Bias refers to a methodological flaw in a research study which prevents a generalization of a sample population out to the entire population. It is a systematic error.
Errors in radiology research studies fall into one of two categories:
random error
systematic error/bias
Random error cannot be controlled, but it can be accounted for with the correct statistical technique. For example, blood pressure varies throughout the day: an individual’s blood pressure may be under- or overestimated depending on when the measurement was obtained and a study subject may be incorrectly labeled as hypo- or hypertensive due to this random variation. An appropriately low p-value improves our confidence that the results are not due to random error (usually set at <0.05 (5%) probability).
Systematic error/bias, on the other hand, cannot be accounted for with statistics. Returning to the blood pressure example, a sphygmomanometer that is incorrectly calibrated may consistently result in falsely elevated blood pressure measurements. Non-random clustering of variable attributes can flaw our ability to generalize to the general population. So can non-random gathering of the data by the radiology researcher.
Non-differential and differential bias
Bias can be categorized as non-differential or differential.
Non-differential bias occurs when members of all groups in a study are equally likely to be misclassified concerning the presence or absence of exposure or the health outcome of interest. This type of misclassification most often biases the study towards the null hypothesis of no difference in the effect of the interventions/exposures under evaluation 2.
Differential bias occurs when members of one group in a study are more often misclassified with respect to the presence or absence of exposure or the health outcome of interest. This type of misclassification may bias the study toward or away from the null hypothesis of no difference in the effect of the interventions/exposures under evaluation. The direction of the bias will depend on the specifics of the misclassification. For example, a faulty sphygmomanometer that overestimates blood pressure in patients receiving a more effective new antihypertensive than those in the control group will bias the study towards the null. However, a faulty sphygmomanometer that underestimates blood pressure in patients receiving a less effective new antihypertensive than those in the control group will bias the study away from the null.
There are multiple opportunities for bias to creep into a radiology study, some obvious and some subtle. It is the researcher's goal to eliminate any large biases and control or account for any smaller ones 3.