up). The data were analyzed using a person-time competing risk survival model (see Gibbons et al., 2003), in which the competing risks were death or transplant failure. Transplant failure included failure to engraft or autologous reconstitution (i.e., engraftment with the recipients own cells). In those cases in which engraftment time was not recorded, it was set to time of autologous reconstitution, or time of backup transplant, or time of death. If none of these events occurred (n = 43 cases), engraftment time was set to 31 days, which is the value recommended by NYBC. The primary variables of interest were the degree of HLA match (low/intermediate at HLA-A and HLA-B and high resolution for HLA-DRB1), and the total nucleated cells per kilogram (TNC/kg), expressed as low (<2.5 × 107 cells/kg), medium (2.5 to 5.0 × 107 cells/kg), and high (>5.0 × 107 cells/kg). The primary hypotheses were that (1) an imperfect match decreases survival time by increasing transplant failure and mortality, and (2) an increase in cell dose increases survival time by decreasing transplant failure and decreasing the rate of mortality. To adjust for case mix, the covariates of time (in months), age, weight (kg), disease type (leukemia versus genetic or other), risk (0 to 3 on the International Bone Marrow Transplant Registry [IBMTR] scale), race, sex, and CMV infection were also included in the analysis. Site was added to the model to adjust for overall differences among the three sites in both transplant failure and mortality rates. The NYBC Blood Bank provided all transplants from 1993 through 1998 (n = 562). NMDP provided all transplants for their blood banks for which consent for outcomes research was obtained, in which presumed consent was assumed for patients who died before they could provide consent (n = 192). This occurred because for many cases consent was obtained retrospectively. NHLBI provided all on-study (COBLT study) protocol cases that received a COBLT study cord blood unit (n = 210). The final data set comprised 755 cases (NYBC, n = 384; NMDP, n = 165; NHLBI, n = 206) after exclusion of patients with prior transplants, HLA matches of less than 4/6, cases with missing data, and transplants that were not conducted in the United States.
Two model parameterizations were used. The first was a main-effects model, and the second included the interaction between cell dose and HLA mismatch. The second, more flexible model allows the effect of cell dose to vary depending on the degree of HLA mismatch. Both models included case-mix adjustment for the previously mentioned covariates. The main-effects model provides a more easily interpretable test of the two primary null hypotheses, whereas the model with cell dose-by-HLA interaction better reproduces the observed proportions of patients who survived, had transplant failures, and died.
One-tailed statistical tests were used to test HLA and TNC/kg effects; otherwise, two-sided tests were used to test the case-mix covariates.