COmplex Statistical Approaches for modelling and quantifying the impact of drug exposure from medico-administrative data: application to PREGnancy

Background

Prescribing drugs during pregnancy is common. However, the information available to prescribers in terms of adverse effects is scarce. Population studies conducted using large and complex administrative health databases such as the French Health Insurance system database (SNIIRAM) are particularly appealing in such respect. These databases contain a great wealth of information regarding, notably, exposure to reimbursed drugs. On the down side, they are usually hampered by limited information on classical risk factors.

Objectives

The overall objective of this project is to evaluate, develop, and calibrate statistical tools to perform pharmacoepidemiological studies from medico-administrative databases. In particular, this project revolves around two methodological axes: (i) accounting for complex relationship between drug exposure and the outcome of interest, and (ii) accounting for measured and unmeasured confusion. This project also encompasses a strong applied component as it is specifically targeted toward the pharmacoepidemiology of pregnant women.

Methods

The quasi-exhaustive recording of drug deliveries will enable us to develop and implement models accounting for the complexity of the relation between drug exposure and the outcome of interest, e.g., by accounting for cumulative past exposures. Methods for estimating the risk of outcome attributable to drug exposure will also be investigated to assess the impact of drug exposure.
We will approach the issue of confounding from two complementary perspectives. First, capitalizing on the thousands of measured variables in the SNIIRAM, we will investigate procedures such as the high dimensional propensity score and we will think about their extensions to account for time-dependent exposure. The second perspective will be dedicated to account for unmeasured confounding through the use of instrumental variables in the context of survival analysis.
Data for this project are already available and concern a SNIIRAM extraction for all pregnancies in 2012-2014. In order to develop our methodological framework, this project will rely on a set of reference associations chosen a priori. This set will contain drug – outcome pairs with different degrees of certainty as regards the plausibility of their associations. The main outcome of interest will be the duration of pregnancy, which includes prematurity. The constitution of this reference set is ongoing and is based on a panel of experts (clinicians and pharmacologists) and on literature review. It will be made of (i) positive controls, (ii) negative controls to be used in the validation steps and (iii) drug-outcome currently discussed in the literature and for which our results will contribute to the debate, e.g., selective serotonin reuptake inhibitors and risk of preterm birth.

Perspectives

Over the last years, the participants of this project have developed an expertise on both its methodological and applied aspects. The scientific spin-off from this project could be at 2 levels. First, in the matter of prevention, acquiring refined knowledge on drug-induced risks of prematurity during pregnancy could help practitioners to make decisions concerning possible adjustments of treatments. Second, on a pharmaco-epidemiological point of view, the method used in the field of pregnancy could be transposed to other pharmacoepidemiology studies to assess the risks of drugs, by using National Health Insurance databases, whatever the population or drug studied.