Article Text


043 Adjusting misclassification of outcome in case-control studies
  1. R Gilbert1,
  2. R M Martin1,
  3. J A Lane1,
  4. D E Neal2,
  5. F Hamdy3,
  6. J Donovan1,
  7. C Metcalfe1
  1. 1Department of Social Medicine, University of Bristol, Bristol, UK
  2. 2Department of Oncology, University of Cambridge, Cambridge, UK
  3. 3Nuffield Department of Surgery, University of Oxford, Oxford, UK


Objective Misclassification of outcome may cause biased estimation for associations of potential risk factors with important diseases. For example, case-control studies of localised prostate cancer frequently measure blood levels of circulating prostate specific antigen (PSA) in healthy men and biopsy those with an elevated level. Inevitably, some men with prostate cancer will be misclassified as controls, either because they do not have an elevated PSA level or because cancer was not detected at biopsy. This misclassification may be differential if the risk factor itself influences PSA level.

Design We reviewed the literature for methods that correct for non-differential and differential misclassification of outcome in case-control studies. We apply these methods to estimating the association between two established risk factors and prostate cancer: family history and diabetes. We use published data on prostate cancer risk in men with low PSA levels to inform our estimates of the amount of misclassification in our data.

Results Potential approaches range from simple sensitivity analyses to probabilistic sensitivity modelling and Bayesian models, incorporating estimates of sensitivity and specificity. Simple sensitivity analyses recalculate cell frequencies accordingly to produce corrected odds ratios (OR). One accurate estimate of sensitivity and specificity can be used to produce a “corrected” effect-estimate, or a range of values can be used as a sensitivity analysis to assess the direction and magnitude of potential bias. Probabilistic and Bayesian methods incorporate uncertainty in the estimates of sensitivity and specificity as probability distributions so producing a frequency distribution of corrected results from which a median corrected estimate can be presented. Using varying estimates of sensitivity (fixing specificity 100%, feasible in the current example), the direction of the association between family history and prostate cancer (assuming non-differential misclassification) did not change, although the OR increased. The magnitude and direction of the association between diabetes and prostate cancer (assuming differential misclassification) becomes increasingly inverse when the sensitivity in the exposed group is greater than in the non-exposed group. When the sensitivity in the exposed group is smaller than in the non-exposed group, the OR increased, potentially even changing the direction of the association.

Conclusion Correction for misclassification of disease allows presentation of results that incorporate estimates of systematic error due to misclassification bias. That such misclassification may change both the magnitude and direction of an association is demonstrated in real data. Careful consideration is required as to the objective of applying these methods.

Statistics from

Request permissions

If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.