ESTIMATING DISEASE PREVALENCE AND INCIDENCE USING ADMINISTRATIVE DATA: SOME ASSEMBLY REQUIRED

The prevalence and incidence of a disease are among the most fundamental measures in epidemiology. Prevalence is a measure of the burden of disease in a population in a given location and at a particular time, as represented in a count of the number of people affected. Counts of the number of people affected with a disease are required to plan appropriately for their health care needs. For example, large numbers of patients with rheumatoid arthritis will necessitate training and staffing with more rheumatologists than if the number of patients with rheumatoid arthritis was low; similarly, having many patients with osteoarthritis will require more orthopedic surgeons and surgical beds. Prevalence estimates, which adjust these counts to the size of the source population, are also useful clinically by providing context for diagnostic decision-making. Knowing that coronary artery disease is much more common than myocarditis is helpful in evaluating patients with anterior chest pain. Prevalence may also be used to compare disease burden across locations or time periods. However, because prevalence is determined by not only the number of persons affected but also their survival, prevalence is a less useful measure in studies of etiology than incidence rates. Incidence rates represent the number of new cases of disease among the number of susceptible persons in a given location and over a particular span of time. The primary value of incidence rates is in studies of disease etiology, by comparing how the rates vary among different subgroups or with different exposures.

To provide prevalence and incidence rate estimates that are both reliable and generalizable, studies must include a sample large enough to capture most (if not all) cases and sufficiently distributed, both geographically and sociologically, to be representative of the general population. With uncommon diseases, including most autoimmune rheumatic diseases, the challenge is multiplied because cases are fewer and harder to find. These factors necessitate surveys of even larger populations to achieve stable estimates (as well as longer durations of observation for estimates of incidence), which in turn increase the cost, time, and effort involved in executing such studies. Because of these issues, studies using primary data collection to determine the prevalence and incidence of diseases such as systemic lupus erythematosus (SLE) are not common [1]. Understandably, investigators have sought ways to circumvent these issues while still obtaining valid and reliable estimates. In many ways, administrative data that include diagnosis codes fit the bill. Administrative data are data collected for monitoring, reimbursement, or regulatory purposes, most often by government agencies or insurers, and not primarily collected for research purposes. However, because administrative data often cover large proportions of the population (with near universal coverage in Canada) and systematically collect similar data elements over years, administrative data are an attractive resource for epidemiological studies. Administrative data have previously been used to obtain estimates of the prevalence or incidence of SLE in Canada, Denmark, Finland, Hong Kong, Iceland, Italy, Norway, Sweden, Taiwan, and the United States [2–11].

Studies of disease prevalence and incidence involve three main activities: assembling the cohort to study; sorting people into affected and unaffected groups (case ascertainment); and counting the number affected. Most methodological work in the use of administrative data to estimate disease prevalence and incidence has focused on the validity of case ascertainment in these data sources [12]. Coding errors, “rule-out” diagnoses, and limits in the number of diagnoses included in a dataset can contribute to errors in administrative data, and the accuracy may vary from data source to data source. Although counting may seem straightforward, capture-recapture methods have been developed to test whether there may be a substantial undercount, and to estimate the number of missing cases [13]. In contrast, relatively little attention has been given to how assembly of the cohort may affect the rates.

The type of data source (e.g. hospitalization, outpatient billing), the geographic locations, age or other demographic limits, and number of years of data to include are the main elements to consider in assembling a cohort. In this issue, Ng and colleagues studied whether estimates of the prevalence and incidence of SLE were sensitive to the number of years of administrative data examined [14]. Estimates of SLE prevalence and incidence were based on both hospital discharges and physician billing codes. The authors defined prevalent cases as those ever coded as having SLE, and patients maintained the diagnosis until death; incident cases were labeled at the first occurrence of a diagnosis of SLE in the database. Using a 15-year period as the reference standard, the authors demonstrated that examining successively shorter time periods resulted in substantially lower prevalence estimates. For example, the prevalence was 46/100,000 when five years of data were examined, compared to 60/100,000 when the full 15 years of data were examined. The undercount presumably represents persons who at one time had a health care encounter with a diagnosis of SLE but were not captured as such when a shorter time window was used. Conversely, incidence was higher with shorter time windows compared to the full 15-year period, because in the shorter window prevalent cases were misclassified as new diagnoses. The incidence was estimated to be 8/100,000 with five years of data, but only 5.6/100,000 with 15 years of data. The authors suggest that more than five years of data are likely needed to avoid these issues and provide valid estimates.

This study highlights the importance of cohort assembly in studies using administrative data, and in particular, the need to think carefully about not only the data included in the study, but also about the data omitted. This study shows that the omitted data can have a major impact when studies examine only a small segment of time. Administrative data are susceptible to these effects because all persons with the disease are not identified continuously, but rather only flagged when they use health care services, and because incident cases can’t be distinguished from prevalent cases in a given year, but only by looking in previous years. Despite the importance of these observations, some caveats remain. Ng and colleagues assessed period prevalence over 15 years, which is quite long. Point prevalence studies or studies with period prevalences of 5 years or less are more common, because these studies provide more contemporary estimates. A potential limitation of using 10-year or 15-year periods is that they may not accurately reflect the current status of the disease, particularly if temporal changes have occurred, due for example, to the introduction of a more sensitive diagnostic test or a remission-inducing treatment. The underestimate of prevalence during successively shorter time windows noted by Ng and colleagues was due to previously diagnosed patients not having a health care encounter coded as SLE in these windows. Their absence from the shorter time windows indicates they are not currently consuming health care resources related to SLE. If a major objective of prevalence estimation is to aide health care planning, these patients may be less relevant. In addition, studies of incidence rates based on administrative data tend to be of limited value to understand disease etiology, because these sources generally lack data on potential exposures. Comparisons are often limited to demographic characteristics and geography, which can provide only crude suggestions about disease etiology. One exception to this generalization is that temporal trends in incidence rates may provide clues to etiology if these can be linked to trends in other data sources.

Despite the greater availability of administrative data, researchers may not have access to a decade or more of data. Can the principles outlined by Ng and colleagues be incorporated in studies that include only a few years of data? In prevalence studies, a short window will only capture a proportion of the all patients, for example those who happen to be hospitalized or who have a physician visit coded as SLE-related in the years included. If the proportion of patients who are hospitalized or who are treated in a given year can be obtained from other sources or from clinic data, an estimate of the number of prevalent patients in the population can be derived by dividing the number observed in the administrative dataset by the proportion hospitalized (or treated) per year. This inflation adjustment results in prevalence estimates that are quite accurate, particularly when the estimate of the proportion hospitalized per year is also population-based [15]. For incidence studies, a one- or two-year lag period can be included, so that any patients who appear in this period are not counted as incident. Incident cases would only be counted among those who were known to have at least some years of follow-up without a prior qualifying diagnosis. This lag helps to approximate observational studies of incidence by assembling a cohort “without disease” at entry. Consideration of these issues will result in studies with more valid estimates of disease prevalence and incidence.