Evaluating the Construct Validity of Implicit Association Tests using Confirmatory Factor Analytic Models

The Implicit Association Test (IAT) is the most widely used method for assessing implicit bias and prejudice. By avoiding the need for introspection, the IAT is suggested to be a more valid indicator of prejudice than explicit measures of attitudes (i.e. questionnaires). However, implicit attitudinal literature has demonstrated highly variable associations between IAT scores and various outcomes. Such inconsistencies imply IAT scores may be significantly influenced by measurement error, which could thwart efforts to accurately estimate underlying attitudes. The aim of the present thesis was to examine the construct validity of the IAT using Confirmatory Factor Analytic models (CFA) to account for the confounding influences of measurement error.

Three studies examined various aspects of the validity of IATs using data from 198 student participants of the University of Tasmania, Australia. Study One assessed the internal consistency and internal convergent validity of traditional verbal IATs, fully pictorial IATs and Affective Priming Tasks (APTs) using single-group CFA. The study revealed high amount of random error variance in the implicit attitudinal data, comprising around 55% of IAT scores and 95% of APT scores. Despite the high proportion of random error, the IATs appeared to consistently assess the trait attitude constructs, though this was not true for the APTs. The APTs were consequently deemed invalid measures of implicit attitudes.

Study Two added to the findings of Study One by further accounting for method variance in the IAT data using the CTCM CFA-MTMM analytical approach. This study indicated that method variance accounts for a further third of the IAT scores, suggesting that an average IAT score is comprised of around 80% error variance (random error and method variance). Notwithstanding this, after accounting for measurement error, strong convergence was evident between the verbal and pictorial IAT formats and two of the four IATs were found to possess good construct validity. Such findings provided some optimism for the future development of psychometrically robust implicit attitude techniques.

Study Three examined the application of IATs to assess implicit attitudes whilst using latent modelling techniques to account for the significant error component of the scores. Specialised CFA models were used to reveal anti-Arab/pro-European bias in the present sample, as well as determine the effect of certain participant characteristics, such as age, on the IAT scores. In summary, the studies of this thesis suggest that IAT scores are likely to be confounded substantially by error variance at the individual level. However, if random error and method variance are partialled out, IAT scores can provide an adequate assessment of implicit attitudes. This suggests that future IAT applications would profit from analysing sample data using CFA or other latent modelling techniques to account for the significant error component of IAT scores.