Evaluation of Methods for Unmeasured Confounding Control in Pharmacoepidemiology Using Healthcare Databases: An Empirical Study Concerning Drugs Inducing Prematurity

Background: Using medico-administrative databases in pharmacoepidemiological studies is becoming common nowadays. Nevertheless, their use is often hampered by unmeasured confounding. In recent years, it has been hypothesized that part of this unmeasured confounding could be recovered by an automated screening of a vast number of covariates in these databases; extension of causal inference methods have been proposed, such as the high-dimensional propensity score (HDPS), the G-computation (GC), and Targeted Maximum Likelihood Estimation (TMLE) but studies comparing their performance on medico-administrative databases are lacking.

Objectives: The main objective of this work was to compare the performance of these methods in terms of bias reduction using an empirical study concerning the potential risks of prematurity of drug exposure during pregnancy.

Methods: We used an extraction from the French National Health Data System (SNDS) from January 2012 to December 2014, which contains socio-demographic information, medication uses, and comorbidities of nearly 2.7 million deliveries. For each delivery, three trimesters (13-week periods) were considered: the trimester before the start of pregnancy (T0), the first trimester (T1), and the trimester before the delivery (TBD). A delivery was considered exposed to a drug during a given trimester if at least one dispensation of that drug had been recorded during that trimester. Three causal inference methods were considered: (i) HDPS with LASSO regression for covariates selection, (ii) GC with LASSO regression as outcome model, and (iii) TMLE. For HDPS, the scores were incorporated through weighting (overlap weights (OW) and matching weights (MW)) or regression adjustment. These methods were also compared to a reference model which includes socio-demographical covariates and drug exposures of interest during previous trimesters. The methods were firstly compared in terms of average bias on a reference set of 42 negative control drugs pertaining to preterm birth. Furthermore, 13 positive control medications were used to test the ability of the methods to reject correctly false null hypotheses.

Results: Overall, our results show that the average bias of all methods were lower than that of the reference model. Moreover, bias tended to be close between HDPS and GC and higher with TMLE. Regarding HDPS implementations, the average bias of OW and MW were similar and lower than that of regression adjustment. As for the positive controls, the number of significant associations found by HDPS and GC was 6 (45%) and 9 for TMLE (70%).

Conclusions: Our results confirm the interest of these methods in accounting for a part of the unmeasured confounding. They also indicate more comparable performance between HDPS and GC than with TMLE.