Probabilistic master lists: integration of patient records from different databases when unique patient identifier is missing

Abstract

We show how Bayesian probability models can be used to integrate two databases, one of which does not have a key for uniquely identifying clients (e.g., social security number or medical record number). The analyst selects a set of imperfect identifiers (last visit diagnosis, first name, etc.). The algorithm assesses the likelihood ratio associated with the identifier from the database of known cases. It estimates the probability that two records belong to the same client from the likelihood ratios. As it proceeds in examining various identifiers, it accounts for inter-dependencies among them by allowing overlapping and redundant identifiers to be used. We test that the procedure is effective by examining data from the Medical Expenditure Panel Survey (MEPS) Population Characteristics data set, a publicly available data set. We randomly selected 1,000 cases for training data set--these constituted the known cases. The algorithm was used to identify if 100 cases not in the training data set would be misclassified in terms of being a case in the training set or a new case. With 12 fields as identifiers, all 100 cases were correctly classified as new cases. We also selected 100 known cases from the training set and asked the algorithm to classify these cases. Again, all 100 cases were correctly classified. Less accurate results were obtained when the training data set was too small (e.g., less than 100 records) or the number of fields used as identifiers was too small (e.g., less than seven fields). In a test of performance of the algorithm, when the ratio of testing to training data set exceeds 4 to 1, the accuracy of the algorithm exceeded 90% of cases. As the ratio increases, the accuracy of algorithm improves further. These data suggest the accuracy of our automated and mathematical procedure to merge data from two different data sets without the presence of a unique identifier. The algorithm uses imperfect and overlapping clues to re-identify cases from information not typically considered to be a patient identifier.