1. Causal inference with RCTs of chemotherapy
Thank the audience for joining this session
Thank the organizers for the opportunity of presenting this work
the importance of well-documented side-effects
C. Lancia, Leiden Univ. Medical Center

2. Summary
Question
Effect of treatment adaptations on mortality in osteosarcoma, and cancer treatment in general
Problem
Treatment-adjustment bias, i.e. toxicities are time-dependent confounders
Mehods
Data from MRC BO06, RCT of chemotherapy in osteosarcoma
Inverse Probability of Treatment Weighing (IPTW) [Hernán & Robins, J Epi Comm, 2006]
Adjusted Kaplan-Meier estimators (AKME) [Therneau, Crowson & Atkinson, 2015]

3. MRC BO06
The data I am going to present next were collected during the Randomized Clinical Trial MRC BO06
Patients were randomised to two regimens having the same cumulative dose for both Doxorubicin, here represented by an A, and Cisplatin, represented by a P
The only difference in the two regimens is the intended duration,
From the first to the last course, Regimen 1, the control arm, was intended to last 19 weeks in the control arm, while Regimen 2, the dose-intense arm, was intended to last 14 weeks
Thus, the intended difference in treatment time was 5 weeks

4. Treatment trajectories

5. Treatment adaptations and outcome
Event-free survival in a landmark analysis at 180 days since randomisation
Patients that completed all 6 cycles of BO06
Treatment adaptations: dose reduction of at least 0.15 standardized units
We use the number of full cycles received as a proxy of the actual dose received
We are interested in its relation with event-free survival
Note: the intended dose received in BO06 is the same in the two arms!
The idea of this reweighing is that, the decision of reducing the following cycle is independent of the toxicities reported provided that these are sufficiently severe

6. Treatment-adjustment bias
This picture shows the so-called treatment-adjustment bias
Treatment administration, in terms of cycle delay and course reductions, predict future side effects
The latter predict in turn future delays and reductions
Everything is a risk factor for the outcome
Therefore, toxicities are time-dependent confounders for the effect of treatment adaptations on survival

7. Inverse Probability-of-Treatment Weighing
IPTW reweighs each patient with the proability of observing a treatment adaptation 𝐴 𝑡 given the reported toxicities 𝑳 𝑡 and his/her baseline characteristics 𝐵
In formulas, the ith patient contributes with 𝑠 𝑤 𝑖 = ∏ 𝑡=1 𝑇 P A 𝑖,t 𝐴 𝑖,𝑡−1 ) P A 𝑖,t 𝐴 𝑖,𝑡−1 , 𝑳 𝑖,𝑡−1 , 𝐵 i ) copies of himself
This techinque creates a pseudopopulation where
the effect of treatment adapatations on mortality is the same
treatment is adapted independently of the reported toxicities
Why is it appealing: a crude analysis on this pseudopopulation enables causal inference because in absence of confounding, association means causation
Inverse Probability-of-Treatment Weighing is a technique for reweighing patients with the inverse probability of observing the allocated treatment given the past medical history and the baseline characteristics
Theoretically, IPWT creates a pseudopopulation where, in each strata of baseline variable and toxicity severity, treatment adaptations are independent of treatment side-effects
Further, the effect of treatment adaptations on survival is the same as in the original population
As such, a crude analysis on this pseudopopulation enables causal inference as in absence of confounding, associations means causation

8. The role of side effects
The model for adaptations given the side effects is a crucial aspect of IPTW
Different modelling strategies can lead to completely different results
Simulations show that unfortunate modelling choices can yield bias enhancement
Attention: If toxicities are not well-documented, the association between side effects and therapy adaptations can be distorted; this might lead to uncorrect reweigthing
Attention: if toxicity and treatment histories are not recorded at the same timepoints, the causal dependencies might not be accurately captured and many modelling choices will be dictated by the less informative history

9. Construction of IPTW
Weights: pooled logistic regression for the first reduction of 0.15 standardized units
Patients are weighted with the regression until the first reduction
Acute toxicity: any occurrence of
Low WBC (CTCAE ≥ 3)
Low plateletes (CTCAE ≥ 3)
Infection (CTCAE ≥ 3)
Mucositis (CTCAE ≥ 3)
Cardiotoxicity (CTCAE ≥ 3)
Ototoxicity (CTCAE ≥ 3)
Baseline confounders: Age (paediatric/adult patient) and regimen (2-weekly/3-weekly)
The idea of this reweighing is that, the decision of reducing the following cycle is independent of the toxicities reported provided that these are sufficiently severe

10. Crude vs Adjusted Kaplan-Meier
EFS at two years is 2.5 years is basically the same in the right plot
Possible confounding left
Treatment duration
Patients’tolerability

11. Conclusions
The effect of the delivered treatment on survival is confounded by toxicity
IPTW + AKME is an attempt of dealing with such confounding
Well-reported side effects are the key to the application of IPTW
Reporting only the most severe grade over a cycle or a period masks causal relationships
Addressing the effect of delays on outcome requires detailed reporting of myelotoxicity
Patient’s tolerability is an important aspect that is currently left out
It could be possibly included by stratifying according to the reported toxicities

12. Acknowledments Marta Fiocco, Leiden University Medical Center
Dr. Jakob Anninga, Radboud University Nijmegen
Cristian Spitoni, University Medical Center Utrecht
Matthew R. Sydes, Medical Research Council
Gordana Jovic, Medical Research Council
Dr. Jeremy Whelan, University College London Hospital
This work was supported by Dutch Foundation KiKa