Study Design, Setting, and Data Sources
We analyzed the 2014 State Inpatient Database (SID) and State Emergency Department Databases (SEDD) of four states: Florida, Maryland, Massachusetts, and New York. These data were collected at each state in the Healthcare Cost and Utilization Project of the Agency for Healthcare Research and Quality (AHRQ).20, 21 The SID includes all the inpatient discharge records from community hospitals (short-term hospitals accessible by the general public, including academic medical centers), regardless of the source of hospital admission; and the SEDD includes all the discharge records on ED visits at hospital-affiliated EDs that did not result in hospital admission. Taken together, these data include all patients who were admitted to a hospital or presented to ED in a given state. We selected these four states because of their socio-demographic diversity of the population and the availability of the homeless indicator (only seven states [four states included in our analysis plus Georgia, Utah, and Wisconsin] included both the homeless indicator and a unique patient linkage number for SID/SEDD. Three states were not included in our analyses because our internal investigation identified an underreporting of the homeless indicator for Utah and Wisconsin, and the hospital identifier was not available for Georgia).
We identified all adults aged 18 years or older who were admitted to the acute care hospitals in these four states in 2014 and then examined their rates of readmissions and ED visits after discharge. Discharges occurring in December 2014 for MA and NY, and the last quarter for MD and FL, were excluded to ensure a full 30 days of follow-up (as only information on discharge quarter was available for MD and FL [and the data on discharge month were unavailable]). From the 4,519,374 hospitalizations in our initial sample, we excluded 193,877 hospitalizations (4.3%) of patients who died in the hospital or were discharged against medical advice; 450,394 hospitalizations (10.0%) with the primary discharge diagnoses related to delivery (Clinical Classification Software [CCS] single-level codes: 177–192, 194–196, 218–220, or 222–224); and 347,720 (7.7%) hospitalizations with missing data on key variables. Our final analytic sample consisted of 3,527,383 hospital discharges.
To obtain information on the hospital characteristics, we linked the SID/SEDD with the 2016 American Hospital Association Annual Survey database 22 to identify information on several hospital characteristics: profit status, Rural-Urban Commuting Area (RUCA) classification, teaching status, hospital size, the presence of medical intensive care unit (MICU) or cardiac intensive care unit (CICU), and the proportion of hospitalized patients with Medicaid or Medicare.
The primary exposure variables of interest were (1) homeless status and (2) the site of care. The SID/SEDD provides the indicator variable for the homeless status that was directly reported by hospitals, as has been used in reports by the AHRQ 23, 24 and in prior research.5 In this study, patients recorded as homeless any time in 2014 were defined as homeless (period prevalence counts rather than point-in-time counts) to account for the dynamic status of the homelessness and to focus on the groups whose access to secured housing was not guaranteed.4
As for the site of care, we calculated the proportion of homeless patients for each hospital and categorized hospitals in the highest decile of the proportion of homeless patients as homeless-serving hospitals (HSH). The remaining 90% of hospitals were categorized as non-HSH. A similar approach has been used to identify minority-serving hospitals in previous studies.25, 26
Our outcome variables of interest were (1) 30-day all-cause readmission captured according to the methodology recommended by the HCUP27 and (2) 30-day all-cause ED visit after hospital discharge.
We adjusted for the potential confounders: patient characteristics, discharge disposition, length of stay (LOS), and hospital characteristics. Patient characteristics included the primary diagnosis for the admission (an indicator variable for Medicare Severity Diagnosis Related Group [MS-DRG] codes), age at the point of admission (18–29, 30–39, 40–49, 50–59, 60–69, 70–79, and 80 years or older), sex, race/ethnicity (Non-Hispanic White, Non-Hispanic Black, Hispanic, and others), primary payers (Medicare, Medicaid, private insurance, self-pay, and others), and the indicator variables for 29 comorbidities included in the Elixhauser comorbidity index.28
Discharge disposition was categorized as routine discharge, transfer to skilled nursing facility/intermediate care facility, home health care, and others. LOS was used as a continuous variable. Hospital characteristics included profit status (for-profit, non-profit, and public), RUCA (urban, suburban, large rural, and small rural), teaching status (major, minor, and non-teaching), hospital size (large [400+ beds], medium [100–399 beds], and small [1–99 beds]), and the presence of MICU/CICU.
First, we compared the characteristics of homeless patients vs. non-homeless patients and hospital characteristics of HSH and non-HSH.
Second, we examined how the rates of readmission and ED visit differ by the homeless status (homeless vs. non-homeless) of patients and by the site of care (HSH vs. non-HSH). We constructed two regression models to control for potential confounders. Model 1 was adjusted for patient characteristics, and quarter and state indicator variables (i.e., dummy variables for quarter and state, effectively comparing patients within the same quarter and state). Model 2 was adjusted for all the variables included in Model 1 plus hospital characteristics, discharge location, and LOS, to investigate if these factors explain the observed differences in patient outcomes. We used multivariable logistic regression models, with standard errors clustered at hospital level to account for a potential correlation of patients treated at the same hospital. To calculate risk-adjusted rates of 30-day readmission and ED visit, we used marginal standardization (i.e., predictive margins or margins of responses). For each hospital discharge, we calculated predicted probabilities of patient readmission and ED visit with either homeless status or homeless-serving status fixed at each category and then averaged over the distribution of covariates in our sample.29
Finally, using the same regression models, we examined whether patient outcomes between the homeless and non-homeless patients vary by the site of care, by including the interaction term between the homeless status and HSH in our regression models. We used a Wald test adjusted for clustering (to approximate a likelihood ratio test because standard likelihood-based tests are not available with the clustered data) to formally test the interaction between the homeless status and HSH.30 For each of four patient groups (homeless patients at HSH, non-homeless patients at HSH, homeless patients at non-HSH, and non-homeless patients at non-HSH), we calculated the risk-adjusted rates of 30-day readmission and ED visit using marginal standardization. To account for multiple comparisons (six pair-wise comparisons between four groups), we considered two-tailed p value < 0.0083 to be statistically significant.
We conducted a series of sensitivity analyses. First, to examine whether the observed relationships depend on our threshold for defining HSH, we used an alternative definition, the highest quintile, to define HSH. Second, we tested whether our findings vary based on the primary diagnosis of hospital admissions. We selected four major conditions in our data and also used as target conditions in the HRRP 13: acute myocardial infarction (AMI) (CCS single-level code: 100), pneumonia (108), chronic obstructive pulmonary disease (COPD) (122), and heart failure (HF) (127). We used the same regression models except for including the primary diagnosis as an adjustment variable.
All analyses were conducted using Stata version 15 (College Station, TX; StataCorp LLC.).31 The study was approved by the UCLA Institutional Review Board.