Database and Study Population
We examined acute care utilization for adults (age > 18 years) using Healthcare Cost and Utilization Project (HCUP) 2012-2015 (quarter 3) discharge data from Florida and New York from the State Inpatient Database (SID), State Emergency Department Database (SEDD), and State Ambulatory Surgery and Services Database (SASD) from the Agency for Healthcare Research and Quality (AHRQ). Florida and New York comprise 12% of the US population and their discharge data had encrypted patient identification numbers that permitted linkage across facilities and hospitals. The years 2012-2015 were selected because they occurred after the Hospital Readmission Reduction Program took effect and before ICD-10 changes in coding were implemented. The SID, SEDD, and SASD are all-payer databases, containing discharges from nonfederal, non-psychiatric community hospitals and emergency departments. They contain more than 100 clinical and nonclinical variables including principal and secondary diagnoses and procedures, admission and discharge status, patient demographics, and length of stay (LOS) (24, 25).
We combined patient-level data from the 2012-2015 SID, SEDD, and SASD with hospital-level data from the 2015 American Hospital Association (AHA) Annual Survey, community-level data from the 2015 Area Health Resources Files (AHRF) and with ZIP-level data for 2015 from the University of Wisconsin Neighborhood Atlas® Area Deprivation Index (ADI) (26).
We included all non-psychiatric medical-surgical admissions. Index admissions for patients aged 18 years and older were considered if they occurred at nonfederal general medical-surgical hospitals. Using AHRQ Clinical Classification Software (CCS) codes, all discharge records for patients with comorbid diagnoses of SMI were identified (Supplementary File 1). Patients were excluded from the cohort if they: 1) did not survive to discharge; 2) had an admission with a primary diagnosis of SMI; 3) had an index length of stay of less than or equal to one day; 4) discharged against medical advice; and 5) were admitted for a diagnosis category or a procedure considered planned (i.e. primary diagnosis of cancer or procedures including obstetrical delivery, transplant surgery, maintenance chemotherapy, rehabilitation; ICD9 and AHRQ CCS categories used to identify planned admissions in the Supplementary File 1) (27).
Community Socioeconomic Disadvantage
2015 data from the University of Wisconsin Neighborhood Atlas® ADI (26) were used to characterize patients’ community socioeconomic disadvantage. The ADI allows for rankings of neighborhoods by socioeconomic status disadvantage in a region of interest (e.g. at the state or national level). It uses American Community Survey (ACS) Five Year Estimates in its construction. The 2015 ADI used the ACS data for 2015, which is a 5-year average of ACS data obtained from 2011-2015. It includes factors from the theoretical domains of income, education, employment, and housing quality. In the present study, ADIs were aggregated to the ZIP-level for communities in Florida and New York. Communities were compared independently for Florida and New York—the 50% least disadvantaged communities were grouped together; the 45% middle disadvantaged communities were grouped together; and the 5% most disadvantaged communities were grouped together.
The revisit was defined as the first revisit for a physical health condition within 30 days of discharge. For the purpose of this study, revisit types include ED visits, observational stays and readmissions. The primary outcomes of interest were three types of medical-surgical revisit: 1) having an ED visit within 30-days of discharge; 2) having an observation stay within 30 days of discharge; and 3) having a readmission within 30 days of discharge. An additional outcome included having any type of revisit within 30 days of discharge (i.e. ED visit or observation stay or readmission). ED visits, observation stays, and readmissions were identified using the HCUP supplemental variables for revisit analysis, which provide a unique visit link to allow for each patient to be tracked at subsequent inpatient visits across time and institutions (28).
Individual patient-level demographic (i.e. age, sex, primary payer [Medicare, Medicaid, private, self-pay, no charge, or other]) and clinical characteristics (length of stay of the index admission, admission type [emergency, urgent, elective, or trauma center] an indicator for if they had a surgical procedure, Elixhauser comorbidity readmission risk score, and DRG of the index admission), hospital-level characteristics (teaching status of the hospital [member of Council of Accredited Teaching Hospitals], total number of hospital beds, technology status of the hospital [i.e. capable of performing heart transplant or adult interventional cardiac catheterization], the hospital’s nurse-to-bed ratio, and the ownership status of the hospital [i.e. non-federal government, private for profit, or private not-for-profit]), and community characteristics (rurality, health care provider supply [ratio of primary care physicians to county population and ratio of nurse practitioners to county population]) were compared by community socioeconomic disadvantage category (least 50% disadvantaged; middle 45% disadvantaged; and 5% most disadvantaged) using chi-squared tests for categorical variables and analysis of variance for continuous variables.
Multivariate logistic regression models with state and year fixed effects were used to examine the relationship between a patient’s community socioeconomic disadvantage category and revisits (any revisit and by revisit type) after adjusting for patient, hospital, and community-level factors. Among patients who experienced a revisit, a second set of logistic regression models were used to examine the relationship between community socioeconomic disadvantage and revisit type after adjusting for patient, hospital, and community-level factors. Post-estimation commands were used to estimate the adjusted revisit rates of each type of revisit by community socioeconomic disadvantage. Statistical analyses were performed using STATA statistical software, version 16.1 (StataCorp, College Station, TX). For all analyses, p values of <0.05 were considered statistically significant.