The Implicit Association Test

The IAT measures the strength and direction of automatic associations between objects and attributes by means of the speed and accuracy with which respondents sort stimuli into their belonging category. The IAT (Table 1) is usually composed of seven blocks. Three blocks (B1, B2 and B5 in Table 1) are pure practice blocks in which either object stimuli or attribute stimuli are sorted in their reference categories. The remaining blocks are the associative blocks that constitute the two IAT mappings conditions (e.g., Mapping A and Mapping B). Usually, the practice and associative practice blocks are composed of 20 trials each, while the associative test blocks are composed of 40 trials each. The two mapping conditions result in 60 trials each.

During IAT administration, respondents might be given a feedback. In case a stimulus is sorted into the incorrect category, a red cross appears on the screen and the response must be corrected to proceed with the experiment. If a feedback strategy is not included, the error response is not signaled to participants, and they can go on with the experiment.

The IAT effect is given by the difference between respondents’ performance in the two associative conditions, under the assumption that the categorization task is easier (i.e., lower response times and higher response accuracy) in the associative condition consistent with respondents’ automatic association. The strength and direction of the IAT effect is usually assessed by means of D-score (Greenwald, Nosek, and Banaji 2003), for which different algorithms are available (Table 2). The differences between the algorithms lie in the strategies used for treating error and fast responses, while the core procedure remains the same.

Irrespective of the algorithm, trials with a response latency exceeding 10,000ms must be excluded. When participants are given a feedback for their responses, the algorithms using a built-in correction (D1 and D2) must be used, according to which error responses are replaced by the time at the incorrect response increased by the time needed to correct it. All the other algorithms (D3 to D6) are used when IAT administration does not include a feedback. The error responses are replaced by the average response time of the block in which the error occurred increased by a standard penalty (i.e., either 600ms or twice the standard deviation of the block). The other feature distinguishing the algorithms regards the decision to include responses faster than 300ms or not.

Once the algorithm has been chosen and the error and fast responses have been treated accordingly, it is possible to compute the D-score following a 3-step procedure:

1. Compute the D-score for the associative practice block (i.e., \(D_{practice}\)) as the difference between the average response time in the two contrasting associative practice blocks. This difference is divided by the standard deviation of the pooled blocks trials.

2. Compute the D-score for the associative test blocks (i.e., \(D_{test}\)) as the difference between the average response time in the two contrasting associative test blocks. This difference is divided by the standard deviation of the pooled blocks trials.

3. Compute the actual D-score as the mean of \(D_{practice}\) and \(D_{test}\).

implicitMeasures also includes a function (IATrel()) for computing IAT reliability as the correlation between the D-scores computed on practice and on test blocks (Gawronski et al. 2017).

The Single Category Implicit Association Test

The SC-IAT (Table 3; Karpinski and Steinman 2006) has been proposed as an alternative to the IAT when the aim is to obtain a measure of the strength of the automatic association that is not dependent to the contrasted object. It results from a modification of the IAT procedure itself in which one of the object categories is dropped. The associative practice blocks are usually composed of 24 trials, while the associative test blocks are composed of 72 trials.

During the SC-IAT administration, respondents’ usually receive a feedback signaling whether the response was correct (a green “O”) or incorrect (a red “X”). A response time window (RTW) at 1,500ms is usually included in the administration, after which the stimulus on the screen disappears. Responses exceeding the RTW are considered non-response.

As in the IAT, the SC-IAT effect results from the difference in respondents’ performance between the two associative conditions, under the assumption that the task will be easier in the condition consistent with the automatically activated association. A modification of the D-score has been proposed to interpret the SC-IAT effect, the D (Karpinski and Steinman 2006). Only the two associative test blocks are considered for the computation of the D. Fast responses (responses under 350ms) are discarded, as well as non-responses (responses exceeding the rtw). Error responses are replaced by the average response time of the block increased by a standard penalty of 400ms. The difference between the average response time in the two contrasting conditions is then divided by the standard deviation of the correct trials in both conditions.

The raw_data data set

The raw_data contains the data from 162 respondents that completed a Chocolate IAT, one Milk Chocolate SC-IAT and one Dark Chocolate SC-IAT, along with their demographic information. Data will be published in Epifania, Anselmi, and Robusto (n.d.).

The IAT administration did not include a feedback strategy, hence only the D-score algorithms not including a built-in correction should be computed.

# upload the data set
data(raw_data)

# explore the dataset
head(raw_data)
#>   Participant latency correct trialcode                blockcode  response
#> 1           4    2592       1  reminder practice.sc_dark.Darkbad trial_imp
#> 2           4     628       1  darkleft practice.sc_dark.Darkbad trial_imp
#> 3           4     808       1   badleft practice.sc_dark.Darkbad trial_imp
#> 4           4     783       1 goodright practice.sc_dark.Darkbad trial_imp
#> 5           4    2059       1   badleft practice.sc_dark.Darkbad trial_imp
#> 6           4    1114       1 goodright practice.sc_dark.Darkbad trial_imp

str(raw_data)
#> 'data.frame':    84726 obs. of  6 variables:
#>  $ Participant: int  4 4 4 4 4 4 4 4 4 4 ...
#>  $ latency    : int  2592 628 808 783 2059 1114 608 663 771 676 ...
#>  $ correct    : int  1 1 1 1 1 1 1 1 1 1 ...
#>  $ trialcode  : Factor w/ 32 levels "age","alert",..: 31 5 3 20 3 20 5 3 20 3 ...
#>  $ blockcode  : Factor w/ 13 levels "demo","practice.iat.Milkbad",..: 4 4 4 4 4 4 4 4 4 4 ...
#>  $ response   : Factor w/ 46 levels "","0","1","19",..: 43 43 43 43 43 43 43 43 43 43 ...

The data set contains respondents IDs (Partcipant), the response times in milliseconds for each trial (latency), the response accuracy (correct, 1 for correct responses and 0 for incorrect responses), the trials and blocks labels of the implicit measures (trialcode and blockcode labels, respectively), and the type of response (response, identifying the trials of the implicit measures or of the demographic information).