Subset Analysis: Insights and Pitfalls

An Example of Subset Analysis

In a clinical trial of the effect of reduced blood transfusion on postoperative morbidities after cardiac surgery, the target population is the patients undergoing coronary artery bypass grafting or heart valve surgery. A representative sample of 1,500 eligible patients is recruited from a clinical center. Since the number of transfused red blood cell units during an operation can not be predetermined, these patients are randomized to two transfusion triggers: (1) a liberal transfusion trigger that requests a unit of blood being transfused whenever the patient's hematocrit level (%) drops below 28 during the operation and (2) a conservative transfusion trigger that requests a unit of blood whenever the hematocrit level (%) drops below 24. Under this design, the conservative transfusion group would receive less transfused blood than the liberal group. The primary analysis is to compare, between the two treatment groups, the mean rate of the primary end point, a postoperative composite morbid outcome. In this way investigators are able to draw a conclusion on whether the conservative transfusion strategy is more beneficial than the liberal transfusion strategy for the target population.

In the primary analysis, the two randomized groups of patients are treated as a whole, and the within-group heterogeneity is ignored. However, the investigators speculate that the treatments may have different effects on patients with different body sizes (body mass index <18.5, 18.5–24.9, 25–30, >30), as patients of smaller body size are at higher risk of anemia if insufficient blood is transfused. The primary analysis cannot tell investigators whether the treatment effect is the same across different subsets and, if different, which patient subsets might benefit more from the conservative transfusion strategy. Subset analysis is the statistical analysis exploring such heterogeneity. People sometimes use the term quantitative heterogeneity for the case where one treatment is always better than the other across all levels of the subset variable (e.g., body mass index levels) but the magnitude of benefit varies; the term qualitative heterogeneity is often used for the case where one treatment is better than the other in some levels of the subset variable but worse in other subsets. Qualitative heterogeneity is rare in clinical studies.

...