Example of Misclassification Bias
Outcome Variable - Cohort Study
Lead Author(s): Jeff Martin, MD
Misclassification bias of the outcome variable does not reduce the measure of association.
Diagram
As you can see in the diagram below the risk ratio is 2.0 in the true scenario.
If specificity of the outcome measurement is 100%
- but there is only 70% sensitivity of the classification of the outcome,
- the risk ratio is unaltered at 2.0.
This is because all that you have done is to decrease both this cell and this cell by the same percentage.
- Therefore, the ratio between exposed and unexposed will not be affected.
This little trick is worth knowing about when you are using cutoffs for continuous variables using ROC curves - it is a recommendation to choose cutoffs which provide very high specificity.
Specificity of Outcome 100%
As noted in this figure from Copeland (1977) when specificity is 100%,
- you can actually get an unbiased risk ratio regardless of the sensitivity of the outcome measurement.
- Here the true risk ratio is 2.0 and when specificity is 100% you can get a risk ratio of 2.0 regardless of the sensitivity.
ROC Curve with Specificity of Outcome 100%
As shown in the figure below 100% specificity in the outcome measurement preserves unbiased risk ratios even in the face of less than perfect sensitivity.
- This is worth knowing when you are considering which outcomes to use or where to make cutoffs for certain diagnostic tests that are measured in their most raw form with a continuous variable.
Choosing the Cutoff
Choosing the most specific cutoff or the cutoff associated with 100% specificity will lead to least biased ratio measures of effect.
- In the ROC curve you can see that you have many choices in terms of where you can make your cutoff for positivity to remind.
- This is important, for example, when you have a diagnostic test for outcome, say an antibody test for an infectious disease, you have many choices in terms of where you can make your cutoff for positivity.
References
Copeland, K. T., Checkoway, H., McMichael, A. J., & Holbrook, R. H. (1977). Bias due to misclassification in the estimation of relative risk. Am J Epidemiol, 105 (5), 488-495.