Longitudinal Ex Post Facto Design — Tracking Pre-Existing Groups Over Time

The terms overlap substantially in practice. Both follow participants over time and observe outcomes without experimental manipulation. The ex post facto label emphasizes that the defining group characteristic (the 'independent variable') is a past event or condition that pre-dates the study's measurement start, whereas prospective cohort studies measure baseline exposure status at enrollment and follow participants forward. In a longitudinal ex post facto design the exposure may be identified retrospectively from records, while a prospective cohort assembles participants before outcomes occur and measures exposure prospectively. Both face the same core limitation: no random assignment.

A minimum of two waves is required to call a design longitudinal. However, two waves only allow you to measure net change; they cannot distinguish different patterns of change (e.g., rapid early gain versus steady linear growth). Three or more waves are needed for growth-curve modeling. The required number of waves depends on the shape of the expected change trajectory — consult substantive theory and prior literature to determine an appropriate measurement schedule.

Propensity score matching and related methods reduce observed confounding and make the study more rigorous, but they only balance on measured covariates. If important confounders are unmeasured, matched estimates remain biased. Causal language ('X caused Y') is therefore not warranted; you may say the findings are consistent with a causal effect, or that a causal interpretation is plausible given the controls applied.

First, analyze and report attrition patterns — compare baseline characteristics of completers versus dropouts overall and by group. If dropout is not random (missing not at random), listwise deletion produces biased estimates. Use full information maximum likelihood (FIML) estimation or multiple imputation to handle missing data under the missing-at-random assumption, and consider sensitivity analyses for non-random dropout.

Yes, if existing archival or administrative data are available or if a short longitudinal window (6–18 months) is feasible within the dissertation timeline. Collecting primary longitudinal data over multiple years is rarely practical for a dissertation; exploiting existing panel datasets (e.g., PISA, NLSY, BHPS, national educational databases) is the standard strategy. Ensure your institution's IRB/ethics committee approves secondary data use and that data access agreements are in place before committing to the design.