Estimating generalized propensity scores with survey and attrition weighted data

Area of Science:

• Causal inference
• Observational data analysis
• Statistical modeling

Background:

• Prior research established the utility of survey weights in binary treatment causal inference.
• Extending these methods to continuous treatments and incorporating attrition weights remained unexplored.
• Generalized propensity score (GPS) analyses are increasingly used for continuous treatment effects with observational data.

Purpose of the Study:

• To extend prior work on survey weights to continuous treatments using GPS.
• To investigate the impact of survey sampling and attrition weights on GPS estimation and outcome modeling.
• To assess the robustness and necessity of using weights in different stages of GPS analysis.

Main Methods:

• Developed analytic results extending prior work to continuous treatments and weights.
• Conducted a simulation study to evaluate different weighting strategies in GPS analysis.
• Examined the role of weights in both propensity score and outcome model stages.

Main Results:

• Using survey or attrition weights in GPS estimation and outcome modeling yields more robust continuous treatment effect estimates.
• While weighting in both stages is sufficient for robustness, it is not always necessary for unbiased estimation.
• Simulation results indicate potential bias reduction when weights are applied in both stages, offering a safeguard against unknown conditions.

Conclusions:

• Incorporating survey and attrition weights is crucial for robust causal inference with continuous treatments in observational studies using GPS.
• The findings provide practical guidance for analysts handling weighted data in complex causal analyses.
• Applying weights in both GPS estimation and outcome modeling serves as a prudent strategy to mitigate potential bias.