Graphical Examples of Misclassification Bias
Lead Author(s): Jeff Martin, MD
The effects of reduced sensitivity of the exposure measurement and/or reduced specificity of the exposure measurement on measurement bias can be seen in the following charts.
Graphing Imperfect Sensitvity and Specifcity of Exposure
As show in the below figure from Copeland different scenarios for imperfect sensitivity and imperfect specificity have been worked out.
Explanation of Graph
The graph assumes a case-control study where the true OR is 2.67,
- which is a decent size odds ratio today (Now that the many of the odds ratios of 10, like smoking and lung cancer, have already been found.)
- The prevalence of exposure in the controls is 0.2.
- On the y axis is the observed or apparent odds ratio and the line shows what happens as specificity is varied from 50% to 100% under 3 different scenarios of sensitivity.
Note especially how there are some pretty substantial hits on the apparent odds ratio as you move away from 100% specificity and that this is accentuated, noted by the steeper slopes, as sensitivity falls.
- Note how the slope is steeper in the sensitivity of 50% curve.
Charting Decreasing OR
Again using Copeland we look at the resulting scenarios for odds ratios under 2.0,
- which is often the smallest odds ratio that many of our studies can pick up.
If sensitivity is 90%, then specificity can be no less than about 87% before the OR drops below 2.
If sensitivity is 70%, then specificity can be no lower than about 94%.
If sensitivity is as low as 50%, then specificity can be no lower than about 98%.
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.