Search

Background Epidemiologists increasingly work with linked “big data”. Uncertainty in record linkage may lead to biased inferences but is often overlooked. We evaluate the impact of linkage uncertainty on statistical inference in linked big data. Methods We developed a graphical framework for describing linkage uncertainty when linking multiple representations of the same entity, applied to de-identified data from South Africa’s national laboratory database. Through simulation, we systematically introduced linkage errors and measured their impact on overall accuracy (sensitivity, positive predictive value (PPV)). We evaluate how linkage errors affect bias and variance in point estimates for a hypothetical parameter of interest in clinical epidemiology: 24-month retention in care for HIV patients. We compare the roles of sampling error vs. linkage error as fundamental sources of uncertainty in datasets of varying sizes. Results We simulated a population of 14,393 HIV patients, with a “true” 24-month retention of 38.7%. There were 338,056 true links. Introducing 4,200 false links reduced PPV by 5%. Removing 21,500 existing links decreased sensitivity by 5%. From 10 simulation runs, a 95% sensitivity led, on average, to a 7.4% overestimate in entries to care and a 2.2% (range: 2.1-2.4%) underestimate in 24-month retention. A 95% PPV resulted, on average, in a 7.5% underestimate in entries to care and a 1.8% (range: 1.5-2.0%) overestimate in 24-month retention. Conclusion We observe that in a large sample, linkage uncertainty minimally impacts variance in point estimates but has a potentially large influence on the magnitude and direction, distinguishing it from typical sampling errors.

Introduction Linked administrative data are widely used in epidemiology to capture patient data across multiple databases. Linkage error rates, critical to measure linkage performance, are rarely reported due to difficulty in obtaining representative gold standard. We propose a training and validation approach for linkage procedures that yield unbiased performance estimates even with a non-representative gold standard. Methods We linked patient records from two non-deduplicated databases for HIV monitoring in South Africa, TIER.Net and NHLS laboratory database, using a network-based probabilistic linkage and deduplication approach. National IDs (gold standard) were available for a non-representative minority of records (10%). We calculated sensitivity (Sen, share of true matches identified by the algorithm) and positive predictive value (PPV, share of algorithm-identified matches that were true matches). We adjusted for bias due to informative missingness in National IDs using inverse probability weights to break the link between missingness and match probability. Results 111,755 record pairs were considered. National IDs were not missing completely at random. Match probabilities for National ID record pairs exhibited substantially less uncertainty (mid-range match probabilities), inflating Sen and PPV. Before bias correction, Sen and PPV were estimated at 97.0% and 97.8% respectively. After bias correction for missing National IDs, Sen and PPV were estimated at 95.7% and 96.6%. Failure to address this bias understated the overlinkage rate (100% - PPV) by 35% and the underlinkage rate (100% - Sen) by 30%. Conclusion Failure to adjust for informative missingness in the gold standard may lead to biased validation metrics and over/underconfidence in linked data.

Background HIV cohort data are crucial for understanding the effects of HIV infection and treatment. Linkage between HIV clinical and population records unlocks potential to understand the cause and consequences of social-economic factors on HIV-care seeking behaviors and mortality outcomes. We created HIV data infrastructure by linking HIV treatment records with routinely collected records. Methods We linked HIV cohort data(n=119,434) to Health and Demographic Surveillance(n=256,425), laboratory tests(n=617,314), clinic visits and hospital admissions(n=170,325) from KwaZulu-Natal, an HIV-endemic setting in rural South Africa. We used graph-based probabilistic record linkage algorithm to link the data for HIV patients accessing care at 17 clinics within the Hlabisa sub-district. Africa Health Research Institute (AHRI) has collected health and demographic surveillance data since 2000. AHRI has set up a hospital information system in the district to capture ICD-10 code admissions and discharges. Results From the 5% sample of 462,524 linkage clusters, the computed F-score was 0.88, sensitivity was 88% and positive predictive value was 91%. 38.5% of the patient ID clusters contained records linked across multiple databases. 36% of the HIV patient ID clusters contained records linked with laboratory tests. 19.5% of the HIV patient ID clusters contained data on demographic surveillance. 12.5% of the HIV patient ID clusters contained records linked with clinic visits and hospital admissions. 4% of patient IDs contained records across all four databases. Conclusion We have successfully created a cross-sectorial HIV cohort through probabilistic record linkage methods. The cohort provides a platform to answer policy-relevant questions.