Rheumatoid arthritis (RA) is an autoimmune disease characterised by pain, morning stiffness, joint swelling, deformity and functional impairments. Patients with RA have an increased risk of mortality and several adverse outcomes such as infections and cancer compared with those who do not have RA 1234. Several studies, however, suggest that complications in RA patients may not be attributable to the disease itself, but to the use of disease-modifying anti-rheumatic drugs (DMARD). For instance, tumour necrosis factor (TNF) α blocking agents have an association with specific types of infections and may be related to an excess risk of lymphomas and neurological complications 56789. Conventional DMARDs may also increase the incidence of lymphoma 1011.

In studies that seek to determine the relationship between drug therapy and adverse events, disease severity is an important confounder. That is, disease severity is known to increase the risk of many adverse events and is probably associated with a higher likelihood of receiving more immunomodulating DMARDs. Failure to adjust for such confounding by indication can create false associations between the exposure and study outcome 12.

Health care utilisation ('claims') data are routinely collected for insurance and are commonly used in health services research 1314. Because adverse outcomes of RA are relatively rare, health care utilisation databases are an ideal source of information for studies of the relationship between DMARDs and adverse events such as cancer and infections. Thus, the development of an RA disease severity measure from claims merits high priority. We believe that health care claims data contain information such as physician visits, surgeries and laboratory tests that correlate with RA disease severity. Thus, to develop a claims-based severity index, we first created an RA medical records-based index of severity (RARBIS) from ratings by a Delphi panel on potential markers of RA severity commonly found in medical charts 15. We then assessed the performance of the RARBIS in a cohort of Veteran's Administration (VA) patients and showed that the RARBIS correlated moderately well with RA treatment intensity and thus exhibited construct validity 16. Next, we established the convergent validity of the RARBIS against a widely-used and accepted RA clinical measure, the Disease Activity Score (DAS28) 17. The goal of the present study was to develop a claims-based severity index (claims-based index for RA severity [CIRAS]) using the previously validated RARBIS, not the DAS28. If validated as a measure of RA disease severity, the CIRAS may serve as a potentially important tool in adjusting for RA severity in pharmacoepidemiology studies of RA treatment and complications using health care utilisation data.

Characteristics of the study population are summarised in Table 2. The study cohort was predominantly male with a mean age of 71 years. During the chart review study period, most had no functional limitations (78%) and did not require a device or wheelchair for ambulatory purposes (66%). About one-half of the population had swollen joints, morning stiffness that lasted less than one hour, but did not have an arthritis flare. The mean score for the RARBIS with medications was 4.4 (range 0 to 11) and without medications was 3.0 (range 0 to 8).

Table 3 provides the unadjusted Spearman correlations for the claims-based RA severity variables and the RARBIS with and without the medication sub-scale using data from one year before the chart review study period. The variables for rheumatology visits, inflammatory markers and other laboratory markers yielded the highest correlation with the RARBIS. In our analysis using administrative data from one year before the chart review period, the highest correlation between the RARBIS and the medication variable were obtained using the medication variable created from the sum of all DMARD prescriptions in method one. For both the RARBIS with and without medication scale, having data from two years before the chart review period did not substantially increase the Spearman correlation coefficients and, in some cases, even decreased the value of the coefficients (data not shown).

Table 4 presents the adjusted correlations between the claims-based RA severity variables and the RARBIS with and without the medication sub-scale with data from one year before the chart review study period. The forward selection models yielded the highest model R² for both the RARBIS with the medication sub-scale (R² = 0.31) and the RARBIS without the medication sub-scale (R² = 0.26). Using two years of data resulted in lower model R²s (data not shown).

Table 5 includes the means and ranges for the CIRAS scores and the Spearman correlation coefficients between the CIRAS and the RARBIS. The CIRAS score with the highest correlation with the RARBIS included the following components: orders for inflammatory markers, rehabilitation visits, age and gender, rheumatoid factor, presence of Felty's syndrome, number of platelet counts and chemistry panels ordered, and rheumatology visits. Figure 1 is a graphic representation of this CIRAS score in tertiles versus the median and interquartile range for the RARBIS with medication sub-scale. Table 6 presents the suggested scoring method for the CIRAS.

We developed a claims-based RA severity index (CIRAS) that demonstrated moderate correlation with a previously validated medical records-based index, the RARBIS. The RARBIS has been previously shown to have good construct validity and moderate convergent validity with the DAS28 1617. Because health care utilisation databases are a valuable source of data for studying health outcomes, other investigators have also used medical records-based indices to create indices for administrative databases. For instance, Deyo and colleagues adapted the Charlson Comorbidity Index, a well-validated index designed for medical records, so that International Classification of Diseases, Ninth Revision codes could be used to calculate the Charlson Comorbidity Index from administrative data 18. Components of the administrative-based index we developed for RA include orders for inflammatory markers, number of platelet counts and chemistry panels ordered, rheumatoid factor, rehabilitation visits, age and gender, presence of Felty's syndrome and number of rheumatology visits. If the CIRAS is found to be valid in other populations, then it might be used to partially adjust for an important confounder, disease severity, in claims-based epidemiology studies. In our analysis, we used data taken from one and two years before the chart review study period. However, using two years of data resulted in lower R² and Spearman correlation values. Including another year of older data might have caused a dilution effect. Additionally, to compute scores on the CIRAS, we used weights from the regression models with the RARBIS. Other methods of weighting could have been chosen, for example, assigning a value of one to administrative variables that had significant correlations with the RARBIS. However, we believe that the method we selected, using beta coefficients as weights, better captures the relationship between the CIRAS and the RARBIS.

Because administrative data are collected primarily for reimbursement purposes, some question the use of claims data for clinical research regarding disease severity 19. However, administrative data are gaining increasing acceptance in health care research, because they represent typical populations, contain large cohorts of patients with given conditions and are readily available. We also demonstrate in the present study that indicators of RA severity from claims data are moderately well related to clinical indicators of RA severity. Thus, it is possible to capture RA disease severity to some degree in claims data. Other proxies for severity of illness measures using claims data such as the diagnosis related group, the all patients refined diagnosis related group and the International Classification of Diseases Ninth Revision-Based Illness Severity Score have been developed 20. Unlike the CIRAS, these other measures are not specific to RA.

The present study has important limitations. Our data source for this study was the New England VA Health system. The VA's population is mostly older men. Older male patients with RA may not represent typical RA patients. This highlights the need to consider these findings as preliminary and requiring replication in other settings. Additionally, data from the VA might be gathered differently from other health care systems, again highlighting the preliminary nature of our findings. However, because the VA contains rich data from both medical record and health care utilisation databases, it is a unique and ideal data source for our analysis. Additionally, the RARBIS, which we used to create the CIRAS, was developed using standard nominal group technique methods, followed by assessing its convergent validity with the DAS28. However, the DAS28 is a measure of disease activity not disease severity. While disease activity is an important component of disease severity, it is not the same. Currently, there is no standard RA disease severity measure.

In our cohort of 120 VA patients, the CIRAS showed moderate correlations with a validated medical records-based index and can be used for improved adjustment of RA disease severity in claims data studies. We do not believe that the value of the CIRAS will be limited to the VA population. We plan on assessing its validity in other populations, such as Medicare patients, and will examine its ability to adjust for confounding and predictive validity for outcomes known to be associated with severe RA, such as future joint surgeries, higher medical care costs and use of combination DMARDs. Additionally, we will explore whether different variations of the CIRAS should be used depending on the study outcome of interest. Ultimately, the CIRAS may be an important methodological tool for researchers studying RA treatment and complications using health care utilisation data, but further tests need to be conducted in other populations.

We developed a claims-based severity index (CIRAS) from a previously validated medical records-based index (RARBIS). The CIRAS can potentially be used for improved adjustment of RA severity in studies of RA medication use and adverse outcomes using claims data, but future studies should examine its validity in other populations.

ACR = American College of Rheumatology; CIRAS = claims-based index of rheumatoid arthritis severity; CRP = C-reactive protein; DAS = disease activity score; DMARD = disease modifying anti-rheumatic drug; ESR = erythrocyte sedimentation rate; RA = rheumatoid arthritis; RARBIS = rheumatoid arthritis records-based index of severity; TNF = tumour necrosis factor; VA = Veterans Administration

The authors declare that they have no competing interests.

GT analysed the data and drafted the manuscript. SS provided support on the statistical analyses, interpretation of data and helped edit the manuscript. RS provided access to the data and helped edit the manuscript. JNK and MEW provided advice on the conceptual design and helped edit the manuscript. MY provided access to the data and helped edit the manuscript. JA contributed conceptual advice and helped edit the manuscript. DHS provided conceptual design, analytic support, access to the data and helped edit the manuscript.