Chapter 13 discussed many of the obstacles inherent in the analysis of data from observational studies, with bias due to confounding arguably the largest of these concerns. Confounding can be seen as an issue of “mistaken identity,” in which the cause of an observed effect is attributed to the wrong party. As an example, consider a cohort study undertaken to assess the efficacy of a treatment. In this cohort, younger people are more likely to receive the treatment and are less likely to experience the outcome of interest (Figure 15–1). If a treatment effect is observed, it is unclear whether the observed effect is due to the treatment or to the younger age of the patients receiving the treatment. That is, age and treatment are said to be confounded; equivalently, age is said to be a confounder. Formally, a confounder is any variable related to both the outcome of interest and the treatment under study. In our example, age affects both the event rate and which treatment the person receives.
Illustration of confounding. In this hypothetical scenario, age is a confounder because it is related to both the treatment of interest and the outcome under study.
Throughout this chapter, we will refer to nonrandomized “treatment.” This term is not restricted to a medication; it can refer to a medical procedure or any variable of interest. The key characteristic of this “treatment” is that it was not assigned at random during the study. To facilitate discussion, we will typically consider a binary treatment; that is, patients can be divided into two treatment (or exposure) groups. Everything presented in this chapter can be generalized to the case of multiple treatment groups.
Whenever the exposure to a “treatment” is not due to randomization (such as in an observational study), confounding is likely. If we fail to address the confounding when conducting the analysis, we could mistakenly attribute an observed difference to the treatment under study when, in actuality, the difference is due to a second factor. In this chapter, we introduce several analytical approaches commonly used to address confounding. We also discuss the clinical assumptions underlying each method and common pitfalls to avoid.
Suppose a patient with a recent diagnosis of hypertension asks for a treatment recommendation. Should the physician make this recommendation without meeting with the patient? Or would he or she first want to consider the patient's history, collect vital signs, etc.? This latter approach—making treatment decisions given certain patient characteristics—suggests a “conditional” approach to addressing confounding. That is, if treatment decisions are made conditionally, perhaps the treatment effect we are interested in estimating is the effect given other patient characteristics. This is the motivation behind regression adjustment. Although the details of regression models were covered in Chapter 14, we will briefly consider how a regression model can be ...