Differential Recall Bias, Intermediate Confounding, and Mediation Analysis in Life Course Epidemiology: An Analytic Framework with Empirical Example.

The mechanisms by which childhood socioeconomic status (CSES) affects adult mental health, general health, and well-being are not clear. Moreover, the analytical assumptions employed when assessing mediation in social and psychiatric epidemiology are rarely explained. The aim of this paper was to explain the intermediate confounding assumption, and to quantify differential recall bias in the association between CSES, childhood abuse, and mental health (SCL-10), general health (EQ-5D), and subjective well-being (SWLS). Furthermore, we assessed the mediating role of psychological and physical abuse in the association between CSES and mental health, general health, and well-being; and the influence of differential recall bias in the estimation of total effects, direct effects, and proportion of mediated effects. The assumptions employed when assessing mediation are explained with reference to a causal diagram. Poisson regression models (relative risk, RR and 99% CIs) were used to assess the association between CSES and psychological and physical abuse in childhood. Mediation analysis (difference method) was used to assess the indirect effect of CSES (through psychological and physical abuse in childhood) on mental health, general health, and well-being. Exposure (CSES) was measured at two time points. Mediation was assessed with both cross-sectional and longitudinal data. Psychological abuse and physical abuse mediated the association between CSES and adult mental health, general health, and well-being (6-16% among men and 7-14% among women, p < 0.001). The results suggest that up to 27% of the association between CSES and childhood abuse, 23% of the association between childhood abuse, and adult mental health, general health, and well-being, and 44% of the association between CSES and adult mental health, general health, and well-being is driven by differential recall bias. Assessing mediation with cross-sectional data (exposure, mediator, and outcome measured at the same time) showed that the total effects and direct effects were vastly overestimated (biased upwards). Consequently, the proportion of mediated effects were underestimated (biased downwards). If there is a true (unobserved) direct or indirect effect, and the direction of the differential recall bias is predictable, then the results of cross-sectional analyses should be discussed in light of that.