Validity of Data in Thailand Cancer-Based Registry with the Application of Probabilistic Record Linkage

Abstract

Objectives: To develop and validate probabilistic data linkage models linking data in cancer registry and data in one tertiary care hospital, and to validate data in cancer registry by using such method. Method: The study linked the data from two sources by using deterministic record linkage with 1-1 matching with newly encoded identification numbers and 17 probabilistic record linkage based models. Each model employed either identifiers or non-identifiers including encoded identification numbers, encoded hospital number, first name, last name, contact address in house registration, date of birth, date of death, zip code of residence in house registration issued by Department of Provincial Administration, sex, and ICD-10. Each model employed 3 to 8 variables, and the optimal cutoff points were determined. Result: There were 7,243 matched patients in both databases accounting for 89.72% of 8,073 patients in cancer registry and 36.72% of 19,725 patients in two databases. All models of probabilistic record linkage had positive predictive values between 99.83 to 99.94%, negative predictive values between 97.04 to 99.98%, sensitivity between 94.74-99.97%, and specificity between 99.90-99.97%. Conclusion:  Linking data on cancer patients between two databases using probabilistic record linkage led to a higher number of matched patients than using deterministic record linkage. Therefore, it could be applied in any work or research where data anonymity and confidentiality are important. However, the validity of models using probabilistic record linkage largely depends upon selected variables in the model. Therefore, researchers should carefully select the variables used in the prediction equation.