Linkage of a De-identified United States Rheumatoid Arthritis Registry with Administrative Data to Facilitate Comparative Effectiveness Research

Abstract

Background

Linkages between registries and administrative data may provide a valuable resource for comparative effectiveness research. However, personal identifiers that uniquely identify individuals are not always available. We describe methods to link a de-identified arthritis registry and U.S. Medicare data. The linked dataset was also used to evaluate the generalizability of the registry to the U.S. Medicare population.

Methods

Rheumatoid arthritis (RA) patients participating in the Consortium of Rheumatology Researchers of North America (CORRONA) registry were linked to Medicare data restricted to rheumatology claims or claims for RA. Deterministic linkage was performed using age, sex, provider identification number, and geographic location of the CORRONA site. We then searched for visit dates in Medicare matching visit dates in CORRONA, requiring at least 1 exact matching date. Linkage accuracy was quantified as a positive predictive value (PPV) in a sub-cohort (n=1581) with more precise identifiers.

Results

CORRONA participants with self-reported Medicare (n=11,001) were initially matched to 30,943 Medicare beneficiaries treated by CORRONA physicians. A total of 8,431 CORRONA participants matched on at least 1 visit; 5,317 matched uniquely on all visits. The number of patients who linked and linkage accuracy (from the subcohort) was high for patients with >2 visits (n=3458, 98% accuracy), exactly 2 visits (n=822, 96% accuracy) visits, and 1 visit (n=1037, 79% accuracy) visit that matched exactly on calendar date. Demographics and comorbidity profiles of registry participants were similar to non-participants, except participants were more likely to use DMARDs and biologics.

Conclusion

Linkage between a national, de-identified outpatient arthritis registry and Medicare data on multiple non-unique identifiers appears feasible and valid.

Background

All data sources have strengths and weaknesses that must be considered in selecting one to conduct comparative effectiveness research. Registry or cohort data typically excel in capturing clinical and phenotypic information. However, they are sometimes limited in the frequency and scope of data collection due to concerns about participant burden and cost. Large administrative claims databases have been shown to be valuable and have been widely used to study comparative effectiveness research questions, especially for safety outcomes (1). Among their strengths are large sample sizes that allow for examination of rare exposures and outcomes, and their comprehensive capture of healthcare services provided to an individual, as well as healthcare costs. However, they generally lack detailed clinical information, and thus, concerns for residual confounding due to disease activity, severity, and other unmeasured factors need to be considered. A `hybrid' approach that would link different types of data sources together would therefore likely be valuable to overcome the limitations inherent to any type of data source (2–4). For example, an analysis conducted using an arthritis registry with data collected at physician office visits would be benefitted by linkage to an administrative database to have a more complete understanding of patients' medication adherence, both to their arthritis drugs and non-arthritis related medications. It also would allow for more complete long term follow-up of safety events of interest (e.g. hospitalization for myocardial infarction, incident malignancy) for safety follow-up and pharmacovigilance purposes, even if patients changed arthritis providers or no longer had any engagement with the registry.

A number of published examples exist where administrative data has been linked to a cohort or registry (5). In terms of the technical requirements for such linkages, having unique identifiers that are considered protected health information (PHI) such as social security numbers, date of birth (DOB), and sex are generally sufficient to permit linkages with high validity (6–8). Several examples of linkages between administrative data and inpatient procedure or device registries where unique PHI is not available also have been published. Several of these examples have linked on hospital center and other non-unique information (e.g. dates of admission/discharge) (6, 9, 10). It has been shown that this deterministic linkage using multiple non-unique identifiers produce highly accurate linkage with >95% sensitivity and specificity >98% compared to linkage using unique identifiers (11). However, in an outpatient registry, many individuals participating may not undergo hospitalization or a medical procedure, and these methods may not be sufficient to link ambulatory cohorts and registries to administrative claims data. Moreover, little guidance exists about the optimal methods or success of linking an outpatient cohort or registry to administrative claims data when unique identifier is not available.

In light of this evidence gap in linkage methodology, we sought to link a large, de-identified, outpatient registry of patients with rheumatoid arthritis (RA) with national Medicare administrative claims data. The purpose of this report is to describe the methods and validity of this linkage. Further, administrative data were used to compare characteristics of registry participants to 1) non-participants with RA, treated by the same rheumatologists, to assess for a potential selection bias in those who were enrolled in the registry and 2) non-participants with RA treated by rheumatologists not participating in the registry, to assess the national generalizability of RA patients enrolled in the registry compared to the remainder of the U.S.

Methods

Data Sources

We used data from the Center for Medicare and Medicaid Services (CMS) from 2006–2010 and selected all beneficiaries with at least one outpatient physician diagnosis of RA (International Classification of Diseases, Ninth Revision [ICD-9] 714.X). Only a single claim with an RA diagnosis code was required intentionally to maximize sensitivity. Medicare enrollees were classified as “observable” in the Medicare data if they had both Medicare part A (inpatient) and part B (outpatient) coverage and were not enrolled in Medicare Advantage. Claims for Medicare Advantage enrollees and for those with only part A or part B are typically incomplete. These data are considered “research identifiable” given the possibility that individuals could be uniquely identified, either in the data by itself or through linkage procedures of the type that we implemented. Governance of the data is under the aegis of a Data Use Agreement (DUA) from CMS that specifies what may be done with the data and who has access to it. The default DUA language generally prohibits linkage to any external data source without explicit approval from CMS, which was obtained for this project. The DUA also prohibits public disclosure of any information about groups of fewer than 11 people to avoid concerns about patient privacy.

We also used 2006–2010 data for RA patients participating in the Consortium of Rheumatology Researchers of North America (CORRONA) arthritis registry (12). CORRONA is a registry of patients with RA and psoriatic arthritis (PsA) as confirmed by rheumatologists; only participants with RA were selected. Patients explicitly consent to participate in the registry, although the possibility that the registry data might be linked to external data sources was not envisioned at the time the study protocol was created. For the purposes of this analysis, patients were considered eligible for linkage to Medicare after they first self-reported that they had Medicare coverage (of any type, including Medicare Advantage) at a CORRONA visit. Both that visit and all subsequent CORRONA visits were eligible for inclusion in the linkage procedure.

Linkage Procedures

Deterministic linkage, or rule-based record linkage, finds matches based on a set of rules about linkage variables, and was performed using multiple non-unique variables. The key variables included year of birth, sex, physician specialty (rheumatologist or not), state that the CORRONA provider practiced in, specific CORRONA provider identification number (UPIN/NPI) (13), and calendar dates of CORRONA provider office visits. The initial approach required exact matches between the calendar dates of all provider visits in the Medicare data and corresponding dates in the CORRONA data.

If a registry participant matched to more than one specific Medicare beneficiary, or vice-versa, priority was given to the matched pair with the greatest number of provider visits. For example, if a given registry participant matched 2 Medicare beneficiaries, one with 5 matched visits and another with only 1 matched visit, the pair with matched 5 visits would be considered correctly matched. If a registry participant matched to more than one Medicare beneficiary with the same number of visits, these `tied' pairs were considered unmatched. After selecting matches as described above, we implemented a procedure that allowed for `fuzzy matching' on visit dates with CORRONA providers and that accepted a mismatch of visit dates by up to ±4 days, and these were included in the primary analysis. Additional variations on mismatch in visit dates (e.g. allowing mismatch in a single digit of visit month or day) were explored but did not appear fruitful in preliminary analyses and were not considered further. Other factors that were considered as potential linkage variables and were available within both the registry and Medicare data (e.g. specific drug exposures, such as biologic medications) were either used too infrequently or imposed further restrictions on the sample size (e.g. because Medicare Part D prescription drug coverage was required) that they were deemed impractical.

Linkage accuracy was evaluated in a subcohort of 1,581 CORRONA participants with more complete information that included full date of birth and/or social security number. These individuals explicitly consented to the linkage and made this additional PHI available for this purpose. Individuals who matched on both month and year of birth, or who matched on full social security number, were considered as accurately matched.

Statistical analysis

Descriptive statistics for the cohort were used to evaluate patients' characteristics according to the accuracy of linkage. Linkage accuracy was defined at the level of groups of similar patients who were classified by the linkage procedure into the same category, as depicted in a flow diagram figure. Information from the validation subcohort with exact date of birth (not just year) and/or SSN available was used as the gold standard for linkage accuracy in the terminal categories in the flow diagram, each of which was shown using both colors and shading. Accuracy was reported as a positive predictive value (PPV) based upon results from the subcohort, and 95% confidence intervals were calculated using a binominal distribution. Characteristics of CORRONA patients who were linked with high accuracy (i.e. they were linked by the matching algorithm with an accuracy >= 80%), lower accuracy (<80%), or not linked were examined descriptively using CORRONA registry data to evaluate whether there were differences between patients who were accurately linked and those who were not.

Evaluation of the generalizability of the registry used the most recent twelve months of Medicare data to examine characteristics of the CORRONA patients (linked with >= 80% accuracy) to patients of the same CORRONA rheumatologists who were not enrolled in the registry (and did not even have a single exactly matched visit) and to RA patients of rheumatologists who did not participate in CORRONA. This comparison was restricted to patients with 3 or more Medicare claims with RA diagnoses (714.0) from a rheumatologist office visit to improve the specificity of the RA diagnosis in the Medicare data. In this analysis, 98% of all linked CORRONA registry participants met this additional criterion. Standardized mean differences (SMDs) were computed to evaluate the magnitude of differences between groups. SMDs of <0.1 were considered clinically unimportant (14). Analyses were conducted in SAS 9.3. The linkage was approved by the CORRONA central institutional review board and by the UAB institutional review board (IRB) without the requirement to obtain additional patient consent.

Results

In the Medicare data, 64,229 patients had at least one RA diagnosis from one of the 387 rheumatologists registered with CORRONA. A total of 11,001 CORRONA participants with rheumatologist-confirmed RA had self-reported Medicare coverage of any type and were potentially linkable. Of these, 10,050 CORRONA participants were preliminarily linked to 1 or more Medicare beneficiaries from a pool of 30,943 Medicare beneficiaries based on matching on year of birth, sex, and CORRONA provider number, as shown in the flow diagram in Figure 1.

Figure 1.

Flow Diagram Describing the Approach to Link Medicare Data to an Outpatient Arthritis Registry on Multiple, Non-Unique Factors

In order to improve the specificity of the linkage, Medicare claims were subsequently restricted to those from a rheumatologist or where the primary diagnosis was RA. Visits with providers were restricted to those occurring in the same state as the CORRONA provider. Following this restriction, there were 1,619 CORRONA participants who did not have at least one visit that matched exactly on calendar date in both the Medicare data and the CORRONA data, and these individuals were removed from further consideration.

The remaining 8,431 CORRONA participants were potentially linkable to a pool of 12,877 Medicare enrollees based upon having at least one provider visit match exactly on calendar date. Of these, 5,756 CORRONA participants linked to 5,979 Medicare enrollees on every visit that occurred in both CORRONA and Medicare data. The median (IQR) number of matched visits was 3 (1, 7). Not all matches were unique, as 502 patients tied with another CORRONA participant or Medicare beneficiary. After prioritizing those with the greatest number of matched visits over potential matches with fewer matched visits, 51 and 12 patients had at least 2 provider visits and were moved to the uniquely matched population. The remaining 439 tied individuals were considered unmatched.

After resolving ties, 5,317 patients were considered uniquely linked. Among these individuals, 3,458 matched exactly on more than 2 visits, shown in the red horizontal-shaded category in Figure 1. Of these, 972 were in the validation subcohort with additional identifying information. Based upon the results for patients classified into this category from the validation subcohort (n=972 individuals), linkage accuracy (i.e. PPV) for patients in this category was 956/972=98.4% (95% CI 97% – 99%). For those who matched on exactly 2 visits (n=822 in the entire cohort, 212 in the validation subcohort; blue vertically-shaded category in Figure 1), PPV was 96.0% (95% CI 93% – 98%). For those who matched on only 1 visit (n=1037, 212 in the validation subcohort; green diagonal-shaded category in Figure 1), linkage accuracy was 78.9% (95% CI 72.4% – 84.4%). If one required that patients match on month, day and year of birth rather than only on month and year, linkage accuracy was approximately 3–4% lower.

Among the 2,675 patients who matched on exact calendar dates for some but not all visits, an additional 524 matched pairs were matched on all visits after allowing mismatch by up to 4 calendar days. For those matching on more than 2 visits (n=434, 159 in the validation subcohort; orange hatched category in Figure 1), linkage accuracy was 88.7% (95% CI 82.7% – 93.2%). For those matching on exactly 2 visits (n=90, with 26 in the validation subcohort; yellow speckled category in Figure 1), linkage accuracy was 34.6% (95% CI 17.2% – 55.7%).

Characteristics of Medicare enrollees who linked with high accuracy (linkage accuracy >= 80%, n=3458+822+434), lower accuracy (< 80%, n=1037+90), or not at all (n=1619) are shown in Table 1. Linked patients were somewhat older and more likely to be female. They had a longer disease duration of RA and were more likely to have used biologics. In total, although differences were observed in RA-related and other clinical characteristics, the magnitude of other differences was generally small. Characteristics of the remaining patients not categorized as linked or definitively not linked is shown in the Appendix Table. In general, these patients' characteristics were intermediate between patients linked with high accuracy and those linked with lower accuracy.

Table 1.

Descriptive Characteristics from Registry Data of Medicare Enrollees with Rheumatoid Arthritis Who Were Linked to Registry Data with High Accuracy, Lower Accuracy, or Not at AN^*

		Patients who were not linked**	SMD vs. Linked ≥ 80% column	Patients Linked with < 80% Accuracy	SMD vs. Linked ≥ 80% column	Patients Linked with ≥ 80% Accuracy
N		1,619		1,127		4,714
Age	Mean (std)	67.47 (11.93)	0.5301	69.47 (10.61)	0.368	73.17 (9.3)
Gender	Female	1097 (67.8%)	0.1644	820 (72.8%)	0.0547	3543 (75.2%)
Race	White	1313 (81.8%)	0.1368	945 (83.9%)	0.0793	4052 (86.7%)
	Hispanic	109 (6.8%)	0.1555	49 (4.4%)	0.0503	158 (3.4%)
	Black	134(8.3%)	0.0332	91 (8.1%)	0.0237	348 (7.4%)
	Asian	21 (1.3%)	0.0303	18 (1.6%)	0.0544	46 (1.0%)
	Other	29 (1.8%)	0.0277	23 (2.0%)	0.0448	68 (1.5%)
			0.1672		0.1593
Body Mass Index	Mean (std)	29.10 (7.05)	0.5301	29.05 (7.1)	0.368	27.93 (6.87)
RA-Related Characteristics
Duration of RA in years +	Mean (std)	13.15 (10.37)	0.3167	12.78 (11.3)	0.335	16.63 (11.60)
CDAI	Mean (std)	11.93 (11.75)	0.1629	12.66 (12.02)	0.2259	10.1 (10.17)
HAQ-DI	Mean (std)	0.46 (0.55)	0.0065	0.53 (0.56)	0.1136	0.47 (0.55)
ESR	Mean (std)	27.41 (25.19)	0.0546	27.87 (23.02)	0.0376	28.75 (23.92)
Anti-CCP antibody	%	360 (64.3%)	0.0694	272 (65.7%)	0.0398	992 (67.6%)
History of biologic use	%	876 (54.1%)	0.1527	579 (51.4%)	0.2079	2905 (61.6%)
RF positive	%	626 (71.7%)	0.0556	421 (72.7%)	0.0332	2094 (74.2%)
History of DMARD use	%	1553 (95.9%)	0.204	1059 (94.0%)	0.2825	4671 (99.1%)
Patient Fatigue	Mean (std)	34.42 (30.17)	0.0243	36.97 (29.11)	0.0864	34.78 (28.82)
Patient Global Assessment	Mean (std)	33.64 (28.54)	0.0426	35.48 (28.11)	0.1137	31.80 ( 27.02)
Patient pain	Mean (std)	35.59 (29.41)	0.0472	38.41 ( 29.32)	0.153	33.19 (28.03)
Comorbidities
Chronic Heart Failure	%	32 (3.3%)	0.0334	26 (3.4%)	0.0265	88 (3.9%)
COPD	%	61 (4.6%)	0.131	59 (7.1%)	0.232	94 (2.3%)
Diabetes	%	185 (11.4%)	0.0257	153 (13.6%)	0.0909	500 (10.6%)
History of hospitalized infection while under observation	%	86 (6.5%)	0.0587	75 (9.0%)	0.035	337 (8.1%)
MI	%	99 (6.1%)	0.0034	57 (5.1%)	0.0426	284 (6.0%)
Stroke	%	65 (4.0%)	0.0653	54 (4.8%)	0.0274	254 (5.4%)
Smoker	%	216 (21.6%)	0.1132	134 (17.2%)	0.0002	396 (17.2%)

SMD = standardized mean difference; CDAI = Clinical Disease Activity Index; HAQ-DI = Health Assessment Questionnaire Disability Index; ESR = Sedimentation Rate; CCP = Cyclic Citrullinated Peptide; COPD = Chronic Obstructive Pulmonary Disease

Note: All factors were measured at the most recent CORRONA visit

^*high accuracy describes individuals in the colored and shaded categories of the Figure with a linkage PPV >80%. Lower accuracy describes participants in all other terminal categories of the Figure where a PPV is reported.

^**not considered matched by linkage procedure

The comparison of the generalizability of CORRONA RA patients to non-CORRONA RA patients of CORRONA physicians, and to RA patients of non-CORRONA rheumatologists, is shown in Table 2. CORRONA patients had similar demographics to the other two groups of RA patients although were more likely to be recruited from the northeast region of the U.S. They were similar in their prevalence of various comorbidities and in healthcare utilization. However, RA patients of CORRONA doctors were more likely to receive DMARD and biologic therapies. Indeed, 85% of CORRONA-enrolled RA patients used a DMARD or biologic, as compared with 73% of RA patients treated by non-CORRONA rheumatologists.

Table 2.

Characteristics^* of RA Patients of CORRONA doctors in CORRONA vs. not in CORRONA vs. Remainder of U.S as Identified in Medicare Administrative Data

		RA Patients of non-CORRONA doctors	SMD vs. patients of CORRONA doctors, linked to CORRONA	RA Patients of CORRONA doctors, not linked to CORRONA	SMD vs. patients of CORRONA doctors, linked to CORRONA	RA Patients of CORRONA doctors, in CORRONA
Number of RA patients, n (%)
		446274		23511		5532
Number of treating rheumatologists, n		5,111		310		280
Age	mean(std)	72.62 (10.82)	0.0219201	72.40 (12.27)	0.0002204	72.39 (9.85)
Sex	Female	332784 (74.7%)	0.0023114	16763 (71.4%)	0.0716897	4123 (74.6%)
Race	White	378675 (85.0%)	0.1366321	20226 (86.2%)	0.1033576	4949 (89.5%)
	Black	42520 (9.5%)	0.067486	2438 (10.4%)	0.0955293	423 (7.7%)
	Hispanic	11230 (2.5%)	0.1695285	318 (1.4%)	0.0931023	26 (0.5%)
	Native American	2891 (0.6%)	0.0187789	98 (0.4%)	0.0131362	28 (0.5%)
	Asian Other	4815 (1.1%)	0.0398825	172 (0.7%)	0.0032143	39 (0.7%)
		5375 (1.2%)	0.007899	220 (0.9%)	0.0182765	62 (1.1%)
Region	MidWest	102673 (24.5%)	0.0139384	3249 (14.5%)	0.2693429	1350 (25.1%)
	North East	69875 (16.7%)	0.3212634	7030 (31.3%)	0.025934	1619(30.1%)
	South	177853 (42.4%)	0.2147475	10134 (45.1%)	0.2695225	1727 (32.1%)
	West	68825 (16.4%)	0.1051288	2055 (9.1%)	0.1145216	684 (12.7%)
Rural residence		105156 (24.9%)	0.0756817	5187 (23.1%)	0.0324153	1169(21.7%)
Comorbidities
COPD		64665 (14.5%)	0.0268574	3377 (14.4%)	0.0232578	750 (13.6%)
Diabetes		94805 (21.2%)	0.0825734	4919 (20.9%)	0.0746882	994 (18.0%)
Cancer		50530 (11.3%)	0.0122488	2709 (11.5%)	0.0059731	648 (11.7%)
MI		6855 (1.5%)	0.0151046	376 (1.6%)	0.0201834	75 (1.4%)
Charlson Comorbidity Index	mean(std)	0.94 (1.75)	0.0720511	0.95 (1.74)	0.0785728	0.82 (1.63)
Health Care Utilization
Number of physician visits		14.00 (10.63)	0.0128949	14.52 (10.73)	0.0627751	13.87 (10.12)
Any hospitalization, %		130276 (29.2%)	0.0666433	7050 (30.0%)	0.0840562	1450 (26.2%)
Any joint surgery		28216 (6.3%)	0.0234481	1522 (6.5%)	0.0172803	382 (6.9%)
Patients with full-year part D, n (%)		246133 (55%)		12698 (54%)		2812 (51%)
Any non-biologic DMARD**, %		156650 (63.6%)	0.2079756	8119 (63.9%)	0.2017908	2060 (73.3%)
Methotrexate		114052 (46.3%)	0.2351051	6083 (47.9%)	0.2033205	1631 (58.0%)
Sulfasalazine		15620 (6.3%)	0.0124709	643 (5.1%)	0.0428719	170 (6.0%)
Leflunomide		23747 (9.6%)	0.0631645	1089 (8.6%)	0.1003173	326 (11.6%)
Hydroxychloroquine		52122 (21.2%)	0.0682692	2233 (17.6%)	0.0226667	519 (18.5%)
Any Biologic**, %		59920 (24.3%)	0.2006718	3364 (26.5%)	0.1510876	939 (33.4%)
Abatacept		9971 (4.1%)	0.174194	746 (5.9%)	0.0915313	231 (8.2%)
Adalimumab		11692 (4.8%)	0.0483425	672 (5.3%)	0.0235619	164 (5.8%)
Certolizumab		1416 (0.6%)	0.0642908	130 (1.0%)	0.0143671	33 (1.2%)
Etanercept		15071 (6.1%)	0.0601676	808 (6.4%)	0.0502699	215 (7.6%)
Golimumab		1080 (0.4%)	0.0759459	68 (0.5%)	0.0629324	31 (1.1%)
Infliximab		25151 (10.2%)	0.2619993	1622 (12.8%)	0.1824123	547 (19.5%)
Rituximab		5141 (2.1%)	0.0791542	307 (2.4%)	0.0572912	95 (3.4%)
Tocilizumab		219 (0.1%)	0.0156767	33 (0.3%)	0.0262631	4 (0.1%)
Any DMARD or Biologic**, %		179716 (73.0%)	0.2893637	9425 (74.2%)	0.2617548	2382 (84.7%)
Prednisone**, %		163678 (66.5%)	0.0825578	8663 (68.2%)	0.1193972	1759 (62.6%)

SMD = Standardized mean difference; n/a = not applicable

^*all factors described were measured during the most recent 1 year of Medicare data available.

^**assessed in Medicare part D pharmacy data in the subgroup of patients with 12 months of part D coverage

Discussion

Using a largely de-identified outpatient arthritis registry of participants with RA, we successfully linked almost 6,000 individuals to national Medicare administrative claims data from CMS. Linkage accuracy was shown to be 96–98% for patients who linked uniquely on 2 or more visits occurring both in the registry and the Medicare data. For patients who matched on exactly 1 visit, 79% of patients were accurately linked. For patients who linked on some but not all visits but had mismatch of only a few days between visits, those who had more than 2 visits in the registry and linked on all visits were matched with 89% accuracy. Based upon the overall linkage accuracy that we found for large numbers of patients in the registry, these results support the validity of using such methods to harness the power of a disease registry and administrative data together, even if PHI is not readily available.

The potential uses of linkages between disparate data sources are myriad. Administrative data can serve to complement a registry to enhance the spectrum of exposures (e.g. non-rheumatology-related medication exposures) and outcomes (e.g. complete healthcare utilization, costs) that can be analyzed but which the registry cannot feasibly collect due to participant burden. It can also enhance case ascertainment for events of interest (7) to capture both underreported events while patients are under follow-up by the registry. It would also allow continued observation of patients who are otherwise lost to follow-up from the perspective of the registry and support pharmacovigilance efforts and large pragmatic trials. Conversely, the linked registry data may be used to complement administrative data analyses to provide information about potentially confounding variables (e.g. seropositivity, disease activity, extent of disability, disease duration, tobacco use) that would otherwise be missing or poorly reported. Registry data can also overcome limitations due to `left-censoring' of administrative data whereby events occurring prior to the start of observation in administrative data are unknown, since a registry can capture a lifetime history of key health events and exposures of interest.

Prior literature has shown similarly high success as we found in linking registries to administrative data and examining outcomes, although most examples linked to an inpatient hospital or procedure-based registry deterministically based upon factors including date of birth, sex, and hospital admission/discharge dates or procedure dates (15–17). Examples of additional data elements that may be useful for linking to external data include social security numbers, name, and full residential address. Linkage rates may also be improved if selected additional elements are available such as specific insurance plan name (rather than simply specifying `commercial insurance'), insurance plan number, and similar details for prescription pharmacy plan coverage captured distinct from patients' medical insurance information.

With linkage to a nationally-representative data source like Medicare, linkages can be used to compare the generalizability of participants in the registry to non-participants, as we did in Table 2. In general, the demographics and comorbidity profiles were quite comparable for patients enrolled in CORRONA versus those not enrolled. However, CORRONA patients were more likely to be treated more with DMARDs and biologics compared to RA patients of non-CORRONA physicians. Some misclassification may exist in classifying a patient as having RA based only a single RA diagnosis code from a rheumatologist (18, 19), although one study found that both the sensitivity and the positive predictive value of this approach was 90% (19). However, our analysis that required to 3 RA diagnosis codes from a rheumatologist in one year in part addressed this issue and showed that 73% of RA patients of non-CORRONA rheumatologists were treated with DMARDs or biologics. While even this proportion receiving treatment may seem low, this estimate is comparable to and higher than national results derived from Medicare managed care plans, where only 63% of RA patients received these therapies (20). In that study, large variability in treatment was observed based on patient and health-plan related characteristics. It was also similar to results from RA patients represented in the Medicare Current Beneficiary Survey(21). The higher prevalence of biologic and DMARD use in linked CORRONA registry participants (85%) may reflect design features of the registry itself. CORRONA participants typically have visits at approximately 4 month intervals unless they are starting or changing therapies for RA, in which case they can be enrolled with an earlier visit interval. Results of the prevalence of DMARD and biologic use (74%) of RA patients treated by CORRONA rheumatologists but not enrolled in the registry were remarkably similar to the corresponding prevalence of RA patients of non-CORRONA rheumatologists (73%). This finding suggests that patients with more aggressively treated disease may be somewhat more likely to be enrolled than patients with more quiescent disease who are not making any changes to their RA medication regimen.

Our study is novel in that it links together a largely de-identified outpatient arthritis registry with 100% of CMS data for RA patients over a four year period. Importantly, we were able to validate our linkage procedures in a large subcohort of more than 1500 participants who explicitly provided consent and made available additional information (e.g. full date of birth, social security number) to confirm the linkage. Despite these strengths, some limitations to this work need to be considered. First, the arthritis registry did not ask participants to self-report traditional fee-for-service Medicare as distinct from Medicare Advantage. Thus, some participants who said they had Medicare coverage were not linkable in as much as only traditional fee-for-service enrollees have outpatient claims-level information available within CMS data. Additionally, we did not evaluate accuracy using unique identifiers even for patients in our subcohort but relied on matching date of birth; however, a previous validation study yielded >95% sensitivity and >98% specificity for correct matches using date of birth via a similar approach (11). We also recognize that deterministic linkage on multiple factors yielded high accuracy for individuals with 2 or more visits (red, horizontal shaded cell; and blue, vertical shaded cell), but was lower for individuals who did not match uniquely on all visits, or who had only a single matched registry visit. High specificity of linkage will be more important than high sensitivity for most applications using linked data, since it is better to not link an individual (yielding `false negatives') if it results in incorrect linkages (`false positives'). Requiring linkage accuracy of 90–95% or greater may be a reasonable requirement, and this level of accuracy was achievable for large numbers of individuals in this cohort.

Other methods for linkage exist, such as estimating a person-level probability of being accurately linked, rather than using a rules-based approach where accuracy was determined at a group level for all individuals meeting those rules(22, 23)). Using these methods in conjunction with additional data elements might be useful in some applications, but this was not feasible in this study given that few additional data elements were available within both the registry and administrative data with high enough prevalence. Additionally, not all patients ended up in classifiable groups with sufficient numbers from our validation subcohort to characterize these individuals as validly matched or not. For example, a participant with 5 visits in the registry that still had at least one mismatch in visit dates after the fuzzy matching procedure was not included in any classifiable group since we could not estimate the accuracy for these individuals without additional information. Attempting to represent these individuals as either validly matched or not was therefore not feasible. We expect some of these patients would be linkable if additional information were available. The Appendix Table showed their characteristics were similar to linked patients.

We found that linked patients were older and included proportionally more women than unlinked patients, which may reflect left-censoring of the Medicare data. This may be a consequence of younger patients being more likely to choose insurance other than Medicare fee-for-service coverage. It also may reflect that it may take several years for patients to accrue enough matchable visits in both data sources to allow for linkage using these methods. While the age and sex differences between matched and unmatched patients should not create bias, it could impact the external generalizability of the matched cohort. Finally, although we showed that patients linked by the algorithm were generally linked accurately, the sensitivity of the approach was not able to be assessed, in as much as it was unclear how many patients should have been linkable. There were 10,050 CORRONA patients with RA who reported any type of Medicare coverage, and 5,317 (53%) were uniquely linked. However, CORRONA does not distinguish between traditional Medicare fee-for-service coverage and Medicare Advantage, and Medicare Advantage claims are not available in these CMS data. In 2011, 75% of all U.S. Medicare beneficiaries were enrolled with Medicare fee-for-service coverage (24). Applying that proportion to the 10,050 CORRONA patients reporting any Medicare Coverage, the sensitivity of the method tentatively might be estimated as (0.53 / 0.75) = 0.71. A more refined approach to calculate sensitivity could be possible if patients are able to accurately self-report whether they have Medicare fee-for-service or Medicare Advantage coverage.

In conclusion, linkage between a largely de-identified national U.S. outpatient arthritis registry and U.S. Medicare claims data using multiple, non-unique identifiers appears both feasible and valid. The approach that matches partially de-identified data based upon on healthcare utilization within physician practices is likely to be useful in future studies even when individual PHI is unavailable.

Significance and Innovation.

• Methods to link de-identified outpatient registry data with health plan data (e.g. administrative claims, electronic health records) have not been well studied

• Using deterministic linkage on multiple, non-unique identifiers (or variables), the CORRONA rheumatoid arthritis patient registry was linked with Medicare administrative claims data. High accuracy of the linkage was demonstrated in a subcohort of patients that had more detailed (or precise) information on linkage variables available.

• The representativeness of the registry population was evaluated by comparing characteristics of registry participants to Medicare RA patients treated by non-CORRONA physicians. The registry participants had good representativeness of general Medicare RA patients, except that registry participants were somewhat more likely to use biologics and DMARDs.