Pervasiveness and correlates of implicit attitudes and stereotypes

Abstract

http://implicit.harvard.edu/ was created to provide experience with the Implicit Association Test (IAT), a procedure designed to measure social knowledge that may operate outside awareness or control. Significant by-products of the website's existence are large datasets contributed to by the site's many visitors. This article summarises data from more than 2.5 million completed IATs and self-reports across 17 topics obtained between July 2000 and May 2006. In addition to reinforcing several published findings with a heterogeneous sample, the data help to establish that: (a) implicit preferences and stereotypes are pervasive across demographic groups and topics, (b) as with self-report, there is substantial inter-individual variability in implicit attitudes and stereotypes, (c) variations in gender, ethnicity, age, and political orientation predict variation in implicit and explicit measures, and (d) implicit and explicit attitudes and stereotypes are related, but distinct.

Acknowledgments

This research was supported by the National Institute of Mental Health (MH‐41328, MH‐01533, MH‐57672, and MH‐68447) and the National Science Foundation (SBR‐9422241, SBR‐9709924, and REC‐0634041). The authors are grateful for technical support from N. Sriram, Ethan Sutin, and Lili Wu. Related information is available at http://briannosek.com/ and http://projectimplicit.net/

Notes

¹DEMO (Demonstration site): http://implicit.harvard.edu/ (moved from http://www.yale.edu/implicit/in January, 2003), SPLC (Southern Poverty Law Center site): http://tolerance.org/, and UP (Understanding Prejudice site): http://understandingprejudice.org/

²This demographics scheme does not follow 2000 US Census norms that distinguish between ethnicity (Hispanic, non-Hispanic) and race. More recent data collected at these websites follows this practice. Initial analysis of that data finds more than 9% of the web sample report Hispanic ethnicity when it is distinct from reporting race.

³The fifth block was 20, 25, 30, or 35 trials. This manipulation was reported by Nosek and colleagues (2005), who found that 40 trials in this reverse practice condition reduce an extraneous influence of task order in which the first combined pairing (B3, B4) interfered with performance of the second combined pairing (B6, B7).

⁴The seven blocks were presented in the order shown above, or with the sorting combinations of B1, B3, and B4 exchanged with B5, B6, and B7. Attribute labels and exemplars were presented in white font and concepts in green, all on a black background, to emphasise that concept items were to be categorised by their category membership, not whether they were liked or disliked. If a participant made an error in sorting during any of the response trials, a red “X” appeared just below the exemplar and remained there until they corrected the error. IATs were scored with the D effect size algorithm proposed by Greenwald and colleagues (2003). The difference between a person's mean response latencies in the two stimulus-pairing conditions (i.e., blocks B3/B4 versus B6/B7) is scaled by the standard deviation of his or her latencies pooled across the two conditions. This algorithm results in D scores with a possible range from −2.0 to 2.0, with zero representing no difference in response latency between the conditions. Standard deviations included both correct and incorrect response latencies. Effect size d and η_P ² reports are based on the D and its variability across individuals. Because there was a period during the 6-year study span in which errant response latencies were misrecorded, all error latencies were replaced with the mean of the correct response latencies in its response block plus a 600-ms penalty prior to calculating the SD (see Nosek et al., 2006b). For about 6% of participants, valid scores could not be calculated because of missing data—stemming either from technical data-transfer problems or participant dropout. Among the remaining 94%, scores were not calculated if any of several speed and accuracy thresholds were exceeded, thus signalling careless performance: these criteria were (1) going too fast (<300 ms) on more than 10% of responses across all critical blocks, (2) 25% of responses too fast in any one of the critical blocks, (3) 35% too fast in any one of the practice blocks, (4) making more than 30% erroneous responses across the critical blocks, (5) 40% errors in any one of the critical blocks, (6) 40% errors across all of the practice blocks, or (7) 50% errors in any one of the practice blocks. Together, across all 17 tasks, these criteria resulted in a median disqualification rate of 7%, with a range of 5 – 15%. Beyond these disqualification criteria, individual trial response latencies were not included in the calculation if they were too fast to be authentic (<400 ms) or so slow as to indicate interrupted attention (>10,000 ms).

⁵A variety of category labels were used for the categories “Black People” and “White People” including: Black Americans/White Americans, Black People/White People, African Americans/White Americans, and African Americans/European Americans. Variation of these labels had minimal effects on IAT performance.

⁶An alternative explanation is that the ambiguous “other people” category may be especially sensitive to other influences compared to coherent, specific contrast categories. However, this alternative does not explain the stronger explicit preference for fat people compared to thin people for the weight task.

⁷An important factor for age-related comparisons of IAT data is the potential impact of age-related slowing in average response latency. Slower responding can have an extraneous influence on IAT effects artefactually making them larger (Hummert, Garstka, O'Brien, Greenwald, & Mellott, 2002). The D score, an individualised effect size used here, mitigates the influence of average response latency on IAT effects (Greenwald et al., 2003).

⁸None of the data collected for the Child-race task included a 6-point political orientation measure. However, for the child-race task, and the other tasks, a similar pattern was observed when a 7-point political orientation item was used. In this case, conservatives showed stronger pro-White children biases than liberals did. Data using 7-point scale results are included in the effect size estimates in Table 6. More detailed summary data with the 7-point scale appears in the supplementary materials.

⁹Across the aggregated social group attitude topics, the IAT and self-report were also strongly correlated, r = .72, and across the aggregated social group stereotypes, the IAT and self-report correlation was r = .50.