Bias in Epidemiological Studies

While the results of an epidemiological study may reflect the true effect of an exposure(s) on the development of the outcome under investigation, it should always be considered that the findings may in fact be due to an alternative explanation¹.

Such alternative explanations may be due to the effects of chance (random error), bias or confounding which may produce spurious results, leading us to conclude the existence of a valid statistical association when one does not exist or alternatively the absence of an association when one is truly present¹.

Observational studies are particularly susceptible to the effects of chance, bias and confounding and need to be considered at both the design and analysis stage of an epidemiological study so that their effects can be minimised.

Bias

Bias may be defined as any systematic error in an epidemiological study that results in an incorrect estimate of the true effect of an exposure on the outcome of interest.¹

• Bias results from systematic errors in the research methodology.
• The effect of bias will be either above or below the true value, depending on the type of systematic error.
• Limited scope exists for the adjustment of most forms of bias at the analysis stage. As a result careful consideration and control of the ways in which bias may be introduced during the design and conduct of the study is essential in order to limit the effects on the validity of the study results.

Common types of bias in epidemiological studies

More than 50 types of bias have been identified in epidemiological studies, but for simplicity they can be broadly grouped into two categories: information bias and selection bias.

1. Information bias

Information bias results from systematic differences in the way data on exposure or outcome are obtained from the various study groups.¹ This may mean that individuals are assigned to the wrong outcome category, leading to an incorrect estimate of the association between exposure and outcome.

Errors in measurement may be introduced by the observer (observer bias), the study participant (responder bias) or the instruments being used to make the measurement, such as weighing scales (instrument bias).

Errors in measurement are also known as misclassifications, and the magnitude of the effect of bias depends on the type of misclassification that has occurred. There are two types of misclassification – differential and non-differential – and these are dealt with separately in the section on errors in epidemiological measurement.

• Observer bias occurs when there are systematic differences in the way information is collected for the groups being studied. This may be a result of the investigator’s prior knowledge of the hypothesis under investigation or knowledge of an individual's exposure or disease status. Such information may influence the way information is collected, measured or interpretation by the investigator for each of the study groups.

For example, in a trial of a new medication to treat hypertension, if the investigator is aware which treatment arm participants were allocated to, this may influence their blood pressure measurements. Observers may underestimate the blood pressure in those who have been treated, and overestimate it in those in the control group.

Minimising observer bias:

• Where possible, observers should be blinded to the exposure and disease status of the individual.
• Blind observers to the hypothesis under investigation.
• In a randomized controlled trial blind investigators and participants to treatment and control group (double blind randomised controlled trial).
• Development of a protocol for the collection, measurement and interpretation of information.
• Use of standardized questionnaires or calibrated instruments, such as sphygmanometers
• Training of interviewers.
• Loss to follow up is a particular problem associated with cohort studies. Bias may be introduced if the individuals lost to follow-up differ with respect to the exposure and outcome from those persons who remain in the study.
• Recall bias - In a case-control study data on exposure is collected retrospectively. The quality of the data therefore, is determined to a large extent on the patient's ability to accurately recall past exposures. Recall bias may occur when the information provided on exposure differs between the cases and controls. For example an individual with the outcome under investigation (case) may report their exposure experience differently than an individual without the outcome (control) under investigation. That is, cases may tend to have a better recall on past exposures than controls.

Recall bias may result in either an underestimate or overestimate of the association between exposure and outcome.

Methods to minimise recall bias include:

• Collecting exposure data from work or medical records
• Blinding participants to the study hypothesis.

2. Selection bias

Selection bias occurs when there is a systematic difference between either:

• Those participating in the study and those who do not
• Those in the treatment arm of a study and those in the control group.

That is, there are differences in the characteristics between those who are selected for a study and those who are not, and those characteristics are related to either the exposure or outcome under investigation

This can occur, for example, if participants are asked to volunteer for a study. People who volunteer are not likely to be representative of the general population, threatening the generalisability of the study results. Volunteers tend to be more health conscious than the general population.

• Selection bias in case-control studies

• Selection bias is a particular problem inherent in case-control studies, where it gives rise to non-comparability between cases and controls. In case-control studies, controls should be drawn from the same population as the cases, so they are representative of the population which produced the cases.

  Controls are used to provide an estimate of the exposure rate in the population. Therefore, selection bias may occur when those individuals selected as controls are unrepresentative of the population that produced the cases.

  The potential for selection bias in case-control studies is a particular problem when cases and controls are recruited exclusively from hospital or clinics. Such controls may be preferable for logistic reasons. However, hospital patients tend to have different characteristics to the wider population, for example they may have higher levels of alcohol consumption or cigarette smoking. Their admission to hospital may even be related to their exposure status, so measurements of the exposure among controls may be different from that in the reference population. This may result in a biased estimate of the association between exposure and disease.

  For example, in a case-control study exploring the effects of smoking on lung cancer, the strength of the association would be underestimated if the controls were patients with other conditions on the respiratory ward, because admission to hospital for other lung diseases may also be related to smoking status. More subtly, the effect of alcohol on liver disease could potentially be underestimated if controls are taken from other wards: higher than average alcohol consumption may result in admission for a variety of other conditions, such as trauma.

  As the potential for selection bias is likely to be less of a problem in population based case-control studies, neighbourhood controls may be a preferable choice when using cases from a hospital or clinic setting. Alternatively, the potential for selection bias may be minimised by selecting controls from more than one source. For example the use of both hospital and neighbourhood controls.

• Selection bias in cohort studies

Selection bias is less of problem in cohort studies compared with case-control studies, because exposed and unexposed individuals are enrolled before they develop the outcome of interest.

However, selection bias may be introduced when the completeness of follow-up or case ascertainment differs between exposure categories. For example, it may be easier to follow up exposed individuals who all work in the same factory, than unexposed controls selected from the community. This can be minimised by ensuring that a high level of follow-up is maintained among all study groups.

The healthy worker effect is a potential form of selection bias specific to occupational cohort studies. For example, an occupational cohort study might seek to compare disease rates amongst individuals from a particular occupational group with individuals in an external standard population. There is a risk of bias here because individuals who are employed generally have to be healthy in order to work. In contrast, the general population will also include those who are unfit to work. Therefore, mortality or morbidity rates in the occupation group cohort may be lower initially than in the population as whole.

In order to minimise the potential for this form of bias, a comparison group should be selected from a group of workers with different jobs performed at different locations within a single facility¹.For example a group of non-exposed office workers.

Alternatively, the comparison group may be selected from an external population of employed individuals.

• Selection bias in randomised trials

Randomised trials are theoretically less likely to be affected by selection bias, because individuals are randomly allocated to the groups being compared, and steps should be taken to minimise the ability of investigators or participants to influence this allocation process. However, refusals to participate in a study, or subsequent withdrawals, may affect the results if the reasons are related to exposure status.

References

1. Hennekens CH, Buring JE. Epidemiology in Medicine, Lippincott Williams & Wilkins, 1987.
2. Breslow NE & Day NE. Statistical Methods in Cancer Research. Vol. 1: The Analysis of casecontrol studies, IARC, 1980.

© Helen Barratt, Maria Kirwan 2009