To the Editor.
—Dr Ellerbeck and colleagues1 support the accuracy of their abstraction of clinical indicators from medical records using κ statistics based on duplicate reviews of 912 records. However, the κ statistic is inappropriate and uninformative in this application, since κ measures agreement, corrected for chance, between reviewers when the true finding is not known,2 whereas in chart abstraction the presence or absence of an indicator can be ascertained with certainty.More appropriate measures of accuracy are sensitivity and specificity.3 In this setting, sensitivity is the proportion of cases correctly identified when the indicator is present in the medical record, while specificity is the proportion of cases correctly identified as having no indicator present.Using κ can mislead. For example, suppose the prevalence of the indicator is 10% in a sample of 1000 records, the sensitivity of the record abstraction is 75% (75 correct of 100
Localio AR, Landis JR. Quality of Chart Review for Quality of Care. JAMA. 1995;274(20):1585-1586. doi:10.1001/jama.1995.03530200021021