Lead Investigator: Ulrich Mansmann, Ludwig-Maximilian’s University München
Title of Proposal Research: Application of Estimands and Causal Reasoning in the Reanalysis of Endpoint Measures in Advanced/Metastatic Non-Small Cell Lung Cancer Clinical Trials
Vivli Data Request: 7384
Funding Source: Amani AlTawil is a scientists at the IBE LMU with a fulltime working contract with the University of Munich which ends 31.12.2024; Prof. Ulrich Mansmann has a permanent position at the University of Munich; Prof. RM Huber is retired
Potential Conflicts of Interest: Dr. Al Tawil was an employee at AstraZeneca from 2014 to 2016 & from 2018 to 2020. Dr. Huber reports potential conflicts of interest: Clinical trials with therapeutics in Thoracic Oncology, Advisory boards for companies, including Roche, Takeda and AstraZeneca. All conflicts of interest will be disclosed and mitigated as deemed necessary. conflicted individuals will not be involved in decision-making functions.
Summary of the Proposed Research:
Lung cancer (LC) is the most frequently diagnosed cancer and the leading cause of cancer-related deaths globally. Non-Small Cell Lung Cancer (NSCLC) is the most common subtype, accounting for 85–90% of LC cases.
The effect of treatment is assessed from randomized clinical trials that aim to describe how the outcome of treatment compares to what would have happened to the same subjects under different treatment conditions. Analysis of clinical trials could be done using different strategies. One way is the application of the intention to treat (ITT) analysis. The other way is the Per-Protocol (PP) analysis. Both of these
methods have strengths and flaws.
The ITT analysis is a strategy for analyzing data where subjects initially allocated to a treatment group should be followed up and analyzed as members of that group whether or not they completed the intervention given to that group. This is achieved by ignoring changes in treatment caused by intercurrent events (ICEs). ICEs are events occurring after treatment initiation that affect either the interpretation or the existence of measurements associated with the clinical question of interest. Some examples of ICEs are switching from placebo to active treatment or death. Applying the ITT principle, may not certainly result in estimating the causal effect of receiving the treatment but rather the effect of assignment to intervention. The ITT estimate often gives a pragmatic and conservative estimate of the causal effect influenced by the effects of the so called ICEs.
The other way of analyzing clinical trials is to exclude patients who did not comply with the Clinical trial protocol, this is called ‘per-protocol’ (PP) analysis. The PP population is a specific subset of the trial population who complied with the protocol. The selection of this subset of population for the analysis though many times is of clinical interest causes disruption of randomization and may introduce confounding resulting in biased effect estimates.
Treatment switching after first progress in cancer trials is common. Treatment switching and other ICEs though unavoidable can introduce complexities in estimating treatment effects for longer-term effects, most notably overall survival (OS). Hence, it is often desirable to adjust OS estimates to reflect what would have been observed had control group patients not switched treatments.
The International Conference on Harmonization (ICH), a global organization that issues and maintains numerous guidance documents, has recognized that a precise definition of the scientific question of interest is required to ensure alignment between trial objectives, trial design, data collection, analysis and interpretation. Recently, the ICH issued the final version of an addendum which introduces a new framework that precisely describe the treatment effect of interest (Estimand). Estimand is the quantity estimated by a statistical analysis to address the scientific question of interest posed by the trial objective. It is of interest to operationalize this abstract concept to specific clinical trial settings. We plan to operationalize different estimand concepts for OS in NSCLC clinical trials regarding treatment switching along with other ICEs. We aim to reanalyze the outcomes in advanced/metastatic NSCLC patients using phase III randomized clinical trials following the Estimands framework and based on causal statistics.
To meet our research objectives a multidisciplinary team will be involved in the discussion around all possible ICEs and suitable strategies to account for these events. We as an academic institution with statistical background will be mainly interested in applying methodological best practice approaches in clinical trials based on causal reasoning. To comply with the estimand concept, a clinician will be looking if the clinical problem at hand is transformed into the correct methodological concept.
Until now, academic oncology doesn’t consider routinely estimands as an opportunity. The criticality of the disease, its poor prognosis and the costs of innovative therapies request a careful methodological approach to provide analyses that correctly answer relevant clinical questions regarding the effects of novel therapies. By developing specific objectives relating to evaluating the patient’s perspective of treatment and embracing the estimand framework, we can guarantee useful collection of data and deliver clear interpretation and conclusions to all stakeholders, thereby providing clear evaluation on the impact of treatment on patients’ lives.
Statistical Analysis Plan:
Reason for study selection:
Due to our research interests, we selected studies from the vivli.org website using the PICO search strategy by Participants (P). We chose “Non-small cell lung cancer” as the targeted population. Using the advanced search, we further selected from the study design filter “parallel studies” and excluded cross-over and single group. We also selected studies that were at stages of development III from the study phase filter and excluded not applicable (NAs) and earlier phases and phase IV because our objectives are related to the efficacy of drugs instead of their pharmacokinetics, pharmacodynamics, or safety properties. Our search retrieved 43 studies in total. For the selected 43 studies, the study links at clinicaltrials.gov were checked to ensure that the studies have measured as primary endpoints at least one of our efficacy outcomes listed in our objectives. From the 43 studies, we wanted to only include advanced/metastatic NSCLC, “targeted therapies” & “Immunotherapy” as the intervention of interest and first-line treatments. The final number of selected studies is 5.
Objective one:
Estimation of the overall treatment effect (Overall survival (OS), progression-free survival (PFS), Quality of life measures (QoL)) while accounting for Intercurrent events using the below Estimands strategies:
Analysis: Primary
Population: Advanced/metastatic NSCLC
Variable: OS: Time between randomization and the date of death from any cause until data cut-off
Intervention: Apply the below strategies to account for intercurrent events:
• Hypothetical Strategy for Cross Over, Prohibited medication, subsequent therapy, changes in background treatment, Administrative Loss to follow up
• Treatment policy Strategy for loss of efficacy (LoE) /Adverse events (AE) / Rescue medication
Effect Summary measure: Adjusted Hazard ratio (HR) of OS and log-rank test stratified by mutation.
Rationale: Hypothetical: Interest would be in treatment effect attributable to treatment not confounded by other biological treatment
Analysis: Primary
Population: Advanced/metastatic NSCLC
Variable: Qol: Time to deterioration
Intervention:
• Hypothetical Strategy for cross over
• Composite Strategy for Death
• Treatment policy Strategy for AE/ Rescue medication
Effect Summary measure: Joint Modelling of Cox regression
Rationale: Hypothetical: What outcomes would have been observed if patients had survived and not progressed for one year
Analysis: Primary
Population: Advanced/metastatic NSCLC
Variable: PFS: Time between randomization and the date of progression or death until data cut-off
Intervention:
• Hypothetical Strategy for Cross Over, prohibited medication, subsequent therapy, changes in background treatment
Effect Summary measure: Adjusted HR of PFS and log-rank test stratified by mutation
Rationale: Hypothetical: Interest would be in treatment effect attributable to treatment not confounded by other biological treatment
Estimand for OS
What is the treatment effect of targeted therapies/immunotherapies compared to chemotherapy or other targeted therapy on OS in patients with stage IV NSCLC regardless of treatment discontinuation & use of rescue for any reason and in the hypothetical situation of no prohibited medication, subsequent medication and cross over?
Estimand for PFS
What is the treatment effect of targeted therapies/immunotherapies compared to chemotherapy or other targeted therapy on PFS in patients with stage IV NSCLC regardless of treatment discontinuation & use of rescue for any reason and in the hypothetical situation of no prohibited medication, subsequent medication and cross over?
Estimand for Time to deterioration
What is the treatment effect of targeted therapies/immunotherapies compared to chemotherapy or other targeted therapy on QoL measures in patients with stage IV NSCLC regardless of treatment discontinuation & use of rescue for any reason and in the hypothetical situation of no prohibited medication, subsequent medication, cross over, death or progression?
Plan, to avoid bias:
• Selection of well-designed randomized controlled trials
• Adjustment of baseline covariates.
• To avoid bias introduced by ICEs regarding the treatment effect on Qol, the concept of joint modelling will be applied. Joint models allow a simultaneous look on time course data of pain and events over time that let the further measurements of pain impossible. This analysis will model the time course of Qol together with tumour-related events and mortality (PFS, OS). A U-shaped curve models the time course of pain, allowing for a decrease and later increase of pain intensity. The estimator derives from the difference between the areas under the curves (AUCs) specified by the main effects of the joint modelling for the longitudinal data over one year.
Objective two
To put the PP analysis into a causal framework, we will describe the specific trial setting by a causal graph to understand pre-and post-randomization confounders as well as the time structure. We will use the graph to define the respective methodological approach in terms of inverse probability weighting, the G-formula, or the g-estimation of structural nested models to account for measured time-varying confounders in the presence of treatment-confounder feedback.
The application of 2-stage adjustment method requires that the data for all potential confounders are collected at the time of disease progression, i.e., secondary baseline data. Thus, our analysis will be restricted to requested studies where secondary baseline data are collected. Having said so, we will not apply the 2 stage method analysis to ALTA-1L Study.
To increase the probability that PP analysis yields a valid estimate of the PP effect, we will follow the following general rules:
• Data from participants should not be censored upon treatment withdrawal for clinical reasons
• Data from participants should be censored when participants stop their treatment
• Adjustment to account for confounding due to incomplete adherence should be done.
Plan, to avoid bias:
For treatment response prediction, we will use only RCT data.
Objective three
The relationship between adverse events and QoL evolves within a complex time structure. Regarding possible confounding the problem studied in objective 3 resembles the setting studied under objective two. Therefore, methodological guidance comes from the experience in handling objective 2. We will apply comparable methodological strategies (formulating a causal graph, defining specific regression models based on the graph) within the time structure of the trial (when did the AE happen, when was the QoL measured).
Plan, to avoid bias:
For treatment response prediction, we will use only RCT data.
General aspects concerning all objectives:
• Data Management will be planned and performed by the data manager after having received the data
• When needed, missing values will be handled as part of the statistical analysis and sensitivity analyses depending on the choice of estimand.
• To conduct our proposed analyses, we require a lot of clinically relevant variables associated with NSCLC. We are convinced that all our criteria regarding data quantity and quality are met with this clinical data application.
Requested Studies:
A Phase 3 Multicenter Open-label Study of Brigatinib (AP26113) Versus Crizotinib in Patients With ALK- positive Advanced Lung Cancer
Data Contributor: Takeda
Study ID: NCT02737501
Sponsor ID: AP26113-13-301
A Phase III, Open-Label, Multicenter, Randomized Study Evaluating the Efficacy and Safety of Atezolizumab (MPDL3280A, Anti-PD-L1 Antibody) in Combination With Carboplatin+Paclitaxel or Atezolizumab in Combination With Carboplatin+Nab-Paclitaxel Versus Carboplatin+Nab-Paclitaxel in Chemotherapy-Naive Patients With Stage IV Squamous Non-Small Cell Lung Cancer
Data Contributor: Roche
Study ID: NCT02367794
Sponsor ID: GO29437
A Phase III Multicenter, Randomized, Open-Label Study Evaluating the Efficacy and Safety of Atezolizumab (MPDL3280A, Anti-PD-L1 Antibody) in Combination With Carboplatin+Nab-Paclitaxel for Chemotherapy-Naive Patients With Stage IV Non-Squamous Non-Small Cell Lung Cancer
Data Contributor: Roche
Study ID: NCT02367781
Sponsor ID: GO29537
Randomized, Multicenter, Phase III, Open-Label Study of Alectinib Versus Crizotinib in Treatment-Naive Anaplastic Lymphoma Kinase Positive Advanced Non-Small Cell Lung Cancer
Data Contributor: Roche
Study ID: NCT02075840
Sponsor ID: BO28984
A Phase III Randomised, Double Blind, Placebo Controlled, Parallel, Multicentre Study to Assess the Efficacy and Safety of Continuing IRESSA 250 mg in Addition to Chemotherapy Versus Chemotherapy Alone in Patients Who Have Epidermal Growth Factor Receptor (EGFR) Mutation Positive Locally Advanced or Metastatic Non-Small Cell Lung Cancer (NSCLC) and Have Progressed on First Line IRESSA
Data Contributor: AstraZeneca
Study ID: NCT01544179
Sponsor ID: D791LC00001
Public Disclosures:
Al Tawil, A., McGrath, S., Ristl, R. and Mansmann, U., 2024. Addressing treatment switching bias in the ALTA-1L trial with g-methods: exploring the impact of model specification. BMC Medical Research Methodology. Doi: 10.1186/s12874-024-02437-6