Generative models provide a powerful interpretive framework in neural reinforcement learning. However, given the complexity of behavior, models are often proposed in the absence of structurally different alternatives, poten-tially weakening their explanatory force. We employ a data-driven methodology (choice-wide behavioral associa-tion study; CBAS; Kastner et al. BioRxiv 2024), to examine such limitations. We use a dataset of human subjects performing a two-step task that was designed to test the interplay between model-based and model free learning (Gillan et al. eLife 2016). The subjects also answered psychiatric symptom questionnaires. Using a rigorous mod-eling framework, the authors concluded that compulsive behaviors and intrusive thoughts (CBIT) negatively cor-relate with model-based learning. CBAS breaks down the choices of subjects into short subsequences, and then, using resampling-based multiple comparisons correction, identifies those subsequences that significantly correlate with a covariate of interest. In this case, CBAS identifies multiple subsequences of choices that correlate with CBIT severity; however, the logic underlying those subsequences do not simply fit with the interpretation that dif-ferences in model-based planning underlie the correlation. We find that there are multiple ways for such sequences to occur, only some of which fit the interpretation of model-based planning. To understand features in the data that are not captured by the model, we use CBAS to compare the subjects’ choices to outputs generated by the model (using parameters fit to the subjects’ data). CBAS identifies many subsequences that consistently differ between the model and data, providing targets for future model design. Ultimately, we find that the strong relationship be-tween performance on the two-step task and CBIT does not uniquely support the interpretation that this stems from differences in model-based planning.