Modeling stimulus variation in three common implicit attitude tasks.

We explored the consequences of ignoring the sampling variation due to stimuli in the domain of implicit attitudes. A large literature in psycholinguistics has examined the statistical treatment of random stimulus materials, but the recommendations from this literature have not been applied to the social psychological literature on implicit attitudes. This is partly because of inherent complications in applying crossed random-effect models to some of the most common implicit attitude tasks, and partly because no work to date has demonstrated that random stimulus variation is in fact consequential in implicit attitude measurement. We addressed this problem by laying out statistically appropriate and practically feasible crossed random-effect models for three of the most commonly used implicit attitude measures-the Implicit Association Test, affect misattribution procedure, and evaluative priming task-and then applying these models to large datasets (average N = 3,206) that assess participants' implicit attitudes toward race, politics, and self-esteem. We showed that the test statistics from the traditional analyses are substantially (about 60 %) inflated relative to the more-appropriate analyses that incorporate stimulus variation. Because all three tasks used the same stimulus words and faces, we could also meaningfully compare the relative contributions of stimulus variation across the tasks. In an appendix, we give syntax in R, SAS, and SPSS for fitting the recommended crossed random-effects models to data from all three tasks, as well as instructions on how to structure the data file.