Data sources
The HRS is a nationally-representative longitudinal panel study of US residents 50 years of age and older [34, 35]. Approximately 20,000 participants are surveyed every 2 years. New cohorts are added to the study every 6 years, and participants are followed from entry until voluntary withdrawal or death [34]. The present study sample included respondents surveyed in the 2016 wave of the HRS, the most recent survey year that provides linkable Medicare data. The HRS and the current study including access to sensitive Medicare files was approved by the University of Michigan Health Sciences/Behavioral Sciences Institutional Review Board (HUM00061128, HUM00152177). Informed consent was obtained from all study subjects. No study subjects were under the age of 18. All methods were performed within the relevant confidentiality guidelines and regulations of the Institutional Review Board and the Health and Retirement Study.
The HRS includes information from Medicare-covered health services events for the 78–84% of respondents who authorize linkage across survey years [34]. Medicare billing claims record the reason for a healthcare provider visit listed as International Classification of Diseases, 9th edition, Clinical Modification or 10th edition (ICD-9-CM & ICD-10) codes, Health Care Common Procedure Coding System (HCPCS), and Current Procedural Terminology (CPT-4) codes. To identify HRS respondents with RA, we linked fee-for-service (FFS) Medicare Part A inpatient, outpatient, skilled nursing facility, home health files, and Part B carrier files in the 2 years (2014–2015) preceding the 2016 survey wave. Part C claims, also called Medicare Advantage or Medicare + Choice, are not available for HRS linkage. We addressed differences in respondents’ Medicare enrollment through the exclusionary criteria discussed below.
Sample eligibility
From the initial 20,890 HRS respondents in 2016, we excluded those who were Medicare-ineligible or who did not consent to Medicare linkage (n = 12,046 excluded). To avoid missing data biases affecting the availability of RA diagnoses, we further excluded those with Medicare linkage who did not have full FFS parts A & B coverage from 2014 to 2015, defined as 11 months or more per year (n = 4382 excluded). This last step excluded anyone with Part C claims greater than 1 month per year, leaving a total of 4462 respondents in our final sample. We did not exclude respondents between the ages of 50–64 who had Medicare benefits due to disability, end-stage-renal disease (ESRD), or amyotrophic lateral sclerosis (ALS).
Identifying RA
The validity of identifying RA via ICD code-based algorithms varies by the population under consideration and the methods used. A systematic review of ICD-9 code-based algorithms for the detection of RA in administrative databases found that the highest positive predictive values (PPV; the proportion of true positives out of all algorithm-identified positives) come from algorithms that include a minimum of two RA diagnostic codes and additional information related to whether a rheumatologist made the claim or if the RA patient received DMARDs, the most common class of medications for people with RA [36]. However, additional requirements for RA classification come with generalization limitations and tradeoffs. For instance, over the course of 2 years, approximately 34% of people with RA see a rheumatologist at least once [37]. Further, estimates from two population based studies showed that less than half of people with RA had associated DMARDs prescriptions [32, 37]. Though additional requirements of having a rheumatologist make the claim, or including DMARD prescriptions increases the PPV of an algorithm, these requirements also identify specific and narrow RA populations that are unlikely to represent all adults with RA. Therefore, we conducted our analysis using an algorithm that is likely to include the most people with RA, and then conducted two sensitivity analyses with increasing PPV’s but higher restrictions, discussed in more detail below.
We identified cases of RA via participants’ Medicare claims by requiring a minimum of two billing diagnoses of ICD-9CM codes 714* or ICD-10 codes M05*or M06*, between study years 2014–2015. We included any code listed either as the principal diagnosis or in one of the 25 primary/secondary diagnostic fields from Medicare Part A files or the 12 fields from routine clinical visits in Part B carrier files. Claims had to be more than 1 day apart. We excluded claims from non-licensed health care providers, such as durable medical equipment providers and ambulance services.
For the sensitivity analyses, we applied identical methods described above using 1). an algorithm requiring, in addition to two RA claims, a minimum of one from a rheumatologist, and 2) a different algorithm requiring two RA codes from any provider, and one DMARD prescription. We identified rheumatology clinic-based claims using CMS provider specialty code “66” listed in at least one of the 13 specialty billing fields in the Part B carrier files. We identified DMARDs using generic names (see appendix) from Medicare part D summary files in 2014–2015.
Measurement of cognitive impairment
HRS respondents’ cognitive status was measured with the Langa-Weir Classification [38]. The Langa-Weir measure provides a 27-point scale of cognition for self-respondents (the modified Telephone Interview for Cognitive Status; or TICS-m) and an 11 point scale for proxy respondents, representing cognition at the time of the 2016 HRS interview [38, 39]. The use of proxy respondents in the HRS allows people who are either physically or cognitively incapable of completing the survey to participate, which ensures adequate representation of the older adult population and reduces bias related to study attrition from low levels of cognitive ability [39, 40]. In 2016, proxies represented 4.5% of all HRS respondents and 3% of those with CIND or dementia.
The Langa-Weir classification assesses cognitive function for self-respondents in the HRS using an adapted version of the Telephone Interview for Cognitive Status (TICS). The adapted TICS consists of immediate and delayed 10-noun free recall (respondents immediately recall a list of 10 words, then remember the list after a delay) and serial 7’s subtraction tests (respondents subtract seven from 100, then continue to remove 7 five more times) to assess memory, and backward count to evaluate attention and processing speed [38, 39, 41]. Proxy measures of CI include caregivers’ assessment of the person’s cognition in the areas of memory (excellent, very good, good, fair, or poor) and instrumental activities of daily living limitations (IADL, scored 0–5). The proxy measure also includes the trained interviewer’s overall estimation of risk for CI (No CI, may have CI, has CI) [38].
The Langa Weir classification yields three groups: normal cognition, “cognitive impairment non-dementia” (CIND), or dementia [42]. Cut points in the Langa-Weir classification produced the same population distribution of cognitive ability as estimated in the Aging, Demographics, and Memory Study (ADAMS), a subsample of HRS respondents who underwent extensive neuropsychological testing and clinical assessment [38, 43]. Using the 27-point self-report scale, cut points for the three categories are normal cognition (12–27), CIND (7–11), and dementia (0–6) [38]. For the 11-point proxy scale, cut points are normal (0–2), CIND (3–5), and dementia (6–11). Both the self-report and proxy scores use imputed values for missing data [38]. Documentation of imputation methods for all cognitive measures is available at the HRS website [44]. Due to low counts of dementia in the RA group, we collapsed the Langa-Weir classification into a binary yes-or-no variable with normal cognition versus CIND or dementia.
Covariates
Additional variables include respondent age at the time of the survey, educational attainment (less than high school, high school graduate/GED, any college or more), gender (male/female), and race (White/Non-White).
Statistical methods
We calculated differences in the sociodemographic characteristics of subgroups with and without RA and the proportion of respondents in each group with CI, defined as CIND or dementia per the Langa-Weir classification. We then examined the relative odds of cognitive impairment among people with RA versus no RA using unadjusted and adjusted logistic regression using RA as the predictor of interest and CI as the outcome. The adjusted model included age (centered at its mean), educational attainment, gender, and race as controls for confounding.
The HRS uses a national probability sample and provides the appropriate weights for complex survey design analysis and national estimates. Our criteria requiring 2 years of complete FSS linked Medicare parts A & B claims reduced the original HRS sample by 79%. This reduction in sample size resulted in having no population members from 24 out of the 80 strata from which HRS samples are drawn (strata in the HRS are non-overlapping Metropolitan Statistical areas, single counties, or groups of small counties used to stratify the population). Because our reduced sample was not nationally representative, and we could not determine if our sample’s weighting reflected the original study’s probability distribution, we did not employ survey design weighting in our analysis. We performed all analyses using STATA 16.1 MP (College Station, TX).