Download presentation
Presentation is loading. Please wait.
Published byTeresa Gallagher Modified over 5 years ago
1
Subgroup analysis on time-to-event: a Bayesian approach
August 2019 Duy Ngo, Richard Baumgartner, Joseph Heyse, Shahrul Mt-Isa, Jie Chen, Dai Feng & Patrick Schnell (Ohio State University)
2
Overview Identifying patient subgroups with favorable benefit from treatment is of interest to regulators and health technology assessment agencies worldwide Evaluate Bayesian credible subgroups in a setting of survival analysis to identify the baseline covariate profiles of benefiting patients Treatment effect for benefiting subgroup estimated using two summaries log hazard ratio restricted mean survival time (RMST) Application of Bayesian credible subgroups in this setting offers several advantages Avoids practice of testing for interaction Addresses multiplicity Does not require pre-specification of subgroups and work directly with the covariate space Naturally makes statistical inferences from the full posterior distribution. Methods applied to a case study of prostate carcinoma and simulated large clinical dataset. PROPRIETARY ICONS HERE
3
Outline Background and Motivation
Bayesian Credible Subgroups for Survival Endpoints Simulation Study The Prostate Cancer Dataset The Simulated Dataset from a Large Clinical Trial Discussion and Conclusion PROPRIETARY ICONS HERE
4
Background and Motivation
Subgroup analysis for survival endpoints from a simulated data. PROPRIETARY ICONS HERE
5
Concepts Goal: Finding subgroups of population that benefiting the treatment, i.e. estimate B. B: unknown true benefit D: evidence of benefit S \ D: insufficient evidence (uncertainty region) π πΆ : no benefit. PROPRIETARY ICONS HERE
6
Existing methods Existing methods to estimate B:
Testing for treatment-covariate interaction. Limitation: not appropriate for subgroup identification. Tree-based methods. Limitation: instability. Main challenges of subgroup analysis: Lack of power to detect the overall main effect difference in response between treatment groups, Multiplicity arising from simultaneous inference on all subpopulation members (predictive covariate points), Post hoc analysis (unplanned analyses or data dredging), Interpretation and conclusions (inference for each individual separately). Schnell (2016) proposed Bayesian credible subgroup for continuous endpoints with several advantages: Control multiplicity, Easily make statistical inferences from the full posterior distribution. PROPRIETARY ICONS HERE
7
Notation πβ₯0 denotes the response variable (failure or survival time).
Suppose that the observed time-to-event data consist π independent subjects of ( π π , π₯ π , π§ π , π
π , π π ) where π π =minβ‘( π π , πΆ π ) and πΆ π is a random variable for censoring (right censoring), π₯ π and π§ π be px1 and qx1 vectors of prognostic and predictive covariates, respectively, π
π is a censoring indicator, i.e. π
π =1 for π π β€ πΆ π and 0 otherwise, π π ={0, 1} is a treatment indicator. PROPRIETARY ICONS HERE
8
Personalized Treatment Effects
Average treatment effect (ATE) is the average over the entire population of the individual treatment effects. Personalized treatment effect (PTE) is the treatment effect for a patient given their baseline characteristics. Optimal treatment regimes find a set of decision rules to provide optimal treatment for a given patient. They are related to the subgroup analysis, but since their focus is on prediction for a single individual, they are not concerned with the multiplicity issues. PROPRIETARY ICONS HERE
9
A Log of Hazard Ratio as a PTE
In a two-arms study with censoring, a traditional Cox regression model has the form: π π‘ π₯, π§, π = π 0 π‘ exp π₯ β² π½+π π§ β² πΎ , where π 0 (π‘) is a baseline hazard, π½ and πΎ are regression coefficients. The PTE for a patient with covariate π§ is Ξ π» = π π‘ π₯, π§, π=1 π π‘ π₯, π§, π=0 = exp π§ β² πΎ < πΏ π» , where πΏ is a predetermined clinical significance. Alternatively, the log of hazard ratio as a PTE is log Ξ π» = π§ β² πΎ< log πΏ π» . PROPRIETARY ICONS HERE
10
A Difference in Restricted Mean Survival Time (RMST)
The RMST is the area under a survival curve π(π‘) between π‘=0 and π‘= π: π= π π π π‘ ππ‘ . The difference in RMST between two arms up to time point π is Ξ π
π = π π=1 β π π=0 = π π [π π‘| π=1 βπ π‘ π=0)] ππ‘ , The PTE for a patient π§ is Ξ π
π > πΏ π
π where πΏ π
π is a predetermined threshold of clinical significance. We use conventional Cox proportional hazard model to estimate the two survival functions on the grid of subgroup-defining covariates in order to compute Ξ π
π . PROPRIETARY ICONS HERE
11
Construct Bayesian Credible Subgroup
The two-step regression-classification procedure: Step 1: Define a model, fit a regression and obtain the joint posterior of coefficients of predictive covariates (interacting with treatment choice), Step 2: Computing the bounds and obtain a pair π·, π where π· is an exclusive credible subgroup and π is an inclusive credible subgroup. PROPRIETARY ICONS HERE
12
π π‘ π₯ π , π§ π , π π = π 0 π‘ exp π₯ π β² π½+ π π π§ π β² πΎ ,
Simulation Study Suppose that the hazard function for π π‘β subject is π π‘ π₯ π , π§ π , π π = π 0 π‘ exp π₯ π β² π½+ π π π§ π β² πΎ , where π₯= π₯ 1 , π₯ 2 β² and π§= 1, π§ 1 , π§ 2 β². Let π₯ 1 = π§ 1 ={0, 1}, and π₯ 2 = π§ 2 uniformly distributed on interval (-3, 3). Let π= 0,1 , and π 0 π‘ =ππ ππ‘ πβ1 is a Weibull baseline hazard with π=0.05 and π=1.1. Perform diagnostic test for credible subgroup for sample size π={50, 100, 500, 1000}, credible level at 0.8, and different settings for π½ and πΎ: The prognostic features are with no or small effect π½= 0, 0 , set πΎ= 0,0,0 and (1,β1,3). The prognostic features have moderate effect π½= 0.2, 0.2 , set πΎ= 1, 1, 1 . The prognostic features have higher effect π½= 1, β2 , set πΎ= 1, 0.1, 1 . Each scenario, we simulate 1000 datasets, and for each dataset, we use 1000 posterior draw kept after 500 burn- in iteration when we perform Bayesian method for Cox model. . PROPRIETARY ICONS HERE
13
Simulation Study We use πΏ π» =1 and πΏ π
=0.
We report the performance of Bayesian credible subgroup in the following criteria: Total coverage: the frequency with which π·βπ΅βπ under a fixed value πΎ. Pair size: proportion of the population included in the uncertainty region, i.e. π π§βπβπ· π·, π) with uniform measure on π§. Specificity and Sensitivity of D: how well the credible subgroup align with the benefiting group. PROPRIETARY ICONS HERE
14
Simulation Results πΏ π» =1, πΏ π
=0, and credible level 0.8.
PROPRIETARY ICONS HERE
15
Simulation Results πΏ π» =1, πΏ π
=0, and credible level 0.8
PROPRIETARY ICONS HERE
16
The Prostate Cancer Dataset
The prostate cancer dataset has been analyzed in literature for exploratory subgroup analysis. Ballarini et al. (2018) proposed a multiple regression model with a Lasso-type penalty to estimate benefiting subgroups. The dataset includes 475 patients who were randomly assigned either to a combination of placebo and the lowest does level of diethyl stilbestrol (control group) or the higher doses (treatment group). The interest covariates are: existence of bone metastasis (bm), disease stage either 3 or 4 (stage), performance (pf), history of cardiovascular events (hx), age and weight (wt). Denote rx as treatment indicator, and include the two interactions bm:rx and age:rx in the model. PROPRIETARY ICONS HERE
17
Result A log of hazard ratio as a PTE: PROPRIETARY ICONS HERE
18
Result A difference in RMST as a PTE: PROPRIETARY ICONS HERE
19
The Simulated Dataset Motivated by a Large Clinical Trial.
The simulated dataset is motivated by a large clinical trial reported in Scirica et.al. (2012) A simulated dataset pertains to patients of whom 8898 were assigned to treatment and 8881 were assigned to placebo. There are 5 variables to consider: age at entry (years), baseline weight (kilograms), history of hyperlipidemia, smoking status and prior coronary revascularization. For each treatment group, we randomly selected 20% of subjects and added a Gaussian noise with zero mean and standard deviation of 1 and 5 for continuous covariates age and baseline weight, respectively. The primary efficacy endpoint is the time of first cardiovascular death, myocardial infarction or stroke. The median followβup was 2.5 years (IQR years). Goal: search for benefiting subgroups without prespecified subgroups of interest. PROPRIETARY ICONS HERE
20
Baseline characteristics
Continuous variable: Median (IQR). Categorical variable: percentage. PROPRIETARY ICONS HERE
21
Result A log of hazard ratio as a PTE PROPRIETARY ICONS HERE
22
Result A RMST differences PROPRIETARY ICONS HERE
23
Discussion We introduced a Bayesian credible subgroup for time-to-event data by using a log of hazard ratio and the difference in RMST as PTEs. Limitations: parametric model, model selection and missing covariates. Research in subgroup analysis is mainly focused on assessment of benefit. It is desirable to assess both benefit and risk in subgroup analysis. Potential approach: Extend Bayesian credible subgroup method for multiple treatments and multiple endpoints. Multiple endpoints can include both benefit and risk. PROPRIETARY ICONS HERE
24
Thank you Questions PROPRIETARY ICONS HERE
25
BACK UP PROPRIETARY ICONS HERE
26
Bayesian Credible Subgroups
Let π be a covariate space, a goal of Bayesian credible subgroups searches the covariate points π§ such that π΅ π» = π§βπ: Ξ π» π§ < πΏ π» , for HR case, and π΅ π
π£ = π§βπ: Ξ π
π£ π§ > πΏ π
for a RMST difference case. In Bayesian framework, common estimators are π΅ π», πΌ = π§βπ: π(Ξ π» π§ < πΏ π» ) | π·ππ‘π)>1βπΌ π΅ π
π£, πΌ = π§βπ: P(Ξ π
π£ π§ > πΏ π
π·ππ‘π >1βπΌ PROPRIETARY ICONS HERE
27
Construct Bayesian Credible Subgroups
Step 1: Define the model, fit a regression and obtain the marginal posterior of πΎ Model: π π π‘ π₯ π , π§ π , π π = π 0 π‘ exp π₯ π β² π½+ π π π§ π β² πΎ Priors: We specify a joint priors for all unknown parameters (π½, πΎ, π 0 ) as followings: π π½, πΎ, π 0 =π π½, πΎ π( πΎ 0 ), π½, πΎ βΌπ π 0 , Ξ£ 0 where π π 0 , Ξ£ 0 is he multivariate normal distribution with π+π Γ1 mean vector π 0 and a π+π Γ(π+π) covariance matrix Ξ£ 0 , We choose a nonparametric gamma process prior on the baseline hazard π 0 . Obtain the posterior of πΎ by using Gibbs sampling (from an R package spBayesSurv) Obtain the posterior for PTEs Ξ. PROPRIETARY ICONS HERE
28
Construct Bayesian Credible Subgroups
Step 2: Compute the bounds and obtain a pair (π·, π) The simultaneous credible bands for Ξ(z) on a covariate space π is Ξ π§ β Ξ π§ Β± π πΌ πππ(Ξ(π§)) , where π πΌ is the 1 βπΌ quantile of the distribution of π= sup π§βπ Ξ π§ β Ξ π§ 2 πππ(Ξ(π§)) and Ξ z is the posterior mean of Ξ π§ . In a case of Ξ π§ β‘ Ξ π» (π§), the exclusive credible subgroup π·= π§βπ: Ξ π§ + π πΌ πππ Ξ π§ , and the inclusive credible subgroup S= π§βπ: Ξ π§ β π πΌ πππ Ξ π§ PROPRIETARY ICONS HERE
29
Simulate Time-to-event Data
Suppose that the hazard function of the π π‘β individual is π π π‘ π₯ π , π§ π , π π = π 0 π‘ exp π₯ π β² π½+ π π π§ π β² πΎ . We assume a Weibull baseline hazard, i.e. π 0 π‘ =ππ ππ‘ π β1 where π and π are the scale and shape parameters, respectively. If π is uniformly distributed on [0, 1], the survival time π π =β log π π exp π₯ π β² π½+ π π π§ π β² πΎ π . Suppose that πΆ π βΌπΈπ₯π(π) is censoring time. Due to censoring, we observe π π = min ( π π , πΆ π ) with a censoring indicator π
π . PROPRIETARY ICONS HERE
30
Simulation study under nonproportional hazard assumption
When the PH assumption is violated, the HR may not accurate represent PTEs and RMST is an alternative approach. We simulated two groups with different hazard rates. The treatment group ( π π =1) had a constant exponential hazard with rate π 0 π‘ =0.01. The control group had a piecewise exponential hazard with rate π 0 π‘ =0.01 for 0β€π‘< π‘ π , and π 0 π‘ =0.1 for π‘ π β€t. We use similar settings for the prognostic and predictive covariates in simulation study under PH assumption. We consider π½= 0.7, 0.7 and πΎ= 0.5, β0.5, β0.5 . We chose πΏ π
π =0 and π‘ πΆ =30. The RMSTs were computed at the change point π‘ πΆ up to π‘ πΆ +50. PROPRIETARY ICONS HERE
31
Simulation study under nonproportional hazard assumption
Average summary statistics for 80% credible subgroup pairs under nonproportional hazard assumption PROPRIETARY ICONS HERE
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.