Properties of a confounder

Properties of a confounder

Confounding is an important issue in observational epidemiological studies. It refers to a situation in which a noncausal association between a given exposure and an outcome is observed as a result of the influence of a third variable (i.e., the confounder).¹ 

A confounder must be: 

1. associated with the exposure
2. causally related to the outcome
3. not affected by the exposure

Figure 4. Causal diagram representing the confounding effect on the observed association between exposure and outcome.

Let’s take the example of the case-control study on the relationship between coffee and pancreatic cancer.¹⁰ If it’s found that an association between coffee drinking and pancreatic cancer exists, we need to ask whether a causal relationship between the two actually exists (Figure 5), or whether it might be explained by the influence of a confounding variable. 

Figure 5. Causal diagram showing a hypothetical causal relationship between coffee drinking and pancreatic cancer.

Consider a third variable: smoking. Does it meet the characteristics of a confounder? Using the three characteristics of a confounder:

1. smoking is associated with coffee drinking
2. smoking is a risk factor for pancreatic cancer
3. smoking is not affected by (or caused) by coffee drinking.

Smoking meets the three characteristics of a confounder, given that smokers are more likely to drink coffee, smoking is a risk factor for pancreatic cancer, and smoking in itself is not affected by (or caused by) coffee drinking. Therefore, the observed association between coffee drinking and pancreatic cancer could be said to be confounded by smoking (Figure 6).

Figure 6. Causal diagram showing smoking as a confounding factor in the observed association between coffee drinking and pancreatic cancer.

Confounding is especially an issue in observational studies because the allocation of exposure groups is nonrandom. In comparison, the use of randomization in experimental studies reduces the likelihood that treated and untreated groups differ with regard to known and unknown confounding factors, though confounding due to random differences can still occur.¹ Observational studies are more likely to be impacted by known and unknown factors related to the exposure of interest.