Keywords: administrative data, state transition models, prediction, prostate cancer, criterion function
Health administrative data are a rich source of population-based information, useful for building state transition models for medical decision making. These models require estimation of transition probabilities and identification of the times that patients move through health states. Indirect prediction methods are needed to infer this information, as it is rarely available in administrative data. Various algorithms have been validated for identifying patients in a specific disease state but not the dates they entered it. Here, we consider a set of algorithms to identify the dates when prostate cancer patients become metastatic, utilizing dates, in administrative data, of: a) ICD9/10 codes for secondary malignancy, b) palliative radiation therapy, c) chemotherapy, d) ICD9/10 codes for bone disorders or procedures. We evaluate the methods using medical charts of 200 patients containing the true date of evidence of metastasis, linked to administrative data at the Institute of Clinical Evaluative Sciences in Ontario, Canada. For the evaluation, we consider criteria using both misclassification and difference between true and predicted dates for the true positives.