Manuel Ato García

Classical and causal inference approaches to statistical mediation analysis

BACKGROUND Although there is a broad consensus on the use of statistical procedures for mediation analysis in psychological research, the interpretation of the effect of mediation is highly controversial because of the potential violation of the assumptions required in application, most of which are ignored in practice. METHOD This paper summarises two currently independent procedures for mediation analysis, the classical/SEM and causal inference/CI approaches, together with the statistical assumptions required to estimate unbiased mediation effects, in particular the existence of omitted variables or confounders. A simulation study was run to test whether violating the assumptions changes the estimation of mediating effects. RESULTS The simulation study showed a significant overestimation of mediation effects with latent confounders. CONCLUSIONS We recommend expanding the classical with the causal inference approach, which generalises the results of the first approach to mediation using a common estimation method and incorporates new tools to evaluate the statistical assumptions. To achieve this goal, we compare the distinguishing features of recently developed software programs in R, SAS, SPSS, STATA and Mplus.

The two-way mixed model: A long and winding controversy

BACKGROUND With the 2-way mixed model, one a fixed factor and the other random, the procedure followed to test statistical significance of the random factor has been the focus of a heated controversy in theoretical and applied statistics, and the debating continues even now. One of the main consequences of this controversy is that the position defended in the classical ANOVA texts on the hypothesis of the significance of the random effect is not the same as that defended in almost all of the professional statistical software programs. METHOD In this paper, we deal with a detailed analysis about the controversy of mixed model and the decision about one of two basic options, the non restrictive and the restrictive model. RESULTS Three key questions we consider to go beyond the controversy are: (1) the two classical models are equivalent, (2) the marginality principle do not allow to test main effects in presence of interactive significant effects and (3) the relevance of linear mixed approach to analyze models with fixed and random effects. CONCLUSIONS We propose the simple solution of using the mixed linear approach with REML estimation instead of the classical linear approach, which is really unapplicable in this context.