Euan MacPherson

Report
Treatment Switching and
Overall Survival in
Oncology
PSI Conference 2014
Euan Macpherson
AstraZeneca
Acknowledgements: Claire Watkins
Disclaimer
Euan Macpherson is an employee of AstraZeneca
LP. The views and opinions expressed herein are
my own and cannot and should not necessarily
be construed to represent those of AstraZeneca
or its affiliates.
2
Euan Macpherson | May 2014
Background
Payer Requirements, Regulatory Requirements, Clinical Trial Design
Payer
Requirements
This is what
payers really
want
Most likely
with some
evidence
of this as
well
3
Euan Macpherson | May 2014
Regulatory &
Trial Design
Requirements
Often, this is
an acceptable
endpoint for
approval
From a
regulatory
approval point
of view these
factors are
also attractive
Demonstrating Overall Survival Benefit in
Oncology Clinical Trials
Challenges
Overall survival is a hard, indisputable endpoint
BUT
• duration may be too long to be practical
• active treatment received as a later line of
therapy – spontaneous crossover or switching
•dilution of any real treatment effect
4
Euan Macpherson | May 2014
A Case Study in NSCLC
Randomisation
Experimental
Therapy
Standard of Care
(SOC)
Progression Free
Survival
Overall Survival
5
Euan Macpherson | May 2014
Available later line treatments
already include the
experimental therapy and
other agents with same
mechanism of action
ITT OS Results
HR: 0.78
95% CI (0.50, 1.20)
P-value = 0.26
• 49% SOC patients subsequently received a treatment of same mechanism of action
as Experimental Therapy
• ITT estimate of OS effect of Experimental vs SOC likely to be confounded by
switch treatment
• Question of interest:
6
- to estimate overall survival benefit of experimental therapy if later
therapy with same mechanism of action had not been available to
SOC patients
Euan Macpherson | May 2014
Simple Analysis Methods To Adjust for
Treatment Switch
Disadvantages
Method
Disadvantage
Intention To Treat
analysis
Compares efficacy in treatment groups as
randomised but if experimental treatment
efficacious in later line therapy, analysis will be
biased in favour of control arm
Exclude switchers from
control arm
Assumes control arm switchers and nonswitchers have the same prognosis – i.e. there is
no confounding between treatment switch and
survival, an unlikely assumption leading to bias.
Breaks randomisation, post-randomisation
selection of treatment groups.
Censor switchers at time
of switch
Standard analysis assumes censoring is
independent of the outcome, an unlikely
assumption, leading to bias
Time-varying covariate
7 Euan Macpherson | May 2014
for treatment or switch
Also relies on the no unmeasured confounders
assumption, unlikely to hold, leading to bias.
Case Study Survival Data
Control Data By Switch Status
CAUTION: COMPARISONS NOT ON A RANDOMISED
BASIS - SUBJECT TO SELECTION BIAS
Switch
Experimental
Non Switch
• Direct treatment comparisons for non-switchers versus Experimental are not on
a randomised basis
• Hazard ratios would be subject to selection bias and not directly interpretable
8
Euan Macpherson | May 2014
Complex Methods
Introduction
• More
sophisticated methods required to
overcome the limitations of simple, naive
methods
• 2 methods will be briefly described here
• Rank Preserving Structural Failure Time
Model (RPSFTM)
• Inverse Probability of Censoring Weighting
(IPCW)
9
Euan Macpherson | May 2014
Complex Methods (1)
Rank Preserving Structural Failure Time
• What would survival have been if active therapy not available?
Observed Overall Survival
Time Off Active Therapy
Time On Active Therapy
Multiply this by acceleration factor
(< 1 if active therapy extends life )
Time Off Active Therapy
Accelerated
Survival
Counterfactual Overall Survival
Assumption: Active treatment effect constant regardless of when
given
Arms balanced due to randomisation
10
Euan Macpherson | May 2014
Choose factor to
balance survival
across both
treatment arms
Compare
counterfactual from
control vs actual OS
from active
Rank Preserving Structural Failure Time
Re-censoring
• Assumption of non-informative censoring in
time-to-event analysis
• Whether a patient switches or survives long
enough to be censored both dependent on
prognosis
• Without recensoring, argument that patients
who actually received less active treatment after
switching and died will be over-represented and
cause bias
• Re-censoring can lead to loss of many events
in RPSFT analysis
• See Korhonen et al for strategies to mimimise this when
planning a study
11
Euan Macpherson | May 2014
Complex Methods (2): IPCW (weight non-switched times)
Control arm survival
Compare to observed experimental arm survival
Observed (ITT) control arm
IPC weighted
WEIGHT
Non
switchers
Non
switchers
WEIGHT
S
Switchers
Switchers
S
Key
Death time
Censor time
S Switch time
12
WEIGHT
WEIGHT
Assumption: The variables in the weight
calculation fully capture all reasons for
switching that are also linked to survival
Weights represent how “switch-like” a patient is that has not yet switched
Calculated using observational propensity score methods, vary by time and patient
Euan Macpherson | May 2014
Slide by Claire Watkins
IPCW
More on Calculation of Weights
• Propensity score model based on pooled logistic regression
- Analogous to time dependent cox regression
• All confounding factors predictive of switching and survival
outcome must be included
• Weights for each subject at time t are basically
Wi (t ) 
P(not switched at t | baselinecovariates)
P(not switched at t | tim e dep' d & baselinecovars)
• Pseudo control population that would have been observed
without switch constructed by weighting the contribution of
patient not yet switched by Wi(t)
• Larger weight given to patients similar to switchers but not
yet switched (or censored for survival)
13
Euan Macpherson | May 2014
Case Study Data
Summary of Results From Complex Methods
1
Observed Experimental vs Control
Control ITT
C RPSFTM
C IPCW
.25
0
0
5
10
15
20
25
Months
Number at risk
Month
0
5
10
15
20
25
Experimental
132
125
94
47
17
0
C ITT
129
121
83
39
15
0
C RP
129
88
2
0
0
0
128.97
103.32
42.42
9.51
5.20
0
C IPCW
14
Euan Macpherson | May 2014
95% CI
0.500-1.202
RPSFTM adjusted OS, all E
vs all C without switch
0.36
0.062,2.095
IPCW adjusted OS, all E vs
all C, without switch
0.611
0.362, 1.031
ITT OS, all E vs all C
Survivor function
.5
.75
Experimental
Hazard ratio
0.776
•After adjusting for treatment
switch, both methods
suggest enhanced
(numerical) treatment benefit
in favour of Experimental
• Challenges in applying
methods noted (next slide)
Challenges In Complex Methods To Adjust For
Treatment Switch
• Retrospective application of methods in the Case Study
• Rank Preserving Method
• Loss of events and power due to “recensoring”
- A process to avoid biasing conclusions too much towards
early events
• Difficulty in communicating the interpretation of this
• Inverse Probability of Censoring Weighting
• Concern that we don’t have the data on all confounders
- Data collection stopped at progression of disease
- Quality of life, tumour growth data no longer collected
- Missing data predictive of both switching treatment or death
• Upfront planning could have mitigated some issues
15
Euan Macpherson | May 2014
Software
• Rank Preserving Structural Failure Time Model
• strbee package in stata
• Not easily applied currently in SAS, R – need for
development
• Inverse Probability of Censoring Weighting
• Implemented using pooled logistic regression modelling in
Stata
• Can be done in SAS
16
Euan Macpherson | May 2014
Considerations For Planning Future Studies
1. Define what is considered to be a crossover treatment
2. Consider if long term treatment effect in absence of crossover is
a relevant question for payers (or regulators)
3. Assess the potential for spontaneous crossover
4. Assess the case for built in crossover
•
•
Patient perspective
Payer (society) perspective
5. Consider randomized controlled and observational trial design
and analysis options to answer the relevant question
17
Euan Macpherson | May 2014
References
Watkins C, Huang X, Latimer N, Tany Y, Wright EJ. Adjusting overall survival for treatment switches:
commonly used methods and practical application. Pharm Stat. 2013 Nov-Dec;12(6):348-57.
Morden JP et al. Assessing methods for dealing with treatment switching in randomised controlled trials: a
simulation study. BMC Medical Research Methodology 2011, 11:4
Robins JM and Finkelstein DM. Correcting for noncompliance and dependent censoring in an AIDS clinical
trial with Inverse Probability of Censoring Weighted (IPCW) Log-Rank tests. Biometrics 2000 (56) 779-788
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; 20:2609-2631.
P.Korhonen, E. Zuber, M. Branson, N. Hollaender, N. Yateman, T. Katiskalahti, D. Lebwohl & T. Haas (2012)
Correcting Overall Survival for the Impact of Crossover Via a Rank-Presevring Structural Failure Time
(RPSFT) Model in the RECORD-1 Trial of Everolimus in Metastatic Renal-Cell Carcinoma, Journal of
Biopharmaceutical Statistics, 22:6, 1258-1271
18
Euan Macpherson | May 2014

similar documents