Panel Data Marginal Structural Model (MSM)

Standard regression conditions on time-varying confounders by including them as covariates, which blocks causal pathways when those confounders are themselves affected by prior treatment. MSMs instead re-weight observations so confounders are balanced without conditioning, correctly estimating the causal effect even in the presence of treatment-confounder feedback.

Sequential exchangeability (no unmeasured confounding at each time point) states that, conditional on the observed history of treatment and covariates up to time t, treatment assignment at t is independent of potential outcomes. If this fails at any wave, the IPTW weights do not remove all confounding and the causal estimate is biased. It cannot be tested from data alone.

Inspect the weight distribution at each wave. Common strategies include stabilized weights (dividing numerator and denominator propensity scores), weight truncation at a chosen percentile (e.g., 99th), and trimming units with near-zero propensity scores. Each approach trades bias for variance, so sensitivity analyses under different thresholds are advisable.

Technically yes, but with only two waves there is little treatment-confounder feedback to correct for. In that case, a doubly-robust estimator, DiD with propensity score weighting, or a simpler IPW approach at a single time point may be more efficient and equally valid.

Ordinary regression standard errors from the weighted model are incorrect because they do not account for variability in the estimated weights. Use robust sandwich standard errors via GEE, or bootstrap the entire procedure — both the weight estimation and the outcome model fitting — to obtain valid confidence intervals.