Sloan FA, Brown DS, Carlisle ES, Ostermann J, Lee PP. Estimates of Incidence Rates With Longitudinal Claims Data. Arch Ophthalmol. 2003;121(10):1462-1468. doi:10.1001/archopht.121.10.1462
PAUL P.LEEMDFrom the Department of Economics (Dr Sloan, Mr Brown, and Ms Carlisle); Center for Health Policy, Law, and Management (Drs Sloan and Ostermann); and Department of Ophthalmology, School of Medicine (Dr Lee), Duke University, Durham, NC. The authors have no relevant financial interest in this article.
Copyright 2003 American Medical Association. All Rights Reserved. Applicable FARS/DFARS Restrictions Apply to Government Use.2003
To estimate incidence rates of the 3 major chronic eye diseases—diabetic retinopathy (DR), glaucoma, and age-related macular degeneration (ARMD)—by using longitudinal claims data from Medicare.
Longitudinal cases were ascertained by using a national probability sample of Medicare beneficiaries aged 65 years and older in 1991 who initially had none of the eye diseases documented. After adjusting for death and enrollment in a health maintenance organization, claims filed by optometrists or ophthalmologists with an International Classification of Diseases, Ninth Revision, Clinical Modification code for all forms of DR, glaucoma, and ARMD were used to indicate diagnosis.
Annual incidence rates for the 3 conditions after the first year of observation ranged from 14.3% to 17.7% (higher earlier) across an 8-year longitudinal follow-up. Incidence rates among those with diabetes mellitus for any form of DR varied between 3.8% and 6.5%, while those for glaucoma varied between 4.6% and 7.8% and those for ARMD varied between 7.5% and 9.3%.
Longitudinal claims data after the first year provide relatively stable estimates of incidence rates on an annual basis. These estimates are comparable with those of the few population-based studies available.
THE LITERATURE contains only limited information on the incidence of the major eye diseases, or even chronic diseases more generally. Longitudinal surveys exist only for specific geographically circumscribed populations (eg, Beaver Dam, 1 Blue Mountains, 2 Barbados, 3,4 Wisconsin5), are costly and effort intensive, and have limited external validity. They require recruitment, efforts to promote retention of the subjects, and physical examinations—all reasons for limiting their scope. High costs of data acquisition limit sample size and diminish ability to obtain adequate numbers for reliable estimates of annual incidence rates, which is an especially serious consequence for certain groups of interest, such as the oldest and racial minorities.
In this study, we used longitudinal Medicare claims data from 1991 through 1999 to estimate annual incidence rates for the 3 major eye diseases: diabetic retinopathy (DR), glaucoma, and age-related macular degeneration (ARMD). Because the data were from the same cohort of Medicare beneficiaries, we were able to examine the stability of incidence estimates. Further, we were able to adjust rates to compare them with currently available population-based incidence rates to determine the relative rates obtained with different methods.
The study population consisted of the same nationally representative sample of 21 644 Medicare beneficiaries that was used to estimate the prevalence of DR, glaucoma, and ARMD in a cohort of Medicare beneficiaries between 1991 and 1999, as in our previous article.6 The methods described in this section specifically pertain to our analysis of incidence rates.
We specified disease onset by using the earliest diagnosis code found in the claims. Once an individual was identified as having a particular disease, he or she remained in a disease cohort for the remainder of the analysis but was excluded from subsequent analysis of incidence rates. For a case to be identified as incident, the beneficiary had to have undergone an eye examination with an ophthalmologist or optometrist during the year or undergone a procedure with an ophthalmological procedure code. For each year, we began with the number of persons who did not have a disease diagnosed previously; determined whether the person had undergone an eye examination during the year; and, conditional on the person having undergone an eye examination, determined whether the disease was diagnosed. We allowed for disease progression into subgroups. We also calculated transitions to a more severe condition within each major eye disease category. For example, once a person with diabetes with background DR (BDR) had a claim with proliferative DR (PDR), we moved the individual into the subgroup of those with diabetes with PDR. Additionally, we allowed for coexistence among the 3 diseases. Annual disease-specific incidence rates were calculated by using a denominator of the number of persons alive and not in a health maintenance organization (HMO) for 6 or more months during the year who underwent eye examination.
Because of concerns that incidence rates based only on the number of persons with examinations may be biased, we also conducted a sensitivity analysis. We assumed a rate of eye disease in the unexamined population that was 50% of the rate of those who underwent eye examination in the same year. Alternatively, we assumed a 0% incidence rate in the unexamined population, yielding the lowest possible bound on total incidence. We thus calculated estimated lower bounds to incidence that were weighted averages of the 2 populations (unexamined and examined).
We identified eye examinations by using Current Procedural Terminology, Fourth Edition and International Classification of Diseases, Ninth Revision, Clinical Modification procedure codes. For individuals without any eye disease, we used the codes shown in panel A in Table 1. Any instance of 1 or more codes on a claim was counted as an eye examination. For the subsample with 1 or more of the 3 eye diseases, we used additional disease-specific codes to identify other forms of contact with an eye-care provider during the analysis interval (Table 1, panel B).
In 1991, 31.2% of beneficiaries without diagnoses underwent eye examination(Table 2). Of these persons, 29.6% had at least 1 disease diagnosed: 1.8% had DR, 17.1% had glaucoma, and 12.7% had ARMD. In 1992, an almost identical percentage (32.9%) of persons without diagnoses underwent examination, with an overall yield of newly diagnosed cases of 17.7%: 1.3% had DR, 7.9% had glaucoma, and 9.4% had ARMD. By 1999, the number of beneficiaries with undiagnosed disease who underwent eye examination had increased to 38.9%, and the number of persons alive, not in HMOs, and without diagnosed disease decreased considerably, to less than one third of the 1991 number. During these years, annual incidence rates did not differ dramatically within disease category after the initial 2 years.
Incidence of DR is conditional on diabetes having been diagnosed. Thus, despite attrition due to mortality and diagnosis of DR, the number of patients with diabetes and no previous DR diagnosis increased until 1993, as a result of more cases of diabetes but not of DR being diagnosed, and declined monotonically thereafter (Table 3). This change did not occur in Table 2 because all Medicare beneficiaries in the sample were included in the denominator. The percentage of those with diabetes with undiagnosed DR who underwent eye examination in each year increased across time. By 1999, the rate was 54.5%. Most patients had DR diagnosed at an early stage (BDR). Diagnosing PDR as the first DR diagnosis was relatively rare and occurred even less frequently across time. The annual incidence rate of BDR declined as eye examination rates increased.
At baseline, 20 019 individuals in our sample did not have glaucoma diagnosed (Table 4). By 1999, because of deaths, movement into an HMO, and diagnosis of glaucoma, only 7988 did not have glaucoma diagnosed. The percentage of persons with no diagnosed glaucoma who underwent eye examination increased appreciably across time, from 31.4% in 1991 to 44.2% in 1999. However, incidence rates decreased across time. After the first 2 years, most of the new diagnoses were for suspected glaucoma, followed by primary open-angle glaucoma (POAG). Narrow-angle glaucoma(NAG) was diagnosed about one fourth as frequently as was POAG.
In 1991, 20 215 beneficiaries did not have ARMD diagnosed (Table 5). As with glaucoma, the number of persons without a diagnosis remaining in the sample declined markedly during the observation period. The percentage of persons without ARMD diagnosis who underwent eye examination increased across this period. In 1992, 3.0% of persons who underwent eye examination had a diagnosis of dry ARMD; incidence rates for dry ARMD reached a peak in 1998 at 4.6%. A first-time diagnosis of wet ARMD was relatively rare but did occur in approximately 1.0% of previously undiagnosed cases.
The severity of eye disease progressed for many individuals in our sample(Table 6). For DR, between 1992 and 1999, rates of first diagnosis of PDR varied from 1.4% to 4.0%. Of the cohort of 118 persons with diabetes with unspecified DR or BDR who were examined in 1992, 4 (3.4%) had a diagnosis of PDR in 1992. Of these 118 persons, 16(13.6%) eventually had a diagnosis by 1999. For glaucoma, transitions to POAG or NAG from suspected glaucoma were appreciably higher than for transitions from less severe forms of DR to PDR. For ARMD, only about 2% of persons with dry or unspecified ARMD progressed to wet ARMD in a given year. However, across time, rates of progression to wet ARMD in this aging population were substantial. For those with a dry or unspecified ARMD diagnosis in 1992, 2.1% of persons who underwent eye examination had a diagnosis of wet ARMD in that year. However, by 1999, 12.7% of persons with other forms of ARMD had progressed to wet ARMD.
We assessed our findings' sensitivity to the definition of the denominator. Assuming that the unexamined population had an incidence rate half that of the examined population, our mean estimated lower bounds were 3.7% vs 5.1% for DR, 4.2% vs 6.2% for glaucoma, and 5.9% vs 8.3% for ARMD (Table 7). Alternatively, substituting an assumed incidence rate of 0% for the unexamined sample, the mean absolute estimates of lower bounds were 2.4% for DR, 2.3% for glaucoma, and 3.5% for ARMD.
The onset of major eye diseases in an older population is a common occurrence, as is progression of these diseases. In this analysis of Medicare data, we found that the annual incidence rates among Medicare beneficiaries ranged from 3.8% to 6.5% for DR for persons aged 65 years and older who had a diagnosis of diabetes mellitus, 4.6% to 7.8% for glaucoma, and 7.5% to 9.3% for ARMD. These estimates are left-censored in that some individuals had developed at least 1 of the 3 eye diseases prior to the beginning of the study. All rates for 1991, and to some extent the next year, were artificially high because these estimates for the initial year reflected prevalence and incidence. For this reason, rates for 1993 through 1999, when the rates more likely reflect incidence alone, are more valid.
Only a minority of Medicare beneficiaries underwent eye examination in a given year, although the percentage of persons without a previous diagnosis who underwent examination increased across time. On average, 53% of persons with a diagnosis of diabetes did not undergo eye examination in any given year after diagnosis, which is far short of the American Academy of Ophthalmology and American Diabetes Association recommendations of eye examinations at diagnosis with annual follow-ups.7- 10 As a greater percentage of persons with diabetes without a DR diagnosis underwent eye examinations each year, the incidence rate of DR decreased. One interpretation is that undergoing eye examination is indicative of better glucose control. However, in contrast to DR incidence rates, conditional on undergoing examination, incidence rates for glaucoma and ARMD in this cohort of beneficiaries were high, measured both in terms of first diagnosis of a particular eye disease and progression (see the following comparison with other studies). Of those individuals still alive in 1999, only 15% had not undergone eye examination during the study.
Our national estimates of incidence rates were based on physician diagnoses as reported on Medicare claims rather than on the explicit diagnostic criteria used in the population-based studies (Table 8). To compare our estimates with those of population-based studies, we normalized the other studies' estimates to our population's age distribution, then aggregated our incidence rates to match the cumulative rates in other studies. To determine incidence rates across multiple years, all possible periods were calculated to provide a range of estimated incidence rates (eg, to compare with 4-year incidence estimates, we aggregated years 1992 through 1995, 1993 through 1996, and so on). For 2 studies (Denmark11,12 and San Luis Valley13) in which age ranges included young people, we were unable to adjust for age owing to lack of information about the study population's age distribution.
The incidence rates for DR in our study were generally in the same range as those of previous studies, although the rates in the other studies varied markedly around the mean difference of −0.4% (Table 8, panel A). The age-adjusted aggregate 4-year incidence rate from the Wisconsin Epidemiological Study on Diabetic Retinopathy5 was 37%, and for the San Luis Valley study, 13 the 4-year unadjusted rate was 22.5%. In contrast, our 4-year aggregate incidence rates averaged 19.1%, with a range of 23.0% (1992-1995) to 16.0% (1996-1999). Our estimates were high compared with those of the Framingham study14—roughly 5% in the Framingham study, as compared with our mean of 23.7% and range of 27.4% (earliest) to 20.2% (latest) for 5-year incidence rates. Our estimates were similar to those of the study conducted in Denmark, 11,12 with 4.3% per year in Denmark, not adjusted for age, as compared with a mean of 5.4% and range of 6.5% (earliest) to 4.3% (latest).
One reason for lower estimates of DR incidence rates is that providers sometimes miss PDR and BDR, depending on the diagnostic method used.15 Also, if a comparison study had a higher proportion of participants with insulin-independent diabetes than there was in our study, we would expect to see a higher incidence rate for complications such as DR. Our estimates were much higher than the Framingham estimates, but the authors of that study acknowledged that its prevalence rates were much lower than previously published estimates.14
Available comparison estimates for glaucoma were limited to POAG. Our incidence rates of POAG were slightly higher than those in other studies (Table 8, panel B). The 4-year incidence rate determined from the Barbados Eye Study, 3,4 with a black population, normalized to our population was approximately 4.5%. The mean of our 4-year rates was 8.2%, with a range of 10.3% (earliest) to 6.0%(latest). Our mean 5-year incidence rate was 9.8%, with a range of 11.7% (earliest) to 8.5% (latest); in contrast, the age-normalized Melbourne 5-year incidence rate16 was about 8%, and the normalized Framingham14 5-year incidence rate was about 1%.
The comparison estimates for this disease were for highly specific definitions of POAG, which partially explains the gap. Clinicians in the community may overdiagnose glaucoma; one study indicated that about 50% of patients receiving glaucoma medications on a long-term basis had only elevated pressures documented, without documentation of either optic nerve or visual field changes or damage.17 We also computed incidence rates restricting the definition of POAG (not shown), which lowered our estimates by 15%. Another potential reason for the difference is that our cohort was much older than that in the other studies. Those studies showed a dramatic increase as the population aged, which would be consistent with higher estimates for the older persons in our study. In the comparison studies, individuals in their late 70s and 80s were at the high end of the age distribution, in samples already limited in size, whereas in our sample, the median age was 74 years in 1991 and 81 years in 1999.
Our estimates of incidence rates for ARMD were high, as compared with those in other studies (Table 8, panel C). The age-adjusted 2-year incidence rate for the Rotterdam study18 was less than 1%; our mean incidence rate was 16.0%, with a range of 17.0% (earliest) to 15.1% (latest). Our 5-year incidence rate was 33.0%, with a range of 34.7% (earliest) to 32.3% (latest). The Blue Mountains, 2 Framingham, 14 and Beaver Dam1 studies had incidence rates lower than 23%. Variation in diagnostic criteria accounts for some of this difference, because clinicians might diagnose even a few drusen as ARMD. In contrast, population-based studies, on which the comparison estimates are based, tend to have strict criteria for ARMD diagnosis that often require more than the presence of drusen. For example, the Framingham study14 required that senile macular degeneration entirely account for loss in visual acuity to 20/30 or worse for the case to be counted as incident. Although guidelines for clinicians exist, individual professional judgment varies, and criteria for diagnosis may be less stringent on the whole. Changes in Medicare financial incentives and resulting changes in coding are not likely to influence our incidence rates on average, because such changes would have had to occur throughout this period and influence our incidence estimates each year. Another consideration is that our sample included individuals who were institutionalized; those in nursing homes may be expected to have higher rates of eye disease than do the elderly living in the community. Finally, as with the glaucoma studies, our population was older than those of the previous ARMD studies, and ARMD tends to increase dramatically with age.
Our incidence rates could be overstated for 2 reasons. First, as mentioned earlier, in the first year and perhaps the next year, we captured prevalent and incident cases. Second, selection could lead to overstatement. We excluded persons in HMOs for more than 6 months during a year, but they reentered our analysis when they switched back to fee-for-service programs. If patients in HMOs are healthier, we might have overstated incidence rates for Medicare beneficiaries overall. The literature offers mixed opinions on whether patients in HMOs are indeed healthier; the prevalent findings are that enrollees in Medicare HMOs tend to be healthier, 19,20 although no such conclusion is supported in the private insurance sector.21- 24 More comprehensive coverage of vision services in Medicare HMOs may disproportionately attract persons with vision problems. Also, incidence rates in this study are conditional on the individual undergoing eye examination during the period. Persons who chose not to be examined might have been less likely to have specific eye diseases. On balance, the selection effect due to HMO membership is likely much smaller than is selection due to eye examinations, because enrollment in HMOs was less than 20% and in the earlier years much less than this.
Our sensitivity analysis indicates a substantial reduction in rates by using the assumption that the incident rate in the unexamined is only half that in the examined population. Such an assumption is likely to be warranted more in the case of DR and ARMD than in the case of glaucoma. With DR and ARMD, the macula and central vision are commonly affected relatively early in the disease course, so patients are likely to come in for eye examinations. Thus, those not coming in for eye examinations could be less likely to have the condition. Our alternative lower-bound rates reduce our incidence rates and make our estimates more comparable with those of previous studies.
Glaucoma, on the other hand, less often has visual symptoms early in the course of disease. Although authors of a recent article emphasized that symptoms are common at initial examination for diagnosis, 25 no controlled study has been performed to assess symptoms in general among all patients seeking examination. Thus, we cannot assume that beneficiaries would notice glaucoma symptoms enough to schedule an examination. Nevertheless, the rates of glaucoma we found in this representative sample of Medicare beneficiaries are appreciably higher than those in population-based studies. Even the sensitivity analysis estimates were higher than published rates, except for the Melbourne study.16 High incidence rates in all 3 diseases emphasize the benefit of regular eye examinations in this vulnerable population. For example, the incidence of diabetes mellitus in general is increasing.26,27 Thus, the number of potential complications to be monitored is also increasing.
The large Medicare database is an excellent source of information about chronic disease incidence, given the similarity of the estimates to known population-based study estimates. In particular, conditions with well-defined and standardized diagnostic criteria, such as DR, appear to be well-suited for estimating disease incidence and complementing the results of population-based studies. For conditions or population segments in which the incidence rates are too low on an annual basis to allow for economically viable study on a population basis, longitudinal claims analysis may provide an important information source. For conditions in which the diagnostic criteria are not well established in a meaningful way, such as glaucoma, substantial differences may exist between claims-based estimates and population-based studies that use stricter definitions. Estimates of incidence rates of glaucoma in the population-based studies may well be higher and similar to those of our analyses if their definitions included cases of ocular hypertension or those with merely suspicious-looking optic nerves.28
Eye care is unusual within medicine and health care in that it has several excellent population-based studies of the incidence of the major eye diseases. However, data about less common conditions are lacking, and estimates are subject to variation in precision because of the usually small sample sizes of population-based studies. The potential benefit of large representative claims databases, such as Medicare's, is a much larger sample size in a national cohort of individuals and a long follow-up. The calculation of population-level incidence rates depends in large part on the extent to which disease incidence is reflected in provider contacts. Estimating the number of undiagnosed cases among persons with no provider contact is the greatest challenge. Another challenge for the provider community is to use more specific criteria for the diagnosis of conditions to ensure that we all describe the same condition.
Corresponding author and reprints: Frank A. Sloan, PhD, Center for Health Policy, Law, and Management, Box 90253, Duke University, Durham, NC 27708 (e-mail: firstname.lastname@example.org).
Submitted for publication September 24, 2002; final revision received April 1, 2003; accepted June 8, 2003.
This study was supported in part by grant RO1-AG-17473 from the National Institute on Aging, Bethesda, Md, and by a Lew Wasserman Merit Award (Dr Lee) from Research to Prevent Blindness, New York, NY.