Objective: Previous research on inpatient care for children and adolescents with emotional or behavioral problems indicates that patient-level factors predict length of stay (LOS) poorly. This analysis examines whether patient-level factors are poor predictors of LOS, because LOS is primarily determined by facilities rather patients. Study Design: This study uses Tennessee Medicaid claims data from 1996 to 2001. The data include information on 14,162 observations related to 8400 patients (age 12-21) from 163 hospitals. We estimate log LOS using a cross-classified model. Covariates include admission-level characteristics (age, diagnosis, qualification for Medicaid, year), patient-level characteristics (gender, race), and facility characteristics (facility type). Principle Findings: Our results suggest that variation in LOS is attributable to facility-level factors (51%), time-invariance patientlevel factors (5%), factors that vary across admissions (42%), and a correlation between patient-level and facility-level factors (5%). Conclusions: About half of the variation in LOS is explained by facility-level factors. Given the vulnerable nature of youth who are in need of inpatient psychiatric care, it may be particularly important to monitor provider-level processes and outcomes. Measuring facility or provider level quality is complicated because of difficulties in adjusting for case-mix severity across providers. The methodology presented here represents a general framework that can be widely used in health services research. Potential applications include broadening models of utilization to simultaneously include patient, provider, geographic and community level variations, as well as provider profiling. (Med Care 2008;46: 240 -246)