Data Source and Study Population

The probability of the events of interest (e.g., incidence of a comorbid condition, admission to a health service) and the annual cost of care of someone with the condition (other than a nursing home admission or the annual cost of pharmaceuticals) were estimated from a 5% random sample of Medicare beneficiaries in the Physician/Supplier Part-B files from January 1, 1999, through December 31, 2005 (the 5% Standard Analytical Files of the Medicare Limited Data Set; see http://www.resdac.umn.edu/Medicare/data_file_descriptions.asp ). A sample of beneficiaries 65 years of age or older was selected. For this study, the POAG cohort included beneficiaries in the 5% Standard Analytical Files with a claim that included the International Classification of Diseases, Ninth Edition (ICD-9), code of 365.11 in 2001, and a claim with the same code in 1999 or 2000. By doing this, we attempted to control for incorrect coding of cases in 2001. We excluded people with moderate or severe visual impairment from this sample. Therefore, anyone with a claim including the ICD-9 code for visual impairment (i.e., ICD-9 code 369.XX) from 1999 through 2001 was excluded. We created the control cohort by matching 1:1 with the POAG cohort for age, gender, race (black vs not black) and the participant’s Centers for Medicare and Medicaid Services region. In addition, we controlled for comorbidity between the 2 groups by calculating an unweighted Charlson index score. We matched participants by their Charlson score, based on a score of 0, 1, or 2 or more. Inclusion in the control cohort required that the person be without POAG in any year from 1999 through 2005 and without moderate or severe visual impairment from 1999 through 2001.

Additional exclusion criteria for both cohorts included a claim in 2001 for depression, dementia, hip fracture, or other fracture (nonhip; see Technical Appendix for an ICD-9 definition of each condition at http://ceem.wustl.edu/LinkClick.aspx?fileticket=OghOpmsKJCg%3d&tabid=85 ). We excluded people who filed claims for services at a nursing home or a skilled nursing facility or for home health care services during 2001, as well as anyone enrolled in a Medicare health maintenance organization from 1999 through 2005.

During the period considered in the study, Medicare did not provide pharmacy benefits, nor did the program provide substantial coverage for nursing home care. However, state Medicaid programs did provide coverage for medication and nursing home services during the study period. Therefore, we estimated the annual cost of pharmaceuticals for people with the conditions of interest and the health care costs of people with a nursing home admission using data from the California Medicaid (Medi-Cal) claims database. We made the assumption, supported by the work of Hunt and associates that this would represent a conservative estimate of the cost of medication compared with Medicare and private payors. The purpose of our investigation was to determine the incremental cost of nonvision care for people with glaucoma resulting from an increased risk of other systemic conditions. Therefore, both the control and POAG cohorts were assumed to begin with the same baseline costs, and no consideration was given to the increased cost of POAG because of ocular medications.

Markov Decision Model

We modeled the lifetime cost of visual impairment resulting from POAG with an attributable cost approach by using a Markov decision model and a microsimulation process to compare the cost of care for people with POAG with that of a control group of people who did not have POAG. In this study, we defined attributable cost as the difference in cost between those beneficiaries with POAG and those without POAG that remains after accounting for other identifiable sources of difference other than POAG. For the purpose of this investigation, POAG is coded with 365.11 on Medicare claims using the ICD-9 convention. We modeled the increased risk and associated costs of the occurrence of the events reported to be associated with increased visual impairment. Specifically, we examined depression, dementia, hip fracture, fractures other than of the hip, nursing home admission, and home health use. We also considered skilled nursing facility admission and increased incidence of moderate and severe visual impairment and blindness. For this study, we defined moderate and severe visual impairment according to the standards of the ICD-9 codes (369.x). According to these standards, moderate vision loss is considered to be present if there is severe impairment in one eye and normal to near-normal vision in the other, whereas severe vision loss is considered to be present if there is severe impairment in the best-seeing eye. Blindness is considered to be present if there is profound impairment in both eyes.

The Markov model is an analytic tool that allows an investigator to estimate the costs and consequences of a disease process. We modeled the disease progression in a cohort of people with a particular condition from the time of diagnosis. During a Markov cycle (e.g., 1 year), each living cohort member is exposed to the risks, benefits, and costs associated with the progression and treatment of a disease as they progress through associated health states. During each cycle, a proportion of people die as estimated from actuarial sources. The Markov model used in this investigation is illustrated in the Figure .

Illustration showing the Markov model used to estimate the economic burden of glaucoma. The circles indicate the health states the participant might enter during a Markov cycle (i.e., 1 year). A participant can experience any of the comorbidities, places of service, or increasing severity listed below the health state. If the participant dies during the cycle, the participation in the trial ends and no events are incurred during that cycle, or any subsequent one.

Estimation of the Markov model was conducted using a microsimulation approach. In such an approach, the model is iteratively estimated at the level of the individual behavioral entity, in this case the patient. A hypothetical cohort of patients is followed throughout the study time horizon (in this case, the remainder of the hypothetical person’s life span), thus mathematically replicating the conditions of a clinical trial with each individual in the cohort preceding through the model on an individual basis. The microsimulation approach includes randomized individual patient characteristics and tracks the progress of individual patients through particular health states in terms of health outcomes and costs. Thus, it better replicates the degree of random variation seen in a natural setting than the traditional decision analytic modeling process.

The Markov model was stratified by age and gender, with the initial age and gender distribution of the individuals in the model equal to those of the people identified with POAG in the Medicare claims database. Each person in the Markov model began in a health state without moderate or severe visual loss, and we assumed that the distribution of severity of POAG among members of the POAG cohort at baseline reflected that seen in Medicare beneficiaries, and therefore we could rely on the rate of incidence of visual impairment seen in the Medicare sample to estimate the rate of progression from simple glaucoma to glaucoma with visual impairment. During a cycle, each cohort member was permitted to transition to a new health state that represents visual impairment, one or more of the following events, or both: depression, dementia, hip fracture, other fracture, home health use, skilled nursing facility admission, or nursing home admission. As there has been no strong evidence reported indicating differential mortality associated with glaucoma; the mortality rates used in the Markov model reflected those observed in the control sample drawn from the Medicare claims database. We made the assumption that those who die during a cycle die at the beginning of the year. All transition probabilities were calculated as Kaplan-Meier estimates of the number of people in the cohort subgroup having at least 1 claim for the condition of interest during the follow-up (2001 through 2005). To create a probabilistic estimate for the microsimulation model, the transition probability was parameterized using a binomial distribution. With the exception of those cohort members in whom dementia develops, the transition probability for a health event did not depend on whether the individual experienced any health events in the previous year. Because there is no known cure for dementia, a person who advanced to that condition remained in that state until death (while still facing risk of other events). In the same manner (and for the same reason), cohort members who progressed to moderate or severe visual impairment could not improve from that condition for their remaining lifetime.

When a condition first developed in a cohort member, they were assigned a cost that we refer to as the incidence cost. Cohort members in whom the condition developed a second or third time in a later cycle or, as in the case of dementia, continued to experience the disease in later cycles were assigned a lower cost (prevalence cost). We did this to recognize the increased cost in management of a condition when it is first diagnosed, as opposed to maintenance costs of a chronic condition. We made the conservative assumption that the cost of managing someone with a chronic condition is not influenced by their underlying POAG. Therefore, the cost associated with the comorbid conditions and other events were calculated from the claims records in the control cohort. Incidence costs were calculated based on the costs of care for 12 months after the first claim for an event. Prevalence costs were calculated by taking the average cost of care in 1 year for a cohort of people with a claim for that condition. Additionally, we calculated the incremental costs of each health event. Because costs for the health event included all costs that the beneficiary incurred in a given year, it would be easy to double-count healthcare costs if a person experienced 2 health events in 1 year. To account for this, we subtracted the cost of living 1 cycle without experiencing a health event from the cost of living a cycle with one of the aforementioned health events. This is an estimate of the cost of care related only to the health event.

The two simulated cohorts, each with 10 000 members individually walked through the Markov model, repeating cycles until they died. The cost and events experienced by each cohort member were recorded, and the mean and median difference in costs between the cohorts was estimated along with the 95% confidence interval (for the mean difference) and 95% range of values (for the median difference).

Further details of the methods and analyses in this study are provided in the Technical Appendix to this report, available at http://ceem.wustl.edu/LinkClick.aspx?fileticket=OghOpmsKJCg%3d&tabid=85 .