Methods to control for unmeasured confounding in pharmacoepidemiology: an overview

Abstract

Background Unmeasured confounding is one of the principal problems in pharmacoepidemiologic studies. Several methods have been proposed to detect or control for unmeasured confounding either at the study design phase or the data analysis phase. Aim of the Review To provide an overview of commonly used methods to detect or control for unmeasured confounding and to provide recommendations for proper application in pharmacoepidemiology. Methods/Results Methods to control for unmeasured confounding in the design phase of a study are case only designs (e.g., case-crossover, case-time control, self-controlled case series) and the prior event rate ratio adjustment method. Methods that can be applied in the data analysis phase include, negative control method, perturbation variable method, instrumental variable methods, sensitivity analysis, and ecological analysis. A separate group of methods are those in which additional information on confounders is collected from a substudy. The latter group includes external adjustment, propensity score calibration, two-stage sampling, and multiple imputation. Conclusion As the performance and application of the methods to handle unmeasured confounding may differ across studies and across databases, we stress the importance of using both statistical evidence and substantial clinical knowledge for interpretation of the study results.

Association of Oral Anticoagulants and Proton Pump Inhibitor Cotherapy With Hospitalization for Upper Gastrointestinal Tract Bleeding

“…Hospitalizations for other gastrointestinal bleeding (predominantly lower gastrointestinal with some gastrointestinal hemorrhages for which site not indicated, eAppendix §3) were analyzed as a negative outcome control. 13 …”

“…The potential magnitude of confounding by unmeasured factors associated with PPI co-therapy was assessed by considering both the association of former co-therapy with upper gastrointestinal bleeding hospitalizations (negative exposure variant) and that of current co-therapy for hospitalizations for gastrointestinal bleeding at other sites (negative outcome). 13 …”

Addressing unmeasured confounding in comparative observational research

“…Weak confounders include: (1) a confounder is weakly related to the outcome but is a strong predictor of the treatment assignment; (2) a confounder is weakly related to the treatment assignment but is a strong predictor of the outcome; and (3) a confounder barely predicts the outcome and the treatment. In this stage, the rule‐out approach can help quantify how strong and imbalanced a confounder would need to be in order to explain (“rule out”) an observed (or expected) result (apparent or observed treatment relative risk, notated as “ARR”). Specifically, this approach attempts to find all combinations of (1) the confounder‐outcome relationship (“ RR CD ”) and (2) the confounder‐treatment relationship (“ OR EC ”) that are necessary to move the observed (expected) treatment relative risk to 1.…”

“…Methodological options of such sensitivity analyses are determined by 2 critical components: (1) availability/origin of information on the unmeasured confounders; (2) the goal of the unmeasured confounding assessment. For the first component, we follow the general strategy of Schneeweiss and there are 3 scenarios: No information: no additional information nor reasonable assumptions on the unmeasured confounder(s); Internal information: information on the unmeasured confounder(s) is available from internal data; External information: external data may contain information regarding the unmeasured confounder(s). …”

“…Groenwold et al applied 3 analytical methods to 2 case studies to assess the impact of an unmeasured confounder, but all 3 methods rely on the availability of unmeasured information for a subset of the original study population. Uddin et al and Streeter et al provided more recent methodological reviews on the methods controlling for unmeasured confounding in observational research, where several new methods since Schneeweiss's paper were included and discussed. However, neither review focused on providing guidance for selecting appropriate methods across a variety of complex research scenarios.…”