CausalGPS: Matching on Generalized Propensity Scores with Continuous Exposures

Naeem Khoshnevis, Xiao Wu, Danielle BraunOpenalex Unknown 2021 · DOI · OpenAlex

Abstract

Provides a framework for estimating causal effects of a continuous exposure using observational data, and implementing matching and weighting on the generalized propensity score. Wu, X., Mealli, F., Kioumourtzoglou, M.A., Dominici, F. and Braun, D., 2022. Matching on generalized propensity scores with continuous exposures. Journal of the American Statistical Association, pp.1-29.

Introduction: Beyond Binary Treatments

In the realm of causal inference, traditional methods have largely focused on binary treatments—situations where a subject either receives a treatment or a placebo. However, many real-world scenarios, particularly in environmental epidemiology and economics, involve continuous exposures. For instance, researchers may need to determine how varying levels of air pollution (PM2.5) or different dosages of a medication impact health outcomes.

The CausalGPS R package addresses this need by implementing a novel framework for matching and weighting on Generalized Propensity Scores (GPS). By moving beyond the "treatment vs. control" paradigm, the package allows for the estimation of dose-response relationships, providing a more nuanced understanding of causal effects across a sliding scale of exposure.

The CausalGPS Methodology

The package provides a comprehensive pipeline for causal analysis, designed to transform observational data into a balanced "pseudo-population" that mimics the conditions of a randomized controlled trial. The workflow consists of four primary stages:

  1. GPS Estimation: Using the SuperLearner library, the package employs ensemble learning to estimate the GPS. By combining multiple machine learning algorithms (such as XGBoost and Ridge regression), it ensures a robust prediction of exposure based on observed covariates.
  2. Pseudo-population Generation: This is the core innovation of the package. It offers two main strategies: GPS Caliper Matching, which matches observations within a specific distance of their GPS values, and Weighting, which assigns weights to observations to balance covariate distributions.
  3. Balance Assessment: Before estimating effects, the package facilitates a "diagnostic-first" workflow. Functions like check_covar_balance allow researchers to calculate absolute correlations between covariates and exposure to verify that confounding bias has been effectively removed.
  4. Exposure-Response Function (ERF): Once balance is achieved, the package estimates the ERF using parametric, semiparametric, or nonparametric models, visualizing how outcomes change across the full exposure spectrum.
Key parameters for generating pseudo-populations and assessing covariate balance in CausalGPS.

As illustrated in the parameter definitions above, the package gives users granular control over the causal inference approach (ci_appr) and the specific metrics used to evaluate covariate balance, such as the balance threshold (covar_bl_trs) and type (mean, median, or maximal).

Data Resources and Validation

To support reproducible research and testing, CausalGPS includes two primary data resources:

  • generate_syn_data: A flexible utility for generating synthetic datasets under various GPS specifications (e.g., linear or quadratic relationships). This allows users to benchmark the package's performance against known ground truths.
  • synthetic_us_2010: A high-fidelity dataset merging 2010 US Census, CDC, and mortality data. It is specifically designed for air pollution studies, using annual PM2.5 estimates as the primary exposure variable.

Key Innovations and Practical Considerations

The package introduces several critical advancements for practitioners:

  • The Caliper Innovation: By applying caliper matching to continuous exposures, the authors provide a solution to the "curse of dimensionality" that often makes matching difficult in continuous spaces.
  • Automated Ensemble Learning: Integrating SuperLearner discourages model-specific bias and promotes more scientifically sound causal estimates.
  • Scalability: The package includes multi-threaded support (nthread) to handle the computational intensity of matching on large-scale observational datasets.

However, users must remain mindful of certain limitations. The success of the matching process is highly sensitive to hyperparameter choices like the "caliper" and "scale." Furthermore, like all propensity score methods, CausalGPS cannot account for unmeasured confounders; the validity of the causal estimate remains dependent on the quality and completeness of the input covariates.

Conclusion

CausalGPS provides a robust computational toolkit for researchers grappling with complex, continuous exposure data. By prioritizing covariate balance and offering flexible matching and weighting strategies, it bridges the gap between theoretical causal inference and practical, large-scale data analysis.

Summary updated

Abstract

Provides a framework for estimating causal effects of a continuous exposure using observational data, and implementing matching and weighting on the generalized propensity score. Wu, X., Mealli, F., Kioumourtzoglou, M.A., Dominici, F. and Braun, D., 2022. Matching on generalized propensity scores with continuous exposures. Journal of the American Statistical Association, pp.1-29.
BibTeX
@inproceedings{khoshnevis2021causalgps,
  author = {Naeem Khoshnevis and Xiao Wu and Danielle Braun},
  title = {{CausalGPS: Matching on Generalized Propensity Scores with Continuous Exposures}},
  booktitle = {Openalex Unknown},
  year = {2021},
  doi = {10.32614/cran.package.causalgps},
}

Loading related content...

Figure