1. Adjusting effects for treatment switching in HTA
Why, when, which and how?
Claire Watkins
Director and Consultant Statistician, Clarostat Consulting Ltd
PSI Conference, 16th May 2017

2. Standard clinical practice (inc gef)
Background
What do we mean by treatment switch/crossover?
We might build this into the trial protocol
e.g. Sunitinib GIST trial (Demetri, 2012)
Or it might happen spontaneously due to clinical practice in the region, if the treatment is already on the market
e.g. Gefitinib IPASS trial (Fukuoka, 2011)
Patients in a parallel group RCT may switch or “crossover” to the alternative treatment at some point before an endpoint of interest occurs.
Disease prog
Randomise
Sunitinib
Survival
Placebo
Sunitinib
Disease prog
Standard clinical practice (inc gef)
Randomise
Gefitinib
Survival
Doublet chemo

3. Switching – Regulatory vs HTA viewpoint
Regulatory agency
Evaluate efficacy in clinical trial
Switch happened in clinical trial
Do not adjust OS to remove switch
ITT primary
May consider switch adjusted OS as a supportive analysis
Primary endpoint may not be OS anyway
Health Technology Assessment (HTA) Agency
Evaluate effectiveness in real world setting
Switch ≠ real world (mostly)
Use plausible methods* to adjust OS to remove switch
If no plausible methods, use ITT
Key endpoint for lifetime cost effectiveness calculations is OS
* Views of “plausible methods” differ by agency!

4. Why does switching matter for HTA?
It all depends on the decision problem
In HTA, the decision problem is often to compare:
If there is switching and the new therapy is effective, ITT underestimates this difference
 How to estimate long term efficacy without switch?
Current clinical practice without new therapy
Potential future clinical practice including new therapy vs

5. Commonly used methods to estimate control arm survival in absence of switch (Latimer 2014, Watkins 2013)
“Naive” methods
Exclude switchers
Censor at switch
Time varying covariate
“Complex” methods
Inverse Probability of Censoring Weighting (IPCW; observational)
Rank Preserving Structural Failure Time (RPSFT; randomisation based)
Iterative Parameter Estimation (IPE; randomisation based)
Two-stage Accelerated Failure Time (AFT)
External data
Simple to apply
High levels of bias
Assumption: no confounders (variables that influence switch and survival)
Harder to apply
Try to reduce bias

6. Complex 1: IPCW (weight non-switched times)
Control arm survival
Compare to observed experimental arm survival
(Robins 2000)
Observed (ITT) control arm
IPC weighted
WEIGHT
Non switchers
Non switchers
WEIGHT
WEIGHT S
Switchers
Switchers
WEIGHT S
Key
Death time
Censor time
Switch time S
Assumption: The variables in the weight calculation fully capture all reasons for switching that are also linked to survival
(i.e. no unmeasured confounders)
Weights represent how “switch-like” a patient is that has not yet switched

7. Complex 2/3: RPSFT/IPE (adjust post switch times)
(Robins 1991, White 2002, Branson 2002)
Control arm survival
Compare to observed experimental arm survival
Observed (ITT) control arm
RPSFT/IPE adjusted
Non switchers
Non switchers S
Switchers
Switchers S
Time off experimental
Treatment multiplier
Time on experimental x
Time off experimental
Time on experimental
Key
Death time
Censor time
Switch time S
Assumption: Each cycle of treatment extends survival by a constant amount.
(i.e. constant/common treatment effect)
Estimated by non-parametric G-estimation (RPSFT) or parametric model (IPE)

8. Complex (4): 2-Stage AFT (observational study)
Control arm survival
Compare to observed experimental arm survival
(Latimer 2014)
Observed (ITT) control arm
2-stage AFT adjusted
Treat control arm as observational study post progression
Re-baseline at progression
Collect covariate data at progression
Calculate effect of switch treatment adjusting for covariates
Adjust switcher data and compare randomised arms P
Non switchers P P S
Switchers P S
Assumptions: The covariates fully capture all reasons for switching that are also linked to survival
No time-dependent confounding between P and S
Key
Death time
Censor time
Switch time
Progression time S
(i.e. no unmeasured confounders) P
Not valid if switch can occur before progression

9. Method selection process
Specific process proposed in NICE DSU Technical Support Document 16 (Latimer 2014)
Key steps (in general):
Can the model be fitted with available data?
If yes, is the model appropriate given the switching mechanism?
If yes, are the assumptions reasonable?
If yes, are the results plausible?

10. 1. Data collection requirements for commonly used switch adjustment methods (Watkins 2016)
Data required
Exclude switchers
Censor switchers
Time varying covariate
RPSFTM /IPE
2-stage
IPCW
Date of starting switch treatment 
Date of death/censoring
Date of stopping switch treatment *
All baseline covariates that may influence switch decision and OS
All time varying covariates that may influence switch decision and OS, collected until switch or death/censoring
All time varying covariates that may influence switch decision and OS, collected until secondary baseline
Date of secondary baseline (e.g. disease progression) ^
* for “on-treatment” approach only where efficacy stops at treatment end
^ if switch is only allowed after this point
Increasing data collection burden

11. 2. Methods appropriate to switching mechanism
Exclude switchers
Censor switchers
Time varying covariate
RPSFTM /IPE
2-stage
IPCW
<10% switch ✗
>80% switch/perfect switch predictor
Switch occurs before progression for some patients
Switch occurs a long time after progression for some patients
Time on/off treatment or ITT survival is similar between arms (HR ≈ 1)
✗ = method not appropriate

12. 3. Summary of key assumptions for switch adjustment methods
ITT
Switch treatment ineffective
Exclude switchers
No confounders (unlikely)
Censor switchers
Time varying covariate
IPCW
No unmeasured confounders
2-stage
No unmeasured confounders (stronger assumption than IPCW as fewer covariates in model)
RPSFTM
Constant treatment effect
IPE
Parametric distribution

13. A D B E C F The 6 core analyses
Produce as routine to understand data and switch patterns
When switch occurred relative to randomisation, progression, stopping randomised treatment, death/censoring, e.g. via patient profile plots
Number/% switched per arm, overall & of patients eligible for switch (e.g. exclude censored/died without prog) (IPCW/2-stage unreliable if too high)
Control arm patient characteristics split by switching status, to determine covariates that influence switch
Analyse covariates that influence survival across all patients in the study regardless of treatment
Compare control arm switchers and non-switchers for endpoints linked to survival but not switch-affected, e.g. progression-free survival
Summarise time on and off treatment by randomised arm (if similar, RPSFT/IPE unreliable) A D B E C F

14. Patient profile plot
Example – control arm patients

15. Assessing the assumptions: No confounders
Naïve methods
Determine if core Analysis C (covariates that influence switch) and Analysis D (covariates that influence survival) find any of the same covariates
Determine if core Analysis E (non-switch affected endpoints linked to survival) shows a difference between switchers and non-switchers
Ask a medical expert if disease progression or other variables are likely confounders.

16. Assessing the assumptions: No unmeasured confounders
IPCW, 2-stage
Determine if Analysis E (non-switch affected endpoints linked to survival) adjusted for the covariates (measured potential confounders) in the statistical model shows a difference between switchers and non-switchers
Review patient profile plots from Analysis A for large time gaps between last key covariate data collection and switch (concern for IPCW,) or secondary baseline and switch (concern for 2-stage)
Ask a medical expert and review literature for potential unmeasured confounders

17. Assessing the assumptions: Constant/common treatment effect
RPSFT, IPE
This is the hardest assumption to assess quantitatively, so seek medical expert opinion.
“Tipping point” sensitivity analysis relaxing this assumption – set the treatment effect in switchers to be smaller than for those initially randomized, until the adjusted survival gets close to the ITT result. Determine if that reduction in effect for switchers is clinically plausible.
For trials with previous interim analyses, determine if the switch adjusted treatment effect from these methods is different for earlier and later analysis times
Assuming no unmeasured confounders holds, determine if the treatment effect from second stage of 2-stage model is different to that from RPSFTM/IPE

18. Summary – choice of switch adjustment method
First things first
Be clear on the decision problem(s) and target audience(s)
Switch adjustment is not always necessary
Define switch treatment carefully
Determine what data were collected
No method is universally “best”
Be methodical and justify your choice
Can the model be fitted with available data?
If yes, is the model appropriate given the switching mechanism?
If yes, are the assumptions reasonable?
If yes, are the results plausible?

19. References
Demetri GD et al. Complete longitudinal analyses of the randomized, placebo-controlled, Phase III trial of sunitinib in patients with gastrointestinal stromal tumor following imatinib failure. Clin Cancer Res 2012; 18:
Fukuoka M et al. Biomarker analyses and final overall survival results from a Phase III, randomized, open label, first-line study of gefitinib versus carboplatin/paclitaxel in clinically selected patients with advanced non-small cell lung cancer in Asia (IPASS). J Clin Oncol 2011; 29(21):
Watkins C et al. Adjusting overall survival for treatment switches: Commonly used methods and practical application. Pharm Stats 2013 Nov-Dec;12(6):348-57
Latimer NR and Abrams KR. NICE DSU Technical Support Document 16: Adjusting survival time estimates in the presence of treatment switching. (2014). Available from
Robins J and Finkelstein D. Correcting for Noncompliance and Dependent Censoring in an AIDS Clinical Trial with Inverse Probability of Censoring Weighted (IPCW) Log-Rank Tests. Biometrics 2000; 56(3):
Robins JM and Tsiatis A. Correcting for non-compliers in randomised trials using rank-preserving structural failure time models. Communications in Statistics - Theory and Methods 1991; 20:
White IR et al. strbee: Randomization-based efficacy estimator. The Stata Journal 2002; 2(2):
Branson M, Whitehead J. Estimating a treatment effect in survival strudies in which patients switch treatment. Statistics in Medicine 2002; 21:2449–2463.
Watkins CL, Latimer N, Wang J, Wright EJ. Guidance on selecting appropriate methods when considering adjusting overall survival for treatment switch in oncology studies. Value in Health 2016; 19(7): A398

20. Q & A