A trainable model to assess the accuracy of probabilistic record linkage

Abstract

Record linkage (RL) is the process of identifying and linking data that relates to the same physical entity across multiple heterogeneous data sources. Deterministic linkage methods rely on the presence of common uniquely identifying attributes across all sources while probabilistic approaches use non-unique attributes and calculates similarity indexes for pair wise comparisons. A key component of record linkage is accuracy assessment — the process of manually verifying and validating matched pairs to further refine linkage parameters and increase its overall effectiveness. This process however is time-consuming and impractical when applied to large administrative data sources where millions of records must be linked. Additionally, it is potentially biased as the gold standard used is often the reviewer’s intuition. In this paper, we present an approach for assessing and refining the accuracy of probabilistic linkage based on different supervised machine learning methods (decision trees, naïve Bayes, logistic regression, random forest, linear support vector machines and gradient boosted trees). We used data sets extracted from huge Brazilian socioeconomic and public health care data sources. These models were evaluated using receiver operating characteristic plots, sensitivity, specificity and positive predictive values collected from a 10-fold cross-validation method. Results show that logistic regression outperforms other classifiers and enables the creation of a generalized, very accurate model to validate linkage results.