Journal Code Article ID Dispatch: 24.05.12 CE: Dorio, Lynette P E R 1 8 6 5 No. of Pages: 23 ME: 1 European Journal of Personality, Eur. J. Pers. (2012) 65 2 Published online in Wiley Online Library (wileyonlinelibrary.com) DOI: 10.1002/per.1865 66 3 67 4 68 5 69 6 70 7 OPEN PEER COMMENTARY 71 8 72 9 73 10 Bias due to Controlling a Collider: A Potentially Important Issue for Personality Research 74 11 75 12 76 JENS B. ASENDORPF 13 77 14 Department of Psychology, Humboldt University Berlin, Germany 78 15 [email protected] 79 16 80 17 Abstract: I focus on one bias in correlational studies that has been rarely recognised because of the current taboo on discussions of 81 fi 18 causality in these studies: bias due to controlling a collider. It cannot only induce arti cial correlations between statistically 82 19 independent predictors but also suppress or hide real correlations between predictors. If the collider is related to selective sampling, a 83 particularly nasty bias results. Bias due to controlling a collider may be as important as bias due to a suppressor effect. Copyright © 20 84 2012 John Wiley & Sons, Ltd. 21 85 22 86 23 In his stimulating paper that is unfortunately sometimes hard Self-efficacy is an important resource, so the association of im- 87 24 to read, Lee (this issue) touches a taboo topic in current migrant status with self-efficacy provides important 88 25 personality publications: causal relations among variables information. Do these immigrants have lower self-efficacy 89 26 that describe between-person differences. For many years, expectations than their Greek peers? The answer is yes (the 90 27 authors were educated by reviewers and editors to avoid zero-order correlation between dummy-coded immigrant 91 28 causal language because of the many pitfalls in causal inter- status in a sample of 969 adolescent immigrant students along 92 29 pretations of correlations. These pitfalls granted dismissing with their Greek classmates was À.15, p < .001). 93 30 causality altogether are like throwing out the baby with the In studies of immigrant adaptation, skills in the host 94 31 bathwater. As humans, we cannot avoid thinking in terms language are often routinely controlled because they may 95 32 of causality, and therefore, tabooing this topic in publications already explain most or all effects of other predictors of 96 33 does not prevent readers and mass media from their own causal adaptation (although in many cases, suppressor effects may 97 34 interpretations guided by implicit rules such as ‘A correlates occur because the effect of language skills on adaptation is 98 35 with B’ means ‘A causes B’ but ‘B correlates with A’ means relatively strong). In the aforementioned case, if one controls 99 36 ‘B causes A’. the correlation between immigrant status and self-efficacy for 100 37 Although causality is a difficult concept in correlational the ability to speak Greek, the resulting partial correlation is 101 38 studies, scientists should and actually can do better than this À.03 and not significant any more. The control of Greek speak- 102 39 if they can be pressed to explicate the causal model, or alterna- ing skills induces a bias due to a collider because these skills are 103 40 tive causal models, underlying their research questions. The very likely causally influenced by both immigrant status and 104 41 directed acyclic graph (DAG) method described by Lee (this self-efficacy. Indeed, the respective correlations were À.37, 105 42 issue) is a valuable method of achieving such an explication p < .001 and .23, p < .001. Thus, controlling for host language 106 43 (see Foster, 2010, for an excellent discussion of causality based skills is highly problematic in studies of the adaptation of immi- 107 44 on DAGs for developmental psychologists). My comment here grants where a resource and/or an adaptation outcome influence 108 45 focuses on a key concept in the DAG approach: the collider. these skills because in such cases one controls a collider. 109 46 If one starts with explicit causal models before decisions 110 47 Bias due to explicit control of a collider: Example from are made on the statistical control of variables in the model, 111 48 research on adaptation one will rarely commit this kind of erroneous over control. 112 49 But if no causal analysis is made and the models involve many 113 50 A collider is an outcome of two joint predictors that may be variables, or variables where it is not clear whether they should 114 51 correlated or not. If one statistically controls for a collider, be considered a predictor or an outcome, researchers can easily 115 52 the resulting correlation between the predictors will be neces- be lost in covariation, relying on traditional routines designed 116 53 sarily biassed. Although this bias is most often discussed for the control of certain predictor variables although they 117 54 only for the case where two predictors are uncorrelated such might be outcomes in the present context. 118 55 that the bias consists of a spurious correlation, the bias is in 119 56 fact general: any correlation will be biassed by the adjust- Bias due to implicitly controlling a collider through 120 57 ment. As Lee (this issue) has correctly observed, the bias is sampling: Example from research on achievement 121 58 obvious but rarely noticed by researchers. 122 59 For an example, let us consider data on risks and resources If a collider is related to sampling such that the sample of 123 60 for adaptation of immigrant youth in Greece to the Greek participants is restricted in variation on the collider, this is 124 61 culture (Motti-Stefanidi, Asendorpf, & Masten, in press). equivalent to statistically controlling part of the variation of 125 62 126 63 127 64 Copyright © 2012 John Wiley & Sons, Ltd. 128 1 2 Discussion 65 2 66 3 the collider and therefore also introduces a bias. This is a Note that this bias is different from effects of restricted 67 4 particularly nasty case because the researcher did not variance that can inflate or suppress a correlation but cannot 68 5 explicitly control for the collider—the collider was implicitly induce a spurious correlation or change signs of a correlation. 69 6 controlled through selective sampling. Bias due to implicit control of a collider can do this and may 70 7 Asurprisingfinding from research on achievement may be even more common than bias due to explicit control of a 71 8 illustrate this bias (I am grateful to Marco Perugini who alerted collider. Many personality researchers are aware of biases 72 9 me to this case). Studies that relate IQ and conscientiousness to due to restricted sampling, and the possibility to correct for 73 10 achievement regularly find the expected positive correlations them, but there is no tradition to consider biases due to the 74 11 of IQ and conscientiousness with achievement but at the implicit control of a collider. 75 12 same time nonsignificant or even negative correlations 76 13 between IQ and conscientiousness (see, e.g. the meta-analysis Importance for personality research 77 14 by Ackerman & Heggestad, 1997), and authors who tried to 78 15 explain this unexpected result had difficulty finding post hoc Right now it is hard to judge the importance of biases in 79 16 explanations. A causal analysis suggests a bias introduced by personality research that are due to explicit or implicit 80 17 controlling a collider related to sampling. Most of these studies control of colliders because both biases are largely 81 18 used university students or samples biassed toward high unexplored. For the time being, a working hypothesis is 82 19 achievement, and this bias in sampling alone induced a that they may be as important as the better known 83 20 negatively biassed correlation because achievement is a suppressor effects between multiple predictors of the 84 21 collider of IQ and conscientiousness. same outcome. 85 22 86 23 87 24 88 25 89 26 What Kind of Causal Modelling Approach Does Personality Research Need? 90 27 91 28 DENNY BORSBOOM1, SOPHIE VAN DER SLUIS1,2, ARJEN NOORDHOF1, MARIEKE WICHERS3, NICOLE 92 29 GESCHWIND3, STEVEN H. AGGEN4,5, KENNETH S. KENDLER4,5 AND ANGÉLIQUE O. J. CRAMER1 93 30 94 31 1 Department of Psychology, University of Amsterdam 95 32 2 Complex Trait Genetics, Department of Functional Genomics and Department of Clinical Genetics, Centre for Neurogenomics and 96 33 Cognitive Research (CNCR), FALW-VUA, Neuroscience Campus Amsterdam, VU University Medical Centre (VUmc) 97 34 3 European Graduate School for Neuroscience, SEARCH, Department of Psychiatry and Psychology, Maastricht University Medical Centre 98 35 4 Virginia Institute for Psychiatric and Behavioural Genetics 99 36 5 Department of Psychiatry, Virginia Commonwealth University 100 37 101 38 [email protected] 102 39 103 40 Q1 Abstract: Lee (2012) proposes that personality research should utilise recent theories of causality. Although 104 41 we agree that such theories are important, we also note that their empirical application has not been very 105 42 successful to date. The reason may be that psychological systems are frequently characterised by feedback, 106 43 nonlinearity and individual differences in causal structure. Such features do not preclude the application of 107 44 causal modelling but do limit the usefulness of the approach for the analysis of typical personality data. To 108 45 adequately investigate personality, intensive time series of repeated measurements are needed. Copyright © 109 46 2012 John Wiley & Sons, Ltd. 110 47 111 48 112 49 We agree with Lee that recent theories of causality (Pearl, to showing that there is a problem in either the causal 113 50 2000) are important additions to the methodological litera- assumptions or the data, or both.