Machine learning mega-analysis applied to the Response Time Concealed Information Test: No evidence for advantage of model-based predictors over baseline

The response time Concealed Information Test (RT-CIT) can help to reveal whether a person is concealing the knowledge of a certain information detail. During the RT-CIT, the examinee is repeatedly presented with a probe, the detail in question (e.g., murder weapon), and several irrelevants, other details that are similar to the probe (e.g., other weapons). These items all require the same keypress response, while one further item, the target, requires a different keypress response. Examinees tend to respond to the probe slower than to irrelevants, when they recognize the former as the relevant detail. To classify examinees as having or not having recognized the probe, RT-CIT studies have almost always used the averaged difference between probe and irrelevant RTs as the single predictor variable. In the present study, we tested whether we can improve classification accuracy (recognized the probe: yes or no) by incorporating the average RTs, the accuracy rates, and the SDs of each item type (probe, irrelevant, and target). Using the data from 1,871 individual tests and incorporating various combinations of the additional variables, we built logistic regression, linear discriminant analysis, and extra trees machine learning models (altogether 26), and we compared the classification accuracy of each of the model-based predictors to that of the sole probe-irrelevant RT difference predictor as baseline. None of the models provided significant improvement over the baseline. Nominal gains in classification accuracy ranged between –1.5% and 3.1%. In each of the models, machine learning captured the probe-irrelevant RT difference as the most important contributor to successful predictions, or, when included separately, the probe RT and the irrelevant RT as the first and second most important contributors, respectively.