(140) Application of capture-recapture method to estimate prevalence with registry data

Background: Population registries use compulsory routine data collection that can be used for public health and research purposes. The number of individuals identified with a specific clinical condition is often retrieved from complementary available data sources with incomplete overlap. Consequently, there may be some individuals not captured by any of the data sources (false negatives) and hence the frequency of the condition may be underestimated. This potential problem may affect the measurement of exposures, confounders, and outcomes and can be explored using capture-recapture-derived methods.

Objectives: We aim to apply these methods to estimate the total number of individuals with a specific clinical condition using incomplete overlapping data sources.

Methods: We included all individuals with permanent addresses in the Stockholm region to identify relevant clinical conditions using ICD-10 codes, including hypertension, type-2 diabetes, obesity, anxiety, depression, dementia, heart failure, and ischemic heart disease among others. We used 3 overlapping data sources between 2014 and 2018: 1. Primary outpatient care, 2. Specialized outpatient care, and 3. Hospital inpatient care. All data sources contain pseudo-anonymized individual-level data. The total 2,326,546 inhabitants in the Stockholm region during 2018 was the denominator for prevalence. We used Poisson regression models to estimate the number of false negative individuals. Pairwise interactions were included to model the lack of independence between data sources. Akaike Information Criteria was used to choose which interactions were included. The predicted number of individuals with a particular condition was calculated as the total amount of observed individuals with the condition plus the predicted false-negative individuals. Observed and predicted prevalences for each condition are presented per 100 inhabitants, with the percentage of increase.

Results: Observed hypertensive prevalence was 13.8% vs a predicted prevalence of 16.1%, showing a percentage increase of 17%. The observed and predicted prevalences respectively are presented as follows for the other medical conditions: type-2 diabetes 4.1% vs 4.3% (5%), obesity 3.3% vs 8.8% (167%), anxiety 8.7% vs 14.7% (69%), depression 7.6% vs 13.4% (77%), dementia 0.8% vs 0.8% (7%), heart failure 1.3% vs 1.6% (18%), and ischemic heart disease 2.2% vs 2.7% (23%).

Conclusions: Differences in the observed and predicted prevalences for a particular condition vary. Capture-recapture methods can be useful to estimate the total frequency of a condition when available data sources are overlapping and incomplete while exploring the data quality by the predicted size of the false negative group.