The heterogeneous effects of confounder bias: Evaluating the implications for social behavioral research

Social scientists often rely on non-experimental data to examine associations between variables. This reliance, however, coincides with an inability to adjust estimated associations for unobserved confounding mechanisms. Importantly, adjusting estimates for a confounder can decrease the probability of committing a Type 1, Type 2, and/or Type S error. Nevertheless, the similarities and differences in the probability of committing these errors across linear regression models with distinct distributional assumptions remains largely unexamined in the literature. Four randomly specified directed equation simulation analyses and four real data examples were estimated to examine the probability of committing these hypothesis testing errors, as well as the degree of bias reduction generated when adjusting for a theorized confounder. These results demonstrated that Type 1, Type 2, and/or Type S errors occur more frequently than expected, suggesting that non-experimental research examining social and behavioral processes should implement techniques to reduce confounder bias across linear regression models.