The study explored to what extent recruitment bias and attrition may have affected the representativeness of our sample and the results of the biometric twin analyses. Results from comparisons between complete pairs and single responders indicate a clear but moderate selection effect in the first stage of the study, Q1, among male twins for symptoms of anxiety and depression (mean difference approximately ¼ SD
between single and pairs) and somatic illnesses (difference
), but not among females. There was no difference between complete pairs and singles regarding co-twin closeness.
An array of factors could potentially influence response to broad-based health questionnaires and participation in a mental health interview. In our Norwegian, population-based program of twin research, socio-demographic factors — primarily female sex, higher education, and zygosity — were the most important predictors of participation. There was a clearly higher participation rate in the first questionnaire study (1992) than among twins invited for the first time to the second questionnaire study. This difference is probably reflecting both a time trend of decreased willingness to respond to questionnaire studies and less willingness to complete the much longer second, compared to the first, quite short, questionnaire.
Few health variables recorded at the twins’ birth predicted participation, and the effects were weak. Whereas health disadvantages mostly predicted low participation, one, placenta previa, predicted high participation. Good somatic and mental health and a healthy life style reported in the first questionnaire (1992) quite moderately predicted participation in the next (1998). There are virtually no health selection effects between the second questionnaire and the interview study.
It is worthwhile noting that none of the analyses indicated selection towards low — or a priori more likely — high physical and emotional co-twin closeness. Such selection might have had particular consequences for phenotypic co-twin similarity and, thus, for results from biometric genetic analyses.
A series of quantitative genetic analyses did not show evidence of differences between interview study participants and nonparticipants in the genetic and environmental covariance structure for a broad range of mental health indicators.
Previous studies of bias in twin studies due to selection of demography and health related phenotypes are rather scarce. A longitudinal study examined sampling bias in the Australian twin cohort born from 1944 to 1963, using a reverse design to identify correlates of nonresponse at a previous occasion (Heath et al., 1998
). All twins invited to participate in a questionnaire study were also recruited to a telephone interview conducted approximately 10 years later. Analyses of the interview data revealed findings highly similar to ours. Sociodemographic correlates, including education below university level, male sex, being a dizygotic twin and membership in the youngest birth cohort, showed the strongest effects explaining nonresponse to the earlier questionnaire study. The psychiatric measures yielded much more modest effects with nonresponse associated with a history of alcohol dependence, childhood conduct disorder, and social anxiety. The authors conclude that sampling biases are strongest for the sociodemographic measures and relatively minor for the psychiatric measures. Another Australian study of somewhat younger twins also showed results similar to ours (Heath et al., 2001
). A British study of twin children showed around 0.2 SD
lower mean scores for aggressive and delinquent behavior in a group whose parents responded to a questionnaire compared to nonresponder families (Taylor, 2004
). Model fitting results showed lower genetic effect and higher shared environmental effect for aggressive behavior, and the opposite trend (increased genetic effect at the expense of shared environmental effect) for delinquent behavior, in responder families compared to the full sample. However, the data were not tested for differences in covariance structure between participant and nonparticipant families. The results are also not fully comparable with ours, examining effects of parental response rather than the twins’ own responses, and because of the twins’ young age. A study of Virginian twins aged from 8 to 16 years and their parents showed no selection towards high SES from the start of the study. Unlike with our sample, however, a selection took place during later phases of the study as families living in low-income communities dropped out of the study. This demographic selection had almost no effect on the prevalence estimates for parental mental disorders (Meyer et al., 1996
). Likewise, another Virginian study of twin data, including a broad set of variables on demographic information and mental health, showed only moderate effects on participation in subsequent studies of the following variables: female sex, higher education, older age, Protestant religious affiliation, and an absence of drinking problems (Kendler & Prescott, 2006
). A Swedish study of selection bias in elderly twins (Simmons et al., 1997
) concluded: ‘The results of the present study suggest that although a selection bias may exist, it is neither pervasive nor large in population based samples’ (p. 565). A Norwegian study of intelligence, comparing twins having participated or not participated in a questionnaire study, showed higher IQ among the participants, but there were no difference between participants and nonparticipants in covariance structure or heritability of IQ (Tambs et al., 1989
). Likewise, our results, showing no recruitment bias for the genetic covariance structure for mental health, is in agreement with results from the Minnesota Twin Family Registry. These results showed no differences in co-twin correlations between twins responding after first invitation and twins only responding after offered incentives to participate (Lykken et al., 1990
). Although differences in samples and phenotypes make comparisons difficult, our results seem to be consistent with most previous results.
Our results should not be interpreted in too much detail. Multiple testing may have resulted in a few results reaching significance by chance — perhaps in nonexpected direction — and which true effects that reached significance and which did not are to a large extent random. In general, however, strong selection over a broad range of somatic and mental health variables would have resulted in more and stronger effects than observed in our sample.
Another important limitation is related to the appropriateness of the medical variables from the MBRN used as predictors to the entry of the study. A lot of evidence exists for the impact of fetal factors on adult health (Barker, 1998
), including some pertaining to mental health (Cheung et al., 2004
; Thompson et al., 2001
). There is also evidence of an association between low birth-weight and risk for epilepsy in males and with refractive disorders, chronic ear infections and stomach problems in women in our Q1 data (Harris et al., 1997
), and with nearsightedness and minimal brain dysfunction (MBD) in the Q2 data (Grjibovski et al., 2005
). Nonetheless, such observed associations are typically weak, and prenatal factors typically predict illness among people older than our twin cohort. This indirect and relatively insensitive way of testing associations between health and participation may well have left selection for health related factors undetected. The absence of evidence of differences in somatic and mental health problems between MZ and DZ twins is somewhat reassuring regarding undetected selection, however.
Perhaps the most important results derive from the biometric twin study of mental health indicators testing for differences in covariance structure between pairs who did and did not participate in the interview. One interpretation of these results is that attrition bias in our data affects prevalence rates but not the genetic and environmental variance estimates. However, this interpretation is not without caveats; the clearest selection effects for the mental health measures seem to have already occurred by Q2, and not very much selection appears to have taken place from Q2 to the interview. Furthermore, the variance component findings should be evaluated in light of power to reject equivalence in the variance components across the two groups. Simulation studies clearly demonstrate that only quite substantial differences in parameter estimates can be expected to be detected with our sample size (Neale et al., 1994
). Rather than proof of absence of recruitment bias, the results — showing no trends of differences between covariance structures for a large number of phenotypes — should be understood as suggestive of no strong bias effect.
Regardless of the described limitations, the results indicate that twin studies of health are not strongly selected towards poor or strong health generally, and mental health specifically. No definite conclusion can be drawn regarding selection bias at the entry of the study, but we are confident that drop-outs during follow-ups do not seriously threaten the representativeness of our sample. The main question is whether individual or pair-wise selection affects the estimates from the genetic analyses. The tentative answer is no.