Tuesday, Aug 5: 2:00 PM - 3:50 PM
4131
Contributed Papers
Music City Center
Room: CC-209A
Main Sponsor
Biopharmaceutical Section
Presentations
Randomized controlled trials (RCTs) are increasingly being used to model conditional average treatment effects (CATEs) using baseline patient characteristics as predictors, generating evidence that could be used to inform personalized treatment strategies. However, missing data in the baseline predictors is ubiquitous in RCT datasets, creating challenges for model derivation and evaluation in this context. We develop a family of extensions to the Bayesian Causal Forest (BCF) model that incorporates missing data priors and is compatible with continuous and binary outcomes. Furthermore, we incorporate flexibility beyond the usual missing at random assumption by exploiting the binary decision trees that form the basis of BCF's model structure. Compared to existing strategies for missing data in CATE modeling, we show via simulations that our approach leads to improved accuracy and quantification of uncertainty when summarizing the evidence for heterogeneous CATEs in RCTs with missing values in the predictor variables. Finally, we demonstrate the practical utility of our method via analysis of existing trial datasets.
Keywords
Bayesian statistics
Missing data
Conditional average treatment effects
Randomized controlled trials
Statistical learning
Background: Personalized estimates of treatment benefit allow for informed decision making.
Methods: Using the addition of ADT to RT in prostate cancer as an example, we calculate and compare measures of treatment benefit using differences in probability vs differences in mean times. We utilize a novel data integration approach to calculate the absolute risk of PCSM (cancer mortality) and MFS (mets free survival) within 15 years using patient level covariates for both cancer outcomes and competing mortality risk. We calculate Mean MFS times unrestricted and restricted to 15 years. We calculated each of these measures for individual patients in a contemporary cohort of >1000 patients enrolled in a statewide quality consortium.
Results: The 15-year risk of PCSM for a stage IIC patient treated with RT+ADT varies from 7% to 15% at the 10th vs 90th percentile of competing mortality risk. For men in the same UIR risk group, ADT reduced the risk of mets at 10 years by an average of 4%, but the 10th and 90th percentiles were 1% and 14%, respectively.
Conclusions: Accounting for individual competing risk levels is important when estimating treatment benefit.
Keywords
Data Integration
Survival Analysis
Competing risks
Treatment efficacy
Personalized medicine
Several attempts have been made to estimate causal effect on principal stratum. These methods deal with causal inference on principal stratum in terms of the population means of potential outcomes. However, sometimes the median or other quantiles of outcomes in a principal stratum according to potential occurrence of intercurrent events are of interest in randomized clinical trials. Regarding the estimation of quantile treatment effects on principal stratum, some authors proposed an estimator under the assumption of "exclusion restriction" (ER). However, Ding and Lu (2017) argued that the ER assumption seems unfeasible in some applications. Therefore, we propose principal quantile treatment effect estimators that can nonparametrically estimate the distribution of potential outcomes by principal score weighting without relying on the ER assumption. Our simulation studies show that the proposed method works in situations where the median or quantiles may be regarded as the preferred population-level summary over the mean. We illustrate our proposed method by using simulated data based on a randomized controlled trial conducted on patients with nonerosive reflux disease.
Keywords
estimand
principal stratum
principal stratification
monotonicity
principal ignorability
non-normal
The Bayesian Optimal Interval (BOIN) design has recently become popular in oncology dose-finding trials (Liu, 2015). The local BOIN design has successfully passed the FDA's Fit-for-Purpose (FFP) evaluation, receiving a determination letter certifying its FFP status under a non-informative prior. However, this paper highlights several potential limitations of the local BOIN design. Notably, the criteria for escalation, retention, and de-escalation are independent of sample size, which contradicts both clinical and statistical reasoning. Furthermore, the design is based on questionable hypotheses and the minimization of the sum of Type I and Type II error rates. The Type I error rate can fluctuate between 0 and 1, with a rate of 1 leading to the abandonment of the current dose regardless of trial outcomes. It is essential to consider these limitations when implementing the BOIN design in clinical trials.
Keywords
drug label
LOCF
MMRM
MMRM
Trajectory of Mean (Median) and Observation Rate (TMOR) plot
go/No-go
First Author
Bob Zhong, Regeneron Pharmaceuticals
Presenting Author
Bob Zhong, Regeneron Pharmaceuticals
Multiple imputation is a common approach for analysis in the presence of missing data in clinical trials. For continuous outcomes with missing data, a δ-adjusted approach on the imputed outcomes is commonly used to assess the tipping-points at which the observed results become insignificant. For binary outcomes, the primary analysis can be performed naturally by deriving the binary outcome from the (imputed) underlying continuous variable(s). Then, for each imputation, the δ-adjusted sensitivity analysis can be performed using the following steps: impute the underlying continuous variables and estimate the probability for the binary outcome for each subject based on the primary analysis model; employ an additive δ adjustment on the logit scale; the missing binary outcome is simulated from a Bernoulli distribution with the δ-adjusted probability; the "complete" data after imputation are analyzed using the same statistical model as the primary analysis; repeat above steps M times and the results from the M imputations are then combined using traditional methods. The tipping points are then generated and summarized. We apply this method to data from a real clinical trial.
Keywords
Missing data
Binary outcome
Sensitivity Analysis
Multiple Imputation
Tipping points
Hierarchical composite outcomes (HCO) provide a useful alternative to conventional composite outcomes by emphasizing the relative importance of the outcome components. Treatment effects on these HCOs are commonly described using win statistics such as Pocock's win ratio. This framework is often used with multiple censored time-to-event outcomes, which poses an issue due to differential censoring between subjects. Comparisons on these HCOs can lead to intransitive relationships since analyses are based on common follow-up time between pairs of subjects. Multiple imputation circumvents the problem of intransitivity. We propose a multiple imputation (MI) approach to augment a more complete discussion about estimating treatment effects on HCOs by leveraging the parallelism between the win ratio and the cumulative logit odds ratio. We demonstrate the greater utility and interpretability of the MI approach through simulations and practical examples.
Keywords
Hierarchical Composite Outcome
Multiple Imputation
Win Ratio
Proportional odds logistic regression
Missing data