BNP-For-Unmeasured-Confounding: A Bayesian Nonparametric Method to Adjust for Unmeasured Confounding with Negative Controls
Unmeasured confounding bias is among the largest threats to the validity of observational studies. Negative control variables have been introduced in the causal inference literature as a promising approach to account for unmeasured confounding bias by using auxiliary information. In this repo, we provide tools based on a Bayesian nonparametric method to estimate a causal exposure-response function (CERF) that, under assumptions, leverages information from negative control variables to address unmeasured confounding. We model the CERF as a mixture of linear models. A mixture model has the nice feature of capturing the potential nonlinearity of the shape of the CERF while maintaining computational efficiency and leveraging closed-form results that are available under the linear assumption.
In the src folder, we provide functions to fit a gaussian linear mixed model based on Bayesian nonparametric methods with and without using negative control exposures and outcomes.
In the simulation folder, we assess the performance of this method by simulation studies and demonstrate that it can recover the true shape of the CERF in the presence of unmeasured confounding.
In the the pm25-cardiovasculardisease-exposure folder, we apply this new estimation procedure to adjust for a potential unmeasured confounder when evaluating the relationship between long-term exposure to ambient
We welcome contributions and feedback about BNP-For-Unmeasured-Confounding. If you have any suggestions, please open an issue or submit a pull request.
The companion paper is hosted at https://arxiv.org/abs/2309.02631