The NHEFS-Medicare linked dataset used in this paper allows us to evaluate the identification of chronic conditions by the CCW algorithm. We examined the CCW algorithm from two perspectives: (1) ability to identify preexisting conditions; and (2) ability to distinguish between preexisting and newly diagnosed conditions. The CCW has the potential to facilitate health services and epidemiological research on chronic conditions. However, researchers should be aware of its limitations to avoid drawing incorrect conclusions about the population of beneficiaries being analyzed. In particular, users of Medicare data with chronic conditions identified by the CCW algorithm should consider how the definitions (specifically the reference or look-back period) may affect the detection of chronic conditions.
The survey data linked to Medicare claims give us an opportunity to estimate the length of claims history needed for the CCW algorithm to identify a preexisting condition. For example, if the study subject was diagnosed with a chronic condition in 1985 and his/her available Medicare claims history starts in January 1992, what is the probability that the CCW algorithm will identify a claim for this condition in the first year of enrollment? In the second year? The answers to these questions depend on the severity of the condition and the need to utilize health care services. Thus, a person with arthritis may not have claims related to arthritis for a long time, while a person with IHD is more likely to visit a doctor regularly and therefore will more quickly generate a claim with the corresponding ICD code. For three of the conditions (diabetes, IHD, and dementia), a relevant claim was identified on average within the CCW reference period. Yet there was variation in the proportion of preexisting cases that were identified. The CCW algorithm and reference period identified a higher proportion of preexisting cases of diabetes (69 percent) and IHD (63 percent) compared with the three other conditions. (See .)
The intersection of disease etiology and health care utilization influences the interpretation of the chronic conditions identified by the CCW algorithm. If the CCW reference period is strictly followed, we have shown that approximately 84 percent of arthritis cases identified were preexisting compared with approximately 50 percent of IHD and COPD cases. (See .) Thus, while most of the arthritis cases identified by the CCW algorithm in the reference period were preexisting cases, only a small proportion of the beneficiaries who actually had previously diagnosed arthritis (about 17 percent shown in ) were captured by the algorithm.
Conclusions drawn from our study should keep in mind several limitations in data and methods, which may affect the ability of the CCW algorithm to identify preexisting chronic conditions. The sample sizes in the linked dataset were not sufficient to do a detailed analysis by age, sex, or race. It is possible that the identification of chronic conditions varies by demographic characteristics. Because only about half of the original NHEFS participants born in 1935 or earlier survived until 1991 and could be linked to Medicare claims, our analysis was conducted on a subsample that is not fully representative of the original NHEFS participants. In addition, the NHEFS data are from a longitudinal survey and include information from different points in time; not all respondents have the same amount of information.
Even though NHEFS covers approximately 20 years, the data may not be complete. NHEFS diagnoses are based on several sources: interview responses, examination results, medical records (hospital discharges and nursing home admission records), and death records. Each source has its own limitations. For example, the self-report of conditions could contain errors. In addition, the study definition for some chronic conditions changed between waves of the follow-up study, in some cases because the medical definition of the condition changed, and in other cases because of changes in the questionnaire. For example, during the study period, the definition of diabetes changed and the previously recommended oral glucose tolerance test was replaced with a recommendation that the diagnosis of diabetes mellitus be based on two fasting plasma glucose levels of 126 mg/dl or higher (Mayfield 1998
). New tests for dementia continue to be developed leading to advances in understanding and detection (Mani et al. 1999
; Cummings 2000
). The follow-up interviews did not ask about conditions such as COPD or IHD, and in these cases, the study data rely only on medical records for overnight facility stays. In addition, the hospital and nursing home records provided ICD-9-CM codes for the various conditions; finding a chronic condition using these records depends on the accuracy of the clinical coding and may be less precise for comorbid conditions than for primary diagnoses (Kern et al. 2006
). Similarly, mortality records contain on average one to two comorbid conditions, in addition to the underlying cause of death, and they may not include all contributing chronic conditions among older deceased participants (Gorina and Lentzner 2008
In contrast to NHEFS data, claims data (with the CCW algorithm) identify chronic conditions based only on utilization of health care for that particular condition, such as a doctor visit or hospitalization. Therefore, patients in remission or with conditions that do not require health care for some period of time may be missed by the claims data analysis (Joyce et al. 2005
). Beneficiaries may underutilize health care due to limited access to care (e.g., in rural areas) and thus not generate claims for a particular chronic condition or may utilize services not covered by Medicare. Previously uninsured or under-insured new Medicare enrollees may utilize health care at a greater rate in the first few years after enrollment than before; thus, some conditions may be identified by the CCW algorithm at a greater rate in the first years after enrollment than in later years. Enrollees in this analysis with a claims history that starts at the time of enrollment in Medicare may have a greater chance of being diagnosed with chronic conditions than those whose early claims history is not available in the analytic sample. In addition, the accuracy of Medicare claims has to be taken into consideration (Taylor, Fillenbaum, and Ezell 2002
; Losina et al. 2003
; Handlon and Cleverley 2006
). Finally, the CCW algorithm is applicable only to the claims of beneficiaries who were enrolled in fee-for-service Medicare Part A and Part B and not enrolled in a Medicare HMO, and therefore may be subject to selection bias (Mello et al. 2003
Our findings differ from the results of Katz et al. (1997)
, who find high levels of agreement for certain arthritic conditions between medical records and Medicare physician claims. The medical records in that study, however, were taken from visits to rheumatology specialists, where one would expect arthritis to be reported more accurately than it might be over the full range of providers, services, and facilities.
We also do not find the same high levels of agreement with claims data for dementia as found by Taylor, Fillenbaum, and Ezell (2002)
, who compared Medicare claims with an Alzheimer's disease registry. The identification of dementia in NHEFS is confounded by a variety of issues. Mortality selection plays a major role in who survives to be linked to Medicare records. Some cases of dementia may have been missed because the participants did not survive to the time period of the Medicare claims linkage. Because only 13 percent of our analytic sample was older than 65 years at baseline, it is not surprising that conditions especially associated with aging (e.g., dementia) may be harder to detect in longitudinal data. Dementia is difficult to diagnose in its early stages and may not be accurately reported in survey data or be the primary reason for health care utilization. Taylor, Fillenbaum, and Ezell (2002)
do find that less severe cases of dementia are less likely to be captured by claims data.
The size of the Medicare beneficiary population and the high prevalence of chronic conditions among this population necessitates its continuing study to improve health and health care policy. The creation of the CCW offers potential for expanded use of Medicare claims data for analysis. Comparing conditions identified by the CCW with outside data, we have shown wide variation in the ability of the CCW algorithm to identify chronic conditions. The reference periods embedded in the CCW algorithm may not be sufficient for all analyses. Depending on the research question, users need to consider how the intersection of time since enrollment, specific condition, age of enrollee, number of years of claims available, and whether a condition is preexisting or newly diagnosed, may affect a study's results and interpretations. For example, analyses using the CCW algorithm's identification of arthritis cases may underestimate the overall cost of treatment since the population captured in a given year misses many beneficiaries with previously diagnosed arthritis. Analyses using beneficiaries recently enrolled may miss cases where more years of data are needed to identify conditions. We encourage continued research and development on the CCW to refine this valuable resource.