Academic journal article Demographic Research

Confounding and Control

Academic journal article Demographic Research

Confounding and Control

Article excerpt

Abstract

This paper deals both with the issues of confounding and of control, as the definition of a confounding factor is far from universal and there exist different methodological approaches, ex ante and ex post, for controlling for a confounding factor. In the first section the paper compares some definitions of a confounder given in the demographic and epidemiological literature with the definition of a confounder as a common cause of both treatment/exposure and response/outcome. In the second section, the paper examines confounder control from the data collection viewpoint and recalls the stratification approach for ex post control. The paper finally raises the issue of controlling for a common cause or for intervening variables, focusing in particular on latent confounders.

(ProQuest: ... denotes formulae omitted.)

1. Introduction

In a recent paper on trends in cardiovascular diseases (CVD) in Europe (Kesteloot, Sans and Kromhout 2006), the authors attribute the dramatic current decline in mortality in the Baltic States to a change in dietary habits, i.e. the greater consumption of vegetable oil for cooking and the progressive replacement of butter by low-fat margarine. Though nutrition does indeed play a significant role on the incidence of CVD, one may also postulate (Gaumé and Wunsch 2003) in the case of the Baltic States that the tremendous change from a communist regime to a liberal one has led to systemic repercussions in the whole society which have brought about both modifications in cardiovascular mortality (through major fluctuations in stressful events), in economic conditions, and in behaviours including nutrition. Societal transformations and contextual changes in the economic, social (including public health), and political spheres would therefore be a confounder masking the true relation between changes in mortality patterns and in nutritional ones. To put it simply, correlation of time series does not imply causation.

To take another example that will be discussed later in this paper, medical reports in the 1920s already pointed out the suspected links between tobacco and cancers, and a 1938 article in the journal Science suggested that heavy smokers had a shorter life expectancy than nonsmokers. In 1939, F.H. Müller also published a paper in German on the relationship between smoking and lung cancer (Bartecchi, MacKenzie and Schreier 1995, Freedman, 1999). Though one now knows that smoking is bad for one's health, actually it is only in the early 1950s that two influential case-control studies, by Wynder and Graham in the U.S. and by Doll and Hill in the U.K., showed that cigarette smoking was a plausible cause of lung cancer. The relationship was confirmed by two prospective studies conducted by Doll and Hill in the U.K. and by Hammond and Horn in the U.S. (Schlesselman 2006). Even then the relationship between lung cancer and cigarette smoking was hotly disputed by a respectable scientist such as R.A. Fisher who argued as above that correlation was not causation. When is a correlation causal and when is it the result of confounding? This is the issue tackled here, by recalling some well-known and lesser-known facts. On the other hand, I will not be concerned with the diagnosis of causation and the assessment of evidence (see Elwood 1988, chapter 8).

Confounding is not a new issue. In philosophy of science, Hans Reichenbach (1956) already showed the presence of screening-off effects between two variables in conjunctive forks, due to the existence of a common cause. For Reichenbach, a conjunctive fork is a causal structure where two (or more) effects have a common cause and where the effects are conditionally independent given the common cause: the association between the effects disappears when one controls for or conditions on the common cause. The example of the drop in atmospheric pressure causing both a storm and a barometer dip is well-known. Simpson' s paradox too is a classic example of confounding (Rouanet 1985). …

Search by... Author
Show... All Results Primary Sources Peer-reviewed

Oops!

An unknown error has occurred. Please click the button below to reload the page. If the problem persists, please try again in a little while.