Objective: There is a need for rigorous methods to study the mechanisms that lead to individual-level change (i.e., process-outcome research). We argue that panel data (i.e., longitudinal study of a number of individuals) methods have 3 major advantages for psychotherapy researchers: (1) enabling microanalytic study of psychotherapeutic processes in a clinically intuitive way, (2) modeling lagged associations over time to ensure direction of causality, and (3) isolating within-patient changes over time from between-patient differences, thereby protecting against confounding influences because of the effects of unobserved stable attributes of individuals. However, dynamic panel data methods present a complex set of analytical challenges. We focus on 2 particular issues: (1) how long-term trajectories in the variables of interest over the study period should be handled, and (2) how the use of a lagged dependent variable as a predictor in regression-based dynamic panel models induces endogeneity (i.e., violation of independence between predictor and model error term) that must be taken into account in order to appropriately isolate within-and between-person effects. Method: An example from a study of working alliance in psychotherapy in primary care in Sweden is used to illustrate some of these analytic decisions and their impact on parameter estimates. Results: Estimates were strongly influenced by the way linear trajectories were handled; that is, whether variables were "detrended" or not. Conclusions: The issue of when detrending should be done is discussed, and recommendations for research are provided. What is the public health significance of this article? This article provides recommendations on how to study psychotherapy processes using dynamic panel data models to strengthen causal inferences. Accurate estimates of what drives individual development in psychotherapy are needed to generate recommendations on what therapists should focus on in therapy. Using the alliance-outcome association as an example, we show that estimated effect sizes may vary greatly depending on which modeling approach is used, with the decision on whether to remove time-trends from the outcome variable making the largest difference.