Destruction, Disinvestment, and Death: Economic and Human Losses

Following Environmental Disaster

Jesse K. Anttila-Hughes

Solomon M. Hsiang

Abstract
The immediate physical damages caused by environmental disasters are conspicuous and often the focus of media and government attention. In contrast, the
nature and magnitude of post-disaster losses remain largely unknown because
they are not easily observable. Here we exploit annual variation in the incidence
of typhoons (West-Pacific hurricanes) to identify post-disaster losses within Filipino households. We find that unearned income and excess infant mortality in
the year after typhoon exposure outnumber immediate damages and death tolls
roughly 15-to-1. Typhoons destroy durable assets and depress incomes, leading to broad expenditure reductions achieved in part through disinvestments in
health and human capital. Infant mortality mirrors these economic responses,
and additional findings that only female infants are at risk, that sibling competition elevates risk, and that infants conceived after a typhoon are also at risk
indicate that this excess mortality results from household decisions made while
coping with post-disaster economic conditions. We estimate that these posttyphoon economic deaths constitute 13% of the overall infant mortality rate in
the Philippines. Taken together, these results indicate that economic and human
losses due to environmental disaster may be an order of magnitude larger than
previously thought and that adaptive decision-making may amplify, rather than
dampen, disasters social cost. JEL Codes: J13, O12, Q54, Q56.

We thank Douglas Almond, Chris Barrett, Scott Barrett, Marshall Burke, Mark Cane, Janet Currie, Tatyana
Deryugina, Ram Fishman, Joshua Graff Zivin, Michael Hanemann, Rema Hanna, Kyle Meng, Gordon McCord, John
Mutter, Suresh Naidu, Daiju Narita, Nicole Ngo, Michael Oppenheimer, Cristian Pop-Eleches, Jeffrey Sachs, Bernard
Salani
e, Nicholas Sanders, Wolfram Schlenker, Kerry Smith, Reed Walker and seminar participants at Columbia, Cornell,
Harvard, Princeton, UC Berkeley, UC San Diego, University of San Francisco, the NBER Summer Institute, the NBER
HC/EEE Fall Meeting, the Advanced Graduate Workshop on Poverty, Development and Globalization and the American
Geophysical Union Fall Meeting for helpful comments and suggestions. We thank Vincent Olaivar of the National
Statistics Office of the Philippines for assistance obtaining data. This work is funded in part by a grant from the
Center for International Business Education and Research; J.K.A.H. was funded by an IGERT grant from the National
Science Foundation, S.M.H. was funded by STAR Research Grant (FP-916932) from the US Environmental Protection
Agency, and Postdoctoral Fellowships in Applied Econometrics (NBER) and Science, Technology & Environmental
Policy (Princeton). This version: February 18, 2013.

Corresponding author: 2130 Fulton St., Department of Economics, Cowell 407, San Francisco, CA 94117. Phone:
(415) 422-6453. Fax: (415) 422-6983. Email:jkanttilahughes@usfca.edu

Electronic copy available at: http://ssrn.com/abstract=2220501

Introduction

It is obvious that natural disasters cause immediate destruction and death. In theory,
documenting the direct physical damages caused by hurricanes, earthquakes, and other
catastrophes is straightforward, although the logisitics of doing so are often difficult. At
the same time, even our theoretical understanding of disasters aftereffects, particularly
on economic outcomes, remains limited by a paucity of empirical observations. The
few facts we have about post-disaster economics come primarily from studies that link
macroeconomic data with country-level estimates of disaster impacts (see Stromberg
(2007) and Cavallo and Noy (2009) for reviews of the literature). Thus even fairly basic
questions about disasters economic effects, such as whether household incomes rise or
fall in a disasters wake, remain unsettled (Albala-Bertrand (1993); Benson and Clay
(2004); Caselli and Malhotra (2004); Hallegatte and Ghil (2008); Horchrainer (2009);
Loayza et al. (2009); Dercon and Outes (2009); Noy (2009); Fomby, Ikeda and Loayza
(2009); Hsiang (2010); Strobl (2011); Deryugina (2011)).
Improving our understanding of post-disaster economic outcomes is important for
several reasons. Designing effective disaster management policies and institutions requires that we understand the full cost of disasters (Kunreuther et al. (2009); United
Nations (2009)); if a sizeable portion of a disasters costs manifest after the event itself,
then models of humanitarian intervention which focus on immediate damages may need
to be reassessed or expanded. Secondly, the wealth of evidence suggesting that disasters
immediate death and destruction is most acute in low-income countries (Kahn (2005),
Mutter (2005), Yang (2008), Hsiang and Narita (2012)) indicates that disasters might
plausibly influence economic development. Of particular concern is disasters potential
to alter long-run outcomes due to short-run losses: if poor households have a limited
ability to mitigate disaster-induced losses, disaster incidence may cause them to sacrifice
valuable investment (Udry (1994); Jacoby and Skoufias (1997); Duflo (2000); Maccini
2

Electronic copy available at: http://ssrn.com/abstract=2220501

and Yang (2009); Banerjee and Mullainathan (2010)) for short-run needs. Lastly, recent evidence suggests that global climate change is expected to increase the frequency
of certain types of environmental disaster, (IPCC 2007; Knutson et al. (2010)). This
implies that any improvement in estimates of disasters costs will necessarily inform
estimates of climate changes anticipated damages (Narita, Tol and Anthoff (2009);
Mendelsohn, Emanuel and Chonobayashi (2012)), and in turn the formulation of climate change policy in general (Stern (2006); Nordhaus (2008); Tol (2009); Weitzman
(2009); Pindyck (2011)).
In this paper we measure the post-disaster economic and health effects of a specific
type of environmental disaster: typhoons. Typhoons are tropical cyclones in the West
Pacific region1 large, fast-moving storms which form over the oceans and cause physical damage via intense winds, heavy rainfall, and ocean surges. We focus on typhoons
both because they are one of the most common and costly types of natural disaster
(Bevere, Rogers and Grollimund (2011)) and because their variation in timing and spatial distribution allow us to identify their effects using quasi-experimental techniques
(Holland (1986)). Typhoons are relatively brief, usually affecting a given location for at
most 1- 2 days. They are also sharply defined in space, with a core 100-200 kilometers
across and traveling distances ranging from a few hundred to a few thousand kilometers
in length. The intensity of a locations typhoon exposure is also variable, both because
the storms themselves vary in frequency and intensity and because different locations
are exposed to different parts of the same storm, another feature that we exploit in our
econometric analysis.
The Philippines is situated in one of the most intense typhoon climatologies on the
planet, a fact that both improves our identification strategy and differentiates this study
from analyses of one-off or infrequent natural disasters. To capture spatial and tempo1

Typhoon is the name for a tropical cyclone that occurs in the western Pacific Ocean. The same
storms are called hurricanes in the Atlantic Ocean and simply cyclones in the Indian Ocean.

ral variations in typhoon exposure within the Philippines we use a physical model of
typhoon winds developed in Hsiang (2010) to create a unique panel dataset of provincelevel incidence. This dataset allows us to adopt a difference-in-differences approach
which takes advantage of each provinces year-to-year variation in typhoon exposure.
We combine physical storm data with two household survey files: the Family Income and Expenditure Survey (FIES), a repeated cross sectional survey of household
economic outcomes conducted by the Filipino government every three years; and the
Demographic and Health Survey (DHS), a suite of cross sectional household-level health
and fertility surveys. The FIES data allow us to identify the impact of storms on households physical assets, income, and consumption2 , while the DHSs retrospective data
on mothers fertility allow us to reconstruct a mother-by-year panel dataset of infant
births and mortality. Infant mortality constitutes a sensitive3 measure of health itself
as well as an indicator of general household well being. When linked together, these
three datasets allow us to characterize the multidimensional response of households to
typhoons.
We begin by demonstrating that our empirical model indeed captures typhoons
direct destructive impact. We verify that our measure of typhoon exposure, spatiallyweighted maximum typhoon wind speed (henceforth wind speed), is a good predictor
of damages and deaths at the national level. We demonstrate that these nationallyaggregated losses are also apparent at the household-level in the form of lost capital
assets, such as televisions, toilets, and walls.
Turning next to household income, we find that typhoons reduce average income
the year after they strike, presumably due to storms direct physical damages as well
2

We note that expenditures alone do not infer quantity of consumption in the absence of prices,
and thus perform a variety of checks on storms impact on prices, which we find to be negligible; see
Section 5 for details.
3
As Chay and Greenstone (2003) point out, infant mortality minimizes problems of cumulative exposure and a host of other potentially confounding identification concerns that emerge when examining
other human capital measures.

as their more indirect disruption of economic activity. We find that household income
drops linearly by 0.39% per meter per second of wind speed exposure. Given the average
annual exposure at the province level of 16.9 m/s during our sample period, we estimate
that the average short-run effect of the countrys typhoon climate is to depress incomes
by 6.6%. This effect occurs across a variety of income sources and subsamples, affects
both richer and poorer households, and is net of public and private transfers. It is
also large compared to estimates of immediate damages the average annual value of
immediate national aggregate typhoon damage is 0.4% of GDP, only one-seventeenth as
large. This suggests that for the average household, roughly 94% of the economic loss
to a disaster accrues as unearned income long after the destructive event has ceased, in
disagreement with the hypothesis that immediate damages adequately summarize the
full economic cost of these disasters (Mendelsohn, Emanuel and Chonobayashi (2012)).
The income losses we measure translate nearly one-for-one into a reduction of household expenditures, which decrease 7.1% for the average household in the average year.
These expenditure reductions track total income losses closely when they are examined
across years (relative to the storm), across space, across typhoon intensities, across income groups and across subsamples. This tight relationship suggests that households
do not mitigate storm-induced losses via consumption-smoothing strategies, such as
in-kind transfers, savings, or borrowing. Instead, we observe that households make
large adjustments to their relative spending on different types of consumption and investment. In general, households reduce their spending the most on expenditures that
most closely resemble human capital investments, such as medicine, education and high
nutrient foods that include meat, dairy, eggs and fruit. In contrast, expenditures decline
much less on pure consumption goods, such as recreation, alcohol and tobacco.
We next examine whether typhoons impact household health outcomes by examining infant mortality rates. We find no evidence that infant mortality rises during

or immediately following typhoon exposure, implying that deaths from physical exposure to the storm itself, which we term exposure deaths, are few. However, we find
that typhoons cause infant mortality to rise the calendar year after the storm itself
has passed. The vast majority of infant female deaths manifest well after the typhoon
event; moreover, many of the infants who die in the aftermath of the storm were not
even conceived until after the storms are gone, implying that the direct mortality impact of the storm is minimal. We estimate that in our sample of mothers who never
migrate, these typhoon-associated deaths amount to an annual average of 1,130 female
infant deaths per million households, corresponding to 55% of the baseline infant female
mortality rate in this non-migrant sample.
Multiple aspects of our findings suggest that deteriorating economic conditions and
disinvestments in human capital are responsible for these female infant deaths. We fingerprint patterns of economic contraction and disinvestment across many dimensions
by looking at their timing relative to storms, their nonlinear response to storm intensity,
their short and medium-run lag structure at different locations in the income distribution, their structure across subsamples and their spatial patterns. We then demonstrate
that patterns of female infant mortality exhibit an almost identical fingerprint across
these same dimensions. Furthermore, we find that mortality is highest in households
where infant daughters face competition from other children over resources, particularly
if those siblings are male. These findings together suggest that female infant deaths
following typhoons are economic deaths resulting from economic losses and the resulting household decisions regarding human capital investments and within-household
resource allocation4 . Extrapolating these estimates to the entire non-migrant population suggests that approximately 11,300 female infants suffer post-typhoon economic
4

This conclusion that female infants bear a differentially large share of the burden from income
loss is consistent with findings from a variety of other contexts (Rosenzweig and Schultz (1982); Rose
(1999); Duflo (2000); Duflo (2005); Bhalotra (2010)).

deaths in the Philippines every year, constituting roughly 13% of the overall infant
mortality rate in the Philippines. In contrast, there was an average of 743 exposure
deaths per year (across all ages groups) according to official reports for the same period (OFDA/CRED 2009). This suggests that the mortality resulting from economic
behavior in reaction to typhoons is roughly 15 times as large as the immediate mortality
caused by direct exposure to these storms.

The remainder of the paper is structured as follows. Section 2 presents background

on typhoons and household adjustments to income loss. Section 3 presents the data
and Section 4 presents our quasi-experimental design. Section 5 presents our results.
We conclude in Section 6, which discusses our findings and some of their implications
for policy.

Background

The Typhoon Climate of the Philippines

Figure I shows a map of the Philippines annual expected typhoon exposure, or its
typhoon climate. The Philippines possesses one of the most active cyclone climatologies
in the world, on average experiencing over ten typhoons each year ranging in intensity
from mild to severe. Because the Philippines is large compared to typhoons, different
regions within the country may experience entirely different levels of storm exposure in
the same year.
The Philippines active typhoon climate provides additional benefits for analysis
compared to other idiosyncratic destructive events such as earthquakes5 or wars, in
5

Note that while earthquakes certainly have a spatial and temporal incidence structure similar
to typhoons the interarrival time of destructive earthquakes is orders of magnitude longer than for
destructive typhoons. See, for example, Triep and Sykes (1997).

that typhoons in the Philippines are a regular and expected occurrence, with median
exposure to typhoon-strength winds above zero for all but a handful of provinces. While
destructive and unpredictable, typhoon exposure itself is thus not surprising, and households almost certainly incorporate typhoon risk into their economic decisions. We can
thus plausibly infer that any impacts that we observe occur in spite of all the adaptive
responses that households employ to mitigate typhoon impacts.
Our estimate of typhoons costs adds to the rapidly growing literature on the natural
environments impact on health outcomes (Deschenes and Moretti (2009); Maccini and
Yang (2009); Deschenes and Greenstone (2011)), as well as the literature on the economic and health impacts of disasters (Toya and Skidmore (2007); Stromberg (2007);
Cavallo and Noy (2009); Simeonova (2011); Hsiang (2010); Deryugina (2011)) as well as
the more general literature exploring climatic influence on economic outcomes (Gallup,
Sachs and Mellinger (1999); Acemoglu, Johnson and Robinson (2002); Bloom, Canning and Sevilla (2003); Easterly and Levine (2003); Miguel, Satyanath, and Sergenti
(2004); Nordhaus (2006b); Schlenker and Roberts (2009); Dell, Jones, and Olken (2009);
Hsiang (2010); Graff Zivin and Neidell (2010)). It also augments the literature on the
economic consequences of physically destructive shocks (Davis and Weinstein (2002);
Vigdor (2008); Miguel and Roland (2011)) though it differs from much of that literature
in that typhoons, rather than being idiosyncratic events like bombings, are a persistent
and common state of the climate.

Household Adjustments to Income Loss

Reductions in household income have the obvious potential to cause deleterious effects:
consumption of goods and services, investments in health and education, and savings
for future use are all potential margins of adjustment which may suffer following income
loss. There are a variety of means by which households seek to minimize these costs.
8

Firstly, households may attempt to smooth their income or consumption over time,
thereby spreading costs out and attenuating the immediate impact of income loss. This
smoothing can come in the form of within-household adjustments such as accumulating
precautionary savings (Paxson (1992); Kazarosian (1997)), directly supplanting income
through adaptive labor market activity (Kochar (1995); Jacoby and Skoufias (1998);
Kochar (1999)), or selling assets during times of duress (Rosenzweig and Wolpin (1993).
It may also come in the form of extra-household adjustments such as accessing credit
markets (Rosenzweig(1988); Cochrane (1991); Morduch (1995)) or relying on transfers
(Foster and Rosenzweig (2001); Fafchamps and Lund (2003); Yang and Choi (2007)).
It is important to note that the income and consumption smoothing literature differentiates between relatively easily insured-against idiosyncratic shocks (i.e., income losses
affecting different households at different times) and less easily mitigated aggregate
shocks affecting many houses (Cochrane (1991); Townsend (1995)). One might thus
expect that income losses due to typhoons, which are particularly large and common
aggregate shocks, might be particularly difficult to smooth over time.
If income losses cannot be smoothed then households may adjust by altering their
expenditure patterns. This adjustment may manifest in altered consumption, e.g., via
changes in eating habits (Subramanian and Deaton (1996); Jensen and Miller (2008)),
or it may manifest as a reduction in investments, such as to human capital (Mincer
(1958); Jensen (2000); Banerjee and Mullainathan (2010))6 . If losses to income result
in disinvestment in human capital, particularly among children, then the potential
costs of a shock may far exceed its immediately observable effects in the long run via
worsened later-life outcomes (Strauss and Thomas (1998); Maccini and Yang (2009);
Banerjee et al. (2010)). Moreover these losses may become compounded if households
6

Note that in many instances consumption and investment cannot be disentangled; expenditures
on nutritious food, for example, can be equally viewed as consumption as well as investment in future
human capital.

differentially disinvest in children by type, for example due to gender biases (Sen (1990);
Duflo (2005)).
This paper expands upon the literature documenting household disinvestments in
childrens human capital following income loss (Jacoby and Skoufias (1998); Strauss
and Thomas (1998); Jensen (2000)), particularly disinvestments in girls human capital
(Rose (1999); Bhalotra and Heady (2003); Maccini and Yang (2009); Chen (2011)).
More broadly, this paper adds to the growing body of research documenting the excess
risk burden borne by female household members in developing contexts (Horton (1986);
Sen (1990); Duflo (2005); Qian (2008); Robinson and Yeh (2011)).

Data

Our analysis requires data describing household assets, income, expenditures, health
outcomes, and typhoon exposure. Summary statistics of these data are presented in
Tables I, II, and III. For reference, Supplementary Figure A3 displays an administrative
map of the 82 provinces (smaller units) and 17 regions (larger units) we include in our
data.

Typhoon data
A central innovation of our analysis is the development of a comprehensive data file
describing a physical measure of typhoon incidence over time. We develop this measure to ensure that our typhoon data are sufficiently precise to describe meaningful
variations in typhoon exposure in a climate where typhoons are common. We begin by reconstructing the wind field for every West Pacific cyclone in the International
Best Track Archive for Climate Stewardship (IBTrACS) database (Knapp (2009)) using
the Limited Information Cyclone Reconstruction and Integration for Climate and Eco-

10

nomics (LICRICE) model (see Hsiang (2010) for a detailed description of the model7 ).
LICRICE only reconstructs wind fields and does not explicitly account for rains, flooding, or storm surges because wind fields are less influenced by topography and are thus
more generalizable. However, our wind field measures describes these other typhoon
impacts to the extent that they are correlated with wind speed.
We use LICRICE to reconstruct the wind field as a translating vortex for all 2,246
storms recorded in the West Pacific Basin between 1950-2008 by interpolating among
72,901 6-hour observations over every 1/34 1/34 pixel of the Philippines (1/34
0.0294 2.02 miles 3.26 kilometers). Figure II (left panel) illustrates a snapshot
of a storms wind field for an example storm, with the height of the surface depicting
the speed of the surface winds. The right panels in Figure II show how LICRICE
combines this kernel with data on storm trajectories and intensities to reconstruct
annual exposure levels during four example years from our sample. Using this approach,
we find that 837 storms affected the Philippines8 between 1950-2008 (13.72 storms per
year). Of these storms, 411 occurred during 1979-2008 (13.70 storms per year), the
period for which we have overlapping economic and health data. For reference, annual
maps of raw LICRICE output for the period 1979-2008 are presented in Supplementary
Figures A1 and A2.
To match typhoon exposure with annualized socioeconomic data files, our continuous physical measure of typhoons must be summarized to form a single observation
for each location in every year. We summarize annual typhoon exposure for provinces
and regions by computing the maximum wind speed achieved at each pixel and then
taking the average across pixels within an administrative unit. We opt for this measure because it allows us to capture storm intensity while controlling for variations in
7

Since Hsiang (2010), version 2 of LICRICE was built (used in this study), substantially improving
upon the models original accuracy. However these improvements were focused on numerical methods
and the heuristic description in Hsiang (2010) remains accurate.
8
That is, 837 storms registered non-zero wind speeds over at least one 1/34 1/34 pixel.

11

the physical size of regions and provinces9 ; a storm that passes over the entirety of a
geographic area would thus register as stronger treatment than one that merely passed
over a small portion of it. For succinctness, we refer to this statistic as wind speed
and it is presented in the units of meters per second (1 m/s = 3.6 km per hour 2.24
miles per hour)10 .
Figure III displays medians, inter-quartile ranges and extreme values of typhoon
exposure for each of the 82 provinces in our sample during 1950-2008. The figure
illustrates that there is strong variation in typhoon exposure between provinces as well
as strong year-to-year variations in exposure at the level of an individual province. Note
that no province completely escapes typhoon exposure in the period of observation and
there are many provinces that are exposed to typhoons every year. Approximately half
of the provinces have median annual exposures in excess of 20 m/s and many provinces
are exposed to events exceeding 50 m/s. As shown in Table I, the average province was
exposed to wind speeds of 17.6 m/s (s.d. = 12.0 m/s) between 1950 and 2008, or 16.9
m/s (s.d. = 11.6 m/s) between 1979 and 2008.
9

Hsiang and Narita (2012) discuss the variety of tropical cyclone measures that have been employed
in previous econometric studies, such as windspeed at landfall, minimum central pressure and total
energy dissipated. As Hsiang and Narita demonstrate, the spatially-weighted maximum wind speed
measure that is employed in this study is well-supported by theory and outperforms alternative measures in a country-by-year panel analysis. Briefly, the theoretical basis for this measure rests on two
observations. First, the stress-strain relationship for most materials is highly non-linear, with catastrophic failure occurring at a critical level of stress (Nordhaus (2006a)). Thus, for a given material, only
the maximum level of stress that the material is exposed to, i.e. the maximum wind speed, is relevant
for determining whether failure is expected. Secondly, people and capital are distributed across space
within a province, making it necessary to construct some sort of spatial average for wind exposure. We
follow Hsiang (2010) and Hsiang and Narita (2012) and adopt area-weights for our averages because
they cannot be endogenous in the same ways that population, capital or income weights might be. For
further discussion see the Supporting Online Material.
10
It is important to note that because reported wind speed values are area-averages, actual wind
speeds at the center of storms are substantially greater than the values we report and cannot be directly
compared.

12

Household Asset, Income and Expenditure Data

Information on household assets, income and consumption are obtained from the crosssectional Family Income and Expenditure Survey (FIES) conducted by the National
Statistics Office (NSO) of the Philippines (Ericta and Fabian (2009)). In 1957, the
government of the Philippines began conducting the FIES irregularly (approximately
every five years) to understand the distribution of income, spending patterns and the
prevalence of poverty, as well as to benchmark consumer price indices. In 1985, the
survey was completely restructured and the NSO began conducting it at regular three
year intervals. In this analysis, we obtain and use FIES Public Use Files for the years
1985, 1988, 1991, 1994, 1997, 2000, 2003 and 2006.
The FIES provide us with data on each households assets across several different
categories, household income by source, total income net of any transfers and subsidies,
and household expenditures on different goods and services.
We note that there are important timing issues to contend with in analyzing the
FIES data that arise from the manner in which the survey is administered. FIES
data are collected twice for each household, just after the middle of the year (July)
and just following the end of the year (the following January), with responses for each
survey reflecting economic behaviors over the preceding six months. Responses for each
household are then averaged between the two surveys to construct annual estimates;
however, if a household cannot be found in either round of the survey they are dropped
from the sample11 . Figure IV shows the FIES survey timeline overlaid with mean
monthly typhoon strikes to indicate why this is a potential concern. Typhoon activity
in the Philippines is concentrated late in the year, so estimates of typhoon impacts
during the year of exposure may be somewhat attenuated because first phase responses
11

Surveyors attempt to revisit households two additional times if the household head cannot be
located in the first visit. Only after three unsuccessful interview attempts is a household dropped.

13

are recorded prior to the bulk of typhoon events. This motivates us to focus on capital
losses the year following typhoon exposure, since it seems unlikely that capital can be
replaced immediately following a storm12 .
Also concerning is the NSOs policy of dropping second round non-respondents, since
typhoons may cause households to migrate, the obliteration of participants physical
homes, or villages becoming inaccessible due to flooding or infrastructure damage13 .
This results in observations being dropped from our sample based on extreme values
of the treatment effect we are interested in, a fact that probably biases our estimate of
treatment effects towards zero. We attempt to minimize this attenuation by including
a vector of observable household covariates in all our models. In addition, we explicitly
test for balance on treatment in Section 4.
Lastly, we note that the lowest unit of geographic designation in the FIES surveys
changed between the 2000 and 2003 waves; early years include province level identifiers
(more detailed) while later years only report regional identifiers (less detailed). Thus,
our baseline models analyze outcomes at the province level but omit 2003 and 2006. We
then reintroduce these years in region level estimates as a check on our main results. For
province level models we are able to match 142,789 household observations contained in
the period 1985-2000, whereas our sample expands to 174,896 observations in regional
level models that span the period 1985-2006.
12

We note that a typhoons contemporaneous effect on capital is roughly 53% of its effect in the
following year. This is consistent with our concern that contemporaneous effects will be smaller because
phase 1 responses occur before roughly 65% of storm events.
13
The NSO explicitly states that a major cause of second phase survey attrition is the inability
to locate households when the physical structure they inhabited during the first phase interview is
destroyed by a typhoon before the second phase. When areas become inaccessible due to flooding
or infrastructure damage, the NSO generally tries to postpone surveys within the affected region.
Unfortunately, the NSO does not provide statistics on these types of attrition.

14

Infant Mortality Data

Our infant mortality data are taken from the 1993, 1998, 2003, and 2008 waves of the
Demographic and Health Surveys (DHS)14 for the Philippines. The DHS are crosssectional surveys with questions related to population, health, and nutrition, particularly pertaining to maternal and child health. The DHS program is highly standardized
with changes between surveys documented and propagated, allowing for comparison of
surveys both across countries and within countries across time. Samples are designed
to be representative at the national and regional levels. Within the Philippines, each
households location is identified according to its administrative region and provincial
identifiers are not available.
The primary interview targets of the DHS are women between the ages of 15 and 49.
A wide suite of questions are asked on topics ranging from HIV awareness to nutritional
practices to each womans full fertility history. The latter provides us with a source
of our infant mortality data, as each woman is asked to provide detailed information
about every child she has ever born, including any children who have died. We are
thus able to construct a time series for each womans fertility and mortality events
over the duration of her life up until the survey, echoing recent research that uses
the DHS in a similar way (Kudamatsu (2011); Kudamatsu, Persson, and Stromberg
(2011); Chakravarty (2011)). We follow these authors in excluding migrant mothers
from our sample, thereby minimizing the sorting and migration concerns that arise in
the FIES. The 24,841 non-migrant mothers in our sample yield 265,430 mother-byyear observations, or nearly 11 years of longitudinal data per woman. Table III shows
summary statistics.
14

The DHS are administered by the Measure DHS project (funded largely by USAID) and are available for free download online at http://www.measuredhs.com/. Started in 1984, the DHS program
has collected survey data on 84 countries as of late 2011, with many of those countries having been
subject to multiple survey waves.

15

The DHS data include several variables aside from infant mortality events which are
particularly useful for this analysis. Of particular note are: a measure of each womans
prior migration history, captured by whether she has ever lived anywhere else and, if
so, when she moved to her current location of residence; educational attainment of
both the woman and her husband, if any; the geographical region in which the woman
resides; and the womans age at time of survey. While there is no direct questioning on
each womans or households income, a variety of socioeconomic status (SES) indicators
are collected, ranging from whether the household has electricity to whether anyone in
the househole owns a car. We construct a proxy for socioeconomic status from these
data for comparison of distributional impacts in a process outlined in the Supplemental
data.

Other Data
Emergency Events Database (EM-DAT) Nationally aggregated data on economic losses and deaths from tropical cyclones are obtained from the Emergency Events
Database data file commonly referred to as EM-DAT (OFDA/CRED 2009). The
EM-DAT data file contains information provided by national governments, international organizations, NGOs, and private companies (e.g., re-insurance companies) on
a self-reporting basis15 . EM-DAT data of economic losses are an estimate of negative
economic impacts that may include lost consumption goods, lost productive capital or
cost of business interruption, depending on the protocols of the reporting institution.
EM-DAT is the database used in most previous cross-country studies of post-disaster
economics, with some of its limitations discussed in the review by Cavallo and Noy
(2009).
15

EM-DAT is provided for free by the Centre for Research on the Epidemiology of Disasters (CRED)
at www.emdat.be, Universite Catholique de Louvain, Belgium.

16

Temperature and Rainfall We control for mean annual temperature and rainfall in
all of our analyses to minimize potentially confounding climate behaviors that might be
correlated with typhoon incidence (Auffhammer et al. 2010). Temperature observations
are extracted from the gridded reanalysis of the Climate Data Assimilation System I
(CDAS1) produced by the National Center for Environmental Prediction (NCEP) and
the National Center for Atmospheric Research (NCAR) (Kalnay et al. (1996). Rainfall
estimates are obtained from the Climate Prediction Center (CPC) Merged Analysis
of Precipitation (CMAP) which merges station readings on the ground with available
satellite data (Xie and Arkin 1996). Both temperature and precipitation data are
spatially averaged over each region or province.

Crop Prices Province-level data on annual commodity retail prices are obtained
from the Bureau of Agriculture in the Philippines16 . Data are available for the period
1985-2008.

Identification

To empirically identify the impact of typhoons on household outcomes we use a differencein-differences approach that exploits random variations in each locations typhoon incidence. Identifying the treatment effect of typhoons requires that we must only utilize
variations in typhoon exposure that are randomly assigned to households (Holland
(1986); Freedman (1991)). Because the formation of typhoons and their trajectories
have strong spatial patterns, some locations have relatively higher or lower levels of
average typhoon exposure (recall Figure I). However, these cross-sectional variations in
mean exposure might be correlated with cross-sectional differences in the unobservable
characteristics of different locations, for example culture. For this reason, we do not
16

Details and data are available at http://countrystat.bas.gov.ph

17

utilize the cross-sectional variation in average exposure and instead rely only on random
year-to-year variations in exposure at each specific location. To achieve this, we include
province (or region) fixed-effects in all of our regressions to absorb any cross-sectional
variation in typhoons exposure or losses. If there are unobservable reasons why some
locations have higher (or lower) incomes or infant mortality on average, these fixedeffects will non-parametrically account for this difference and it will not contaminate
our estimates of the typhoon treatment effect (Greene (2003)).
Randomness in typhoon exposure arises because both the location and timing of
storm formation as well as storm trajectories themselves are stochastic. One might
be concerned that annual variations in storm exposure might not be entirely random
because households could make location choices based on seasonal typhoon forecasts.
Yet, while it is now possible to predict average storm frequencies for each storm season
in a given basin with moderate skill (Heming and Goerss (2010); Smith et al. (2010)),
these forecasts have almost no predictive power if one were to try forecasting locationspecific seasonal risk. Thus, it is reasonable to assume that annually varying risk
differentials are imperceptible for individuals on the ground, since these differentials
still cannot be predicted by scientists. In contrast to seasonal prediction, it is possible
to forecast typhoon exposure a few days before a storm strikes17 (Heming and Goerss
(2010)), a fact that often allows individuals to evacuate and protect some of their
assets. This is important for interpreting our results, because the treatment effect that
we estimate is the effect of typhoons after households have employed the full range of
adaptive behaviors available to them, such as evacuation. But it does not seem plausible
that short-term evacuations based on short-term forecasts lead to the reorganization of
populations on annual time-scales, so it is unlikely that forecast-based sorting affects
17

For example, Willoughby et al. (2007) note: In the past, a forecast was considered successful if it
predicted the hurricanes position and intensity 12 - 72 h into the future. By the 1990s, forecast users
came to expect more specific details such as spatial distributions of rainfall, winds, flooding, and high
seas. In the early 21st century, forecasters extended their time horizons to 120 h.

18

our annualized estimates.

We wish to avoid spurious correlations, so we must avoid correlating trends in typhoon incidence and our outcomes of interest. To do this, we flexibly account for
common trend behaviors by including year fixed-effects in all of our models (Greene
(2003)). These fixed effects also account for any unobservable common climatic shocks,
such as the El-Ni
no-Southern Oscillation, which could be correlated with typhoon exposure (Camargo and Sobel (2005)) as well as socio-economic outcomes (Hsiang, Meng
and Cane (2011)).
The primary threat to the validity of our study is the potential for household sorting
in the wake of typhoon exposure. As we explained, sorting due to typhoon risk should
not be a major concern since we include location fixed-effects and annual changes in
risk are imperceptible to households. Sorting on typhoon incidence, however, could be
problematic if the passage of a storm causes families to migrate away for long periods,
altering the household composition of different locations. This is of particular concern
for the FIES data given their survey methodology (discussed in Section 3).
To address this concern, we test for balance in the FIES data by regressing observable household characteristics on typhoon exposure, presenting results in Table IV.
This approach checks whether observable household characteristics vary with the intensity of the previous years typhoon exposure. We allow household composition to vary
nonlinearly in response to typhoon exposure by including indicator variables for prior
years maximum wind speed18 . In support of our approach, we find almost no evidence
of sorting. Out of 49 parameters estimated, six are statistically significant at the 10%
level and one is significant at the 5% level; this is very close to what we would expect if
household composition were random (five and two respectively). If one interprets these
coefficients literally, they might provide suggestive evidence that typhoon exposure is
18

This exact model is used throughout the paper to identify the effect of typhoons on time-varying
outcomes. It is explained in greater detail in the next section.

19

positively associated with total family size and negatively with the probability that the
household head has finished primary school. However, in neither case does the intensity of cyclone exposure matter in a systematic way, suggesting that these correlations
are probably random19 . Nonetheless, to be certain that bias from sorting along these
covariates is minimized, we control for all of these all of them in our main regression
models.
We are less concerned about sorting in the DHS data for two reasons. First, households in the DHS are asked whether they have ever lived anywhere other than their
current location. This allows us to directly avoid sorting behavior by restricting our
sample to non-migrant mothers. The second reason is that, unlike FIES data, the DHS
consist of panel data that allow us to follow specific women over time. Thus, there are
no compositional changes in the DHS panel that can be driven by typhoon exposure
during the mothers adult life.

Results

We structure the presentation of our results as follows: We first demonstrate that our
measure of typhoon incidence accurately predicts physical damage at both the macro
and micro level in Section 5.1. We then demonstrate in Section 5.2 that the legacy of
this physical destruction leads to losses to income the year following storms, which are
closely matched by expenditure and consumption losses as detailed in Section 5.3. We
then demonstrate the infant mortality impacts of typhoons in Section 5.4 and provide
evidence supporting the argument that they stem from economic losses. Lastly we
explore cross-sectional evidence of adaptation in Section 5.5.
19

See Supplemental data for additional discussion.

20

5.1

Physical damages

It may seem obvious that typhoons are physically destructive, but measuring the economic importance of this destruction is not trivial. The first studies that used aggregate
measures of tropical cyclone (including typhoon) losses were unable to detect any effect
of storm intensity on losses (Kahn (2005); Noy (2009)). If this result were accepted
at face value, it would imply that variations in the intensity of cyclone climates have
no effect on economies. In this section we show evidence that our measure of typhoon
incidence predicts physical damages using both national data from EM-DAT as well as
asset loss data from FIES.

Evidence from national data

We begin by presenting prima facie evidence that aggregate losses scale with typhoon
intensity in the Philippines. Using standard EM-DAT estimates for all the economic
losses and deaths attributed to typhoons in each year, we estimate whether national
losses increase with wind speed exposure, averaged over the entire country. National
losses and their bivariate dependance on wind speed are shown in Figure V. In Table
V we present several ordinary least-squares estimates for the time-series regression

Zt = Wt + + 1 t + 2 t2 + t

(1)

where Z is the log of total deaths or total economic losses, W is typhoon wind speed,
is a constant, 1 and 2 are trend terms and  is variation that we do not explain.
Following Pielke et al. (2003) and Hsiang and Narita (2012), we also present models
where the dependent variable Z is normalized by the size of the economy (GDP) or the
countrys population.
We find that national average typhoon exposure explains about a third of the varia-

21

tion in EM-DATs estimates for both total typhoon damages and total typhoon deaths.
In all models, the intensity of wind exposure is highly significant, with an increase in
wind exposure by one meter per second increasing losses roughly 22%. We note that
the economic damages estimated by EM-DAT include capital losses, lost revenue and
any other economic cost that is associated with a storm, but it is impossible with
these data to uncover finely-grained structure that might indicate the mechanism by
which either damages or, for that matter, deaths occur.

Household asset losses

To estimate typhoons impact on household assets, we use ordinary least-squares regression to estimate the linear probability that a household has each of several different
types of physical capital recorded in the FIES data. We control for unobserved household attributes common across households in a given year or province by including
province and year fixed effects. We further augment the model with controls for households observable characteristics, namely: the total number of household members; the
number of members above fourteen years old; and age, gender and education level of the
household head. Finally, we control for the annual mean temperature and rainfall observed in each province in each year, since these variables are known to affect economic
conditions (Miguel, Satyanath, and Sergenti (2004); Nordhaus (2006b); Schlenker and
Roberts (2009); Dell, Jones, and Olken (2009); Hsiang (2010)) and they are driven
by many of the same climatological factors that affect typhoon incidence20 . Thus our
complete regression model is

Zhprt =

5
X

[L Wp,tL + L Tp,tL + L Rp,tL ] + t + p + Xh + rt + ht

(2)

L=0
20

For example, typhoon activity in the West Pacific is affected by the El Ni

no-Southern Oscillation
(Camargo and Sobel (2005)), which also influences temperatures and rainfalls in the Philippines .

22

where h indexes households, p indexes provinces, r indexes regions and t indexes years.
Z is a one if a household has an asset and zero otherwise while W is typhoon wind speed,
T is temperature, R is rainfall, is a year fixed-effect, is a province fixed-effect, X is
the vector of observable household characteristics, rt is a shock affecting all households
in a region and ht is a household level disturbance. We employ a distributed lag model
to examine the effect of typhoon exposure for the five years prior to the survey, with lags
indexed by L. In addition, because region-level shocks may exhibit unknown patterns
of serial correlation and household-level shocks may exhibit spatial correlations at a
sub-regional but supra-provincial scale, we cluster our estimated standard errors at the
region level following Bertrand, Duflo, and Mullainathan (2004)21 and Conley (1999).
Table VI presents estimates of for four of the most general and widely owned
household assets: a closed toilet (eg. not a pail or open pit), a television, walls constructed with primarily strong materials (compared to light or salvaged materials) and
access to electricity. We find that for all these assets response to previous years typhoon
treatment is negative and significant, varying between a 0.11% and 0.16% probability
of loss per m/s of typhoon treatment, or 1.9 - 2.7 % given the average annual provincial treatment of 16.9 m/s. Cars, which we also show, do not respond at all, possibly
because they are a valuable asset that can easily be moved quickly when typhoon forecasters warn populations about an impending storm22 . We show additional results of
typhoon treatment for an array of other household assets in Supplementary Table A1.
We note that the average coefficient across all 14 assets in year zero is -0.036, versus
an average coefficient in the first year lag of -.069. We estimate that the asset response
in year 0 is thus 52.5% of the response in year 1, consistent with our observation that
21

Because regions are aggregations of provinces and provinces are the level of treatment, clustering
by region means we are also clustering at the level of treatment. Clustering at the province level does
not appreciably change our results.
22
It is also possible that the coefficients for cars is small because there are a limited number of
households in our sample that ever possess a car.

23

year-of typhoon impacts estimated using FIES data will be biased downwards due to
averaging across the two waves of the survey and possible attrition.

Non-Linear Estimates of Asset Losses We now relax, and test, the assumption
that physical damages, and hence asset losses, are linear in windspeed. It is plausible
that damage is highly non-linear in windspeed; for example, Nordhaus (2010) and
Mendelsohn et al. (2012) argue that losses are a power function of windspeed at landfall.
However, these papers examine aggregate storm damages, similar to Equation 1, so it
is not obvious whether estimates using our micro-data should have similar functional
forms. Thus, we estimate the losses to wind speed non-parametrically, allowing the
response function to have an arbitrary functional form, and examine whether it is
approximately linear or not. To do this, we construct dummy variables that are one if
exposure falls within a five meter per second range and zero otherwise
[x,x+5)
p,t1
W
= 1 [Wp,t1 [x, x + 5)]

leaving events with 0-5 meters per second as the dropped bin. We then run the regression from Equation 2 where the inner product of these dummy variables and their
coefficients replace the linear term 1 Wp,t1 . To limit the number of estimated parameters, we keep the remaining terms unchanged and focus our attention on the coefficients

24

of these dummy variables. The full nonlinear model that we estimate is

[5,10)

Zhprt =1

[5,10)
[10,15) [10,15)
[15,20) [15,20)
[20,25) [20,25)
p,t1
W
+ 1
Wp,t1 + 1
Wp,t1 + 1
Wp,t1 +

[25,30)

[25,30)
[30,35) [30,35)
[35,) [35,)
p,t1
W
+ 1
Wp,t1 + 1
Wp,t1 +

L Wp,tL +

L{0,[2,5]}

5
X

[L Tp,tL + L Pp,tL ] +

L=0

t + p + Xh + rt + ht

(3)

and it is estimated using the same method and sample as Equation 2. Panel A of
Figure VI displays these coefficients for six of the main asset types. The probability
that households lose electricity, a closed toilet, walls made of strong materials, their
television or their refrigerator increase approximately linearly with typhoon wind speed
exposure. In contrast, the probability that a household loses a car to typhoon exposure
remains near zero.
The linearity of these response functions indicates that our earlier estimate for the
average number of households missing an asset due to typhoons was a good approximation. Furthermore, the coefficient for the 15-20 meter per second bin is generally
in the range of 1.5-3%, matching our earlier linearized estimates. Finally, it is worth
noting that exposures exceeding 35 meters per second (spatially-averaged) are not uncommon, recall Figure III, and these stronger events cause 4-7% of households to lose
their immobile assets.

5.2

Income Losses

Our approach to estimating income losses mirrors our approach to estimating physical
damages. We focus our attention on income earned by households the year following storm exposure, partly to minimize the aforementioned attenuation risk in the

25

FIES data and partly because that is where the result manifests most strongly. We
again note that the measures of total household income collected by FIES include all
reported transfers from other households and the government. Prior work by Yang
(2008) demonstrated that tropical cyclone strikes increased remittances to some countries, suggesting that transfers provided a mechanism for income insurance. In addition,
Fafchamps (2003) used Filipino micro data to show that some income shocks lead to
inter-household transfers that partially compensate for losses23 . These previous studies
suggest that transfers might be important for mitigating household income losses, so it
is fortunate that our measures of total income account for them.

Household Income Losses

To estimate the effect of typhoons on income, we estimate Equation 2 replacing Z with
the natural logarithm of household income. Table VII presents these results24 in in
columns 1-4. Including all all our control variables, we find that household income
falls by 0.39% for each additional meter per second of windspeed exposure the year
prior. This implies that under average exposure levels (16.9 m/s), average household
income is depressed 6.6%. In column 5 we estimate the same model except we match
households to the average exposure of its region (the larger administrative unit) rather
than its province. Doing this allows us to include the 2000 and 2006 waves of the
FIES which we could not do otherwise because households in these waves lack province
identifiers. Using this longer sample with coarser measures of exposure, we continue to
find a large effect of typhoon exposure on household income. In a final specification
check presented in column 6, we collapse our household data to the province level,
dramatically reducing our number of observations from 142,779 to 367. This allows us
23

Deryugina (2011) finds similar results for Federal transfers in response to hurricanes in the United
States.
24
Similar to our results for capital losses, province fixed-effects are the most important control for
limiting bias.

26

to conduct two additional checks: (1) whether we are over-estimating our true number of
independent observations (Bertrand, Duflo, and Mullainathan (2004)) and (2) whether
spatial correlations in  cause us to underestimate our standard errors25 (Conley (2008)).
Estimating an analog26 of Equation 2 using this collapsed data set and estimating
spatially-robust standard errors27 we find our coefficient of interest unchanged and that
our standard errors increase only slightly.
In Table VIII, we examine whether wage or entrepreneurial income responds more
strongly to cyclone exposure. Entrepreneurial income is income from self-employed
activities, including own agricultural cultivation, whereas wage income is income earned
by selling labor to firms or other households. We find that self-employed entrepreneurial
income responds strongly and negatively in the year following storms, falling -0.28%
per m/s or 4.7% for average treatment. Non-agricultural wages also fall by an average
of 3.2% per year, although the effect is not significant, and it is worth noting that both
coefficients are negative, but not significant, in the second lagged year as well. This is
in stark contrast to agricultural wages, which do not respond negatively in the first lag
and exhibit little systematic variation in response to typhoon exposure.
Table VIII also presents the estimated value of 1 for different categories of entrepreneurial income ranked by the number of respondents claiming any income from
that source. None of these estimates are statistically different from zero, probably because the sample size declines rapidly for each subcategory. However, we find that
the point estimates for lost income are consistently negative with only two exceptions:
earnings from gambling and income in the transport and storage industry. We thus are
25

In theory, we could explicitly account for spatial correlation in errors using our micro-data, however
it is not computationally cost effective.
26
Using our collapsed data set requires that we introduce a lagged dependent variable into Equation
2 because aggregated output measures are highly correlated over time. We do not do this in the model
with household data because that data set is not a true panel, so we do not know what household
incomes were in the last period of observation.
27
For technical reference, see Conley (1999).

27

confident in stating that income losses seem to not be driven by losses confined to a
single sector.

Non-Linear Estimates of Income Losses We verify that our linear approximation

of the income response is reasonable by estimating Equation 3 for household income.
Panel B in Figure VII plots the coefficient for each wind speed bin along with its
confidence interval, and panel B of Figure VI presents coefficients for the entrepreneurial
and non-agricultural wage components of income. All of theses measures of household
income loss are approximately linear in wind speed exposure. This linearity suggests
that Equation 2 is a good approximation of the response function and it agrees with
previously measured GDP responses to tropical cyclone exposure (Hsiang (2010)).

Subsamples We examine whether the loss of income that follows typhoon exposure
differs across subpopulations by estimating Equation 2 on stratified subsamples of our
data. We separate the sample based on (1) whether a households location is urban or
not, based on the NSOs scoring of a location, (2) the sex of the household head, and (3)
whether the household head completed primary school. Across all subpopulations the
overall lag structure is similar to the pooled estimate, with income loss in the first year
after exposure being large and subsequent labs being smaller or zero. In the first row of
Table IX, we tabulate the estimated effect of typhoons on income in the first year after
exposure. Households headed by a male and households headed by an individual with at
least a primary school education are estimated to suffer income losses that are slightly
larger, but none of the differences in estimated effect-sizes are statistically significant.
In general, the income loss following typhoon exposure appears to be pervasive across
these subpopulations.

28

Income Losses at Different Locations of the Income Distribution

Kahn (2005), Skidmore and Toya (2007), Noy (2009), Hsiang (2010), Hsiang and Narita
(2012), the United Nations (2009) and the World Bank (2010) have all suggested that
poor populations experience larger relative losses to natural disasters, including tropical cyclones. Though compelling, these analyses have been based on country-level
comparisons of income which may be tainted by an array of confounding factors as well
as subject to omitted variable bias. We explore whether this relationship is plausible
within-country by comparing losses for high and low-income households that inhabit
the same Filipino province and are subject to the same institutional environment.
The FIES are not a true panel, so we cannot condition household income (or capital)
losses on income in the previous period. We overcome this limitation by comparing how
the income distribution in each province responds to typhoon exposure. To do this, we
first collapse the data by province-year, retaining estimates of Yptq , the household income
at the q-quantile of the income distribution for province p in year t. Thus, for each
value of q we have a panel of province by year observations that we use to estimate the
model

q
Yptq = q Yp,t1
+

5
X

[Lq Wp,tL + Lq Tp,tL + Lq Rp,tL ] + tq + qp + qpt

(4)

L=0

where all coefficients are q-quantile specific versions of those described in Equation
2 and q is a q-quantile specific autocorrelation coefficient that we introduce because
our collapse of the dataset generates substantial autocorrelation (recall column 6 of
Table VII). Similar to before, our variable of interest is 1q , the relative shift of income
observed at the q-quantile following the previous years typhoon exposure.
Panel A of Figure VIII presents OLS estimates of 1q for q {10, 20, ...90} along
with confidence intervals. Strikingly, the semi-elasticity of income to wind speed is

29

practically constant at all points along the income distribution, with the response always
near the average household response (horizontal line). These results seem at odds with
earlier cross-country studies that found different short-run responses for high and lowincome populations. Yet, a completely different picture emerges when we examine the
cumulative impact of typhoons on the shape of the income distribution. Panel B shows
P
estimates for the cumulative effect of one additional meter per second in wind: 5L=0 Lq .
When we sum coefficients for all the years following a storm, we see that incomes below
the median suffer much larger cumulative losses than income above the median, which
actually exhibit no cumulative losses. This occurs because losses at low ends of the
income distribution persist for several years after the storm, whereas incomes at the
high end of the distribution actually rise slightly above average a few years after the
storm, allowing these groups to recover previously unearned income. Thus, when we
look at typhoon-induced income losses beyond the first year we find strong evidence
that the income distribution widens, with low-income households suffering differentially
larger cumulative losses when compared to high-income households.

5.3

Consumption and Investment Effects

Having found strong evidence that household incomes respond negatively to typhooninduced economic losses, we examine whether consumption and investment expenditures also adjust. To determine whether expenditures fall, we implement the same
analysis that we conducted for income, but instead examine the consumption and investment variables available in FIES. Broadly, these variables are all expenditures,
although it is not always possible to clearly distinguish whether a specific variable represents consumption or investment, since many expenditures represent a combination
of the two. For example, recreation is clearly a consumption good, while education
is mostly an investment in human capital, but food is probably some combination
30

since food consumption increases immediate utility but also augments health, a specific
dimension of human capital.
We note that a change in expenditures absent information about a potential change
in prices does not allow one to infer changes in consumption. We check for and find
little evidence that typhoons affect regional prices for food in Supplementary Tables
A2-A4.

Household Losses to Consumption and Investment

We estimate Equation 2, replacing Z with the logarithm of household expenditures,
and present the results in Table X. In column 1 we show that household expenditures
fall 0.42% for each additional meter per second in the prior years typhoon wind speed.
This implies that under the average level of exposure (16.9 m/s), total expenditures are
7.1% lower than they would otherwise be due to the transient impact of the typhoon
climate, mirroring the average income loss of 6.6%. In columns 2-11, we estimate
Equation 2 for different expenditure subcategories. For eight out of ten subcategories we
observe similar patterns of losses with the exceptions being recreational expenditures,
which declines insignificantly, and repairs to the households capital assets, which rise
slightly. Notably, some of the largest relative declines in spending occur in categories
related to human capital investments: personal care (0.74% per m/s), medical services
(0.85% per m/s), and education (0.79% per m/s). Food expenditures, another type
of investment in human capital, do not decline as strongly. However, clear reductions in
food expenditure seem to extend over a longer period of time, beginning immediately
in the year of storm exposure and continuing for three years afterwards.
In Table XI we decompose the response of food expenditures into its different subcategories. The strongest declines are clearly in the purchase of meat, with strong
responses also appearing in the fish, dairy & eggs and fruit categories. Purchases of

31

cereal also decline, but much less than the more nutritious foods. The overall structure
of this response is consistent with previous observations by Subramainian and Deaton
(1996) and Jensen and Miller (2008) that real income losses lead to a shift in food
consumption that protects overall calorie intake at the expense of nutrients.
The last three columns of Table XI present the expenditure response for nonalcoholic
beverages, alcoholic beverages and tobacco. All three types of purchases decline in the
year of typhoon exposure, however their responses after that year diverge: nonalcoholic
purchase remain low for up to three more years, alcoholic purchases mostly recover but
also have long but statistically insignificant declines, while tobacco purchases become
insignificantly positive in the year following exposure but then return to their original
level. This relatively lower income elasticity of alcohol and tobacco, pure consumption
goods, relative to more nutritious foods, partially human-capital investments, agrees
with our earlier observation that other non-food varieties of human capital investments
decline more rapidly than expenditures that more closely resemble pure consumption
goods (Table X). This finding is consistent with the hypothesis that wealth shocks
directly alter the utility function of household members by increasing the marginal
utility of immediate consumption (e.g., Banerjee and Mullainathan (2010)).

Non-Linear Estimates of Consumption and Investment Losses Following our

earlier estimates, we verify that our linear model is a good approximation of the expenditure response by non-parametrically estimating the impact to typhoon exposure.
We estimate Equation 3 for total expenditures and present our coefficients in Panel D
of Figure VII. We find that the response of expenditures is almost exactly linear and
mirrors the response to income, which we illustrate by overlaying the two responses
in Panel C of Figure VI. When we examine the various subcategories of expenditures,
which we show in Panel D of Figure VI, we continue to observe responses that are linear

32

in typhoon wind speed. Inspecting subcategories of food purchases, displayed in Panel

E, we see the same linear structure.

Subsamples We examine whether the reduction of expenditures following typhoon

exposure differs based on whether the household is urban or not, the household heads
sex and household heads level of education. We stratify the sample along these dimensions and estimate Equation 2, tabulating the estimated effect of typhoons on
lagged expenditures in the second row of Table IX. Male headed households reduce
their expenditure slightly more than female headed households, mirroring the 0.1%
per m/s differential in average income loss for these subpopulations. However, none
of the differences in expenditure responses are statistically significant for any of these
stratifications.

Expenditure Losses at Different Locations of the Expenditure Distribution

We next examine how the expenditure distribution responds to typhoon incidence,
demonstrating that it mirrors the response of in the income distribution. We estimate
Equation 4 for both total expenditures and food expenditures, displaying our results in
panels C-F of Figure VIII. Identical to our results for the income distribution, the total
expenditure and food expenditure distributions shift coherently the year after typhoon
exposure. This shift persists for the following years at q-quantiles below the median,
generating large cumulative impacts. In contrast, total expenditure actually rises in
later years for households above median expenditure, leading to a cumulative impact
near zero. Households above median food expenditure do not consume extra food in
later years, however, so cumulative effects are observable throughout the distribution.
It is plausible that this occurs because the benefits or utility from food consumption
are not substitutable over time periods longer than a year.

33

5.4

Infant mortality

Our final analysis of household outcomes centers around estimating infant mortality.
We first show that the entirety of the child and infant mortality response is driven
by infant female deaths, and then provide evidence suggesting that these deaths are
attributable to typhoon-induced economic losses and the resulting household decisions.

The Response of Infant Mortality

We estimate the effect of typhoons on child mortality by altering Equation 2 to reflect
the structure of the DHS data. We arrive at the model

Zwrt =

5
X

[L Wr,tL + L Tr,tL + L Rr,tL ] + t + w + Xwt + rt + wt

(5)

L=0

where w indexes a woman. Here, Z is one if a woman reports that a child of the relevant
demographic category died in year t and zero otherwise. Xwt are the time-varying traits
of a womans age and age-squared. w is a woman-specific fixed-effect that controls for
any time-invariant woman-specific traits. Estimates with this model benefit from the
fact that our reconstructed panel contains more years (24) than the FIES data (6);
they suffer, however, in that DHS reports only the region a woman lives in and not her
province, substantially shrinking the number of distinct treatment groups that we have
in any given year (13). To account for the fact that all women in a region are coded
as receiving the same typhoon exposure, as well as to account for any serial correlation
within or between women that in same region, we cluster our standard errors at the
region-level.
We present estimates for our parameter of interest, , in Table XII. The coefficients
report the number of additional women, out of one-million, who report the death of a
child in association with an increase in wind speed of one meter per second. The first

34

column shows that there is a detectable increase in child mortality the year following
a typhoon, with roughly 80 additional deaths (per one-million women) for an increase
in exposure by one meter per second. For the period of observation, regional mean
exposure was 15.3 meters per second. This suggests that in an average year roughly
1,220 women out of one-million would report a child dying due to typhoon exposure the
prior year. Columns 2 and 3 decompose this response into deaths of male and female
children, revealing that the bulk of these deaths are among females. Columns 4 and 5
examines whether theses female deaths are from young children and infants, and we see
that almost all of the additional deaths are infant females: of the 80 child deaths per
m/s, 73 of them are female infants. In contrast, examination of infant males in column
6 reveals that they do not contribute to the observed mortality response in the first
lagged year at all.

Attributing Infant Mortality to Economic Conditions and Household Decisions

We chose to examine child mortality partly because we claimed that they reflect human
capital stocks, and we interpret these female infant deaths as evidence that households
are disinvesting in the health of their female infant children. However, it is possible
that these deaths are not a result of disinvestment, but instead result from the physical
trauma of exposure to the typhoon itself or the typhoons aftereffects on the ambient
environment, e.g., disease ecology. We lack data on the proximate cause of death, so
we cannot completely rule out these hypotheses. However, we are able to demonstrate
that this mortality response mirrors the economic response across multiple dimensions,
strongly suggesting that economic conditions are causing these infant deaths. Further,
we are able to demonstrate that the patterns of infant mortality are consistent with our
understanding of how households reallocate resources in response to a wealth shock.

35

Temporal Structure The timing of female infant deaths does not suggest that they
are a result of direct exposure to typhoons. We illustrate this point in Figure IX,
where we estimate a version of Equation 5 using regional average mortality rates at
the monthly level. The black line shows cumulative monthly infant female mortality
impact of typhoon exposure, i.e., the rolling sum of coefficients for months 12 months
before typhoon impact through month n, normalized such that the cumulative effect
evaluated at the month preceeding typhoon impact is 0. We see that the coefficient
of typhoon impact on month-of mortality is near zero, and the increase in mortality
rates does not manifest until nearly a year after the storm has hit. Deaths continue
to accumulate past the 12 month mark, beyond which all infants in the sample have
been born after typhoon exposure. We can thus conclude that the bulk of deaths occur
significantly later than any immediate traumatic impact of the storm could plausibly
be acting.
It is reasonable to question whether fetal exposure to typhoons may be partly driving our results. The recent explosion of literature in fetal origins28 , including natural
disasters impacts on them (Simeonova (2011)), suggests that exposure to shocks while
in utero can have serious deleterious effects on later health. One might thus posit that
in utero weakening contributes to or perhaps even drives the increase in death rates. We
explore this claim in Figure IX where the dotted grey line shows the number of female
births resulting in an infant deaths per million households. We see that in utero effects
may be contributing somewhat to the increase in deaths, as evinced by the increase
in births ending in infant death immediately after typhoon impact. We nonetheless
note that a large portion of the total cumulative births resulting in deaths occur 9 or
more months after typhoon exposure, when in utero effects are strictly impossible. We
conclude that while in utero impacts may be accelerating the increase in infant deaths,
28

For a detailed overview see Almond and Currie (2011).

36

they can at most be an auxiliary. This observation is further supported by the lack of
a similar death pattern among infant males as detailed in the Supplementary data.
Returning to the annual data, we note that there is a striking agreement between
the timing of depressed economic conditions and female infant mortality. The left
panels of Figure VII display the timing of income losses, expenditure reductions and
female infant mortality. The spike in mortality coincides with the sharp reductions
in income and expenditure described earlier. Both mortality and economic conditions
remain abnormal two years after a typhoon, although effects are smaller and are only
marginally significant. Differences only arise three and four years after the storm, when
mortality remains slightly elevated but average income and expenditure return to their
baseline values.

Non-linear Structure We look for nonlinear structure of the mortality response

by altering Equation 5 so that 1 is decomposed following Equation 3. Panel F of
Figure VI displays the response function for all children, all infants and female infants.
The responses of the larger samples are noisy but approximately linear, but when we
isolate the infant female deaths that are driving the pooled response we see an almost
exactly linear response. Presenting the income, expenditure and female infant mortality
responses in the right panels of Figure VII, we see that all three all three match in their
linear responses to typhoon wind speed exposure.

Structure Across Subsamples We examine whether female infant mortality following typhoon exposure differs based on whether the household is urban or not, the
household heads sex and household heads level of education. In the third row of Table
IX, we tabulate the estimated effect of typhoons on female infant mortality for these

37

subsamples29 . Similar to the response of income and expenditure, there are no statistically significant differences in the mortality response across these subpopulations
and the point-estimates for the mortality response within male-headed households and
households headed by more educated individuals is somewhat higher. We note that
the pattern of heterogenous mortality effects across subsamples matches the pattern
of expenditures but the relative difference in effect-sizes tend to be greater for mortality (35-63% larger for the more affected group) compared to expenditures (7-29%)
unfortunately, it is difficult to know if these larger differences are meaningful or arise
because the standard errors of our mortality estimates are substantially larger than in
our estimates for economic responses. Nonetheless, the overall pervasiveness of infant
mortality across subsamples matches the pervasiveness of income loss and expenditure
reduction.

Distributional Structure The DHS data lack income information, so to examine

distributional aspects of our results we instead examine the mortality response at different locations in the wealth distribution. We outline our method for inferring household
wealth in the Supplementary data. In Table XIII we present the response of female
infant mortality for women above and below the median for assets, as well as for the
bottom and top deciles. We find that in the year immediately following storm exposure,
mortality is slightly higher in the lowest wealth groups, and moreover remains elevated
for several years. This pattern of relative uniformity in the year after the storm, with a
slower recovery for poor households, matches the response of income, expenditures and
human capital investments, as we illustrate in Figure 4.
29

Unlike earlier stratifications based on the FIES (for income and expenditure), we were unable to
construct the education level of the household head at the time of typhoon exposure. Instead, we
used the level of education for the mother. This alternative variable provided a very similar relative
partitioning of the data, with 34.8% of non-migrant mothers (in DHS) having completed primary
school compared to 36.0% of household heads (in FIES).

38

Spatial Structure Up to this point, we have only estimated response functions that
pool all Filipino observations, however it is possible that some regions are more or
less susceptible to typhoon-induced economic losses. If this is true, and if economic
losses are the mechanism through which typhoons increase female infant mortality,
then regions suffering larger typhoon-induced economic losses should also exhibit larger
typhoon-induced infant mortality. To examine whether this is the case, we estimate
region-specific versions of the coefficient 1 . We do this by modifying Equations 2
and 5 so that 1 is interacted with a vector of region dummies. In the top panel of
Figure X, we plot 1f ood expenditure against 1income for each region. The strong positive
correlation verifies that locations with larger typhoon-induced income losses are also
the regions with larger reductions in typhoon-induced food purchases, one of the most
important inputs to human capital. In the bottom panel, we plot 1inf ant mortality against
the coefficient for income, finding that the regions with stronger economic responses to
typhoons are also the regions with stronger mortality responses.

Gender Bias A striking feature of the response is that it is completely restricted to

female infants, with no similar response in male infants, as can be seen in Table XII.
Differentially worse health outcomes for female children in times of economic duress are
a common result in the development literature, see e.g., Rose (1999) for a specific case
and Duflo (2005) for an overview. This pattern is generally thought to arise because
parents give less weight to girls outcomes when making decisions about intrahousehold
resource allocations. Senauer, Garcia, and Jacinto (1988) provide supportive evidence
that this dismal situation applies to the Philippines as well. Thus, our finding that
typhoon-induced infant mortality is a strictly female phenomenon is consistent with
previous work on the within-household allocation of resources following income shocks.
We note that it is possible that the gender differential in mortality could be partly

39

driven, at least shortly after impact, by the commonly documented tendency of males
to die in utero at higher rates than females (Almond and Currie (2011); Sanders and
Stoecker (2011)). We examine this claim in the Supplementary data and note that while
it may be occurring in our data, it cannot explain more than a portion of unobserved
male deaths in the first year.
Resource Competition Among Siblings If the female infant mortality that we
observe occurs because of disinvestment in female children, it is plausible that this disinvestment will be larger if the female infant faces greater competition for resources via
older siblings, particularly older brothers30 . We look for evidence that female children
who must compete for resources with other children are more likely to die in the year
following a typhoon by estimating the mortality response of four subsamples of female
infants: those who are the first born to their mother, those who have only older sisters,
those who have only older brothers, and those who have both older sisters and brothers. Table XIV presents our results. We find that mortality among first born females
is moderate, but it doubles when infants have older sisters and nearly doubles again if
there are any older brothers. We interpret these findings as strong evidence that female
infant mortality is driven by resource scarcity within households and not by physical
exposure to typhoons themselves.
Economic Deaths Exceed Exposure Deaths
Our finding that infant mortality (1) mirrors the structure of economic losses in time,
space, subpopulations, income/wealth, and storm intensity combined with our finding
that it is (2) gendered, (3) enhanced by sibling competition, and (4) apparent in children that were not conceived at the time of exposure, strongly suggests that these infant
deaths are caused by economic conditions that deteriorate in the wake of typhoons and
30

See, for example, Butcher and Case (1994).

40

not by physical exposure to the typhoons themselves. We thus term the lagged mortality we observe in our data economic deaths to distinguish it from the exposure
deaths resulting from direct physical exposure to the storm itself, e.g., via drowning or
blunt injuries. We estimate that in the average year, the prior years typhoon climate
causes 1,130 additional female infant deaths in every one-million households, roughly
55% of female infant mortality in our sample of non-migrant mothers. These economic
deaths among female infants constitutes 13% of the overall infant mortality rate in the
Philippines. A back-of-the-envelope calculation31 suggests that across the entire country this amounts to approximately 11,261 economic deaths annually. This number
exceeds 721, the annual average32 number of exposure deaths reported by EM-DAT
across the entire population, by more than a factor of 15. These findings indicate that
the bulk of the Filipino mortality from typhoons does not result from physical exposure to the storm. Rather, 94% of mortality occurs after the storm has passed and is
attributable to the deterioration of economic conditions and subsequent disinvestment
in health and human capital. To date, the authors know of no previous study that
attributes these economic deaths to typhoon exposure.

5.5

Evidence of Adaptation to Typhoon Climates

Households should suffer positive typhoon loses only if adaptation to their typhoon
climate is costly. Here, we briefly examine whether there is evidence of adaptation to
typhoons using cross-sectional variation in typhoon climates (recall Figures I and III).
31

We observe a death rate per woman-household of 1,130 deaths per million; 44.8% of
women in the sample are non-migrants; and there were 22.3 million women aged 15-49 (the
DHS age range) in 2007 according to the Philippine National Statistics Office as detailed at
http://www.census.gov.ph/data/pressrelease/2010/pr10162tx.html.
32
Between 1985-2006.

41

Optimal Adaptation in Theory

We imagine that households can exert costly adaptive effort e to reduce their losses if a
typhoon strikes33 . If the cost function over e is convex, then households will exert adaptive efforts only until their marginal costs of effort equal the expected marginal benefits.
Because adaptive efforts only provide benefits when a typhoon actually strikes, locations that have more frequent or more intense typhoons should have greater returns to
adaptation. Thus, theory predicts that households located in relatively intense typhoon
climates will invest more in costly adaptation, reducing their marginal losses when a
typhoon actually strikes. Denoting a households optimal level of adaptive effort e , we
expect
e
>0
W

(6)

is expected typhoon wind exposure, a summary statistic for a locations

where W
typhoon-climate. Unfortunately, we cannot directly observe whether this is true because
we do not observe e . However, increasing effort reduces marginal losses (Y /W ) to
a fixed level of actual typhoon exposure

Y (e)
< 0.
e W

(7)

This enables us to infer that adaptation is occurring if we see that marginal losses
decline as climates intensify. Assuming households optimize, we multiply equations 6
and 7 to obtain
Y (e )

<0
W W
33

For example, households could reinforce the walls of their home.

42

(8)

a result that we now investigate empirically. For a more complete treatment of optimal
adaptation to tropical cyclone climates, as well as empirical evidence from around the
world, we refer readers to Hsiang and Narita (2012).

Cross-Sectional Evidence of Adaptation

We test Equation 8 by examining whether typhoon-induced losses vary with the typhoon climatologies of different Philippine regions. In the top left panel of Figure XI,
we plot the negative semi-elasticity of income (Y /W ) for each region against its
). Consistent with Equation 8, the marginal effect of tyaverage typhoon exposure (W
phoon exposure declines with increasingly intense typhoon climates. This suggests that
populations do invest adaptive effort in response to their typhoon climate. However, we
note that all regions have positive marginal losses, indicating that no region undertakes
complete adaptation by driving their marginal damages to zero. In the lower left
panel we provide suggestive evidence that this adaptive response can also be observed
in female infant mortality responses.
Two points regarding the left panels of Figure XI are worth noting. First, the average
losses due to cyclones remain high even for regions that exhibit high levels of adaptation.
This occurs because average exposure necessarily increases with the climatological wind
speed, so more intense average exposure counteracts falling marginal losses. In the right
panels of Figure XI we plot estimates for average annual losses34 and find that average
total losses are large across all climatologies35 . The second point of note is that the
slope of the OLS fit in the upper left panel, representing the response of adaptive effort
to climatological conditions, is 0.04. This implies that marginal income losses decline
by roughly 2.8% with each one meter per second increase in climatological wind speed.
34

We compute a regions average annual losses to be that regions marginal loss times its average
.
exposure level: (Y /W ) W
35
We fit a quadratic curve to both panels because total average cost should be quadratic if the
response in the left panels are linear.

43

This number, estimated using only the within-Philippines cross-section, almost exactly
matches Hsiang and Naritas (2012) earlier estimate (3%) which used the cross section
of all countries in the world36 .

Summary and Discussion

We have shown that the typhoon climate of the Western Pacific imposes major economic
and human costs on Filipino households. We observe this impact directly in the form
of lost physical assets; measure its economic effect of depressing household income
and reducing consumption and human capital investments; and lastly show evidence
that these disinvestments have irreversible consequences, which we demonstrate by
examining infant mortality. We discuss some implications of these results and policy
options below.

The Magnitude of Typhoon Losses

In Table XV we summarize our estimates for the average effects of the prior years
typhoon exposure on an average household in an average year. On average, typhoons
reduce household income by 6.6% and household expenditures by 7.1%. Total food expenditure declines similarly (5.9%) but larger declines appear in critical human capital
investments such as meat (12.5%), education (13.3%) and medical (14.3%) expenditures. Typhoons, and the disinvestments in human capital which they cause, elevate
female infant mortality by 18.1 deaths per 1,000 live births. In aggregate, this corresponds to roughly 11,300 additional deaths per year, constituting 13% of the overall
36

Using the same measure of cyclones exposure (spatially averaged maximum wind speed), Hsiang
and Narita (2012) found that increasing average exposure by 1 m/s led to adaptive adjustments which
reduced marginal damages by approximately 3% of their baseline value (when average exposure was 0
m/s). Our results, presented in Figure XI, indicate that increasing average exposure by 1 m/s reduces
marginal income losses by roughly 2.8% of the analogous baseline value (marginal income losses are
0.0143 log points per m/s when average exposure is set to 0 m/s in the upper left panel of Figure XI).

44

infant mortality rate of the Philippines.

In Figure XII we compare these effects with officially reported losses to typhoons
(same data as Figure V), losses which are almost exclusively incurred during physical
exposure to the storm. We estimate average annual damages due to tropical cyclones
as reported in the EM-DAT file to be 0.39% of average real Filipino GDP over the
years 1979-2000; in contrast, we estimate that typhoons cause the average households
income to decline 6.6% in the following year, a loss that is 16.9 times larger. (To
place these numbers in context, the Philippines National Statistical Office estimates
that the average familys savings rate was 15.0% in 200937 indicating that these
losses represent a substantial portion of a households budget.) Similarly, we find that
annual exposure deaths due to typhoons during 1979-2008 average 743; in contrast,
our analysis indicates that 11,261 economic deaths among female infants occur in
the year following typhoon exposure, a loss that is 15.1 times as large. Together these
results suggest that the majority of disaster costs may be incurred well after the disaster
has past.
These results are large, and it is important to be clear when interpreting them. We
calculate these losses based on mean typhoon incidence in the average province/region;
so we interpret them as capturing the expected difference in outcomes between a
typhoon-free year and a year experiencing average typhoon exposure (16.9 m/s). They
can thus be thought of as mean losses conditional on the Philippines having the typhoon
climate that it has; any major shift in that typhoon climate would necessarily lead to a
host of adaptive responses that are impossible to estimate given lack of an observable
counterfactual.
37

Calculated using an average family income of 206,000 pesos and average savings of 31,000 pesos for
2009. Data available at http://www.nscb.gov.ph/secstat/d income.asp; see table Total Number
of Families, Family Income and Family Expenditure and Gini Coefficient: 2009 and 2006.

45

Implications for Economic Development

Our results indicate that typhoons destroy existing capital and reduce investments in
new capital. Both of these effects are a concern for economic development. As discussed
in Hsiang (2010), Dell, Jones, and Olken (2012), and Pindyck (2011), climatic conditions that interfere with capital accumulation and economic growth are particularly
pernicious because their effects are compounded over time. The repeated exposure of
populations to tropical cyclones, both in the Philippines and elsewhere, probably slows
the accumulation of capital stocks at the household level. Unfortunately, such long run
effects are difficult to identify empirically because cross-sectional variations in cycloneclimates are correlated with unobservable omitted variables; hopefully, future research
will address this challenge. Nonetheless, given our evidence, tropical cyclones should
be added to the list of geographically-varying factors which may be contributing to
spatial patterns in global economic development (Gallup, Sachs and Mellinger (1999);
Nordhaus (2006b)).

Implications for Climate Change

As previously discussed and shown in Figure I, the Philippines has one of the most
active typhoon climatologies in the world. The frequency with which typhoons impact
the Philippines suggests that households must understand and incorporate typhoon risk
into their economic decisions (Mendelsohn (2000); Hsiang and Narita (2012)). Thus,
we interpret our estimates as conditional on households having already exploited the
full range of adaptive behaviors available to them. This assumption is supported by
our cross-sectional evidence that levels of adaptation vary across typhoon-climates, results that match the cross-country findings of Hsiang and Narita (2012) with striking
precision. The fact that we continue to observe large typhoon impacts in one of the

46

worlds most intense typhoon climates, where populations have already adapted optimally, suggests that adaptation costs are so high that populations find they are better
off suffering typhoon losses rather than investing in additional adaption. This has
unsettling implications for future climate change policy.
In the design of climate change policy, adaptation to climatic changes is viewed as
a substitute for efforts to mitigate climate changes directly (Stern (2006); Nordhaus
(2008); de Bruin, Dellink and Tol (2009); Aldy et al. (2010); Patt et al. (2010)). If
adaptation is generally inexpensive compared to mitigation, then the cost-minimizing
strategy is to not invest heavily in mitigation and instead to rely primarily on adaptation. However, if adaptation is very costly, then mitigation should be utilized more
vigorously. Our findings suggest that adaptation to tropical cyclones is extremely costly;
thus policies cannot assume that adaption to changes in the future cyclone climate38
will be cheap. This should increase the estimated social cost of greenhouse gas emissions, and concomitantly the value of mitigating these emissions. We may speculate
that technological advances will reduce the future cost of adaptation, but until further
evidence is martialled this remains an assumption. Moreover, if adaptation costs to
tropical cyclones are representative of adaptation costs to a broader class of climatological phenomena, this would suggest that current models of future adaptation are too
optimistic (de Bruin, Dellink and Tol (2009)).
38

Knutson et al. (2010), a recent review of this topic, conclude

[F]uture projections based on theory and high-resolution dynamical models consistently
indicate that greenhouse warming will cause the globally averaged intensity of tropical cyclones to shift towards stronger storms, with intensity increases of 211% by 2100. Existing
modeling studies also consistently project decreases in the globally averaged frequency
of tropical cyclones, by 634%. Balanced against this, higher resolution modelling studies
typically project substantial increases in the frequency of the most intense cyclones, and
increases of the order of 20% in the precipitation rate within 100 km of the storm centre.
(p. 157)

Thus the entire distribution of tropical cyclone events is expected to shift on average, with fewer low
intensity storms but more frequent high intensity storms. However, there remains extensive uncertainty
and the relationship between tropical cylcones and warming is an area of active research.

47

Policy Options in the Current Climate

Setting aside issues surrounding the future climate, it is important to note that there are
a variety of targeted policies that might increase the welfare of Filipino households, or
other typhoon-afflicted populations, in the current climate. At present, large-scale postdisaster management is almost entirely an ad hoc process that is strongly influenced
by political concerns and the media (Besley and Burgess (2002); Garrett and Sobel
(2003); Eisensee and Stromberg (2007); Yang (2008); Kunreuther et al. (2009); United
Nations (2009)). However, our findings provide insight into systematic policies that
could address typhoon-induced welfare loss.

Insurance Social insurance allows Filipino households to smooth their consumption

over some, but not all, income shocks (Fafchamps (2003), Yang and Choi (2007)). Our
observation that consumption responds strongly to typhoons, reflecting income changes,
indicates that current insurance networks are not well-diversified against these events.
Perhaps this occurs because typhoons are large with respect to insurance networks, a
fact that would reduce the idiosyncratic component of the income shock (Townsend
(1995)). Expanding insurance networks over larger spatial scales should reduce the
uninsurable aggregate component of typhoon shocks; however this must be done carefully as even wealthy countries have struggled to sustainably insure tropical cyclone
risk (Kunreuther et al. (2009)).

Credit Without looking specifically at credit markets, we cannot say exactly how
they behave in the wake of Typhoons. However, we observe that income and consumption in low income households recover more slowly than in high income households.
It is plausible that this differentially slow recovery persists because poor households
are credit constrained, preventing them from efficiently rebuilding their capital stocks

48

(Duflo and Banerjee (2007); Noy (2009)). If this is true, subsidizing the development
of credit markets for low-income households may increase their resilience.

Information It seems unlikely that the households in which female infants die are
intentionally allowing these infants to perish. It is more plausible that parents believe
their newborn can cope with higher-than-average levels of neglect, and that there will
be limited permanent damage (Duflo and Banerjee (2011)). Unfortunately, for a small
number of unlucky families, this assumption proves false. It may be the case that simply
educating parents about the risks of post-typhoon neglect will be enough to mitigate a
large portion of typhoons effect on infant mortality.

Targeted Subsidies If household decisions were made to perfectly optimize household welfare, than post-disaster economic decisions would be efficient. Unfortunately,
it seems that childrens long-term welfare, which depends in part on their human capital, is differentially neglected in comparison to short-term consumption of goods like
recreation, tobacco and alcohol. In this situation, it may be optimal to tax adults
to finance human capital subsidies that specifically target children. To avoid political
manipulation of these subsidies, it might be possible that they be indexed to verifiable
measures of typhoon exposure (Hellmuth (2009)).

Technology Standards Because it is difficult for consumers to verify the quality of

infrastructure, construction quality may not be properly priced into markets (Olken
(2007)). This could introduce additional uncertainty into households calculation of the
economic risk they bear in a particular typhoon-climate. To correct for this market
imperfection, it may be optimal for governments to enforce building codes or other
technology standards that mandate a specified level of robustness to typhoon exposure.

49

Research and Develop Adaptation Technologies We find a continuous gradient

in levels of adaptation that reflects the gradient in cyclone risk. This suggests that
adaptation technologies are effective, however the cost of adopting additional adaptation technology is binding throughout the Philippines (Hsiang and Narita (2012)).
Thus, research and development that raises the effectiveness or reduces the cost of adaptive technologies should induce households to employ greater levels of self-protection.

UNIVERSITY OF SAN FRANCISCO

PRINCETON UNIVERSITY AND UNIVERSITY OF CALIFORNIA, BERKELEY

50

References
Acemoglu, Daron, Simon Johnson, and James A. Robinson, Reversal Of Fortune: Geography And Institutions In The Making Of The Modern World Income Distribution,
The Quarterly Journal of Economics, 117 (2002), 12311294.
Albala-Bertrand, J. M., The Political Economy of Large Natural Disasters (Clarendon
Press, 1993).
Aldy, Joseph E., Alan J. Krupnick, Richard G. Newell, Ian W. H. Parry, and William A.
Pizer, Designing Climate Mitigation Policy, Journal of Economic Literature, 48
(2010), 903934.
Almond, Douglas and Janet Currie, Killing Me Softly: The Fetal Origins Hypothesis,
Journal of Economic Perspectives, 25 (2011).
Auffhammer, Maximilian, Solomon Hsiang, Wolfram Schlenker, and Adam Sobel, Global
Climate Models and Climate Data: a User Guide for Economists, Working paper ,
(2010).
Banerjee, Abhijit and Sendhil Mullainathan, The Shape of Temptation: Implications for
the Economic Lives of the Poor, Working Paper 15973, NBER, 2010.
Benjamin, Marito Garcia Senauer and Elizabeth Jacinto, Determinants of the Intrahousehold Allocation of Food in the Rural Philippines, American Journal of Agricultural
Economics, 70 (1988), 170180.
Benson, C. and E. Clay, Understanding the Economic and Financial Impacts of Natural
Disasters, Manuscript, 2004.
Bertrand, Marianne, Esther Duflo, and Sendhil Mullainathan, How Much Should We
Trust Differences-in-Differences Estimates?, The Quarterly Journal of Economics,
119 (2004), 249275.
Besley, Timothy and Robin Burgess, The Political Economy of Government Responsiveness: Theory and Evidence from India, The Quarterly Journal of Economics, 117
(2002), 14151451.
Bevere, Lucia, Brian Rogers, and Balz Grollimund, Sigma - Natural catastrophes and manmade disasters in 2010: a year of devastating and costly events (Swiss Reinsurance
Company Ltd., 2011).
Bhalotra, Sonia, Fatal fluctuations? Cyclicality in infant mortality in India, Journal
of Development Economics, 93 (2010), 719.
and Christopher Heady, Child Farm Labor: The Wealth Paradox, World Bank
Economic Review , 17 (2003), 197227.
51

Bloom, David E., David Canning, and Jaypee Sevilla, Geography and Poverty Traps,
Journal of Economic Growth, 8 (2003), 355378.
Butcher, Kristin F. and Anne Case, The Effect of Sibling Sex Composition on Womens
Education and Earnings, Quarterly Journal of Economics, 109 (1994), 531563.
Carmargo, Suzana J. and Adam H. Sobel, Western North Pacific Tropical Cyclone Intensity and ENSO, Journal of Climate, (2005).
Caselli, Francesco and Massimo Morelli, Bad politicians, Journal of Public Economics,
88 (2004), 759782.
Cavallo, Eduardo and Ilan Noy, The Economics of Natural Disasters: A Survey, RES
Working Papers 4649, Inter-American Development Bank, Research Department,
2009.
Chakravarty, Abhishek, Dams and infant mortality in Africa, Working Paper, University College London, 2011.
Chay, Kenneth Y. and Michael Greenstone, The Impact Of Air Pollution On Infant
Mortality: Evidence From Geographic Variation In Pollution Shocks Induced By A
Recession, The Quarterly Journal of Economics, 118 (2003), 11211167.
Chen, Joyce, Dads, disease, and death: determinants of daughter discrimination,
Journal of Population Economics, 24 (2011).
Cochrane, John H., A Simple Test of Consumption Insurance, Journal of Political
Economy, 99 (1991), 957976.
Conley, Timothy, GMM Estimation with Cross Sectional Dependence, Journal of
Econometrics, 92 (1999), 145.
, Spatial Econometrics (2008).
Davis, Donal R. and David E. Weinstein, Bones, Bombs, and Break Points: The Geography of Economic Activity, American Economic Review , 92 (2002), 12691289.
de Bruin, Kelly C., Rob B. Dellink, and Richard S. J. Tol, AD-DICE: an implementation
of adaptation in the DICE model, Climatic Change, 95 (2009), 6381.
Dell, Melissa, Benjamin F. Jones, and Benjamin A. Olken, Temperature and Income:
Reconciling New Cross-Sectional and Panel Estimates, American Economic Review:
Papers and Proceedings, 99 (2009), 198204.
, , and , Climate Change and Economic Growth: Evidence from the Last Half
Century, American Economic Journal: Macroeconomics, 4 (2012), 6695.

52

Dercon, S. and I. Outes, Income Dynamics in Rural India: Testing for Poverty Traps
and Multiple Equilibria, Manuscript, 2009.
Deryugina, Tatyana, The Dynamic Effects of Hurricanes in the US: The Role of NonDisaster Transfer Payments, Job Market Paper, M.I.T., 2011.
Deschenes, Olivier and Enrico Moretti, Extreme weather events, mortality and migration, The Review of Economics and Statistics, 91 (2009), 659681.
, Robin Burgess, David Donaldson, and Michael Greenstone, Climate Change and
Mortality in India, Mimeo, (2011).
Duflo, Esther, Child Health and Household Resources in South Africa: Evidence from
the Old Age Program, American Economic Review , 90 (2000), 393398.
, Gender Equality in Development, Working Paper, Abdul Latif Jameel Professor of Poverty Alleviation and Development Economics, Massachusetts Institute of
Technolog, 2005.
and Abhijit Banerjee, The Economic Lives of the Poor, The Journal of Economic
Perspectives, 21 (2007), 141168.
and

, Poor Economics (PublicAffairs, 2011).

E., E. N. Rappaport Willoughby H. and F. D. Marks, Hurricane Forecasting: The State

of the Art, Natural Hazards Review , 8 (2007), 4550.
Easterly, W. and R. Levine, Tropics, germs and crops: how endowments influence economic development, J. of Monetary Econ., 50 (2003), 339.
Eisensee, Thomas and David Stromberg, News Droughts, News Floods, and U.S. Disaster Relief, The Quarterly Journal of Economics, 122 (2007), 693728.
Ericta, Carmelita and Emma Fabian, A Documentation of the Philippines Family Income and Expenditure Survey, Discussion Paper 2009-18, Philippine Institute for
Development Studies, 2009.
Fafchamps, Marcel and Hwa Jung Choi, Risk-sharing networks in rural Philippines,
Journal of Development Economics, 71 (2003), 261287.
Fomby, T., Y. Ikeda, and N. Loayza, The Growth Aftermath of Natural Disasters,
Manuscript, 2009.
Foster, Andrew D. and Mark R. Rosenzweig, Imperfect Commitment, Altruism, and the
Family: Evidence from Transfer Behavior in Low-Income Rural Areas, Review of
Economics and Statistics, 83 (2001), 389407.

53

Freedman, David, Statistical Models and Shoe Leather, Sociological Methodology, 21

(1991), 291313.
Gallup, J. L., J. D. Sachs, and A. D. Mellinger, Geography and Economic Development,
Consulting Assistance on Economic Reform II, Discussion Papers, (1999).
Garrett, Thomas A and Russell S Sobel, The Political Economy of FEMA Disaster
Payments, Economic Inquiry, 41 (2003), 496509.
Greene, William H., Econometric Analysis, Fifth Edition (Prentice Hall, 2003).
Hallegatte, S. and M. Ghil, Natural Disasters Impacting a Macroeconomic Model with
Endogenous Dynamics, Ecological Economics, 68 (2008), 58292.
Hellmuth, Molly E., Daniel E. Osgood, Ulrich Hess, Anne Moorhead, and Haresh Bhojwani,
Index insurance and climate risk: Prospects for development and disaster management (International Research Institute for Climate and Society, 2009).
Heming, Julian and Jim Goerss, Track and Structure Forecasts of Tropical Cyclones
(2010).
Hochrainer, S., Assessing Macro-economic Impacts of Natural Disasters: Are There
Any?, Manuscript, 2009.
Holland, Paul W., Statistics and Causal Inference, Journal of the American Statistical
Association, 81 (1986), 945960.
Horton, Susan, Child Nutrition and Family Size in the Philippines, Journal of Development Economics, 23 (1986), 161176.
Hsiang, Solomon M., Temperatures and Cyclones strongly associated with economic
production in the Caribbean and Central America, Proceedings of the National
Academy of Sciences, 107 (2010), 1536715372.
and Daiju Narita, Adaptation to Cyclone Risk: Evidence from the Global CrossSection, Climate Change Economics, 3 (2012).
, Kyle C. Meng, and Mark A. Cane, Civil conflicts are associated with the global
climate, Nature, 476 (2011), 438441.
IPCC, Climate Change 2007 - The Physical Science Basis (Cambridge University Press,
2007).
ISDR, Global Assessment Report on Disaster Risk Reduction (United Nations, 2009).
Jacoby, Hanan G. and Emmanuel Skoufias, Risk, Financial Markets, and Human Capital
in a Developing Country, Review of Economic Studies, 64 (1997), 311335.

54

and , Testing Theories of Consumption Behavior Using Information on Aggregate

Shocks: Income Seasonality and Rainfall in Rural India, American Journal of
Agricultural Economics, 80 (1998), 114.
Jensen, Robert, Agricultural Volatility and Investment in Children, American Economic Review , 90 (2000), 399404.
Jensen, Robert T. and Miller Nolan H, Giffen Behavior and Subsistence Consumption,
American Economic Review , 98 (2008), 15531577.
Kahn, Matthew E., The Death Toll from Natural Disasters: The Role of Income,
Geography, and Institutions, The Review of Economics and Statistics, 87 (2005),
271284.
Kalnay, E., M. Kanamitsu, R. Kistler, W. Collins, D. Deaven, L. Gandin, M. Iredell, S. Saha,
G. White, J. Woollen, Y. Zhu, A. Leetmaa, B. Reynolds, M. Chelliah, W. Ebisuzaki,
W. Higgins, J. Janowiak, K. C. Mo, C. Ropelewski, J. Wang, R. Jenne, and D. Joseph,
The NCEP/NCAR 40-Year Reanalysis Project, Bulletin of the American Meterological Society, 77 (1996), 437471.
Kazarosian, Mark, Precautionary Savings A Panel Study, The Review of Economics
and Statistics, 97 (1997), 241247.
Knapp, Kenneth R., Archive Compiles New Resource for Global Tropical Cyclone Research, Eos, 90 (2009), 4647.
Knutson, Thomas R., John l. Mcbride, Johnny Chan, Kerry Emanuel, Greg Holland, Chris
Landsea, Isaac Held, James P. Kossin, A. K. Srivastava, and Masato Sugi, Tropical
Cyclones and Climate Change, Nature Geoscience, 3 (2010), 157163.
Kochar, Anjini, Explaining Household Vulnerability to Idiosyncratic Income Shocks,
American Economic Review , 85 (1995), 159164.
, Smoothing Consumption by Smoothing Income: Hours-of-Work Re- sponses to
Idiosyncratic Agricultural Shocks in Rural India., Review fo Economics and Statistics, 81 (1999), 5061.
Kudamatsu, Masayuki, Has Democratization Reduced Infant Mortality in Sub-Saharan
Africa? Evidence from Micro Data, Journal of the European Economic Association,
forthcoming (2011).
Kunreuther, Howard C., Erwann O. Michel-Kerjan, Neil A. Doherty, Martin F. Grace,
Robert W. Clein, and Mark V. Pauly, At War With the Weather (MIT Press, 2009).
Loayza, Norman, Eduardo Olaberria, Jamele Rigolini, and Luc Christiaensen, Natural
Disasters and Growth: Going beyond the Averages, Policy Research Working Paper
4980, The World Bank, 2009.
55

Maccini, Sharon and Dean Yang, Under the Weather: Health, Schooling, and Economic
Consequences of Early-Life Rainfall, American Economic Review , 99 (2009), 1006
1026.
Masayuki, Torsten Persson Kudamatsu and David Stromberg, Weather and Infant Mortality in Africa, in review, 2011.
Mendelsohn, Robert, Efficient Adaptation to Climate Change, Climatic Change, 45
(2000), 583600.
Miguel, Edward and Gerard Roland, The Long Run Impact of Bombing Vietnam, Journal of Development Economics, 96 (2011), 115.
, Shanker Satyanath, and Ernest Sergenti, Economic Shocks and Civil Conflict: An
Instrumental Variables Approach, Journal of Political Economy, 112 (2004), 725
753.
Mincer, Jacob, Investment in human capital and personal income distribution, The
Journal of Political Economy, 66 (1958), 281302.
Morduch, Jonathan, Income Smoothing and Consumption Smoothing, The Journal
of Economic Perspectives, 9 (1995), 103114.
Mutter, J. C., The Earth Sciences, Human Well-Being, and the Reduction of Global
Poverty, Eos, 86 (2005).
Narita, Daiju, R. S. J. Tol, and D. Anthoff, Damage costs of climate change through
intensification of tropical cyclone activities: an application of FUND, Climate Research, 39 (2009), 8797.
Natural Hazards, UnNatural Disasters
Natural Hazards, UnNatural Disasters, (The World Bank, 2010).
Nordhaus, William D., The Economics of Hurricanes in the United States, Annual
Meetings of the American Economic Association, (2006).
, Geography and macroeconomics: New Data and new findings, Proceedings of
the National Academy of Sciences, 103 (2006).
, A Question of Balance: Weighing the Options on Global Warming Policies (Yale
University Press, 2008).
, The Economics of Hurricanes and Implications of Global Warming,
Change Economics, 1 (2010), 120.

Climate

Noy, Ilan, The macroeconomic consequences of disasters, Journal of Development

Economics, 88 (2009), 221231.
56

OFDA/CRED, The International Disaster Database (2009).

Olken, Benjamin, Monitoring Corruption: Evidence from a Field Experiment in Indonesia, Journal of Political Economy, 115 (2007), 200249.
Patt, Anthony G., Detlef P. van Vuuren, Frans Berkhout, Asbjorn Aaheim, Andries F. Hof,
Morna Isaac, and Reinhard Mechler, Adaptation in integrated assessment modeling:
where do we stand?, Climatic Change, 99 (2010), 383402.
Paxson, Christina H., Using Weather Variability To Estimate the Response of Savings
to Transitory Income in Thailand, American Economic Review, 82 (1992), 1533.
Pielke, R. A. Jr., J. Rubiera, C. Landsea, M. Fernandez, and R. Klein, Hurricane Vulnerability in Latin America and the Caribbean: Normalized Damage and Loss Potentials,
Natural Hazards Rev., (2003).
Pindyck, Robert S., The Economics of Climate Change: Adaptations Past and Present
(2011).
Qian, Nancy, Missing Women and the Price of Tea in China: The Effect of Sex-Specific
Income on Sex Imbalance, The Quarterly Journal of Economics, 123 (2008), 1251
1285.
Robert, Kerry Emanuel Shona Chonabayashi Mendelsohn and Kaura Bakkensen, The
Impact of Climate Change on Global Tropical Cyclone Damage, Nature Climate
Change, 2 (2012), 205209.
Robinson, Jonathan and Ethan Yeh, Transactional Sex as a Response to Risk in Western
Kenya, American Economic Journal: Applied Economics, 3 (2011), 3564.
Rose, Elaina, Consumption Smoothing and Excess Female Mortality in Rural India,
Review of Economics and Statistics, 81 (1999), 4149.
Rosenzweig, Mark R., Risk, Private Information, and the Family,
Economic Review, 78 (1988), 245250.

The American

and Kenneth I. Wolpin, Credit Market Constraints, Consumption Smoothing, and the
Accumulation of Durable Production Assets in Low-Income Countries: Investments
in Bullocks in India, Journal of Political Economy, 101 (1993), 223244.
and T. Paul Schultz, Market Opportunities, Genetic Endowments, and Intrafamily
Resource Distribution: Child Survival in Rural India, The American Economic
Review, 72 (1982), 803815.
Sanders, Nicholas J. and Charles F. Stoecker, Where Have All the Young Men Gone?
Using Gender Ratios to Measure Fetal Death Rates, Working Paper 17434, NBER,
2011.
57

Schlenker, Wolfram and Michael Roberts, Nonlinear temperature effects indicate severe
damages to U.S. crop yields under climate change, Proceedings of the National
Academy of Sciences, 106 (2009), 1559415598.
Sen, Amartya, More than 100 million women are missing, The New York Review of
Books, (1990), 6166.
Simeonova, Emilia, Out of Sight, Out of Mind? The Impact of Natural Disasters on
Pregnancy Outcomes, Working Paper 2814, CESIFO, 2011.
Smith, Doug M., Rosie Eade, Nick J. Dunstone, David Fereday, James M. Murphy, Holger
Pohlmann, and Adam A. Scaife, Skilful multi-year predictions of Atlantic hurricane
frequency, Nature Geoscience, 3 (2010), 846849.
Stern, Nicholas, Stern Review: The Economics of Climate Change (Cambridge University Press, 2006).
Strauss, John and Duncan Thomas, Health, Nutrition, and Economic Development,
Journal of Economic Literature, 36 (1998), 766817.
Strobl, Eric, The Economic Growth Impact of Hurricanes: Evidence from U.S. Coastal
Counties, The Review of Economics and Statistics, 93 (2011), 575589.
Stromberg, David, Natural Disasters, Economic Development, and Humanitarian Aid,
Journal of Economic Perspectives, 21 (2007), 199222.
Subramanian, Shankar and Angus Deaton, The Demand for Food and Calories, Journal
of Political Economy, 104 (1996), 133162.
Tol, Richard S. J., The Economic Effects of Climate Change, Journal of Economic
Perspectives, 23 (2009), 2951.
Townsend, Robert M., Consumption Insurance: An Evaluation of Risk-Bearing Systems
in Low-Income Economies, The Journal of Economic Perspectives, 9 (1995), 83
102.
Toya, Hideki and Mark Skidmore, Economic development and the impacts of natural
disasters, Economics Letters, 94 (2007), 2025.
Triep, Enrique G. and Lynn R. Sykes, Frequency of occurrence of moderate to great earthquakes in intracontinental regions, Journal of Geophysical Research, 102 (1997),
99239948.
Udry, Christopher, Risk and Insurance in a Rural Credit Market: An Empirical Investigation in Northern Nigeria, Review of Economic Studies, 61 (1994), 495526.
Vigdor, Jacob, The economic aftermath of Hurricane Katrina, Journal of Economic
Perspectives, 22 (2008), 135154.
58

Weitzman, Martin L., On modeling and interpreting the economics of catastrophic climate change, The Review of Economics and Statistics, 91 (2009), 119.
Xie, P. and P. A. Arkin, Analyses of Global Monthly Precipitation Using Gauge Observations, Satellite Estimates, and Numerical Model Predictions, Journal of Climate,
9 (1996), 840858.
Yang, Dean, Coping with Disaster: The Impact of Hurricanes on International Financial Flows, 1970-2002, Advances in Economic Analysis & Policy, 8 (2008),
19031903.
and Hwa Jung Choi, Are Remittances Insurance? Evidence from Rainfall Shocks in
the Philippines, World Bank Economic Review, 21 (2007), 219248.
Zivin, Joshua Graff and Matthew J. Neidell, Temperature and the allocation of time:
Implications for climate change, Manuscript 15717, NBER, 2010.

59

Figure I
Annual maximum wind speed (in meters per second) averaged over 1950-2008. Data: LICRICE.

61

100

kilometers North

200
0
100
200
300
300

200

100

100

kilometers East

200

300

1989

1987

10
20
30
40
50
60
Maximum typhoon wind speed at surface

1990

1988

>70

Figure II
Example snapshot of the parametrized typhoon wind field used by LICRICE to reconstruct exposure at the surface (left)
and examples of reconstructed surface-level wind exposure for four consecutive years in the sample. Figures A1 and A2 in
the Supplementary data show surface-level data for all years in the study (1979-2008).

0
300

10

20

30

40

50

Maximum azimuthal wind speed = 37.9 m/s

Eastward translational velocity = 1 m/s
Northward translational velocity = 5 m/s

Total wind speed at surface

(meters per second)

62

40

Annual area-averaged maximum typhoon wind speeds during 1950-2008 (m/s)

20

60

Figure III
Box-whisker plot of provincial annual maximum wind speed measure, 1950-2008. Provinces (y axis) are ranked by
median annual maximum wind speed (white dots). Shaded boxes show interquartile range of wind speed distribution;
grey dots show outlier years. Data: LICRICE.

Provinces of the Philippines ordered by median exposure

Phase 1
interviews

Fraction of strikes (%)

Phase 1 recall period

Phase 2
interviews

Phase 2 recall period

20

10

Distribution of
provinces-level
typhoon strikes
(19792008)

Month of year 0
Figure IV
Timeline of FIES data collection overlaid with a histogram of typhoon events by
month.

63

Reported Deaths
6

deaths (thousands)

hundred-million US$ (1990)

Reported Damages

6
4
2
0

4
2
0

1960

1980

2000

1960

1980

year

year

16

14

Log(deaths)

Log(damages)

2000

12
10
8

6
4
2
0

10

15

20

25

30

10

wind speed (m/s)

15

20

25

30

wind speed (m/s)

Figure V
Country-level losses to typhoons for the Philippines. Reported damages and deaths
from the EM-DAT dataset are shown as time series from 1950-2008. Scatter plots
show log losses as a function of wind speed averaged nationally. Fit is OLS with
confidence interval.

64

Probability lost (%)

Expenditure reduction (%)

20-25

25-30

>35

0-5

0-5

5-10

15-20

20-25

25-30

20-25

25-30

30-35

30-35
Annual maximum wind speed (m/s)

10-15

15-20

Annual maximum wind speed (m/s)

10-15

cereal
meat
fish & marine
fruit
dairy & eggs

5-10

nonagricultural wages
entreprenurial income

>35

>35

500

500

1000

1500

2000

10

12

14

16

18

010

0-5

15-20

20-25

25-30

2025

2530

30-35

3035
Annual maximum wind speed (m/s)

1015

1520

Annual maximum wind speed (m/s)

10-15

infant (female)
infant (all)
any age (all)

5-10

income
expenditure

Figure VI
Household losses as a nonlinear function of previous calendar years typhoon exposure. Losses are on the y-axis, x-axis
shows coefficient on indicator variables for previous years maximum wind speed.

Annual maximum wind speed (m/s)

15-20

10

15

20

25

30

35

40

10-15

>35

5-10

food
medical
fuel
education
transport &
communication

30-35

1015 1520 2025 2530 3035

Annual maximum wind speed (m/s)

510

10

12

14

16

18

0-5

05

10

15

20

25

30

35

40

electricity
closed toilet
strong walls
car
television
refrigerator
Losses (%)
Expenditure reduction (%)

Losses (%)
Deaths (per million HH)

>35

>35

typhoon

Lost household income

0.4
0.3
0.2
0.1
0

10

0.2
0

years after storm season

typhoon

Household expenditure reduction

0.5
0.4

05

maximum wind speed (m/s)

>35

20

0.3
0.2
0.1
0

15
10
5

0.1

0.2
0

05

years after storm season

Female infant mortality

80

Female infant mortality

>35

3000
deaths per million HH

100

510 1015 1520 2025 2530 3035

maximum wind speed (m/s)

typhoon

deaths per million HH (per m/s)

510 1015 1520 2025 2530 3035

Household expenditure reduction

losses (%)

losses (% per m/s)

15

0.1

60
40
20
0
20

20

losses (%)

losses (% per m/s)

0.5

Lost household income

2500
2000
1500
1000
500
0

500

years after storm season

010

1015

1520

2025

2530

3035

>35

maximum wind speed (m/s)

Figure VII
Similarities in the structure of typhoon losses across categories. Dynamic (left column)
estimates of typhoon impacts show responses for 5 years starting in the calendar year
of exposure (year 0). Nonlinear (right column) estimates show coefficients on indicator
variables for previous years (Lag=1) maximum wind speed. Thin lines are 90% CI.
66

Effect in first lag year: income

Six year cumulative effect: income

income loss (% per m/s)

1.2

1.2

0.8

0.8

0.4

0.4

0
20

40
60
income centile

80

20

Effect in first lag year: expenditure

40
60
income centile

80

Six year cumulative effect: expenditure

expenditure reduction (% per m/s)

C
1.2

1.2

0.8

0.8

0.4

0.4

0
20

40
60
80
expenditure centile

20

food expenditure reduction (% per m/s)

Effect in first lag year: food consumption

40
60
80
expenditure centile

Six year cumulative effect: food consumption

1.2

1.2

0.8

0.8

0.4

0.4

0
20

female infant deaths per million HH (per m/s)

40
60
80
food expenditure centile

20

Effect in first lag year: female infant mort.

40
60
80
food expenditure centile

Six year cumulative effect: female infant mort.

800

800

400

400

15

35

55
75
SES centile

95

15

35

55
75
SES centile

95

Figure VIII
The effect of of typhoons on the shape of income, expenditure, and wealth distributions. Left column
shows the effect of the previous years typhoon exposure at different q-quantile cutoffs (A, C, E) and
inferred socioeconomic status (G). Right column shows the cumulative impact over six years.
Whiskers are 2- for each point-estimate. Thick horizontal lines in A, C, E, and G are the main
effect from the baseline model, shaded regions are the 90% CI. Curves in B, D, F, and H are simple
OLS fits to the estimated coefficients. Each coefficient is estimated in a separate province-by-year
panel containing province fixed effects, year fixed effects, controls for temperature and rainfall and a
lagged dependent variable.

68

-12

-6

Month of
typhoon

12

18

24

Births resulting
in mortality

30

Months relative to typhoon strike

9 months

Mortality

36

concieved after typhoon

42

48

Figure IX
Cumulative impact of typhoons on number of infant female deaths (grey line) and number of female births resulting in
infant deaths (dotted black line) per million households. Observation is at the regional aggregate level. Cumulative effect
normalized such that month prior to typhoon impact is zero. Includes region, year, and month fixed effects, lagged
temperature and precipitation controls, and lagged dependent variable.

12

16

exposed
in-utero

Cumulative female infant deaths (per million HH per m/s)

 food exp. / wind speed

(% per m/s)
inf. mort. / wind speed
(deaths per 1M HH per m/s)

SMD

CVS

WMD
ILC
CGYWVS
CLZ
BIC
EVS

SLZ

NMD

(slope = 0.91)

1
2
income / wind speed (% per m/s)

200

SMD

100

BICEVS
CGY

SLZ

CLZWVS
WMD

(slope = 68)

ILC

NMD
CVS

1
2
income / wind speed (% per m/s)

Figure X
Cross-sectional correlation of region-specific coefficients. Regions that suffer relatively
larger income losses (x-axis) also suffer larger reductions in food expenditure (top)
and larger increases in infant mortality (bottom), holding the intensity of typhoon
exposure fixed.

69

70

- income/ wind speed

(% per m/s)

100

200

10

NMD

20

STG
WVS

CVS

EVS
CLZ
BIC

ILC

10

NMD
WVS

CVS

20

STG
CLZ

ILC

EVSBIC

30

CGY

(slope = -2.3)

climatological wind speed (m/s)

WMD

SMD

30

CGY

(slope = -0.04)

climatological wind speed (m/s)

WMD

SMD

-1000

1000

2000

10

20

10

NMD
WVS

20

BIC

EVS
CLZ

ILC

10

NMD

WVS

CVS

20

ILC

CLZ

climatological wind speed (m/s)

WMD

SMD

STG

EVSBIC

climatological wind speed (m/s)

WMD

SMD

CVS STG

30

CGY

30

CGY

Figure XI
Cross-sectional evidence and effect of adaptation to sub-national typhoon climates. Regions with higher mean exposure
(x-axis) generally suffer smaller losses to income (top left) and infant life (bottom left) when the intensity of typhoon
exposure is held fixed. Multiplying these linearly declining marginal effects against climatological exposure results in
average expected losses that are quadratic (top right, bottom right).

inf. mort./ wind speed

(deaths per 1M HH per m/s)

2
Average income lost
(%)
Average infant mortality
(deaths per 1M HH)

Year of typhoon

Year after
(this study)

16

12

4
0.39%

743

6.59%

Thousand deaths (19792008)

Percent of income (19792000)

(EM-DAT estimate)

11,261

0
economic
loss
(national est.)

exposure
mortality
(all ages)

economic
economic
loss
mortality
(avg. household)
(infant)

Average effects in an average year

Figure XII
Comparison of average annual economic and human losses to typhoons in the
Philippines. Year of typhoon estimates show immediate economic damage
normalized by contemporaneous GDP and total mortality as reported by EM-DAT.
Year after estimates show average household income loss and total infant mortality
in the calendar year following a typhoon event. Economic and human losses the year
after a typhoon exceed the immediate losses by factors of 16.9 and 15.1 (resp.).
Whiskers denote 95% confidence intervals.

71

Table I
Typhoon exposure (maximum wind speed) summary statistics.
Unit of observation
Province
Province
Region
Region
Nation
Nation

Years
1950
1979
1950
1979
1950
1979

2008 4838
2008 2460
2008 885
2008 450
2008 59
2008 30

Mean

Std. Dev.

Min

Max

17.6
16.9
16.1
15.3
17.3
16.5

12.0
11.6
11.5
11.0
4.6
4.0

0.0
0.0
0.0
0.0
9.2
9.2

62.1
53.5
47.4
45.9
30.5
23.6

Notes: Maximum wind speed measured in meters per second.

72

Table II
Summary averages for FIES households
VARIABLES
Total number of household members
Number of household members above age 15
Age of household head (yr.)
Household head is male (%)
Household head completed no school (%)
Household head completed primary school (%)
Household head completed secondary school (%)
Total household:
Income (PHP)
Expenditures (PHP)
Food expenditures (PHP)
Education expenditures (PHP)
Medical expenditures (PHP)

Observations:

5.2
[2.27]
4.1
[1.78]
47.6
[14.16]
85.2
[35.5]
5.7
[23.19]
64.0
[48.01]
33.6
[47.2]
127500
[157600]
103700
[106000]
50500
[33100]
4000
[11900]
2200
[11300]
142789

Household has:
Electricity (%)

62.6
[48.4]
72.2
[44.8]
54.5
[49.8]
39.1
[48.8]
6.7
[24.9]

Closed toilet (%)

Strong walls (%)
Television (%)
Car (%)

Observations:

107620

Notes: Standard errors shown in parentheses. Income and expenditures shown in year
2000-equivalent Philippine Pesos.

73

Table III
Summary averages for DHS households
Variable
Age
Married (%)
Wife has no education (%)
Wife has post-secondary education (%)
Husband has no education (%)
Husband has post-secondary education (%)
Wifes total children born
Wifes total sons born
Wifes total daughters born

Observations

28.65
[10.08]
48.32
[49.97]
3.2
[17.6]
28
[44.9]
4.17
[20]
22.5
[41.7]
2.0
[2.57]
1.02
[1.49]
0.95
[1.42]
24841

Notes: Standard errors shown in parentheses. Non-migrant households only.

74

75
142,789
0.0145

Observations
R-squared

142,789
0.02

-0.81
[1.27]
-1.12
[1.60]
-0.68
[1.77]
-0.91
[1.84]
-0.15
[1.87]
0.81
[2.03]
-0.49
[2.09]
142,789
0.01

0.30
[0.53]
0.68
[0.66]
0.50
[0.65]
0.70
[0.77]
0.28
[0.81]
0.44
[1.05]
0.67
[0.98]

(5)

(6)

(7)

142,789
0.01

-0.30
[0.42]
-0.17
[0.65]
0.25
[0.63]
0.76
[0.78]
0.85
[0.81]
0.53
[1.04]
1.55*
[0.90]
142,789
0.08

1.02
[0.81]
0.67
[0.85]
0.83
[0.86]
0.96
[1.02]
0.24
[1.06]
0.97
[1.11]
0.73
[1.16]

142,789
0.14

-3.06**
[1.27]
-2.69*
[1.55]
-3.00*
[1.64]
-3.26
[2.05]
-1.41
[2.20]
-2.14
[2.09]
-1.21
[2.29]

142,789
0.24

-0.45
[0.97]
-0.24
[1.27]
-1.04
[1.39]
0.40
[1.68]
2.64
[1.88]
2.59
[1.82]
3.13
[1.94]

Household head characteristics

Only
Only
Male
primary secondary
(%)
Uneducated school
school

(4)

Notes: Standard errors clustered at the treatment (province) level in brackets. *** p<0.01, ** p<0.05, * p<0.1. Lags 0 and 2-5 estimated but
not shown. Includes province and year fixed effects and lagged temperature and precipitation controls.

35+ m/s

30-35 m/s

25-30 m/s

20-25 m/s

15-20 m/s

10-15 m/s

0.0594
[0.0603]
0.1480*
[0.0781]
0.1205
[0.0864]
0.1968*
[0.1000]
0.0748
[0.1099]
0.0808
[0.1139]
0.1046
[0.1171]

Married
(%)

Single
family
dwelling
(%)

Total
family
size

Max windspeed 5-10 m/s

VARIABLES

(3)

(2)

(1)

Table IV
Checking for balance on typhoon exposure in FIES data

76
44
0.31
-

(3)

44
0.34
Y

37
0.41
-

Damages
Damages / GDP
(log)
(log)
23.65*** 26.34***
[5.05]
[4.89]

(2)

(5)

37
0.48
Y

53
0.21
-

Damages
/ GDP
Killed
(log)
(log)
22.23*** 16.63***
[5.75]
[3.54]

(4)

Notes: Typhoon treatment at national level. White standard errors in brackets. *** p<0.01, ** p<0.05, * p<0.1.

Observations
R-squared
Linear and quad trends

Damages
(log)
Maximum wind speed (m/s) 20.62***
[4.24]

(1)

Table V
Country level typhoon damages.

53
0.35
Y

Killed
(log)
22.39***
[3.65]

(6)

48
0.30
-

Killed
/ pop.
(log)
20.62***
[3.53]

(7)

48
0.34
Y

Killed
/ pop.
(log)
22.00***
[3.84]

(8)

Table VI
Typhoon impact on assets

VARIABLES
Max wind speed, T=0
T+1
T+2
T+3
T+4

Observations
R-squared

(1)

(2)

(3)

(4)

(5)

Has
electricity
(%)

Has
closed
toilet
(%)

Strong
walls
(%)

Has
TV
(%)

Has
car
(%)

-0.21*
[0.11]
-0.14**
[0.06]
0.01
[0.07]
0.04
[0.06]
-0.07
[0.04]

-0.04
[0.10]
-0.16***
[0.03]
-0.09
[0.06]
-0.12
[0.09]
-0.11
[0.06]

-0.08
[0.10]
-0.11**
[0.04]
-0.09
[0.08]
0.12
[0.10]
0.01
[0.05]

-0.17*
[0.09]
-0.12**
[0.05]
0.11*
[0.06]
-0.10*
[0.05]
-0.15***
[0.03]

-0.03
[0.03]
0.01
[0.02]
-0.03
[0.05]
-0.01
[0.03]
0.01
[0.02]

107,620
0.28

107,620
0.21

107,620
0.21

107,620
0.34

107,620
0.07

Notes: Notes: Standard errors clustered at the treatment (province) level in brackets. *** p<0.01, **
p<0.05, * p<0.1. Lags 0-5 estimated but only shown for 0-4. Includes province and year fixed effects,
lagged temperature and precipitation controls, and household controls consisting of the number of members,
working and non, in household as well as household heads gender and education level.

77

78
province
Y
2000
-

Exposure:
Province FE
Region FE
HH controls
Lagged temp, precip
Lagged dependent var.
SE clustered at region
Years 1985 to ...
Conley (spatial) SE

province
Y
Y
2000
-

-0.10
[0.14]
-0.33***
[0.09]
0.01
[0.10]
-0.14
[0.14]
-0.08
[0.09]
142,789
0.38
province
Y
Y
Y
2000
-

0.02
[0.15]
-0.35***
[0.09]
0.00
[0.08]
-0.17
[0.13]
-0.09
[0.11]
142,779
0.57

Income
(%)

(3)

province
Y
Y
Y
Y
2000
-

-0.00
[0.18]
-0.39***
[0.10]
-0.16
[0.12]
0.04
[0.15]
-0.06
[0.10]
142,779
0.57

Income
(%)

(4)

region
Y
Y
Y
Y
2006
-

0.27
[0.16]
-0.58***
[0.14]
-0.17
[0.14]
-0.22
[0.15]
-0.03
[0.09]
174,896
0.55

Income
(%)

(5)

province
Y
Y
Y
2000
Y

-0.01
[0.15]
-0.39***
[0.14]
0.02
[0.14]
0.29
[0.18]
-0.05
[0.11]
367
0.95

Income
(%)

(6)

Notes: Percent change calculated as log points *100 per m/s of max wind speed. Standard errors in brackets. *** p<0.01, ** p<0.05,
* p<0.1. Lags 0-5 estimated but only shown for 0-4. Spatial SE calculated using a distance of 250km and uniform weights. Column
6 is collapsed to the province level. Income includes all wages, salary, and net transfers.

Observations
R-squared

T+4

T+3

T+2

T+1

0.26
[0.26]
-0.88***
[0.19]
1.04***
[0.32]
0.16
[0.17]
-0.58**
[0.25]
142,789
0.32

Income
(%)

Income
(%)

Max wind speed, T=0 (m/s)

VARIABLES

(2)

(1)

Table VII
Household income as a function of typhoon exposure and covariates

Table VIII
Typhoon impact on income sources
Dependent variable

Max wind speed, T + 1 (m/s)

% change

SE

Obs.

Agricultural wages

0.04

[0.27]

30,773

Non-Agricultural wages

-0.19

[0.15]

77,754

Entrepreneurial income

-0.28**

[0.11]

96,989

[0.21]
[0.18]
[0.44]
[0.46]
[0.24]
[0.42]
[0.26]
[0.45]
[0.71]
[0.96]
[1.24]

52,193
30,479
17,158
10,776
10,258
8,715
7,855
7,011
2,537
1,454
870

Subcategories of entrepreneurial income

Crop farming / gardening income
Trade income
Livestock / poultry income
Gambling winnings
Fishing income
Manufact. income
Transport / storage income
Services income
Forestry / hunting income
N.A. / entrep. income
Construct. income

-0.29
-0.18
-0.46
0.28
-0.40
0.08
0.15
-0.10
-0.44
-1.04
-1.56

Notes: Percent change calculated as log points *100 per m/s of max wind speed. Standard errors
clustered at the treatment (province) level in brackets. *** p<0.01, ** p<0.05, * p<0.1. Lags 0-5
estimated but only shown for 0-4. Includes province and year fixed effects, lagged temperature and
precipitation controls, and household controls consisting of the number of members, working and
non, in household as well as household heads gender and education level.

79

80
60.73
[38.30]
111,518

Female infant mortality

(per million HH)
Observations

-0.30**
[0.10]
71,157

(4)

(5)

55.14
[41.51]
36,845

-0.34***
[0.11]
21,101

-0.31**
[0.12]
21,101

74.40**
[24.85]
228,101

-0.44***
[0.09]
121,678

-0.41***
[0.11]
121,678

52.59
[44.77]
92,493

-0.44***
[0.11]
51,438

-0.35***
[0.11]
51,438

85.62***
[20.00]
172,937

-0.47***
[0.10]
91,341

-0.46***
[0.11]
91,341

Notes: Percent changes to income and expenditures calculated as log points *100 per m/s of max wind speed. Mortality shown per million
households per year. Standard errors clustered at the region level in brackets. *** p<0.01, ** p<0.05, * p<0.1. Lags 0-5 estimated but
only shown for lag 1. Income and expenditures models include province and year fixed effects, temperature and precipitation controls, and
household controls consisting of the number of members, working and non, in household as well as household heads gender (except models
3-4) and education level (except models 5-6). Mortality models include mother and year fixed effects, lagged temperature and precipitation
controls, and mothers age and age squared. In income and expenditure models, the sample is based on whether the household head (HHH)
at the time of the interview had completed primary school. In mortality models, the sample is stratified based on whether the mother had
completed primary school.

97.91**
[33.71]
82,519

-0.35***
[0.12]
71,622

-0.27**
[0.10]
71,622

(6)

Household head sex HHH/mothers education

Female
Male
<Primary
Primary

(3)

Independent variable: Max wind speed, T + 1 (m/s)

-0.30***
[0.07]
71,157

Observations

(2)

Household location
Nonurban Urban

Expenditures (% change)

Observations

Income (% change)

Dependent variable

Stratification by:

(1)

Table IX
Typhoon impacts by subpopulation

81
142,779
0.61

Observations
R-squared

142,779
0.66

-0.22**
[0.09]
-0.35***
[0.09]
-0.23**
[0.11]
-0.12
[0.11]
0.03
[0.07]
142,757
0.54

0.14
[0.14]
-0.21*
[0.12]
0.08
[0.13]
0.19
[0.15]
-0.01
[0.11]

Fuel
(%)

(3)

140,384
0.53

-0.01
[0.20]
-0.74***
[0.15]
0.11
[0.21]
0.13
[0.18]
-0.24*
[0.13]

Personal
Care
(%)

(4)

136,030
0.29

-0.18
[0.22]
-0.32*
[0.16]
-0.18
[0.22]
0.20
[0.23]
0.13
[0.15]

Clothing
(%)

(5)

135,077
0.37

0.06
[0.24]
-0.84***
[0.17]
-0.31
[0.24]
0.17
[0.25]
0.02
[0.16]

Travel /
Comm.
(%)

(6)

127,364
0.17

-0.20
[0.26]
-0.85***
[0.21]
-0.31
[0.29]
0.10
[0.27]
-0.10
[0.16]

Medical
(%)

(7)

100,268
0.30

0.20
[0.28]
-0.79***
[0.16]
-0.26
[0.26]
-0.24
[0.23]
-0.20
[0.15]

Education
(%)

(8)

88,440
0.28

-0.15
[0.22]
-0.57***
[0.18]
-0.34
[0.24]
-0.04
[0.26]
-0.34**
[0.15]

Special
Events
(%)

(9)

60,893
0.24

-0.54*
[0.31]
-0.11
[0.23]
0.79**
[0.37]
0.20
[0.32]
-0.36
[0.26]

Recreation
(%)

(10)

34,923
0.21

-0.11
[0.85]
0.09
[0.33]
0.78*
[0.46]
-0.30
[0.79]
-0.46
[0.40]

Repair
(%)

(11)

Notes: Percent change calculated as log points *100 per m/s of max wind speed. Standard errors clustered at the treatment (province) level in brackets. *** p<0.01, ** p<0.05, *
p<0.1. Lags 0-5 estimated but only shown for 0-4. Includes province and year fixed effects, lagged temperature and precipitation controls, and household controls consisting of the
number of members, working and non, in household as well as household heads gender and education level.

T+4

T+3

T+2

T+1

-0.03
[0.10]
-0.42***
[0.10]
-0.14
[0.11]
-0.01
[0.11]
-0.09
[0.08]

Food
(%)

Total
Exp.
(%)

Max wind speed, T=0 (m/s)

VARIABLES

(2)

(1)

Table X
Typhoon impact on major expenditure categories

82
137,495
0.46

Observations
R-squared

142,361
0.41

-0.49**
[0.19]
-0.14
[0.13]
-0.37*
[0.20]
-0.21
[0.20]
0.12
[0.12]
133,913
0.36

-0.38**
[0.18]
-0.75***
[0.19]
-0.18
[0.22]
-0.18
[0.25]
-0.08
[0.13]

Dairy
(%)

(3)

142,421
0.50

-0.03
[0.20]
-0.83***
[0.15]
-0.37**
[0.16]
0.03
[0.20]
-0.02
[0.12]

Fruit
(%)

(4)

142,684
0.69

-0.07
[0.07]
-0.22***
[0.06]
-0.08
[0.09]
-0.01
[0.09]
-0.05
[0.05]

Cereal
(%)

(5)

115,746
0.31

79,381
0.11

-0.64**
[0.29]
-0.17
[0.28]
-0.21
[0.33]
-0.38
[0.33]
0.40*
[0.21]

Alcoholic
Beverages
(%)

Nonalcoholic
Beverages
(%)
-0.47**
[0.19]
-0.45**
[0.18]
-0.27
[0.25]
-0.62**
[0.27]
0.28*
[0.16]

(7)

(6)

89,148
0.16

-0.35
[0.29]
0.32
[0.23]
-0.03
[0.28]
-0.25
[0.30]
0.34
[0.21]

Tobacco
(%)

(8)

Notes: Percent change calculated as log points *100 per m/s of max wind speed. Standard errors clustered at the treatment (province) level in
brackets. *** p<0.01, ** p<0.05, * p<0.1. Lags 0-5 estimated but only shown for 0-4. Includes province and year fixed effects, lagged temperature
and precipitation controls, and household controls consisting of the number of members, working and non, in household as well as household heads
gender and education level.

T+4

T+3

T+2

T+1

-0.73***
[0.22]
-0.74***
[0.22]
-0.55**
[0.23]
-0.52**
[0.23]
-0.01
[0.17]

Fish
(%)

Meat
(%)

Max wind speed, T=0 (m/s)

VARIABLES

(2)

(1)

Table XI
Typhoon impact on food expenditures

83

-43.16
[30.16]
79.68*
[36.61]
-32.82
[39.64]
-5.614
[32.21]
63.93
[52.90]
265,430
0.106

Max wind speed, T=0 (m/s)

Observations
R-squared

265,430
0.090

-8.765
[19.39]
26.11
[29.76]
-21.35
[34.04]
-0.695
[21.82]
43.47
[35.96]

Male

(2)

265,430
0.086

-36.82
[32.92]
58.82**
[26.03]
-5.694
[22.42]
-4.957
[24.56]
26.07
[29.56]

Female

(3)

265,430
0.083

265,430
0.082

21.16
[23.29]
73.37***
[20.71]
28.54
[17.79]
37.99**
[15.89]
41.84
[25.27]

Infant
female

Under
5 y.o.
female
-24.96
[25.76]
54.46**
[22.50]
-4.821
[22.27]
12.47
[15.39]
30.36
[26.65]

(5)

(4)

265,430
0.084

-5.697
[16.44]
1.717
[16.72]
-21.54
[22.19]
-20.31
[17.46]
23.61
[22.94]

Infant
male

(6)

Notes: Mortality shown per million households per year. Standard errors clustered at the treatment (region) level in brackets.
*** p<0.01, ** p<0.05, * p<0.1. Lags 0-5 estimated but only shown for 0-4. Includes mother and year fixed effects, lagged
temperature and precipitation controls, and mothers age and age squared.

T+4

T+3

T+2

T+1

Any age

VARIABLES

(1)

Table XII
Typhoon impact on child mortality

Table XIII
Typhoon impact on child mortality by SES group
(1)
VARIABLES
Max wind speed, T=0 (/s)
T+1
T+2
T+3
T+4

Observations
R-squared

(2)

Poorest
decile Below med.

(3)

(4)

Above med.

Top decile

98.33
[62.12]
146.1*
[70.92]
140.8*
[73.40]
164.8*
[91.13]
173.4**
[74.24]

9.214
[32.81]
82.15**
[31.36]
31.53
[35.88]
54.30**
[23.82]
58.67
[34.35]

36.08
[25.34]
63.44**
[23.42]
28.99
[20.53]
21.39
[16.06]
20.27
[22.20]

87.85
[58.01]
62.74
[56.10]
-5.418
[47.83]
27.98
[34.26]
13.61
[45.81]

26,637
0.084

142,216
0.086

123,214
0.071

20,587
0.072

Notes: Mortality shown per million households per year. Standard errors clustered at the treatment (region)
level in brackets. *** p<0.01, ** p<0.05, * p<0.1. Lags 0-5 estimated but only shown for 0-4. Includes mother
and year fixed effects, lagged temperature and precipitation controls, and mothers age and age squared.

84

Table XIV
Typhoon impact on infant female mortality by sibling gender
Max wind speed,
T=1 [m/s]

VARIABLES

First born
Marginal impact of having older siblings

Only older sisters

Marginal impact of having older brothers

Only older brothers

Marginal impact of having older sisters

35.32
[23.48]
75.70**
[25.35]
75.80
[78.97]
53.18
[67.54]
121.2*
[57.10]
0.405
[40.07]

Notes: Mortality shown per million households per year. Standard errors clustered at the treatment (region) level in brackets. *** p<0.01, ** p<0.05, * p<0.1. Lags 0-5 estimated but only
shown for 0-4. Includes mother and year fixed effects, lagged temperature and precipitation
controls, and mothers age and age squared.

85

Table XV
Summary of household losses due to previous years mean typhoon exposure
Outcome

Average annual loss

6.6%

Household income

7.1%
5.9%
12.5%
13.3%
14.3%

Household expenditures
Food expenditure (all)
Meat expenditure
Education expenditure
Medical expenditure
Female infant mortality per 1,000 births
National total female infant deaths (based on 2007 census)
Fraction of infant mortality for non-migrant females (sample)
Fraction of national infant mortality rate (male & female)

+18.1 deaths
+11,261 deaths
55.0%
13.0%

Notes: Losses are calculated using mean typhoon exposure between 1979 and 2008 as shown in Table I for the province level
(region level for infant mortality) using coefficient on T + 1 from estimates in Tables VII (col. 4),X (cols. 1, 2, 7, and 8) and
XII (col 5). Losses for income and expenditures are log points times 100.

86