2.5. Statistical Analysis

Characteristics of groups with or without Hb decline at baseline were compared by chi-squared tests, independent-group t-tests, and Wilcoxon rank-sum tests, depending on the variable levels and distribution of measurements after testing for normality by the Kolmogorov–Smirnov test. The missing rates were less than 5% for each variable, and the attribute values were missing at random. The default missing-data analysis was complete-case analysis. Cox multiple regression analysis was performed to evaluate the independent contributions of baseline clinical characteristics, medical history, and laboratory data to the occurrence of all-cause mortality at 2 years with forward stepwise selection. Patients were grouped according to the quintile of discharge Hb to permit comparison of the relationship with 2-year outcomes between the groups (quintile 1: 5.4–11.3 g/dL; quintile 2: 11.3–12.5 g/dL; quintile 3: 12.5–13.5 g/dL; quintile 4: 13.5–14.6 g/dL; quintile 5: 14.6–18.8 g/dL). Hb levels of quintile 3 were used as a reference for odds ratio (OR) analyses. Logistic regression was performed to find whether a dose-response effect exists. Receiver-operating characteristic area under the curve (ROC-AUC) was constructed to further illuminate the best cutoff values per 1 g/L decrease in discharge Hb to predict 2-year all-cause mortality. Kaplan–Meier survival curve analysis compared survival among the severity of bleeding groups using the log-rank test. All analyses were conducted using Stata software, version 12 (Stata Corp., College Station, TX), and a two-sided P < 0.05 was considered statistically significant.