0022-3018/01/1897–413 Vol. 189, No.

7
THE JOURNAL OF NERVOUS AND MENTAL DISEASE Printed in U.S.A.
Copyright © 2001 by Lippincott Williams & Wilkins

A Population-Based Twin Study of Generalized Anxiety Disorder in

Men and Women
JOHN M. HETTEMA, M.D., PH.D.,1 CAROL A. PRESCOTT, PH.D.,1 and KENNETH S. KENDLER, M.D.1,2

This study aimed to a) assess whether genetic or environmental effects are of

similar magnitude in the etiology of GAD in men and women, and b) investigate
whether familial (genetic or common environmental) risk factors are the same in
men and women, or whether there are gender-specific effects. We obtained a
lifetime history of DSM-IIII-R GAD, via face-to-face and telephone interviews, from
3100 complete male-male, female-female, and male-female twin pairs, ascertained
through a population-based registry. Biometrical twin modeling was utilized to
estimate the relative contributions of genetic and environmental factors to liability
for GAD, allowing for gender-specific effects. The familial aggregation of GAD in
this sample was only modest. In the best-fitting models, the heritability of GAD was
the same in men and women, estimated at about 15% to 20%, with no effects of
gender-specific genes detected.
—J Nerv Ment Dis 189:413–420, 2001

Generalized anxiety disorder (GAD) has under- blind study, found similar results in their sample
gone significant revision as a diagnostic entity over (morbidity risk 8.9% vs. 1.9%). In the first twin study
the last two decades. It was first defined as a syn- of DSM-III GAD, Torgerson (1983) found only 2 of 20
drome separate from panic disorder in the Research dizygotic (DZ) and 0 of 12 monozygotic (MZ) con-
Diagnostic Criteria in 1975 as a refinement of the cordant pairs. In a sample of 446 twin pairs from the
earlier concept of “anxiety neurosis” (Spitzer et al., Australian Twin Registry, Andrews et al. (1990) ob-
1975). Based upon increasing clinical and research tained a trend for higher MZ than DZ concordance
data, further revisions in the areas of symptom com- rates (21.5% vs. 13.5%, NS) for DSM-III GAD without
plex, duration, and diagnostic hierarchy have oc- diagnostic hierarchy. Skre et al. (1993) studied a
curred. Elucidating the etiological role of genetic sample of 49 twin pairs diagnosed with DSM-III-R
and environmental factors will potentially further anxiety disorders, finding a nonsignificant MZ-DZ
enhance our evolving understanding of GAD. concordance difference (3/5 vs. 1/7). These latter
Previous studies have produced results consistent two studies, though not definitive, suggest that the
with familial aggregation of GAD. In a nonblind fam- familial aggregation of GAD is most likely due to
ily study, Noyes et al. (1987) found higher rates of genetic effects.
DSM-III GAD in the first-degree relatives of GAD In 1992, we reported DSM-III-R lifetime diagnoses
probands than relatives of control subjects (19.5% of GAD in 1033 personally interviewed female-fe-
vs. 3.5%). Mendlewicz et al. (1993), in a controlled, male twin pairs identified through the population-
based Virginia Twin Registry (Kendler et al., 1992a).
That study estimated the heritability of GAD in
1
Virginia Institute for Psychiatric and Behavioral Genetics, women to be on the order of 20% to 30%, depending
Department of Psychiatry, Medical College of Virginia of Virginia
Commonwealth University, P.O. Box 980126, Richmond, Virginia on diagnostic definition. A bivariate study of GAD
23298 – 0126. Send requests for reprints to Dr. Kendler. and major depression in a combined population-
2
Department of Human Genetics, Medical College of Virginia based and clinically ascertained Swedish sample of
of Virginia Commonwealth University, Richmond, Virginia.
This work was supported by NIH grants MH-40828, and MH/ 1484 male and female twins (including 486 complete
AA-49492 and a Research Scientist Award (MH-01277) to Dr. pairs) estimated the heritability at 14.3% and 49% for
Kendler. Dr. Hettema is partially supported by an NIMH Training narrow and broad definitions, respectively (Roy et
Fellowship (MH-20030).
We acknowledge the contribution of the Virginia Twin Regis- al., 1995), with no evidence for sex differences.
try, now part of the Mid-Atlantic Twin Registry (MATR), to Scherrer et al. (2000) conducted a study of GAD and
ascertainment of subjects for this study. The MATR, directed by panic disorder in 3362 male-male twin pairs from the
Drs. L. Corey and L. Murrelle, has received support from the
National Institutes of Health, the Carman Trust, and the W. M. Vietnam Era Twin Registry, obtaining a heritability
Keck, John Templeton, and Robert Wood Johnson Foundations. estimate of 37.2% in their univariate model for a

413
414 HETTEMA et al.

broadly defined GAD. None of these studies re- ity and analyzed using 8 RFLP markers (Spence et
ported a significant role for common family environ- al., 1988). More recently, we have performed PCR
ment, though our original study found that it ac- zygosity tests on an additional 269 twin pairs, over-
counted for twin resemblance in two of nine sampling those where our prior zygosity assignment
diagnostic definitions. was questionable. On the basis of these tests (where
In this paper, we investigated the role of genetic the mean number of markers tested per pair was
and environmental factors in the etiology of GAD in 17.5 [SD ⫽ 8.4]), zygosity was changed for 12 pairs
men and women. We examined the impact of dura- (4.5% of those tested). In the MM/MF sample, zygos-
tion and diagnostic hierarchy with major depression ity was initially determined by an algorithm based
on these results. In particular, analyzing the com- on standard questions, validated against the zygosity
bined sample of male-male (MM), female-female diagnoses in the FF sample. Application of the algo-
(FF), and opposite-sex (MF) twins allowed us to a) rithm to this male sample was validated by analysis
assess whether genetic or environmental effects are of PCR polymorphisms (mean of 11.5 [SD ⫽ 11.9]
of similar magnitude in the etiology of GAD in men markers per pair) in a random sample of 196 twin
and women, and b) investigate whether familial (ge- pairs. The algorithm classified 186 pairs correctly,
netic or common environmental) risk factors are the an error rate of 5.1%. Lifetime GAD was assessed via
same in men and women, or whether there are gen- telephone and face-to-face structured psychiatric in-
der-specific effects. This study has several potential terview based on the Structured Clinical Interview
advantages over that of Roy et al. (1995), in that our for DSM-III-R (Spitzer and Williams, 1985). However,
sample is about four times the size, the assessments we assessed last-year history of GAD and a lifetime
are by direct interview, and varying diagnostic defi- history of GAD before the last year in two separate
nitions are tested. sections, combining these to estimate lifetime GAD.
DSM-III-R hierarchy rules with major depression
(MD) differentiated those affected with full overlap
Methods
of GAD and MD episodes (i.e., the two syndromes
The twins in this report derive from two interre- always occurred contemporaneously) from those
lated projects utilizing the population-based Virginia with no or only partial overlap. Interviewers, who
Twin Registry—formed from a systematic review of were carefully trained and supervised, had at least a
all birth certificates in the Commonwealth of Virgin- master’s degree in a mental health-related field or a
ia—which now constitutes part of the Mid-Atlantic bachelor’s degree in such a field and 2 years of
Twin Registry. The details of the samples in these clinical experience. Two senior staff reviewed each
two projects are described elsewhere (Kendler and interview for completeness and consistency. Data
Prescott, 1999). This study results from the fourth were double-entered to minimize data entry errors.
interview wave with the sample of white female- Members of a twin pair were interviewed by differ-
female twins (FF4) and the second interview wave ent interviewers, who were blind to clinical informa-
with the sample of white male-male and male-female tion about the co-twin.
twins (MM/MF2). In FF4, we assessed, via telephone Due to the relatively low prevalence of strictly
interview, both members of 832 pairs, 505 of whom defined GAD (1 to 3%), the statistical power neces-
were MZ and 327 of whom were DZ. The mean age of sary to perform model fitting was poor (Neale et al.,
the participating twins in FF4 was 36.3 (SD ⫽ 8.2) 1994). To improve this, we examined the separate
years and ranged from 21 to 62. From the MM/MF2 effects of reducing the minimum duration from 6
sample, we interviewed both members of 708 male- months to 1 month and/or broadening the diagnostic
male MZ pairs, 491 male-male DZ twins, and 1070 criteria. In our previous work (Kendler et al., 1992a),
opposite-sex male-female DZ pairs. Face-to-face in- we found that a multiple threshold model, which
terviews were conducted in 80% of the MM/MF2 proposes that 1-month and 6-month GAD lie on a
sample, with the remaining interviews conducted via single continuum of liability, was an appropriate
telephone. At the time of interview, subjects in MM/ way to view the minimum duration criterion. We use
MF2 ranged in age from 20 to 58 years, with a mean the term “narrow GAD” to refer to the full syndrome
of 36.9 years (SD ⫽ 9.1). as defined in DSM-III-R (American Psychiatric Asso-
As outlined previously (Kendler et al., 1992b), in ciation, 1987) and “broad GAD” to refer to a GAD-
the FF sample, zygosity was initially determined by like syndrome that requires anxiety that persists for
a blind review of two experienced twin researchers a minimum duration of 1 or 6 months, the requisite
that utilized standard questions (Eaves et al., 1989) minimum number 6 of 18 associated symptoms, and
and photographs. Blood samples were obtained hierarchical exclusion with MD. Finally, we loos-
from both members of 119 pairs of uncertain zygos- ened the diagnostic criteria in the broad GAD cate-
 GAD IN MEN AND WOMEN 415
gory further by dropping the hierarchy rules with on the liability to GAD in male and female subjects.
MD (“nonhierarchical GAD”). When rg ⫽ 0 or 1, this corresponds to male and
We present information about twin resemblance female subjects sharing none or 100% of genes pre-
in several ways. The odds ratio (OR) is the ratio of disposing to GAD, respectively. Similarly, parameter
the risk of being affected among co-twins of affected rc is the correlation in the common environment on
twins to the risk of being affected among co-twins of the liability. These two parameters cannot be esti-
unaffected twins. The difference in ORs between MZ mated simultaneously with these models.
and DZ twins is assessed by a one-tailed Breslow- Given that the risk to GAD is correlated in twin
Day test (Breslow and Day, 1980), given our direc- pairs, twin studies provide a method of testing for
tional hypothesis of greater MZ resemblance. We the possible impact of differential cooperation. We
also present the tetrachoric correlation, defined as examined this potential ascertainment bias in two
the correlation in members of twin pairs for the ways. First, we tested whether co-twin cooperation
liability to GAD (Falconer, 1965; Pearson, 1901). was predictive of GAD. If noncooperation is posi-
We used a liability-threshold model to estimate tively correlated with the risk for GAD, then the
the genetic and environmental contributions to twin rates of GAD should be higher in twins whose co-
resemblance. For categorical characteristics like twin refused versus cooperated in the study. In ad-
GAD, the estimates are for the resemblance of twins dition, we reanalyzed some of our models by using
in a pair for their liability to develop the disorder the raw data option in Mx, including unpaired twins,
(Falconer, 1965). Liability is assumed to be contin- to see whether their inclusion influenced the results.
uous and normally distributed in the population, The equal environment assumption requires that
with individuals who exceed a theoretical threshold MZ and DZ twin pairs be equally similar in their
expressing the disorder. shared environments that are of etiological rele-
Under the basic twin model, individual differences vance for the trait being studied. If the environments
in liability are assumed to arise from three sources: of DZ twin pairs were less similar than those of MZ
additive genes (abbreviated A), common or familial twin pairs, this could lead to an underestimation of
environment (C), and individual-specific environ- shared environment and a proportional overestima-
ment (E) (see Kendler, 1993, and Neale and Cardon, tion of genetic influence. We investigated this pos-
1992, for more detailed descriptions). In addition, by sibility by assessing both childhood similarity mea-
comparing estimates for A, C, and E in MM and FF sures (Loehlin and Nichols, 1976) and the frequency
pairs, we can examine whether the relative impor- of current contact of the twin pair in adulthood.
tance of genetic or environmental risk factors for Using logistic regression and controlling for zygos-
GAD differ in men and women. Furthermore, by ity, we examined whether the mean level of child-
analyzing all five twin-zygosity groups, we can ad- hood or adult environmental similarity reported by the
dress an additional question: to what extent do the twin pair interacted with the diagnosis of GAD in one
same genetic and/or environmental factors influence twin in predicting the risk for GAD in the co-twin.
GAD in men versus women? The twin models em-
ployed also assume independence and additivity of
the latent variables, absence of assortative mating, Results
equality of shared environmental effects for MZ and
DZ twins, and no age effects. Table 1 exhibits the prevalence data for male-male
We fit the twin models with maximum likelihood (MM), female-female (FF), and male-female (MF)
estimation by using the Mx structural modeling pro- opposite sex twins for the various diagnostic defini-
gram (Neale, 1991). Model fit is evaluated according tions. When looking at members of same-sex pairs
to the principle of parsimony. Models with fewer first, only about 1.5% of male and 2% to 3% of female
parameters are considered preferable if they do not subjects satisfied full criteria (narrow) for 6-month
provide significantly worse fit. We operationalize minimum duration lifetime GAD. Broadening the
parsimony by using the AIC statistic (Akaike, 1987; definition and removing the diagnostic hierarchy
Williams and Holahan, 1994), calculated as the with MD increased the proportion affected suffi-
model ␹2 minus two times the degrees of freedom ciently to allow modeling of the data. Similar in-
(df). We then present parameter estimates from the creases resulted from decreasing the minimum du-
best fitting model, where a2, c2, and e2 equal the ration criterion from 6 months to 1 month. No
proportion of variance in liability due to additive significant differences between MZ and DZ preva-
genetic effects, common environment, and individ- lence were found for the same-sex pairs for any
ual specific environment, respectively. The parame- definition except for nonhierarchical 1-month GAD
ter rg is the correlation in the additive genetic effects in the MM pairs (13.3% vs. 16.2%,␹21 ⫽ 4.41, p ⫽ .036).
416 HETTEMA et al.
TABLE 1
Lifetime Prevalence of GAD
Lifetime Prevalence (%)
Minimum Male-Male Pairs Female-Female Pairs Male-Female Pairs
Duration
GAD Definition (mo) MZ DZ MZ DZ Male Female
Narrow 1 3.3 3.5 6.4 6.3 3.8 8.2
6 1.5 1.4 2.2 2.9 1.7 3.7
Broad 1 9.0 10.6 12.6 13.6 10.3 20.0
6 3.2 3.3 3.9 4.5 4.1 9.1
Nonhierarchical 1 13.3 16.2 19.6 21.5 16.4 29.1
6 5.0 5.4 7.2 7.5 8.2 15.0

TABLE 2
Odds Ratios and Tetrachoric Correlations
Odds Ratio Tetrachoric Correlation
(95% Confidence Interval) (Asymptotic Standard Error)
Male-Male Female-
Minimum Male-Male Pairs Female-Female Pairs Male- Pairs Female Pairs Male-
Duration Female Female
GAD Definition (mo) MZ DZ MZ DZ Pairs MZ DZ MZ DZ Pairs
Broad 1 1.44 0.93 2.19 1.49 1.24 0.10 ⫺0.02 0.23 0.12 0.06
(0.66–3.17) (0.35–2.47) (1.11–4.32) (0.61–3.63) (0.78–1.98) (0.11) (0.13) (0.10) (0.14) (0.07)
6 1.37 — 5.21 1.53 1.60 0.06 — 0.38 0.09 0.11
(0.18–10.6) (1.38–19.6) (0.19–12.5) (0.66–3.89) (0.21) (0.17) (0.24) (0.11)
Nonhierarchical 1 2.18 1.37 3.58 3.84 1.42 0.24 0.10 0.42 0.44 0.12
(1.28–3.72) (0.73–2.57) (2.19–5.86) (2.12–6.97) (1.01–2.01) (0.08) (0.10) (0.08) (0.09) (0.06)
6 1.71 0.74 9.69 3.53 1.62 0.12 ⫺0.06 0.59 0.33 0.14
(0.50–5.85) (0.09–5.68) (4.31–21.8) (1.20–10.4) (0.93–2.81) (0.15) (0.21) (0.10) (0.15) (0.08)

Prevalence differences between opposite sex Breslow-Day test. The OR for male subjects tend to
members of the MF pairs are all significant at the be lower than for female subjects. Similar observa-
p ⫽ .005 level, with female-to-male ratios of about tions apply to the tetrachoric correlations. Control-
2:1. Comparing members of same sex DZ with op- ling for zygosity and age, resemblance between
posite sex pairs, for male twins the only significant twins for GAD was not predicted by similarity of the
difference in prevalence was for the 6-month nonhi- twin pair’s childhood environment or by frequency
erarchical GAD definitions (5.4% vs. 8.2%, ␹21 ⫽7.38, of contact as adults.
p ⫽ .007). However, prevalence for lifetime GAD One way to evaluate whether differences in re-
was significantly higher (p ⬍ .005) in female mem- sults for different definitions of illness (obtained by
bers of MF pairs compared with DZ members of FF varying the minimum duration from 6 months to 1
pairs for all of the GAD definitions except the nar- month) are meaningful is to fit multiple threshold
row 1-month and 6-month definitions. We found no models to the data. These models test whether the
significant relationship between lifetime history of
data can be explained by assuming these discrete
GAD in one twin and interview cooperation status
categories lie on the same underlying continuum of
(agreed/refused) of his or her co-twin after control-
liability to GAD. In this case, the continuum is of
ling for zygosity and age.
Twin resemblance in the form of ORs and tetra- duration, with three categories and two thresholds
choric correlations are shown in Table 2. Dashes subdividing them: unaffected; affected with minimal
indicate that the two-by-two tables used to calculate duration between 1 month and 6 months; or affected
these measures contained a zero cell, rendering the with duration 6 months or greater. If this is found to
statistic ill-defined. The ORs for most of the defini- be consistent with the data, then these multiple
tions exceeded unity (though reaching significance definitions of illness can be collapsed for modeling
only for the broadest definitions), predicting an in- purposes into these categories to increase statistical
creased risk in the co-twin of an affected twin com- power, obtaining results that should be more gener-
pared with that of an unaffected twin. Also, for most ally applicable across duration. For all of the defini-
of the definitions, the OR in MZ pairs exceeded that tions, we were unable to reject the multiple thresh-
seen in DZ pairs. However, the 95% confidence in- old models (p ⬎ .05) and therefore used the multiple
tervals are substantial, with no differences seen by threshold diagnoses in a subset of our analysis.
 GAD IN MEN AND WOMEN 417
TABLE 3
Model Fitting Results
Minimum Model 1 Model 2 Model 3 Model 4 Model 5 Model 7
Duration Fit ACE-ACE AE-AE AE-AE AE-AE AE-AE Model 6 AE-ACEc
GAD Definition (mo) Index (rg ⫽ F, rc ⫽ 1b) (rg ⫽ F) (rg ⫽ 1.0) (rg ⫽ 0) (rg ⫽ 1.0) E-E (rg ⫽ 1)
Broad 1 ␹2 2.56 2.56 2.56 3.34 3.72 8.98
AICa ⫺11.44 ⫺15.44 ⫺17.44 ⫺16.66 ⫺18.28† ⫺13.02
6 ␹2 3.59 3.59 3.59 4.38 5.37 10.71
AIC ⫺10.41 ⫺14.41 ⫺16.41 ⫺15.62 ⫺16.63† ⫺13.29
Threshold ␹2 6.38 6.38 6.38 7.48 7.98 13.38
AIC ⫺21.62 ⫺25.62 ⫺27.62 ⫺26.52 ⫺28.02† ⫺24.62

Nonhierarchical 1 ␹2 5.46 10.19 10.74 14.12 16.29 60.97 6.59

AIC ⫺8.54 ⫺7.81 ⫺9.26 ⫺5.88 ⫺5.71 36.97 ⫺13.41†
6 ␹2 1.33 1.38 1.38 4.11 11.22 33.89 2.95
AIC ⫺12.67 ⫺16.62 ⫺18.62† ⫺15.89 ⫺10.78 9.89 ⫺17.05
Threshold ␹2 15.55 20.88 20.89 24.67 27.44 76.73 16.43
AIC ⫺12.45 ⫺11.12 ⫺13.11 ⫺9.33 ⫺8.56 38.73 ⫺17.57†
a
Akaike’s Information Criterion, where † denotes the best-fit model by AIC.
b
rg ⫽ genetic correlation between genders, rc ⫽ common environment correlation between genders, F means that the parameter was freely
estimated by the model.
c
AE-ACE denotes the male-female latent variables included in the model, with the underline signifying that the A parameters for male and
female subjects are constrained to be equal.

Table 3 displays our model fitting results for the definitions. In model 6, we tested the hypothesis of
various definitions of GAD. Our models assumed no familial resemblance by then setting all additive
three thresholds within a definition: all male sub- genetic parameters to zero. We were able to reject
jects, female members of male-female DZ pairs, and this model by the ␹2 difference test for all GAD
all other female subjects, consistent with observed definitions.
prevalence differences. Model 1 separately esti- For GAD diagnosed without MD hierarchy (bot-
mated A, C, and E in male and female subjects and tom half of Table 3), another level of modeling had
allowed the correlation between the genders for the to be invoked to improve the fit over the full ACE
effects of additive genetic factors to be freely esti- model. Without outlining all of the sequential details
mated. In model 1 only, we assumed that the gen- in Table 3 as was done for the other models, we
ders were perfectly correlated for the effect of com- obtained model 7 by first setting the male common
mon environmental factors influencing the liability environment parameter equal to zero, then con-
to GAD. For all definitions, this model fit well (p ⬎ straining the genetic risk factors in male and female
.34). In model 2, we set all the parameters for com- subjects to be perfectly correlated as in model 3,
mon environment to zero. This resulted in no signif- then finally constraining them to have the same
icant deterioration in fit for the broad definitions magnitude, as in model 5. These simplifications over
and a substantial improvement in the AIC. For the the full ACE model resulted in a minimal deteriora-
nonhierarchical 1-month and threshold models, tion in fit (⌬␹23 ⬵ 1) and a substantial improvement
however, the fit was significantly worse under this in parsimony by AIC for the 1-month and threshold
assumption of no common environmental source of definitions.
variance. For the next steps, we restricted our dis- Table 4 shows the parameter estimates for both
cussion to the broad (hierarchical) definitions (up- the full ACE model and the best-fitting models for
per portion of Table 3). Models 3 and 4 obtain from each of the definitions obtained from Table 3. The
model 2 by constraining the genetic risk factors for best-fitting models estimated that the heritability of
male and female subjects to be either perfectly cor- broad (hierarchical) lifetime GAD ranges from 14%
related (rg ⫽ 1) or uncorrelated (rg ⫽ 0), respec- to 21% in both male and female subjects. The re-
tively. Model 3 fit as well as model 2 and better than maining approximately 80% of the variance in liabil-
model 4. Working then from model 3, we con- ity was attributed to individual specific environment,
strained the magnitude of the impact of the genetic with no role for common familial environment. The
risk factors to be the same in male and female genetic correlation (rg) between male and female sub-
subjects, obtaining model 5. This produced improve- jects was estimated to be 1.0.
ments in AIC due to increased parsimony, with little For the nonhierarchical definitions of GAD, the
deterioration in fit over model 3, and proved to be best-fitting models again estimated the heritability
the best-fitting model by AIC for the broad GAD to be the same in both male and female subjects and
418 HETTEMA et al.
TABLE 4
Parameter Estimates
Parameter Estimates (95% Confidence Interval)
Minimum Male Female
Duration
GAD Definition (mo) Model a2 c2 e2 a2 c2 e2 a’2
Broad 1 Full .08 (.00–.28) 0.0 (.00–.21) .92 (.72–1.0) .21 (.00–.41) 0.0 (.00–.33) .77 (.58–.96) .02 (.00–.40)
Best-fit .15 (.02–.27) — .85 (.73–.98) .15 (.02–.27) — .85 (.73–.98) —
6 Full .07 (.00–.39) 0.0 (.00–.30) .93 (.61–1.0) .37 (.00–.64) 0.0 (.00–.52) .63 (.36–.95) 0.0 (.00–.51)
Best-fit .21 (.00–.42) — .79 (.58–1.0) .21 (.00–.42) — .79 (.58–1.0) —
Threshold Full .06 (.00–.25) 0.0 (.00–.18) .94 (.75–1.0) .22 (.00–.41) .01 (.00–.33) .77 (.59–.96) 0.0 (.00–.39)
Best-fit .14 (.02–.27) — .86 (.73–.98) .14 (.02–.27) — .86 (.73–.98) —

Nonhierarchical 1 Full .21 (.00–.38) .02 (.00–.28) .77 (.62–.93) .01 (.00–.43) .42 (.00–.54) .57 (.44–.69) 0.0 (.00–.49)
Best-fit .22 (.09–.34) — .78 (.66–.91) .22 (.09–.34) .25 (.09–.40) .53 (.42–.66) —
6 Full .10 (.00–.36) 0.0 (.00–.28) .90 (.64–1.0) .53 (.00–.76) .07 (.00–.60) .40 (.24–.61) 0.0 (.00–.70)
Best-fit .11 (.00–.32) — .89 (.68–1.0) .60 (.41–.76) — .40 (.24–.59) —
Threshold Full .21 (.00–.37) .01 (.00–.17) .77 (.62–.93) .03 (.00–.42) .40 (.04–.52) .57 (.44–.69) 0.0 (.00–.41)
Best-fit .21 (.09–.33) — .79 (.67–.91) .21 (.09–.33) .25 (.10–.39) .54 (.43–.65) —

of similar magnitude (11% to 22%), except for nonhi- dence intervals displayed in Table 4. Therefore,
erarchical 6-month GAD, which contains a value for although the genetic correlation between men and
the female additive genetic variance that signifi- women rg was estimated to be unity, its uncertainty
cantly exceeds that of the male subjects in the best- is large (95% confidence interval: .0 to 1.0).
fitting model. In the 1-month and threshold defini- Our models for hierarchically defined GAD sug-
tions, the only difference lies in the possible added gest that genetic factors alone (i.e., no common
significance for common environment to explain environment) are responsible for twin resemblance
about 25% of the remainder of the variance for the in both genders, whereas two of the nonhierarchical
female subjects that had been attributed entirely to models reflect roles for additive genetics alone for
individual specific environment in the hierarchical the male subjects and both genetics and common
definitions of GAD. environment for the female subjects. An examina-
Although not elaborated upon in detail here, we tion of the 95% confidence intervals of our full model
also tested models that included a nonadditive parameters in Table 4 suggests that we have insuf-
source of genetic variance. In all cases, its potential ficient power to reject one source of familial resem-
contribution to the total variance was small (less blance in favor of the other. In our previous analysis
than 5%) and not statistically significant. of GAD in these female-female twins, the best fit
models of two of nine definitions (6-month GAD
Discussion with and without hierarchy with MD) had common
In this study, we found evidence for the role of environment responsible for twin resemblance, sug-
genetics in the etiology of GAD in both women and gesting that in that analysis as well as this one, this
men, consistent with results obtained in our previ- inconsistency is most likely due to lack of power.
ous analysis of female-female twin pairs and other This is a ubiquitous problem in twin analysis of
twin studies of GAD. Model fitting results were con- categorical variables (Neale et al., 1994). If this were
sistent with genetic factors alone explaining twin a real effect, that is, if only GAD with comorbid MD
resemblance in GAD diagnosed with hierarchical required an etiological role for common family en-
exclusion of MD. When hierarchical exclusion rules vironment, this is inconsistent with other data from
were relaxed, two of three models tested produced this population, which show that common environ-
results that suggested roles for both genetics and ment plays little or no role in the liability to MD, with
common environment, at least in the female sub- or without GAD (Kendler et al., 1992c; Kendler and
jects. The magnitude of the role genetics plays was Prescott, 1999). One way to improve our ability to
found to be equal in men and women in five of six of differentiate these competing sources of resem-
the models, in the modest effect range of 15% to 20%. blance and reduce measurement error that inflates
Our results are consistent with the hypothesis that individual environmental variance at the expense of
men and women share all of the genes that predis- genetic or common environment is by applying a
pose to GAD. However, the power to detect a gen- measurement model. This incorporates information
der-specific factor a’ was low, as seen by the confi- from multiple, sequential assessments and includes
 GAD IN MEN AND WOMEN 419
measurement error as an additional variance source about three, similar to the approximately fivefold
separately from the ones of etiological interest. This change (9.1% to 45%) reported by Breslau and Davis
has been successfully applied to data on major de- (1985) in their geographically based sample of 357
pression (Kendler et al., 1993) and phobias (Kendler mothers. Examination of opposite sex pairs showed
et al., 1999) for our FF sample, and we plan to apply a significant increase in prevalence in female mem-
this model to our GAD data in the future. bers of MF pairs compared with DZ members of FF
A slightly higher heritability estimate (20% to 30%, pairs for some GAD definitions. Andrews et al.
depending on diagnostic definition) was obtained in (1990) found a similar increase (29% vs.16%). This
our previous study of female-female twins (Kendler raises the question of what impact gender might
et al., 1992a). Scherrer et al. (2000) recently pub- have on the psychiatric diagnosis of one’s co-twin.
lished a heritability estimate of 37.2% (95% CI: 27.7% The results of this analysis should be interpreted
to 46.1%) in the male-male twins of the Vietnam Era in the context of several potential limitations. First,
Twin Registry for a broadly defined GAD similar to despite the size of our sample, our modeling efforts
our 1-month, nonhierarchical definition. We are were limited by insufficient statistical power due to
aware of only one other twin study that compared the low prevalence of strictly defined GAD. To over-
the genetic liability for GAD between male and fe- come this difficulty, we broadened both the diagnos-
male subjects (Roy et al., 1995). That bivariate study tic and duration criteria and utilized multiple thresh-
of GAD and major depression also supported the old testing, producing overall results that are
hypotheses that the genetic factors that predispose supportive of GAD as a syndrome with partially
to GAD are of similar magnitude in male and female genetic etiology. However, their applicability to
subjects and are completely correlated. It estimated whatever the “true” GAD phenotype is may be ques-
the heritability of GAD at 14.3% for their narrow tionable, particularly in light of its continually evolv-
definition and 49% for their broad definition. Both of ing diagnostic definitions (Wolk et al., 1996). To look
these were similar to our 1-month nonhierarchical at the impact of revisions in diagnostic criteria, we
definition but differed from each other by the cer- also constructed DSM-IV-based GAD definitions cor-
tainty of diagnosis. responding to those analyzed herein, using instead
Our modeling results were, within the limitations the revised minimum of three of six associated
of our statistical resolution, unchanged by reducing symptoms. The modeling results were essentially
the minimum duration from 6 months to 1 month. the same.
This agrees with the findings of Bienvenu et al. Second, our analyses were limited to complete
(1998), who examined five separate syndromal cat- twin pairs. As discussed in previous sections, we
egories of generalized anxiety in a follow-up of the saw no evidence for cooperation bias. In addition,
Baltimore cohort of the ECA. They concluded that we reexamined some of our models using the raw
the main differentiating factor that produced a par- data option in Mx, allowing us to include unpaired
ticular epidemiological profile was requiring the twins in the analysis and found no difference in
presence of six associated symptoms (DSM-III-R) results.
and that neither the nature of the subjects’ worries Third, female members of opposite sex DZ twin
nor the duration of symptoms influenced this. Also, pairs had a higher prevalence for GAD than other
in a previous analysis of this sample of FF twins, we female twins. It is unclear whether this difference is
found that only the total number of associated symp- real or methodological, for example, as a result of
toms and comorbidity with bulimia in a proband multiple sampling history of the female-female pairs
twin with GAD uniquely predicted risk of GAD in her compared with the opposite sex pairs. Regardless,
co-twin (Kendler et al., 1995). because we sampled the twins independently of
For this sample, we found that about 1.5% of male their psychiatric status providing “built in” control
and 2% to 3% of female subjects met the full criteria groups, such differences in prevalence do not inval-
for lifetime DSM-III-R (6-month, narrow). When hi- idate our results when incorporated into our analy-
erarchical exclusion of MD was relaxed, this in- ses, as we have done.
creased to 5% to 7% of male and 7% to 8% of female Fourth, lifetime history of GAD was assessed at
subjects. This is consistent with Judd et al.’s (1998) one time-point, which potentially confounds effects
review of the ECA and NCS data, in which they of individual specific environment and measurement
conclude that, although the lifetime prevalence of error, reducing the corresponding estimates of her-
GAD is 4% to 7%, only about one-third of those are itability. We have found, for MD in our FF sample,
“pure,” with the majority having comorbid MD. De- that improving the diagnostic reliability by reducing
creasing the minimum duration from 6 months to 1 error via multiple, sequential assessments increased
month increased our prevalence rates by a factor of the heritability estimate substantially (Kendler et al.,
420 HETTEMA et al.

1993). As mentioned already, we plan to apply this Kendler KS, Neale MC, Kessler RC, Heath AC, Eaves LJ (1992b)
measurement model to our data in the future. A population-based twin study of major depression in women:
The impact of varying definitions of illness. Arch Gen Psychi-
Fifth, because the sample was made up entirely of atry 49:257–266.
Caucasians, the results may not generalize to other Kendler KS, Neale MC, Kessler RC, Heath AC, Eaves LJ (1992c)
ethnic groups. Major depression and generalized anxiety disorder: Same
genes, (partly) different environments? Arch Gen Psychiatry
49:716 –722.
Conclusions Kendler KS, Neale MC, Kessler RC, Heath AC, Eaves LJ (1993)
The lifetime history of major depression in women: reliability
This study in a large population-based sample of of diagnosis and heritability. Arch Gen Psychiatry 50:863– 870.
male and female twins supports the findings of past Kendler KS, Neale MC, Kessler RC, Heath AC, Eaves LJ (1995)
Clinical characteristics of familial generalized anxiety disor-
twin and family studies of GAD, that there is modest der. Anxiety 1:186 –195.
familial aggregation of GAD and that genetic factors Kendler KS, Prescott CA (1999) A population-based twin study of
are largely responsible for resemblance in relatives. lifetime major depression in men and women. Arch Gen Psy-
chiatry 56:39 – 44.
In addition, the relative importance of genetics in Loehlin JC, Nichols RC (1976) Heredity, environment and per-
the etiology of GAD was similar in male and female sonality: A study of 850 sets of twins. Austin, TX: University
subjects, in the range of 15% to 20%, with male and of Texas Press.
female subjects sharing 100% of the genes for GAD. Mendlewicz J, Papadimitriou GN, Wilmotte J (1993) Family study
of panic disorder: comparison with generalized anxiety disor-
In this sample, we found an inconsistent role for der, major depression and normal subjects. Psychiatr Genet
common family environment in the female subjects 3:73–78.
that may account for another 25% of the variance Neale MC (1991) Statistical modeling with Mx. Richmond, VA:
Department of Psychiatry, Medical College of Virginia of Vir-
when GAD is diagnosed without hierarchy under ginia Commonwealth University.
major depression. Further work to clarify this incon- Neale MC, Cardon LR (1992) Methodology for genetic studies of
sistency is warranted. twins and families. Dordrecht, The Netherlands: Kluwer Ac-
ademic.
Neale MC, Eaves LJ, Kendler KS (1994) The power of the clas-
References sical twin study to resolve variation in threshold traits. Behav
Genet 24:239 –258.
Akaike H (1987) Factor analysis and AIC. Psychometrika 52:317– Noyes R, Clarkson C, Crowe RR, Yates WR, McChesney CM
332. (1987) A family study of generalized anxiety disorder. Am J
Amreican Psychiatric Association (1987) Diagnostic and statis- Psychiatry 144:1019 –1024.
tical manual of mental disorders (rev, 3rd ed). Washington, Pearson K (1901) Mathematical contributions to the theory of
DC: Author. evolution. VIII. On the correlation of characters not quantita-
Andrews G, Stewart G, Allen R, Henderson AS (1990) The genet- tively measurable. Proc R Soc 66:241–244.
ics of six neurotic disorders: a twin study. J Affect Disord Roy MA, Neale MC, Pedersen NL, Mathe AA, Kendler KS (1995)
19:23–29. A twin study of generalized anxiety disorder and major depres-
Bienvenu OJ, Nestadt G, Eaton WW (1998) Characterizing gen- sion. Psychol Med 25:1037–1049.
eralized anxiety: Temporal and symptomatic thresholds. Scherrer JF, True WR, Xian H, Lyons MJ, Eisen SA, Goldberg J,
J Nerv Ment Dis 186:51–56. Lin N, Tsuang MT (2000) Evidence for genetic influences
Breslau N, Davis GC (1985) DSM-III generalized anxiety disorder: common and specific to symptoms of generalized anxiety and
An empirical investigation of more stringent criteria. Psychi-
panic. J Affect Disord 57:25–35.
atry Res 14:231–238.
Skre I, Onstad S, Torgersen S, Lygren S, Kringlen E (1993) A twin
Breslow NE, Day NE (1980) Statistical methods in cancer re-
search (Vol 1). Lyon: International Agency for Research on study of DSM-III-R anxiety disorders. Acta Psychiatr Scand
Cancer. 88:85–92.
Eaves LJ, Eysenck HJ, Martin NG, Jardine R, Heath AC, Feingold Spence JE, Corey LA, Nance WE, Marazita ML, Kendler KS,
L, Young PA, Kendler KS (1989) Genes, culture and personal- Schieken RM (1988) Molecular analysis of twin zygosity using
ity: An empirical approach. London: Academic Press. VNTR DNA probes (Abstract). Am J Hum Genet 43(3):A159.
Falconer DS (1965) The inheritance of liability to certain dis- Spitzer RL, Endicott J, Robins E (1975) Research diagnostic
eases, estimated from the incidence among relatives. Ann criteria for a selected group of functional disorders. New
Hum Genet 29:51–76. York: New York Psychiatric Institute.
Judd LL, Kessler RC, Zeller PV, Wittchen H-U, Kunovac JL (1998) Spitzer RL, Williams JBW (1985) Structured Clinical Interview
Comorbidity as a fundamental feature of generalized anxiety for DSM-III-R (SCID). New York: Biometrics Research De-
disorders: Results from the National Comorbidity Study partment, New York State Psychiatric Institute.
(NCS). Acta Psychiatr Scand 98:6 –11. Torgersen S (1983) Genetic factors in anxiety disorders. Arch
Kendler KS (1993) Twin studies of psychiatric illness: Current Gen Psychiatry 40:1085–1089.
status and future directions. Arch Gen Psychiatry 50:905–915. Williams LJ, Holahan PJ (1994) Parsimony-based fit indices for
Kendler KS, Karkowski LM, Prescott CA (1999) Fears and pho- multiple-indicator models: Do they work? Structural Equation
bias: Reliability and heritability. Psychol Med 29:539 –553. Modeling 1:161–189.
Kendler KS, Neale MC, Kessler RC, Heath AC, Eaves LJ (1992a) Wolk SI, Horwath E, Goldstein RB, Wickramaratne P, Weissman
Generalized anxiety disorder in women: A population based MM (1996) Comparison of RDC, DSM-III, DSM-III-R diagnostic
twin study. Arch Gen Psychiatry 49:267–272. criteria for generalized anxiety disorder. Anxiety 2:71–79.