Bias is an important issue in observational studies, and it should be carefully considered when interpreting the results of such studies. In this article, I will highlight the main types of bias encountered in observational studies.
Bias is a systematic error in the design and methods of the study, leading to an incorrect interpretation. It is important that bias is considered during the design and conduct of the study because it cannot be corrected afterward. Bias should be distinguished from random error, which is related to the variability in the sampled population and can be reduced by increasing the sample size.
Selection bias occurs when the selected study participants are systematically different in characteristics from eligible participants who are not selected for the study. Additionally, when the exposed and unexposed groups are different in important outcome predictors. the results might be biased.
If a study is assessing the frequency of visits to the dentist among participants in wealthy and poor neighborhoods via a questionnaire, no response (self-selection), especially by participants less likely to visit the dentist, will result in biased estimates caused by a special form of selection bias called nonresponse bias . It is possible that the nonresponders are less likely to be reached because they live in a poor neighborhood, or that they are embarrassed to respond since they do not visit the dentist regularly.
Selection bias also arises from what has been termed the healthy-worker effect , which indicates that working participants, who are more likely to be healthy, are compared with less healthy participants, introducing bias in the results. This is usually an issue when the general population is selected as the comparison group.
Information bias refers to systematic errors in measurement (or misclassification) of the exposure or outcome ( Table ).
|Nondifferential misclassification||Differential misclassification|
|Consequences of information bias (reporting/recall and observer bias)||Reduces the strength of the association||Can reduce or increase the strength of the association|
Spencer et al examined the association between exposure to water fluoridation and the increase of dental caries in 2 Australian states, South Australia (SA) and Queensland (Qld). “Children were enrolled between 1991 and 1992 (SA: 5-15 yrs old, n = 9,980; Qld: 5-12 yrs old, n = 10,695). Follow-up caries status data for 3 years (± 1/2 year) were available on 8,183 children in SA and 6,711 children in Qld.” The authors reported that “Baseline data on lifetime exposure to fluoridated water, use of other fluorides and socio-economic status (SES) were collected by questionnaire, and tooth surface caries status by dental examinations in school dental service clinics.”
In this study, a potential source of bias can be related to the use of the questionnaire and the possibility that information on exposure is incomplete and inaccurate ( reporting bias or recall bias ). If for some reason the completeness and accuracy of the responses are associated with the outcome—ie, more accurate among participants with healthier dentitions or the opposite—then the results of the study might be biased because of differential misclassification . Differential misclassification can overestimate or underestimate the association of interest. If the inaccuracies are similar among the exposed and the unexposed subjects, then we have what we call nondifferential misclassification , which tends to underestimate potential associations by making the groups more similar.
We can also have information bias from biased recording of the outcome ( observer bias ): dental examiners might record dental caries incorrectly, especially if they are not blind to the exposure.