Systematic error, or 'bias' is of particular importance in any epidemiological investigation, and should be avoided wherever possible. Biases will reduce the validity of any results obtained, whether it be by overestimating or underestimating the frequency of disease in a population or the association between an exposure and disease. The forms of bias covered here can only be minimised through careful study design and execution - they cannot be accounted for in the analysis. Although confounding is considered by many authors as a form of bias, it can be accounted for during analysis, and so is covered separately.

Bias can be introduced into a study through the selection of participants (selection bias), or through errors made in the classification of measurement of exposures/outcomes of interest (information bias). Many of the examples given here relate to observational studies rather than experimental studies, as the process of randomisation of treatment in experimental studies is intended to minimise biases. However, it should be noted that even experimental studies may be affected by bias.

Validity

As mentioned earlier, the validity of an estimate of disease occurrence is a measure of how well it can be extrapolated to the source population (or, in the case of external validity, to other populations). The same concept of validity applies to analytic studies, although in these cases it relates to the estimation of the association between an exposure and an outcome of interest. Both selection and information biases act to reduce the validity of an estimate.

Selection bias

Selection bias affects inclusion of individuals in the study and results in the study sample not being representative of the source population. This may occur in the initial selection process, or may be a result of nonresponse or losses to follow-up during the study period. As alluded to above, the process of bias introduction into descriptive analytic studies differs from that in analytic studies, due to different aims in these investigations. In descriptive studies, any process which results in the selection of a population which is not representative of the source population can bias the results of the study. However, in the case of analytic studies, the representation of the source population is less important, as long as the measure of association between the exposure and outcome is not affected. In these cases, if selection is associated with both exposure and outcome (for example, if 'diseased' individuals who were exposed to the factor of interest were underrepresented in the sample), then bias will result. Note, however, that if selection is 'only' associated with 'either' exposure status or outcome (for example, if diseased individuals in general were underrepresented, regardless of whether they were exposed to the factor or not), then bias will not result. These concepts are similar to the concept of differential and non-differential information bias described below.

Descriptive studies

In descriptive studies, the precise definition of the source population and the use of an appropriate method of selection of individuals from this is of paramount importance in order to minimise bias. However, even when this is achieved effectively, a lack of compliance amongst selected individuals ('non-response') can introduce considerable bias. Increasing compliance, through the use of reminders or incentives to participate, can help reduce this source of error.

Analytic studies

Sources of selection bias which are applicable to all analytic studies are:

  • appropriate definition of the source population (noting that selection pressures may be in place in any commercial environment in order to remove less healthy animals from the population)
  • non-response bias (if the association between exposure and outcome differs between responders and non-responders)
  • inappropriate selection of comparison groups (both groups should be from the same source population)

Some examples of additional types of selection bias which may affect particular study designs are described below. Note that case-control studies are not mentioned here as the predominant source of selection bias in these cases is non-response bias.

Case-control studies

Case-control studies are particularly susceptible to selection bias. A major source of selection bias in these studies is through the selection of the control group. As mentioned above, it is of vital importance that the control group comes from the same population as the case group (that is, if they happened to experience the outcome of interest during the study period, they would have been classified as a case instead of a control). Admission risk bias is a type of selection bias associated with hospital-based studies where cases and controls are selected from hospitals (or similar establishments), and so may have a different exposure profile to the general population (to which the results are generally intended to be extrapolated). However, selection of controls from other sources can also introduce biases (for example, these controls may not be members of the same population as the hospital-based cases). Detection bias is a form of selection bias which results from differential classification of disease status according to exposure status (and so is closely related to information bias).

Cohort studies

Selection bias in cohort studies is generally less likely than in case-control studies, as selection for participation in the study generally precedes the development of the outcome of interest. However losses to follow-up during the study can introduce selection bias. For example, a study may be conducted in order to investigate the effect of farm biosecurity on introduction of disease. It is plausible that farmers with poor biosecurity will be less likely to remain in the study, meaning that these farms individuals (which may be more likely to experience disease problems) will be lost from the study. Another type of selection bias may be seen in 'occupational' cohort studies in human epidemiology. In these studies, members of a particular occupation (who are easier to identify and recruit) are followed up over time and monitored for disease and death. However, these people are from a distinct subpopulation of 'healthy workers', and so may not be comparable to members of the general population. This 'healthy worker effect' is also seen in veterinary epidemiology - as distinct selection processes are in place in any commercial industry (whether it be a farm or a horse racing yard) in order to remove less healthy animals (both through culling and through death).

Information bias

Information bias results from errors in exposure/outcome measurement or classification amongst the individuals included in the study. In the case of analytic studies, this may be further classified as differential or non-differential, depending on whether the bias is associated with an individual's exposure to factors of interest, or experience of the outcome of interest.

Descriptive studies

Errors in measurement of the outcomes of interest (known as measurement error in the case of continuous variables, and misclassification bias in the case of binary or categorical variables) are a common example of non-differential bias. For example, misclassification bias may be seen if a test with imperfect sensitivity is used in a study, resulting in some diseased animals being incorrectly classified as non-diseased. This can result in errors in the measurement of the outcome of interest in the population (although it should be noted that techniques are available to adjust for imperfect test sensitivity and specificity, if the values of these test characteristics can be estimated).

Analytic studies

As with selection bias, biases introduced into analytic studies differ form those in descriptive studies to some degree due to the aims of the study, although the same concepts are present.

Non-differental information bias

Non-differential bias occurs when the chance of bias is not affected by the group the individuals belong to. Measurement error and misclassification bias are common examples of non-differential bias. An example of measurement error may be seen if the scales used in a study are not correctly calibrated, resulting in the incorrect recording of the weight of all animals sampled. As this form of bias affects both exposed and unexposed (or diseased and non-diseased, depending on the study type) similarly, it will reduce the strength of any associations present, and will increase the probability of a type II error.

Differential information bias

Differential bias occurs when the misclassification/measurement error for exposures or outcomes is different for the different groups being compared. This may take place through a number of stages in the investigation - common examples are incorrect results given by the participants, by the investigator or by other people involved in the study. Reporting bias refers to the purposeful misrepresentation of the truth by participants (i.e. lying), whereas recall bias may be seen in relation to exposure classification for retrospective studies (where participants may have differential recollection of exposures, dependant on whether the outcome of interest was experienced or not). Observer bias may result from differences in the classification of exposures or outcomes by the investigator or person assigning the classification, and can be reduced through the use of blinding techniques. For example, Boxer dogs may be more likely than other dog breeds to be diagnosed by vets as having mast cell tumours (even if they do not have them), due to a postulated breed predisposition. Differential information bias may lead to either a strengthening or a weakening of the estimated strength of association between exposure and outcome.