Background: Immune checkpoint inhibitors are now standard-of-care for patients with esophageal squamous cell carcinoma(ESCC). The predictive significance of PD-L1 expression in ESCC has been investigated by different antibodies, scoring algorithms and cutoff values in several phase III trials. However, it remains controversial whether PD-L1 is a predictive biomarker for response to PD-1/PD-L1 blockade. Evaluating the concordance of the commercially available PD-L1 assays is essential to understand the discrepancy in the ESCC clinical trials, and helpful to further investigate the predictive value of PD-L1 expression. Methods: 145 archival tumor samples were obtained from 131 patients with ESCC (stage I-IV) at National Taiwan University Hospital. Formalin-fixed, paraffin-embedded archival tumor samples were assessed by three commercially available PD-L1 assays: VENTANA SP263, Dako 22C3 and Dako 28-8 assays. Assays were performed in a College of American Pathologists accredited central laboratory and scored for PD-L1 staining by using multiple metrics including tumor cell score (TC), immune cell score (IC) and combined positive score (CPS) or tumor area percentage (TAP). Analytical concordance was calculated pairwise between assays using the Spearman (ρ) rank correlation coefficient. Classification concordance, including agreement between clinically relevant scoring algorithms, was investigated using positive percent agreement (PPA), negative percent agreement (NPA), and overall percent agreement (OPA) at multiple cutoff values to assess the overlap between populations by using different assays. Results: SP263, 22C3 and 28-8 assays showed good analytical correlation for TC staining (Spearman’s rank correlation coefficient 0.8–0.9). Correlation was lower for IC (Spearman’s rank correlation coefficient 0.59-0.61). There was moderate overlap between populations identified by SP263 and 28-8 or 22C3 based on CPS or TAP algorithm at multiple cutoff values. OPAs for these 3 assays ranged from 68%–88% at various matched algorithms. The SP263 and 22C3 PD-L1 assays appeared relatively more sensitive, assigning a higher proportion of patients as PD-L1 positive or high, compared to 28-8 assay. When using assay-specific clinically relevant algorithm, moderate classification agreement was seen for SP263 versus 22C3 or 28-8. Differences were observed between patient populations with tumor classified as PD-L1 high versus PD-L1 low/negative using CPS≥10, TAP≥10% and TC≥1%. The PPA between 22C3 CPS≥10 and SP263 TAP≥10 or 28-8 CPS≥10 or 28-8 TC≥1% were 75%, 57% and 68% respectively. Conclusions: This study is the first dataset to compare various PD-L1 assays in ESCC. Differences in classification of patients with PD-L1 high versus low/negative using clinically relevant algorithms suggest that caution should be taken when comparing data across the trials.