Statistics from Altmetric.com
If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.
In both experimental and observational studies, many researchers attempt, often implicitly, to identify causal relations among variables. In trying to understand the possible causal processes that might have generated their data, the concepts of confounding and mediation play a prominent role. The two phenomena are often confused, and indeed are not always readily distinguishable. In the present paper, I will present a brief, somewhat simplified, introduction to confounding and mediation. I will present basic defining criteria, how to distinguish the two and also the problem of cases in which the distinction is not clear, along with some final caveats.
The term confound arises from the Latin confundere, to pour together or mix.1 The English word confuse arises from the same Latin root (http://www.merriam-webster.com/). In the context of empirical research, the term confounding is most often encountered in situations where some “predictor” of interest, let’s call it a, is presumed to be associated causally with some outcome, say, c. However, there may an additional variable, b, that also is associated with the predictor of interest, a, and the outcome, c. In the broadest application of the term, the effects of a and b are said to be confounded—that is, mixed. (There are in fact several ways in which the term confounder or confounding are currently used in research methodology but for our purposes here we will focus on the most common one2.) This, however, is only a very broad use of the term. In order to distinguish it from mediation, we will use the more specific definition offered by methodologists, which is as follows. Confounding is present when the following conditions occur: (1) both the predictor of interest and the potential confounder must be associated with the outcome (a and b are …
Competing interests: None.