A Practical Guide to Causal Inference in Three-Wave Panel Studies

Causal inference from observational data poses considerable challenges. This guide explains an approach to estimating causal effects using panel data focussing on the three-wave panel design. **Part 1: Pre-specification of Causal Estimands for a Target Population** considers the first step: how to ask a causal question by clearly pre-specifying a causal contrast for a well-defined exposure on well-defined outcomes in the population of interest. **Part 2: Three-Wave Panel Design** discusses the methodology for obtaining causal effect estimates from three-wave panel studies and discusses issues of bias from sampling, attrition/missing responses, measurement error, and unmeasured confounding. Here we discuss methods for avoiding these biases, which should be considered in advance of data collection, but which is inevitable, even with the best-made plans. We describe strategies for handling these biases. **Part 3: Statistical Estimands and Estimators** describes the process of converting observational data into consistent causal effect estimates for the targeted causal estimands. Here we consider conventional parametric estimators, as well as more recently developed non-parametric and semi-parametric machine learning methods. **Part 4: Pre-registration, Data Analysis, and Reporting** describes the protocols for pre-registering analyses, conducting these data analyses, and clearly and accurately communicating scientific findings. **Part 5: Addressing Complex Causal Questions** discusses methods for addressing complex causal questions relating to treatment-effect heterogeneity, causal interactions, causal mediation, and longitudinal treatment strategies. We examine how the approaches discussed in Parts 1 - 4 may be cautiously adapted to handle these complex causal questions, and why social scientists should tread lightly before attempting to answer them. Overall, we hope to provide a clear, step-by-step guide that applied researchers may use to obtain robust causal inferences using three waves of longitudinal data.