Causal learning with interrupted time series data

Abstract

People often test changes to see if the change is producing the desired result (e.g., does taking an antidepressant improve my mood, or does keeping to a consistent schedule reduce a child’s tantrums?). Despite the prevalence of such decisions in everyday life, it is unknown how well people can assess whether the change has influenced the result. According to interrupted time series analysis (ITSA), doing so involves assessing whether there has been a change to the mean (‘level’) or slope of the outcome, after versus before the change. Making this assessment could be hard for multiple reasons. First, people may have difficulty understanding the need to control the slope prior to the change. Additionally, one may need to remember events that occurred prior to the change, which may be a long time ago. In Experiments 1 and 2, we tested how well people can judge causality in 9 ITSA situations across 4 presentation formats in which participants were presented with the data simultaneously or in quick succession. We also explored individual differences. In Experiment 3, we tested how well people can judge causality when the events were spaced out once per day, mimicking a more realistic timeframe of how people make changes in their lives. We found that participants were able to learn accurate causal relations when there is a zero pre-intervention slope in the time series but had difficulty controlling for nonzero pre-intervention slopes. We discuss these results in terms of 2 heuristics that people might use.