Statistics in Brief: Instrumental Variable Analysis: An Underutilized Method in Orthopaedic Research

Abstract

Background The randomized controlled trial (RCT) design is considered the most robust approach of the primary research study designs. However, RCTs are expensive, often are highly selective (that is, inclusion criteria are restrictive and not always generalizable), and are not always feasible because of ethical concerns or because impractically large sample sizes are needed. Observational studies that use routinely collected data (such as from electronic medical records or larger national datasets from claims and registries) provide an alternative approach and may address some of these limitations. However, observational data often are not captured with a future study in mind, and for that reason and others, studies drawn from them are susceptible to confounding, which occurs when a third variable is related both to a study’s dependent and independent variables. For example, when evaluating the association between sex and the risk of an anterior cruciate ligament injury, the type of sport involved is an obvious confounding factor. While various methods exist that address confounding in observational studies, instrumental variable analysis can address both measured and unmeasured confounding. In general terms, an instrumental variable analysis mimics an RCT in that an “instrumental variable,” defined as a variable that is strongly correlated with the treatment but not the outcome, is used to represent a mechanism for assigning a treatment to patients. Following this theoretical treatment assignment, standard statistical methods are used to calculate treatment effects. When adequately performed, an instrumental variable analysis can be superior to more commonly used multivariable regression or propensity score analyses as it addresses confounding factors that are not always available in observational datasets. Moreover, it provides an additional opportunity for researchers to demonstrate robustness of results across varying statistical approaches. This Statistics in Brief feature accompanies a study by Chan et al. [6], in which we used population-based data to evaluate closed wound drainage and outcomes in 105,116 patients undergoing shoulder arthroplasty. In that study, we used three statistical methods to account for confounding and demonstrate similar results across these methods: mixed-effects regression [10], propensity score-based matched analysis [2], and instrumental variable analysis [11]. Although the first two methods are commonly used, studies applying an instrumental variable analysis are relatively rare, especially in orthopaedics research. A PubMed search of the years 2016 to 2018 with “Instrumental Variable” as a keyword resulted in 25 publications in oncology, 47 in cardiology, and six publications in orthopaedics. Despite its underuse in medical and surgical observational research, instrumental variable analysis has some important advantages over other multivariable statistical approaches. The purpose of this article is to provide a brief overview of instrumental variable analysis, its underlying assumptions, its value, and its limitations in the context of the study by Chan et al. [6] guided through the following two questions: (1) What is an instrumental variable analysis? (2) What is its importance in orthopaedic research? Discussion Unlike RCTs [15], observational studies must address, and ideally control for, confounding; this can be addressed by multivariable approaches such as regression modeling [10] and propensity score analysis [2]. These, however, can only address known confounding factors, not unobserved confounders such as information on surgeon decision-making or local postoperative protocols. By contrast, instrumental variable analysis aims to address both known and unknown confounders, a major advantage. The basic principle behind instrumental variable analysis [3, 11] is choosing an instrumental variable—also referred to as the “instrument”—to represent a mechanism for assigning treatment to pa