Faking of the Implicit Association Test Is Statistically Detectable and Partly Correctable

Abstract

Male and female participants were instructed to produce an altered response pattern on an Implicit Association Test measure of gender identity by slowing performance in trials requiring the same response to stimuli designating own gender and self. Participants' faking success was found to be predictable by a measure of slowing relative to unfaked performances. This combined task slowing (CTS) indicator was then applied in reanalyses of three experiments from other laboratories, two involving instructed faking and one involving possibly motivated faking. Across all studies involving instructed faking, CTS correctly classified 75% of intentionally faking participants. Using the CTS index to adjust faked Implicit Association Test scores increased the correlation of CTS-adjusted measures with known group membership, relative to unadjusted (i.e., faked) measures.

ACKNOWLEDGMENTS

This research was supported in part by NIMH grants MH-57672 and MH-01533, as well as a grant from the National Science Foundation (OMA-0835854) to the LIFE Science of Learning Center.

We thank Klaus Fiedler and Matthias Bluemke for providing the data used in Reanalysis 1.

Notes

Note. For half the participants, the positions of Combined Task 1 Blocks in each IAT were switched with those of Combined Task 2 Blocks, respectively. IAT = Implicit Association Test.

¹Index e was excluded in these simultaneous regressions because of a linear dependency on a combination of the other four. The dependency could have been solved by dropping other indexes, but the results gave no compelling reason to retain Index e, so it was dropped.

²More details for this reanalysis (and other reanalyses reported next) can be found in the original publications on which each reanalysis was based.

³The sample consisted of another group, which was composed of offenders committing violent and sexual assaults against adolescents (hebephiles; n = 14), but had not been convicted of a sexual offense against children. Although there were no differences between hebephiles and controls in their child–sex IAT scores, t(60) = 0.29, p = .77, the difference between child–sex IAT scores of hebephiles and pedophiles was marginally significant, t(45) = 1.94, p = .06. Given this ambiguity about hebephiles' IAT scores, the hebephile offenders were omitted from further analyses.

⁴In the three Fiedler and Bluemke's (2005) studies, this was Index a, and in the Brown, Gray, & Snowden study (see Brown, 2005) the next strongest predictor was Index b.

Note. Coefficient a = Unstandardized slope of the regression of an unfaked D score (in D units) on a previous baseline Implicit Association Test (IAT; in D units), reflecting the reliability of IAT measures. Coefficient b = Slope of the regression of the measure of faking success (D change; in D units) on the combined task slowing (CTS) index (in seconds), indicating the expected distortion of D measures that is predictable from CTS. Constant c = Intercept of the regression of CTS on D change, indicating the values of CTS (measured in seconds) associated with no change in IAT scores from unfaked to a faked IAT. Sample sizes reflect the faking as well as the non-faking group used in each coefficient computation (see text for details). Parentheses indicate the Slope a value that was not used in the calculation of the weighted row average (see text for details).

^a Present study.

^b Brown (2005).

^c Fiedler and Bluemke (2005).

Note. IAT = Implicit Association Test.

^a Present study.

^b Brown (2005).

^c Fiedler and Bluemke (2005).