Content area
Full Text
In a previous issue of Evidence Based Mental Health , we discussed the role that selection bias can have in introducing systematic error into studies (see Evid Based Ment Health 2007;10 :67-8). In this article we discuss measurement (or information) bias-the other major type of systematic error commonly encountered in epidemiological research (fig 1). Important general points about bias include the following:
Bias may be introduced by poor study design or poor data collection.
Bias cannot be "controlled for" at the analysis stage.
Bias risks researchers and readers drawing conclusions that are systematically different from the truth.
Bias can lead to an over or underestimation of an effect.
An awareness of how bias may be acting helps a reader interpret study findings more accurately.
Measurement bias occurs when information collected for use as a study variable is inaccurate. The incorrectly measured variable can be either a disease outcome or an exposure. Measurement bias can be further divided into random or non-random misclassification. We are more concerned with non-random misclassification, as this can spuriously inflate or reduce estimates of effect. Non-random misclassification can itself be divided into subtypes, including observer bias and recall bias .
RANDOM VERSUS NON-RANDOM MISCLASSIFICATION
Random misclassification (also known as non-differential misclassification) is often thought of as less worrying than non-random misclassification. It occurs when either an exposure or a disease outcome is classified incorrectly in equal proportions for any subject group in a study-that is, it is random. Therefore if there are errors in the classification of a disease (for example, having schizophrenia), then for random misclassification to have occurred this must be unrelated to any exposures being examined (for example, being from an ethnic minority). Conversely any misclassification of exposure must be unrelated to disease status. When random misclassification occurs, it leads to an underestimation of an effect, thus making two groups seem more alike than they actually are. Therefore the main consequence of random misclassification is that the effective sample size is reduced, estimates are biased towards the null hypothesis, and as a result associations that truly exist are not revealed. Most epidemiologists are less concerned with this type of error than when estimates are spuriously inflated-leading to the nickname "clean dirt" by some. However the role...