We used data from the National Health and Aging Trends Study (NHATS), a nationally representative survey of Medicare beneficiaries aged 65 and older that began in 2011.21 NHATS conducts annual in-person interviews with individuals or proxy respondents. Information is collected on demographics, living arrangements, health conditions, functional status, healthcare use, and economic status. The survey oversamples adults 90 years and older and Black non-Hispanic individuals. We included data from NHATS rounds one through six, collected from 2011 to 2016, the most recent years for which subsequent linked claims data (from 2011 to 2017) were available. Our sample was limited to individuals 70 and older as per NHATS technical guidelines for analyzing repeated survey rounds because, as the sample ages, the representativeness of people aged 65 to 69 decreases.22 We excluded those residing in nursing homes or with fewer than 12 months of Fee-for-Service (FFS) Medicare coverage pre-NHATS survey. The Johns Hopkins University Institutional Review Board approved the NHATS protocol. The Icahn School of Medicine at Mount Sinai’s Institutional Review Board and the Centers for Medicare & Medicaid Services (CMS) Privacy Board approved the study.
Homebound status was the primary exposure of interest for this study. We defined homebound status in accordance with previously published constructs based on responses to NHATS questions.2 Participants were asked “How often did you go out in the last month?” Participants who responded that they never or rarely (no more than 1 day/week) went out were considered homebound for that year.
We included demographic, clinical, and geographic measurements in our analyses to characterize the homebound population and to adjust for factors that confound the association between homebound status and healthcare utilization and spending. These included age, sex, race/ethnicity (Black non-Hispanic, Hispanic, White non-Hispanic, and other race [American Indian/Asian/Native Hawaiian]), education level, marital status, income quartile, functional status as defined by receiving help with activities of daily living (ADL), living alone, living in an assisted living facility, speaking a language other than English, Medicaid status, self-reported general health, and Charlson Comorbidity Index.23 In addition, we included measures of sensory loss. Following Simning et al., we defined auditory impairment as self-reported inability to “hear well enough to carry on a conversation in a room with a radio or TV playing,” even with a hearing aid.24 We similarly defined visual impairment as self-reported inability to “see well enough to read newspaper print,” even with glasses or contacts. We then categorized observations as having either auditory impairment, visual impairment, dual sensory impairment, or no auditory or visual impairment.
Presence of depressive symptoms was classified based on Patient Health Questionnaire (PHQ-2) score greater than 3, and anxiety was classified based on the Generalized Anxiety Disorder (GAD-2) score greater than 3.21,22 We defined probable dementia using criteria established by NHATS which incorporate self-report of dementia, proxy responses to the Alzheimer’s disease (AD)-8 screening tool, and a cognitive interview that assessed memory, orientation, and function both by self-report and direct cognitive assessment conducted by NHATS.25 Observations were classified into metropolitan or non-metropolitan area (per Rural-Urban Continuum Code classification) based on the county in which the respondent resided at the time of interview.
We obtained information on healthcare utilization and costs from linked Medicare FFS claims data. Utilization outcomes included rates of inpatient admissions, emergency department (ED) visits, skilled nursing facility (SNF) visits, home health visits, hospice visits, primary care visits, and specialist visits in the year following the interview. Primary care and specialist visits were identified using Healthcare Common Procedure Coding System and Provider Specialty codes (Appendix Table 1).26 Furthermore, in order to identify potentially preventable hospitalizations, we used ICD-9 and ICD-10 codes for a list of fourteen Ambulatory Care Sensitive Conditions as defined by the Agency for Healthcare Research and Quality, such as hypertension, dehydration, and diabetes complications (Appendix Table 2).27
We examined spending both in total per observation per year and by claim type. These claim types included inpatient, carrier (professional provider), SNF, home health, outpatient, hospice, and durable medical equipment. We obtained information on Medicare reimbursements by hospital referral region from the Dartmouth Atlas.28 All dollar amounts were inflation adjusted to 2017 dollars using the CPI-U Index.
Our unit of analysis was person-year and individuals were allowed to have repeat observations. This framework allowed us to increase our number of observations and account for the fact that an individual’s homebound status can change from year to year.29
We examined demographic and clinical characteristics of our observations by homebound status using bivariate linear or logistic regression. We used logistic and zero-inflated negative binomial regressions to compare adjusted and unadjusted differences in utilization. We used a generalized linear model with a log distribution to compare the adjusted and unadjusted differences in overall, carrier, and outpatient spending, as well as two-part models for other expenditures with frequent zeros. In our adjusted models, we included sex, race, age, education, marital status, geographic region, metropolitan area, functional status, Medicaid enrollment, probable dementia, and Charlson Comorbidity Index. These covariates were selected based on our conceptual model of the determinants of homebound status and their potential relationship to healthcare utilization and spending.10 Our adjusted spending models also included quintile of Medicare reimbursements by hospital referral region.
We examined the sensitivity of our findings to high end-of-life spending by excluding from our sample those who died within 12 months of their NHATS interview. We explored the sensitivity of our findings to place of residence by excluding those residing in assisted living facilities. We also investigated the sensitivity of our findings to using person-years as our unit of observation by limiting the sample to one survey year (2015), using the survey year with the largest sample size. We also used 2015 data to estimate the proportion of total annual Medicare FFS spending attributable to the homebound. Finally, because those who are homebound over longer periods of time may differ from those who may only be homebound temporarily, we also compared healthcare utilization and spending among individuals who were persistently homebound (i.e. had been homebound in 2015 and remained so in 2016) to those who were not homebound.
All analyses adjusted for NHATS analytic weights that consider survey design and differential probabilities of selection and non-response. All person-year analyses take into account clustering at the respondent level to account for repeat observations per respondent.22 All analyses were conducted using Stata17.
Role of the Funding Source
The National Institute on Aging played no role in the design, conduct, and analysis of the study or in the decision to submit the manuscript for publication.