A New Approach to Handle Missing Covariate Data in Twin Research: With an Application to Educational Achievement Data

The often-used ACE model which decomposes phenotypic variance into additive genetic (A), common-environmental (C) and unique-environmental (E) parts can be extended to include covariates. Collection of these variables however often leads to a large amount of missing data, for example when self-reports (e.g. questionnaires) are not fully completed. The usual approach to handle missing covariate data in twin research results in reduced power to detect statistical effects, as only phenotypic and covariate data of individual twins with complete data can be used. Here we present a full information approach to handle missing covariate data that makes it possible to use all available data. A simulation study shows that, independent of missingness scenario, number of covariates or amount of missingness, the full information approach is more powerful than the usual approach. To illustrate the new method, we applied it to test scores on a Dutch national school achievement test (Eindtoets Basisonderwijs) in the final grade of primary school of 990 twin pairs. The effects of school-aggregated measures (e.g. school denomination, pedagogical philosophy, school size) and the effect of the sex of a twin on these test scores were tested. None of the covariates had a significant effect on individual differences in test scores.