Patients
Patients discharged between 1 January 2012 and 30 June 2013 by the hospitalist groups were identified by searching AAMC’s Crimson Continuuum of Care (The Advisory Board, Washington, DC), a software analytic tool that is integrated with coded clinical data. Adult patient hospitalizations determined by Crimson to have a medical (non-surgical, non-obstetrical) APR-DRG code as the final principal diagnosis were included. Critically ill patients or those appropriate for “step-down unit” care were cared for by the in-house critical care staff; upon transfer out of critical or step-down care, patients were referred back to the admitting hospitalist team. A diagnosis (and its associated hospitalizations) was excluded for referral bias if the diagnosis was the principal diagnosis for at least 1% of a group’s discharges and the percentage of patients with that diagnosis was at least two times greater in one group than the other. Hospitalizations with a diagnosis of “ungroupable” (APR-DRG 956) were also excluded.
Measurements
Demographic, insurance status, cost of care, length of stay (LOS), APR-DRG (All Patient Refined Diagnosis-Related Group) severity of illness (SOI) and risk of mortality (ROM), consultant utilization, 30-day all-cause readmission (“readmission rate”), and mortality information was obtained from administrative data and exported into a single database for statistical analysis. Readmissions, inpatient mortality, and cost of care were the primary outcomes; consultant use and length of stay were secondary outcomes. A hospitalization was considered a readmission if the patient returned to inpatient status at AAMC for any reason within 30 days of a previous inpatient discharge. Inpatient mortality was defined as patient death during hospitalization. The cost of care was measured using the case charges associated with each encounter. Charge capture data from both groups was analyzed to classify visits as “physician-only,” “physician co-visit,” and “PA-only” visits. A co-visit consists of the physician visiting the patient after the PA has already done so on the same day, taking their own history and performing their own physical exam, and writing a brief progress note. These data were compared against the exported administrative data to find matching encounters and associated visits, with only matching visits included in the analysis. If a duplicate charge was entered on the same day for a patient, any conflict was resolved in favor of the physician visit. A total of 49,883 and 28,663 matching charges were identified for the conventional and expanded PA groups.
Statistical Methods
Odds of inpatient mortality were calculated using logistic regression and adjusted for age, insurance status, APR-DRG ROM, and LOS. Odds of readmission were calculated using logistic regression and adjusted for age, LOS, insurance and APR-DRG SOI. Cost of care (effect size) was examined using multiple linear regression and adjusted for age, APR-DRG SOI, insurance status and LOS. This model was fit using the logarithmic transformations of cost of care and LOS to correct deviation from normality. Robust regression using MM estimation was used to estimate group effects due to the existence of outliers and high leverage points. Length of stay (effect size) was assessed using the log-transformed variable and adjusted for APR-DRG SOI, age, insurance status and consultant use. Finally, category logistic regression models were fit to estimate the odds of consultant use in the study groups and adjusted for age, LOS, insurance status and APR-DRG SOI.