Epidemiologists have a lot of words that start with "C": Causal (inference). Cohort. Collider. Collinearity. Counterfactual. Case-Control. However, a very important word that requires a lot thought in study design and statistical analysis is "Confounding".

In many epidemiological studies, we are interested in studying & quantifying if an exposure is associated with an outcome. However, to do so, you also need to think about other variables that may confound (distort) your association of interest.

Modern Epidemiology (a go-to textbook for epidemiologists) defines a confounders as factors that “explain or produce all or part of the difference between the measure of association and measure of effect that would be obtained with a counterfactual ideal.” (3rd ed, page 58)

Sometimes, you’ll hear a confounder described as a variable that is associated with both the exposure and outcome of interest, but does not fall on the causal pathway between the two (i.e., the exposure does not cause the confounder).

In theory, if a researcher has collected all confounding factors of interest and adjusts for them when quantifying the association between the exposure and outcome, then the measure of association will approximate the effect measure we wish to study (i.e., be unbiased).

However, many epidemiologists conduct studies with secondary data (i.e., data that has already been collected) and sometimes there are unmeasured confounders. So how do we deal with these unmeasured confounders?

One way to combat this is to calculate an “E-value” ( http://bit.ly/3qjFcny ), which describes the minimum strength or magnitude a confounder would need to nullify an association of interest.

If your E-Value is quite small, this would suggest the unmeasured confounder may be an important threat to the magnitude and significance of your findings.

Hope you enjoyed this thread on #ConfoundingFactors and #EValues!