Misclassification bias in chronic disease case ascertainment algorithms: a reclassification approach

Use of administrative health data to identify chronic disease cases can cause misclassification bias. Reclassification-based exit rules may reduce misclassification bias.

Manitoban administrative health data (1995-2022) were used to ascertain multiple sclerosis (MS) and "juvenile diabetes" (JD) prevalence. We constructed multivariable logistic regression model-based algorithms and used a model-predicted probability exit rule to reclassify JD and MS case status annually. Sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV) and reclassification rates were estimated. Linear regression tested for differences in prevalence estimates for the model-based algorithm with an exit rule and an existing Canadian Chronic Disease Surveillance System (CCDSS) algorithm without an exit rule.

The MS cohort included 60 228 individuals (608 cases, 59 620 non-cases) and the JD cohort 44 125 individuals (2506 cases, 41 619 non-cases). Model-based algorithm sensitivity was 0.62 to 0.85 for MS and 0.87 to 0.95 for JD. PPV for MS was 0.21 to 0.60 and for JD was 0.92 to 0.95. Specificity and NPV were consistently high (0.98-1.00). Non-cases were frequently misclassified; reclassification rates for non-cases were higher than for cases for MS (0.22-0.33 vs. 0.14-0.28) and JD (0.18-0.65 vs. 0.13-0.15). The model-based algorithm with an exit rule for MS, but not for JD, had a slower increase in prevalence than the CCDSS algorithm.

Case ascertainment algorithms with an exit rule can address misclassification bias when estimating chronic disease prevalence using administrative health data. Improvements are disease dependent.