Evaluation of Methods Adjusting for Unmeasured Confounding Using Large Healthcare Databases: An Empirical Study...

With the growing availability of large healthcare databases for clinical science, mitigating unmeasured confounding has emerged as a major issue in pharmacoepidemiologic studies. Extensions of causal inference methods to high-dimensional settings could help address this problem, but studies comparing their performance in real-world databases are still lacking. This study aims to compare the ability to reduce the measured and indirectly measured confounding of three causal inference methods adapted to a real-world high-dimensional database using a machine learning LASSO algorithm: G-computation (GC), Targeted Maximum Likelihood estimation (TMLE) and Propensity Score with overlap or stabilized inverse probability treatment weighting. This large-scale empirical study was based on the French National Healthcare Claims Database (SNDS), consisting of 2,172,702 pregnancies $\ge$  22 weeks of gestation over the period 2011-2014. We used a set of 42 negative and 13 positive reference drugs related to prematurity risk. For each reference drug, the logarithm of the odds ratio for prematurity and its 95% confidence interval were estimated using each method. The proportions of false positive and true positive associations were calculated and compared between the methods. All methods yielded fewer false positives than a crude model based on a minimal set of adjusted covariates. TMLE produced the lowest proportion of false positives (45.2%), followed by GC (47.6%). GC yielded the highest proportion of true positives (92.3%). Our results confirm the interest of causal inference methods exploiting the wealth of data in healthcare databases, especially GC in terms of performance and ease of implementation.