Link to video transcript in a Word file

Most health outcomes are multifactorial, meaning that many factors can influence the risk of developing an outcome, and this complexity often distorts the measurement of association for a relationship we are trying to assess. For example, there are many established risk factors for coronary heart disease (CHD), but suppose we wanted to conduct a prospective cohort study to understand the impact of physical activity on risk of developing CHD. We could enroll a cohort of adults without known CHD and assess their physical activity and perhaps divide the cohort into quintiles based on their physical activity. We could then follow the cohort over time and eventually do an analysis to compare the incidence of CHD among the groups. The problem is that the subjects who exercise the most are probably systematically different from those who exercise the least, and the groups probably differ with respect to other factors that affect the likelihood of CHD developing. People who exercise regularly are likely to be more health conscious, and they are more likely to eat a healthy diet, maintain a healthy weight, get their blood pressure checked regularly, and take vitamins. In addition, they are less likely to smoke and less likely to have diabetes or hypertension. All of these things put them at less risk for CHD, but we want to measure the independent effect of physical activity, i.e., independent of all these other risk factors. When other risk factors that affect the outcome are unequally distributed among our comparison groups, the measure of association will be biased as a result of confounding.

Yet another problem is that the magnitude of association between an exposure and a health outcome might differ depending on the presence of another risk factor. For example, the association between heavy smoking and lung cancer might have a risk ratio of 20 in the general population, but the risk ratio for smoking and lung cancer might be about 60 among shipyard workers who smoke heavily and used to install asbestos in ships as fireproofing. This is a separate phenomenon from confounding called effect measure modification. In essence, it means that the effect of an exposure on an outcome is modified by another factor. In this example, it means that the effect of heavy smoking on risk of lung cancer is greatly magnified among people who also worked with asbestos.

In this module we will take a closer look at confounding and effect measure modification, and you will learn methods of identifying them. Effect measure modification is a biological phenomenon that should be described and reported. In contrast, confounding is just a mathematical distortion, and you will learn ways to prevent confounding and ways to adjust for its distorting effects when it cannot be prevented.

Essential Questions

  1. How can we take into account other factors besides the exposure under investigation that influence the exposure-outcome relationship?
  2. What happens if there are multiple causes that affect an exposure-outcome relationship?
  3. What does confounding mean?
  4. How can we prevent it? How do we know if it is a problem in our study?
  5. How should we deal with confounding in data analysis?
  6. What is effect measure modification (EMM)? How do we evaluate for EMM in a study?

Learning Objectives

After completing this section, you will be able to: