Machine Learning to Analyze Single-Case Graphs: A Replication and Extension with Nonsimulated Data

Machine learning algorithms may adequately control for Type I error rate and power when analyzing single-case AB graphs, but the most promising models have mainly been evaluated on simulated data. Moreover, the characteristics of the graphs that contribute to decision errors remain undocumented. To address these issues, we applied two machine learning models to a previously published nonsimulated dataset containing nearly 17,000 AB graphs showing no change to examine the proportion of false positives. On average, one of the two models (i.e., support vector classifier) produced lower proportions of false positives than well-established methods to analyze AB graphs (i.e., the dual-criteria methods). Larger mean differences between the two phases, lower standard deviations, and negative trends all led to more false positives. These results further support the use of machine learning to analyze single-case graphs, but further replications by independent research teams using educational and clinical data remain necessary.