Dr Roseanne McNamee, School of Community Based Medicine.
Confounding should be always be a concern for any researcher who is using observational data to address questions about causes, whether these questions arising from the health sciences, social sciences or other disciplines.
Confounding is a study bias and, if not addressed, will lead to an incorrect conclusion about the size of the effect of X on Y. The bias can be in either direction – the effect can be underestimated or overestimated - and an effect can be suggested when none exists in reality.
In the health sciences for example, it has been claimed that the results from observational studies suggesting that hormone replacement therapy reduced the risk of heart attacks in women suffered from confounding bias and that there was no such benefit in reality. Many headline claims about exposures supposedly harmful to health may also turn out to be erroneous for similar reasons.
In this talk the underlying issue behind confounding are explored and a working definition of a confounding factor in a particular study is given. The utility of causal graphs in helping us to explore what is, and isn’t a confounder emerges.
The methods used – either at the design or analysis stage of a study - to try to prevent confounding, or to reduce it, are discussed. Finally, some residual problems – which often mean that we can never exclude confounding - are emphasised.
Download PDF slides of the presentation 'What is confounding?'