Our study sample draws from the National Health and Aging Trends Study (NHATS, www.nhats.org) [16], a nationally representative study of US Medicare beneficiaries ages 65 years and older, with NHATS data linked to Medicare claims [17]. The sampling probabilities of the original NHATS cohort were designed to yield equal probability samples by age group and race/ethnicity. This linked data set was available for approved use (Johns Hopkins Medicine IRB00077995). Our analytic sample included 2,549 community-dwelling older adults from the first wave of NHATS in 2011 (i.e., baseline visit), who had frailty and cognition measured at baseline, and were continuously enrolled in fee-for-service Medicare for at least 1 year prior and 2 years after baseline visit. Given that Medicare enrollees could enter and exit the program at any time for different reasons, the two-year study follow-up was chosen to maximize sample size and data capture of hospitalization events while also considering the availability of the Medicare claims data at the time of the analysis. Supplemental Fig. 1 displays exclusions: non-continuous Medicare coverage over this 3-year period; any hospitalization 1 year prior to baseline, which may exacerbate risk for further hospitalizations; and history of stroke or depression, due to possible overlapping symptoms with frailty or cognitive impairment [18].
Study variables included the physical frailty phenotype, as measured in NHATS, with 5 main criteria: exhaustion, low physical activity, slowness, weakness, and weight loss [19]. In brief, participants met criteria as follows: exhaustion, by self-reporting low energy or ease of exhaustion for limiting activities; low activity, by self-reporting never walking for exercise or engaging in vigorous activities; slowness, by walking speed (first of 2 tests of 3 m walk) at or below the 20th percentile of the population distribution by sex and height categories; weakness, by grip strength (maximal value of two tests using a handheld dynamometer) at or below the 20th percentile of the population distribution by sex and body mass index (BMI) categories; and weight loss, by BMI < 18.5 kg/m2, or reported unintentional loss of ≥ 10 pounds in the last year. Those meeting 3 + criteria were classified as frail; 0–2 as not-frail.
Cognitive impairment was determined by meeting at least one of three criteria [7]: 1) scored in bottom quintile in either executive function (clock-drawing test) or memory (10-item immediate and delayed recall batteries); 2) self- or proxy report of doctor’s diagnosis of dementia or Alzheimer’s disease (AD); or 3) scored of 2 or higher on the AD8 Dementia Screening Interview [20]. Probable dementia was defined as having met at least one of three criteria: [21] (1) self-respondents’ test performance less than 1.5 SD below the mean in at least two of three domains: memory, orientation, and executive functioning; (2) self- or proxy-report of dementia or AD by doctor’s diagnosis; (3) a score of 2 or higher on the AD8 administered to proxy respondents [22]. The first item in the probable dementia definition is distinct from our cognitive impairment definition; the other two items are the same. CI intends to capture a broader spectrum of cognitive impairment that includes probable dementia.
Covariates included age (years, in 5-year increments); race/ethnicity: (white non-Hispanic, Black non-Hispanic, Hispanic, or other); education: (eighth grade or less, ninth to twelfth grade (no diploma), or high school graduate or higher); and total personal annual income (income quartiles, < $15,000, $15,000–$30,000, $30,000–$60,000, or > $60,000 USD per year).
Health characteristics included body mass index (BMI), categorized as underweight (< 18.5 kg/m2), normal (≥ 18.5 and < 25 kg/m2), overweight (≥ 25 and < 30 kg/m2), or obese (≥ 30 kg/m2). Comorbidities included history of heart disease, hypertension, arthritis, osteoporosis, diabetes mellitus type II, lung disease, cancer, and hip fracture; and comorbidity burden by the number of chronic diseases (0,1,2,3,4 +). Activities of daily living (ADLs) (using the toilet, getting cleaned up, dressing, and eating) and Mobility Disability (getting out of bed, going outside, and getting around inside) were scored using an ordinal scale (“fully able,” “modification,” “difficulty,” or “assistance”), with dependency defined as requiring “assistance” [23, 24].
Outcomes included all-source hospitalizations; ED-admission hospitalizations, where an ED visit directly preceded hospitalization (typically unplanned, acute hospitalizations); and direct admission hospitalizations, where hospitalization occurred without a preceding ED visit (typically planned procedures). ED-admission hospitalizations were defined as inpatient claims with revenue codes 0450, 0451, 0452, 0456, 0459 and 0981. Centers for Medicare & Medicaid Services (CMS) data identifying hospitalizations after ED visits typically excludes ED visits within 24 h of another ED visit, to avoid double-counting via clerical errors, and considers a hospitalization to have resulted from an ED visit if it occurred on the same day or the next day compared to the ED visit, to allow for ED visits that crossed midnight before the admissions decision was made. Inpatient claims without revenue codes 0450, 0451, 0452, 0456, 0459 and 0981 were defined as direct admission hospitalizations. We defined a recurrent event as one or more hospitalizations of the same study subject within the two years following baseline assessment.
Statistical analyses: Descriptive statistics on sociodemographic and health characteristics were reported by analytic groups: 1) Unimpaired: those with no frailty or cognitive impairment; 2) Cognitive Impairment (CI) only: those with CI but not frailty; 3) Frailty only: those who were frail but not cognitively impaired; 4) CI + Frailty: participants who were both cognitively impaired and frail. In addition, we also compared the baseline characteristics of study subjects who had at least one hospitalization during the two-year follow-up by type of their first hospital admission (i.e., direct-admission vs. Emergency Department (ED)-admission). Frequency percentages were used to summarize categorical variables. Chi-square and Kruskal–Wallis tests were used to assess the difference in sociodemographic and health factors among CI/frailty groups. Our analyses focused on the association of baseline CI and physical frailty, separately and jointly, with the recurrence of all-source vs. ED-admission vs. directly admitted hospitalizations over 2 years analyzed in separate models using linked NHATS and Medicare claims data. We used a recurrent events model, the marginal means/rates model, where all-source, ED and direct admission hospitalization were treated as recurrence event outcomes with effect size reported as a rate ratio (RR). This approach considers all hospitalizations of the same subject as a single counting process and corrects for dependency among recurrent event times within a subject without the need to parameterize the dependence structure, therefore making it particularly appealing to applications where the dependence structure is complex and unknown, or the nature of the dependence is not of primary interest [25]. We first performed an unadjusted recurrent events model to examine the association between CI/Frailty groups and the risk of recurrent hospitalizations (analytic model 1). Secondly, we performed an adjusted model adding covariates: age (continuous), gender, race/ethnicity, education, income, and obesity (analytic model 2). Lastly, additional covariates of comorbidities (continuous) and dependency were included to assess the associations independent of multimorbidity and disability (analytic model 3). To examine the reasons for direct-admission vs. ED-admission hospitalizations, we analyzed the primary/principal diagnosis code established to be chiefly responsible for ED- or direct-admission hospitalizations. Diagnoses were classified by body system or condition using chapters from the International Classification of Diseases, ninth Revision, Clinical Modification (http://www.icd9data.com/2015/Volume1/default.htm). Percentage distribution of primary diagnosis code categories with the highest proportions (≥ 20%) of use among individuals who had one or more hospitalizations was tabulated by hospital admission type and by Frailty/CI group membership. Statistical analyses were performed using SAS (v.9.4; SAS Institute Inc, Cary, North Carolina) and Stata (v.15; StatCorp LLC, College Station, Texas). A p-value < 0.05 was considered statistically significant.
