Who Has Diabetes? Best Estimates of Diabetes Prevalence in the Department of Veterans Affairs Based on Computerized Patient Data

This study was conducted using DEpiC databases for estimation of diabetes rates over 3 years, from October 1997 through September 2000 (fiscal years, from October through September 1998, 1999, and 2000). These databases consist of linked individual patient-level data obtained from a number of sources.

National VA patient medical and administrative data were obtained from the Austin Automation Center in Texas. This center routinely obtains selected data abstracts from each VA facility, including information on patient demographics and eligibility for care, medical diagnoses and procedures, outpatient visits, hospital and long-term care stays, and vital status. Quality control is maintained centrally and is facilitated by system-wide standardization of data collection and processing. Medical diagnoses for conditions under treatment are recorded as ICD-9-CM codes (21) and medical procedures are coded as ICD-9-E or CPT4 codes (22). The accuracy and completeness of demographic data were maximized using custom algorithms that made use of data points across multiple years and from other sources. For this analysis, we used VA patient medical data from October 1996 through September 2002.

Many VA patients also obtain some of their care outside of the VA system with the majority being covered by Medicare (23). For this reason we obtained Medicare claims data from the Centers for Medicare and Medicaid Services for all VA patients with any indication of diabetes in VA data, supplemented by any additional VA patients with any codes for diabetes in the claims data alone. A verification process identified dependents and others receiving care and claimed under each beneficiary’s number to assure that we applied claims correctly to the VA patients. We used Medicare files for institutional inpatient and outpatient care (Part A), physician care (Part B), skilled nursing facilities, home health care, hospice, and durable medical equipment. Data elements include vital status, demographics, encounter dates, and diagnostic and procedure codes (ICD-9-CM and CPT4). These files are known to be relatively complete for care covered by Medicare that is provided on a fee-for-service basis (24). For this analysis, we used Medicare data from October 1996 through December 2000.

VA pharmacy data were obtained from two sources. The VA Health Care Analysis Information Group used a software package in place at the majority of VA facilities to routinely collect data on diabetes medications prescribed through VA pharmacies. These data were available for prescriptions dispensed from October 1997 through September 2000, although, in the first year, data were missing for a number of sites representing 5.8% of the patient population. The second source of VA pharmacy data is the Pharmacy Benefits Management Strategic Health Group, which has maintained a national VA prescription database for all VA patients since October 1998. They extract prescription data from every VA facility on a monthly basis, verify its accuracy, and process it centrally into a national database. For this analysis, we used VA prescription data from October 1997 through September 2000, with the first year (FY1998) coming from the Health Care Analysis Information Group, and the last 2 years (FY1999 and FY2000) coming from the Pharmacy Benefits Management Group.

The VA Health Care Analysis Information Group also obtained and processed selected laboratory values, including HbA_1c, from a majority of VA facilities. For each of the 3 years of the study, these data were not available from facilities representing 26.5, 9.7, and 7.1% of the VA patient populations, although we found no apparent bias in rates obtained from other measures for this subset of the population. Given the variability in median values among facilities, we explored standardization of HbA_1c values to the normal range provided by each facility’s laboratory, but—since this resulted in only small differences—we used unadjusted HbA_1c in this study. For this analysis, we used HbA_1c data from October 1997 through September 2000.

Vital status of all VA patients was determined from three sources. The Beneficiary Identification and Record Location files maintained at the Austin Automation Center is a national database of veterans who applied for death benefits. It has been reported that this file is 95% complete for the VA patient population (25,26), and recently the VA has conducted routine cross checks with Social Security Administration files to identify any missing deaths among veterans. Using an algorithm developed and tested for this purpose, we supplemented this database with indicators of mortality from VA inpatient records and Medicare files to improve completeness and accuracy of death records through September 2000.

The 1999 National Health Survey of VA Enrollees collected information directly from a large sample of VA patients, including self-report measures of diabetes and other diseases and use of VA and non-VA services (18). The diabetes question asked “Has a doctor ever told you that you had diabetes or high blood glucose?” For this survey, a stratified random sample of over 1.4 million enrollees (41% of all enrollees as of March 1999) were mailed surveys in July of 1999, and a total of 877,775 returned completed surveys (63.1% response rate). Medical and administrative records for all enrollees in the sampling frame were analyzed and used to derive weights to adjust for response and sampling bias. These data were linked with other DEpiC data for the evaluation.

Fixed-period prevalence of diabetes was estimated for each of the 3 fiscal years. All patients alive at the beginning of the fiscal year and receiving any VA health care during that year were included in the denominator. Nonveteran patients were further excluded since they were mostly relatives and VA employees who received sporadic, emergency care. Different definitions of diabetes using the various data elements were evaluated in those patients in the denominator to estimate 1-year period prevalence.

Definitions of diabetes under consideration included combinations of ICD-9 codes for diabetes (250, 357.2, 362.0, 366.41), dispensing of medications for glycemic control (insulin, sulfonyureas, biguanides, thiazolidinediones, other hypoglycemic medications) or supplies for glucose monitoring, and orders for or high values from HbA_1c testing. Other factors considered in the definition included the use of various data sources for these indicators, the period of time used for surveillance of indicators of diabetes, and the source of the diagnostic codes. We recognized that prescriptions for diabetes medication would be a valid and specific indicator of diabetes, given that these medications are used very rarely for other indications. Other indicators must be used, however, since a large number of patients with diabetes are not treated with medications. Diagnostic codes require that the patient have a medical encounter involving evaluation or treatment of the condition and that it be coded accurately and entered into the information system. For these reasons, we assumed that a long period of surveillance would be needed to identify diabetes codes, beginning with 24 months, and that there may be errors in coding, with variation according to the source of the code. Codes from inpatient and outpatient visits were evaluated separately, and outpatient codes were restricted to those visits involving a face-to-face encounter with a clinician. Thus, we excluded diagnostic codes from laboratory, screening, and telephone encounters in VA outpatient files and in claims from the Medicare physician care files. Diabetes also may be identified through HbA_1c testing, but there are concerns over its validity as a sole indicator because a number of other conditions may raise HbA_1c levels, and such testing is used increasingly for screening purposes.

Various definitions of diabetes were evaluated in a number of ways. In the subset of 737,340 patients who completed a survey and were in our denominator for FY1999 from facilities with laboratory data, we used response to the diabetes question as an indicator of true diabetes status and calculated positive predictive value, specificity, and sensitivity for the various definitions as test criteria. Overall diabetes prevalence based on these definitions were calculated and compared to estimate the incremental effects of various factors. Cumulative prevalence was calculated on a monthly basis, and trends were examined as another means to evaluate effects of varying definitions on the estimates of prevalence. As a type of face validity, the proportions of patients receiving pharmacologic treatment and HbA_1c testing were examined and compared in those classified as having diabetes based on different definitions. Finally, VA patients in FY1998, the first year of the study, who were classified as having diabetes were followed forward in time to evaluate persistence of their assignment based on data from subsequent years, taking into account mortality and attrition from the health care system.

The working definition of diabetes that emerged from this evaluation process was used in estimating 1-year period prevalence of diabetes for each of the three years of the study. Rates are presented overall and in population subsets defined by sex, age, and race. We also present diabetes prevalence based on affirmative response to the diabetes question in the survey, with estimates adjusted for survey weights to account for sampling and response bias (18). All estimates are presented as crude rates and as rates directly standardized to the sex, age, and race distribution of the general U.S. adult population in 2000 (27).

In addition to prevalence, we estimated diabetes incidence over a 2-year period. VA patients in FY1998, the first year of the study, who were classified as not having diabetes were followed forward in time, and those who subsequently met the working definition of diabetes were considered as incident cases. We estimated incidence by both cumulative monthly enumeration of new cases and total number with the definition in subsequent years, taking into account mortality and attrition from the health care system.

There were a total of ∼3 million veteran VA patients in each of the 3 years of the study (Table 1). Over this period, the population increased by 6–8% per year. Overall, most (95.3%) VA patients are men. The age distribution is irregular, reflecting the different theaters of war, with ∼50% of patients aged 65 years or older. The racial and ethnic distribution of VA patients is similar to the overall adult U.S. population, except for underrepresentation of Hispanics, Native Americans, and Asian/Pacific Islanders. After excluding the patients with unknown racial/ethnic status, most VA patients are white (75.9%) or African American (17.9%).

Table 2 presents results from our evaluation of diabetes definitions based on diagnostic codes. In this evaluation, we used codes identified over a 24-month period from both VA and Medicare data with selected exclusions of outpatient codes. Varying combinations of numbers of codes and whether they come from inpatient or outpatient sources were considered and compared with the survey response to the diabetes question. The top part of the table presents the percent of those classified as having diabetes based on the various definitions that affirm diabetes in the survey. If the survey is considered as the standard, these are the predictive value positives for each group. Among patients with no diabetes diagnostic codes, 6.8% indicated diabetes in the survey. The percentage increases substantially with the presence of codes, to 37.4% for a single outpatient code, 50.8% for a single inpatient code, and >70% for all combinations of two or more codes (indicated in bold type). Based on these findings, we considered two or more diagnostic codes to be part of our working definition for diabetes.

The bottom of Table 2 shows a number of parameters for varying definitions based on codes, with our choice of two or more codes in bold. Prevalence varies predictably from a low of 6.9% for any inpatient codes to a high of 22% for any codes. As the number of codes required increases, the prevalence and sensitivity decrease. Specificity varies less across the definitions but is generally higher with more codes required. The other columns in this table show the percent of patients who were on diabetes medications or had HbA_1c testing with various test results over the prior 9-month period. The percentages tend to increase with increasing numbers of codes required.

Our choice of two or more codes represents a reasonable compromise. Including those with a single inpatient code, as has been recommended by Hebert et al. (15) based on analyses of Medicare data, would make little difference in prevalence and sensitivity but would include patients with a positive predictive value of only 50% and a medication prescription rate of only 38% (data not shown). Using our definition, the diabetes prevalence is 18.7%; specificity and sensitivity are 98.3 and 73.1%, respectively; and the medication treatment and HbA_1c testing proportions seem reasonable, at 80.9 and 75.4%, respectively.

Figure 1A and B show results of cumulative prevalence calculations based on diagnostic codes for the 3-year period from October 1996 through September 1999. Figure 1A shows how rates vary over time with variation in the number of codes required. As the number of codes required in the definition increases, the length of surveillance required to obtain nearly stable rates increases. For the definition using two or more codes, 24 months are needed to attain the maximum, stable rate, although >90% of that rate is attained at 12 months. These rates are unadjusted since they accumulate both numerator (diabetes cases) and denominator (VA patients) incrementally among veteran patients without accounting for death or attrition or restricting to patients alive and using services in a given year. Figure 1B shows cumulative prevalence for the same period using two or more codes as the definition but with adjustment incrementally for death and attrition; attrition is defined as no use of services for 12 months following the last visit or stay.

Based on these analyses and on the knowledge that diabetes medication use is a specific indicator of the disease, our working definition for the identification of patients with diabetes is as follows: prescription for diabetes medication in the current year and/or two or more ICD-9-CM codes for diabetes in current and past years (24 months) from inpatient stays and/or outpatient visits on separate days, excluding codes from lab tests and other nonclinician visits.

We applied this definition in estimating diabetes prevalence for the 3 years of the study; the estimates, along with additional estimates of how these rates vary with modifications of the definition, are presented in Table 3. The prevalence estimates for the 3 years are 16.7, 18.6, and 19.6%. Diagnostic codes from outpatient visits are the most important source of indicators of the disease, since missing medication or inpatient data results in decrements in the prevalence estimates of no more than 0.8%. Missing Medicare data would be a more critical element since its inclusion accounts for ∼2% in the estimates or nearly 60,000 cases in each year. Furthermore, given the larger decrement in rates observed for lack of medication data in the absence of Medicare data, Medicare data seem to identify many patients with VA prescriptions but minimal use of VA services. Attempts to improve the accuracy of outpatient diagnostic codes by either requiring 7 days between days with codes or restricting to codes from primary care visits made small differences in the rates, and our decision to exclude codes from laboratory tests and other nonclinician visits also had a very small effect on the rates. Classifying patients with a single inpatient code as diabetic (15) increases prevalence by no more than 0.3%, but considering those with a single outpatient code as diabetic increases rates substantially; if both are included in the definition, rates increase by ∼3%. We have already shown in our evaluation how this does not seem to be justified. Classifying patients with a sole indicator of either multiple HbA_1c tests or high values from these tests adds few additional patients and does not seem justified.

Prevalence of diabetes in the VA patient population is summarized in Table 1. We present both crude and standardized rates, overall and within subsets defined by sex, age, and race. Given that the VA patient population is older and predominantly male, it is not surprising that standardization resulted in lower rates. The crude prevalence estimates for the 3 years are 16.7, 18.6, and 19.6%, indicating rates that are both high and trending upward. Even after standardization, the rates are 9.1, 10.2, and 11.4%, indicating a 25% increase in prevalence over the 3-year period. Diabetes prevalence is higher in men and increases with age up to 65–74 years, with modest declines in rates at older ages. As expected, prevalence is higher in Hispanic and racial minority patients than in whites, with the highest standardized rates in patients who are Native American or Asian/Pacific islanders.

Diabetes prevalence estimates based on survey responses to the diabetes question are presented for comparison purposes. Since the survey was completed in the last quarter of FY1999, it should be compared with FY2000 estimates from the medical and administrative data. Survey estimates of prevalence are similar and somewhat higher, particularly after standardization. Standardization made less of a difference in the survey estimates because of the much lower proportion of unknown race in the survey sample (4.1%) as compared with the total population (19.8%), although unknown race from administrative data was <5% among patients with diabetes.

VA patients from FY 1998, stratified according to diabetes status, were followed forward in time to estimate persistence of diabetes status in medical and administrative data, mortality among diabetic patients, and incidence of new diabetes among those without it at baseline. This is summarized in Table 4 and cumulative incidence is presented in Fig. 1C and D. Mortality among patients with diabetes averaged 5.0% per year over the 2-year period of follow-up as compared with 2.6% among patients without diabetes. A small percent (1.2%) of diabetic patients with no registration of death had no record of health care from VA or Medicare records. These patients may be receiving health care with other sources of coverage, and many of them are likely to be in long-term care, probably with Medicaid coverage. Another 4.7% of diabetic patients had no VA service use in that period but had Medicare claims only. Of the remaining 411,492 patients with diabetes in FY1998 and subsequent use of VA services, most (90.8%) met our definition for diabetes in the subsequent 2-year period, 3.1% had a single diabetes diagnostic code, and 5.9% had no indication of diabetes. Persistence was higher (96.4%) among surviving and VA service-using patients taking diabetes medication.

Diabetes incidence was estimated in the 1,946,413 VA patients who did not meet our definition for diabetes in FY1998 but survived and continued using VA services. Of these patients, 3.9% subsequently met the diabetes definition, providing an estimate of annual incidence of almost 2%. Another 2.7% of these patients subsequently had a single diabetes diagnostic code. Those patients who we considered to be free of diabetes in FY1998 included nearly 4% with a single diabetes diagnostic code in the baseline year. While a large percentage of these patients received the code in error or as part of screening for diabetes, many of them were likely to have been diabetic or at least pre-diabetic. Therefore, it is not surprising that a substantial fraction of these patients met our definition for diabetes in the subsequent 2 years (26.7%) or had a single diabetes diagnostic code in that period (9%). It is evident from the cumulative incidence curves (Fig. 1C) that the number of patients enumerated with a single diabetes diagnostic code increases more steeply than the number with two or more or with use of diabetes medication (Fig. 1D). This suggests a combination of both increasing incidence and increasing screening for diabetes with assignment of initial diagnostic codes for diabetes.

This study has demonstrated the value of using computerized medical and administrative data for disease-rate estimation in a patient population, but also shows the importance of considering a number of potential limitations in using and interpreting these data. Our best definition of diabetes using diagnostic codes and prescription records showed good performance in a number of ways. Validity was evident from the evaluation using survey self-report of diabetes as the standard, in that our diabetes diagnostic code definition was found to be a good predictor (93%) of diabetes report in the survey, a highly specific (98%) indicator of diabetes with few false positives, and reasonably sensitive (73%) for identifying all patients with diabetes. Our best definition also had good face validity in that prevalence rates showed expected variation according to sex, age, and race, and patients identified as having diabetes had near-expected levels of medication treatment (81%), HbA_1c diagnostic testing (75%), and mortality (5% per year). The definition also appeared to have reasonably good reliability in that nearly all (91%) patients classified with diabetes who survived and continued using VA services had their status confirmed in subsequent years.

Our best definition was chosen only after careful consideration of the trade-offs of various modifications. For example, any modifications to our definition in terms of number of codes required or source of codes resulted in less favorable trade-offs in terms of sensitivity, specificity, and resulting prevalence. One variation that we did consider carefully was the inclusion of patients with a single inpatient code since this definition has been used by others based on an analysis of Medicare data (15). It is interesting that, in our evaluation, sensitivity (74%) and specificity (98%) for this definition were identical to those reported from the Medicare analysis, while positive predictive value was higher (93 vs. 84%), reflecting the higher prevalence of diabetes in the VA. Nevertheless, we judged this definition to be less optimal then ours since diabetes was confirmed in only 50% of the added patients and relatively few of them received diabetes treatment or HbA_1c testing. It is possible that the Hebert definition (15) performs better in patients covered by Medicare as opposed to the VA, but there is no evidence for this. While there may be some overcoding of diagnoses in VA records (12), coding of diagnoses in Medicare records may be susceptible to gaming or “diagnosis-related group or DRG creep” (10,28–30), resulting in more error and less reason to accept a single code in the definition.

Another modification that we considered that has been reported by others was the broader inclusion of all patients with any diagnostic code for diabetes (18,31). Clearly many of those patients with a single code in the 24-month period did have diabetes, as suggested by the high proportion (36%) of these patients with such codes and/or diabetes medication prescriptions in subsequent years. Not including them in the definition results in some underestimation of the true prevalence by perhaps 1%. Including them would introduce significant error, however, in falsely classifying patients as diabetic who may be pre-diabetic or not have the disease at all. In our evaluation, specificity only goes down to 95.7% for this definition, but it is likely that a proportion of patients affirming diabetes in the survey did not actually receive a diagnosis but were only screened for the disease or discussed it with their doctor.

A noteworthy finding from our analysis is the importance of including Medicare data in estimating disease rates in the VA. Most VA patients over the age of 65 years and a substantial proportion of those who are younger are eligible for Medicare benefits. It is known from both administrative records and patient reports that the majority of VA patients use non-VA services for part of their health care (23,32). In our analysis, without Medicare data, we would have missed identification of nearly 60,000 cases per year, and prevalence rates would be lower by ∼10%. Without Medicare and VA medication data, the rates would be 15% lower. This finding highlights the importance of considering dual coverage in VA patients and the broader problem of complexity in health care coverage and provision for many in the U.S. in using medical and administrative records for disease estimation.

Other modifications to our definition made less difference. This includes the period of surveillance used to capture indicators of disease, various steps that may be taken in attempting to improve accuracy of diagnostic codes, and the use of laboratory data as a sole indicator of disease. Our analysis informed our decisions in how to best estimate diabetes prevalence in the VA, and the findings of others may vary depending on differences in the systems of care studied and the data used in the estimation process. Our analysis and findings should be instructive, however, to those who plan to estimate disease rates using medical and administrative records.

Our findings confirm the relatively high prevalence of diabetes in the VA patient population. Based on our findings, nearly one in five patients in the VA have diabetes, and nearly one in four patients who are a racial or ethnic minority have the disease. Furthermore, the prevalence appears to have increased substantially from 1998 to 2000.

Our estimate for diabetes prevalence in the VA differs from those of previous reports. Pogach et al. (19) reported that, in 1994, 11.8% of patients at 62 VA medical centers had diabetes based on prescriptions for diabetes medication or glucose monitoring strips. Kazis et al. (18) found that 16.9% of VA enrollees who were in the sampling frame for the 1999 survey had any diagnostic codes for diabetes in VA patient files over the past 2 years. These earlier and lower estimates are not in disagreement with our findings, but were calculated using incomplete data sources (e.g., no Medicare data) and relatively simple, untested methods. In this sense, the methodology developed as part of DEpiC and the prevalence estimates reported here represent a state-of-the-art approach to disease-rate estimation using computerized medical and administrative data.

The prevalence of diabetes among VA patients is substantially higher than that of the general population and most other patient populations. It has been reported that 7.2% of the adult population in the U.S. had diabetes in 2000 (33). This compares with our FY2000 standardized estimate of 11.4% based on medical and administrative data or, perhaps more appropriately, with our standardized estimate from the 1999 survey of 14.2%, since the general population estimate comes from a very similar survey question in the Behavioral Risk Factor Surveillance System. Recently published estimates of diabetes prevalence based on insurance claims or medical data tend to be lower, including estimates of 4.4% in Hawaii (34), 4.5% in Ontario (31), 6.7% in Manitoba (35), and 3.6% in the Northern California Kaiser Permanente Medical Care Program (36). Hebert et al. estimated a prevalence rate of 13.6% among Medicare beneficiaries >66 years of age in 1993 (15), as compared with our standardized estimate of 24.0% in VA patients in 1998.

The higher prevalence of diabetes in the VA may be due in part to more intense screening for the disease by VA clinicians. This may be particularly true for the more recent years following the change in diagnostic criteria for diabetes by the American Diabetes Association (37) and the subsequent increase in use of HbA_1c diagnostic testing. Based on analysis of data from the Third National Health and Nutrition Examination Survey, it has been reported that over a third of diabetes is undiagnosed and that there are a substantial number of additional people with impaired glucose tolerance (fasting plasma levels of between 110 and 126 mg/dl) (38). Intense screening may have resulted in higher proportions of patients with diabetes receiving a diagnosis in the VA, but, given that rates are nearly double those in other populations, it is still likely that the disease is more common in VA patients.

The heavy burden of diabetes among VA patients and the trend of increasing prevalence of this condition present challenges for VA clinicians and health system management in the provision of health care. Continued monitoring of prevalence, incidence, mortality, and measures of disease severity and progression, using the best available methodologies, will continue to be of great value in this process.