Biostatistics and Epidemiology for the Toxicologist: Information Bias—Differential and Non-Differential Misclassification (Part I)

Bias is the term used in research to describe systematic, non-random errors that can distort the results of a study and potentially lead to incorrect conclusions [1]. Notably, bias violates assumptions of statistical models by incorporating non-random error and thus renders statistical results invalid. Though limiting bias is a hallmark of fundamentally sound research, random errors (or “noise”) will inevitably occur.

Bias can occur during any part of the research process including study design, data collection, data analysis, interpretation, and publication. While many types of bias exist, certain types are particularly relevant to Medical Toxicology; understanding and recognizing sources of potential bias is essential for researchers as well as for “consumers” of research who must be able to critically evaluate literature.

To start, study design and patient recruitment can be subject to bias even before a single patient is enrolled. A study that contains unrecognized bias by the investigators could generate data with systematic flaws that cannot be addressed later on in the research process when data collection is complete. For example, choosing data collection methods is an important area where bias can be introduced. When feasible, objective and validated measuring tools should be used as they have higher inter-rater reliability than subjective measures which inherently have a larger observed variance [2]. In scenarios where outcomes do not have validated measuring tools, it is imperative to develop unbiased methods of assessment to minimize non-random error. Blinding study personnel to a patient’s exposure and/or outcome status and blinding data analysts to treatment allocation groups are a few common techniques used to limit bias. Moreover, standardizing data collection protocols also limits variability when multiple parties are gathering and entering data. Lastly, bias can occur even at the conclusion of a study after patient enrollment has stopped. Study sponsors and researchers may choose to only publish “favorable” results. As a result, positive findings are more likely to be submitted for publication over negative or inconclusive findings (i.e., publication bias). [2]

Information bias (also known as misclassification bias) refers to the non-random errors that occur during the measurement of an exposure or outcome and is one of the most common types of bias seen in Medical Toxicology research. When two study groups are being compared, misclassification can occur in either group even if measured errors are equal between the exposed and unexposed groups or between participants who have or do not have the health outcome of interest. There are several known mechanisms that result in misclassification. The key distinction is whether the errors are non-differential or differential with respect to the comparison groups.

Non-differential misclassification occurs when the error is equal in both study groups. In other words, there is equal misclassification of exposure between exposed and unexposed or equal misclassification of the health outcome between exposed and unexposed subjects. There are several mechanisms by which non-differential misclassification can occur: equally inaccurate memory of exposures, recording and coding errors, using surrogate measures of exposure, using clinically relevant cut-off points for thresholds, and non-specific definitions of exposure and/or outcome.

In Fig. 1, we provide an example of non-differential misclassification by creating a hypothetical example of subjects exposed to benzene, an organic solvent linked to the development of hematologic disorders (eg. leukemia). Subjects who have been exposed to benzene are outlined in black while those who have not been exposed have no outline; subjects who develop a blood disorder (i.e., have the health outcome of interest) are represented by blood cells in black, while those who do not are represented by blood cells in red. Figure 1 depicts non-differential misclassification bias as there are two misclassified cases in each outcome category. In both the “Blood Disorder” and “No Blood Disorder” categories, there are two patients who have been misclassified as “No Benzene Exposure.”

Fig. 1.

Example of non-differential misclassification bias

Differential misclassification occurs when the degree of misclassification is not the same between the comparison groups. In other words, the misclassification of exposure is not equal between subjects who have been exposed and those who have not been exposed or when the misclassification of the health outcome (diseased vs. not diseased) is unequal. Differential misclassification can incorrectly alter test statistics such as the risk ratio, odds ratio, etc. Consequently, the subsequent bias that results may alter the association either towards or away from the null, depending on the magnitude and direction of the misclassification in the study groups (See paper by Flegal for additional detail and examples regarding the impact of misclassification bias on the test statistic) [3]. There are several mechanisms by which differential misclassification can occur including: recall bias, interviewer bias, and surveillance bias. These forms of bias will be discussed in detail in part two of this series. Figure 2 illustrates differential misclassification bias again using benzene as the exposure and blood disorder as the outcome.

Fig. 2.

Example of differential misclassification bias

Conclusion

Classifying which subjects are exposed to a xenobiotic and which subjects are diseased (i.e., poisoned) is a challenge we constantly face in Medical Toxicology. Information (misclassification) bias may be introduced into studies without malintent by the investigators as a result. Correct classification of exposure or disease status are essential since misclassification can profoundly affect or alter the analysis of a study to distort the true nature of any observed (or unobserved) associations. In the realm of evidence-based medicine, preventing and limiting bias are vital to perform quality and meaningful research; identifying information bias (and other forms of bias) is paramount for those who aim to critically appraise literature.

Sources of Funding

None

Declarations

Conflicts of Interest

None