Shorter Isn't Always Better: Identifying the Relationship between Length of Hospital Stay and the Probability of Readmission

Abstract

The length of patient stays in U.S. hospitals has been in steady decline since the advent of Medicare’s Prospective Payment System in 1983. This decline, incentivized by diagnosis-based flat fee reimbursements, has been found to be largely benign in its effect on patient outcomes by 30 years of medical research. The literature, however, has relied on analysis of broad aggregate trends in mortality and readmission rates concomitant with the decline in length of stay (LOS), unable to identify the effects of trends in LOS separately from advances in medicine and public health. Further, the reverse causal relationship between LOS and patient outcomes, via bias from omitted patient health characteristics, while given cursory acknowledgment, is never controlled for in patient level studies. We analyze the records of 395,828 adult patients with a primary diagnosis of heart failure hospitalized in California between 2005 and 2011 and estimate the effect of hospital LOS on probability of 30-day readmission, controlling for demographics, comorbidities, procedures, and medical complications. We use hospital occupancy rates and emergency vehicle diversions as instrumental variables to identify the effect of patient LOS on the probability of readmission within 30 days. We find a U-shaped relationship between LOS and the probability of hospital readmission. The average probability of readmission initially decreases with LOS, and does not begin to increase until LOS exceeds 9 days, relative to a mean LOS of 4.55 days. Our results suggest that studies that fail to control for omitted variable bias significantly underestimate the benefits from the early stages of a patient’s hospital stay. The endogenous relationship between LOS and readmission rates has implications for patient-care initiatives, such as Accountable Care Organizations under the Affordable Care Act, where reimbursement policies are tied to both metrics.