Estimating misclassification error in a binary performance indicator: case study of low value care in Australian hospitals

Objective Indicators based on hospital administrative data have potential for misclassification error, especially if they rely on clinical detail that may not be well recorded in the data. We applied an approach using modified logistic regression models to assess the misclassification (false-positive and false-negative) rates of low-value care indicators. Design and setting We applied indicators involving 19 procedures to an extract from the New South Wales Admitted Patient Data Collection (1 January 2012 to 30 June 2015) to label episodes as low value. We fit four models (no misclassification, false-positive only, false-negative only, both false-positive and false-negative) for each indicator to estimate misclassification rates and used the posterior probabilities of the models to assess which model fit best. Results False-positive rates were low for most indicators—if the indicator labels care as low value, the care is most likely truly low value according to the relevant recommendation. False-negative rates were much higher but were poorly estimated (wide credible intervals). For most indicators, the models allowing no misclassification or allowing false-negatives but no false-positives had the highest posterior probability. The overall low-value care rate from the indicators was 12%. After adjusting for the estimated misclassification rates from the highest probability models, this increased to 35%. Conclusion Binary performance indicators have a potential for misclassification error, especially if they depend on clinical information extracted from administrative data. Indicators should be validated by chart review, but this is resource-intensive and costly. The modelling approach presented here can be used as an initial validation step to identify and revise indicators that may have issues before continuing to a full chart review validation.