Interaction

What It Is

When a researcher is assessing the nature of a relationship between a treatment and an outcome, or an exposure and an outcome, it is of interest to determine whether the relationship is the same (e.g., same direction, same strength) across all the subgroups of a population.  If it isn’t the same, then there is said to be an “interaction” between the treatment (exposure) and the factor that determines the subgroups of the population.  For example, perhaps a new treatment works very well to achieve a target hemoglobin A1c in younger patients and not so well in older patients with diabetes.  That means there is an interaction between the treatment and age in terms of the treatment’s effect on diabetes control.  In other words, “what’s the effect of this treatment on hemoglobin A1c in patients with diabetes?”  Answer: “it depends on the patient’s age.”

Why It’s Important

Assessing whether interactions exist helps researchers and clinicians have a better understanding of whatever phenomenon they are studying.  In the example above, it would be helpful to know that a diabetes treatment does not work (or work well) in older patients.  More effective treatments can then be sought for them.  Also, sometimes in a simpler statistical analysis a factor will appear to be not related to an outcome when, in fact, it is under certain circumstances.  This would be missed if the researcher doesn’t check for interactions. 

There are some nuances around whether we are looking at an additive measurement scale or a multiplicative measurement scale which are not addressed here.

Conducting many tests for interaction can produce spurious significant results (type 1 error).  When examining data for potential interactions, a solid understanding of background information on the topic and a thoughtfully constructed putative causal pathway should be used to identify interactions more likely to have biological meaning.