Confounding versus Effect Modification: Essential differences

In the recent past I have received questions regarding the difference between confounding and effect modification. In this article I will attempt to provide a simplified explanation of both the terms and mention important differences between the two. This article may be useful for Post Graduates, researchers, and those pursuing the Basic Course in Biomedical Research.

Background Information:

Often, one observes that individuals diagnosed with a particular condition (say A) tend to have common attributes. (For instance, those with cardiovascular disease are often physically inactive, smokers, consume an unhealthy diet, and may be overweight or obese.) The frequent occurrence of these attributes with a condition makes one wonder about the relationship between the attributes and the condition under consideration- do these attributes have a role in causing the condition?

Fundamentally, epidemiological studies aim to determine the relationship between attributes or suspected risk factors and various conditions. This knowledge helps develop strategies and interventions for both prevention and control of the condition.

The problem is that epidemiological studies may generate results that suggest one kind of relationship whereas the true relationship is something else. This happens when there is bias (systematic error that distorts the measure of effect of an exposure on outcome), effect modification, or interaction.

I have discussed bias in previous articles, so will limit this article to confounding bias and effect modification.

Key Messages:

Confounder: A confounder is a factor that is associated with the exposure under scrutiny, is not part of the causal pathway, and can independently cause the outcome.

To use an analogy: A man (M) and a woman are in a relationship and the woman’s best friend (F) (who always accompanies W and knows M well) also has feelings for M (unknown to W). Eventually, F marries M. Here, F is a confounder as not only was she associated with both M and W, she can have a relationship with M independently of W.

The problem is that we do not always know if a factor is a confounder or not. When we are aware that a factor is a confounder, we refer to it as a known confounder. When we are unaware that a factor is a confounder, it is called an unknown confounder. Confounding distorts the effect of exposure on outcome, creating a bias. Therefore, we try to minimize confounding. Known confounders are dealt with by matching while unknown confounders are dealt with by randomization (randomization deals with both known and unknown confounders).

To cause confounding bias, a variable must be a risk factor for the disease among non-exposed persons, must be associated with the exposure of interest in the population from which the cases derive, but must not be an intermediate step in the causal pathway between the exposure and the disease.

Examples of Confounding:

A study plans to investigate the association between L-tryptophan (the precursor of serotonin) and eosinophilia-myalgia syndrome. However, L-tryptophan is indicated to treat depression and insomnia- both conditions associated with myalgias including eosinophilia-myalgia syndrome. Here, L-tryptophan is a confounder (confounding by indication).
A crude analysis suggests that coffee drinking causes lung cancer. However, when adjusted for cigarette smoking the association disappears. The effect was on account of coffee drinkers being more likely to smoke cigarettes than abstainers. It is cigarette smoking that increases the risk of lung cancer, not coffee drinking.

Confounding is study specific and results from the particular study design. Thus the effect of confounding is likely to vary even among studies investigating the same exposure and health outcome.

Effect modification: When an exposure has different effect in different subgroups or strata, effect modification is suspected. An effect modifier is associated with the outcome but not the exposure.

Unlike confounding, effect modification is not a bias as it does not result in systematic error. On the contrary, effect modification provides important information- that the magnitude of the effect of (an) exposure on (an) outcome will vary according to the presence of a third factor.

Effect modification is a true characteristic of the association between an exposure and an outcome. Therefore, it is constantly present regardless of study design and with all else being equal must be present to the same extent in all cohorts.

Example of Effect Modification: A study investigating the relationship between low-level lead exposure and child (mental) development discovered that mental development index (MDI) scores were lower for children with higher cord blood lead levels. On further analysis investigators discovered that cord blood lead levels were associated with lower MDI scores if children belonged to lower social class. While children belonging to high social class manifested adverse MDI scores only if the lead levels were ‘high’, lower social class children adverse MDI scores at medium blood lead levels. Here, social class is an effect modifier.

Summary:

A confounder is associated with both exposure under consideration and outcome.

Confounding is a bias and thus must be minimized.

The extent of confounding depends on the particular study (design) and may vary from study to study even if the same exposure and outcome are being investigated.

Effect Modification reflects a true relationship between exposure and outcome in which a third factor modifies the effect of exposure on outcome. The effect modifier is associated with outcome, not exposure.

It is identified when exposure has different effects in different subgroups/ strata of a third variable.

Effect Modification is not a form of bias and cannot be minimized.

The extent of Effect Modification is expected to be similar across cohorts and not influenced by study design.