Addressing confounding

Addressing confounding

Confounding can be addressed in the study design and the data analysis. Let’s go over these approaches.

Study design

Individual or group matching on potential confounding variables can eliminate the impact of confounding.[10] For instance, to eliminate age as a confounder in a cohort study, one could match exposed and unexposed individuals to have the same distribution of age groups in both groups. If an association between an exposure and outcome were observed, one would know that the association is likely not due to differences in age between the exposed and unexposed groups because matching was performed in the study design.

Data analysis (stratification)

We can address confounding in the analysis of data by stratification or adjustment.

Stratifying on a potential confounding factor enables us to assess whether the association between an exposure and outcome differs by levels of a third variable. In the absence of confounding, we would expect that the same direction and a similar magnitude of association between the exposure and outcome in each stratum as in the crude analysis. A difference in the direction and/or magnitude of association in the strata would indicate the presence of confounding. Essentially, by examining each stratum separately (i.e., holding the confounding variable constant), we’re able to assess whether the crude measure of association persists by levels of the confounding variable.

For example, let’s say that in a hypothetical case-control study, we observed an association between gender and malaria, with an odds ratio of 1.71 (Table 7).[1] We’d like to assess whether work environment (indoor or outdoor) is a potential confounder in the association.

Table 7. Hypothetical case-control study of male gender as a risk factor for malaria infection

From Szklo M, Nieto FJ. Epidemiology: beyond the basics. Burlington, MA: Jones & Bartlett Learning; 2014.

First, we observe strong associations of the outdoor work environment with both exposure and outcome (Table 8). This is characteristic of a confounder.

Table 8a. Associations of the confounder with the exposure and outcome – Confounder vs exposure

Table 8b. Associations of the confounder with the exposure and outcome – Confounder vs outcome

From Szklo M, Nieto FJ. Epidemiology: beyond the basics. Burlington, MA: Jones & Bartlett Learning; 2014.

To further investigate, let’s conduct a stratified analysis by work environment. If the outdoor environment is a confounder, we would expect that the OR should be different in either direction or magnitude in the two stratum compared with the crude OR of 1.71. We can see that the odds ratio of the gender-malaria association is closer to 1.00 in each stratum (Table 9). This would suggest that work environment is a confounding factor that can explain the positive association between male gender and malaria.

Table 9a. Association between gender and malaria stratified by work environment – Mostly outdoor

Table 9b. Association between gender and malaria stratified by work environment – Mostly indoor

From Szklo M, Nieto FJ. Epidemiology: beyond the basics. Burlington, MA: Jones & Bartlett Learning; 2014.

Adjustment

Confounding can also be addressed by adjusting (or controlling) for the confounding variable.[1] Various statistical approaches exist for estimating adjusted measures of association. For example, the Mantel-Haenszel method produces adjusted odds ratios or relative risks, in the form of an overall weighted average of stratum-specific effect estimates.  Other methods for confounding adjustment include: standardization (direct and indirect), multiple regression models, instrumental variables, propensity scores (Module 10), and marginal structural models.