Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 14-Day Trial for You or Your Team.

Learn More →

Capture‐recapture using multiple data sources: estimating the prevalence of diabetes

Capture‐recapture using multiple data sources: estimating the prevalence of diabetes D iabetes is an increasingly important public health problem worldwide, requiring significant health care resources and has recently been described as a ‘pandemic’. Accurate and timely estimates of diabetes prevalence are important for monitoring the epidemic and informing health care provision and resource decisions. Frequently, such information is not readily available. In New Zealand, where biennial (two‐yearly) national health surveys are undertaken, diabetes prevalence is based on self‐report. Periodic workplace and regional prevalence surveys have used oral glucose tolerance tests (OGTTs) or fasting plasma glucose tests to identify those with diabetes. While diagnostic testing identifies the undiagnosed population, such surveys are generally labour intensive, costly and tend to have a low response rate. An alternative, but under‐utilised, approach is to use administrative diabetes data such as, hospitalisations, mortality, pharmaceutical usage and general practice registries. For any one list, the ascertainment of all diagnosed diabetes cases is usually incomplete; under‐reporting of diabetes on hospitalisation and mortality lists is well recognised, and people with diabetes treated by diet only are not included on pharmaceutical lists and can be missed on primary care diabetes registries. Record linkage of incomplete lists through a common identifier usually under‐estimates prevalence because those not included on any list are missing. However, capture‐recapture methods can be used to estimate the number of cases not on any list, allowing better estimates of prevalent and incident cases. Capture‐recapture methods, originally developed for estimating populations of wild animals which were captured, marked, released and recaptured, have been applied to human populations to estimate the prevalence of different diseases and health behaviours including dementia, alcohol related problems, drug use, salmonellosis and diabetes. Capture‐recapture has been promoted as a cost‐effective means to monitor the diabetes epidemic, but there can be problems of model uncertainty and bias when estimates are derived from a single model. To minimise these problems, we investigated capture‐recapture in conjunction with model averaging using four data sources to estimate the prevalence of diagnosed diabetes in the Otago region, New Zealand. Methods A capture‐recapture technique using four different lists of Otago residents with diabetes and two covariates (age and sex) was used. All diabetes types were included. Ethical approval was obtained from the Lower South Regional Ethics Committee. Research Setting The study area was the Otago District Health Board region of New Zealand, a predominantly rural area on the east coast of the lower South Island. In 2005, the estimated population was 181,660 (estimated by Statistics NZ and sourced from the Otago District Health Board), of whom about two‐thirds lived in Dunedin, the main city. Approximately 125 general practitioners (GPs) delivered primary medical care with practice nurses in group or solo general practices, most of which were affiliated to the region's only Independent Practitioners Association, South Link Health ( http://www.southlink.co.nz ). Most diabetic patients attend their GP for treatment and are referred to a hospital‐based diabetes clinic and the regional retinal screening service as appropriate. A nationwide ‘Free Annual Get Checked’ program for diabetic patients was launched in 2000. Data Sources Lists of diabetic patients resident in Otago with clinical data recorded in 2005 were obtained from four sources: the Otago Diabetes Project, South Link Health, the Dunedin Hospital Diabetes, Podiatry and Diabetes Education Clinics and the Otago Diabetic Eye Monitoring Service. The common data elements were the National Health Index (NHI), sex and date of birth. The NHI, which is intended to be a unique identifier, was used to link the lists. 1. The Otago Diabetes Project (L1) A quality improvement initiative, the Otago Diabetes Project, established a regional diabetes registry in 1997 to monitor and evaluate diabetes care in the region. Details of how general practice diabetes registries and the regional registry were established have been described elsewhere. In brief, identified diabetic patients were invited to participate and sign a consent form prior to enrolment. Consent was subsequently obtained opportunistically when patients attended their GPs or the local retinal screening program. GP records were the primary source of data, although diabetes hospital outpatient or eye department records were checked for missing data. Nine GPs declined to participate in the project. Data were collected annually. The project provided annual GP audit reports, which included lists of patients who had not had recommended tests, for example, retinal examination. These recall lists prompted completion of missing examinations or tests. 2. South Link Health (L2) South Link Health implemented the nationwide ‘Free Annual Get Checked’ program for the Otago region at the end of 2000. The purpose was to monitor and improve care and outcomes for people with diabetes by providing a free annual diabetes consultation with their GP. In Otago, GPs invited identified diabetic patients to attend an initial diabetes review and to give written consent for their data to be used for research and structured feedback reports to practices. At each annual review, required data were recorded on a standard paper form and sent to South Link Health for regional collation. Reimbursement of the clinic visit was linked to provision of annual diabetes review data. 3. Dunedin Hospital Clinics (L3) The Diabetes Clinic, Podiatry Clinic and Diabetes Specialist Nurse Educator Clinic were located at Dunedin Hospital, a publicly funded hospital. Diabetic patients were usually referred to these clinics by their GP, and occasionally by hospital medical or nursing staff or other health professionals. Clinical data and test results pertaining to the consultation were recorded electronically at each visit and collated by Dunedin Hospital Information Systems. 4. Otago Diabetic Eye Monitoring Service (L4) The Otago Diabetic Eye Monitoring Service (ODEMS), located at Dunedin Hospital, was responsible for diabetic retinal photography screening for most diabetic patients in the region. Screening photography was done at Dunedin Hospital or as part of a mobile service provided in the rural areas. Some patients were under the care of an ophthalmologist at the Eye Department because of diabetic retinopathy or another ophthalmic problem. A negligible number of diabetic patients opted to have retinal screening performed by a private ophthalmologist. In New Zealand, retinal screening at least every two years is recommended. Statistical Analysis Loglinear models, within a closed capture‐recapture framework, were used to estimate the number of people with diagnosed diabetes who were not on any of the four lists in 2005. Eighty‐one loglinear models were fitted to the data, using PROC GENMOD in SAS. These models were chosen to allow for dependencies among the lists and for any effects of sex and age on the probability of an individual appearing on a list. We first fitted the model containing sex, age (divided into nine groups: 0–14, 15–24, 25–34, 35–44, 45–54, 55–64, 65–74, 75–84 and 85+ years), the interaction between sex and age, the main effects for each list and all two‐ and three‐way interactions. The four‐way interaction (L1·L2·L3·L4) was excluded because people who do not appear on any list will never be counted. To explore associations between the lists, we fitted models obtained by removing the following terms from the full model: • One of the three‐way interactions (4 models) • A pair of three‐way interactions (6 models) • A pair of three‐way interactions plus all two‐way interactions that could be removed without violating the hierarchy‐principle (6 models) • All three‐way interactions (1 model) • All three‐way interactions plus one of the two‐way interactions (6 models) • All three‐way interactions plus two of the two‐way interactions (15 models) • All three‐way interactions plus three of the two‐way interactions (20 models) • All three‐way interactions plus four of the two‐way interactions (15 models) • All three‐way interactions plus five of the two‐way interactions (6 models) • All three‐way interactions plus all two way interactions (1 model) For each model, a maximised log‐likelihood ( L ) was estimated and used to calculate a modified Akaike's Information Criterion (QAICc) which, in turn, was used for model selection (to determine the best models from the full set of 81) and model averaging. QAICc allows for over‐dispersion and sample size, and is calculated as follows: where ĉ is a variance inflation factor, K is the number of parameters in the model and n is the number of people included in the model. For each model, a QAICc weight, W, was calculated using the formula: where ΔQAICc is the difference in QAICc between a particular model and the model with the lowest QAICc and the sum is over all models. There may not be a single ‘best’ model (assessed through QAICc) so model averaging is used to provide a weighted estimate of period prevalence (and 95% confidence interval) over several candidate models. The models were ordered by their weight and candidate models containing most of the model uncertainty were chosen. For each selected model, the number of males and females with diabetes who were not on any of the four lists was estimated. This was done for each age group. The model‐averaged result was obtained by weighting each estimate by its QAICc and summing over all the models. Models with a very low weight (i.e. below the chosen cutoff) make a negligible contribution to the overall estimate and were excluded. A related calculation is done for the standard error. The ascertainment rate was calculated for each list by dividing the number on the list by the estimated total number of people in Otago with diagnosed diabetes × 100. Diabetes prevalence rates were calculated and age‐standardised to the WHO standard population using the direct method. Results The number of individuals enrolled on one or more of the four lists in 2005 was 5,716. Table 1 shows the number and demographic characteristics of the people on each list. The Otago Diabetes Project and South Link Health had the largest lists. Diabetes Clinic attendees were younger than those on the other lists. A total of 379 (6.6%) people appeared on all four lists and 1,670 (29.2%) people appeared on one list only. 1 Demographic characteristics of the four lists.* L1 N=3,693 L2 N=3,807 L3 N=1,727 L4 N=2,928 % of total observed (5,716) 64.6 66.6 30.2 51.2 Female (%) 48.9 48.4 51.9 48.8 Median age (years)** 65.8 (55.7,74.3) 66.7 (56.8,74.9) 60.7 (46.5,71.8) 63.8 (53.8,72.6) Ethnicity (%) NZ European 91.4 ‐ ‐ ‐ NZ Maori 3.3 ‐ ‐ ‐ Other 3.3 ‐ ‐ ‐ Not recorded 2.0 ‐ ‐ ‐ Diabetes Type (%) Type 1 9.0 7.1 ‐ ‐ Type 2 88.5 92.2 ‐ ‐ Gestational 0.2 0.1 ‐ ‐ Secondary 1.3 0.0 ‐ ‐ Other 1.1 0.6 ‐ ‐ * L1=Otago Diabetes Project; L2=South Link Health; L3=Dunedin Hospital Diabetes Clinics; L4=Otago Diabetic Eye Monitoring Service ** Lower quartile and upper quartile are presented. Table 2 shows the 18 candidate models selected and the estimated diabetes period prevalence for each model. Selection of the candidate models was based on the QAICc weight being at least 0.01, this weight being the probability that a particular model provides the best approximation to ‘truth’. These 18 models had a combined weight of 0.99, thus accounting for most of the model uncertainty. The two best models had a combined QAICc weight of only 0.41. Of the 18 candidate models, only three did not exclude at least L1·L3·L4, L2·L3·L4 or L3·L4 indicating the Dunedin Hospital Clinics (L3) and ODEMS (L4) were independent lists. The estimated number of people with diagnosed diabetes was 6,721 (95% CI: 6,097–7,346). 2 The 18 candidate models with a QAICc weight of at least 0.01 and estimated diabetes prevalence (%) for each model. The models are described in terms of interactions that were excluded. Terms excluded in addition to L1·L2·L3·L4* K** ΔQAICc QAICc Weight Prevalence (95% CI) 1. L1·L3·L4, L2·L3·L4, L3·L4 29 0 0.25 3.65 (3.53, 3.77) 2. L1·L2·L3, L1·L2·L4, L1·L3·L4, L2·L3·L4, L3·L4 27 0.93 0.16 3.64 (3.59, 3.69) 3. L1·L3·L4, L2·L3·L4 30 2.43 0.07 3.62 (3.53, 3.71) 4. L1·L2·L3, L2·L3·L4, L2·L3 29 2.53 0.07 4.05 (3.90, 4.20) 5. L1·L2·L3, L1·L2·L4, L1·L3·L4, L2·L3·L4, L2·L3, L3·L4 26 2.53 0.07 3.72 (3.66, 3.77) 6. L1·L2·L3, L1·L3·L4 30 2.98 0.06 3.94 (3.80, 4.07) 7. L1·L2·L3, L1·L2·L4, L1·L3·L4, L2·L3·L4 28 3.26 0.05 3.63 (3.57, 3.68) 8. L1·L2·L4, L2·L3·L4 30 3.41 0.05 3.50 (3.45, 3.55) 9. L1·L2·L4, L1·L3·L4 30 3.59 0.04 3.53 (3.47, 3.59) 10. L1·L2·L3, L2·L3·L4 30 3.80 0.04 3.90 (3.76, 4.04) 11. L1·L3·L4 31 4.39 0.03 3.73 (3.59, 3.87) 12. L1·L2·L3 31 4.66 0.02 4.11 (3.90, 4.32) 13. L1·L2·L3, L1·L2·L4, L1·L3·L4, L2·L3·L4, L2·L3 27 4.80 0.02 3.70 (3.65, 3.76) 14. L2·L3·L4 31 4.92 0.02 3.66 (3.53, 3.80) 15. L1·L2·L4 31 5.91 0.01 3.51 (3.37, 3.65) 16. L1·L2·L3, L1·L2·L4, L1·L3·L4, L2·L3·L4, L1·L3, L3·L4 26 6.45 0.01 3.54 (2.82, 4.25) 17. Full model 32 6.73 0.01 3.85 (3.61, 4.10) 18. L1·L3, L2·L3, L3·L4, L1·L2·L3, L1·L2·L4, L1·L3·L4, L2·L3·L4 25 7.52 0.01 3.60 (3.57, 3.64) * L1=Otago Diabetes Project; L2=South Link Health; L3=Dunedin Hospital Diabetes Clinics; L4=Otago Diabetic Eye Monitoring Service ** K is the number of parameters The estimated age‐standardised prevalence of diagnosed diabetes was 3.70% (95% CI: 3.36%‐4.04%) for the total population and 4.45% (95% CI: 4.03%‐4.86%) for adults aged 15+ years. Table 3 shows that estimated age‐specific prevalence rates were slightly higher for males than females, except in the youngest age groups. Rates increased with increasing age, peaking in the 65–84 year age group for both sexes. For both men and women, the greatest difference between the observed and estimated counts was in the 55–64 and 75–84 age groups. The ascertainment rates for the four lists (based on the model averaged estimate) were: The Otago Diabetes Project, 54.9%, South Link Health, 56.6%, Dunedin Hospital Clinics, 25.7% and the Otago Diabetic Eye Monitoring Service, 43.6%. 3 Age‐specific and age‐standardised diabetes prevalence by sex. Age groups (years) Observed Number Otago Population Model Averaged Estimated Number Model Averaged Estimated Prevalence (%) (95% CI) Female 0–14 29 15,400 34 0.22 (0.21–0.25) 15–24 59 17,010 69 0.40 (0.40–0.41) 25–34 94 10,400 110 1.05 (1.04–1.07) 35–44 201 12,830 234 1.82 (1.80–1.85) 45–54 355 12,390 413 3.33 (3.29–3.38) 55–64 616 9,880 716 7.24 (7.15–7.35) 65–74 734 6,960 853 12.25 (12.09–12.43) 75–84 597 5,550 694 12.50 (12.33–12.68) 85+ 130 2,290 151 6.61 (6.51–6.73) 15+ 2,786 77,310 3,276 4.24 (3.78–4.70) Total 2,815 92,710 3,310 3.57 (3.18–3.96) Male 0–14 25 16,600 29 0.18 (0.17–0.18) 15–24 69 16,830 80 0.48 (0.47–0.49) 25–34 63 10,070 74 0.73 (0.72–0.75) 35–44 188 11,760 219 1.86 (1.83–1.89) 45–54 422 12,440 491 3.94 (3.89–4.00) 55–64 721 9,740 838 8.60 (8.49–8.72) 65–74 836 6,530 971 14.87 (14.67–15.08) 75–84 519 4,090 603 14.75 (14.55–14.97) 85+ 58 890 68 7.60 (7.46–7.77) 15+ 2,876 72,350 3,382 4.67 (4.18–5.17) Total 2,901 88,950 3,411 3.84 (3.43–4.24) Discussion The estimated number of people with diagnosed diabetes in the Otago region was 6,721, and the estimated period prevalence for adults aged 15 years and older was 4.2% for females and 4.7% for males. Capture‐recapture methods use the degree of overlap between lists to estimate the size of an unknown population, therefore allowing an estimate of the undercount. In our study, 5,716 individuals with diagnosed diabetes appeared on at least one of the four lists, and we estimated a further 1,005 people (15%) with diagnosed diabetes were not on any of the lists we obtained. This proportion is slightly less than that in a similar study in Northern Italy, where the estimated undercount was 20%, suggesting better overall ascertainment of diabetes cases on lists for the Otago region. The method used in our study is an improvement on the standard approach to capture‐recapture, and is clearly better than the record linkage method which simply provides a count of people on the combined list (5,716 in our study). We extended the scope of the standard capture‐recapture analysis by refining the model selection procedures. We used model averaging which allows for the uncertainty of choosing a particular model to generate estimates. As no one or two models appeared appreciably better than the others according to the QAICc weight, our selection of the 18 models was based on the QAICc weight being at least 0.01. The combined weight of the 18 models is 0.99 (and that of the other 63 models was 0.01) suggesting that the 18 selected models are most likely to provide the best approximation of the ‘truth’. Our diabetes prevalence estimates are comparable to the national self‐reported diabetes prevalence rates from the New Zealand 2006/07 Health Survey for adults aged 15 years and older. The national prevalence for females (3.7%) was slightly less, while that for males (4.7%) was the same. Both our actual count and our estimated count were higher than was estimated by the New Zealand Ministry of Health for 2005 for people aged 25 years and older. We counted 5,534 individuals with diagnosed diabetes aged 25 years and older who appeared on at least one list and estimated 6,435 had diagnosed diabetes in the Otago region in 2005 compared with 5,108 cases estimated by the Ministry of Health using a multi‐state life model. These discrepancies highlight the constraints of the model identified by the authors who recognised limitations inherent in the model design and underlying assumptions. One of the strengths of our study is that we used four lists. While the use of two lists is the simplest approach to using capture‐recapture, it is assumed that the two lists are independent. When this assumption of independence is violated, the resulting estimates are an underestimate. Having two independent lists is seldom possible in the health sector because being on one list is often associated with being on another. For example, in our study, the Otago Diabetes Project audit reports listed patients who had not had a retinal screen, triggering a referral to the Otago Diabetic Eye Monitoring Service (ODEMS). When more than two lists are available, this assumption of independence is not required, as loglinear models allow dependencies among the lists to be explicitly modelled. We assumed, in using closed population capture‐recapture methods, that the study population was closed. During our 12‐month study period, the population would have changed through new diagnoses, regional migration and death. As the Otago population is fairly static and the annual number of incident diabetes cases is similar to the annual number of deaths among the diabetes population, it is reasonable that we considered the study population was closed. We also assumed that members of the population appearing on lists were able to be reliably matched using a common identifier. We used the National Health Index (NHI), which is intended to be unique for each New Zealander. This is not always the case as duplicate NHIs for individuals have been identified. However, since 2003 an intensive national programme has worked towards reducing this problem. Because of these dedicated national efforts, the number of individuals with more than one NHI was probably small, and so any overestimate in our results is likely to be negligible. Another assumption in capture‐recapture is that the study population is homogeneous in the sense that all members have the same probability of being included on each list. In our study, heterogeneity was introduced by influences such as of age, sex, ethnicity, diabetes type, diabetes treatments, and disease severity, which can be modelled through the use of loglinear models. In our analysis we were able to model age and sex, but not all four lists included ethnicity, diabetes type, diabetes treatment or glycaemic control data. While it is unclear what effect these additional data would have had on our prevalence estimates, the inclusion of at least ethnicity and diabetes type would have informed estimates of diabetes prevalence by ethnicity and diabetes type, both of which would be useful. In conclusion, we successfully applied capture‐recapture methods taking into account recognised issues of dependence between lists and uncertainty of model selection. We consider this to be an improvement to standard methods, thus providing better prevalence estimates. Our estimate of diabetes prevalence in the Otago region was better than earlier modelling work using a multi‐state life model and similar to results from a national prevalence survey. This method provides a relatively easy way of estimating diabetes prevalence using routinely collected diabetes information, thus providing the opportunity to monitor the diabetes epidemic and inform planning decisions and resource allocation. Acknowledgements We acknowledge and thank the organisations who provided data for this study, in particular Megan Boivin and Patrick Manning (Otago District Health Board), John Gillies, Jim Mann and Joe O'Neill (Otago Diabetes Project Trustees), Susan Dovey and Murray Tilyard (South Link Health). We also acknowledge the help of Andrew Gray (Preventive and Social Medicine, (University of Otago) who assisted with downloading the data. Thanks also to Sheila Williams (Preventive and Social Medicine, University of Otago) and Jim Mann (Edgar National Centre for Diabetes Research) for their input on the manuscript. http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Australian and New Zealand Journal of Public Health Wiley

Capture‐recapture using multiple data sources: estimating the prevalence of diabetes

Loading next page...
 
/lp/wiley/capture-recapture-using-multiple-data-sources-estimating-the-c7WQSDJkVT

References (31)

Publisher
Wiley
Copyright
© 2012 The Authors. ANZJPH © 2012 Public Health Association of Australia
ISSN
1326-0200
eISSN
1753-6405
DOI
10.1111/j.1753-6405.2012.00868.x
pmid
22672027
Publisher site
See Article on Publisher Site

Abstract

D iabetes is an increasingly important public health problem worldwide, requiring significant health care resources and has recently been described as a ‘pandemic’. Accurate and timely estimates of diabetes prevalence are important for monitoring the epidemic and informing health care provision and resource decisions. Frequently, such information is not readily available. In New Zealand, where biennial (two‐yearly) national health surveys are undertaken, diabetes prevalence is based on self‐report. Periodic workplace and regional prevalence surveys have used oral glucose tolerance tests (OGTTs) or fasting plasma glucose tests to identify those with diabetes. While diagnostic testing identifies the undiagnosed population, such surveys are generally labour intensive, costly and tend to have a low response rate. An alternative, but under‐utilised, approach is to use administrative diabetes data such as, hospitalisations, mortality, pharmaceutical usage and general practice registries. For any one list, the ascertainment of all diagnosed diabetes cases is usually incomplete; under‐reporting of diabetes on hospitalisation and mortality lists is well recognised, and people with diabetes treated by diet only are not included on pharmaceutical lists and can be missed on primary care diabetes registries. Record linkage of incomplete lists through a common identifier usually under‐estimates prevalence because those not included on any list are missing. However, capture‐recapture methods can be used to estimate the number of cases not on any list, allowing better estimates of prevalent and incident cases. Capture‐recapture methods, originally developed for estimating populations of wild animals which were captured, marked, released and recaptured, have been applied to human populations to estimate the prevalence of different diseases and health behaviours including dementia, alcohol related problems, drug use, salmonellosis and diabetes. Capture‐recapture has been promoted as a cost‐effective means to monitor the diabetes epidemic, but there can be problems of model uncertainty and bias when estimates are derived from a single model. To minimise these problems, we investigated capture‐recapture in conjunction with model averaging using four data sources to estimate the prevalence of diagnosed diabetes in the Otago region, New Zealand. Methods A capture‐recapture technique using four different lists of Otago residents with diabetes and two covariates (age and sex) was used. All diabetes types were included. Ethical approval was obtained from the Lower South Regional Ethics Committee. Research Setting The study area was the Otago District Health Board region of New Zealand, a predominantly rural area on the east coast of the lower South Island. In 2005, the estimated population was 181,660 (estimated by Statistics NZ and sourced from the Otago District Health Board), of whom about two‐thirds lived in Dunedin, the main city. Approximately 125 general practitioners (GPs) delivered primary medical care with practice nurses in group or solo general practices, most of which were affiliated to the region's only Independent Practitioners Association, South Link Health ( http://www.southlink.co.nz ). Most diabetic patients attend their GP for treatment and are referred to a hospital‐based diabetes clinic and the regional retinal screening service as appropriate. A nationwide ‘Free Annual Get Checked’ program for diabetic patients was launched in 2000. Data Sources Lists of diabetic patients resident in Otago with clinical data recorded in 2005 were obtained from four sources: the Otago Diabetes Project, South Link Health, the Dunedin Hospital Diabetes, Podiatry and Diabetes Education Clinics and the Otago Diabetic Eye Monitoring Service. The common data elements were the National Health Index (NHI), sex and date of birth. The NHI, which is intended to be a unique identifier, was used to link the lists. 1. The Otago Diabetes Project (L1) A quality improvement initiative, the Otago Diabetes Project, established a regional diabetes registry in 1997 to monitor and evaluate diabetes care in the region. Details of how general practice diabetes registries and the regional registry were established have been described elsewhere. In brief, identified diabetic patients were invited to participate and sign a consent form prior to enrolment. Consent was subsequently obtained opportunistically when patients attended their GPs or the local retinal screening program. GP records were the primary source of data, although diabetes hospital outpatient or eye department records were checked for missing data. Nine GPs declined to participate in the project. Data were collected annually. The project provided annual GP audit reports, which included lists of patients who had not had recommended tests, for example, retinal examination. These recall lists prompted completion of missing examinations or tests. 2. South Link Health (L2) South Link Health implemented the nationwide ‘Free Annual Get Checked’ program for the Otago region at the end of 2000. The purpose was to monitor and improve care and outcomes for people with diabetes by providing a free annual diabetes consultation with their GP. In Otago, GPs invited identified diabetic patients to attend an initial diabetes review and to give written consent for their data to be used for research and structured feedback reports to practices. At each annual review, required data were recorded on a standard paper form and sent to South Link Health for regional collation. Reimbursement of the clinic visit was linked to provision of annual diabetes review data. 3. Dunedin Hospital Clinics (L3) The Diabetes Clinic, Podiatry Clinic and Diabetes Specialist Nurse Educator Clinic were located at Dunedin Hospital, a publicly funded hospital. Diabetic patients were usually referred to these clinics by their GP, and occasionally by hospital medical or nursing staff or other health professionals. Clinical data and test results pertaining to the consultation were recorded electronically at each visit and collated by Dunedin Hospital Information Systems. 4. Otago Diabetic Eye Monitoring Service (L4) The Otago Diabetic Eye Monitoring Service (ODEMS), located at Dunedin Hospital, was responsible for diabetic retinal photography screening for most diabetic patients in the region. Screening photography was done at Dunedin Hospital or as part of a mobile service provided in the rural areas. Some patients were under the care of an ophthalmologist at the Eye Department because of diabetic retinopathy or another ophthalmic problem. A negligible number of diabetic patients opted to have retinal screening performed by a private ophthalmologist. In New Zealand, retinal screening at least every two years is recommended. Statistical Analysis Loglinear models, within a closed capture‐recapture framework, were used to estimate the number of people with diagnosed diabetes who were not on any of the four lists in 2005. Eighty‐one loglinear models were fitted to the data, using PROC GENMOD in SAS. These models were chosen to allow for dependencies among the lists and for any effects of sex and age on the probability of an individual appearing on a list. We first fitted the model containing sex, age (divided into nine groups: 0–14, 15–24, 25–34, 35–44, 45–54, 55–64, 65–74, 75–84 and 85+ years), the interaction between sex and age, the main effects for each list and all two‐ and three‐way interactions. The four‐way interaction (L1·L2·L3·L4) was excluded because people who do not appear on any list will never be counted. To explore associations between the lists, we fitted models obtained by removing the following terms from the full model: • One of the three‐way interactions (4 models) • A pair of three‐way interactions (6 models) • A pair of three‐way interactions plus all two‐way interactions that could be removed without violating the hierarchy‐principle (6 models) • All three‐way interactions (1 model) • All three‐way interactions plus one of the two‐way interactions (6 models) • All three‐way interactions plus two of the two‐way interactions (15 models) • All three‐way interactions plus three of the two‐way interactions (20 models) • All three‐way interactions plus four of the two‐way interactions (15 models) • All three‐way interactions plus five of the two‐way interactions (6 models) • All three‐way interactions plus all two way interactions (1 model) For each model, a maximised log‐likelihood ( L ) was estimated and used to calculate a modified Akaike's Information Criterion (QAICc) which, in turn, was used for model selection (to determine the best models from the full set of 81) and model averaging. QAICc allows for over‐dispersion and sample size, and is calculated as follows: where ĉ is a variance inflation factor, K is the number of parameters in the model and n is the number of people included in the model. For each model, a QAICc weight, W, was calculated using the formula: where ΔQAICc is the difference in QAICc between a particular model and the model with the lowest QAICc and the sum is over all models. There may not be a single ‘best’ model (assessed through QAICc) so model averaging is used to provide a weighted estimate of period prevalence (and 95% confidence interval) over several candidate models. The models were ordered by their weight and candidate models containing most of the model uncertainty were chosen. For each selected model, the number of males and females with diabetes who were not on any of the four lists was estimated. This was done for each age group. The model‐averaged result was obtained by weighting each estimate by its QAICc and summing over all the models. Models with a very low weight (i.e. below the chosen cutoff) make a negligible contribution to the overall estimate and were excluded. A related calculation is done for the standard error. The ascertainment rate was calculated for each list by dividing the number on the list by the estimated total number of people in Otago with diagnosed diabetes × 100. Diabetes prevalence rates were calculated and age‐standardised to the WHO standard population using the direct method. Results The number of individuals enrolled on one or more of the four lists in 2005 was 5,716. Table 1 shows the number and demographic characteristics of the people on each list. The Otago Diabetes Project and South Link Health had the largest lists. Diabetes Clinic attendees were younger than those on the other lists. A total of 379 (6.6%) people appeared on all four lists and 1,670 (29.2%) people appeared on one list only. 1 Demographic characteristics of the four lists.* L1 N=3,693 L2 N=3,807 L3 N=1,727 L4 N=2,928 % of total observed (5,716) 64.6 66.6 30.2 51.2 Female (%) 48.9 48.4 51.9 48.8 Median age (years)** 65.8 (55.7,74.3) 66.7 (56.8,74.9) 60.7 (46.5,71.8) 63.8 (53.8,72.6) Ethnicity (%) NZ European 91.4 ‐ ‐ ‐ NZ Maori 3.3 ‐ ‐ ‐ Other 3.3 ‐ ‐ ‐ Not recorded 2.0 ‐ ‐ ‐ Diabetes Type (%) Type 1 9.0 7.1 ‐ ‐ Type 2 88.5 92.2 ‐ ‐ Gestational 0.2 0.1 ‐ ‐ Secondary 1.3 0.0 ‐ ‐ Other 1.1 0.6 ‐ ‐ * L1=Otago Diabetes Project; L2=South Link Health; L3=Dunedin Hospital Diabetes Clinics; L4=Otago Diabetic Eye Monitoring Service ** Lower quartile and upper quartile are presented. Table 2 shows the 18 candidate models selected and the estimated diabetes period prevalence for each model. Selection of the candidate models was based on the QAICc weight being at least 0.01, this weight being the probability that a particular model provides the best approximation to ‘truth’. These 18 models had a combined weight of 0.99, thus accounting for most of the model uncertainty. The two best models had a combined QAICc weight of only 0.41. Of the 18 candidate models, only three did not exclude at least L1·L3·L4, L2·L3·L4 or L3·L4 indicating the Dunedin Hospital Clinics (L3) and ODEMS (L4) were independent lists. The estimated number of people with diagnosed diabetes was 6,721 (95% CI: 6,097–7,346). 2 The 18 candidate models with a QAICc weight of at least 0.01 and estimated diabetes prevalence (%) for each model. The models are described in terms of interactions that were excluded. Terms excluded in addition to L1·L2·L3·L4* K** ΔQAICc QAICc Weight Prevalence (95% CI) 1. L1·L3·L4, L2·L3·L4, L3·L4 29 0 0.25 3.65 (3.53, 3.77) 2. L1·L2·L3, L1·L2·L4, L1·L3·L4, L2·L3·L4, L3·L4 27 0.93 0.16 3.64 (3.59, 3.69) 3. L1·L3·L4, L2·L3·L4 30 2.43 0.07 3.62 (3.53, 3.71) 4. L1·L2·L3, L2·L3·L4, L2·L3 29 2.53 0.07 4.05 (3.90, 4.20) 5. L1·L2·L3, L1·L2·L4, L1·L3·L4, L2·L3·L4, L2·L3, L3·L4 26 2.53 0.07 3.72 (3.66, 3.77) 6. L1·L2·L3, L1·L3·L4 30 2.98 0.06 3.94 (3.80, 4.07) 7. L1·L2·L3, L1·L2·L4, L1·L3·L4, L2·L3·L4 28 3.26 0.05 3.63 (3.57, 3.68) 8. L1·L2·L4, L2·L3·L4 30 3.41 0.05 3.50 (3.45, 3.55) 9. L1·L2·L4, L1·L3·L4 30 3.59 0.04 3.53 (3.47, 3.59) 10. L1·L2·L3, L2·L3·L4 30 3.80 0.04 3.90 (3.76, 4.04) 11. L1·L3·L4 31 4.39 0.03 3.73 (3.59, 3.87) 12. L1·L2·L3 31 4.66 0.02 4.11 (3.90, 4.32) 13. L1·L2·L3, L1·L2·L4, L1·L3·L4, L2·L3·L4, L2·L3 27 4.80 0.02 3.70 (3.65, 3.76) 14. L2·L3·L4 31 4.92 0.02 3.66 (3.53, 3.80) 15. L1·L2·L4 31 5.91 0.01 3.51 (3.37, 3.65) 16. L1·L2·L3, L1·L2·L4, L1·L3·L4, L2·L3·L4, L1·L3, L3·L4 26 6.45 0.01 3.54 (2.82, 4.25) 17. Full model 32 6.73 0.01 3.85 (3.61, 4.10) 18. L1·L3, L2·L3, L3·L4, L1·L2·L3, L1·L2·L4, L1·L3·L4, L2·L3·L4 25 7.52 0.01 3.60 (3.57, 3.64) * L1=Otago Diabetes Project; L2=South Link Health; L3=Dunedin Hospital Diabetes Clinics; L4=Otago Diabetic Eye Monitoring Service ** K is the number of parameters The estimated age‐standardised prevalence of diagnosed diabetes was 3.70% (95% CI: 3.36%‐4.04%) for the total population and 4.45% (95% CI: 4.03%‐4.86%) for adults aged 15+ years. Table 3 shows that estimated age‐specific prevalence rates were slightly higher for males than females, except in the youngest age groups. Rates increased with increasing age, peaking in the 65–84 year age group for both sexes. For both men and women, the greatest difference between the observed and estimated counts was in the 55–64 and 75–84 age groups. The ascertainment rates for the four lists (based on the model averaged estimate) were: The Otago Diabetes Project, 54.9%, South Link Health, 56.6%, Dunedin Hospital Clinics, 25.7% and the Otago Diabetic Eye Monitoring Service, 43.6%. 3 Age‐specific and age‐standardised diabetes prevalence by sex. Age groups (years) Observed Number Otago Population Model Averaged Estimated Number Model Averaged Estimated Prevalence (%) (95% CI) Female 0–14 29 15,400 34 0.22 (0.21–0.25) 15–24 59 17,010 69 0.40 (0.40–0.41) 25–34 94 10,400 110 1.05 (1.04–1.07) 35–44 201 12,830 234 1.82 (1.80–1.85) 45–54 355 12,390 413 3.33 (3.29–3.38) 55–64 616 9,880 716 7.24 (7.15–7.35) 65–74 734 6,960 853 12.25 (12.09–12.43) 75–84 597 5,550 694 12.50 (12.33–12.68) 85+ 130 2,290 151 6.61 (6.51–6.73) 15+ 2,786 77,310 3,276 4.24 (3.78–4.70) Total 2,815 92,710 3,310 3.57 (3.18–3.96) Male 0–14 25 16,600 29 0.18 (0.17–0.18) 15–24 69 16,830 80 0.48 (0.47–0.49) 25–34 63 10,070 74 0.73 (0.72–0.75) 35–44 188 11,760 219 1.86 (1.83–1.89) 45–54 422 12,440 491 3.94 (3.89–4.00) 55–64 721 9,740 838 8.60 (8.49–8.72) 65–74 836 6,530 971 14.87 (14.67–15.08) 75–84 519 4,090 603 14.75 (14.55–14.97) 85+ 58 890 68 7.60 (7.46–7.77) 15+ 2,876 72,350 3,382 4.67 (4.18–5.17) Total 2,901 88,950 3,411 3.84 (3.43–4.24) Discussion The estimated number of people with diagnosed diabetes in the Otago region was 6,721, and the estimated period prevalence for adults aged 15 years and older was 4.2% for females and 4.7% for males. Capture‐recapture methods use the degree of overlap between lists to estimate the size of an unknown population, therefore allowing an estimate of the undercount. In our study, 5,716 individuals with diagnosed diabetes appeared on at least one of the four lists, and we estimated a further 1,005 people (15%) with diagnosed diabetes were not on any of the lists we obtained. This proportion is slightly less than that in a similar study in Northern Italy, where the estimated undercount was 20%, suggesting better overall ascertainment of diabetes cases on lists for the Otago region. The method used in our study is an improvement on the standard approach to capture‐recapture, and is clearly better than the record linkage method which simply provides a count of people on the combined list (5,716 in our study). We extended the scope of the standard capture‐recapture analysis by refining the model selection procedures. We used model averaging which allows for the uncertainty of choosing a particular model to generate estimates. As no one or two models appeared appreciably better than the others according to the QAICc weight, our selection of the 18 models was based on the QAICc weight being at least 0.01. The combined weight of the 18 models is 0.99 (and that of the other 63 models was 0.01) suggesting that the 18 selected models are most likely to provide the best approximation of the ‘truth’. Our diabetes prevalence estimates are comparable to the national self‐reported diabetes prevalence rates from the New Zealand 2006/07 Health Survey for adults aged 15 years and older. The national prevalence for females (3.7%) was slightly less, while that for males (4.7%) was the same. Both our actual count and our estimated count were higher than was estimated by the New Zealand Ministry of Health for 2005 for people aged 25 years and older. We counted 5,534 individuals with diagnosed diabetes aged 25 years and older who appeared on at least one list and estimated 6,435 had diagnosed diabetes in the Otago region in 2005 compared with 5,108 cases estimated by the Ministry of Health using a multi‐state life model. These discrepancies highlight the constraints of the model identified by the authors who recognised limitations inherent in the model design and underlying assumptions. One of the strengths of our study is that we used four lists. While the use of two lists is the simplest approach to using capture‐recapture, it is assumed that the two lists are independent. When this assumption of independence is violated, the resulting estimates are an underestimate. Having two independent lists is seldom possible in the health sector because being on one list is often associated with being on another. For example, in our study, the Otago Diabetes Project audit reports listed patients who had not had a retinal screen, triggering a referral to the Otago Diabetic Eye Monitoring Service (ODEMS). When more than two lists are available, this assumption of independence is not required, as loglinear models allow dependencies among the lists to be explicitly modelled. We assumed, in using closed population capture‐recapture methods, that the study population was closed. During our 12‐month study period, the population would have changed through new diagnoses, regional migration and death. As the Otago population is fairly static and the annual number of incident diabetes cases is similar to the annual number of deaths among the diabetes population, it is reasonable that we considered the study population was closed. We also assumed that members of the population appearing on lists were able to be reliably matched using a common identifier. We used the National Health Index (NHI), which is intended to be unique for each New Zealander. This is not always the case as duplicate NHIs for individuals have been identified. However, since 2003 an intensive national programme has worked towards reducing this problem. Because of these dedicated national efforts, the number of individuals with more than one NHI was probably small, and so any overestimate in our results is likely to be negligible. Another assumption in capture‐recapture is that the study population is homogeneous in the sense that all members have the same probability of being included on each list. In our study, heterogeneity was introduced by influences such as of age, sex, ethnicity, diabetes type, diabetes treatments, and disease severity, which can be modelled through the use of loglinear models. In our analysis we were able to model age and sex, but not all four lists included ethnicity, diabetes type, diabetes treatment or glycaemic control data. While it is unclear what effect these additional data would have had on our prevalence estimates, the inclusion of at least ethnicity and diabetes type would have informed estimates of diabetes prevalence by ethnicity and diabetes type, both of which would be useful. In conclusion, we successfully applied capture‐recapture methods taking into account recognised issues of dependence between lists and uncertainty of model selection. We consider this to be an improvement to standard methods, thus providing better prevalence estimates. Our estimate of diabetes prevalence in the Otago region was better than earlier modelling work using a multi‐state life model and similar to results from a national prevalence survey. This method provides a relatively easy way of estimating diabetes prevalence using routinely collected diabetes information, thus providing the opportunity to monitor the diabetes epidemic and inform planning decisions and resource allocation. Acknowledgements We acknowledge and thank the organisations who provided data for this study, in particular Megan Boivin and Patrick Manning (Otago District Health Board), John Gillies, Jim Mann and Joe O'Neill (Otago Diabetes Project Trustees), Susan Dovey and Murray Tilyard (South Link Health). We also acknowledge the help of Andrew Gray (Preventive and Social Medicine, (University of Otago) who assisted with downloading the data. Thanks also to Sheila Williams (Preventive and Social Medicine, University of Otago) and Jim Mann (Edgar National Centre for Diabetes Research) for their input on the manuscript.

Journal

Australian and New Zealand Journal of Public HealthWiley

Published: Jun 1, 2012

There are no references for this article.