[Record-linkage procedures in epidemiology: an Italian multicentre study]

Area of Science:

• Health Informatics
• Biostatistics
• Public Health Data Management

Background:

• Accurate matching of health care data from disparate sources is essential for reliable epidemiological studies and health service evaluation.
• Existing record linkage (RL) procedures vary in their application across different Italian healthcare settings.
• Electronic health care databases, such as hospital discharge (HD) and population registries, require robust linkage for comprehensive analysis.

Purpose of the Study:

• To compare the performance of various record linkage (RL) procedures, including deterministic and probabilistic methods, for matching data from Italian electronic health care databases.
• To evaluate the accuracy, sensitivity, and specificity of different RL approaches in identifying patient records for acute myocardial infarction (AMI) and diabetes mellitus.
• To assess the impact of different RL procedures on the computation of age and gender standardized annual hospitalization rates.

Main Methods:

• Two distinct Italian health care archives (hospital discharges and population registry) from four administrative areas were utilized.
• Exact deterministic, stepwise deterministic, and a standard probabilistic RL procedures were applied to match records for acute myocardial infarction (AMI) and diabetes mellitus.
• Sensitivity and specificity of each RL procedure were calculated following a manual review process, and hospitalization rates were computed and compared.

Main Results:

• The probabilistic RL procedure identified, on average, over 11% more matched pairs compared to deterministic methods.
• Sensitivity of the probabilistic RL approach was found to be similar to or greater than that of other evaluated procedures.
• Differences in computed hospitalization rates were observed between stepwise deterministic RL and the standard probabilistic RL across different areas, highlighting the impact of linkage methodology.

Conclusions:

• Exact deterministic RL is effective when high-quality data and unique identifiers are consistently available.
• The proposed probabilistic RL procedure performs comparably to semi-deterministic RL, particularly when the latter includes data quality control or manual review.
• Deterministic or semi-deterministic RL methods, without rigorous quality control, may introduce classification errors of unknown magnitude and direction, potentially compromising data integrity.