Mediation Analysis in Social Psychology: Current Practices and New Recommendations

A key aim of social psychology is to understand the psychological processes through which independent variables affect dependent variables in the social domain. This objective has given rise to statistical methods for mediation analysis. In mediation analysis, the significance of the relationship between the independent and dependent variables has been integral in theory testing, being used as a basis to determine (1) whether to proceed with analyses of mediation and (2) whether one or several proposed mediator(s) fully or partially accounts for an effect. Synthesizing past research and offering new arguments, we suggest that the collective evidence raises considerable concern that the focus on the significance between the independent and dependent variables, both before and after mediation tests, is unjustified and can impair theory development and testing. To expand theory involving social psychological processes, we argue that attention in mediation analysis should be shifted towards assessing the magnitude and significance of indirect effects. Understanding the psychological processes by which independent variables affect dependent variables in the social domain has long been of interest to social psychologists. Although moderation approaches can test competing psychological mechanisms (e.g., Petty, 2006; Spencer, Zanna, & Fong, 2005), mediation is typically the standard for testing theories regarding process (e.g., Baron & Kenny, 1986; James & Brett, 1984; Judd & Kenny, 1981; MacKinnon, 2008; MacKinnon, Lockwood, Hoffman, West, & Sheets, 2002; Muller, Judd, & Yzerbyt, 2005; Preacher & Hayes, 2004; Preacher, Rucker, & Hayes, 2007; Shrout & Bolger, 2002). For example, dual process models of persuasion (e.g., Petty & Cacioppo, 1986) often distinguish among competing accounts by measuring the postulated underlying process (e.g., thought favorability, thought confidence) and examining their viability as mediators (Tormala, Briñol, & Petty, 2007). Thus, deciding on appropriate requirements for mediation is vital to theory development. Supporting the high status of mediation analysis in our field, MacKinnon, Fairchild, and Fritz (2007) report that research in social psychology accounts for 34% of all mediation tests in psychology more generally. In our own analysis of journal articles published from 2005 to 2009, we found that approximately 59% of articles in the Journal of Personality and Social Psychology (JPSP) and 65% of articles in Personality and Social Psychology Bulletin (PSPB) included at least one mediation test. Consistent with the observations of MacKinnon et al., we found that the bulk of these analyses continue to follow the causal steps approach outlined by Baron and Kenny (1986). Social and Personality Psychology Compass 5/6 (2011): 359–371, 10.1111/j.1751-9004.2011.00355.x a 2011 The Authors Social and Personality Psychology Compass a 2011 Blackwell Publishing Ltd The current article examines the viability of the causal steps approach in which the significance of the relationship between an independent variable (X) and a dependent variable (Y) are tested both before and after controlling for a mediator (M) in order to examine the validity of a theory specifying mediation. Traditionally, the X fi Y relationship is tested prior to mediation to determine whether there is an effect to mediate, and it is also tested after introducing a potential mediator to determine whether that mediator fully or partially accounts for the effect. At first glance, the requirement of a significant X fi Y association prior to examining mediation seems reasonable. If there is no significant X fi Y relationship, how can there be any mediation of it? Furthermore, the requirement that X fi Y become nonsignificant when controlling for the mediator seems sensible in order to claim ‘full mediation’. What is the point of hypothesizing or testing for additional mediators if the inclusion of one mediator renders the initial relationship indistinguishable from zero? Despite the intuitive appeal of these requirements, the present article raises serious concerns about their use.