Journal of Exposure Analysis and Environmental Epidemiology (2005) 15, 225–233
r 2005 Nature Publishing Group All rights reserved 1053-4245/05/$30.00
www.nature.com/jea
Factor analysis of pesticide use patterns among pesticide applicators in the
Agricultural Health Study
CLAUDINE SAMANIC,a JANE A. HOPPIN,b JAY H. LUBIN,a AARON BLAIRa AND MICHAEL C.R. ALAVANJAa
a
National Cancer Institute, Division of Cancer Epidemiology & Genetics, Bethesda, Maryland, USA
National Institutes of Environmental Health Sciences, Epidemiology Branch, Research Triangle Park, North Carolina, USA
b
Exposure to certain pesticides has been linked with both acute and chronic adverse health outcomes such as neurotoxicity and risk for certain cancers.
Univariate analyses of pesticide exposures may not capture the complexity of these exposures since use of various pesticides often occurs simultaneously,
and because specific uses have changed over time. Using data from the Agricultural Health Study, a cohort study of 89,658 licensed pesticide applicators
and their spouses in Iowa and North Carolina, we employed factor analysis to order to characterize underlying patterns of self-reported exposures to 50
different pesticides. Factor analysis is a statistical method used to explain the relationships between several correlated variables by reducing them to a
smaller number of conceptually meaningful, composite variables, known as factors. Three factors emerged for farmer applicators (N ¼ 45,074): (1) Iowa
agriculture and herbicide use, (2) North Carolina agriculture and use of insecticides, fumigants and fungicides, and (3) older age and use of chlorinated
pesticides. The patterns observed for spouses of farmers (N ¼ 17,488) were similar to those observed for the farmers themselves, whereas five factors
emerged for commercial pesticide applicators (N ¼ 4,384): (1) herbicide use, (2) older age and use of chlorinated pesticides, (3) use of fungicides and
residential pest treatments, (4) use of animal insecticides, and (5) use of fumigants. Pesticide exposures did not correlate with lifestyle characteristics such
as race, smoking status or education. This heterogeneity in exposure patterns may be used to guide etiologic studies of health effects of farmers and other
groups exposed to pesticides.
Journal of Exposure Analysis and Environmental Epidemiology (2005) 15, 225–233. doi:10.1038/sj.jea.7500396
Published online 28 July 2004
Keywords: pesticides, farmers, custom applicators, factor analysis, herbicides, insecticides, fungicides, fumigants, Iowa, North Carolina.
Introduction
Factor analysis has been increasingly used in epidemiologic
studies to examine relationships such as dietary factors and
breast cancer risk (Lubin et al., 1981; Terry et al., 2001),
patterns of metabolic factors and heart disease (Marusic,
2000), and environmental factors associated with prevalence
of acute respiratory infection in children (Gupta et al., 1999).
Factor analysis is a statistical method used to explain the
relationships between several correlated variables by reducing
them to a smaller number of conceptually meaningful,
composite variables, called factors (Kleinbaum and Kupper,
1978). These factors may be used as independent or
dependent variables in subsequent analyses, and may be
used to guide subsequent analyses. Therefore, factor analysis
is often used in conjunction with more traditional statistical
methods such as regression analysis.
Using data from the Agricultural Health Study (AHS), a
prospective cohort study of licensed pesticide applicators and
1. Address all correspondence to: Ms. C. Samanic, National Cancer
Institute, Division of Cancer Epidemiology & Genetics, 6120 Executive
Blvd., EPS 8115, Bethesda, MD 20892-7240, USA.
Tel.: þ 1-301-402-7824. Fax: þ 1-301-402-1819.
E-mail: samanicc@mail.nih.gov
Received 28 August 2003; accepted 2 June 2004; published online 28 July
2004
their spouses in Iowa and North Carolina, we employed
factor analysis to examine the underlying patterns of selfreported exposures to 50 pesticides. Exposure to certain
pesticides has been linked with both acute and chronic
adverse health outcomes such as neurotoxicity and certain
types of cancer (Zahm et al., 1997). Two important factors
that make pesticide exposure assessment difficult are that use
of various pesticides often occurs simultaneously, and that
pesticide products registered for specific uses have changed
over time. Therefore, characterizing patterns of exposure
may provide additional insight into disease occurrence than
evaluating single exposures.
Three types of applicators were enrolled into the AHS
cohort: farmer applicators, spouses of farmer applicators,
and commercial applicators. Each group provided information on use of the same 50 pesticides. Since the types
of application activities may vary between farmer and
commercial applicators, and may vary between farmers
and their spouses, we performed separate analyses for each
group.
Methods
The Agricultural Health Study has been described elsewhere
in detail (Alavanja et al., 1996). Briefly, the Agricultural
Samanic et al.
Health Study is a prospective cohort study of 57,311 licensed
pesticide applicators (including 52,395 farmer and 4916
commercial applicators) and 32,347 spouses of farmer
applicators in Iowa and North Carolina. Farmer applicators
and spouses were enrolled from both states, while commercial
applicators were enrolled only from Iowa. Data were
collected by means of a self-administered questionnaire given
at study enrollment. The enrollment questionnaire sought
information concerning use of various pesticides, pesticide
application methods, use of personal protective equipment,
types of crops and livestock raised, smoking and alcohol
consumption, medical history, diet, as well as basic demographic information. Recruitment into the cohort began in
December 1993 and continued through December 1997.
Questionnaires may be obtained from the following website:
http://www.aghealth.org.
In the enrollment questionnaire, farmer and commercial
applicators were asked to provide detailed information on
their use of 22 pesticides, including frequency (average
number of days per year used), duration (total number of
years used), and the decade when they first starting using the
pesticide. For an additional 28 pesticides, exposure was
reported in terms of ever/never use. Farmer applicators were
also given questionnaires for completion by their spouses.
Spouses were asked to report whether they had ever used any
of the same 50 pesticides during their lifetimes. In general,
pesticides are substances used to destroy or prevent any pest,
including insects, animals, weeds, fungi, molds and bacteria.
While insecticides specifically target insect pests, herbicides
are used to destroy weeds or other plant pests, and fungicides
and fumigants target fungi and mold. The group of 50
pesticides evaluated in this analysis included 18 herbicides, 22
insecticides, four fumigants and six fungicides.
In order to include all 50 pesticides in the factor analysis,
and to compare results between applicators and spouses,
reported use of the 50 pesticides was scaled as ever ¼ 1 or
never ¼ 0. Other variables included in the analysis were
demographic variables for state of residence (0 ¼ North
Carolina, 1 ¼ Iowa), subjects over 50 years of age at
enrollment (0 ¼ r50, 1 ¼ 450), and gender (0 ¼ female,
1 ¼ male).
In addition to questions about specific pesticides used,
farmer applicators were asked detailed questions about their
farm activities and pesticide application practices, which
allowed us to perform additional analyses to determine
whether any of the following would cluster with pesticide use:
types of crops grown or livestock raised; methods of pesticide
application generally employed; protective equipment generally used while applying pesticides, and lifestyle characteristics such as smoking and alcohol consumption. Variables
for types of crops included the following: field corn, sweet
corn, soybeans, cotton, peanuts, tobacco, Christmas trees,
strawberries, peaches, other fruit (apples, blueberries, grapes,
watermelon, and other fruits), other vegetables (alfalfa,
226
Factor analysis of pesticide use
cabbage, cucumbers, green peppers, potatoes, snap beans,
sweet potatoes, tomatoes, and other vegetables), and grains
(hay, oats, sorghum, wheat, and other). Six livestock
variables included beef cattle, dairy cattle, poultry, sheep,
eggs, and hogs. There were 11 variables for herbicide or crop
insecticide application methods: aerial, airblast, backpack
spraying, tractor boom, hand spraying, in furrow or banded
application, mist blowing, row fumigation, pouring fumigants from a bucket, use of a gas canister, and powder
dusting. Two additional variables for crop pest control
methods were use of granules/tablets and planting pretreated
seeds. Animal insecticide application methods included
spraying animals, dipping animals, applying ear tags, and
injecting animals. Types of personal protective equipment
generally used while applying pesticides included use of
chemical gloves, use of fabric gloves, wearing disposable
clothing such as Tyveks, use of a face shield or goggles, use
of a respirator or face mask, and none. We also included the
following six lifestyle variables: smoking status (ever/never),
race (white/non-white), education (less than high school/high
school or higher), fruit consumption (1 or more per day/less
than 1 per day), vegetable consumption (1 or more per day/
less then 1 per day), and alcohol consumption (1 or less per
week/more than 1 per week).
All analyses were performed using Statistical Analysis
Software (SAS Institute, 2001), using the principal axis
factoring method (PROC FACTOR). With this method,
factors were extracted, or derived, in descending order of
importance with respect to the proportion of variance in the
observed data accounted for by each factor (Hatcher, 1994).
For example, the first factor derived is the weighted linear
combination of the variables which accounts for the largest
total variation in the data (Kleinbaum and Kupper, 1978).
No other linear combination of variables should have as large
a variance as the first derived factor. The second factor
derived accounts for the largest proportion of the remaining
variance not accounted for by the first factor, and so on.
Each variable included in the analysis contributes one unit
of variance to the total variance in the data set. The
eigenvalue associated with each factor represents the amount
of variance accounted for by that factor. The sum of all
eigenvalues equals the total number of variables in the
dataset. An eigenvalue less than 1.00 indicates a factor that
accounts for less variance than a single variable. Since the
goal of factor analysis is to reduce a large number of variables
down to a relative small number of summary factors, it is not
efficient to retain factors that account for less variance than
what was contributed by the original variables. An
eigenvalue greater than 1.00, however, indicates a summary
factor that accounts for a greater amount of variance than
had been contributed by one variable.
A factor loading score was calculated for every variable in
every factor. Factor loading scores represent the correlations
between each of the variables included in the analysis and
Journal of Exposure Analysis and Environmental Epidemiology (2005) 15(3)
Factor analysis of pesticide use
Samanic et al.
each summary factor, and are equivalent to Pearson
correlation coefficients (Hatcher, 1994). Ideally, few variables in each factor will have a factor loading score above a
certain value that is specified prior to analysis. In general, for
exploratory analyses, factor loading scores are considered to
be meaningful when they exceed 70.30 or 0.40 (Floyd and
Widaman, 1995). Since this was an exploratory technique we
selected 70.40 or higher, in order to be more discriminating.
The initial factor analyses performed for farmer applicators,
their spouses, and commercial applicators included the 50
pesticide variables, age, gender, and state (state was not
included for commercial applicators because all were from
Iowa).
After initial extraction, factors were rotated to aide
interpretation. We employed oblique rotation to allow for
some degree of correlation among the newly derived factors.
Since pesticide and farming activities may change over time,
we assumed that farmer’s activities may be described by more
than 1 factor. Rotation may be defined as a linear
transformation performed on the initial factors, with the
goal of simplifying the factor structure, so that each variable
will load on as few factors as possible (Gorsuch, 1983). In
simpler terms, rotation simplifies the factor patterns so that
the variables in each factor with high factor loading scores are
different for each factor, yielding distinct groupings of
variables that are used to interpret the factors.
To determine how many factors to retain we applied the
following criteria: there must be at least three variables in the
factor with a high factor loading score (70.40 or greater);
factors must have an eigenvalue greater than 1.00; and each
factor must account for at least 5% of the total variance.
Ideally, the number of retained factors will be small, and will
explain the majority of the variance in the observed data
(Floyd and Widaman, 1995).
Table 1. Characteristics of licensed pesticide applicators and spouses
in the Agricultural Health Study cohort (1993–1997) included in the
study
Results
analyses revealed both consistent patterns of ever/never
chemical use across applicator subgroups, as well as patterns
that differed among subgroups.
A total of 8934 (7321 farmer, 532 commercial, 1081 spouse)
participants with one or more missing variables were omitted
from the analyses, resulting in 45,074 farmer applicators and
4384 commercial applicators retained in the factor analyses
(Table 1). For spouses of farmer applicators, we also
excluded an additional 13,778 who reported that they had
never mixed or applied pesticides, leaving 17,488 (56.1%) in
the analysis. Commercial applicators tended to be younger
than farmer applicators, and the majority of all cohort
members were white and lived in Iowa. The majority of
farmer (97.5%) and commercial (95.9%) applicators were
male, while the majority of spouses were female (98.9%).
Correlation coefficients were higher for pesticides within
the same class (i.e., herbicides, insecticides), ranging from
0.30 to 0.70, and lower or close to zero among pesticides of
different classes (data not shown). Results of the factor
Journal of Exposure Analysis and Environmental Epidemiology (2005) 15(3)
Characteristic
Applicators
Farmer (%)
N ¼ 45,074
Commercial (%)
N ¼ 4384
Spouses (%)
N ¼ 17,488
State
Iowa
North Carolina
29,277 (64.9)
15,797 (35.1)
4384 (100)
12,791 (73.1)
4697 (26.9)
Age (years)
o40
40–50
450
Not reported
14,199
13,631
17,240
4
2588 (59.0)
1184 (27.0)
612 (14.0)
4670 (26.7)
5931 (33.9)
6887 (39.4)
Gender
Female
Male
1120 (2.5)
43,954 (97.5)
204 (4.1)
4712 (95.9)
17,299 (98.9)
189 (1.1)
Race
White
Non-white
Not reported
43,115 (95.6)
936 (2.1)
1023 (2.3)
4838 (98.4)
26 (o1.0)
52 (1.0)
17,312 (99.0)
141 (0.8)
35 (0.2)
Highest grade
completed
o12 years
Z12 years
Not reported
3701 (8.2)
39,455 (87.5)
1918 (4.3)
146 (3.0)
4592 (93.4)
178 (3.6)
896 (5.0)
16,244 (93.0)
348 (2.0)
Smoking status
Never
Ever
Not reported
21,869 (48.5)
21,922 (48.6)
1283 (2.8)
2285 (46.5)
2536 (51.6)
95 (1.9)
12,181 (69.6)
4766 (27.3)
541 (3.1)
(31.5)
(30.2)
(38.2)
(o1.0)
Farmer Applicators
For farmer applicators, three factors (F1–F3) were retained
that explained 89% of the total variance in the observed data
(Table 2). Variables loading high on factor F1 (i.e., with
factor loading scores equal to 70.40), included most of the
herbicides, two insecticides (phorate, terbufos,), and Iowa.
Factor F1 explained 44% of variance in the observed data.
Factor F2 explained an additional 31% of the variance, and
was comprised of one herbicide (paraquat), four insecticides
(aldicarb, carbaryl, diazinon, parathion,), one fumigant
(methyl bromide), four fungicides (benomyl, chlorothalonil,
mancozeb, metylaxyl), and North Carolina. Variables
significant to Factor F3 were two herbicides (2,4,5-T and
2,4,5-TP), six chlorinated insecticides (aldrin, chlordane,
227
Factor analysis of pesticide use
Samanic et al.
Table 2. Factor analysis results for farmer applicators (N ¼ 45,074),
including 50 pesticides, age, state and gendera
Factor F1
Herbicides
Insecticides
Fumigants
Fungicides
228
Factor F2
Factor F3
49
59b
50b
48b
54b
55b
43b
31
58b
58b
62b
14
44b
46b
60b
51b
7
6
12
1
10
18
10
23
7
28
21
11
0
52b
13
30
1
1
11
0
2
0
9
11
4
3
0
4
11
11
5
6
12
13
2
8
42b
56b
Aldicarb
8
Aldrin
12
Carbaryl
4
Carbofuran
32
Chlordane
0
Chlorpyrifos
34
Coumaphos
11
DDVP
21
DDT
12
Diazinon
6
Dieldrin
2
Fonofos
32
Heptachlor
13
Lindane
17
Malathion
31
Parathion
7
Permethrin
26
(animal)
Permethrin (crop) 29
Phorate
40b
Terbufos
44b
Toxaphene
8
Trichlorfon
6
61b
6
44b
18
19
23
2
3
10
40b
0
8
14
6
16
41b
3
9
63b
21
15
54b
3
15
18
64b
26
56b
10
62b
36
19
24
4
33
2
0
28
9
8
17
1
41b
2
Aluminum
14
phosphide
Ethylene
4
dibromide
Methyl bromide 11
80/20 mix
3
18
14
36
20
58b
11
3
38
Benomyl
Captan
Chlorothalonil
Mancozeb
Metylaxyl
Ziram
61b
18
54b
56b
61b
23
5
7
12
9
3
18
Alachlor
Atrazine
Butylate
Chlorimuron ethyl
Cyanazine
Dicamba
EPTC
Glyphosate
Imazethapyr
Metolachlor
Metribuzin
Paraquat
Petroleum oil
Pendimethalin
Trifluralin
2,4-D
2,4,5 T P
2,4,5 T
b
1
17
6
10
4
2
Table 2. (Continued)
Demographics Age450 years
State ¼ Iowa
Gender ¼ male
Factor F1
Factor F2
28
46b
17
13
66b
0
Factor F3
55b
8
2
Eigenvalue
6.75
4.79
2.23
% variance
explained
0.44
0.31
0.14
% cumulative
variance
0.44
0.75
0.89
a
For ease of presentation, all values were multiplied by 100 and rounded to
the nearest integer.
b
Indicates factor loading score of 70.40 or higher.
DDT, dieldrin, heptachlor, toxaphene), and age greater than
50 years.
Iowa Commercial Applicators
Five factors were retained (C1–C5) that explained 97% of
the total variance in the observed data for this subgroup
(Table 3). Almost all of the pesticide variables clustered in
one of the five factors, and none overlapped. The first two
factor patterns for commercial applicators were similar, but
not identical to, two of the patterns observed for farmer
applicators. Factor C1 included almost all of the herbicides
(except for 2,4,5-T and 2,4,5-TP), and one crop insecticide
(permethrin). Factor C2 included two herbicides (2,4,5-Tand
2,4,5-TP), six chlorinated insecticides (aldrin, chlordane,
DDT, dieldrin, heptachlor, toxaphene), and age greater than
50 years. Pesticides significant to Factor C3 included
fungicides (benomyl, chlorothalonil, mancozeb, metylaxyl)
and residential insect treatments (carbaryl, diazinon, trichlorfon). Insecticides that are often used for controlling
livestock pests loaded high on Factor C4, including
carbofuran, DDVP, fonofos, permethrin and terbufos.
Factor C5 was comprised of one fungicide (ziram), three
fumigants: aluminum phosphide, ethylene dibromide, and
80/20 mix, a carbon tetrachloride/carbon disulfide mixture.
Spouses of Farmer Applicators
The factor patterns for spouses (S1–S3) were generally
similar to those observed for farmer applicators, and
explained 91% of the total variance in the observed data
(Table 4). Factor S1 explained 62% of the variance, and
included most of the herbicides and four insecticides
(carbofuran, chlorpyrifos, fonofos, terbufos). Aldrin, chlordane, DDT, dieldrin, and heptachlor, five chlorinated
insecticides, comprised Factor S2. Pesticides significant to
Factor S3 were a combination of a fumigant (methyl
bromide) and fungicides (benomyl, chlorathalonil, mancozeb, metylaxyl). For spouses, none of the demographic
Journal of Exposure Analysis and Environmental Epidemiology (2005) 15(3)
Factor analysis of pesticide use
Samanic et al.
Table 3. Factor analysis results for commercial
(N ¼ 4384), including 50 pesticides, age, and gendera
applicators
Factor Factor Factor Factor Factor
C1
C2
C3
C4
C5
Herbicides
Insecticides
Fumigants
Fungicides
b
0
2
13
4
12
12
6
4
6
0
5
6
4
2
6
6
85
74b
77b
48b
89b
89b
85b
52b
85b
52b
79b
58b
11
12
0
2
5
6
8
5
4
12
5
9
2
2
62b
73b
12
24
10
34
10
13
10
12
11
8
0
29
10
6
0
9
2
0
6
3
0
2
10
11
0
6
13
11
2
7
9
12
5
4
4
5
0
2
0
15
3
12
Aldicarb
Aldrin
Carbaryl
Carbofuran
Chlordane
Chlorpyrifos
Coumaphos
DDVP
DDT
Diazinon
Dieldrin
Fonofos
Heptachlor
Lindane
Malathion
Parathion
Permethrin
(animal)
Permethrin
(crop)
Phorate
Terbufos
Toxaphene
Trichlorfon
3
2
19
23
5
20
3
9
6
3
7
17
2
1
31
4
3
21
77b
15
8
49b
3
8
1
71b
15
65b
2
73b
28
3
29
8
19
12
44b
3
19
29
4
5
2
46b
4
2
8
13
26
10
2
3
5
12
42b
8
33
27
48b
2
17
4
50b
7
22
21
9
42b
24
1
7
9
4
7
4
19
2
1
18
9
0
15
2
17
11
45b
7
10
23
4
21
22
2
6
24
2
57b
7
0
1
5
46b
34
52b
4
1
0
7
14
6
Aluminum
phosphide
Ethylene
dibromide
Methyl
bromide
80/20 mix
15
14
3
25
53b
1
8
6
2
62b
13
11
2
32
28
2
12
1
15
53b
6
7
9
1
9
4
2
3
10
1
1
8
10
22
6
11
2
17
16
4
7
26
19
54b
Alachlor
Atrazine
Butylate
Chlorimuron
ethyl
Cyanazine
Dicamba
EPTC
Glyphosate
Imazethapyr
Metolachlor
Metribuzin
Paraquat
Pendimethalin
Petroleum oil
Trifluralin
2,4-D
2,4,5 T P
2,4,5 T
Benomyl
Captan
Chlorothalonil
Mancozeb
Metylaxyl
Ziram
78
83b
66b
86b
b
69b
29
69b
58b
50b
16
Journal of Exposure Analysis and Environmental Epidemiology (2005) 15(3)
Table 3. (Continued)
Factor Factor Factor Factor Factor
C1
C2
C3
C4
C5
Demographics Age450 years 8
Gender ¼ male 19
48b
7
9
7
7
5
13
1
Eigenvalue
12.07
4.99
2.21
1.38
1.09
% variance
explained
0.54
0.22
0.10
0.06
0.05
% cumulative
variance
0.54
0.76
0.86
0.92
0.97
a
For ease of presentation, all values were multiplied by 100 and rounded to
the nearest integer.
b
Indicates factor loading score of 70.40 or higher.
variables (age, state, gender) contributed to any of the
factors.
As mentioned previously, factors are extracted in order of
the proportion of variance accounted for by each factor, and
the factor explaining the greatest proportion of variance is
extracted first. For both farmer applicators and spouses of
farmer applicators, the summary factor that explained most
of observed variables was weighted heaviest by herbicide use.
The second and third factors derived for each group were
similar but differed according to the order in which they were
extracted. For farmer applicators, the second summary
factor (F2) was weighted by insecticides, methyl bromide
and fungicides, while the third factor was weighted by
chlorinated pesticides. This order was reversed for spouses.
Additional Analyses for Farmer Applicators
Because farmer applicators provided detailed information
concerning farming activities that would be correlated with
pesticide use, we performed additional factor analyses to
examine whether any of the following would cluster with
pesticide use: crops or livestock produced, pesticide application methods, protective equipment generally worn while
applying pesticides, and lifestyle characteristics. In this
analysis, an additional 3926 subjects were omitted due to
missing data for one or more of the variables included in the
analysis. The results (Table 5) were similar to what we
observed when we included pesticides only (Table 2), with the
addition of a fourth summary factor. Factor FF1 was
essentially the same as factor F1. In addition to the herbicides
and Iowa residence, the following variables contributed to
this factor pattern: field corn, soybeans, boom application, in
furrow or banded application, and use of chemical gloves.
Factor FF2 was also similar to what we observed when
evaluating pesticides only (Table 2; F2), with the addition of
cotton, peanuts, tobacco, and row fumigation, but with only
one insecticide (aldicarb). Significant contributors to Factor
FF3 included chlorinated pesticides and age greater than 50
229
Factor analysis of pesticide use
Samanic et al.
Table 4. Factor analysis results for spouses of farmers (N ¼ 17,488),
including 50 pesticides, age, state and gendera
Factor S1
Herbicides
Insecticides
Fumigants
Fungicides
230
Alachlor
Atrazine
Butylate
Chlorimuron
Ethyl
Cyanazine
Dicamba
EPTC
Glyphosate
Imazethapyr
Metolachlor
Metribuzin
Paraquat
Pendimethalin
Petroleum oil
Trifluralin
2,4-D
2,4,5 T P
2,4,5 T
Aldicarb
Aldrin
Carbaryl
Carbofuran
Chlordane
Chlorpyrifos
Coumaphos
DDT
DDVP
Diazinon
Dieldrin
Fonofos
Heptachlor
Lindane
Malathion
Parathion
Permethrin
(animal)
Permethrin
(crop)
Phorate
Terbufos
Toxaphene
Trichlorfon
Aluminum
phosphide
Ethylene
dibromide
Methyl bromide
80/20 mix
Benomyl
Captan
Chlorothalonil
Mancozeb
Metylaxyl
Ziram
b
Factor S2
Factor S3
70
70b
52b
57b
1
2
3
5
3
1
2
3
71b
69b
61b
22
75b
76b
65b
21
65b
48b
67b
37
15
13
1
5
5
2
11
12
2
4
11
9
2
14
22
34
5
6
3
6
7
0
2
24
12
5
4
0
2
3
6
4
11
41b
4
43b
6
1
4
0
5
43b
5
3
7
1
6
1
57b
17
27
46b
16
30
45b
33
19
52b
25
57b
33
26
22
25
31
10
24
7
12
17
1
9
3
32
7
3
11
12
21
18
3
13
16
18
36
53b
4
3
25
16
39
17
4
2
6
8
7
7
12
1
6
21
2
1
5
17
49b
7
2
1
5
3
7
0
1
11
2
1
4
3
45b
30
42b
49b
55b
20
Table 4. (Continued)
Factor S1
Factor S2
Demographics Age450 years 8
State ¼ Iowa
15
Gender ¼ male 12
20
7
1
Factor S3
0
39
20
Eigenvalue
8.35
2.37
1.50
% variance
explained
0.62
0.18
0.11
% cumulative
variance
0.62
0.80
0.91
a
For ease of presentation, all values were multiplied by 100 and rounded to
the nearest integer.
b
Indicates factor loading score of 70.40 or higher.
Table 5. Factor analysis results for farmer applicators, including
crops, pesticide application methods, protective equipment, 50
pesticides, demographic and lifestyle variables (N ¼ 41,148)a
Factor
FF1
Factor
FF2
Factor
FF3
Factor
FF4
12
4
12
19
18
14
8
2
4
2
10
12
8
8
7
2
4
14
0
22
14
13
14
4
8
16
11
19
40b
37
39
23
15
11
15
22
Crops/
livestock
Field corn
Sweet corn
Soybeans
Cotton
Peanuts
Tobacco
Trees
Strawberries
Peaches
Other fruit
Other vegetables
Grains
Beef cattle
Dairy cattle
Poultry
Sheep
Eggs
Hogs
59b
6
66b
20
19
11
27
10
13
19
1
5
3
9
4
3
5
24
24
11
8
56b
54b
61b
3
6
7
16
4
3
16
11
10
8
2
21
Application
methods
Aerial
Airblast
Backpack
sprayer
Boom
Hand sprayer
In furrow/
banded
Mist blowing
Row
fumigation
Pour
fumigants
Gas canister
6
12
12
7
10
21
4
8
12
2
9
23
43b
18
48b
19
3
7
9
8
3
3
28
17
5
8
4
52b
8
6
28
4
5
14
5
15
8
32
4
18
Journal of Exposure Analysis and Environmental Epidemiology (2005) 15(3)
Factor analysis of pesticide use
Samanic et al.
Table 5. (Continued )
Table 5. (Continued )
Factor
FF1
Powder
dusting
Tablets
Pretreated
seeds
Spraying
animals
Dipping
animals
Ear tags
Injecting
animals
Protective
equipment
Herbicides
Insecticides
Chemical
gloves
Disposable
clothing
Fabric gloves
Face shield/
goggles
Respirator/
face mask
None
Alachlor
Atrazine
Butylate
Chlorimuron
ethyl
Cyanazine
Dicamba
EPTC
Glyphosate
Imazethapyr
Metolachlor
Metribuzin
Paraquat
Petroleum oil
Pendimethalin
Trifluralin
2,4-D
2,4,5-T
2,4,5-TP
Aldicarb
Aldrin
Carbaryl
Carbofuran
Chlordane
Chlorpyrifos
Coumaphos
DDT
DDVP
Diazinon
Dieldrin
Fonofos
Heptachlor
Lindane
Malathion
Parathion
Factor
FF2
Factor
FF3
Factor
FF4
10
18
7
34
8
28
13
20
8
6
24
28
19
9
0
55b
4
10
3
39
8
18
8
4
2
4
52b
54b
40b
10
5
6
3
5
1
6
9
12
8
1
2
2
11
6
16
12
8
12
13
4
1
47
56b
48b
51b
15
4
11
21
7
5
18
4
1
1
6
7
48b
50b
36
20
58b
58b
60b
9
37
49b
59b
45b
6
5
8
23
8
25
18
16
0
47b
11
36
2
1
6
4
7
6
7
7
5
4
15
14
17
5
8
14
56b
45b
5
4
3
6
10
8
7
7
13
8
12
4
1
2
5
63b
28
19
57b
0
11
64b
16
34
57b
9
65b
38
25
29
11
13
12
8
2
14
30
10
24
14
6
7
5
17
13
7
14
b
16
18
4
26
2
28
0
8
11
4
4
28
16
8
24
4
68
8
36
17
10
23
3
5
7
30
6
7
17
2
10
35
b
Journal of Exposure Analysis and Environmental Epidemiology (2005) 15(3)
Permethrin
(animal)
Permethrin
(crop)
Phorate
Terbufos
Trichlorfon
Toxaphene
Fumigants
Fungicides
Aluminum
phosphide
Ethylene
dibromide
Methyl
bromide
80/20 mix
Benomyl
Captan
Chlorothalonil
Mancozeb
Metalaxyl
Ziram
Demographics State ¼ Iowa
Age450 years
Gender ¼ male
Race
Education
Lifestyle
Ever smoke
Alcohol intake
Fruit
consumption
Vegetable
consumption
Factor
FF1
Factor
FF2
Factor
FF3
Factor
FF4
13
6
0
35
22
32
0
7
42b
42b
2
11
1
3
8
23
17
1
5
44b
12
12
19
4
5
29
25
3
16
58b
2
6
4
40b
2
2
3
6
7
15
4
4
54b
11
58b
48b
61b
13
1
9
3
4
16
13
7
19
5
24
2
16
4
3
1
1
47b
18
21
7
15
67b
16
2
10
2
2
55b
2
6
4
1
23
3
9
13
5
11
4
7
3
11
14
4
17
9
2
9
2
3
11
15
Eigenvalue
8.41
6.85
2.94
2.42
% variance
explained
0.30
0.24
0.10
0.08
% cumulative
variance
0.30
0.54
0.64
0.73
a
For ease of presentation, all values were multiplied by 100 and rounded to
the nearest integer.
b
Indicates factor loading score of 70.40 or higher.
years, with the addition of the fumigant 80/20 mix. This is
also similar to Factor F3 observed for farmer applicators
(Table 2) and factor S2 observed for spouses (Table 4). The
fourth factor pattern included other vegetables, spraying of
animals, use of ear tags, and injecting animals. None of the
lifestyle variables achieved factor loading scores of 70.40 or
higher in any of the factors.
231
Samanic et al.
Discussion
In this exploratory analysis, factor analysis was a simple
and relatively quick tool for examining the complex relationships among a large set of exposure variables, and
summarizing 50 pesticide exposure variables into smaller
groups of exposure patterns that varied among the
three types of pesticide applicators in this cohort. The
variables considered to be significant to each factor
(i.e., variables with factor loading scores of 70.40 or higher)
were all correlated to some degree, because some of the
pesticides in each factor were likely to have been used
simultaneously. The correlations were also supported by an
understanding of pesticide use over time, since changes in
pesticide formulations may have led to changes in recommendations for use, or certain pesticides may have been taken
off the market.
The factors derived for farmer applicators identified three
distinct patterns of pesticide application activities: Iowa
agriculture (soybeans, corn) and herbicide use (F1); North
Carolina agriculture (cotton, peanuts, tobacco), which
requires more intensive insecticide and fumigant applications
(F2); and use of chlorinated insecticides, an exposure pattern
more typical of older farmers (F3). These chlorinated
insecticides are no longer marketed for use in the United
States (California EPA, 2002). Farmers who were 50 years
of age or older at the time of enrollment were more likely to
have greater exposure to these pesticides that were on the
market during the 1970s and 1980s, when they have been in
their 20s and 30s. These three factor patterns were the same
for both the pesticide-only analysis and analysis including
crop-type and other variables.
The additional variables that clustered with the herbicides
in Factor FF1 were also related to Iowa agriculture and
herbicide applications (i.e., field corn, soybeans, boom
application, in furrow or banded application), while the
additional variables that clustered with the pesticides in
Factor FF2 are typical of North Carolina agriculture (i.e.,
peanuts, cotton, tobacco, and row fumigation). Factor FF3
remained virtually unchanged; none of the additional farming variables patterned with chlorinated pesticides. A fourth
factor, which did not include any of the pesticides, was
suggestive of livestock production (i.e., other vegetables,
animal pest control application methods). Two additional
variables (beef cattle and dipping animals) were most likely
important to this fourth factor, although the factor loading
score for each of these variables was 0.39. None of the
lifestyle factors that we were able to include, and known to be
related to cancer, heart disease, and other types of health
problems, were significant components of the resulting factor
patterns. We observed no relationship between pesticide or
farm-related exposures and lifestyle characteristics, such as
smoking. In this cohort, the greatest amount of heterogeneity
seemed to be due to pesticide use, indicated by the fact that
232
Factor analysis of pesticide use
pesticides were the greatest contributors to the resulting
exposure patterns.
There was some overlap between the factor patterns
observed for farmer applicators and those observed for
commercial applicators. The five factors derived for commercial applicators may be described as follows: Iowa
agriculture and herbicide use (C1); chlorinated pesticides no
longer marketed for use in the US (C2); residential pesticide
treatments (C3); treatment of livestock (C4); and fumigation
(C5). It is not surprising that factor C1, which explains the
majority of the variance in the observed data, represents
Iowa agriculture, since all of the commercial applicators were
from Iowa. The factor patterns observed for this group of
applicators suggest a broader range of application activities
compared to farmer applicators, and therefore greater
potential pesticide exposures. These activities should be
considered in etiologic studies. Although farmer and
commercial applicators may receive similar pesticide exposures when applying herbicides or insecticides to crops,
commercial applicators may differ from farmer applicators
with respect to the type of pesticides they apply, the number
of spray jobs performed in a day, the range of pesticides
applied in a day, and the frequency with which other tasks
are performed, such as mixing and loading, and performing
maintenance on spray rigs (Hines et al., 2001).
Factor patterns observed for spouses of applicators were
similar to those observed among farmer applicators (i.e.,
their husbands), and may be described as Iowa agriculture
and herbicide use (Factor S1); chlorinated pesticides no
longer marketed for use in the US (Factor S2); and fumigant
and fungicide use (Factor S3). Although the spouses were
not the licensed pesticide applicators, they may engage in
similar pesticide application activities as the farmer applicators by assisting with field work, or performing additional
application activities such as pest control in the home,
garden, and on pets. Given that the majority of spouses live
in Iowa and share a similar age distribution as their farmer
husbands, use of chlorinated pesticides (Factor S2, related to
age) weights heavier than the pattern of pesticide use
associated with North Carolina (Factor S3), and explains
why among spouses, unlike their farmer husbands, use of
chlorinated pesticides explains a greater proportion of
variance than use of fumigants and fungicides.
In order to compare the factor patterns across the three
types of applicators, we were limited to ever/never response
data. Although farmer and commercial applicators provided
pesticide-specific information concerning number of days per
year and number of years applied, spouses were not asked to
provide such quantitative pesticide exposure information. In
addition, only 40% of farmer and commercial applicators
(N ¼ 24,365) provided quantitative exposure data for all 50
pesticides. Despite this limitation, the factor patterns we
observed for each applicator group were consistent with the
types of farming activities in which they would most likely be
Journal of Exposure Analysis and Environmental Epidemiology (2005) 15(3)
Factor analysis of pesticide use
engaged. In addition, because farmers and their spouses
engage in a variety of tasks to maintain their farms and
equipment, it may be necessary to consider nonpesticide
exposures when investigating exposure-disease associations,
such as the use of solvents, paint, or exposure to welding
fumes (Coble et al., 2002).
The factors observed for each type of applicator have been
used as independent variables in an epidemiologic analysis of
the association between exposure to various pesticides and
the risk of prostate cancer (Alavanja et al., 2003). To do this,
the factor loading scores for each of the variables in the
analysis and the subjects’ responses to each of those variables
(e.g., 0 ¼ No, 1 ¼ Yes) were used in an algorithm to calculate
a factor-based score for each subject, for each factor. These
scores represent each subject’s relative standing with respect
to how closely they ‘‘resemble’’ each factor (Kleinbaum and
Kupper, 1978), and may be used to divide subjects into low,
medium, and high exposure.
If a significant proportion of variables that heavily weight
a summary factor are independently related to a certain
health outcome, a quantitative association may be observed
between that factor and the health outcome of interest. If
only one or a small number of variables that weight a
summary factor are associated with the health outcome of
interest, then factor-based scores may not show a strong,
quantitative association with that particular health outcome.
In addition, the variables that significantly contribute to each
factor may help guide traditional multivariate analyses by
suggesting potential interaction between exposure variables.
Our results indicate that factor analysis was a suitable
method for characterizing patterns of exposure to the
pesticides among the licensed pesticide applicators and their
spouses in the AHS cohort, and should be considered for
similar kinds of epidemiologic analyses. The resulting factor
patterns were clearly interpretable and logical for the study
population, and completely data driven. The next step may
be to further explore these patterns by using quantitative
(e.g., continuous) data such as lifetime days of pesticide use,
which is being collected from all three subgroups (commercial
Journal of Exposure Analysis and Environmental Epidemiology (2005) 15(3)
Samanic et al.
applicators, farmers, spouses of farmers) during Phase II of
the study. The patterns of exposure observed in the present
study may be used to guide more traditional analyses of
pesticide exposure and risk for various health outcomes
among the subgroups of applicators in this cohort.
References
Alavanja M.C.R., Sandler D.P., McMaster S.B., Zahm S.H., McDonnell C.J.,
Lynch C.F., Pennybacker M., Rothman N., Dosemeci M., Bond A.E., and
Blair A. The Agricultural Health Study. Environ Health Perspect 1996: 104:
362–369.
Alavanja M.C.R., Samanic C., Dosemeci M., Lubin J., Tarone R., Lynch C.F.,
Knott C., Thomas K., Hoppin J.A., Barker J., Coble J., Sandler D.P., and
Blair A. Use of agricultural pesticides and prostate cancer risk in the
Agricultural Health Study cohort. Am J Epidemiol 2003: 157: 800–814.
California Environmental Protection Agency. Department of Pesticide Regulation
and United States Environmental Protection Agency. Pesticide Product
Information Query Site, 2002, ( http://www.cdpr.ca.gov/docs/epa/epamenu.htm).
Coble J., Hoppin J.A., Engel L., Elci O.C., Dosemeci D., Lynch C.F., and
Alavanja M.C.R. Prevalence of exposure to solvents, metals, grain dust and
other hazards among farmers in the Agricultural Health Study. J Exp Anal
Environ Epidemiol 2002: 12: 418–426.
Floyd F.J., and Widaman K.F. Factor analysis in the development and refinement
of clinical assessment instruments. Psych Assess 1995: 7: 286–299.
Gorsuch R.L. Factor Analysis. Lawrence Erlbaum Associates, Pubs, Hillsdale,
NJ, 1983.
Gupta R.K., Kumar A., and Singh P. Factor analysis of acute respiratory
infections among under fives in Delhi slums. Indian Pediatrics 1999: 36:
1146–1149.
Hatcher L. A Step-by-Step Approach to Using the SAS System for Factor Analysis
and Structural Equation Modeling. SAS Institute Inc., Cary, NC, 1994.
Hines C.J., Deddens J.A., Tucker S.P., and Hornung R.W. Distributions and
determinants of pre-emergent herbicide exposures among custom applicators.
Ann Occup Hyg 2001: 45: 227–239.
Kleinbaum D.G., and Kupper L.L. Applied Regression Analysis and Other
Multivariate Methods. MA, Duxbury Press, North Scituate, 1978.
Lubin J.H., Burns P.E., and Blot W.J., et al. Dietary factors and breast cancer
risk. Int J Cancer 1981: 28: 685–689.
Marusic A. Factor analysis of risk for coronary heart disease: an independent
replication. Int J Cardiol 2000: 75: 233–238.
SAS Institute. SAS, version 8.2. SAS Institute, Cary, NC, 2001.
Terry P., Suzuki R., Hu F.B., and Wolk A. A prospective study of major dietary
patterns and the risk of breast cancer. Cancer Epidemiol Biomarkers Prev 2001:
10: 1281–1285.
Zahm S.H., Ward M., and Blair A. Pesticides and cancer. Occup Med: State of the
Art Rev 1997: 12: 269–289.
233