Record linkage

Patient data of all HIV‐positive adults aged ≥ 16 years at start of ART attending the Sinikithemba clinic between January 2004 and December 2011 (n = 8,721) were linked with the cancer records of public laboratories in KZN province for the time period January 2004 to December 2010 (32,923 patients and 35,536 cancer cases). Cancers were not systematically recorded in the Sinikithemba clinic database; however, 93 cancer cases were documented in the electronic patients' notes and known prior to record linkage. We linked records probabilistically using the record linkage software G‐Link of Statistics Canada17 based on name and surname, year of birth, gender, date of death and date of cancer diagnoses. The main steps in PRL are selection of linkage variables, data pre‐processing, blocking, matching and post‐processing.7, 13, 18 Details of the record linkage procedure are provided in the Supporting Information. In brief, after pre‐processing the data, we applied blocking criteria to identify potential pairs and to increase the efficiency of the linkage process. In the next step (matching), linkage rules were applied, probability weights estimated and upper and lower thresholds defined. The probability weights and the defined thresholds were used to classify the potential pairs as definite matches, possible matches and impossible matches, which were rejected. In the next step (post‐processing), a thorough clerical review of the possible pairs was conducted and possible pairs were reclassified as definite or rejected pairs. For each definite match, we retrieved data on the cancer diagnosis, date of diagnosis, ICD‐O‐3 topography (site) and morphology (histology and behaviour) code. The additional information from the KZN cancer records was then incorporated into the IeDEA database and all patient identifying information was removed.