You are an emergency physician working at a secondary care hospital. The institution's board wants to provide the best possible treatments for patients with acute myocardial infarction (MI) and is considering expanding its cardiac catheterization facility to better serve candidates for catheterization.1 You are requested to assess the evidence regarding that decision.
You are aware of a systematic review of randomized clinical trials (RCTs) comparing a routine invasive treatment (cardiac catheterization) vs conservative therapies (pharmacological therapy, with selective catheterization only for patients with recurrent symptoms or objective inducible ischemia) and showing a relative decrease of 18% in mortality or MI in the more invasively treated group.2 A colleague argues that in the real world the effect may be much larger, with as much as a 50% relative decrease, as suggested by a cohort study of more than 120 000 Medicare patients hospitalized with MI, some of whom received invasive treatment (ie, catheterization within 30 days) and others medical treatment.3
Unsure from where best estimates of the effect of invasive vs conservative therapy should come, you review the report from the large cohort study. You find that the investigators used several adjustment methods: multivariable risk adjustment, propensity score adjustment, propensity-based matching, and instrumental variable analysis (IVA). These analyses showed variable results, ranging from a 16% relative reduction—very similar to that achieved in the RCTs—to a 50% relative decrease in mortality in the more invasively treated group. You want to determine what inferences you can draw from the RCTs as compared with the observational trials that used differing adjustment methods.
Appraising Studies on Therapy and Harm
Large randomized trials ensure that treatment and control groups are balanced with respect to factors associated with outcomes (typically age, sex, disease severity, and comorbidity—“prognostic factors”) and thus provide the optimal approach to address questions about the benefits and harms of interventions.4,5 However, RCTs may not be available or feasible, particularly regarding questions of harm. For example, one may be interested in long-term outcomes years after exposure to an intervention, or in rare serious adverse effects, detection of which will require data from tens of thousands of patients. Randomization also may be unethical, as in the study of risky behaviors. In such instances, the best evidence will come from observational studies. Sophisticated analyses of longitudinal databases are also increasingly cited in the context of comparative effectiveness research. Enthusiasts claim that such studies address limitations in the generalizability of RCTs and thus may provide best estimates of real-world effectiveness.6-8
However, observational designs can often produce misleading results, because real-life treatment decisions are typically influenced by patient characteristics likely to be associated with the outcomes of interest (ie, they are likely to be prognostic factors).9,10 Sophisticated statistical adjustment mechanisms may overcome the problems of prognostic imbalance. The purpose of this Users' ...