Online Program


Estimation of Osteoporosis Prevalence over Time from Two Incomplete Data Sources using Capture-Recapture Techniques
Christopher Bowman, National Research Council of Canada, Winnipeg 
Shamima Huq, Manitoba Centre for Health Policy, University of Manitoba 
Chel Hee Lee, School of Public Health, University of Saskatchewan 
William D. Leslie, Department of Internal Medicine, University of Manitoba 
*Lisa M. Lix, School of Public Health, University of Saskatchewan 
Douglas G. Manuel, Ottawa Health Research Institute, Ottawa 
Hude Quan, Department of Community Health Sciences, University of Calgary 
Liqun Wang, Department of Statistics, University of Manitoba 
Marina Yogendran, Manitoba Centre for Health Policy, University of Manitoba 

Keywords: administrative data, logistic models, completeness, chronic disease

Background: Estimates of chronic disease prevalence are essential for monitoring population health and planning health services. Administrative databases are a common tool for prevalence estimation, but each database will not capture all disease cases and capture rates may vary over time. Capture-recapture (CR) techniques can be used to assess completeness of administrative databases and reduce bias in disease estimates. In addition to independence, two-source CR models assume homogeneity of capture probabilities, which is unlikely to be satisfied. Therefore, covariate-adjusted analyses are required to control for bias in CR estimates.

Purpose: This study uses CR techniques to estimate osteoporosis prevalence from physician and prescription administrative databases, compares results for models with different covariates, and investigates the association of the covariates with capture probabilities over time.

Method: The data sources were physician claims and prescription drug records from Manitoba, Canada. The number of women (50+ years) with osteoporosis was estimated for fiscal years 1999/00 to 2003/04 using diagnosis and drug codes. Multinomial logistic regression analyses were applied to models with: (A) age covariate, and (B) multiple covariates, including age, region of residence, income quintile, comorbidity, continuity of physician care, and presence of an osteoporosis fracture diagnosis. Prevalence estimates and 95% confidence intervals were computed. Odds ratios (ORs) for model covariates were computed for each year.

Results: Uncorrected prevalence estimates ranged from 6.0% (95% CI: 5.8, 6.3) in 1999/00 to 10.3% (95% CI: 10.1, 10.5) in 2003/04. Unadjusted capture rates ranged from 71% to 77%. Model B resulted in improved fit compared to Model A, as judged by the Aikake Information Criterion, and produced prevalence estimates ranging from 8.4% (95% CI: 8.3, 8.8) in 1999/00 to 13.2% (95% CI: 13.1, 13.4) in 2003/04. Covariates associated with the probability of capture (p < .05) were age, region of residence, income quintile, comorbidity, and continuity of physician care. ORs changed substantially over time for age, continuity of care, and region of residence due to an increasing rate of case capture in prescription drug records.

Conclusions: CR techniques can be used to estimate the number of missing osteoporosis cases and identify the variables associated with bias in capture probabilities from administrative database