Strong instrumental variables biased propensity scores in comparative effectiveness research: A case study in oncology

Abstract

Background and Objectives

Some medications require specific medical procedures in the weeks before their start. Such procedures may meet the definition of instrumental variables (IVs). We examined how they may influence treatment effect estimation in propensity score (PS)-adjusted comparative studies, and how to remedy.

Study Design and Setting

Different covariate assessment periods (CAPs) did and did not include the month preceding treatment start were used to compute PS in the French claims database (Sytème National des Données de Santé-SNDS), and 1:1 match patients with metastatic castration resistant prostate cancer initiating abiraterone acetate or docetaxel. The 36-month survival was assessed.

Results

Among 1, 213 docetaxel and 2, 442 abiraterone initiators, the PS distribution resulting from the CAP [-12; 0 months] distinctly separated populations (c = 0.93; 273 matched pairs). The CAPs [-12;-1 months] identified 765 pairs (c = 0.81). Strong docetaxel treatment predictors during the month before treatment start were implantable delivery systems (1% vs. 59%), which fulfilled IV conditions. The 36-month survival was not meaningfully different under the [-12; 0 months] CAP but differed by 10% points (38% vs. 28%) after excluding month −1.

Conclusion

In the setting of highly predictive pretreatment procedures, excluding the immediate pre-exposure time from the CAP will reduce the risk of including potential IVs in PS models and may reduce bias.

Introduction

Nonexperimental and pharmacoepidemiological studies relying on longitudinal health care databases are considered as a valuable complementary source of evidence to randomized controlled trial (RCT) evidence. Measurement issues aside, inferring causality from nonexperimental studies requires adjustment for confounding, i.e., the potential distortion of the true relationship between the assessed drug and the outcome of interest that may result from imbalance in outcome predictors between the treatment groups [[1], [2], [3]]. Propensity scores (PSs) are frequently used to minimize confounding in the analysis of health care databases as they can adjust for many proxies of the patients’ health status observable in secondary data [4]. A variety of ways to implement PS analyses in nonexperimental studies exists, including matching, weighting, stratification, and modeling [[5], [6], [7], [8]]. A PS is defined as the probability for a patient to start a treatment according to its baseline characteristics. Under some assumptions [5], the PS can act as a balancing score that mimics randomization in RCTs conditional on the observable factors, resulting in comparative cohorts with similar distribution of baseline characteristics.

At the time of comparing the effect of two treatments, pharmacy dispensing records and by extension their associated claims are considered reliable markers for drug exposure [9]. Consequently, in comparative cohort studies based on claims data, dates of initial dispensing are usually used as the beginning of exposure, as well as the end of the covariate assessment period (CAP) essential to the construction of a PS [6,10]. However, it is worth remembering that in its original definition, PS refers to the “probability of assignment to a particular treatment” [5], and so, to the time when the decision to treat was made. In some situations, this decision to treat may occur well before the actual treatment onset because of a specific pretreatment check (e.g., specific laboratory tests), or drug administration-related medical procedure (e.g., implantable chamber for chemotherapy). In such situations, treatment-specific services occurring in patient journeys before the treatment start but after the decision to treat was made may be strongly associated with the exposure (i.e., treatment proxies) but poorly associated with patient characteristics. Measurements of such pretreatment services therefore have properties similar to instruments (or instrumental variables–IVs) [11].

Adjusting for preexposure patient characteristics that have properties of instruments may amplify any residual confounding [[12], [13], [14]], especially strong treatment predictors [15]. In practical settings, it is sometimes difficult to identify such IVs and if so then the impact on treatment effect estimates remains unclear. Although the theoretical process of variable selection in PS construction have been broadly discussed [13,[16], [17], [18]], few illustrative examples of results issued from designs based on an IV-biased PS have been published and contextualized. The objective of this article is to illustrate a situation of strong treatment predictors in oncology through a comparative cohort study conducted in the French Nationwide health care database (Sytème National des Données de Santé–SNDS) [19,20], to exemplify how including IVs in a PS can alter the composition of the comparator groups and impact effect estimates, and to introduce solutions to address this issue.