ABSTRACT
When the size of an association varies by the level of another variable, this is may be called heterogeneity, heterogeneity of effect, interaction, effect modification, or subgroup analysis. Formal tests for variation in the size of an association can be conducted. Several problems are described. Problem 1: Using subgroup P-values for claims of variation. Problem 2: Failure to fully include main effect terms when studying interactions. Problem 3: Incorrectly claiming no variation because P-values for a test of heterogeneity are large. Problem 4: Interaction on the multiplicative scale may mean no interaction on a difference scale. Or vice versa. Problem 5: Failure to report all subgroup estimates in an evenhanded manner – illustrated by a published study regarding racial bias in physician referral patterns.
