Bias due to Controlling aCollider : A Potentially Important Issue for Personality Research

I focus on one bias in correlational studies that has been rarely recognised because of the current taboo on discussions of causality in these studies: bias due to controlling a collider. It cannot only induce artificial correlations between statistically independent predictors but also suppress or hide real correlations between predictors. If the collider is related to selective sampling, a particularly nasty bias results. Bias due to controlling a collider may be as important as bias due to a suppressor effect. Copyright © 2012 John Wiley & Sons, Ltd. In his stimulating paper that is unfortunately sometimes hard to read, Lee (this issue) touches a taboo topic in current personality publications: causal relations among variables that describe between-person differences. For many years, authors were educated by reviewers and editors to avoid causal language because of the many pitfalls in causal interpretations of correlations. These pitfalls granted dismissing causality altogether are like throwing out the baby with the bathwater. As humans, we cannot avoid thinking in terms of causality, and therefore, tabooing this topic in publications does not prevent readers andmass media from their own causal interpretations guided by implicit rules such as ‘A correlates with B’ means ‘A causes B’ but ‘B correlates with A’ means ‘B causes A’. Although causality is a difficult concept in correlational studies, scientists should and actually can do better than this if they can be pressed to explicate the causal model, or alternative causal models, underlying their research questions. The directed acyclic graph (DAG) method described by Lee (this issue) is a valuable method of achieving such an explication (see Foster, 2010, for an excellent discussion of causality based on DAGs for developmental psychologists).My comment here focuses on a key concept in the DAG approach: the collider. Bias due to explicit control of a collider: Example from research on adaptation A collider is an outcome of two joint predictors that may be correlated or not. If one statistically controls for a collider, the resulting correlation between the predictors will be necessarily biassed. Although this bias is most often discussed only for the case where two predictors are uncorrelated such that the bias consists of a spurious correlation, the bias is in fact general: any correlation will be biassed by the adjustment. As Lee (this issue) has correctly observed, the bias is obvious but rarely noticed by researchers. For an example, let us consider data on risks and resources for adaptation of immigrant youth in Greece to the Greek culture (Motti-Stefanidi, Asendorpf, & Masten, in press). Self-efficacy is an important resource, so the association of immigrant status with self-efficacy provides important information. Do these immigrants have lower self-efficacy expectations than their Greek peers? The answer is yes (the zero-order correlation between dummy-coded immigrant status in a sample of 969 adolescent immigrant students along with their Greek classmates was .15, p< .001). In studies of immigrant adaptation, skills in the host language are often routinely controlled because they may already explain most or all effects of other predictors of adaptation (although in many cases, suppressor effects may occur because the effect of language skills on adaptation is relatively strong). In the aforementioned case, if one controls the correlation between immigrant status and self-efficacy for the ability to speak Greek, the resulting partial correlation is .03 and not significant any more. The control of Greek speaking skills induces a bias due to a collider because these skills are very likely causally influenced by both immigrant status and self-efficacy. Indeed, the respective correlations were .37, p< .001 and .23, p< .001. Thus, controlling for host language skills is highly problematic in studies of the adaptation of immigrants where a resource and/or an adaptation outcome influence these skills because in such cases one controls a collider. If one starts with explicit causal models before decisions are made on the statistical control of variables in the model, one will rarely commit this kind of erroneous over control. But if no causal analysis is made and the models involve many variables, or variables where it is not clear whether they should be considered a predictor or an outcome, researchers can easily be lost in covariation, relying on traditional routines designed for the control of certain predictor variables although they might be outcomes in the present context. Bias due to implicitly controlling a collider through sampling: Example from research on achievement If a collider is related to sampling such that the sample of participants is restricted in variation on the collider, this is equivalent to statistically controlling part of the variation of European Journal of Personality, Eur. J. Pers. (2012) Published online in Wiley Online Library (wileyonlinelibrary.com) DOI: 10.1002/per.1865 Copyright © 2012 John Wiley & Sons, Ltd. Journal Code Article ID Dispatch: 24.05.12 CE: Dorio, Lynette P E R 1 8 6 5 No. of Pages: 23 ME: 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24