148-29: An Approach to Record-Linkage Using Propensity Score

Traditional uses of propensity score involve bias reduction in matching a treatment with a non-randomized control group. The propensity score controls for patient demographics and other background covariates while producing a single scalar summary variable. Distance or nearest available matching between the case and the control group is then performed based on the single score with comparative ease. The same scoring method but different matching techniques are proposed as an approach to record linkage between two or more mock datasets. Each dataset is comprised of shared and unshared patient records, and the datasets lack a common single patient identifier. Common multiple key identifiers are used as the covariate vector in the SAS/STAT LOGISTIC procedure, which generates the single score representing the propensity of existing in one dataset relative to the other. The propensity score can then be match-merged to link records belonging to the same patient as defined by the independent fields. A simplified illustration is provided, and a discussion of this technique of record linkage is accompanied by considerations in deployment, including weaknesses, strengths and practical implications. INTRODUCTION Logistic regression assigns each record a propensity score, which is an estimate of the probability of an assignment to a particular group given a vector of observed covariates (Rosenbaum and Rubin, 1984). Popular uses of the propensity score include one-to-one and one-to-many matches that are based on distance-metric methodologies, weighting schemes, or matching on ranges. By controlling for demographic and other characteristics of a patient, two comparison groups are produced, as in a case-control match. An approach for linkage of same rather than comparable subjects is proposed for patients that exist across multiple data sources and for whom no single primary-key identifier exists. For each record, a single score is produced by SAS/STAT LOGISTIC procedure and is derived using identifiers such as demographics, gender and temporal and geographical characteristics of the patient. Although exact matching of the score between data sources is required, the covariates can be selectively fuzzed or imputed prior to running the logistic model. The approach taken is empirical and therefore any conclusions drawn from these results may be limited to the examples provided. OBJECTIVES The goal of this paper is to identify the identical patient, rather than the customary similar patient, in two or possibly more datasets using the propensity score generated by the logistic model. If the propensity score is a unique representation based on the covariate vector of each observation, then an individual match-merge on such a score will yield the same patient as defined by the set of covariates. This is in contrast to the nearest available matching proposed by Rubin (1973). THE DATASETS The three mock datasets found in APPENDIX A offer mock Registry, Clinical, and Hospital Discharge data on patients. Individually they serve their own purpose but when combined they offer synergistic value to research. APPENDIX A contains an example of how 3 datasets are matched. THE COVARIATES AND THE OUTCOME The covariates that uniquely predict the outcome in the logistic model consist of STATE, HOSPITAL ID, ADMIT DATE, DISCHARGE DATE, AGE, GENDER and PATIENT FIRST MIDDLE and LAST INITIAL. In this example, the outcome variable named SOURCE is a field coded as 1 for REGISTRY dataset, 2 for the HSPDISCH dataset and 3 for the CLINICAL dataset. All covariate and outcome fields must be converted into their numeric equivalents. PREPARING THE DATASETS To prepare the datasets for logistic regression, the records in each dataset must be appended but the variables must first undergo a standardized naming and format convention. The system ID is applied to each dataset prior to stacking. In addition to the system ID, the covariate and outcome fields ideally are the only fields to be carried forward in the initial stack. After successful linkage, the original source data can be merged back into the matched dataset using the system ID. Separating unnecessary fields from the variables used in the model reduces the impact of record inflation on system resources. Other methods that ultimately reduce the impact of record inflation and of SUGI 29 Posters