Why a standard IAT effect cannot provide evidence for association formation: the role of similarity construction

ABSTRACT

Moran and Bar-Anan (Moran, T., & Bar-Anan, Y. (2013). The effect of object-valence relations on automatic evaluation. Cognition and Emotion, 27(4), 743–752) demonstrated that evaluations on a direct measure reflected information on both US valence and CS-US relations, whereas evaluations on an indirect measure (IAT) reflected only information on US valence. This dissociation between measures supposedly tapping into propositional and associative processes apparently supports dual process models of EC. In the present study, we present an alternative explanation of this pattern, based on an interpretation of IAT effects in terms of flexible similarity construction processes. According to this account, processing draws on those features that discriminate between target categories, and help to align targets with attributes in the compatible block. Across two experiments, we consistently found that IAT effects did not reflect rigid associations, but instead depended on whichever information could be used for similarity constructions between targets and attributes in different variants of the IAT. The findings are discussed with regard to theoretical models of EC as well as in reference to prominent accounts of IAT performance.

KEYWORDS:

• Evaluative conditioning
• association formation
• propositional learning
• Implicit Association Test
• similarity construction

Disclosure statement

No potential conflict of interest was reported by the authors.

ORCID

Christoph Stahl http://orcid.org/0000-0002-9033-894X

Notes

1 Note that this assessment refers to direct measures that do not restrict controlled processing. By contrast, using a direct measure that imposes time pressure and/or requires participants to base their judgement on spontaneous feelings may allow for diverging predictions by dual process vs. propositional accounts.

2 The reported results are based on the $D_3$ measure. For this measure, all trials from practice as well as main blocks with a response latency greater than 10000 ms are excluded and response latencies on error trials are replaced by the sum of the block mean of correct responses and the block standard deviation of correct responses.

3 Other than that, we followed the specifications provided by Greenwald, Nosek, and Banaji (2003). For each family and IAT alike, the $D_f$ scores were calculated by subtracting the mean response time for trials in which the respective family shared a response key with positive attributes from the mean response time for trials in which it shared a key with negative attributes. This difference was then divided by the pooled standard deviation of both “positive block” trials and “negative block” trials for the respective family and IAT.

4 In this and the following experiment, positive (negative) family-specific D scores indicate that classification was faster when the family shared a response key with positive (negative) attributes.

5 Note that because the shape of this interaction is predicted by our account of the IAT effects, and because of the equivalence of t tests and F tests with one degree of freedom in the numerator (Maxwell & Delaney, 1990), the three-way interaction of US Valence, CS-US Relation, and IAT Type can be considered marginally significant by a one-tailed test (p = .056).

6 Moran and Bar-Anan (2013) found an unqualified EC effect in the “ending” IAT as well as in another indirect measure of evaluation, the sorting paired feature task (SPF). As to whether the unqualified EC effect in the SPF can also be explained in terms of similarity construction, we can only speculate. Nevertheless, a number of procedural similarities between the IAT and the SPF make this possibility somewhat plausible. First of all, the unqualified EC effect in the SPF is also based on a version of the SPF that compares the two “ending” families. As in the “ending” IAT, the relational information is therefore constant which renders it less salient than the US Valence information that varies across the two families. Second, as is the case in the IAT, the SPF requires an active classification of the target stimuli making it more prone to attempts of task simplification than for example the evaluative priming procedure. Last but not least, the SPF requires the classification of pairs of target and attribute stimuli into four categories (the Cartesian product of the two attribute categories and the two target categories) using four keys. It seems plausible to assume that participants would apply similarity construction based on US Valence, e.g. to memorise the key that represents the “ending melody” family and positive attributes as the “positive” key.