Regression models for RFS
. | Bayesian Weibull model . | . | . | Cox model . | . | |||
|---|---|---|---|---|---|---|---|---|
| Covariate* . | Mean posterior effect (SD) . | 95% posterior credible interval for the effect† . | Posterior probability of shorter RFS . | Estimate of effect (SE) . | P‡ . | |||
| CRB | -0.02 (0.13) | -0.28,0.23 | 0.43 | 0.01 (0.13) | .94 | |||
| No G-CSF | 0.22 (0.10) | 0.01,0.42 | 0.98 | 0.17 (0.11) | .12 | |||
| Inv16 or t(8;21) | -0.47 (0.21) | -0.88,0.07 | 0.01 | -0.45 (0.21) | .03 | |||
| Cytogenetics -5/-7 | 0.75 (0.12) | 0.53,0.97 | 1.00 | 0.74 (0.12) | <.0001 | |||
| Log [(WBC) × (% circulating blasts at presentation)] | 0.04 (0.02) | 0.01,0.07 | 0.99 | 0.04 (0.02) | .02 | |||
| Log days to CR-MDA | 0.66 (0.16) | 0.35,0.96 | 1.00 | 0.61 (0.16) | .0001 | |||
| Received TA§ | 0.26 (0.13) | 0.02,0.53 | 0.98 | 0.19 (0.13) | .15 | |||
| Received FA§ | 0.09 (0.14) | -0.16,0.37 | 0.75 | 0.09 (0.14) | .51 | |||
. | Bayesian Weibull model . | . | . | Cox model . | . | |||
|---|---|---|---|---|---|---|---|---|
| Covariate* . | Mean posterior effect (SD) . | 95% posterior credible interval for the effect† . | Posterior probability of shorter RFS . | Estimate of effect (SE) . | P‡ . | |||
| CRB | -0.02 (0.13) | -0.28,0.23 | 0.43 | 0.01 (0.13) | .94 | |||
| No G-CSF | 0.22 (0.10) | 0.01,0.42 | 0.98 | 0.17 (0.11) | .12 | |||
| Inv16 or t(8;21) | -0.47 (0.21) | -0.88,0.07 | 0.01 | -0.45 (0.21) | .03 | |||
| Cytogenetics -5/-7 | 0.75 (0.12) | 0.53,0.97 | 1.00 | 0.74 (0.12) | <.0001 | |||
| Log [(WBC) × (% circulating blasts at presentation)] | 0.04 (0.02) | 0.01,0.07 | 0.99 | 0.04 (0.02) | .02 | |||
| Log days to CR-MDA | 0.66 (0.16) | 0.35,0.96 | 1.00 | 0.61 (0.16) | .0001 | |||
| Received TA§ | 0.26 (0.13) | 0.02,0.53 | 0.98 | 0.19 (0.13) | .15 | |||
| Received FA§ | 0.09 (0.14) | -0.16,0.37 | 0.75 | 0.09 (0.14) | .51 | |||
Age, antecedent hematologic disorder (AHD) of more than 1 month (yes vs no), diagnosis (AML vs RAEB), platelet count, hemoglobin, circulating blast and neutrophil percentages/counts both before treatment and at CR, pretreatment marrow blast percentage, and number of courses of chemotherapy were also considered but dropped from the model.
Under the Bayesian model, a posterior mean of 0 corresponds to a beneficial and a harmful effect being equally likely. Because the 95% posterior credible interval for the effect of CRB versus CR (-0.28, 0.23) is very nearly centered at 0, there is very weak evidence for either a beneficial or a harmful effect on RFS. Because an effect of 0 corresponds to a relative risk (RR) of 1, each estimated mean and 95% credible interval may be converted to a RR by simply expotentiating. Thus, the RR of failure because of CRB versus CR has 95% credible interval of 0.76, 1.26.
A posterior probability of shorter (or longer) RFS close to either 0 or 1 under the Bayesian model corresponds to a small P under the conventional Cox model.
TA indicates topotecan + ara-C-containing regimen; FA, fludarabine+ ara-C-containing regimen. TA and FA were compared with idarubicin + ara-C-containing regimens (IA).10