Anxiety disorders are common; lifetime prevalence for the group of disorders is estimated to be as high as 25% (1). They are highly comorbid with each other and with other psychiatric conditions (2), particularly mood disorders (3). Anxiety disorders carry a substantial burden of distress and impairment that is comparable to that of chronic "medical" disorders, such as diabetes (4–8). As such, their etiology has been, and continues to be, a major target of investigation.
A long history of primary studies has sought to determine whether anxiety disorders are familial and to estimate their heritability. A number of excellent, detailed reviews of the genetic epidemiology of anxiety disorders have also appeared (e.g., reference 9). Much knowledge has been amassed in this area; however, to date, there has been no attempt to summarize the information quantitatively. Therefore, a meta-analytic treatment of the data is timely, particularly as many research groups have begun to search for "anxiety genes."
We have attempted to answer, by means of a meta-analysis of selected epidemiological studies, the following two questions for major anxiety disorders: 1) what is the magnitude of the familial aggregation of anxiety disorders? 2) what is the relative contribution of genetics and environment to their etiology?
We applied this method to the following anxiety disorders: panic disorder, generalized anxiety disorder, phobias, and obsessive-compulsive disorder (OCD). We were limited to family and twin studies, as we know of no available adoption studies of anxiety disorders.
The basic goals of meta-analysis, as applied in this study to the genetic epidemiology of anxiety disorders, are 1) to determine the consistency between differing results by testing for heterogeneity across studies and 2) to combine data from multiple primary studies for the purpose of synthesizing the information, thus providing more reliable, precise, and less biased aggregate estimates of familial risk and heritability.
We performed a keyword-driven MEDLINE search to identify all potential primary studies for inclusion in our analyses. In addition, we searched the reference sections of these manuscripts and existing reviews for additional sources. We limited our analyses to studies involving adult subjects that estimated the risk in relatives for the same anxiety disorder as that diagnosed in the proband. Although some of the studies included in this analysis examined the question of familial and genetic sources of the comorbidity of anxiety disorders, this important issue is outside of the scope of this study.
We used the following inclusion criteria for the studies: 1) use of operationalized diagnostic criteria, to the exclusion of studies assessing only anxiety symptoms and older studies of anxiety neuroses; 2) systematic ascertainment of probands and relatives; 3) direct interviews with the majority of subjects, to the exclusion of family history studies; 4) diagnostic assessment of relatives performed with investigators blind to proband affection status; and 5) for family studies, inclusion of a comparison group. For some disorders, in which fewer studies were available, we relaxed the criterion for blind ascertainment of diagnosis only to increase our data pool for comparison. Many primary studies identified in the literature were rejected for failing one or more of these criteria. In particular, twin studies meeting the inclusion criteria were limited to, at most, one or two large independent samples per disorder.
Family studies may be thought of as a type of case-control study design, in which the patients are affected probands and the comparison subjects are individuals without a history of the disorder (10). The measure of interest is the binary outcome of a relative diagnosed as affected or unaffected, which allows construction of 2×2 contingency tables of affected versus unaffected relatives of patients and comparison subjects. We reported the odds ratio as the measure of association, owing to its desirable statistical property of invariance under study design (10) and greater familiarity to readers than, e.g., the tetrachoric correlation, which is less sensitive to prevalence (11). In addition, stratifying by study allows one to test for homogeneity across studies by means of the Breslow-Day statistic (12) and to calculate a summary odds ratio by means of the Mantel-Haenszel method (13). The SAS routine proc freq (14) was used to make these calculations. For panic disorder and OCD, for which the most data exist, we also calculated an aggregate risk by combining raw data from each of the studies.
For the twin studies, we were limited by the scarcity of studies that met our inclusion criteria; only two large twin samples that met our criteria and were used to study anxiety disorders exist: the Virginia Twin Registry (15) and the Vietnam Era Twin Registry (16). Data from the Virginia Twin Registry have been reported regarding panic disorder, generalized anxiety disorder, and phobias; data from the Vietnam Era Twin Registry have been published on panic disorder and generalized anxiety disorder. Therefore, we attempted meta-analysis using twin data for panic disorder and generalized anxiety disorder only, but we tabulated twin data for the other disorders where it was available.
We report several measures of twin resemblance, depending on availability. The proband-wise concordance is the proportion of affected co-twins of affected index twins. The tetrachoric correlation is between the members of a twin pair regarding their liability for the disorder; it assumes an underlying continuous trait with a threshold that distinguishes between unaffected and affected twins. Structural equation modeling provides estimates of variance in liability to a disorder that are attributable to additive genetic (a2), common familial environmental (c2), and individual-specific environmental (e2) factors (17).
For panic disorder and generalized anxiety disorder, we performed structural equation modeling by combining data from the two large twin studies using the statistical package Mx (18). To this end, we created 2×2 contingency tables of affection status for twin 1 versus twin 2 from the reported sample parameters for each study. The output of the modeling provided the fit of the model and the component estimates of variance. The fit of submodels created by placing limiting assumptions on the model (such as testing the significance of one of the variance components) was then compared to that of the full (saturated) model. Data from the Vietnam Era Twin Registry come from same-sex male twin pairs, whereas data from the Virginia Twin Registry come from same-sex female twins for panic disorder and male-male, female-female, and opposite-sex twin pairs in the most recent assessment of generalized anxiety disorder. Therefore, the most saturated models allowed for differing proportions of genetic and environmental variance between men and women, whereas the submodels tested the hypothesis that these were the same for the two genders. In setting up the model, significant differences in prevalence in the various groups required that different thresholds be assigned.
Finally, for panic disorder and generalized anxiety disorder, we attempted to perform an overall structural equation modeling of the combined twin and family data. In comparison to the population-based twin data, data from all of the family studies were from clinically ascertained probands. Therefore, some restrictive assumptions had to be made in order to combine them. Owing to differing prevalence estimates among the family studies that were obtained by use of unequal comparison groups, a threshold for each family study was not calculated from the sample characteristics as it was for the twin data, but, rather, it was assigned from a base prevalence rate obtained from National Comorbidity Survey data (1). In addition, because the family study data were not reported separately by gender or type of relative, we had to make the rather broad assumption that all of the first-degree relatives in these studies could be modeled as if they had the same genetic and environmental correlations as did their siblings (or dizygotic twins). We tested the validity of this assumption by including a special twin environmental parameter that differentiated between dizygotic resemblance in twin pairs and all first-degree relatives in the family studies. We also had to test the hypothesis that model parameters from clinical and community samples could be equated to obtain aggregate estimates. Since all of the family studies used clinical study groups and the larger twin studies were community-based, it was not possible to test these last two assumptions separately. For several studies, partially or wholly overlapping samples were studied in more than one report. In most cases, we chose the studies that contained the largest sample sizes or the most recent results.
Five family studies of panic disorder, all from clinical populations, met our criteria for inclusion (t1). Studies excluded from our analysis are cited in the References (29–32). (For detailed reviews, refer to references 33–36). All five studies support the familial aggregation of panic disorder. Except for the study by Mendlewicz et al. (20), all of the odds ratios are similar; in no instance do they significantly differ from each other. The results of our meta-analysis show that there is a significant association between panic disorder in the probands and panic disorder in the first-degree relatives (Mantel-Haenszel statistic=47.8, df=1, p<0.0001); the hypothesis that the odds ratios for the five studies were homogeneous could not be rejected (Breslow-Day χ2=2.69, df=4, p=0.61). The summary odds ratio across the five studies was 5.0 (95% confidence interval [CI]=3.0–8.2), which strongly supports a familial component in liability to panic disorder. The unadjusted aggregate risk based on 1,356 total first-degree relatives of panic disorder probands was 10.0%, compared to 2.1% in 1,187 comparison relatives.
Three twin studies of panic disorder met our inclusion criteria (t2). We also tabulated two other (nonblind) studies but did not include them in the meta-analysis. The first three smaller studies support a role for genetic factors in the etiology of panic disorder. The two larger twin studies, performed on different samples and opposite genders, are consistent with each other in attributing 30%–40% of the variance in liability to additive genetics; the remaining variance in liability comes from individual-specific environments. The twin studies do not support a role for common family environment in the etiology of panic disorder. Combining the three blinded studies by means of structural equation modeling in Mx produced a best-fitting model that included only additive genetics and individual environmental factors, which is consistent with the results of the primary studies. In addition, we could constrain the genetic variance to be equal across the studies without a significant deterioration in model fit (Δχ2=3.43, df=2, p>0.10), which suggests that genes affect panic disorder similarly in men and women. The parameters of the best-fitting model are provided at the bottom of t2.
We then tested the broader model, combining the family study data with the twin data. We found that the special twin environmental parameter that differentiated between dizygotic twins and other first-degree relatives could be made equal to zero with no degradation in fit (Δχ2=0.003, df=1, p=0.95), which allowed us to equate parameters across studies. This combined data produced a best-fitting model that required only additive genetics and individual environment to account for liability to panic disorder, with a heritability estimate of 0.48 (95% CI=0.41–0.54).
Generalized Anxiety Disorder
Only one family study of generalized anxiety disorder met all of our inclusion criteria (20). In order to expand our sample, we included one additional study, that of Noyes et al. (40), which met all criteria except blind assessment. As indicated in t3, both studies supported the familial aggregation of generalized anxiety disorder, with an odds ratio greater than 1, and together they showed a significant association between generalized anxiety disorder in the probands and in their first-degree relatives (Mantel-Haenszel statistic=19.1, df=1, p<0.0001). A Breslow-Day test supported homogeneity of the odds ratio (Breslow-Day χ2=0.69, df=4, p=0.61), and we calculated the Mantel-Haenszel summary odds ratio as 6.1 (95% CI=2.5–14.9), which supports familial aggregation of generalized anxiety disorder.
Two large twin studies met our inclusion criteria (16, 41), and we included one smaller controlled nonblinded study (30) in t4. The twin studies excluded are listed in the References (43, 44). The excluded study by Roy et al. (44) reported a bivariate analysis of major depression and generalized anxiety disorder in a sample of 1,484 male and female twins originally ascertained for affective illness from a combination of clinical and population sources. Although it met our inclusion criteria, because this study employed a complex ascertainment scheme, we chose not to use this data in our meta-analysis. Similar to what was done with the twin studies regarding panic disorder, we combined the two larger studies of generalized anxiety disorder by means of structural equation modeling. For consistency with the Scherrer et al. study (16), we used the definition of generalized anxiety disorder of 1 months’ minimum duration from a report by Hettema et al. (41) without diagnostic hierarchy. The best-fitting model predicted that 31.6% (95% CI=24%–39%) of the variance for liability to generalized anxiety disorder was attributable to additive genetics in both genders and that the same genes predispose men and women to generalized anxiety disorder. The model also estimated a small role for common familial environment in the women only, with the remaining variance due to individual-specific environment. Although not included in the meta-analysis, the study by Roy et al. (44) estimated an additive genetic variance of 14.3% and 49% for their narrow and broad definitions of generalized anxiety disorder, respectively, no role for common environmental factors, and no gender differences. Since we could not equate the model parameters across genders as was possible for panic disorder, we were unable to combine the twin and family data into one model.
Although only a few family and twin studies of generalized anxiety disorder exist that were appropriate for our analyses, they provided convincing evidence for familial aggregation. In addition, the twin studies suggested at least a modest role of genetics in familial aggregation and an uncertain role of common environment. (For a review of genetic studies of generalized anxiety disorder, see Woodman .)
For the purposes of this meta-analysis, specific phobias, social phobia, generalized social phobia, and agoraphobia were grouped together owing to the scarcity of these studies in each of the individual categories. Four family studies of phobias met our criteria for inclusion in this analysis; they are listed in t5. All studies support the familial aggregation of phobias. Except for the Stein et al. study (47), all of the odds ratios are similar, and they are statistically consistent with each other. The results of our meta-analysis show that there is a significant association between phobias in probands and in their first-degree relatives (Mantel-Haenszel statistic=52.3, df=1, p<0.0001), and the hypothesis that the odds ratios for the five studies were homogeneous could not be rejected (Breslow-Day χ2=4.03, df=4, p=0.40). The summary odds ratio across the studies was 4.1 (95% CI=2.7–6.1), which strongly supports a familial risk for phobic disorders. (The family studies reviewed that did not meet the inclusion criteria are references 32, 50, and 51.)
The only twin studies of phobias that met our inclusion criteria were reported by our group for the Virginia Twin Registry sample, which was described earlier. Included in t6 are the results from a report by Carey and Gottesman (52) of a nonblinded assessment of 49 twin pairs from the Maudsley Twin Registry made up of twins hospitalized for obsessional neurosis, obsessional personality disorder, or phobic neurosis. It is not clear whether DSM-III criteria for phobias or OCD were used in the study, and the authors provided concordance rates for various levels of treatment criteria. We chose to include their data for twins who had a treatment episode involving phobias (t6), as this appeared to be the definition most consistent with a DSM phobic disorder. (The twin studies we excluded are references 43, 56–60.)
The following discussion focuses on data from the Virginia Twin Registry. Kendler et al. (53) reported best-fitting models for social phobia, agoraphobia, and animal phobias that suggested additive genetics is responsible for twin resemblance in women and that there is no role for common environment. In situational phobias and phobias involving blood, needles, hospitals, or illness, reported in Neale et al. (54), the best-fitting models suggested just the opposite. This may not represent a real difference, however, since there may not have been sufficient power to differentiate between these two sources of variance (61). In fact, in a subsequent analysis that included measurement reliability from two assessments 8 years apart, the models predicted that twin resemblance was due solely to additive genetics for all but animal phobias and that when fears and phobias were combined by using a multiple threshold model, additive genetics alone accounted for twin resemblance in all of the phobic categories (62). We recently obtained similar results in a twin sample of men; the results are displayed in the bottom of t6(55).
Although we know of only one large adult twin sample in which phobia diagnoses were examined, it predicted that additive genetics is largely responsible for the familial aggregation seen in the family studies, which represents roughly 20%–40% of the variance seen, depending on the type of phobia. Its results are supported by a large population-based adolescent twin study of social phobia, major depression, and alcoholism (63), in which additive genetics accounted for 28% of the variance in risk for social phobia, with the remainder due to nonshared environment. (Refer to Carey  for a general review of phobias and Fyer  for a review of social phobia.)
There were four family studies of OCD that met our inclusion criteria, plus one more that we included in t7, although its assessment of relatives was not blind to proband status. Separately, the family study data provided mixed support for the familial aggregation of OCD. Taken together, however, there was a significant association between OCD in the probands and OCD in their first-degree relatives (Mantel-Haenszel statistic=25.1, df=1, p<0.0001). In addition, the hypothesis that the odds ratios across the five studies were homogeneous could not be rejected (Breslow-Day χ2=4.48, df=4, p=0.34), which allowed us to calculate the Mantel-Haenszel summary odds ratio across the five studies, 4.0 (95% CI=2.2–7.1), which lent further support for the familial aggregation of OCD. The unadjusted aggregate risk based on 1,209 total first-degree relatives of the probands with OCD was 8.2%, compared to 2.0% in 746 comparison relatives. The family studies not meeting our inclusion criteria are cited in the References (71–84).
We could not identify any twin studies of OCD that met our inclusion criteria. The study by Carey and Gottesman (52), mentioned in the Phobias section of this report, included twin concordance rates for treatment of obsessive and compulsive symptoms. Other twin studies of obsessive-compulsive symptoms, obsessive-compulsive neurosis, or OCD were found (30, 39, 43, 85–89). Although these studies provided insight into the potential genetic basis of variables that are related to the OCD phenotype, speculations on the common genetic basis between these variables and OCD are outside of the scope of this article. (For reviews of the genetics of OCD, see references 68, 90–93.)
The results of this meta-analysis strongly support the following conclusions:
1. Panic disorder, generalized anxiety disorder, phobias, and OCD aggregate in families; the strongest evidence exists for panic disorder, for which the most data exist. The summary odds ratios are similar across the disorders, ranging from about 4 to 6.
2. The major source of familial risk is genetic. This is supported by two large-scale twin studies of panic disorder and generalized anxiety disorder and one of phobias. Since we know of no adoption studies on anxiety disorders, twin studies are the only available means of differentiating potential genetic versus common familial environmental etiologies for their familial aggregation. The models used to proportion the variance in liability into either genetic or familial environmental sources are based on the assumption that greater intrapair resemblance of monozygotic versus dizygotic twins for a phenotype is due to their greater genetic resemblance rather than any greater similarity of their environment, also known as the equal-environments assumption. Our studies from the Virginia Twin Registry included examination of various measures of common environment as possible sources of violation of this assumption; we came to consistently negative conclusions. Also, there have been several direct investigations of the equal-environments assumption in twin studies supportive of its validity (94–96).
3. Estimated heritabilities across the disorders are in the modest range, 30%–40%, significantly lower than for disorders such as schizophrenia and bipolar disorder. This leaves the largest proportion of the variance in liability to be mostly explained by individual environmental factors. This value, however, most likely represents an underestimation of true heritability. This is owing to the fact that measurement error (including diagnostic reliability and stability) and gene-environment interactions not accounted for in the models lead to inflation of estimates of the impact of specific environment. For example, our group recently performed an analysis of phobias that took into account measurement error by including data from two assessments of the same sample 8 years apart (62). The resulting estimated heritabilities across the phobias were in the 50%–60% range, compared to the point estimates of 30%–40% displayed in t6.
We limited our analysis to studies involving anxiety disorders. The relationship between anxiety symptoms seen in unselected individuals compared to those affected with an anxiety disorder is unclear. If symptoms and disorders exist on a single quantitative continuum based on number, severity, duration, and the like, then studies of anxiety symptoms may be informative regarding the genetics of anxiety disorders. Twin studies have indeed found genetic contributions to nonspecific anxiety symptoms (97), fears (53, 54, 56, 57, 60, 62, 98–100), symptoms of panic-phobia (101), and obsessions and compulsions (52, 88, 102).
There are several potential limitations on the interpretation of these results. First, we may not have included all primary studies, owing either to their unavailability in the literature or to our exclusion criteria. It is unclear what impact this choice of quality versus quantity of information had on our summary statistics, but we predict that correlations between relatives would be diluted by using less rigorous standards, lowering estimates of familial aggregation and heritability. Second, because of the scarcity of twin studies for phobias and OCD, the validity and generalizability of the data for these disorders are limited compared to data for panic disorder and generalized anxiety disorder. Third, the relative contribution of genetic verses family environment is based on twin studies only, as adoption studies were not available and family studies do not provide these estimates. Finally, as previously discussed, since our genetic information is limited to twin samples, any undetected violations of the equal-environments assumption might lower the heritability estimates from those presented.
Received Nov. 22, 2000; revision received March 22, 2001; accepted April 26, 2001. From the Virginia Institute for Psychiatric and Behavioral Genetics, Departments of Psychiatry and Human Genetics, Medical College of Virginia, Virginia Commonwealth University. Address reprint requests to Dr. Kendler, Virginia Institute for Psychiatric and Behavioral Genetics, Department of Psychiatry, P.O. Box 980126, Richmond, VA 23298-0126; firstname.lastname@example.org (e-mail). Funded by NIMH Training Fellowship grant MH-20030 (to Dr. Hettema), grant MH-01458 (to Dr. Neale), and grant MH-49492 (to Dr. Kendler).