🧠 The Problem With Post-Treatment Confounders
Post-treatment confounders make causal inference for time-varying treatments difficult. Conditioning on these variables can block causal pathways or create spurious associations, producing biased marginal effect estimates. Marginal structural models (MSMs) paired with inverse probability weighting (IPW) are commonly used to avoid this bias, but IPW has important drawbacks:
- Requires modeling the conditional distributions of treatment
- Highly sensitive to model misspecification
- Relatively inefficient and prone to finite-sample bias
- Difficult to apply with continuous treatments
🧪 How Residual Balancing Constructs Weights
Residual balancing offers an alternative way to build weights for MSMs by modeling the conditional means of post-treatment confounders rather than the full conditional distributions of treatment. Key features include:
- Models conditional means of post-treatment confounders (not treatment distributions)
- Produces weights used in MSM estimation
- Naturally accommodates continuous treatments
📊 Evidence: Simulations and Empirical Examples
Numeric simulations show that residual balancing is generally more efficient and more robust to model misspecification than IPW and common IPW variants across a range of scenarios. The method is illustrated with two applied examples:
- Estimating the cumulative effect of negative advertising on election outcomes
- Estimating the controlled direct effect of shared democracy on public support for war
Open-source software is available to implement residual balancing.
❗ Why It Matters
Residual balancing provides a practical, more robust alternative to IPW for researchers using MSMs to study time-varying treatments. By shifting modeling effort from treatment distributions to confounder means, the approach improves finite-sample performance and makes analyses with continuous treatments more tractable, while reducing sensitivity to misspecification.
