MAKING THE MOST OF LIMITED DATA WITH UNCERTAINTY USING PROBABILISTIC ANALYSES: A CASE STUDY IN ONCOLOGY

Purpose: Frequently in oncology drug development, model-based projections of treatment outcomes are sought that must be based on limited data from early-phase clinical trials.  The assessment of treatment benefit may be improved by adjusting for baseline risk factors.  Small sample sizes and inclusion of covariates implies small degrees of freedom, which makes probabilistic analyses very important.  If probabilistic analyses are conducted by independently drawing parameters of the survival functions for progression-free survival (PFS) and overall survival (OS), the relationship between PFS and OS may not be realistic. This work presents an illustrative case study using bootstrapping in the context of fitted statistical models to assess treatment benefit and its uncertainty with respect to PFS and OS estimates for decision-analytic models.

Methods: A randomized phase II trial (PRECEDENT) evaluated the efficacy of vintafolide plus pegylated liposomal doxorubicin (V+PLD) vs. PLD alone in platinum-resistant ovarian cancer treatment. PRECEDENT showed the efficacy of V+PLD was present in patients who tested to have 100% folate receptor positive tumors (FR [100%]).  For FR[100%] patients (n=37), median PFS were 24.0 vs. 6.6 weeks for V+PLD vs. PLD patients (HR 0.381; p=0.018).  Our study sought to extend the estimation of the PFS and OS for use in modeling.  Parametric Weibull survival models were estimated including treatment indicator and pre-specified baseline factors.  Survival probabilities were estimated for all patients given a treatment and baseline covariates.  The mean of all patients’ calculated survival probabilities at each time point yielded estimated population survival curves, with area under the curve providing the mean survival time.  Variability was assessed by repeatedly drawing samples with replacement from the trial data (bootstrapping). 

Results: Modeled OS and PFS by group fit Kaplan-Meier curves well. Predicted means with confidence intervals are shown for 2-year results:

Mean time (months)	PFS	95% CI	OS	95% CI
PLD	2.6	(1.4, 5.7)	9.3	(5.3, 15.4)
V+PLD	6.8	(3.9, 10.9)	14.3	(11.0, 17.6)
Diff	4.1	(-0.8, 8.3)	5.0	(-2.9, 10.1)

Conclusions: This case study illustrates an approach to statistically fit and capitalize on bootstrapping to estimate and capture the uncertainty in modeling PFS and OS for a small, randomized Phase II trial.  This methodology can also be used for other outcomes important to decision-analytic models.   The techniques described will help modelers working with limited clinical data.