7.1 Random error and bias

An estimate of a measure of association can be wrong for two kinds of reason. Random error comes from sampling variation; it shrinks as the study grows and is described by a confidence interval. Systematic error, or bias, is a difference between what the study estimates on average and the true value in the source population; it does not shrink with size, and no confidence interval reflects it. A study is internally valid if it is free of bias for the population it sampled. Whether its result applies to other populations, its external validity or generalisability, is a separate and usually more judgemental question.

Biases are traditionally grouped into three kinds:

1.

selection bias, from the way subjects come to be in the study or in the analysis;

2.

information bias, from errors in measuring exposure, disease or other variables;

3.

confounding, from the mixing of the effect of exposure with the effects of other factors.

The first two are flaws in how the data were obtained. The third is a property of the source population, present even in perfectly collected data, and it is the one the methods of the following sections are designed to remove.

Questions on this section

Stuck on something here? Ask below and it stays attached to this topic.