections, OLS and 2SLS tests are performed on selected
variables having a theoretical basis for consideration as determinants of
velocity. The theory is an extension of Eq. 17.2.2a above
The initial OLS model tested estimates the net effects of changes in taxes
and government spending on M1 velocity. Normally, without crowd out,
we would expect a positive sign on the coefficient for government spend-
ing and a negative sign on the tax variable (as we found with our GDP
tests of crowd out earlier in the chapter). This would in fact be the case
if only state and local taxes were collected. For the day a taxpayer reduces
his checking count M1 by paying his taxes, the state or local government
deposits it in its own checking (or possibly savings) account at the com-
mercial bank it uses to pay state bills. However, in either case, paying
taxes generally will not affect the M1 money supply, and certainly not
M2. However, when a taxpayer pays federal taxes, reducing commercial
bank balances of M1, the check is deposited in the U.S. Treasury depart-
ment’s checking account (TGA account) at the Federal Reserve. Though
17.2 THE AGGREGATE DEMAND CURVE AND THE ROLE OF VELOCITY. . . 337
This model shows the stimulus effects of fiscal policy to cut total taxes
or increase total government spending in the form of the coefficients
on those variables (TT , GT&I ) in Eq. 17.3.1. However, these coeffi-
cients are “net” of some (described below) but not all crowd out effects.
Even with only some crowd out accounted for, we see that the effect
of a deficit financed tax cut on velocity (and therefore GDP) is negat-
ive. For government spending, the net effect shown still leaves the effect
of increased spending on velocity positive (V1, and therefore real GDP,
increase 0.43 for each dollar of additional government spending), but not
to the extent expected theoretically in Keynesian mechanics, i.e., 1/(1–
MPC) (GT&I ) where the multiplier is generally assumed to be greater
than one.
But this only partially shows the extent to which crowd out offsets
traditional Keynesian stimulus effects, as indicated in the note below:
(Note: Total crowd out effects include the reduction in private spending equal
to the amount of reduced private borrowing, plus additional negative effects
on spending that may occur because borrowing is typically done for “big ticket”
items, partially financed by consumers and businesses out of income, partially
out of borrowing. When borrowing declines, total spending may decline some
multiple of that. This multiple effect is the only part of crowd out’s effects
that show up in the coefficients on taxes and government spending when bor-
rowing levels are explicitly controlled for in the model, as in 17.3.1 using the
(CB2 + (IB(–1) ) variable. Results can be positive for both the tax and spend-
ing variables if crowd out effects dominate stimulus effects for tax cuts, but
don’t for government spending. In an attempt to disentangle, we have added
a variable (the government deficit) to explicitly pick up crowd out effects and
re-estimated the model in Eq. 17.3.2.
Another way of showing the crowd out effect of deficits’ net negative
effect on V1 is to add the deficit size as a variable explicitly to the model,
retaining variables showing the general effects of taxes and government
spending deficits. This is done in Model 17.3.2 below, which is just Model
17.3.1 with the deficit (TT –GT&I ), net of changes in the size of the pool
of loanable funds, added as the first variable in the model.
Model 17.3.2
OLS V1 Velocity Model (Deficit Variable Added)
Note: All right-hand-side variables multiplied by P/M1. Model includes
explicit deficit variable and private borrowing control variable
17.3 OLS TESTS OF M1 VELOCITY’S DETERMINANTS 341
Model 17.3.3
OLS V1 Velocity Model (Deficit Variable Added, Borrowing
Variable Deleted)
V1 = 0.65(TT – GT&I ) – 0.21(TT(–1) ) + 0.89(GT&I ) + 14.19PR
(t =) (3.1) (–1.5) (2.4) (1.1)
+ 0.39DJ–0 + 0.38DJ–2 – 10.63XRAV – 755.85POP16
(1.6) (0.8) (–1.5) (–1.0)
342 17 CONVERTING THE OLDER KEYNESIAN IS-LM MODEL. . .
Model 17.4.1
2SLS V1 Velocity Model (Deficit Variable Added, Borrowing
Variable Deleted))
Note: All right-hand-side variables multiplied by P/M1. Includes explicit
deficit variable; does not control for private borrowing effects. GT&I and
DEP replaced by strong, nonendogenous instrument
Graph 17.4.1
Specification Robustness
Deleting the last two variables from the model and re-estimating gives
Model 17.4.1.TR.a
2SLS V1 Velocity Model (Deficit Variable Added, Borrowing
Variable Deleted)
(Two Variables Deleted)
V1 =1.40(TT – GT&I ) – 0.64(TT(–1) ) + 1.80(GT&I ) + 5.29DEP
(t =)(6.3) (–3.4) (3.7) (5.1)
2
– 007POP R = 94.0%, D.W. = 1.8, MSE = 0.10
(–2.2) (17.4.1.TR.a)
1.0
0.5
0.0
–0.5
0.2 –1.0
0.1 –1.5
0.0
–0.1
–0.2
–0.3
1965 1970 1975 1980 1985 1990 1995 2000 2005 2010
Residual Actual Fitted
Model 17.4.1.TR.b
2SLS V1 Velocity Model (Deficit Variable Added, Borrowing
Variable Deleted)
(Two Variables Added)
Model 17.5.1
OLS V2 Velocity Model (With Borrowing Variable)
V2 = 0.35(TT – GT&I ) + 0.19(TT(–1) ) + 0.66(GT&I ) – 3.60PR
(t =) (2.2) (1.2) (24) (–0.5)
+ 0.06DJ–0 + 0.29DJ–2 – 10.80XRAV – 443.41POP16
(0.2) (0.6) (–2.3) (–1.1)
+ 0.008POP – 0.45ICC–1 + 2.88 M2AV + 0.31(ACC)
(1.8) (–0.6) (0.1) (5.1)
+ 1.84DEP + 3.18CAP–1 + 4.66PR–2 + 0.081PROF–0
(3.0) (0.7) (0.6) (0.3)
+ 0.32(CB2 + (IB(–1) ) – 0.30(X – M)
(3.3) (–0.7)
R2 = 95.1%, D.W. = 1.6, MSE = 0.015
(17.5.1)
348 17 CONVERTING THE OLDER KEYNESIAN IS-LM MODEL. . .
Model 17.5.2
OLS V2 Velocity Model (Without Borrowing Variable)
The “core time period” robust model is defined as those variables that
were statistically significant in at least three of the four periods sampled.
In addition, we individually add to the core any variable found significant
only two or less times in the original tests of the four-period samples. All
variables now found significant are collectively added to the core model
variables and retested in all four periods. Those found significant in at
least three of the four sample periods, along with the core, become the
final time period robust model, shown below as Eq. 17.5.2.TR:
Model 17.5.2.TR
Time Period Robust OLS V2 Velocity Model
(Without Borrowing Variable)
V2 = 0.64(TT – GT&I ) + 0.48(TT(–1) ) + 0.59(GT&I ) + 0.43 ACC
(t =) (4.9) (3.1) (2.9) (8.3)
+ 1.96DEP + 0.48 AR(1) R2 = 93.1%, D.W. = 1.6, MSE = 0.015
(2.6) (3.1)
(17.5.2.TR)
17.5 OLS TESTS OF M2 VELOCITY’S DETERMINANTS 351
Model 17.5.2.TR.a
Time Period Robust OLS V2 Velocity Model
(Without Borrowing Variable)
(One Additional Variable Deleted)
V2 = 0.77(TT – GT&I ) + 0.62(TT(–1) ) + 1.00(GT&I ) + 0.44 ACC
(t =) (6.3) (4.6) (6.1) (11.3)
+ 0.48 AR(1) R2 = 91.8%, D.W. = 1.6, MSE = 0.016
(3.1) (17.5.4.TR.a)
Significance levels of the remaining variables remain high; coefficients for
the deficit and total government revenues and the accelerator remain fairly
constant. However, the coefficients on the marginal effect of government
spending grow by two-thirds.
Adding the profits variable to the final time period robust model
17.5.2.TR and re-estimating, we get:
Model 17.5.2.TR.b
Time Period Robust OLS V2 Velocity Model
(Without Borrowing Variable)
(One Additional Variable Added)
V2 = 0.60(TT – GT&I ) + 0.49(TT(–1) ) + 0.57(GT&I ) + 0.41 ACC
(t =) (4.5) (2.9) (2.7) (7.0)
+ 1.94DEP + 0.30PROF–0 + 0.43 AR(1)
(2.7) (1.9) (2.5)
R2 = 93.5%, D.W. = 1.7, MSE = 0.015 (17.5.4.TR.b)
The model estimates are highly robust to the addition of the profits
variable.
Overall we conclude that the model is robust to time period sampled
and largely robust to additions and subtractions of variables from the
model, with the exception of the coefficient on the government spend-
ing variable. The time period/specification robust model is graphed in
Graph 17.5.1.
352 17 CONVERTING THE OLDER KEYNESIAN IS-LM MODEL. . .
0.15
0.10
0.05
0.00
–0.05
0.04
–0.10
0.02 –0.15
0.00
–0.02
–0.04
1965 1970 1975 1980 1985 1990 1995 2000 2005 2010
Residual Actual Fitted
Recall that in the initial models of velocity, all variables found signific-
ant determinants of GDP in consumption and investment models, plus
net exports, were included, a total of 17 variables (see 17.3.1, 17.4.1).
17.6 WHICH DETERMINANTS OF GDP ARE ALSO DETERMINANTS OF VELOCITY 353
Though all were significant determinants of GDP, only some were found
to be significant determinants of either M1 or M2 velocity. This may have
been for technical reasons related to multicollinearity or degrees of free-
dom relative to the large number of variables used (as was the case when
we directly statistically estimated the GDP function using these 17 vari-
ables as the explanatory variables. However, it may also have been because
not everything that raises GDP necessarily raises velocity, even when the
money supply is held constant.
Why Certain Variables Affect Velocity
Possible reasons for the findings of significant impact on velocity (or lack
thereof) are noted below, explained in terms of their impact on velocity in
the equation
Which is the same thing as saying that anything that raises the real GDP,
real money supply constant, raises velocity, yet some factors were not
found statistically significantly related to velocity.
Five factors were found to affect both M1 and M2 money supply
velocity:
Tax or spending deficits: Borrowing to finance the deficits is usually done
out of unused M1 funds held by the investment community or by banks
(in the bank’s own corporate checking account). When open market pur-
chases of government bonds are made, the government bonds are paid
for by check, and deposited in Treasury’s bank account at the Fed. This
deposit reduces the money supply, but when this deposit is re-spent by
the government to finance the spending deficit, it is generally deposited
by the recipient of the transfer payment or seller of the goods and services
in their checking account, or kept as cash. This raises the nominal M1
money supply back to its former level. The nominal M1 money supply in
Eq. 17.2.1 or 17.2.1.a is unchanged. If the price level remains constant,
the real money supply must also remain unchanged, and this must mean
M1 velocity increases.
Alternatively, the bond purchase may be made by the investment com-
munity by converting unused (non-M1) M2 money, e.g., savings account
deposits, to M1 money, raising the M1 money supply, but not changing
M2. It is then spent in open market purchases of government bonds issued
354 17 CONVERTING THE OLDER KEYNESIAN IS-LM MODEL. . .
to finance the deficit. Here again, the government deposits this in Treas-
ury’s account at the Fed (reducing M1and M2), and then subsequently
re-spends it on goods, services or transfer payments. The recipient sub-
sequently deposits it in their M1 account (restoring the earlier M1 and M2
balance. Or they may deposit it in their (non-M1) M2 account, restoring
the original M2 balance. If the price level remains constant, the GDP rises,
though M2 (and perhaps M1) remains unchanged. Hence M2 velocity
must increase, since it is but the calculation of (P/M2)*Y. (The increased
M1 money supply would mean M1 velocity would remain constant. Our
statistical results confirm that a change in GDPReal *(P/M1 or 2 ) is posit-
ively correlated with V1 or V2. If the (P/M1 or 2 ) portion is constant, this
must mean V1 or V2 rises with a rise in real GDP.
The financial community uses its M1 to buy government bonds in
open market operations. This M1, when received by the government, is
taken out of circulation and deposited in the Treasury’s checking account,
reducing M1.
However, the tax cuts financed by deficits leave extra money in people’s
M1 or non-M1 M2 accounts, or available for use financing goods and
services purchases. The money used for purchases would be deposited by
sellers in M1 or M2 accounts), If the tax cut just stays in M1, M1 stays
constant despite the deficit. If the tax cut stays in other parts of M2 as
savings, M2 stays constant. In either case, since not used to raise the GDP,
velocity stays constant. If the taxpayer’s unexpected M1 windfall money is
used to purchase goods and services, the GDP rises. If the M1 used to buy
the goods is deposited by the sellers in their M1 accounts, the M1 money
supply again remains constant, but since the GDP rises, so V1 must rise.
Taxes (controlling for deficits and effects on private borrowing) were
found negatively related to velocity, suggesting a cut in taxes raised velocity
(holding money supply constant). Keynesian fiscal stimulus theory sug-
gests tax cuts should positively affect the GDP (controlling for effects on
private borrowing). Fisher’s equation suggests that velocity and GDP are
positively related (holding the real money supply constant). Hence our
finding of a negative relationship between taxes and velocity is consist-
ent with both Fisher and Keynes. Similarly, a rise government spending
controlling for the deficit means an equal rise in taxes, causing a positive
effect on GDP for the same reason. For both tax cuts (+0.48) and spend-
ing increases (+0.59), this is what our empirical results show: for each
$1 of stimulus, split equally between tax cuts and spending increases, a
slight stimulus effect of 1.06 occurs, slightly larger per dollar of spending
increase than the stimulus itself.
17.6 WHICH DETERMINANTS OF GDP ARE ALSO DETERMINANTS OF VELOCITY 355
Exchange rate: If the price of foreign goods drops because the cost
of foreign currency drops (which would show as a rise in the variable
(XRAV0 to –3 ), the price level indicator used in this study (GDP deflator)
may change somewhat since the price of American produced goods
includes the cost of foreign components. But since imported compon-
ents constitute only a portion of the price, the reflection of changes in the
price of foreign goods in the price of American goods should be less than
complete, though some downward change may occur, freeing up money
for expenditure on additional foreign or U.S. goods. Put another way, real
income rises.
This would be in addition to normal income and substitution effects
resulting from foreign goods of all types being cheaper, which would
occur when the price of foreign goods drop. Some of the existing M1
or M2 money supply, no longer needed to purchase foreign goods,
may shift toward purchase of domestic goods (if income effect domin-
ates substitution effect), raising the real GDP. The real money supply
(M1Nom /GDPDeflator) may rise as much (if nominal money is deflated
using a deflator that includes the prices of foreign goods). If it does, no
change in velocity, since (V = YReal /MReal ). If it doesn’t, the resulting rise
in GDP will cause velocity to increase (which is not what we observe,
since the exchange rate variable’s sign is negative, suggesting the substitu-
tion effect is more dominant). If the substitution effect is stronger than the
income effect, the drop in foreign currency prices will result in a net switch
in demand from U.S. to foreign goods, reducing real U.S. GDP. This will
cause a decline in velocity, which is what our empirical results show. (Some
increase in the real money supply may also occur if the money deflator
includes the prices of foreign goods, which would also decrease V since
(V = YReal /MReal ). This study uses the GDP deflator, so the real money
supply does not change, except perhaps because of the effect of cheaper
foreign components in U.S. products noted in the previous paragraph.)
Population Growth: If the population grows, demand for goods and
services increases. If produced, GDP will rise. If P and M remains constant,
velocity must increase. This is what our empirical results show.
Depreciation: Increases in depreciation allowances were also associated
with an increase in V1 and V2 velocity.
The increased allowances tend to occur in periods when investment,
and therefore GDP, is increasing.
Therefore, growth in depreciation should be associated with an increase
in velocity (controlling for the real money supply), which is what this study
finds.
356 17 CONVERTING THE OLDER KEYNESIAN IS-LM MODEL. . .
5% level M1 M2
Variable (tests of first differences) Test criterion Test statistic Test statistic
V1 –1.95 –4.70
V2 –1.95 –6.47
Taxes (TT ) –1.95 –4.82 –5.03
Government Spending (GT&I ) –1.95 –7.03 –4.37
Exchange Rate (XRAV ) –1.95 –4.05 –2.26
Population Size (POP) –1.95 –4.72 –2.43
Consumer and Business Borrowing –1.95 –7.25 –6.54
Depreciation –1.95 –3.97 –4.98
M2 Levels, Past Years Av. (M2AV ) –1.95 –5.28 –
Capacity. Utilization % (CAP–1 ) –1.95 –5.54 –
Accelerator –1.95 – –9.61
Wealth Measure (DJ–2 ) –1.95 –1.47∗ –
Variables deleted in stepwise procedure:
Accelerator (ACC) 1.95 –8.93 –
Prime Rate–2 (DJ–2 ) 1.95 –8.78 –9.87
Prime Rate–0 (DJ–0 ) 1.95 –8.65 –9.99
Net Exports (X-M) 1.95 –3.72 –4.46
Wealth Measure (DJ–0 ) 1.95 – –2.14
Cons. Confidence (ICC–1 ) 1.95 –5.78 –6.70
Young/Old Age Ratio (POP16 ) 1.95 –3.70 –2.24
% Cap. Utilized (CAP–1 ) 1.95 – –3.09
Wealth Measure (DJ–2 ) 1.95 – –2.07
Wealth Measure (DJ–0 ) 1.95 –1.03∗ ∗ –2.14
∗ –1.79 when detrended, using DJAV *P/M1
–2 Nom –0.007@trend. However, failure to meet the ADF
criterion is not a problem since the wealth measure (without trend) is also cointegrated with d(V1), as its
M2Nom form cointegrated with d(V2).
∗∗ (–1.98 when detrended using DJAV *P/M1–0.01@trend. Also, DJAV without trend is cointegrated
–0 0
w/V1, hence, no detrending is needed.)
parameter estimates for these variables that were obtained regressing con-
sumption and investment on their determinants, and adding government
spending and exports. We can incorporate the calculated GDP function
(Eq. 8.1) into our velocity model to see the estimated effects of fiscal
policy variables on velocity as well as others that are determinants of the
GDP. The model, using 8.1, would be as follows:
(PY) P P
i.e., V= = (Y) = fy (determinants of Y) (Repeating Eq. 17.2.2a)
M M M
P P
i.e., V = (Y) = (C + I + G + NX)
M M
17.8 ALTERNATIVE METHOD: CALCULATING IMPACT OF DETERMINANTS OF GDP. . . 361
Clearly the results indicate that the marginal stimulus effect of a cut in
government revenue (–0.41TTot net of crowd out effects +0.86TTot) is
that velocity declines. The effect of an increase in government spending
on velocity, net of crowd out, is positive (+1.41–0.86), though whether
statistically significant is uncertain. This approach does not suffer from the
problem of estimating the GDP directly, i.e., the combination of effects of
a determinant that are correlational as well as causal.
CHAPTER 18
Dynamics
18.1 INTRODUCTION
You can think of the results obtained in Chapters 4–17 as the basic sci-
ence describing how the macroeconomy operates. Parameter estimates
indicate how a change in one explanatory variable can affect that equa-
tion’s dependent variable, holding the other variables in the model constant.
This ceteris paribus methodology allows us to avoid attributing to one
explanatory variable the influence of another.
But it does not deal with the very real fact that changes in the dependent
variable often subsequently cause “feedback” changes in some or all of the
explanatory variables. That is, there are feedback loops that cause a shock
to the system to have a direct effect on (say) GDP. This change in GDP
can then cause a subsequent change to the variable that caused the shock,
or other explanatory variables. These changes cause further changes in
GDP, which causes even further changes in the explanatory variables, etc.
Determining the path of these changes, until a new equilibrium is reached,
is the business of economic dynamics. Shocks we wish to evaluate are typ-
ically fiscal or monetary policy shocks, though shocks in other variables in
the system can also be of interest.
Unlike models with a limited number of explanatory variables, the
Cowles model described in this volume does not need Chomsky decom-
positions or other alternative ways of isolating the effects of changes in
fiscal or monetary policy variables (or any other variables) to simulate their
effects on the GDP. At least in theory, Cowles models, such as the one
presented in Chapters 4–17, explicitly contain all the variables that affect
the economy, and parameter estimates indicating how each of them affects
other variables. Hence, there is no need to observe changes in the error
term in a model before and after a known shock, and try to disentangle
the policy change’s effect on the error from other factors.
For example, the Cowles model developed in this volume provides
regression estimates of the current period effects of tax, government
spending, interest rate or monetary changes on other variables in the sys-
tem. They measure the marginal effects on a model’s dependent variable
of a change in some explanatory variable of interest.
However, the models do not show the feedback loop effects over time
of such initial changes. Estimating such dynamic effects is not a science
problem; the science has been already done (as in Chapters 4–17). It is
a mathematics problem to be solved using these coefficients provided by
the science. This can be a complex problem in a model with a large num-
ber of variables and relations. It becomes all the more complicated in a
model where some of the determinants of a variable (say, the GDP) are
determined in the same period as the dependent variable. For example,
in the model developed in this book, the “IS” model for GDP determ-
ination contains 8 explanatory variables as determinants of current period
GDP that themselves are determined in the current period. Often, in part,
they are determined by what GDP is determined to be in the current
period, as well as by other variables in the system, some of which are also
determined in part by what GDP becomes during this period. Hence, it
appears as though the dependent and eight explanatory variables have to
be “simultaneously” determined. How should we do this?
The clue to how to solve this puzzle lies in the definition of “period”
used. Make each period large enough, say 20 years, and it would seem that
almost every variable is determined in the same period. However, make it
small enough, say a month, and, in simulations, every small change in
the system from the initial shock onward can be examined recursively. For
example, an increase in government spending resulting in more firemen
being hired in this (mini) period may increase the GDP in this (mini)
period. In the next (mini) period, the changed GDP may cause changes
in the inflation and unemployment rates. In the subsequent (mini) period,
the change in inflation and unemployment may change interest rates and
the loanable funds pool, etc. Hence, if we can discover the order in which
18.1 INTRODUCTION 365
these mini steps takes place we can show how feedback loops cause changes
in all variables in the system over time.
How do we know the order in which an initial shock to the system
translates into changes in individual variables? This is a hugely import-
ant question. Romer (2016) has insightfully pointed out that the order in
which these steps are assumed to occur can affect your ultimate conclu-
sions about how big a change in a variable of interest will be in response
to a shock.
We feel the model developed in this volume avoids this problem. As
we have noted in earlier chapters, it is remarkably successful in explaining
the variation in the economy over the past 50 years, and does so about
as well in one decade as another. It is also very successful when used to
explain out-of-sample period behavior of the economy using the parameter
estimates developed with the help of data from earlier periods. Hence, we
feel the model represents good science.
When we apply this science to solving the IS equation (Eq. 8.1.2.1TR),
substituting into it our other models for its endogenously determined
explanatory variables, we find that there is only one ordering of the (mini)
period effects that allows the system to be solved at all! And solving the sys-
tem is needed if a dynamic path is to be plotted out over time showing the
impact, period by period, of the initial shock. We found no other ordering
of effects would work.
Hence we feel we have no major “identification” problem of the type
that involves uncertainty in determining whether in response to a shock,
variable X is determined before variable Y, or vice versa. In our Cowles
model, there is no ambiguity, since the model only solves one way, and
the model, from a scientific point of view, appears to be a good one. One
issue that did remain unresolved, in the period a deficit-induced stimulus
was introduced, was whether to have its stimulus effects or its crowd out
effects occur first. We found the choice only made a marginal difference.
We chose to model the crowd out effects as occurring first.
Our model for GDP determination (Eq. 8.1.2.1TR) tells us that the
effect of a change in government spending on GDP is +.69 times the
change in government spending, or –.41 ( Taxes). The recursive pro-
cess we found had to be used to measure the dynamic path of this shock
on the GDP and its determinants was as follows:
In this way we determine the initial effects of a shock on GDP and other
variables in the system, and also the feedback effects that subsequently
occur until the new equilibrium is established.
We conclude that if all the standard statistical tests are telling you the
structural model is a good model, i. e., provides good estimates of how the
economy operates, then it is probably also telling you the order in which
things have to be resolved in order to get these good estimates.
At the time this book went to press, we were still working on the math
required to accurately articulate the whole IS curve variable set’s dynamic
path. We do have results for a slightly simpler model, given in Table 18.1
below. Unlike our earlier chapters, the results here should be considered
illustrative, not definitive. We have employed some simplifying assump-
tions about how quickly (in “mini” periods) the lagged variables become
effective, and start influencing the GDP. In our simplified model, there
is only one iteration of the model per period. In reality many iterations
18.1 INTRODUCTION 367
occur, and lags do not kick in until after the “many” have occurred. For
example, many iterations of the consumption multiplier effect occur in
the same year as an initial change in income. Lagged effects of changes
in other variables would not start until the following year or later, after
several iterations (“mini” periods) have occurred. A few other simplifica-
tions are employed. A 500 line Excel program was developed to run the
simulations. The dynamic path of GDP is presented in Table 18.1 below
for various monetary and fiscal stimulus policy options. Table 18.2 details
the effects on other key economic variables of these stimulus programs.
The dynamic paths of deficit financed $400 billion tax cuts, and $400
billion increased government spending programs are estimated, as is a
$400 billion change in the money supply. $800 billion combinations of
tax cuts and spending increases, with (and without) an $800 increases in
money supply are also simulated. The numbers were picked to represent
magnitudes associated with the 2009 U.S. stimulus program, The Amer-
ican Recovery and Reinvestment Act. Money supply growth 2007–11 was
also about $800 billion. Other initial condition variables were selected
to roughly represent 2010 values, or 2000–10 averages, as seemed most
appropriate.
Variable definitions follow the usage in earlier chapters, with a few
exceptions:
Total Savings = SPCD = Personal, corporate and depreciation savings
Loanable Funds(LF) = Total savings and Foreign borrowing (FB)
X = exports
M= Imports
PR = Prime Interest rate
Cons.Bor. = Consumer borrowing
The dynamic effects simulations are all based on the empirical model
estimated in Chapters 4–17. Clearly, the 33 period simulations suggest
that the structure of the American economy is such that both fiscal and
monetary stimulus programs are unlikely to be successful, and more likely
than not will have net negative effects. This, of course would change
if “accommodating” monetary policy did its job. But if accommodating
monetary policy worked, we would not have found the strong, statistically
significant crowd out effects so commonly found in Chapters 4–17 mod-
els. It is not clear why accommodating monetary policy does not seem to
work. More work remains to be done on this topic.
368
Time Per. $400 Increase in Gov’t. $400B decrease in Taxes $400B increase in M1 $800B Increase, 1/2 $800B Increase, 1/2
Spending Gov’t.Sp., 1 /2 in Taxes Gov’t.Sp., 1 /2 Taxes and
$800B increase in M1
1 Period Perm. 1Period Perm. 1Period Perm. 1 Period Perm. 1 Period 1 Permanent
Initial Conditions: GDP Without-Stimulus (Billions)
0) $18,000 18,000 18,000 18,000 18,000 18,000 18,000 18,000 18,000 18,000
GDP With Stimulus
1) 18,000 18,000 18,000 18,164 18,000 18,000 18,000 18,000 18,000 18,000
2) 18,091 18,091 17,902 17,902 18,055 18,055 17,991 17,991 18,099 18,099
3) 17,881 18,052 17,821 17,770 18,014 18,098 17,702 17,382 17,738 18,017
4) 17,691 17,987 17,801 17,716 17,938 18,082 17,492 16,950 17,396 17,867
5) 17,639 17,959 17,824 17,736 17,920 18,073 17,463 16,895 17,429 17,846
10) 17,721 17,996 17,843 17,776 17,954 18,078 17,546 17,122 17,540 17,933
20) 17,700 18,017 17,848 17,818 17,957 18,080 17,549 17,184 17,559 18,002
30) 17,675 18,025 17,844 17,834 17,964 18,092 17,520 18,186 18,546 18,051
33) 17,669 18,027 17,844 17,838 17,966 18,095 17,513 17,187 17,543 18,063
Table 18.2 Dynamic Effects of Stimulus Programs on the GDP (Detailed effects on other key economic variables after 33
periods)
Variables GDP T Total G Total UNEM% INFL% PR% Tot.Saving TDEF GDEF M X Cons. Personal Investment M1
PCD = LF-FB Bor Saving
Initial Conditions 18,000 3,600 3,600 5.00 2.00 3.00 2,568 0 0 3,600 3,100 500 275 2,790 1,000
Pre-Stimulus
(Period 0)
Stimulus Results After 33 Periods:
programs:
$400 Govt. 17,669 3,639 3,596 4.91 5.95 4.78 2,581 18 (19) 3,584 3,073 456 260 2,771 1,000
Spending Incr. (1
Yr. Only)
$400 Govt. 18,027 3,725 4,006 4.88 4.19 4.08 2,393 305 587 3,616 3,100 426 286 2,474 1,000
Spending Incr.
(Permanent)
$400 Tax Cut 17,838 3,230 3,596 4.83 3.56 3.88 2,236 (32) 335 3,598 3,094 413 380 2,408 1,000
(Permanent)
$400 M1 Incr 18,095 3,608 3,609 5.16 (0.93) 1.56 2,636 (62) (61) 3,636 3,132 447 326 2,840 1,400
(Non-Accom,)
(continued)
18.1 INTRODUCTION
369
370
18 DYNAMICS
Table 18.2 Dynamic Effects of Stimulus Programs on the GDP (Detailed effects on other key economic variables after 33
periods)
Variables GDP T Total G Total UNEM% INFL% PR% Tot.Saving TDEF GDEF M X Cons. Personal Investment M1
PCD = LF-FB Bor Saving
$400B Tax Cuts, 18,063 3,372 4,020 5.00 (0.09) 2.09 2,174 174 822 3,691 3,161 239 486 2,193 1,800
$400B Govt. Spend-
ing Inc., $800B
Increase in M1 (All
Permanent)
$400B Tax Cuts, 17,507 3,269 3,592 4.73 7.54 5.67 2,248 (12) 317 3,583 3,067 368 385 2,389 1,000
$400B Govt.
Spending Inc., No
Accomodating M1
(T Permanent, not
spending which is 1
year only)
Table 18.2 Dynamic Effects of Stimulus Programs on the GDP (Detailed effects on other key economic variables after 33
periods)
Variables GDP T Total G Total UNEM% INFL% PR% Tot.Saving TDEF GDEF M X Cons. Personal Investment M1
PCD = LF-FB Bor Saving
$400B Tax Cuts, 17,705 3,287 3,611 5.05 1.69 2.81 2,361 (111) 218 3,660 3,134 270 459 2,491 1,800
$400B Govt. Spending
Inc., $800 (non-
Account) Incr. in M1
(All Permanent, except
G Spending which is 1
year only)
18.1 INTRODUCTION
371
CHAPTER 19
Consumption Investment
GDP–2
Unemployment rate GDP(Trad. Ptnrs-2) Inflation rate
Pop growth Exchange RateAV Unemployment rate
Vietnam U.S. imports M1 Money Supply0
Reagan Mil. Build up Prime Interest RateAV-1-2 M1 Money Supply–1
Iraq Mil. Build Up U.S. Inflation RateAV-1-2 Gov’t. deficit (crowd out test)
Fernandez-Villaverde (2010) also found little impact for the money vari-
able in his research, noting “I must admit that I am myself less than totally
convinced of the importance of money outside the case of large infla-
tions” (Fernandez-Villaverde 2010). Our own Taylor rule model indicated
the current year liquidity effect of increasing the money supply was to
lower the prime interest rate, but that the inflation effect 1 year later fully
restored the prime rate to its previous level.
For the equations representing the individual components of invest-
ment, three additional determinants were found to affect the demand for
residential housing: consumer borrowing, the M1 level and (negatively)
the price of housing. Inventory investment was affected by two addi-
tional variables: (negatively) related to same period growth in consumer
spending, and positively related to prior year growth in GDP.
In the prime interest rate model, current period growth in M1 was
negatively related to the prime rate, as expected, but the same growth in
M1 lagged one period had a slightly larger positive effect, thus accounting
for both the liquidity and inflation effects expected with changes in the
money supply. The government-deficit variables were included as a test of
the oft-stated hypothesis that the way crowd out affects the economy is
through its effect on interest rates, We found no such crowd out effect on
the prime rate, which is the interest rate found most systematically affected
the GDP, consumption, and investment. This is perhaps not surprising
since the prime rate is not a market determined rate in the U.S., it is an
administered rate changed by the banking community to reflect changes
in the federal funds rate (also an administered rate).
For total government spending (including transfers), the variable best
explaining the endogenous part was found to be the unemployment rate.
By comparison, government spending for only goods and services was
more dependent on lagged GDP than on unemployment. More exogen-
ous effects included population growth, the Vietnam and Iraq military
build ups, and the 2008 financial crisis.
The exports demand model (unexpectedly) showed U.S. demand for
imports to be overwhelmingly the most important variable determining
the level of foreign demand for U.S. exports the same year. One explan-
ation for this is that foreigners must obtain U.S. currency to buy U.S.
exports, which, in the main, they get from sales of their goods to the
U.S. Another possible reason (probably a component part of the first
explanation) is that trade is so internationalized that much of what any
country exports is assembled from parts or other resources imported from
376 19 SUMMARY AND CONCLUSIONS (PRODUCTION SIDE OF THE NIPA ACCOUNTS)
markets to rise. This new money is not used to raise the GDP, i.e.,
buy real goods and services.
6. Securities sold by banks to the Fed for increased reserves is an excep-
tion. The increased reserves are generally used to give loans that
will be used to buy real goods and services. This fits with our find-
ings that changes in the money supply were significantly related to
spending on residential housing.
7. A new way of achieving the neoclassical synthesis is introduced using
Fisher’s income equation of exchange (MV = PY), instead of the
Phillips curve to meld together Keynesian short-run and neoclassical
long-run mechanics. The method shows that Keynesian fiscal policy
shifts the aggregate demand curve through its effects on velocity, a
key variable in Fisher’s equation.
PART II
income for the factor, the other presents findings on variables related
to variation in the factor’s share of national income over time. These
four factors’ incomes, combined with proprietor’s income, total national
income. Proprietor income is taken as a given because of the difficulty
separating its profit and labor income components.
The focus of this chapter is on the distribution of income, not the pro-
duction of income. Total income is determined by how much product
is produced. In the best year-to-year explanatory models, production of
goods and services is largely a function of the demand for them. In the
long run, technological progress is the key. This chapter contains 11 addi-
tional equations to describe how the distributional of income generated
producing the GDP is determined, given the level of production. Eight
are behavioral and three are identities.
The eight behavioral equations are developed in Section 20.4.
Section 20.4.1 identifies the determinants of labor’s level and share of
national income, Section 20.4.2 does the same for profit’s share and level.
Sections 20.4.3 and 20.4.4 deal with rental and interest income in the
same way. In each part, results obtained are tested exhaustively for robust-
ness over time of results and in different models, to ensure as best as
possible that results for any particular variable thought to be a determ-
inant of factor shares are not merely idiosyncratic to the particular time
period or model tested.
Besides the eight behavioral equations, we include three standard eco-
nomic identities connecting the income and product sides of the GDP to
each other, and to national income:
not the level of labor income. It was the result of increased investment by
U.S. companies in productive capacity overseas. The rest of the long-run
decline was due to a long-term slowdown of productivity growth. Sub-
stantial short-run changes in labor’s share around these long-term trends
are found to be common, driven mostly by Keynesian changes in demand.
Capital is relatively fixed in the short run, and labor is not. Therefore,
increases in demand are initially met by short-run increases in the labor
force. With capital constant, this increases labor’s share.
The decline in labor’s share due to growing foreign investment was pre-
dictable. When a new factory is built (or bought) in a foreign land with
a domestic country’s investment, the profits flowing back home and are
included in the domestic country’s national income, increasing capital’s
share. Labor income is counted in the foreign country’s national income,
since that’s where it is generated. Hence, other things equal, increased for-
eign investment over time by U.S. firms should lead to a larger share of income
for capital, and a lesser share for labor in the U.S.
The major determinants of profits’ growing share were rising profits
from foreign operations, a cheaper dollar making U.S. goods more
competitive, and increases in labor productivity, which appears mainly
to benefit profits. Rental income’s major determinant was house prices.
Growth in housing prices was found to decrease rent’s share, presumably
indicating a shift in housing preferences toward home ownership. Interest
income was driven by prime and bond interest rate levels and the level of
employment relative to GDP (growing employment stimulates borrowing
among those who can now afford to).
Table 20.1.1.1 Index of real profit and labor income growth 1929–2010
(1960 = 1.00)
Sources: EOP (1963), Table C11; (1974), B3; (2010 and 2012), B28, B3
growth in labor income. The profit index grew from 3.09 to 5.77; the
labor index only grew from 4.07 to 4.49. Most to the profit growth
was in 2003–2005, when the index grew from 3.40 to 5.77. However,
profit growth was not because labor income declined; it just grew slower.
The labor income index grew from 4.07 to 4.49 for the whole decade
and achieved most of its growth during 2005–2007, growing from 4.23
to 4.40.
The overall picture shown by Table 20.1.1.1 is one in which the level
of labor’s income grew more rapidly than profits coming out of the
depression, and in the more normal times which characterized the over-
all 1960–2000 period. By comparison, in boom periods characterized by
the World War II decade and the 2001–2010 decade, particularly the
super boom years of 2006–2007, profit income grew more rapidly than
labor income. But profit income did not grow by redistributing labor
income. Labor income also increased during both boom periods, just not
as fast.
Table 20.1.1.2 shows labor and profits’ share of national income, rather
than their level. It indicates that both labor’s and profits share of national
income grew during the recession decade 1930s, because the shares of
other types of income fell more precipitously. Labor and profit income
shares grew again during the boom decade 1940 and 1950. Labor’s share
increased 1950 until 1980, increasing from 64% to 68%, but since then has
declined to 62%. Profits share declined 1950–2000 but increased during
the boom decade 2001–2010.
20.1 INTRODUCTION, THEORY OF FACTOR SHARES, AND SUMMARY OF FINDINGS 385
Table 20.1.1.2 Nominal income levels and shares for labor, profit, rent, and
interest 1930–2010
In general, the percentage increases in the level of income for profits was
greater in boom periods. The percentage increases in the level of income
for labor was greater normal times and rebounding from the depression.
It is much harder to find a pattern for their factor shares. This suggests the
factors driving changes in levels from decade to decade are, at least in part,
different than the factors driving changes in shares.
Rental and interest shares of national income have also varied consid-
erably over the decades, but the variation is not so consistently related to
just the general condition of the economy.
MPK2 MPL2
MPK1 MPL1
K1 K2 L1 L2
MPK2 MPL2
MPL3
Real cost Real
of K = r wage = w
MPK1 MPL1
K1 K2 L1 L3 L2
Graph 3A Graph 3B
Real
wage = w2
MPK2 MPL2
Real cost Real
of K = r wage = w1
20 DETERMINANTS OF FACTOR SHARES
MPK1 MPL1
K1 K2 L1 L3 L2
That said, for the U.S. at least, a long-term change may be occur-
ring which lowers labor’s share of national income, though not its level
of national income. Fifty years ago the U.S. made little foreign invest-
ment; little of the profit component of national income was derived from
foreign operations. In a simple world of two factors, labor and capital, if
investment is only in domestic industries, generally any investment that
increases the level of returns to capital also increases the level of returns to
labor. This is because it takes both capital and labor to make a new factory
produce anything. The increases in both are included in national income.
Even investment that initially only raises capital income eventually raises
labor income (Graphs 1A and 1B).
With foreign investment, the new factory is built (or bought) in a for-
eign land; the profits flow back home and are included in the domestic
country’s national income, increasing capital’s share. Labor income is
counted in the foreign country’s national income, since that’s where it
is generated. Hence, other things equal, increased foreign investment over
time by U.S. firms should lead to a larger share of income for U.S. capital,
and a lesser share for U.S. labor.
For a long time this may reduce labor’s share in the U.S (and other
countries with large foreign investment programs). Our calculations in
Section 3.1 indicate about half of all the decline in Labor’s share since
1980 has been due to growth in profits due to foreign investment, without
the simultaneous growth in labor income that would have occurred if the
investment had been domestic. The decline in labor’s share was not caused
by a decline in labor income in the U.S. (it has grown). It is due in large
part to the much faster growth in profits, due to the rapid growth of profits
on foreign operations.
The opposite is true in the country receiving U.S. foreign investment.
Its national income will only be increased by labor’s income from the new
factory, not by the income earned by the factory’s U.S. owners. Hence,
ceteris paribus, we should see increases in labor’s share of national income
in countries whose growth is financed by foreign investment.
Even so, in the long, long run, the countries now receiving U.S. for-
eign investment will become more and more developed, with large pools
of their own investment capital, and turn, use part of it to make foreign
investments. When the U.S. receives some of this investment, e.g., when
Korea builds a factory in the U.S., labor’s share in U.S. national income
should grow, without growth in the profits portion. This may ultimately
restore the old Cobb-Douglas equilibrium of factor shares.
392 20 DETERMINANTS OF FACTOR SHARES
A Note on Outsourcing
With foreign investment, the discussion above contemplates a new factory
being built (or bought) in a foreign country. This decreases labor’s share,
but not its level of national income. No domestic workers lose their jobs.
At the macro level, this is what seems to have been going on in the U.S.
since 1980. Outsourcing is something different. If a firm closes a domest-
ically located factory and builds or buys a comparable factory in a foreign
country, then labor’s share and level of income will decline since less labor
will be needed domestically.
old levels if the increase demand for GDP persists into the long run,
warranting capital expansion.
The time it takes to adjust to technological innovations can
cause similar cyclical swings. Implementation first boosts one factor’s
share, then increases the other factor’s usage enough to restoring the
old factor share ratio equilibrium.
Permanent, long-term changes in factor shares depend on
changes in the relative elasticity of marginal product curves over
very long periods of time. How long? Even for the 50-year period
studied in this chapter, 1960–2010, we did not have to factor this
in to account for most variation in factor shares. About 85% of
the variation in labor’s share and 93% of the variation in profits
share is shown to occur from cyclical economic factors driven by
changes in the level of aggregate demand, or the growing presence
of profits from foreign operations in profit’s share of national income
(occurring without an offsetting decline in labor income). Labor
productivity was also a factor, positively affecting both labor and
profit’s shares.
Gollin (2002)
Finds wide ranges of labor shares among counties, but mostly due to inac-
curate methods of calculation. When calculated correctly, most counties
labor share falls within a 65% – 80%. The low end of this range is roughly
consistent with this chapter’s finding for the U.S.
Guscina (2006)
Examines causes of declining labor share in 18 OECD industrialized coun-
tries, 1960–2000, using a panel regression. Post-1985 regression results
indicate causes of declining labor share included factor biased technological
progress e.g., information technology development after 1985, and negat-
ive effects of globalization, by reducing labor bargaining power. Regression
results for pre-1985 data indicated positive effects on labor share of (lagged)
productivity growth and employment protection measures and negative
effects of globalization (defined as total trade/GDP). Robustness testing
was extensive: variables were tested in various models and all tests were in
levels and first differences for comparison.
(This study has somewhat similar results for globalization, general pro-
ductivity growth and %unionized. % unionized is similar to Guscina’s
“employment protection” variable.)
Hein (2009)
This is a theoretical, Post-Kaleckian model, not an empirical study. It con-
cludes the most likely outcome of financialization is increasing shareholder
power causing economic contraction. “Financialization” is defined as the
relationship of financial to nonfinancial sectors. The model suggests object-
ives and constraints of firms may be affected by growing financialization:
rising shareholder power may subordinate management’s and workers’ pref-
erence for (long run) growth of the firm, to shareholders’ preference for
(short term) profitability. Greater shareholder power may mean increasing
dividend payments, share buybacks etc., which restricts the availability of
finance for firms’ real investment projects. Distribution of income may be
affected due to changes in power relations between shareholders, managers
and workers, which will then feedback on investment and consumption.
This may lead to a contractive regime, in which higher interest and dividend
payments to rentiers have a restrictive effect on the rates of capacity utiliz-
ation, profit and capital accumulation. These distribution of income results
seem at odds with Piketty [11] who finds growth in maldistribution mainly
caused by growth of labor income at the very top. This model does suggest
slower capital accumulation will occur as a result of financialization. This
396 20 DETERMINANTS OF FACTOR SHARES
is consistent with this study’s finding that labor’s declining share may be
related to declining labor productivity growth.
Imf (2007)
This panel regression of 18 advanced OECD countries using 1982–2002
data, is one of the more widely read studies of labor’s share. Findings indicate
labor’s share significantly related to exports (-), Imports (+), Labor/capital
ratio (+) Offshoring (-), immigration (-) and Information & communica-
tions technology capital (- short term, + long term). (The labor/capital ratio
is similar to this study’s employment/GDP ratio, whose positive finding
may be explained by capital being relatively fixed in the short run, resulting
in most short-term GDP increases occurring from increased employment.
The findings for exports and imports are also consistent with this chapter’s
findings for a related variable, the positive effect of the real exchange rate
on labor’s share: a strong exchange rate will reduce exports and increase
imports.
Stockhammer (2012)
Shows wages in OECD countries, and some developing countries, have
fallen the past 25 years. Statistical tests indicated the decline was attributed
to strong negative effects of globalization, financialization (growth of the
finance industry), and welfare state retrenchment. Technological change is
20.3 METHODOLOGY 397
20.3 METHODOLOGY
OLS or 2SLS, as appropriate, was used to test for variables related to both
the level and share of income received by four factors of production: labor,
profits, interest and rental income. Proprietor’s income was not tested.
398 20 DETERMINANTS OF FACTOR SHARES
Stationarity Issues
All variables described above were tested for nonstationarity using the
ADF test. Nonstationary explanatory variables were then tested for coin-
tegration with the dependent variable. If not found cointegrated, they
were detrended.
Endogeneity Issues
Explanatory variables in the models used were tested for endogenously
with their dependent variable using the Hausman endogeneity test. The
test involves regressing a variable suspected of endogeneity on all the mul-
tifactor model’s exogenous and lagged variables. The residuals from this
regression are then added to labor share model as an additional variable.
If the residual variable’s t-statistic is significant at the 5% level the variable
is replaced with an endogeneity-free, Wald strong instrument in a 2SLS
estimator. See Griffiths et al. (2008, 2011).
The Sargan test was applied to ensure the instrument itself was not
endogenous with the dependent variable.
Heteroskedasticity
Newey-West standard errors were used to address heteroskedasticity
issues.
Robustness Testing
The chapter’s foremost methodological goal is to ensure its findings rep-
resent good science. It is not “just another study” offering a set of
unverified initial findings.
Serious efforts are made to verify the results presented in this chapter are
not idiosyncratic to the time period or particular model tested. Robustness
testing of initial findings is exhaustive. Only one study in the literat-
ure reviewed met this standard (Guscina 2006), who studied labor’s
share.
This robustness testing enhances the credibility of the model consid-
erably, moving it, we believe, from the oft seen (but never intellectually
satisfying) category of “one more study of . . . .” to something more akin
to true, reliable science. Economics long ago developed methodologically
to the point where this kind of robustness testing of findings could be
done with minimal time and effort, and should be a minimal requirement
for publication, if economics is to be considered a twenty-first-century
science. This extensive use of robustness testing is how we meet the
requirement for good science noted by physicist L. M. Krauss [17], i.e.,
of “trying exhaustively, but failing, to disprove” our own initial findings
before submitting them for publication.
Here, more pages are devoted to robustness testing than to developing
the initial findings. Initial results are retested in three additional, though
overlapping, time periods. Only initial results replicated in at least two of
three additional time periods tested are deemed valid, and only then if the
results remain robust when adding or subtracting additional variables to
the model. The “2 of 3” rule was used because some explanatory vari-
ables show little or no movement in some sample periods and can appear
statistically insignificant when in fact they are not. Spuriously high mul-
ticollinearity levels in one period can negatively affect significance levels,
causing the same problem.
Finally, where both OLS and 2SLS are used, we present both results,
though we only rely on the 2SLS results in our analysis. This is done to
allow comparison of results of older studies which used only OLS, a com-
mon problem with older studies. It provides one way of testing whether
differences with findings in older studies were simply a result of them using
400 20 DETERMINANTS OF FACTOR SHARES
lagged variables. The residuals from this regression are then added to labor
share model as an additional variable. If the residual variable’s t-statistic is
significant at the 5% level the variable is replaced with an endogeneity-free,
Wald strong instrument in a 2SLS estimator. See Griffiths et al. [16].
None were found endogenous except the government deficit (T-G),
which was replaced by a Wald strong instrument. The Sargan test was
applied to ensure the instrument itself was not endogenous with the
dependent variable.
Choice of Dependent Variable
The dependent variable was the average of labor’s share of real national
income in the current and prior year (LSAV ). The average was used to
eliminate “white noise” occurring in individual year data on the depend-
ent variable which considerably lowered the amount of variance explained.
The explanatory variables found to most systematically explain year-to-year
variation in labor’s share of national income as shown below (acronyms
given in parentheses).
Equation 20.4.1.1 shows parameter estimate and significance level find-
ings for all the hypothesized determinants of labor’s share except the ratio
of the number of firms employing less than five to the total number of
private firms in the U.S. Data for all these variable was available for 1965–
2010. Model 20.4.1.1 presents the initial results. The final model, robust
to changes in time period sampled and the number of other variables
included in the model, is presented in Model 20.4.1.2.TR further below.
Model 20.4.1.1
OLS Model of Labor’s Share of National Income
LSAV0,–1 = 0.000020 YAV0–1 + 0.03 (Empl / NI) – 1.15 (Wage / NI)
(t =) (2.1) (6.6) (–1.2)
+ 0.001 %Union – 0.64 PROFROW /NI
(1.2) (–1.5)
+0.011 (FinProf / TProf) + 0.72 LProd(av–1–2) /NI
(1.9) (1.9).
+0.14 INFL – 0.0069 UNEMAV–1–2
(1.1) (–4.9)
+0.006 LParRate–3 – 0.00004 (T – G)
(1.8) (–3.1)
R2 = 0.84; DW = 1.5
(20.4.1.1)
402 20 DETERMINANTS OF FACTOR SHARES
Tests indicated the deficit variable was endogenous with the model’s
dependent variable. It was replaced by a strong instrument (Wald test)
which was not itself endogenous with the dependent variable (Sargan
test). Results Of retesting using 2SLS are shown in Model 20.4.1.2
below.
Model 20.4.1.2
2SLS Model of Labor’s Share of National Income
LSAV0,–1 = 0.000019 YAV0–1 + 0.03 (Empl / NI) – 1.10 (Wage / NI)
(t =) (1.7) (5.3) (–1.0)
+ 0.001 %Union – 0.74 PROFROW /NI
(0.5) (–1.4)
+0.008 (FinProf / TProf) + 0.64 LProd(av–1–2) /NI
(1.2) (1.7).
+0.16 INFL – 0.0071 UNEMAV–1–2
(0.9) (–4.3)
+ 0.005 LParRate – 0.000046 (T – G)
(1.6) (–2.6)
R2 = 0.83; DW = 1.5
(20.4.1.2)
20.4 DETERMINANTS OF LABOR, PROFITS, RENT, AND INTEREST FACTOR SHARES. . . 403
Results are very similar for the two models; nearly identical for some
variables. Only two variables significant in the OLS model became
insignificant in the 2SLS model (PROFROW /NI; FinProf/TProf). Aver-
age income, the employment /national income ratio, labor productivity,
the unemployment rate, the labor force participation rate and the govern-
ment deficit were found significant (with the sign on the deficit estimate
indicating deficits have a positive impact on labor’s share).
Since all factor shares must add to unity, a gain in one factor’s share
has to come at the expense of some other factor’s share. Many of the
same variables are in both the profit and labor share models, but with the
opposite sign. This may not occur for economic reasons (of the zero sum
game type), but simply for arithmetic ones. A variable which has, for sound
economic reasons, raised one factor’s share, must (for simple arithmetic
reasons), show a relationship with the opposite sign with another.
Adding the “Firms 1–4” Variable
We next added the variable “% of firms with 1–4 employees as percent of
the total number of firms” to the model and re-estimating it using the
shorter 1978–2010 period for which data was available. Results indicated
the larger the number of firms with 1–4 employees relative to the total
number of firms, the smaller would be labor’s share of national income,
possibly indicating small firms more typically pay less. See Eq. 20.4.1.3.
Model 20.4.1.3
2SLS Model of Labor’s Share of National Income
(Including Firms 1–4 Variable)
LSAV = 0.000028 YAV0–1 + 0.024 (Empl / NI) + 1.46 (Wage / NI)
(t =) (3.2) (4.4) (2.5)
+ 0.001 %Union – 0.61 PROFROW /NI
(1.1) (–1.4)
+ 0.009 (FinProf / TProf) + 4.65 LProd(–1) /NI – 0.15 INFL
(1.4) (1.0) (–0.9)
– 0.006 UNEMAV–1–2 + 0.009 LParRate–3 – 0.000027 (T – G)
(–2.8) (3.6) (–1.8)
– 0.55 Firms1 – 4 R2 = 0.93; DW = 2.2
(–3.8)
(20.4.1.3)
The 2SLS results for the more limited 1978–2010 data set found all the
same variables significant (plus firms 1–4) that were found significant in
404 20 DETERMINANTS OF FACTOR SHARES
the full 1965–2010 initial model, except two: labor productivity, found
insignificant here, and the ratio of average wages to national income, found
significant here.
Robustness of Parameter Estimates Over Time
Physicist Lawrence Krauss (2012) notes that good science requires
researchers to spend as much time trying to disprove a hypothesis as they
spend trying to prove it. Having obtained a significant result in one test,
only if you fail to disprove it in others can you be reasonably sure your
initial results are valid, not spurious. This study uses two methods to try
and disprove its initial results. First, it attempts to verify or disprove results
found statistically significant in one time period sampled by retesting the
same models in three other sample periods. If results cannot be repeated in
other time periods, initial results are judged likely to be spurious and dis-
carded. Second, we also attempt to verify or disprove results for variables
obtained testing one model by retesting them in other plausible mod-
els purporting to explain the same relationship. We do this by adding or
subtracting variables from the original model and retesting. This type of
robustness testing is critical in economic time series models, where signific-
ant multicollinearity is common and can easily change parameter estimates
obtained when even small model changes occur.
Explained Variance and Robustness
Contributions to Explained Variance
Stepwise regression can be used to estimate the contribution of individual
variables to total explained variance. Variables in the initial 2SLS labor
Share Model 20.4.1.2 are tested and their contributions (using both first-
in and first-out methods) are shown in Table 20.4.1.1.
Using the first-out method, the most important contributors to
explained variance were the ratio of employment income to national
income, the unemployment rate and the government deficit. All three were
positively related to labor’s share. Using the first in method, the same three
variables were most important.
Robustness of Initial Results Over Time
All variables were tested in four different, but overlapping, time peri-
ods to determine the replicability of results. Findings are presented in
Table 20.4.1.2.
In Table 20.4.1.2 we attempt to disprove our initial results by testing
the hypothesis that our initial results were spurious. We test this hypothesis
20.4 DETERMINANTS OF LABOR, PROFITS, RENT, AND INTEREST FACTOR SHARES. . . 405
Table 20.4.1.2 Coefficient stability in Eq. 20.4.1.2: 2SLS labor share model
Several of the results from our initial 1965–2010 sample period test
seem to have been based on spurious correlations. Only three variables
significant in the initial sample were also found significant in all three other
time periods sampled. They were 2-year average GDP size, the ratio of
employment to national income, and the unemployment rate. The deficit
variable was significant in three of the four samples. We draw attention to
those found significant in at least three of the four tests because a variable
can be found insignificant in an isolated sample simply because of lack of
adequate variation during the period, or because of multicollinearity, as
well as for substantive reasons. Taking this into consideration, we adopt
three of the four sample periods as our standard of verification of results.
Hence, there are four variables we judge sufficiently robust to changes
in time period sampled to be considered reliable determinants of labor’s
share of national income, likely to again be found significant in any future
tests.by other researchers. The relationship of the other variables tested is
problematic at best.
Reducing the original 11 variable model to the four variables found
significant in at least three of the four samples, and re-estimating using
2SLS, we have:
0.04
0.02
0.00
0.012
0.008 –0.02
0.004
–0.04
0.000
–0.004
–0.008
–0.012
1965 1970 1978 1985 1990 1995 2000 2005 2010
Residual Actual Fitted
Graph. 20.4.1.1 Model of only variables robust in at least three of four sample
periods (Eq. 20.4.1.2.TR)
model containing only the four time period robust variables discussed
above, and one of the variables previously found insignificant in at least
3 of 4 sample period tests and rejected. Three were now found significant
and were added back to the model (labor force participation rate, ratio
of average wage to national income and labor force participation rate).
This new seven variable model was retested again in the other three time
periods. All variables were significant in all four sample periods tested,
except foreign profits which was significant in three. These seven variables
became the “semi-final” model, a model robust to time period sampled,
and is presented below. It is marked time-period robust “final,” but is
really “semi”-final, because we still must test for robustness of parameter
estimates to model changes.
The time-period robust, seven variable semi-final, model of the determ-
inants of labor’s share of national income is given below and shown in
Graph 20.4.1.1, tested using the longest sample period, 1965–2010.
Short term, the labor share models clearly indicate year-to-yearly ups
and downs in labor’s share is overwhelmingly due to more Keynesian
factors, like variation in income, employment, unemployment, labor force
participation rates that cause fluctuation in labor’s share. They have ten-
ded to decline more in downturns than increase in upturns, leading to
some long-term downward net drift in labor’s share (even after account-
ing for the 1/2 of the decline attributable to foreign profit growth).
This net drop we suspect is caused by the decline in productivity growth
(Table 20.4.1.3).
instead of the much smaller decline of only 1.6% that actually occurred. As
we calculated real wages, they were essentially stagnant, suggesting labor
costs relative to other factor costs declined, and that the wage decline kept
the decline in labor’s share much smaller than it would otherwise have
been. Further, if the GDP growth rate had been merely 1/3 higher than
its 2.5% average level during the 1980–2010 period, i.e., 3.33%, labor’s
share would have increased 3.5% instead of declining 1.6%. Hence, the
Study concludes that in addition to foreign profits growth, the other major
source of decline in labor’s share was the decline in GDP growth rates
during this period.
Three other factors would also have made sizable changes: if the labor
force participation rate had not increased, the decline in labor’s share
would have been 3.1%, not 1.6%. Had foreign profits not grown relative to
national income, Labor’s share would have remained virtually unchanged,
dropping only 3/10 of 1%. Had the government deficit not grown, labor’s
share would have dropped 2.2%, not just 1.6%.
The remaining two factors had much smaller, or marginal effects. Had
the growth in the ratio of average wages to national income not occurred,
the loss in labor’s share would have only been 1.3% instead of 1.6%.
The unemployment rate drop, though substantial (3.4%) had virtually
no effect on predicted labor’s share, presumably because the effects of
the social safety net offset much of any change in labor income due to
changing unemployment. The coefficient on the unemployment variable
is (–0.006). Multiplied by the change in unemployment rate, we have
(–0.006)(–0.034) = +0.0002, i.e., a mere 2/100 of 1% change.
Model 20.4.2.1
OLS Model of Determinants of the Level of Labor Income
profits to total profits. This ratio was found positively related to the level
of aggregate income for labor. This was somewhat surprising, since finan-
cialization variables in other factor share studies are more typically found
negatively related.
In robustness tests, however, only the GDP and labor force participa-
tion rates were significant in at least three of the four time periods sampled.
Subsequent modeling of labor income levels as a function of these two
variables alone left only the level of GDP significant. This becomes our
robust model, shown in Eq. 6R. In short, in the levels model, labor income
is driven by Keynesian variables that determine the GDP. The robust
model results with autocorrelation controls are given in Eq. 20.4.2.1.TR.
Our findings with levels in the models above provide some support our
findings with shares. Five of the seven variables found to be significant
determinants of share are the same as (or driven by) changes in GDP and
the labor force participation rate found to be key determinants of the level
of labor income. They include
1. GDP
2. Employment/national income ratio
3. Unemployment
4. Labor force participation rate
5. The deficit (reduced by growing GDP)
in preliminary testing. Its graph is also shown. The final model, robust to
changes in time period sampled and the number of other variables included
in the model, is presented in Model 20.4.3.1.TR further below.
Model 20.4.3.1
OLS Model of Profit’s Share of National Income
where
(Y) = GDP
(Y–Y–1) = the Samuelson accelerator
(Empl/NI) = ratio of employment to real national income
(Wage/NI) = The real wage/real NI ratio
(%Union) = % unionized
(INFL2 ) = the inflation rate (squared):
(LPROD(–1) /NI(–1) )
= ratio of labor productivity growth to real national
income growth, lagged 1 year
(PROFROW/NI) = real U.S. profits derived from operations in the rest
of the world (ROW) as a portion of real national
income
(PROFROW/TProf) = profits on foreign operations as % of total profits
(PRAV–3–4 ) = average of real prime interest rate lagged 3 and 4
years
(CDebt) = consumer debt
(XRAv.0,–1,–2 ) = the real exchange rate average for current and past
two years
418 20 DETERMINANTS OF FACTOR SHARES
0.04
0.02
0.00
0.010
–0.02
0.005
–0.04
0.000
–0.005
–0.010
1965 1970 1975 1980 1985 1990 1995 2000 2005 2010
Residual Actual Fitted
Graph. 20.4.3.1 Graph of the initial profit’s share model (Eq. 20.4.3.1)
20.4 DETERMINANTS OF LABOR, PROFITS, RENT, AND INTEREST FACTOR SHARES. . . 419
Variable
(Y) 0.92 0.04
(ACC) 0.92 0.22
EMPL/NI 0.90 0.12
W/NI 0.92 0.00
%UnionAV 0.92 0.00
LProd(–1 /NI–1 0.92 0.11
Prof ROW /NI 0.80 0.02
ProfROW /TProf 0.70 0.50
PRAV–3–4 0.91 0.01
INFL 0.93 0.00
Cons.Borrowing 0.91 0.18
XRAV0,–1,–2,–2 0.91 0.07
420 20 DETERMINANTS OF FACTOR SHARES
Since factor shares must sum to unity, the growth of foreign profits in
the denominator of the labor share definition, (= Labor income/national
income) without affecting the numerator does explain half the total decline
in labor’s share over the 1960–2010 period, as we noted earlier.
Re-estimating the model using only the explanatory variables found
significant in at least three of the four sample periods, we have our time
period robust model, robust to sample period tested, but not yet tested
for robustness to changes in the model itself:
Model 20.4.3.1.TR
Time Period Robust OLS Model of Profit’s Share of
National Income
The test below compares average changes during the 1970s (the decade
before the labor decline began) with average changes in the most recent
decade in our model, 2000–2009, using the robust five-variable profit’s
share model (Eq. 20.4.3.1.TR, repeated below).
Actual profit’s share of NI (av. for 2000–2009 minus av. for 1980–1999 = 0.0347 (3.5%)
Predicted profit’s share (from Model 3.2.1.1A) = 0.0348 (calculated using actual for
all explanatory variables)
Explanatory variables Reg. coef. (Actual data Pred. profit share after simulation
used in simul. (using actual data except “0” for
2010– 1980) simulated variable)
Model 20.4.4.1
OLS Model of the Determinants of Level of Profit Income
PLevel = + 0.008(Y) – 0.08(Y – Y–1 ) – 0.35.60(EmpI / NI)
(t =) (1.7) (–1.2) (–1.0)
+ 674.58(Wage / NI) – 7.72%Union
(0.2) (–0.5)
+ 56430.86(LPROD(–1) /NI–1 ) + 41221.13PROFROW /NI
(1.1) (3.7)
– 4273.90(FProf / TProf) – 7.44PRav–3–4
(–7.5) (–0.9).
+ 232.06INFL + 0.18(CDebt)
(0.3) (2.9)
– 4.28(XRAV–0,–1,–2,–2 ) R2 = 0.88; DW = 1.8
(–1.4)
(20.4.4.1)
426 20 DETERMINANTS OF FACTOR SHARES
This “levels” model was also tested in three other sample periods to
determine the robustness of findings to time period sampled. Only three
variables were found significant in at least three of the four samples of
profit levels tested: GDP, foreign profits/national income ratio, and for-
eign profits/total profits ratio. These three variables explain 83 of the 88%
of variance explained by the larger levels model above. The robust model
is shown in Eq. 8R.
Model 20.4.4.1.TR
Time Period Robust OLS Model of the Determinants
of Level of Profit Income
PLevel = + 0.10(Y) + 47766.69PROFROW /NI – 5077.98(FProf/TProf)
(t =) (3.6) (7.9) (–13.2)
R2 = 0.83; DW = 1.9
(20.4.4.1.TR)
Previously, we found the same two foreign profits variables are the key
variables determining profit’s share. The levels model leads us to conclude
that the main variables explaining growth in profits since 1965, both in
levels and as a share of national income, is the increase in GDP and the
contribution made by foreign profits to total U.S profits. As was the case
with labor income, the principal driver of profit income in the long run is
GDP growth. This is not totally surprising, since in general income can’t
grow unless output is growing. However, with profits, growth in U.S.-
owned foreign production has led to U.S. profit growth, too. Notice the
exchange rate variable has the opposite sign it had in the labor income
levels model. Here, its negative sign means a strong U.S. dollar reduces
exports and increases imports, both of which hurt company profits.
The growth in profit income levels was not found matched by any
decline in labor’s total income. Hence profit’s gain did not occur because
of a “beggar thy neighbor” effect in which profit’s gain came at the
expense of labor’s loss. However, one could argue that labor’s share might
not have declined at all relative to profit’s if the foreign investment that
led to profit growth had been invested domestically, creating labor jobs in
the states rather than in foreign countries. But this would probably be a
“beggar thy neighbor” approach, since presumably, the reason U.S. busi-
nesses invested abroad was because profits were expected to be better than
what was available from investing the same money domestically.
As the test below shows, the growth in profits on foreign (ROW)
operations as a percent of national income is associated with declining
20.4 DETERMINANTS OF LABOR, PROFITS, RENT, AND INTEREST FACTOR SHARES. . . 427
Model 20.4.4.2
Relationship of Unemployment Rate to Foreign Profits/NI Ratio
Model 20.4.4.3
Okun Unemployment Model After Adding Profits Variable
where
Thus, the revised Okun model confirms our earlier finding that the growth
in foreign-earned profits as a fraction of national income did not come at
the expense of employment in the U.S. in fact appears to have stimulated
428 20 DETERMINANTS OF FACTOR SHARES
it. This may mean that some foreign profits are reinvested in the U.S.,
representing investment that would otherwise have not occurred in the
U.S., increasing employment opportunities.
Summary of Findings – Profit Models
See Table 20.4.4.1.
Table 20.4.4.1 Summary of factors affecting profit’s % share and level of real
national income
Model 20.4.5.1
OLS Estimates of Determinants of Rent’s Share of National Income
RS = – 0.0000094(HPrice /NI) + 26.82(Prop.Inc/NI) + 0.025Mort.Int–1
(t =) (–117.8) (2.6) (2.3)
–19.41(Av.Wage–1 /GDP–1 ) R2 = 0.77; DW = 2.0
(–2.1)
(20.4.5.1)
where here or in the levels equation further below:
Hence, we conclude that our final model, found robust to both changes
in time period sampled and model changes, is the one explanatory vari-
able model (Housing Price /GDP ratio) given in Eq. 20.4.5.1TR, which
explains 70% of the variance in rental income over the 1961–2010 period
tested. With the variables we examined, we were not able to find other sig-
nificant determinants, though with 30% of the variance in rent’s share of
national income still unaccounted for, suggests other factors are at work.
We expect research work in the future will uncover some of them, but
because of the robustness of our own tests, do not expect these future
studies to overturn the results found here, which indicate absolutely noth-
ing is more important in determining rent’s share of national income than
how strong demand for home ownership is, i.e., how high housing prices
are relative to GDP.
Model 20.4.6.1
OLS Model of Determinants of the Level of Rental Income
RILevel = + 0.00027(HPrice /NI) + 2.60Mort.Int–1
(t =) (6.6) (2.9)
– 3769.09(Av.Wage–1 /GDP–1 ) – 0.108Res.Inv.–1,–2
(–1.4) (–4.0)
– 488.51(LaborInc / NI) + 0.72AR(1)
(–2.0) (6.9)
R2 = 0.78; DW = 1.9
(20.4.6.1)
Here, higher house prices are positively related to the level of rental
income, seem to indicate a shift away from home ownership to renting
(earlier, when analyzing shares, it symbolized a reduction in rent’s share
(not necessarily its level) due to growth in labor and profit income when
housing booms occur, but not necessarily a reduction in rental income
levels). Mortgage interest rates effect on rental income is again positive.
Rising wages relative to GDP again is found negatively related, as is growth
20.4 DETERMINANTS OF LABOR, PROFITS, RENT, AND INTEREST FACTOR SHARES. . . 433
Table 20.4.6.1 Summary of factors affecting profit’s % share and level of real
national income
Positive Positive
HPrice /GDP growth HPrice /GDP growth
Real mortgage interest rate growth(–1)
Level of residential investmentAV–1,–2
Growth in labor’s share of NI
∗ Findings not robust for samples without 2001–2010 data
Stationarity Issues
All variables listed above were tested for nonstationarity using the ADF
test. Only one variable was found nonstationary: the stock market index
variable. This variable was tested for cointegration with the interest share
of national income variable, which is the dependent variable in the model
in which the stock market index is used as an explanatory variable. It was
found cointegrated using the Dickey-Fuller test, so no detrending was
needed.
Endogeneity Issues
Explanatory variables in the models above were tested for endogen-
eity with the dependent variable using the Hausman endogeneity test,
described in detail earlier. Results indicated no endogeneity except
between consumer and business debt, and the interest share variable. A
nonendogenous, Wald strong instrument was developed to replace it.
In preliminary tests, all the variables mentioned at the beginning of this
section were tested, but only the variables in the model shown below were
found significant. The others were dropped from the model during this
preliminary testing process. The final, time period and model specification
robust model is given in Model 20.4.7.2 further below.
Model 20.4.7.1
OLS Model of Interest Income’s Share of National Income
IntS = – 0.42Debt0.3
C& I + 27.20PRAV–1,–2,–3 /NI + 0.021Empl / NI
(t =) (–4.1) (9.0) (14.2)
– 0.0005(T – G) + 7.34Baa / NI + –0.27AR(1)
(–13.2) (2.5) (–1.8)
2
R = 0.95; DW = 2.0
(20.4.7.1)
20.4 DETERMINANTS OF LABOR, PROFITS, RENT, AND INTEREST FACTOR SHARES. . . 435
Model 20.4.7.2
2SLS Model of Interest Income’s Share of National Income
IntS = – 0.39Debt0.3
C& I + 27.17PRAV–1,–2,–3 /NI + 0.021Empl / NI
(t =) (–3.9) (8.6) (14.2)
– 0.0005(T – G) + 7.24Baa / NI + –0.28AR(1)
(–12.7) (2.3) (–1.6)
R2 = 0.95; DW = 2.1
(20.4.7.2)
Explained Variance and Robustness
Contributions to Explained Variance
Stepwise regression was used to estimate the contribution to total
explained variance attributable to any one variable. Findings are presented
in Table 20.4.7.1.
With both methods, the variable that explains the most variance is the
ratio of employment to national income; the ratio of the average prime
rate to national income was the second most important.
Robustness Over Time
Table 20.4.7.2 shows tests of interest share model 12 to determine the
stability of coefficients from sample to sample.
All the variables except the government deficit remained significant in
all samples, hence, for them, results are robust to time period sampled,
and our confidence in our model of the determinants of interest’s share
of the NI enhanced. For the government deficit, its effect on interest’s
share here is dominated by the large deficits in the 2001–2010 decade,
which significantly increased interest’s earnings, and positively affected
interest’s share of national income. For the two periods 1968–2000 and
Model 20.4.7.2.TR
Determinants of Interest’s Share of National Income
(Final Time Period and Specification Robust Model)
Model 20.4.8.1
OLS Model of the Determinants of the Level of Interest Income
IntLevel = + 41.14Debt0.3
C& I + 27.63PRAV–1,–2,–3
(t =) (1.8) (5.1)
– 21.16UNEM + 0.034(T – G) + 3.40Baa
(–3.4) (4.1) (1.3) (20.4.8.1)
+5112.72INFLAV0,–1,–2 + 0.30DJAVAV0,–1
(4.0) (2.2)
R2 = 0.73; DW = 2.0
438 20 DETERMINANTS OF FACTOR SHARES
The variables have the same meaning as in the shares model except the
two interest rate variables stand alone; they no longer represent interest
rate/national income ratios. In addition, “UNEM” stands for the unem-
ployment rate and “INFL” the inflation rate. Together, these only explain
73% of the variation in the level of interest income over the 50 years
sampled, 1965–2010. Additional determinants most likely exist, but were
not found.
In testing for levels of interest income, in addition to the variables
shown above, all the other variables described at the beginning of this
section were tested, as well as the exchange rate and the trade deficit,
which were also found insignificant. Also tested were the mortgage and
Aaa bond rates. These were found insignificant when the Baa rate was also
included in the model, but were at least marginally significant if not. This is
essentially the behavior we would expect if they were serving as imperfect
proxies for the Baa rate. Since market interest rates tend to move together
(rrealBaa,Aaa = 0.93), we interpret our Baa finding as capturing the effect
of all such highly correlated market-driven rates on the level of interest
income. The prime interest rate by comparison, also significant, is not a
market-determined rate. It is rigidly linked to the level of the federal funds
rate, which is an administered one established by the Federal Reserve. To
a large extent, it moves independently of market rates. For this reason we
find both the prime rate and the Baa rate significant in the model.
The debt, inflation, and Baa interest rate variables were tested for endo-
geneity with the level of interest income. Only the Baa interest rate variable
was found endogenous and was replaced with a Wald-strong instrument
containing the same variables used as an instrument in the interest’s
share model, but with two lagged values of the Baa variable added to
strengthen it. The model was then re-estimated, yielding the following
results:
Model 20.4.8.2
2SLS Model of the Determinants of the Level of Interest Income
IntLevel = + 44.91Debt0.3
C& I + 30.40PRAV–1,–2,–3 – 21.30UNEM
(t =) (1.9) (5.2) (–3.1)
+ 0.032(T – G) + 5.19Baa + 5669.50INFLAV0,–1,–2
(3.7) (1.6) (4.1)
+ 0.28DJAVAV0,–1 R2 = 0.74; DW = 2.1
(1.9)
(20.4.8.2)
20.4 DETERMINANTS OF LABOR, PROFITS, RENT, AND INTEREST FACTOR SHARES. . . 439
Four of the variables were found significant in all four sample peri-
ods: consumer debt (+), the prime interest rate/national income ratio
(+), inflation (+) and unemployment (–). None of the other three vari-
ables in the initial test was found significant in more than two of the four
sample periods, nor were any of them found significant when individually
retested with the four-variable core robust model. Though two were found
individually significant, when both were added to the core model and
retested, none passed the “3 of 4” test. Hence, the results for these three
variables were not considered time period robust. The four-variable core
model becomes the time period robust model, shown as Eq. 20.4.8.2.TR
below.
IntLevel = + 54.48Debt0.3
C& I + 30.97PRAV–1,–2,–3 – 29.33UNEM
(t =) (2.1) (5.7) (–4.4)
+ 5518.62INFLAV0,–1,–2 R2 = 0.74; DW = 2.1
(3.4)
(20.4.8.2.TR)
Table 20.4.9.1 Summary of factors affecting interest’s % share and level of real
national income
Positive Positive
Real Prime Int. RateAV–1–2–3 /NI Prime Int. RateAV–0–5
Real Baa Int. Rate/NI Unemployment Rate
Employment/NI Ratio Consumer and Business Debt
Inflation
20.5 SUMMARY AND CONCLUSIONS (INCOME SIDE OF THE NIPA ACCOUNTS) 441
four time periods sampled, but not as a substitution effect, i.e., not with a
positive sign. The house price/NI ratio consistently had a positive sign,
suggesting the finding is indicating shifting housing preferences from
rental to home ownership, which would be expected to push up house
prices, is the more fundamental factor causing the decline in rent’s share
of national income.
Barro, R.J. (1987). “Government Spending, Interest Rates, Prices and Budget
Deficits in the United Kingdom, 1701–1918”. Journal of Monetary Economics
20, 221–47. Jan. 1987.
Baumol, W. (1952). “The Transactions Demand for Cash: An Inventory Theoretic
Approach”. Quarterly Journal of Economics 66(4), 546–56.
Bentolila, Samuel, and Saint-Paul, Gilles. (2003). “Explaining Movements in the
Labor Share”. Contribution to Macroeconomics 3(1). Berkeley, CA Berkeley
Press.
Brayton, F., Levin, A., Tryon, R., and Williams, J.C. (1997). “The Evolution of
Macro Models at the Federal Reserve Board”. Originally Prepared for Carnegie
Rochester Conference on Public Policy 47(1), 43–81. Nov. 22–23 1996, 1997.
Revised Feb. 7, 1997.
Brayton, F., Laubach, T., and Reifschneider, D. (2014). “The FRB/US Model:
A Tool for Macroeconomic Policy Analysis”. FEDS Notes. Washington, DC:
Board of Governors of the Federal Reserve System. Apr. 2014.
Brayton, F, and Tinsley, P., ed. (1996). “A Guide to the FRB/US A Macroeco-
nomic Model of the United States”. Division of Research and Statistics, Federal
Reserve Board. Version 1.0. Oct. 1996.
Chari, V. V., Kehoe, Patrick J., and McGrattan, Ellen R. (2009). “New Keyne-
sian Models: Not Yet Useful for Policy Analysis”. American Economic Journal:
Macroeconomics 1(1), 242–66.
Christoffel, K., Coenen, G., and Warne, A. (2010). “Forecasting With
DSGE Models”. European Central Bank Working Paper Series. 1185. May
2010.
Chung, Hess. (2015). “The Effects of Forward Guidance in Three Macro Mod-
els”. FEDS Notes. Washington, DC: Board of Governors of the Federal Reserve
System. Feb. 26.
© The Author(s) 2017 445
J. J. Heim, An Econometric Model of the US Economy,
DOI 10.1007/978-3-319-50681-4
446 BIBLIOGRAPHY
Colander, D., Howitt, P, Kirman, A., Leijonhufved, A., and Mehrling, P. (2008).
“Beyond DSGE Models: Toward an Empirically Based Macroeconomics”.
American Economic Review: Papers and Proceedings 98(2), 236–40.
Colander, D. (2010). “Testimony Presented to U.S. House of Representatives
Science and Technology Committee Hearing Entitled ” Building a Science of
Economics for the Real World, July 20th, 2010.
Cooley, T.F. (1997). “Calibrated Models” August 1997 Lecture at UC Davis,
pp.1–29. Also essentially the same as “The Nature and Use of Calibrated
Models” 7th World Congress of the Econometric Society, Tokyo. Aug. 1995.
Dawkins, C., Srinivasan, T, and Whalley, J. (2001). calibration article authors in
Heckman, J., and Leamer E. eds. Cptr. 58 of Handbook of Econometrics, Vol. 5.
Calibration: North Holland Press.
Del Negro, M, Schorfheide, Smets, F, and Wouters, R. (2007). “On the Fit of
New-Keynesian Models”. Journal of Business & Economic Statistics 25, 123–49.
Del Negro, M., Eusepi, S, Giannoni, M., Sbordone, A., Tambalotti, A., Cocci, M.,
Hasegawa, R., and Linder, M.H. The FRBNY DSGE Model. Federal Reserve
Bank of New York Staff Reports, 647. Oct. 2013.
deLeeuw, F., and Gramlich, E. (1968). “The Federal Reserve – MIT Economic
Model”. Federal Reserve Bulletin 54(1), 11–40. Jan. 1968.
Eckstein, O. (1983). The DRI Model of the U.S. Economy. New York: McGraw-Hill
Book Company.
Eckstein, O. (2014). FRB/US Equation Documentation. Federal Reserve Board.
Available at www.federalreserve.org/econresdata/frbus/us-models-about.htm/frbus/
us-models-about.htm.
Eckstein, O. (2015). “About (the FRB/US Model)”. Federal Reserve Board.
Available at http://www.federalreserve.gov/econresdata/frbus/us-models-about/.
Eckstein, O. (1963, 2002, 2010–2014). Economic Report of the President. Wash-
ington, DC: Government Printing Office.
Eckstein, O. (2011). FRB/US Equation Documentation. Federal Reserve Board.
Flow of Funds Accounts of the United States, Data Series TODNS, TBSDODNS,
CMDEBT, FGSDODNS, SLGSDODNS, AND DODFS. Washington, DC.
Board of Governors of the Federal Reserve System. March 11, 2011 Update.
Available at www.federalreserve.gov/releases/z1/current/.
Edge, R., Kiley, M., and Laforte, P. (2010). “A Comparison of Forecast Per-
formance of between Federal Reserve Staff Forecasts, Simple Reduced Form
Models, and a DSGE Model”. Finance and Economics Discussion Series. Divi-
sion of Research and Statistics and Monetary Affairs, Federal Reserve Board,
Washington, DC.
Edge, R., and Gurkaynak, R. (2011). “How Useful are Estimated DSGE Model
Forecasts for Central Bankers? Brookings Papers on Economic Activity, Fall
2010. Table 2, p. 224 Also available as “How Useful are Estimated DSGE
BIBLIOGRAPHY 447
Mountford, A., and Uhlig, H. (2008). “What Are The Effects Of Fiscal Policy
Shocks?” NBER Working Paper # 14551, Dec. 2008.
Okun, A. (1962). “Potential GNP: Its Measurement and Significance”. Pro-
ceedings of the Business and Economics Statistics Section, American Statistical
Association. Washington, DC: American Statistical Association.
Paccagnini, A. (2011). “DSGE Model Evaluation and Hybrid Models: A Com-
parison”. European University Institute, Florence. Max Weber Program. EUI
Working Paper MWP 2011/11.
Piketty, Thomas. (2014). Capital in the Twenty – First Century. Cambridge, MA:
Belknap Press of the Harvard University Press.
Piketty, Thomas. (2015). “About Capital in the Twenty – first Century”. American
Economic Review: Papers and Proceedings 105(5), 48–53.
Pindyck, R., and Rubinfeld, D. (1991). Econometric Models and Economic Forecasts.
Boston: Irwin McGraw-Hill.
Reiss, A., and Livio, M. (2016). “The Puzzle of Dark Energy”. Scientific American
314(3), 38–43. March 2016.
Romer, D. (1996). Macroeconomics. New York: McGraw Hill.
Romer, P. (2016). “The Trouble With Macroeconomics” Delivered 1/5/16
as the Commons Memorial lecture of the Omicron Delta Epsilon Society.
Forthcoming in The American Economist.
Sbordone, A., Tambalotti, A., Rao, K., and Walsh, K. (2010). “Policy Ana-
lysis Using DSGE Models: An Introduction”. FRBNY Economic Policy Review
16(2), 23–44. Oct. 2010.
Sims, C. (1972). “Money, Income, and Causality”. The American Economic Review
62(4), 540–52. Sept. 1972.
Sims, C. (1980). “Macroeconomics and Reality”. Econometrica 48(1), 1–48. Jan.
1980.
Sims, C. (2006). “Comment on Dl Negro, Schorfheide, Smets and Wouters”.
Journal of Business and Economic Statistics 25 (2) 2, 152–154. Available at
http://sims.prinston.edu/yftp/Ottawa/SWcommentSlides.pdf.
Smets, F., and Wouters, R. (2003). “An Estimated Dynamic Stochastic Gen-
eral Equilibrium Model of the Euro Area”. Journal of the European Economic
Association 1(5), 1123–75. Sept. 2003.
Smets, F, and Wouters, R. (2007). “Shocks and Frictions in U.S. Business Cycles:
A Baysian DSGE Approach”. American Economic Review 97, 568–606.
Smith, N. (2014). “The Most Damning Critique of DSGE”. (Noah-
pinion, 1/10/14) Available at http://noahpinionblog.blogspot.com/2014/01/
the-most-damning-critique-of-dsge.html.
Solow, R. M. (2010). Testimony Presented to U.S. House of Representatives
Science and Technology Committee Hearing Entitled “Building a Science of
Economics for the Real World”, July 20th, 2010.
Solow, R.M. (2016). Letter to J. Heim dated 6/29/2016.
450 BIBLIOGRAPHY
Dynamic Stochastic General Equilib- Exchange rate, 13, 41, 57, 62, 65,
rium (DSGE), 3–6, 8, 9, 11, 12, 67, 86, 123, 156, 158, 160, 164,
13, 39, 40, 48–56, 58, 59, 61, 63, 168, 172, 174, 191, 192, 194,
66, 68, 69, 70–73, 77, 78, 81–88, 198, 204, 205, 208, 211, 213,
103, 108, 109–111, 113, 114, 214, 215, 216, 217, 218, 221,
115, 131, 140, 142, 145, 244, 224, 228, 244, 251, 286, 289,
372 296, 300, 345, 355, 372, 392,
409, 410, 415, 416, 418, 419,
E 420, 422, 434, 438
Eckstein, Otto, 1, 4, 7, 9, 13, 39, 41, Exogenous, 42, 68, 123, 148,
49, 55, 67, 87, 90–92, 110, 111, 149, 186, 187, 229, 253, 255,
115, 125–131, 145, 155 303, 309, 318–320, 371, 394,
Econometric model, 1, 7, 8, 9–14, 39, 396
49, 57, 108, 114, 115, 144, 148, Explained variance, 43, 45, 60, 63, 65,
328, 377 69, 98, 111, 141, 150, 153, 154,
Economic philosophy, 5–9 155, 158, 159, 163, 164, 168,
Economic Report of the President, 40, 169, 172, 173, 174, 178, 181,
47, 61, 77, 285 189, 194, 199, 204, 205, 208,
Economic science, 5–9, 124 212, 213, 215, 217, 226, 256,
Edge, R., 53, 111 261, 267, 277, 278, 280, 288,
Endogeneity, 13, 39, 41, 42, 45, 292, 295, 298, 299, 304, 310,
47, 57, 78, 96, 116, 126, 314, 343, 349, 370, 400, 415,
147, 148, 149, 150, 172, 181, 425, 426, 431, 435
185, 186, 187, 215, 224, 255, Explanatory model, 139, 378
280, 286, 309, 342, 349, 394, Explanatory variable, 7, 8, 11, 14,
396, 397, 410, 412, 424, 45, 46, 49, 52, 57, 63, 90,
430, 434 98, 101, 104, 105, 106, 107,
Endogenous, 10, 41, 42, 57, 108, 116, 113, 116, 131, 140, 148, 150,
147–149, 151, 172, 177, 181, 158, 159, 164, 175, 179, 182,
185, 187–189, 192, 193, 198, 184, 192, 195, 197, 201, 208,
204, 208, 254–255, 276, 286, 212, 216, 232, 258, 261, 267,
292, 303, 310, 322–330, 342, 291, 308, 310, 318, 338, 339,
349, 371, 394, 396, 397, 398, 341, 343, 346, 347, 348, 349,
412, 434 353, 394, 395, 396, 397, 402,
Engineering, 4, 7, 11, 24, 50, 81, 88 404, 405, 407, 409, 412, 417,
Engineering Manual, 4, 11, 24, 418, 419, 424, 428, 430, 433,
50, 81 439
Engineering manual, 4, 11, 24, Export demand, 117, 221, 222,
50, 81 225–228, 233, 372
Equation of Exchange, 2, 332, 333, Exports, 1, 10, 44–45, 82, 93–95,
373 109, 110, 115, 117, 123, 131,
Euler condition, 61 138, 139, 144, 149, 158, 159,
454 INDEX
416, 419, 422, 425–428, 433, Instruments, 8, 13, 41, 42, 116, 147,
435, 439 149, 167, 172, 187, 193, 198,
Gurnayak, R., 53, 111 202, 342
Guscina, A, 391, 395 Interest income, 16, 377, 378, 379,
430, 431, 433, 434, 435, 439,
H 440
Hausman, 42, 47, 57, 148, 149, 161, Interest income, level, 433–435, 440
172, 181, 186–187, 189, 193,
Interest income, share, 430, 431, 439
198, 215, 216, 255, 260, 267,
276, 286, 342, 349, 394, 396, interest rate, 5, 6, 11, 15, 52, 62,
412, 424, 430 63, 65–67, 69, 70, 86, 91,
Heim, J., 47, 56, 57, 61, 87, 112, 141, 96–98, 100, 104, 109, 113–115,
147, 253, 357 117–120, 131–134, 142–145,
Heteroskedasticity, 39, 44, 395 158, 161, 165, 166, 172, 173,
Hill, R., 91, 107 176, 182, 183, 191, 194, 198,
Housing, 41, 82, 113, 117, 120, 122, 199, 211, 216, 217, 221, 222,
128, 129, 134, 135, 138, 143, 224, 227, 232–235, 253–263,
174, 175, 176, 177, 212, 369, 269–271, 297–300, 357, 365,
370, 371, 373, 379, 425, 426, 369–372, 379, 412, 416, 418,
427, 428, 429, 439, 440 425, 427, 428, 433, 434, 436,
439, 440
I Inventory investment, 13, 121, 136,
Identification, 4, 5, 7, 8, 13, 41, 51, 188, 215, 216, 218, 371
96, 147, 185 Investment, 1, 9, 10, 13, 40, 44–
Identities, 2, 9, 14, 126, 378 45, 52, 67, 72–76, 82, 93,
ILO, 392 94, 104–107, 109, 110–117,
IMF, 392 119–122, 126, 129–132, 134–
Imported Consumer goods, 140, 150, 140, 143–145, 148, 149, 172,
151–157, 161 174, 176, 177, 178, 185–219,
Imported investment goods, 188, 197, 229–245, 247–251, 253, 285,
198, 200, 201 291–295, 317, 320, 322, 333,
Imports, 14, 82, 83, 93, 109, 117, 346, 352, 353, 355, 357–362,
123, 124, 137, 138, 140, 144, 369–371, 379, 387, 388, 391,
149, 150, 157, 158, 159, 160, 392, 422–424, 427, 429
161, 163, 197, 198, 199, 200, IS curve, 1, 44, 94, 138, 139, 229,
201, 202, 222, 223, 224, 226, 231, 236, 239, 240, 241, 244,
228, 296, 371, 392, 410, 419, 246, 247, 248, 250, 252, 361
422 IS curve, 1, 94, 138, 139, 229–252
Income distribution, 377, 389, 393
Instruments, 8, 13, 41, 42, 116, 147,
149, 167, 172, 187, 193, 198, J
202, 342 Jaumotte, F., 392
456 INDEX
Modigliani, F., 50, 59, 87, 140 291, 292, 294, 295, 297, 301,
Monetary policy, 2, 67, 112, 113, 115, 302, 304, 307, 310, 312, 313,
332, 333, 335, 367, 372 315, 316, 336, 338, 339, 340,
Monetary policy, 2, 67, 112–113, 115, 341, 342, 346, 347, 348, 350,
332, 333, 335, 372 351, 359, 369, 393, 395, 396,
Mortgage interest rate, 13, 133, 134, 397, 399, 409, 410, 412, 413,
176, 369, 425, 427, 428, 440 417, 421, 422, 424, 425, 428,
Mountford, A., 91, 92, 93, 107 429, 430, 433
OPEC, 125, 267, 275, 277, 280
N Orszag, P., 91
National income, 6, 14, 16, 47,
377–382, 384, 387–390, 393, P
396–400, 402, 403, 405–413, Paccagnini, A., 54
415–420, 422–432, 434, Participation Rate, 396, 399, 403,
436–440 407, 409, 410, 411, 438, 439
National Income and Product % Unionized, 391, 416
Accounts, 14 Phillips Curve, 1, 265–271, 274, 326,
Neoclassical Mechanics, 3 373
Neoclassical models, 3 Piketty, T., 377, 388, 389, 391, 393,
New Keynesian models, 56 437
NIPA, 377, 437 Population, 58, 62, 65, 67, 132, 133,
Nonstationarity, 7, 131, 162, 185, 134, 152, 154, 158, 159, 161,
197, 336, 361, 394, 396, 412, 162, 164, 168, 181, 182, 189,
424, 430 192, 195, 197, 208, 216, 236,
NYSE Composite Index, 414 243, 244, 249, 250, 251, 288,
296, 297, 299, 309, 310, 311,
O 312, 316, 343, 346, 355, 358,
Okun, 273–280, 310, 423 371, 388, 425
Okun’s Law, 273–276, 310 Population Age Distribution, 149
“Old” Keynesian models, 1, 12, Prime Interest Rate, 62, 67, 98, 100,
56–59, 87, 108, 112 113, 132, 166, 172, 182, 194,
OLS, 4, 8, 11, 30, 41, 43, 45, 47, 57, 199, 211, 217, 221, 234, 253–
78, 108, 111, 132, 147, 150, 151, 263, 269, 300, 357, 370–371,
152, 157, 162, 163, 165, 166, 412, 416, 418, 434, 436, 439,
167, 171, 172, 173, 176, 177, 440
180, 181, 183, 184, 185, 188, Profit income, 378, 379, 380, 408,
189, 192, 193, 196, 197, 202, 410, 421, 422, 424, 428, 439,
203, 207, 209, 212, 214, 215, 440
216, 218, 224, 225, 227, 228, Profit income, level, 378–380, 408,
254, 255, 259, 260, 265, 266, 410, 421–424, 428, 439
267, 268, 269, 273, 274, 275, Profit income, share, 378–380, 408,
277, 279, 280, 286, 287, 288, 410, 421, 422, 424, 428
458 INDEX
Profits, 48, 82, 109, 138, 189, 190, 231, 237, 244, 256, 258,
197, 198, 199, 205, 208, 209, 261, 262, 267, 268, 277, 278,
214, 216, 217, 244, 265, 285, 282, 289, 290, 294, 300, 301,
286, 288, 289, 290, 331, 332, 304, 305, 307, 311–313, 315,
351, 358, 379, 380, 381, 384, 343–345, 349, 351, 378, 393,
387, 390, 392, 394, 396, 403, 403, 404, 411, 415, 417–419,
405, 406, 408, 409, 411, 414, 422, 426–429, 431–433, 435,
415, 416, 417, 418, 419, 420, 436, 440
422, 423, 425, 437, 438, 439 Robustness testing, 8, 11, 46, 146,
391, 393–396, 400, 405
R Rule of thumb, 56, 75, 76, 77, 78, 80
R2 , 16, 44, 48, 63, 65, 154, 317
Rational Expectations, 5, 12, 50, 52, S
56, 63, 68, 70, 72, 73, 75, 78, 80, Saint-Paul, G., 390
81, 83, 84, 87, 114, 125, 131, Sample period, 4, 7, 8, 11, 12, 14, 45,
140 53, 71, 78, 91, 98, 99, 103, 106,
Real Wage, 265, 384, 409 150, 154, 155, 156, 157, 159,
Regression Coefficients, 16, 44, 45, 160, 161, 164, 166, 170, 171,
52, 77, 90, 102, 103, 113, 139, 174, 175, 176, 179, 183, 191,
145, 150, 202, 235, 242, 248, 195, 205, 215, 254, 273, 289,
298, 341, 348, 359–362 290, 291, 298, 300, 304, 311,
Reiss, A, 5 315, 317, 318, 350, 395, 400,
Rental income, 16, 377, 378, 379, 402, 403, 404, 417, 418, 422,
394, 424, 425, 427–429, 430, 429, 433, 435, 436, 440
440 Samuelson accelerator, 1, 218, 412
Rental income, level, 16, 377, 378– Sargan, 42, 187–189, 193, 198, 255,
379, 394, 424, 425, 427, 428, 260, 286, 342, 394, 397
429, 438, 440 Savings, corporate, 10, 167, 285–291,
Rental income, share, 377–379, 424, 296
425, 427, 428, 429, 430, 438 Savings, depreciation, 291, 292, 294,
Residential construction, 13, 74, 143, 295
212, 214, 215 Savings, personal, 167, 168, 285,
Residential investment, 134, 172, 174, 296–302, 370
176–178, 212, 215, 425, 427, Sbordone, A., 68, 70, 71, 81
429, 433 Scientific, 4–7, 9, 13, 46, 51, 83, 89,
RHS, 181, 275 96, 104, 110, 111, 112, 125, 235,
Robustness, 4, 30, 36, 42–44, 145, 297, 317
152, 153, 156, 158, 159, 161, Self Evident Truths, 5, 9, 83, 110, 111
163–165, 168–170, 173, 175, Semi-manufactured goods, 123, 124
178, 179, 181–183, 189–191, Serial correlation, 13, 16, 44, 91, 124,
194, 196, 199, 201, 204–206, 139, 394
208–214, 216–218, 226, 227, Shoc08, 275, 276, 278, 280, 423
INDEX 459
149, 150, 151, 152, 155, 161, 221, 222, 224, 226, 254,
162, 163, 167, 168, 169–174, 256, 260, 261, 267, 273, 275,
177–182, 185, 186, 188–195, 277, 278, 280, 286, 288, 291,
197–202, 204, 206, 207, 208, 292, 294–296, 298, 299, 304,
210, 211, 212, 214, 215, 216, 309–311, 314, 317, 343, 349,
217, 230, 231, 236, 240, 255, 357, 370, 396, 397, 400, 405,
256, 257, 258, 260, 262, 263, 409, 415, 419, 422, 425, 426,
267, 276, 277, 278, 280, 282, 428, 431, 435
286, 287–292, 297, 310, 336, VAR methodology, 115
342, 344, 345, 346, 349, 350, Vector autoregressive model (VAR), 3,
369, 393, 394, 395, 396–403, 8, 9, 12, 13, 39, 40, 45, 48–52,
409, 410, 412, 424, 429, 431, 54, 55, 71–73, 75, 84, 87–94,
433, 434 96–100, 101, 103–107, 109, 111,
Tytell, I. 113–114, 115–116, 126, 130,
139, 141, 235, 244, 372
U Velocity, 2, 15, 332–362
Uhlig, H., 91, 92, 93, 107
Unemployment inflation, 116 W
V Wald, 13, 42, 149, 172, 177, 187,
188, 193, 208, 255, 260, 286,
Variance, 8, 10, 14, 43–46, 52, 53,
342, 394, 397, 412, 430, 434
57–63, 65, 69, 71, 74, 77, 78, 80,
81, 87, 98, 106, 108, 109, 111, Warne, A.
112, 114, 115, 131, 132, 137, Wharton Econometric Model, 116
139–141, 145, 150, 153–155, Wickens, M., 55
158, 159, 168, 172–176, 178, Wilcox, D., 72
181, 189, 194, 199, 203–205, Wouters, R., 53, 55, 72, 82, 83, 86,
208, 209, 212, 213, 215–218, 111