Analytics

Concerning the analytics, fixed effects (FE) panel regression exploits the longitudinal data structure of panels as it only takes variations within the individuals’ life courses into account. Thus, the FE estimator is unbiased in the presence of cross-sectional unobserved heterogeneity affecting both the observed covariates and the outcome^(42,43). If the strict exogeneity assumption holds, the FE regression adequately estimates unbiased causal effects of the covariates on the outcome⁽⁴⁴⁾. For instance, this enables the identification of the effect of ageing on dietary quality while controlling for birth cohort categories or whether an increase in dietary quality induces a decrease in BMI. However, it is still common practice in most of the epidemiological prospective cohort studies not to regress changes in the outcome on changes in the covariates. This practice implicitly assumes that diets are a time-invariant exposition and remain stable over the individuals’ life courses⁽²¹⁾. This assumption is not empirically valid, as the KiGGS and NEMONIT data as well as Mertens et al. ⁽¹⁷⁾ demonstrate.

Nevertheless, standard FE models can only estimate the effects for time-varying variables and do not allow the inclusion of time-invariant characteristics. The generalised linear mixed panel regression model (so-called hybrid model)^(45,46) simultaneously estimates fixed, between and random effects. Thus, the hybrid model enables the inclusion of both time-varying (e.g. income) and time-invariant (e.g. migration background) variables in the same model. As it is true for the FE model, the hybrid model’s FE estimates are unbiased, if the strict exogeneity assumption holds⁽⁴⁴⁾. Given the outlined advantages of the hybrid model, it is applied throughout the analyses. The specification of the hybrid model and an overview of the sensitivity analyses performed for all the reported regression results can be found in the supplement.