Statistical Methods
The distribution of patients among the three weight categories (underweight,normal/overweight,andobese)atdiagnosiswascomparedusingthe2
test
for each predictor listed in Table 1. The primary end point for outcome analysis
was EFS, defined as the time from study entry to no achievement of remission,
disease relapse, or death resulting from any cause. Patients without an observed
event were censored at the time of last contact. In addition, conditional analyses
were performed using EFS measured from the start of maintenance among patients
who reached maintenance. The primary method of analysis for impact of
weightcategoryandotherpredictorsonEFSwasCoxproportionalhazardsregression
analysis stratified by protocol treatment arm. The percent time spent at an
extreme weight category during premaintenance therapy cycles was computed as
the course length–weighted proportion of courses during which the patient was
classified as underweight or obese, with patients classified as 50% cumulative
time underweight, 50% cumulative time underweight, normal/overweight
throughout, 50% cumulative time obese, or 50% cumulative time obese.
These categories were then used to investigate the effect of cumulative time at an
extreme weight on EFS from the start of maintenance after adjusting for weight
category at diagnosis. To rule out the possibility that the observed effect of cumulativetimespentatweightextremeswasanartifactbecauseofclusteringofpatients
near the border of the underweight and obese categories at diagnosis (ie, whether
patients who changed weight category were those only slightly obese or underweight),
we repeated the analysis dividing BMI at diagnosis into both seven and
nine weight categories. We also verified the model by including BMI percentile at
diagnosisasacontinuouslinear-quadraticelementinstead.Examinationoffailure
rates (relapse or death) before the start of maintenance in each weight group was
performed to assess differences for patients who did not survive to maintenance.
To evaluate the impact of concurrent weight on acute TRT, each patient’s
weight was classified at the start of each phase into obese, underweight, or
normal/overweight categories. Primary end points for the analysis of TRT
were the occurrences during each course of: grade 3 to 4 nonhematologic
toxicities as defined by the CCG Toxicity and Complications Criteria (TCC);
hematologic toxicity as reflected by the surrogate measure of hematologic
support (ie, transfusion of blood products, elective use of hematopoietic
growth factors); documented infections; use of other supportive care interventions
(antimicrobial treatments, intravenous analgesics, parenteral nutrition);
delay in start of treatment phase; and number of hospital and intensive care
unit (ICU) days. The CCG-TCC define grade 3 performance toxicity as a
Karnofsky score 50% and grade 3 weakness as causing functional impairment.
These end points were analyzed during the period from diagnosis
through the first three cycles of maintenance. The primary method of analysis
was an end point–specific multivariable linear or logistic regression model.
The variables listed in Table 1 were included as candidate predictors in the
multivariable EFS and TRT models (TRT analyses also contained treatment
phase). Predictors were initially screened in univariable analyses, and those
with P .15 via likelihood ratio test were included in a preliminary multivariable
model. The final model for each end point (Appendix Table A1, online
only) was determined by a reverse stepwise selection of predictors until all
remaining covariates had likelihood ratio P .05. This model was then
verified by testing whether reintroducing previously eliminated predictors
would significantly improve the model. Use of formal repeated measure methods
in the TRT analyses had a negligible effect on the results. Unless otherwise
stated, P values refer to two-sided tests. Statistical computations were performed
with STATA software (version 11; STATA, College Station, TX).