Confounding vs Effect Modification — Never forget after this!

I have always struggled with confounding and effect modification. I know the definition — confounders are extraneous variable that cause apparent association between exposure and outcome, when none exists, because the confounders are related to both the exposure and the outcome.

Whereas, effect modifiers are those that also cause an apparent association between the exposure and outcome, when one does exists, but this is by virtue of the effect modifiers’ association with the outcome.

Confused — yeah, me too!

So, let’s clear it out with an example.

A hypothetical prospective study wants to measure the association between use of hormonal therapy in postmenopausal females and Deep Vein Thrombosis (DVT). The results show that the hormonal therapy when compared to controls not on hormonal therapy, does increase the risk of DVT with the relative risk of 2.5 (95% Confidence Interval: 2.1–2.8). Now, a few other cross-sectional studies have shown that smoking among postmenopausal females is also a risk factor for DVT. So, we want to find whether that’s the case in our study.

So, we do a sub-group analysis also known as stratified analysis. For this, we divide our study participants according to this extraneous factor i.e. into smokers and non-smokers.

Let’s say that we get three scenarios.

In scenario A, we find that in the smoker group, the hormonal therapy increases the risk of DVT with the RR of 2.8 (95% CI: 2.5–3.1) but in the non-smoker group, the RR of DVT with hormonal therapy is 1.1 (95% CI: 0.8–1.5). So, clearly in this case the hormonal therapy does not by itself increase the risk of DVT. There is some modification of causal pathway by smoking, that is responsible for the increased in risk of DVT with hormonal therapy. Hence, smoking is an effect modifier in this scenario.

In scenario B, we find that in the smoker group, the RR of DVT with hormonal therapy is 0.8 (95% CI: 0.7–1.2) whereas in the non-smoker it is 0.9 (95% CI: 0.8–1.1). So, this means that the initial result that hormonal therapy increased the risk of DVT was not true. In fact, it was smoking was confounding the association between hormonal therapy and DVT. (In order to find the association between smoking and DVT, we can now ignore the hormonal therapy and divide the entire cohort to smokers and non-smokers and then test the association with DVT.)

In scenario C, we find that in the smoker group, the RR of DVT with hormonal therapy is 2.5 (95% CI: 2.3–2.8) whereas in the non-smoker group it’s 1.8 (95% CI: 1.5–2.1). So, this means that smoking is neither a confounder nor an effect modifier.

In summary, if in a sub-group (based on suspected confounder/effect modifier) analysis, the original association between the exposure and outcome doesn’t hold up in BOTH sub-groups the factor is a confounder. But if the original association holds up in one group but not in another, then the factor is an effect modifier.

Do leave comment if you have a better way of explaining or if anything is not clear!