ABSTRACT
BACKGROUND: Linking Brazilian databases demands the development of algorithms and processes to deal with various challenges including the large size of the databases, the low number and poor quality of personal identifiers available to be compared (national security number not mandatory), and some characteristics of Brazilian names that make the linkage process prone to errors. This study aims to describe and evaluate the quality of the processes used to create an individual-linked database for data-intensive research on the impacts on health indicators of the expansion of primary care in Rio de Janeiro City, Brazil. METHODS: We created an individual-level dataset linking social benefits recipients, primary health care, hospital admission and mortality data. The databases were pre-processed, and we adopted a multiple approach strategy combining deterministic and probabilistic record linkage techniques, and an extensive clerical review of the potential matches. Relying on manual review as the gold standard, we estimated the false match (false-positive) proportion of each approach (deterministic, probabilistic, clerical review) and the missed match proportion (false-negative) of the clerical review approach. To assess the sensitivity (recall) to identifying social benefits recipients' deaths, we used their vital status registered on the primary care database as the gold standard. RESULTS: In all linkage processes, the deterministic approach identified most of the matches. However, the proportion of matches identified in each approach varied. The false match proportion was around 1% or less in almost all approaches. The missed match proportion in the clerical review approach of all linkage processes were under 3%. We estimated a recall of 93.6% (95% CI 92.8-94.3) for the linkage between social benefits recipients and mortality data. CONCLUSION: The adoption of a linkage strategy combining pre-processing routines, deterministic, and probabilistic strategies, as well as an extensive clerical review approach minimized linkage errors in the context of suboptimal data quality.