Confounding in Observational Research

Abstract

Related Article, see p 636KEY POINT: Confounding distorts the relationship between exposures and outcomes in observational research and must be accounted for during data analysis to reduce bias.In this issue of Anesthesia & Analgesia, Maheshwari et al1 report results of an observational study on the association between perioperative hypotension and postoperative delirium. Authors used a multivariable Cox regression to assess this relationship while controlling for potential confounders.2 In observational research, systematic differences between subjects with or without a certain exposure (eg, hypotension) may distort (confound) the relationship between the exposure and an outcome if the variable(s) that differ also affect the outcome. Such variables, related to both the exposure and the outcome, but not involved in the causal pathway between the exposure and the outcome, are termed confounders (Figure).3Figure.: The proposed relationship between the exposure (perioperative hypotension) and outcome (postoperative delirium) in Maheshwari et al.1 A variable related to both the exposure and the outcome (emergency surgery) is termed “confounder.” More specifically, a confounder is a common cause of the exposure and the outcome (pathway “A”). Emergency surgery could be a proxy for factors that cause both hypotension and delirium, and could hence distort the relationship. A factor that is in the causal pathway—such as cerebral hypoperfusion—is a mediator (pathway “B”).For example, in the study by Maheshwari et al,1 patients with postoperative delirium had relatively more often undergone emergency surgery. Emergency surgery may also be associated with hypotension (eg, due to patient-related factors). Hence, the apparent relationship between hypotension and delirium has 2 components: (1) a bona fide effect of hypotension on delirium (if there is any effect) and (2) a spurious effect that is due to the relationship with the confounder(s). An analysis that disregards confounding may give a biased estimate of the effect of the exposure. Multivariable regression techniques can control for confounding4 by holding the confounding variable(s) constant. This is similar to the principle of stratification, in which the effects of hypotension would be assessed in each stratum (subgroup) of patients with or without emergency surgery. Within each stratum, the level of the confounder is constant and can thus no longer confound the relationship. Unlike stratification, multivariable regression can simultaneously adjust for a larger number of confounders, as done by Maheshwari et al.1 Recently, propensity score methods have become a common alternative. The propensity score is the probability of being in a certain exposure group, given a patient’s baseline characteristics.5 This score can then be used to weight the patients’ observations or to match patients, such that the exposure groups are well balanced with respect to the potential confounders. Importantly, all these techniques can only adjust for observed confounders, so that residual bias by unobserved confounders is still quite possible. Whatever technique is being used, researchers must determine which variables should be considered as potential confounders. Various approaches exist, and P value–driven stepwise selection approaches are especially common for regression modeling. However, a confounder does not have to be “statistically significant” to distort a relationship; therefore, this approach carries a risk of excluding relevant confounders. Anesthesia & Analgesia rather advocates a liberal adjustment, for as much confounding as possible. Theoretical considerations about the assumed relationships between variables are a good basis for selection of potential confounders.5 Importantly, a variable in the causal pathway between exposure and outcome is not a confounder, but a mediator. Adjusting for such a variable must be avoided because this diminishes the effect of the exposure of interest.3

Topics & Concepts