Data and study population
We used data from the 2012 AHRQ Healthcare Cost and Utilization Project (HCUP) State Inpatient Databases (SID) to carry out a cross-sectional analysis. In the SID, we identified 1.43 million discharge records, representing 3.58 million (weighted) potentially preventable hospitalizations. We selected a sub-sample of the SID, the SID disparities analysis file, which is used to compute national estimates for the National Healthcare Disparities Report. It consists of weighted records from a sample of hospitals from 38 States participating in HCUP that have high-quality race/ethnicity data in 2012: AK, AR, AZ, CA, CO, CT, FL, GA, HI, IA, IL, IN, KS, KY, MA, MD, MI, MO, NC, NJ, NM, NV, NY, OH, OK, OR, PA, RI, SC, SD, TN, TX, UT, VA, VT, WA, WI, and WY. The SID disparities analysis file contains records representing 91 % of all U.S. hospital discharges. Nationally-representative statistics were computed using discharge weights that were constructed with consideration of different attributes of the hospital, including the number of beds, geographic region, number of discharges and teaching status. We selected records for U.S. adults aged 18 years and older discharged from U.S. community, non-rehabilitation hospitals with a primary diagnosis of an ambulatory care sensitive condition contained within AHRQ PQIs. All investigators completed training and signed a data use agreement for the HCUP SID (https://www.hcup-us.ahrq.gov/tech_assist/dua.jsp). Because data for our analyses did not involve human subjects, IRB approval was not required. HCUP SID data are available via the HCUP Central Distributor (https://www.hcup-us.ahrq.gov/tech_assist/centdist.jsp).
We categorized hospitalizations into subsets based on three PQIs focused on broad composites of potentially preventable hospitalizations: the acute composite (PQI 91), the chronic composite (PQI 92), and the overall composite (PQI 90). The acute composite (PQI 91) included hospitalizations for dehydration, bacterial pneumonia, and urinary tract infection (n = 559,515). The chronic composite (PQI 92) included hospitalizations for diabetes (short-term complications, long-term complications, uncontrolled diabetes, and lower-extremity amputation for diabetes), chronic obstructive pulmonary disease (COPD) or asthma in older adults, hypertension, congestive heart failure, angina without a procedure, and asthma in younger adults (n = 866,668). The Overall Composite (PQI 90) was the union of the three PQI 91 conditions and the nine PQI 92 conditions (n = 1,426,153). Discharge records were scored using version 4.4 of the AHRQ Quality Indicator SAS software to identify hospital stays that fit the definitions of the PQI composites [17].
Primary independent variable
Our primary independent variable was the number of MCCs present in primary and secondary diagnosis codes. We created categories of MCCs grouped as 0–1 condition, 2–3 conditions, 4–5 conditions, and 6+ conditions from the list of conditions defined by the HHS Strategic Framework [7, 18]. To be included in the chronic PQI sample, patients had to have at least 1 chronic condition; therefore the “0–1 condition” reference category for these patients included those with exactly 1 chronic condition, whereas patients with an acute PQI were placed in the “0–1 condition” category if they had either 0 or 1 concurrent chronic condition. The 20 chronic conditions included were arthritis, asthma, autism spectrum disorder, coronary artery disease (CAD), cancer, cardiac arrhythmias, chronic kidney disease (CKD), congestive heart failure (CHF), COPD, dementia, depression, diabetes, hepatitis, HIV, hyperlipidemia, hypertension, osteoporosis, schizophrenia, stroke, and substance use disorders.
Covariates of interest
We selected patient population characteristics including sex, age group (18–39 years, 40–64 years, 65+ years), and race/ethnicity (White non-Hispanic, Black non-Hispanic, Hispanic, Asian Pacific Islander non-Hispanic, and Other). Race/ethnicity measures may be problematic in hospital discharge databases because some states do not collect information on race and ethnicity from hospitals and, within states that collect the information, some hospitals do not code race and ethnicity reliably. To deal with this problem, we used the SID Disparities Analysis File for 2012 designed for the AHRQ National Healthcare Disparities Report to provide national accurate estimates of race and ethnicity. A measure for race/ethnicity in the SID is created using a stratified, weighted sample of hospitals with good reporting of patient race and ethnicity from 38 SID states: Alaska, Arizona, Arkansas, California, Colorado, Connecticut, Florida, Georgia, Hawaii, Illinois, Indiana, Iowa, Kansas, Kentucky, Maryland, Massachusetts, Michigan, Missouri, Nevada, New Jersey, New Mexico, New York, North Carolina, Ohio, Oklahoma, Oregon, Pennsylvania, Rhode Island, South Carolina, South Dakota, Tennessee, Texas, Utah, Vermont, Virginia, Washington, Wisconsin, and Wyoming. It is drawn as a hospital sample of 40 % of community, nonrehabilitation hospitals in the United States (about 2000 hospitals) [19].
Additional patient circumstances explored in descriptive analyses were patient location (largest locales [metropolitan and micropolitan] and smallest locales [nonmetropolitan and nonmicropolitan]) and primary expected source of payment (private insurance, Medicare, Medicaid, other insurance, uninsured [includes uninsured, self-pay, no charge, and other])
Analytic methods
All analyses used discharge-level data rather than hospital-level aggregates. Descriptive statistics were computed by tabulating the number and percentage of discharges observed in categories of MCCs and in categories of demographic characteristics of the sample for each set of ambulatory care sensitive conditions (i.e., within each PQI composite). We then developed multivariate models to examine associations between categories of MCCs and the three outcome variables: 1) inpatient costs per stay (dollars), 2) inpatient costs per day (dollars per day), and 3) length of inpatient hospitalization (days). As our data on costs per stay and per day was over-dispersed, we used a negative binomial 2 regression model to accommodate this skewed cost data. We used generalized linear models to model length of stay (LOS). Hospital costs were derived by converting reported charges data to estimated costs using hospital-specific cost-to-charge ratios contained in the 2012 HCUP SID Cost-to-Charge Ratio files (https://www.hcup-us.ahrq.gov/db/state/costtocharge.jsp#user).
All models included the primary independent variable, number of MCCs. Zero or 1 chronic condition was the reference category in multivariate analyses. We also controlled for the population characteristics using males, 65+ years of age, and White race/ethnicity as the reference categories.
Because inpatient mortality could confound the relationship between cost and MCCs, we adjusted models to assess whether findings were affected by discharge disposition. We used two indicators. The first indicator—dead or alive—was tested in the inpatient hospital cost per stay regression model. The second indicator used eight discharge dispositions—home or self-care, transfer to short-term hospital, transfer to other facility, home health care, against medical advice, died in hospital, discharged alive, and unknown or missing—to examine other potential reasons for the longer LOS for patients with MCCs.