Hepatic graft-versus-host disease (HGVHD) contributes significantly to morbidity and mortality after hematopoietic stem cell transplantation (HSCT). Clinical findings and liver biomarkers are neither sensitive nor specific. The relationship between clinical and histologic diagnoses of HGVHD was assessed premortem and at autopsy. Medical records from patients who underwent HSCT at the National Institutes of Health (NIH) Clinical Center between 2000 and 2012 and expired with autopsy were reviewed, and laboratory tests within 45 days of death were divided into 15-day periods. Clinical diagnosis of HGVHD was based on Keystone Criteria or NIH Consensus Criteria, histologic diagnosis based on bile duct injury without significant inflammation, and exclusion of other potential etiologies. We included 37 patients, 17 of whom had a cholestatic pattern of liver injury and two had a mixed pattern. Fifteen were clinically diagnosed with HGVHD, two showed HGVHD on autopsy, and 13 had histologic evidence of other processes but no HGVHD. Biopsy or clinical diagnosis of GVHD of other organs during life did not correlate with HGVHD on autopsy. The diagnostic accuracy of the current criteria was poor (κ = −0.20). A logistic regression model accounting for dynamic changes included peak bilirubin 15 days before death, and an increase from period −30 (days 30 to 16 before death) to period −15 (15 days before death) showed an area under the receiver operating characteristic curve of 0.77. Infection was the immediate cause of death in 68% of patients. In conclusion, liver biomarkers at baseline and GVHD elsewhere are poor predictors of HGVHD on autopsy, and current clinical diagnostic criteria have unsatisfactory performance. Peak bilirubin and cholestatic injury predicted HGVHD on autopsy. A predictive model was developed accounting for changes over time. Further validation is needed.