Introduction
Lifetime major depression (MD), when assessed at one occasion during adulthood, has a heritability of approximately 40% (Sullivan et al. Reference Sullivan, Neale and Kendler2000; Kendler et al. Reference Kendler, Gatz, Gardner and Pedersen2006). When lifetime MD is evaluated twice several years apart (Kendler et al. Reference Kendler, Gardner, Neale and Prescott2001a ), the latent heritability increases substantially. This is because unreliability of measurement or the impact of short-term environmental risks (which cannot easily be distinguished) set an upper limit on heritability. If such ‘errors’ are substantial, measuring MD more than once and modeling MD as the latent or stable liability assessed at multiple time points would be expected, by standard statistical theory (Falconer, Reference Falconer1989), to increase heritability estimates.
Indeed, questions have been raised about the validity of the lifetime assessment for MD (Coughlin, Reference Coughlin1990; Andrews et al. Reference Andrews, Anstey, Brodaty, Issakidis and Luscombe1999; Patten, Reference Patten2003; Wells & Horwood, Reference Wells and Horwood2004; Kruijshaar et al. Reference Kruijshaar, Barendregt, Vos, de, Spijker and Andrews2005). Many episodes are subject to forgetting (Moffitt et al. Reference Moffitt, Caspi, Taylor, Kokaua, Milne, Polanczyk and Poulton2010). Recall is influenced by the level of current depressive symptoms (Kendler et al. Reference Kendler, Gardner and Prescott2001b ). Finally, twin studies of lifetime MD (Sullivan et al. Reference Sullivan, Neale and Kendler2000; Kendler et al. Reference Kendler, Gatz, Gardner and Pedersen2006) assume that genetic risk factors for MD are stable over adulthood – that is that the same genetic factors would impact on risk for MD at ages 25, 30, 35 years, etc. We are aware of only a single test of this hypothesis conducted over a quite short period (~18 months) (Kendler et al. Reference Kendler, Neale, Kessler, Heath and Eaves1993b ). This is a legitimate concern because genetic risk factors have been shown to change for depressive symptoms over childhood and adolescence (Silberg et al. Reference Silberg, Rutter and Eaves2001; Kendler et al. Reference Kendler, Gardner and Lichtenstein2008; Nivard et al. Reference Nivard, Dolan, Kendler, Kan, Willemsen, van Beijsterveldt, Lindauer, van Beek, Geels, Bartels, Middeldorp and Boomsma2014; Waszczuk et al. Reference Waszczuk, Zavos, Gregory and Eley2016), and for major depression in late adult life (Kendler et al. Reference Kendler, Fiske, Gardner and Gatz2009).
To address these concerns, we examined last-year MD assessed at personal interview in a population-based sample of female–female twin pairs in mid-adult life assessed at four waves spread out over an 8-year time period. We fit a series of structural equation models that address the following questions:
-
(1) Are the genetic influences on last-year MD stable over middle adulthood?
-
(2) What proportion of the environmental contributions to MD are (i) occasion-specific – that is, reflect short-term environmental stressors and/or measurement error v. (ii) temporally stable?
-
(3) Does the heritability of MD change over time?
-
(4) How much does the latent heritability for MD measured over these four interview waves increase over that obtained from each individual assessment? In particular, is it possible that the liability to MD is highly heritable but heritability assessments are substantially attenuated because of the large impact of diagnostic error and/or short-term environmental stressors?
Method
Sample and measures
Participants were the Caucasian female–female subsample from the Virginia Adult Twin Study of Psychiatric and Substance Use Disorders (VATSPSUD; Kendler & Prescott, Reference Kendler and Prescott2006). All subjects were ascertained from the population-based Virginia Twin Registry formed from a systematic review of birth certificates in the Commonwealth of Virginia. These pairs were born during 1934–1974 and were eligible if both members responded to a mailed questionnaire. As detailed in Table 1, this cohort was interviewed four times in person or by telephone over a decade with a total attrition of only 20.7%. At the first wave, the mean (s.d.) of years of education of this cohort was 14.1 (2.3). The % cooperation (of eligible twins) across the four waves equaled: 92%, 93%, 88%, and 85% (Kendler & Prescott, Reference Kendler and Prescott2006).
Table 1. Descriptive statistics for the twin sample
MD, Major depression.
Interviews were separated by a minimum of 1 year and were conducted by trained mental health professionals blind to knowledge about the co-twin. This research was approved by the VCU IRB. Informed consent was obtained prior to personal interviews and assent prior to all phone interviews. Zygosity was determined by discriminate function analyses using standard twin questions validated against DNA genotyping in 496 pairs (Kendler & Prescott, Reference Kendler and Prescott1999).
In all interviews, twins were asked about the occurrence in the last year of 15 symptoms reflecting all DSM-IV A criteria for MD (APA, 1987) which are nearly identical to those in DSM-5 (APA, 2013). As detailed elsewhere (Kendler & Prescott, Reference Kendler and Prescott2006), questions for this section were adapted from the SCID interview (Spitzer & Williams, Reference Spitzer and Williams1985). With the assistance of the interviewer, the twin subjects then aggregated these symptoms in time, reported total number of episodes and dated, to the month, the onset and offset of each episode. In earlier parts of the interview, stressful events in the last year were also reported by month, so the respondent was repeatedly helped by the interviewer to structure in time their last year experiences. Inter-rater reliability for last-year MD was good: κ (Cohen, Reference Cohen1960) = +0.74 (s.e. = 0.08), tetrachoric correlation r = 0.96 ± 0.03. In 375 twins interviewed twice by different interviewers with a mean (s.d.) inter-interview interval of 30 (9) days, the test–retest reliability of the diagnosis of MD in the last year was: κ = +0.66 [95% confidence interval (CI) 0.58–0.74], and tetrachoric correlation r = 0.88 (95% CI 0.82–0.93) (Kendler et al. Reference Kendler, Kuhn, Vittum, Prescott and Riley2005).
Data analysis
We fitted three different models to the four interview waves on the presence/absence of one or more MD episodes meeting DSM-III-R criteria in the last year: independent pathway, common pathway and Cholesky decomposition (CD). The common pathway model is a subtype of the independent pathway model so their relative fits can be directly compared. This is not true for the Cholesky model but we nonetheless relied on a comparison of fit indices as a rough guide, a commonly used approach in structural modeling.
As with traditional twin modeling, we divided the individual differences in liability to MD into three categories: additive genetic effects (A), shared or family environment (C), and unique environment (E) (Guerrini et al. Reference Guerrini, Cook, Kest, Devitgh, McQuillin, Curtis and Gurling2005). Shared environment reflects family and community experiences that increase similarity in twins raised together. Unique environment included environmental experiences not shared by siblings that impact on the phenotype under study and diagnostic error. Both independent and common pathway models further divide the sources of individual differences in liability to last-year MD into those that are in common across all four waves of assessment (indicated by subscript c) and those that are occasion-specific in their effect (subscript s).
The independent pathway model allows the common A, C and E effects to directly impact on last-year MD. The common pathway model, by contrast, constrains the A, C and E effects to impact on a latent liability to MD which in turn influences last-year MD. The full Cholesky model assumes four A, C and E factors the first of which impacts on all four times of assessment. The second and third factors impact, respectively, on the last three and the last two waves. The final factors influence only the fourth wave.
Model fitting was conducted using the classic Mx program (Neale et al. Reference Neale, Boker, Xie and Maes2003). The goal of our model fitting is to achieve the best possible balance of explanatory power and simplicity. This goal is operationalized by Akaike's Information Criterion (AIC; Akaike, Reference Akaike1987), which equals χ2 – 2df, where df is degrees of freedom. We seek to minimize the AIC value.
Errors of measurement should contribute to occasion-specific individual environmental effects (Es). Because we have an estimate of the test–retest reliability of MD in our sample (+0.88), we can subdivide Es into two parts reflecting measurement error and wave-specific unique environmental effects. We assume that measurement error contributes 1 minus 0.88 or 12% of the variance in liability. If the total e s 2 = x, then x – 0.12 equals the wave-specific unique environmental effects.
Ethical standards
The authors assert that all procedures contributing to this work comply with the ethical standards of the relevant national and institutional committees on human experimentation and with the Helsinki Declaration of 1975, as revised in 2008.
Results
At the first wave of interviews, the 2153 cooperative twins reflected 596 complete monozygotic (MZ) pairs, 433 complete dizygotic (DZ) pairs, and 95 unmatched twins who were included in the analysis and provided information about threshold placements. The prevalence of last-year MD was relatively stable over the four interview waves ranging from 9.3% to 10.8% (Table 1). Among individuals who reported an episode of MD in the last year at the first interview, the median and mean (s.d.) age at onset, number of lifetime episodes and longest duration of an episode (in days) were, respectively: 20 and 22.1 (9.7), 3 and 9.0 (17.9), and 90 and 214.0 (568.6).
The tetrachoric correlation matrix for last-year MD (Table 1) shows no clear attenuation over time. Table 1 also contains the mean inter-wave intervals which ranged from 17 to 44 months. There is no evident relationship between inter-wave intervals and correlations in MD occurrence. The mean interval between waves 1 and 4 was 7.7 years.
Table 2 presents model-fitting results. For each of our three models – independent pathway, common pathway and CD – we began with a full model and simplified until we could no longer improve the model fit as indexed by AIC. For the common pathway model, we could constrain the loadings from the common factor to the four waves of last-year MD to equality (model II), eliminate all C effects (model III), constrain the specific A and E effects on the last-year assessments to equality (model IV) and drop the specific genetic effects all with continued improvements of AIC (model V). Model V was the best-fit common pathway model.
Table 2. Model fitting results
AIC, Akaike's Information Criterion; A, additive genetic effects; C, shared environmental effects; E, individual specific environmental effects.
a Best-fit version within model: CP (common pathway), IP (independent pathway) or CD (Cholesky decomposition).
b Best overall fit.
For the independent pathway model, dropping the C effects (model II) improved the AIC value but it deteriorated when we tried to constrain the common A and E loadings to equality (model III). Dropping the occasion-specific A effects (model IV) further improved the fit index but constraining the specific E loadings to equality across the four waves did not (model V). Model IV was therefore the best-fit independent pathway model.
For the CD, we could, with an improvement in AIC, drop all C effects (model II), drop the fourth genetic factors (model III), drop the third genetic factor (model IV) and drop the second genetic factor (model V). We tried to further simplify the model by dropping all the off-diagonal E effects (so that the four E factors only impacted respectively, waves 1, 2, 3, and 4 with no cross-wave loadings) (model VI). Here, however, AIC deteriorated so model V was the best-fit CD.
As seen from Table 2, the best-fit common pathway model (CP-V) had a considerably lower AIC than the best-fit independent pathway (IP-IV) or Cholesky decomposition (CD-V) models. So the CP-V was the overall best-fit model. All three modeling approaches agreed on two key findings. First, shared environmental effects did not substantially contribute to individual differences in MD liability (hence the improved fits seen with models CP-III, IP-II and CD-II). Second, given one common genetic factor impacting on all four assessment waves, there was no need for other genetic influences (improved fits with models CP-V, IP-IV and CD-V). Furthermore, the failure of model CD-VI indicated the existence of important unique environmental effects that were shared across waves.
The best-fitting overall model (CP-V) was very simple (Fig. 1). The latent liability to last-year MD as assessed four times over an 8-year period resulted from only genetic and individual specific environmental factors. The heritability of this latent liability was high (78%) with the remaining 22% resulting from environmental factors that were both unique to individual twins and temporally stable in their effect.
Fig. 1. The best fit structural equation model for last-year major depression assessed four times over 8 years in female–female twin pairs. The path estimates, presented along with their 95% confidence intervals, represent standardized partial regression coefficients and therefore must be squared to equal the proportion of variance accounted for. The following abbreviations are used: A, Additive genetic effects; E, individual specific environmental effects; subscript ‘c’ for common or stable; subscript ‘s’ for occasion-specific.
By far the strongest contribution to risk for last-year MD, however, was our wave-specific E effects accounting for 58% of the overall liability. As detailed in the methods section, using our test-retest reliability results of last-year MD (estimated to equal +0.88) allows us to divide the influences on last-year MD into four components: stable genetic effects 33%, stable unique environmental effects 9%, unreliability of assessment 12%, and wave-specific unique environmental effects 46%.
Discussion
We examined last-year DSM-III-R MD assessed at four waves of personal interviews completed over an 8-year period years in a population-based sample of female twins in middle adult life. We fitted three classes of structural equation models and found one that clearly fitted best. From that model, we sought to answer the four questions we articulated above, the results of which we review in turn.
First, genetic influences on last-year MD were stable over the four waves of assessment. While our independent and common pathway models included wave-specific genetic influences and our Cholesky model allowed for ‘genetic innovation’ over time (that is new genetic effects at later waves), in all cases we could constrain to zero these novel sources of genetic effects with an improvement in model fit. We looked hard but could find no evidence from our results that new genetic influences on MD were activated over the 8 years of our study period. These results provide further support for the hypothesis that contrary to adolescence (Silberg et al. Reference Silberg, Rutter and Eaves2001; Kendler et al. Reference Kendler, Gardner and Lichtenstein2008) and late adult life (Kendler et al. Reference Kendler, Fiske, Gardner and Gatz2009), genetic influences on depression are stable though middle adulthood.
Second, our model allowed us to clarify the nature and timing of the environmental influences on last-year MD. The large majority of these effects are wave-specific – that is, were important only over the short time period of the last year. Excluding the impact of measurement error (which is typically confounded in most), 54% of the liability to last-year MD resulted from environmental influences. Of this aggregate effect, 17% reflected the enduring impact of environmental stressors experienced prior to the first assessment wave that impacted on MD risk over the entire period of the study. However, 83% of this effect resulted from environmental exposures over the last year whose influence on MD risk was short-term and had entirely attenuated by the next interview wave. Prior studies of this twin cohort provide some insight into the nature of these experiences. Consistent with other studies (Bebbington et al. Reference Bebbington, Tennant and Hurry1981; Surtees et al. Reference Surtees, Miller, Ingham, Kreitman, Rennie and Sashidharan1986; Brown & Harris, Reference Brown and Harris1978), we found that typical stressful life events (such as assault, job loss, marital problems, and major illnesses) produce robust elevations in risk for MD that typically last only 1–2 months (Kendler et al. Reference Kendler, Karkowski and Prescott1998). What might be responsible for the enduring environmental effects? Tentative answers come from a qualitative study of MZ twins discordant for lifetime MD conducted in this sample (Kendler & Halberstadt, Reference Kendler and Halberstadt2013). In these pairs, a joint life-history interview revealed four classes of explanations for the discordance in the experiences of the affected v. well twin: (i) severe traumatic events, (ii) chronic relationship difficulties beginning with a traumatic romantic break-up, (iii) repeated, severe occupational difficulties, and (iv) life-changing injuries (Kendler & Halberstadt, Reference Kendler and Halberstadt2013). Each of these groups of experiences likely contributed to the enduring environmental. We also found that, as in other many other samples (Fergusson & Mullen, Reference Fergusson and Mullen1999; Nelson et al. Reference Nelson, Heath, Madden, Cooper, Dinwiddie, Bucholz, Glowinski, McLaughlin, Dunne, Statham and Martin2002), childhood sexual abuse could produce a life-long enduring vulnerability to MD (Kendler et al. Reference Kendler, Bulik, Silberg, Hettema, Myers and Prescott2000).
Third, the heritability of MD was stable over the 8 years of this study. This result, interesting in its own right, also provides information about the time course of the most common environmental exposures that predispose to MD in adult life: stressful life events (Brown & Harris, Reference Brown and Harris1978; Kendler et al. Reference Kendler, Karkowski and Prescott1998). Important for our point is consistent evidence that exposure to most classes of events are only modestly correlated in twin pairs (Kendler et al. Reference Kendler, Neale, Kessler, Heath and Eaves1993a ; Middeldorp et al. Reference Middeldorp, Cath, Vink and Boomsma2005; Boardman et al. Reference Boardman, Alexander and Stallings2011). If these stressors produced long-lasting increases in the liability to MD, this should lead to declining twin resemblance for MD over time in both MZ and DZ twin pairs. This, in turn, would decreases in heritability. That we see no such effect is consistent with studies showing that the impact of typical stressful live events on risk for MD are short-lived (Brown & Harris, Reference Brown and Harris1978; Bebbington et al. Reference Bebbington, Tennant and Hurry1981; Surtees et al. Reference Surtees, Miller, Ingham, Kreitman, Rennie and Sashidharan1986; Kendler et al. Reference Kendler, Karkowski and Prescott1998). To be clear, we cannot rule out some events in middle adulthood which produce enduring effects on depression liability. While they likely occur, they are probably too rare for us to detect in our analyses given the available sample size.
Fourth, we showed a large increase in heritability of MD moving from that observed for a 1-year period (33%) compared to the latent stable estimate acting across all four waves (78%). To be clear, these two estimates are not directly comparable as they apply to two different constructs for ‘MD’. In the first, we examine only the presence or absence of a depressive episode in a 1-year time period. That is, we analyze ‘observed MD’. In the second, we examine the latent liability to MD as measured over four different time periods. That is, we analyze the ‘liability to MD’. The increased heritability of the latter measure results directly from this model's ability to reduce the impact of short-term environment and errors of measurement thereby increasing the proportion of liability to MD due to genetic factors.
The use of multiple measures to increase heritability is an old technique in statistical genetics first developed by animal and plant breeders (Falconer, Reference Falconer1989). In discussing its application milk yield in cattle and litter size in mice, Falconer notes that this method works because of
increasing the number of measurements reduces the amount of variance due to special [aka occasion-specific] environment … When the repeatability is low, multiple measurements may lead to a worthwhile gain in accuracy [of genetic prediction] (Falconer, Reference Falconer1989, pp. 139–144).
He presents formulas for the expected increases in estimated heritability as a function of the repeatability of the trait. For last-year MD (with a ‘repeatability’ of ~+0.40), this formula suggests substantial gains for up to five or six independent measurements.
Our results address another concern with heritability estimates of lifetime MD – the possible impact of problems with long-term recall. Critics can argue, given well-demonstrated problems in the accuracy and bias of long-term memory (Coughlin, Reference Coughlin1990; Andrews et al. Reference Andrews, Anstey, Brodaty, Issakidis and Luscombe1999; Kendler et al. Reference Kendler, Gardner and Prescott2001b ; Wells & Horwood, Reference Wells and Horwood2004; Moffitt et al. Reference Moffitt, Caspi, Taylor, Kokaua, Milne, Polanczyk and Poulton2010), that the heritability of lifetime MD might be substantially ‘contaminated’ by such recall effects. For example, a lifetime history of MD might assess, at least in part, a ‘plaintive set’ of cognitive biases rather than the true occurrence of MD episodes. Our results argue against this hypothesis both because our assessments were based on much shorter time periods of recall (1 year) and we employed interviews that aided the respondent in temporally structuring their last year experiences, an approach shown to improve the accuracy of human recall (Belli, Reference Belli1998).
We have previously demonstrated that two independent reports of lifetime MD also increases heritability substantially (Kendler et al. Reference Kendler, Gardner, Neale and Prescott2001a ). These and the present results have implications for molecular genetic studies of MD which suffer from problems of low power due to small effect-size genetic variants (CONVERGE Consortium, 2015; Major Depressive Disorder Working Group of the Psychiatric GWAS Consortium, 2012). When dealing with low heritability phenotypes like MD with only moderate stability over time especially in general population samples, multiple independent measurements is one potential way to increase, potentially substantially, the ‘genetic signal’.
Limitations
Our findings should be considered in the light of five potential methodological limitations. First, our subjects are all female, white twins born in Virginia. While our sample has been shown to be broadly representative of the US population for a range of relevant traits (Kendler & Prescott, Reference Kendler and Prescott2006), the degree to which our present findings extrapolate to other samples and to males is uncertain. With respect to the rates of psychopathology, twins are probably representative of the general population (Kendler et al. Reference Kendler, Pedersen, Farahmand and Persson1996; Kendler & Prescott, Reference Kendler and Prescott2006). Our 1-year prevalence rates for MD in females were mid-way between those reported in the National Comorbidity Survey (12.9%) (Kessler et al. Reference Kessler, McGonagle, Zhao, Nelson, Hughes, Eshleman, Wittchen and Kendler1994) and the National Epidemiologic Survey on Alcoholism and Related Conditions (6.9%) (Hasin et al. Reference Hasin, Goodwin, Stinson and Grant2005).
Second, in examining a community-based sample, we obtain cases of MD that are representative of the general population. However, their average level of severity and patterns of recurrence and comorbidity likely differ from those seen in clinical settings, especially in-patient treatment. So our findings may not extrapolate to clinically ascertained cases of MD in particular because more severe forms of MD and those with treatment tend to be more reliability reported (Bromet et al. Reference Bromet, Dunn, Connell, Dew and Schulberg1986; Foley et al. Reference Foley, Neale and Kendler1998).
Third, because the power of our sample size of twins was only moderate, we cannot rule out modest changes in genetic influences on MD over time or small shared environmental effects (Neale et al. Reference Neale, Eaves and Kendler1994).
Fourth, there have been traditional concerns about biases in the twin method that might increase heritability estimates, especially the ‘equal environmental assumption’ – that MZ and DZ twins have approximately equally similar childhood environments. This assumption has been tested several times in this sample and shown for MD and other psychiatric disorders to be well justified (Kendler & Prescott, Reference Kendler and Prescott2006).
Fifth, our estimates of the role of environmental risk factors (E) in our analyses were all inferred from the patterns of resemblance for MD across time and in MZ and DZ twin pairs. This has the important advantage of cleanly separating sources of variance in liability based on our twin model assumptions. The disadvantage is that we cannot specify the actions of specific risk factors. We have, however, in previous large-scale path models in this sample, examined the action of both a wide variety of genetic, temperamental and environmental risk factors predicting onset of MD in the last year (Kendler et al. Reference Kendler, Gardner and Prescott2002) and the interrelationship on a month by month basis of a latent depressive vulnerability and aggregate stressful life events (Kendler & Gardner, Reference Kendler and Gardner2016). These analyses provide complementary perspectives on the etiologic processes underlying risk for MD.
Summary
The latent liability to MD as assessed over four waves of measurement in middle-adulthood is highly heritable and developmentally stable. The large majority of the environmental risk for MD (83%) is occasion-specific and produces only transient elevations in MD risk. However, a small proportion of the environmental risk for MD (17%) arises from environmental experiences earlier in life that have an enduring impact on liability. As predicted from genetic theory, given the moderate stability of depressive episodes over adult life, heritability of MD is increased substantially by multiple measurements.
Acknowledgements
This project was supported by grant MH40828. The Virginia Twin Registry is supported by grant UL1RR031990. Dr Carol Prescott played a central role in the design and implementation of this twin study. Dr Michael Neale provided advice on the statistical modeling.
Declaration of Interest
None.
