Confounding: An Enigma?

The concept of statistical confounding is understood intuitively by most practicing epidemiologists and social scientists. However, attempts to define confounding explicitly have encountered numerous difficulties. In Chapter 5, we explained how selection bias can occur when the study groups are formed in a manner that does not entail randomization. Viewed as a problem of selection, this noncomparability of the groups leads naturally to consideration of the assignment mechanism. In the current chapter, the differences between study groups are seen in a somewhat different light. Rather than focusing on the assignment mechanism, we ask how the individuals differ across these groups.

Consider the observational studies that appeared to show a benefit of hormone replacement therapy (HRT) on coronary heart disease (CHD). The women who opted for HRT were thought by many to be different in certain important respects from those who did not. Generally, the HRT-users appeared somewhat healthier and more “health-conscious” than the nonusers. Consequently, the apparent reduction of CHD risk may have been caused by one or more preexisting health-related factors rather than to HRT. That is, the effects of HRT and these other factors could have been confounded.

In many situations like the HRT studies, biases can be ascribed either to selection or to confounding. The choice is to some extent a matter of preference and convention. Indeed, many methodologists, especially ...

Get Bias and Causation: Models and Judgment for Valid Comparisons now with O’Reilly online learning.

O’Reilly members experience live online training, plus books, videos, and digital content from 200+ publishers.