Definition of Bias
Bias is an error in the design, conduct, or analysis of a study that results in a deviation of the test statistic from its true value. This results in an incorrect estimate of an association between an exposure and the target population such that the test statistic generated deviates from its true value.
- Biases can’t be eliminated completely but can be reduced/minimized through proper study design.
- 2 major subtypes of bias:
- Selection bias: bias caused by some preference in the way subjects are chosen that makes the study groups different from the target population
- Information bias: bias that results in a systematic error in the way data are measured or collected
|Masking||Hides an existing relationship between the independent and a dependent variable|
|False relationships||Creates spurious relationships between the independent and a dependent variable|
|Overestimation||Exaggerates the size of an existing relationship between the independent and a dependent variable|
|Underestimation||Diminishes the size of an existing relationship between the independent and a dependent variable|
Selection bias is the error introduced when the study population does not represent the target population due to some selection preference.
- Results in a nonexistent association between the exposure and outcome
- A selection preference can exist:
- Between study groups
- Between the selected and excluded participants for the study
Common examples of selection bias
- Healthy worker effect: bias occurs because subjects recruited from a workplace environment tend to be healthier and more active than the general population
- Tends to lead to an underestimation of morbidity or mortality compared with the general population
- Loss-to-follow-up bias: occurs due to the data loss from participants who are lost during the course of the study due to loss of interest, adverse effect, clinical improvement, etc.
- Leads to a reduction of the effective sample size
- Bias results if the study groups become unmatched or if the subjects who are lost to follow-up differ in exposure and/or outcome from those who remain in the study.
- Susceptibility bias: In cases where a first disease predisposes to a second disease, the treatment of subjects with the first disease erroneously appears to predispose to causing the second disease.
- Survival bias: occurs when data are collected only from subjects who have survived an outcome or have met some other clinical criteria related to it
- Berkson’s bias: occurs due to subjects being chosen from a specific segment of the population that has a more or less intrinsic risk of having the outcome than the general population
- Also called Berkson’s fallacy or Berkson’s paradox
- For example, hospitalized patients have a greater risk for disease than the general population regardless of exposure; thus, in a study looking at the link between smoking and chronic obstructive pulmonary disease, choosing hospitalized subjects for the experimental group will result in a greater association than would be seen in the general public.
- The higher morbidity of hospital patients relative to the general population can create spurious associations between diseases.
Information bias results from systematic errors in the measurement of some exposure, outcome, or variable. The major types of information bias are misclassification bias, recall bias, interviewer bias, response bias, reporting bias, observer bias, ascertainment bias, and confirmation bias.
Examples of information bias
- Misclassification bias: results from the misclassification of exposure or health outcome for subjects in a study. People who do not have a disease are classified as having it, and vice versa.
- Common reasons include inaccurate records, different disease definitions, or different diagnostic criteria.
- Differential: caused by a measurement difference that exists between study groups
- Non-differential: caused by equally inaccurate or random measurements across all study groups
- Recall bias: caused by the differential (typically, enhanced) recall of events by case subjects compared with controls
- Interviewer bias: occurs when the interviewer influences the responses of the participant through the method, content, or style of their questioning
- Response bias: occurs when subjects report what the researchers want to hear because the subject wants to please the interviewer with their response
- Reporting bias: refers to the tendency of researchers to only report statistically significant results, motivated by the desire to publish
- Observer bias: occurs due to the differences that exist between observers in the way an outcome is measured or assessed or the difference in the way a subject acts when they are being observed
- Bias that results from the Hawthorne effect (see below)
- Ascertainment bias: occurs when certain subjects are more likely to be included in the final results than others
- Common sources include a group difference in the screening method or frequency, investigator knowledge of subject group assignments, or allocation preference for subjects into study groups.
- Confirmation bias: occurs due to the tendency of investigators to only include data that agree with their pre-existing beliefs and/or support their hypothesis
- Detection bias: occurs due to a difference in the way an outcome is measured or detected based on a particular subject characteristic (see Image 1).
- Experimenter bias: occurs due to a subject behaving differently based on the investigator’s expectations
- Bias that results from the Rosenthal effect (see below)
Hawthorne and Rosenthal Effects
- Refers to the tendency of subjects in a study to behave or act differently (i.e., work harder) when they know they are being watched
- Results in and is related to observer bias
- Especially common in psychiatric studies
- Bias caused by this effect is difficult to eliminate but can be reduced through hiding from the subjects the knowledge of when/how they are being observed.
- Refers to the tendency of subjects or investigators to behave differently based on other’s expectations
- Also called the Pygmalion effect
- For example, in a study of the efficacy of a new drug to treat depression, subjects in the treatment group report an improvement in depressive symptoms because the investigators who are assessing them expect subjects who are given the treatment to improve.
- Results in and is related to experimenter bias
- Bias that results from this effect can be prevented by blinding.
A confound is an additional variable other than the independent variable that has an effect on the dependent variable, causing an erroneous relationship to be inferred between them.
- Technically, a confounding variable is not a bias.
- Confounds can be accounted for through proper study design. The following techniques help reduce the effect of confounds:
- Randomization: Subjects are randomly assigned to study groups so that confounds are spread equally between groups.
- Controlling: All subjects are chosen so that they have the same confound.
- Matching: Each subject with a confound in a particular group is matched to a subject in the other groups with the same or similar confound.
- Confounds are common in observational studies, as it is more difficult or impossible to control for confounding variables.
- Confounding variables are different from effect modifiers, a third variable that contributes to the true relationship between the exposure and outcome (see Table 2).
|Creates the illusion of association||Yes||No|
|Modifies the nature of the relationship||No||Yes|
- Celentano, David D., ScD., M.H.S., & Szklo, Moyses, MD, M.P.H., DrP.H. (2019). More on causal inference: Bias, confounding, and interaction. In Celentano, David D., ScD, MHS, & Szklo, Moyses, MD, MPH, DrPH (Eds.), Gordis epidemiology (pp. 289-306). doi://dx.doi.org/10.1016/B978-0-323-55229-5.00015-2. Retrieved from https://www.clinicalkey.es/#!/content/3-s2.0-B9780323552295000152
- Althubaiti, A. (2016). Information bias in health research: Definition, pitfalls, and adjustment methods. Journal of Multidisciplinary Healthcare, 9(1), 211-217. doi:10.2147/JMDH.S104807. Retrieved from https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4862344/
- Mansournia, M., Higgins, J. P., Sterne, J. A., & Hernán, M. (2017). Biases in randomized trials: A conversation between trialists and epidemiologists. Epidemiology, 28(1), 54-59. doi:10.1097/EDE.0000000000000564. Retrieved from https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5130591/#S11title