Impact of concomitant proton pump inhibitors in patients treated for advanced or metastatic NSCLC: a target trial emulation protocol

Abstract

Background

Proton pump inhibitors (PPIs) are commonly coprescribed in patients with advanced or metastatic non-small-cell lung cancer (NSCLC), including those treated with tyrosine kinase inhibitors (TKIs) or immune checkpoint inhibitors (ICIs), often without a well-defined indication. PPIs may reduce TKI bioavailability by increasing gastric pH and compromising ICI efficacy through alteration of the gut microbiota. Observational studies have reported an association between concomitant PPI use and poorer survival in these patients. However, causal inference remains limited by confounding and time-related biases, and conducting a dedicated randomized trial is unlikely to be feasible. Target trial emulation provides a structured methodological framework to estimate causal effects using real-world data (RWD).

Materials and methods

Two target trials will be emulated using RWD from the nationwide UNICANCER Epidemiological Strategy and Medical Economics (ESME)-Lung Cancer database, linked to the French National Health Data System. Advanced or metastatic NSCLC patients initiating first-line treatment with a TKI or an ICI between 2015 and 2024 could be included. Two strategies will be compared: initiation of a TKI or an ICI with concomitant PPI dispensation within a 12-week grace period versus without PPIs during this period. Time zero is defined as TKI or ICI initiation. The primary endpoint is overall survival; secondary endpoints include real-world progression-free survival and real-world time to next treatment. Causal effects will be estimated using a cloning, censoring, and weighting approach with inverse probability of censoring weights to address confounding and immortal time bias. Sensitivity analyses will include G-computation, quantitative bias analysis, and E values.

Highlights

• •Survival impact of PPIs use in NSCLC patients under TKIs or ICIs remains unclear.
• •We emulate a target trial using RWD to estimate the causal effect of PPI use on OS.
• •Advanced causal inference methods are used to robustly address observational biases.
• •Sensitivity analyses will assess robustness to residual confounding.

Introduction

Proton pump inhibitors (PPIs) are widely prescribed, particularly to patients treated for cancer with tyrosine kinase inhibitors (TKIs) or immune checkpoint inhibitors (ICIs), especially in non-small-cell lung cancer (NSCLC). Their use is often justified by gastrointestinal symptoms or for gastroprotection, sometimes without formal indication, thereby exposing patients to a potential risk of drug interactions.¹ TKIs, which are orally administered, are weak bases and exhibit reduced bioavailability when gastric pH is increased by PPIs, potentially decreasing absorption and compromising therapeutic efficacy.2, 3, 4 Furthermore, chronic use of PPIs alters gut microbiota composition, a key factor in the response to immunotherapies targeting programmed cell death protein 1 (PD-1)/programmed death-ligand 1 (PD-L1).⁵ These alterations are associated in unplanned retrospective analysis with a substantial reduction in progression-free survival (PFS) and overall survival (OS) in NSCLC patients treated with TKIs or ICIs.6, 7, 8

Numerous observational studies and post hoc analyses of randomized controlled trials (RCTs) have shown an association between PPIs coprescription and reduced survival outcomes.^9,10 However, causal interpretation remains challenging because patients receiving PPIs are often older, more comorbid, and have poorer performance status, raising concerns about residual confounding. Conducting a dedicated RCT to address this question would be ethically and practically questionable.

To overcome these limitations, the target trial emulation (TTE) framework has been proposed by Hernán and Robins as a principled approach to estimate causal effects from real-world data (RWD) by explicitly specifying and emulating the protocol of a hypothetical randomized trial.^11,12 By clearly defining eligibility criteria, treatment strategies, time zero, causal estimands, and follow-up, TTE aims to minimize common biases, including confounding, selection bias, and immortal time bias.

The current proposal aims to apply a TTE methodology on RWD from the Epidemiological Strategy and Medical Economics-Lung Cancer (ESME-LC) database, a large French database including >50 000 patients with advanced or metastatic NSCLC in its September 2025 release. This approach allows for the replication of a randomized trial protocol from observational data while rigorously addressing key biases, including confounding, selection bias, and immortal time bias. Recently, we established proof of concept that this methodology can be appropriately applied in oncology to emulate nine pivotal RCTs using data from the ESME- metastatic breast cancer (MBC) database.^13,14 Our objective is to evaluate the causal impact of PPIs coadministration on survival in NSCLC patients receiving first-line TKIs or ICIs. Due to the different mechanisms of action of the two treatments considered, two separate target trials will be defined: one for TKIs and another for ICIs. The anticipated findings could directly influence clinical practice by supporting a more judicious use of PPIs and encouraging the preferential use of less interactive alternatives—such as antacids or H2 receptor antagonists—when appropriate, with the goal of maximizing the effectiveness of targeted anticancer therapies.

Objectives

The aim of this study is to emulate two different target trials based on the UNICANCER ESME-LC real-world database linked with the French National Health Data System (SNDS) database to estimate the causal effect of concomitant PPIs dispensation on the effectiveness of two therapeutic classes commonly used as a first-line treatment in advanced or metastatic NSCLC patients in France:

• •First target trial (A): TKIs targeting epidermal growth factor receptor (EGFR), and anaplastic lymphoma kinase (ALK)/ROS alterations;
• •Second target trial (B): ICIs targeting PD-1 and PD-L1.

Materials and methods

Data sources

Main baseline patient characteristics and outcomes (see below) are collected in the UNICANCER ESME-LC database (NCT03848052), a nationwide database that routinely collects data on adults treated for lung cancer at one of the 38 participating French sites (comprehensive cancer centers and public hospitals). ESME data are derived from patients’ medical records and were collected retrospectively by trained data managers and Clinical Research Associates, with no involvement of the clinicians who managed the patients. Information is recorded as available in medical charts and clinical reports in a standardized manner across sites.¹⁵ The database gathers detailed clinical information on patient and disease characteristics, including molecular testing, treatment patterns, and outcomes. The database is updated annually, incorporating both follow-up data for previously enrolled patients and newly collected data for patients diagnosed during the previous year. The ESME-LC database was first authorized by the French data protection authority in 2017 (initial authorization no. DE-2017-397) and obtained complementary authorization in 2019 in accordance with European regulation. To date, the UNICANCER ESME program has supported >80 international publications, including 22 for the ESME-LC database.¹⁶

The ESME research program collects data from patients diagnosed with NSCLC since 1 January 2015. We will include all patients with diagnosis from 1 January 2015 to 30 September 2024, followed up to the last release of the ESME database (database locked in October 2025). The 2025 ESME-LC database contains 50 982 patients with advanced or metastatic NSCLC. Among them the patients are as follows:

• •Patients initiated a first-line treatment with TKI: gefitinib (n = 765), erlotinib (n = 1026), osimertinib (n = 1673), afatinib (n = 552), alectinib (n = 384), and crizotinib (n = 582) (total number = 5132).
• •Patients initiated a first-line treatment with ICI: pembrolizumab (n = 11 053), atezolizumab (n = 490), and cemiplimab (n = 184) (total number: 11 727).
• •Patients initiated chemotherapy (CT) (± bevacizumab) without TKI or ICI in first-line treatment (exploratory objective): n = 21 686.

Since 2024, UNICANCER has been managing the Health data to foster Meaningful Oncologic Evaluations (HARMONIE) health data warehouse (HDW), which hosts a subsidiary database of the SNDS, along with data from the ESME-LC database. The SNDS data hosted within HARMONIE constitute a comprehensive extraction of the SNDS, covering breast, lung, and ovarian cancers, with a >15-year lookback period. SNDS encompasses information from the French National Health Insurance (Cnam), the Technical Agency for Information on Hospitalization, and the National Institute of Health and Medical Research (INSERM). It contains pseudonymized data on health care reimbursements, hospitalizations, out-of-hospital care, long-term conditions, and medical causes of death. The dataset is updated annually and aims to promote real-life studies to improve understanding of the overall management of cancer patients and the associated health economic burden. The HARMONIE HDW was authorized by the French data protection authority in July 2024 (authorization DT-2024-015).

HARMONIE HDW was considered a fit-for-purpose data source as it was, to the best of our knowledge, the only European database combining (i) high-quality and high-granularity RWD in NSCLC from a nationwide database, namely the ESME database, and (ii) linkage with a general-purpose database containing information on medications not related to oncology, namely the SNDS dataset, which could provide unique and comprehensive information on PPI use (drug names and dispensation date).

The linkage between the ESME database and the SNDS relies on a strict probabilistic matching algorithm to minimize false-positive matches, and the matching rate was above 89% among patients included in the HARMONIE database. Of note, there were no redundant variables between data sources. In case of uncertainty regarding the validity of the match, the patient was excluded from the HARMONIE dataset. All patients were informed about the use of their data, and an independent scientific and ethics committee specifically approved the present work.

Available data

Within HARMONIE HDW, accessible data include (among others) the following: (i) gender, year of birth, place, date, and medical cause of death, (ii) reimbursed outpatient healthcare expenses such as physician visits or procedures and dispensed drugs with drug anatomical therapeutic chemical (ATC) code, doses, date, and quantity dispensed, as well as (iii) information on hospitalization and hospital procedures from the Programme de Médicalisation des Systèmes d'Information (PMSI, French DRG), which will allow probabilistic matching with the ESME-LC cohort.

Baseline patient characteristics collected from the ESME-LC database include the date of diagnosis of advanced/metastatic NSCLC, age at diagnosis, gender, smoking status, Eastern Cooperative Oncology Group performance status (ECOG PS), disease stage, metastasis location (including brain metastases), molecular testing status (EGFR, ALK, and ROS1 mutation and PD-L1 expression), and treatment details in the (neo)-adjuvant setting and in the advanced/metastatic setting.

Patient oncological outcomes collected from the ESME-LC database include the date of tumor progression, the date of death from any cause, and the date of the last medical information.

PPI dispensation data collected from the SNDS include start and end dates of PPI dispensation episodes and the dose of PPI dispensed, considering the following drugs for trial emulation: omeprazole, esomeprazole, pantoprazole, lansoprazole, and rabeprazole. In addition, the modified Charlson comorbidity index¹⁷ will be inferred from the SNDS data, including the following items identified from the relevant International Classification of Diseases (ICD)-10 codes: myocardial infarction, peripheral vascular disease, cerebrovascular disease or transient ischemic attack, dementia, peptic ulcer disease, liver disease, hemiplegia, and acquired immunodeficiency syndrome. The period during which the components of the modified Charlson comorbidity index are selected will be the year preceding the time zero ‘T0’ according to the definition of Bannay et al.¹⁷ The similarity of the distribution of comorbidities and the modified Charlson comorbidity index will be checked.

Study design

We will conduct two French, multicenter observational studies in patients with first-line advanced or metastatic NSCLC: one among patients receiving TKI and one among patients receiving ICI. Both studies aim to compare two independent treatment strategies: (i) first-line TKI (or ICI) with concomitant PPIs dispensation and (ii) first-line TKI (or ICI) without concomitant PPIs dispensation. Both studies are designed within the TTE framework. Figure 1 illustrates the study design, and Table 1 summarizes the key components of the protocols for the target trial and the emulated trials.

Figure 1.

Study design illustrating the two target trials that will be emulated with two arms in each using the RWD available from theEpidemiological Strategy and Medical Economics Lung Cancer (ESME-LC)and SNDS databases. ECOG PS, Eastern Cooperative Oncology Group performance status; ESME-LC, ESME-Lung Cancer; ICI, immune checkpoint inhibitor; NSCLC, non-small-cell lung cancer; PPI, proton pump inhibitor; RWD, real-world data; SNDS, French National Health Data System; TKI, tyrosine kinase inhibitor.

As an exploratory objective, a third TTE will be conducted to assess the potential lack of substantial impact of concomitant PPIs dispensation on OS and real-world progression-free survival (rwPFS) among patients with advanced or metastatic NSCLC who are not receiving first-line treatment with a TKI or ICI, but are treated with CT alone, with or without bevacizumab.

No data analysis has been initiated to date.

TTE

Time zero

Time zero in the analysis is equivalent to the time of randomization in the target trial, defined here as the time at which individuals meet all the eligibility criteria, i.e. at first initiation of a TKI or an ICI as first-line treatment of advanced or metastatic NSCLC.

Treatment strategies

The primary objective is to assess the impact of concomitant dispensation of PPIs in patients with advanced or metastatic NSCLC and initiating first-line treatment∗ with (A) TKI or (B) ICI, among the following:

• •TKI: gefitinib, erlotinib, osimertinib, afatinib, alectinib, and crizotinib.
• •ICI: pembrolizumab, atezolizumab, and cemiplimab.

∗ ‘Administered as a single agent or in combination with other classes of therapy.’

Strategy 1: Patients initiate first-line treatment with a TKI or an ICI and are subsequently dispensed a PPI. A 12-week grace period was selected to reflect real-world clinical practice in which PPIs are frequently initiated shortly after the start of anticancer therapy, rather than exactly at treatment initiation. This duration was considered sufficient to capture the majority of clinically relevant codispensations while remaining early in the treatment course. In advanced NSCLC, most patients remain progression-free during the first 2-3 months of first-line therapy, as reported in pivotal trials.^19,20 Specifically, the first PPI dispensation must occur within 12 weeks after initiation of the TKI or ICI, and PPI use may continue for any duration beyond this window. A single dispensation within the grace period is considered sufficient for inclusion; only the date of the first dispensation is taken into account.

Strategy 2: Patients initiate first-line treatment with a TKI or ICI and do not receive any PPIs during the 12-week grace period after treatment initiation. This strategy represents a treatment pathway in which PPI use is intentionally avoided during the early phase of anticancer therapy to minimize potential drug–drug interactions or microbiome disruption.

Washout period

A 12-week washout period was applied to ensure absence of PPI exposure at or immediately before treatment initiation. This defined window aims to ensure that patients are not under the effect of PPIs at or immediately before the initiation of first-line TKI/ICI therapy, rather than to define a strict new-user cohort. The 12-week window reflects typical PPI dispensing durations in routine practice and provides a conservative buffer to minimize residual exposure at baseline.

Outcomes

The primary outcome of the two TTEs is OS, defined as the time from the date of initiation of TKI or ICI treatment to death from any cause. Patients who are still alive at the date of database release will be censored at the date of the last available medical information. Most of the OS benefit in advanced or metastatic NSCLC is achieved in the first-line setting, which supports the choice of OS as the primary endpoint. Indeed, the current standard of care in first-line depends on the presence of a gene mutation or rearrangement, i.e. TKI for EGFR mutation, ALK rearrangement or ROS fusion, and ICI (alone or in combination with CT) in the absence of a targetable mutation.

Secondary outcomes are rwPFS and real-world time to next treatment (rwTTNT). rwPFS is defined as the time from the date of initiation of TKI or ICI treatment to disease progression or death from any cause, whichever occurs first. Disease progression is defined as the first event retrieved among the following: regional tumoral, lymph node progression, new metastatic site, progression of a previously diagnosed metastatic site, or end of treatment due to the progression of the disease. rwTTNT is defined as the time from the date of initiation of TKI or ICI treatment to the date of initiation of a second treatment line. For these two outcomes, patients alive without progression/subsequent line at the date of database release will be censored at the date of the last medical information. Patients without any progression/subsequent line who died >9 months after the last medical information will be censored at the date of the last medical information. This rule was implemented to enhance the robustness of rwPFS estimation. In a real-world setting, when a patient has no documented progression date and a death recorded long after the last known follow-up, this situation may reflect a patient lost to follow-up, leading to an artificial overestimation of the PFS. This censoring rule is consistently applied in all projects conducted on the ESME database. The 9-month threshold was discussed with members of scientific committees and validated as a reasonable methodological compromise. Regardless, this rule will apply to only a small number of patients and helps secure PFS estimates.

Main protocol components

Table 1.

Summary of the key components of the protocol for the target trial and emulated trials

Main components	Target trials	Emulated trials using HARMONIE HDW
Eligibility criteria	•Histologically or cytologically confirmed NSCLC. •Advanced or metastatic disease diagnosed between 1 January 2015 and 30 September 2024. •≥18 years of age. •ECOG PS of 0 to 2. •No prior PPIs in the 12 weeks before the inclusion. •TKI target trial: no prior TKI for advanced or metastatic NSCLC. •ICI target trial: no prior ICI for advanced or metastatic NSCLC.	•Same as the target trial. •De novo IIIB/IIIC, de novo IV, regional disease progression, or metastatic disease progression. Date of locally advanced or metastatic diagnosis between 1 January 2015 and 30 September 2024. •Same as the target trial. •Last available ECOG PS at TKI or ICI initiation between 0 and 2. •No prior PPIs in the 12 weeks before initiating TKI or ICI. •TKI emulated trial: no prior TKI before the date of locally advanced or metastatic diagnosis. •ICI emulated trial: no prior ICI before the date of locally advanced or metastatic diagnosis.
Treatment strategies	TKI target trial: •Gefitinib (250 mg/day) or •Erlotinib (150 mg/day) or •Osimertinib (80 mg/day) or •Afatinib (40 mg/day) or •Alectinib (1200 mg/day) or •Crizotinib (500 mg/day). ICI target trial: •Pembrolizumab (200 mg IV q3 weeks or 400 mg/q6 weeks) or •Atezolizumab (840 mg IV q 2 weeks or 1200 mg IV q 3 weeks or 1680 mg IV q 4 weeks) or •Cemiplimab (350 mg IV q 3 weeks). In both trials: TKI and ICI are administered in accordance with current recommendations for first-line treatment in advanced or metastatic settings, i.e. until intolerance or disease progression. In arm A, patients may be prescribed a concomitant PPI (omeprazole, esomeprazole, pantoprazole, lansoprazole, or rabeprazole) as part of routine clinical care, typically in response to gastrointestinal symptoms such as epigastric pain, heartburn, or for gastroprotection, rather than as a systematic comedication. In arm B, patients receive first-line TKI or ICI treatment without concomitant PPI dispensation, and alternative non-PPI medications may be used to manage gastrointestinal symptoms when clinically indicated.	TKI emulated trial: •Arm A: As first-line treatment, initiate any dose of TKI (same as for the target trial), and initiate any dose of PPIs within 12 weeks after TKI initiation. •Arm B: As first-line treatment, initiate any dose of TKI (the same as for the target trial), and do not initiate any PPIs within 12 weeks after TKI initiation. ICI emulated trial: •Arm A: As first-line treatment, initiate any dose of ICI (same as for the target trial), and initiate any dose of PPIs within 12 weeks after ICI initiation. •Arm B: As first-line treatment, initiate any dose of ICI (same as for the target trial), and do not initiate any PPIs within 12 weeks after ICI initiation. Arm A of both trials: To better reflect clinical reality, where the actual dispensation of a PPI may not coincide exactly with the time of initiation of a TKI/ICI, the definition of the dispensing period includes a ‘grace period’ within 12 weeks of the TKI/ICI initiation. There is no upper limit to the length of time that PPIs can be prescribed at the same time as TKI or ICI.
Treatment assignment	Study participants are randomly assigned (1 : 1 ratio) to the treatment strategy as baseline in favor of arm A (concomitant administration of PPIs during the treatment of first line) or B (no concomitant administration of PPIs during the treatment of first line).	Randomization is emulated via CCW procedure. Each participant and their clone will be assigned to a different treatment strategy. Clones were censored when they deviated from the strategy of their assigned arm.
Follow-up period	Start at TKI or ICI assignment (time of randomization) until death, end of the study, or loss to follow-up, whichever occurs first (OS).	Time zero of follow-up is the first time a participant meets all eligibility criteria, i.e. the first TKI or ICI initiation. Follow-up ends at the earliest occurrence of death from any cause or the latest news (OS).
Main outcomes	OS is defined as the time from randomization to death from any cause. The patients still alive at the end of the follow-up will be censored at the date of the last medical information. PFS is defined as the time from randomization until tumor progression according to RECIST v1.1,18 or death from any cause. Patients who have not progressed or died at the time of analysis will be censored at the time of their last evaluable RECIST assessment.	OS is defined as the time from the date of TKI/ICI initiation to death from any cause. The patients still alive at the date of database release will be censored at the date of last medical information. rwPFS is defined as the time from the date of TKI/ICI initiation to disease progression or death from any cause, with disease progression being the first event retrieved among the following: regional tumoral or lymph node progression, new metastatic site, progression of a previously diagnosed metastatic site, or the end of treatment due to the progression of the disease. rwTTNT is defined as the time from the date of TKI/ICI initiation to the date of initiation of a new treatment line. Patients alive without progression at the date of database release will be censored at the date of the last medical information. Patients without any progression reported died >9 months after the last medical information will also be censored at the date of the last medical information.
Causal contrast	Intent-to-treat effect: all eligible patients are included in the efficacy analysis according to their treatment assignment. Per-protocol effect: the analysis includes only participants who strictly adhered to the study protocol as initially assigned.	Primary estimand: The per-protocol effect of adherence to the assigned treatment strategy during the 12-week grace period. This estimand represents the contrast in outcomes that would be observed if all individuals adhered to one strategy versus the other and is estimated as a marginal population-level contrast using weighting and standardization approaches, allowing interpretation as the average treatment effect of these strategies. Secondary estimands: The ATT will be explored in sensitivity analyses as complementary causal contrasts, estimated using G-computation under alternative modeling assumptions.
Analysis plan (key points)	Survival outcomes distributions (OS and PFS) are estimated by KM analysis. The comparison of survival outcomes between treatment arms will be analyzed using an univariable Cox model.	The same as the target trial. The primary causal analysis will rely on the CCW approach to manage scenarios in which a participant’s treatment strategy is indistinguishable at time zero, with IPW to account for informative censoring and adjust for baseline confounders. Confounders will be identified from DAGs. Sensitivity analyses will be carried out to account for potential residual confounding bias through G-computation. Multiple imputation will be used to impute missing data for variables of interest using missMDA package.

ATT, average treatment effect on the treated; CCW, cloning, censoring, and weighting; DAGs, directed acyclic graphs; ECOG PS, Eastern Cooperative Oncology Group performance status; HDW, health data warehouse; ICI, immune checkpoint inhibitor; IPW, inverse probability weighting; KM, Kaplan–Meier; NSCLC, non-small-cell lung cancer; OS, overall survival; PFS, progression-free survival; PPI, proton pump inhibitor; rwPFS, real-world progression-free survival; rwTTNT, real-world time to next treatment; TKI, tyrosine kinase inhibitor.

Data analysis

Population description

We will first describe the baseline characteristics (disease characteristics, clinical and biological data, etc.) of our entire population. Quantitative variables will be described using the number of observations, mean, standard deviation, median, minimum, and maximum values. Qualitative variables will be described using frequency and percentage. The number of missing data will be given, but will not be considered for the calculation of proportions. This description will then be repeated separately for the TKI and ICI populations.

Identification of confounders

The identification of potential confounders will be supported by existing literature and clinical expertise, and the corresponding hypotheses will be represented using causal directed acyclic graphs (DAGs) constructed in the DAGitty browser-based environment.²¹ Suspected variables include demographic characteristics (age and sex), clinical factors (ECOG PS, smoking status, and comorbidities derived from the Charlson index), tumor characteristics (histology and metastatic sites), and molecular features (EGFR, ALK/ROS1 status, and PD-L1 expression), as well as prior treatments. This approach provides a structured framework to distinguish true confounders from other variables that could introduce bias if controlled in the analysis such as colliders or mediators. Confounders will be identified a priori for each TTE. All confounders are defined at baseline, consistent with the specification of baseline-defined dynamic treatment strategies.

Causal hypotheses

To support valid causal inference in the two emulated target trials, the following assumptions are made:

Exchangeability

We assume that, conditional on measured baseline covariates, the potential outcomes under each treatment strategy are independent of the actual treatment received. This implies that all relevant confounders influencing both treatment assignment (i.e. PPIs dispensation) and outcomes (OS, and rwPFS) are adequately measured and adjusted for. Confounders will be identified through literature review and causal diagrams (DAGs) constructed using DAGitty (compare C. Identification and classification of confounders).

Positivity

There must be a nonzero probability of receiving each treatment strategy (with or without PPIs) for all combinations of baseline covariates. This ensures that comparisons between strategies are possible across the full spectrum of patient profiles.

Consistency

The treatment strategies are well-defined (compare § Study design) and consistently implemented. That is, patients classified as exposed to PPIs truly received the intervention as specified (within the grace period), and those classified as unexposed did not receive PPIs during the defined window. The outcome observed under each strategy corresponds to the potential outcome we aim to estimate.

Noninterference (stable unit treatment value assumption)

We assume that the treatment assignment of one individual does not affect the outcomes of another. This is reasonable in the context of pharmacological interventions such as the PPI use in NSCLC patients.

Correct model specification

The statistical models used for inverse probability weighting and outcome estimation will be correctly specified and justified. Sensitivity analyses [e.g. G-computation, E value, and quantitative bias analysis (QBA)] will be conducted to assess the robustness of findings to potential violations of this assumption.

Primary analysis: cloning, censoring, and weighting

The primary causal analysis will rely on the three-step cloning, censoring, and weighting (CCW)²² approach to mimic the randomization of our target trial comparing the effect of initiation of PPIs within 12 weeks after TKI or ICI initiation. We consider this grace period of 12 weeks in order to reflect clinical reality more faithfully in which the actual dispensation of a PPI may not coincide exactly with time zero and the initiation of TKI/ICI.

• •The first step is to clone, at time zero, each eligible patient and assigned to each treatment strategy of interest (with or without concomitant PPI dispensation).
• •Then, when individuals deviate from their assigned strategy, they are censored. That is, clones in arm A (TKI or ICI with concomitant PPIs) are artificially censored at the end of the grace period if they had not initiated PPIs. Clones in arm B (TKI or ICI only arm) are artificially censored at the time they initiate PPIs within the grace period. Artificial censoring is implemented at event time using exact dates of drug dispensing and treatment initiation, with daily resolution given the structure of the SNDS data. No fixed discrete time grid is imposed. Clones are not censored at the time of PPI initiation, as this is consistent with adherence to the assigned strategy. Death during the grace period will be treated as an event.
• •The last step consists of weighting clones using inverse probability of censoring weights to adjust for the potential selection bias induced by this artificial censoring. The weights will be estimated using a multivariable Cox proportional hazards (PH) model for the censoring process, including baseline confounders identified a priori through the DAGs. The baseline hazard function will be estimated using the Breslow estimator.²³ The adequacy of the weighting procedure will be assessed by graphical evaluation of the positivity assumption and by examining standardized differences of patients’ characteristics between the two treatment strategies in the weighted population at 12 weeks after TKI or ICI initiation. Only patients eligible to receive PPIs will be selected, and we expect sufficient covariate overlap, which should limit positivity violations. We will inspect the distribution of weights, and truncation will be considered only in the presence of extreme values or evidence of positivity violations. Positivity will be assessed through inspection of covariate balance at week 12 in the weighted population, evaluation of overlap in estimated propensity score distributions between treatment strategies, and inspection of the distribution of weights over time, with attention to extreme weights as indicators of potential violations.

Estimation of treatment effects

Adjusted survival curves will be estimated using the weighted nonparametric Kaplan–Meier (KM) estimator described by Cole and Hernán.²⁴ The effect of concomitant PPIs will be reported using both relative and absolute measures. Relative effects will be expressed as adjusted hazard ratios (HRs) estimated from weighted Cox PH models. Absolute effects will be estimated using the restricted mean survival time (RMST),²⁵ defined as the average survival time (i.e. the area under the adjusted survival curve) up to a prespecified follow-up time (3 years for TKI and 2 years for ICI). The difference in RMST between treatment strategies quantifies the average gain or loss in survival over the fixed period.

Complementary analysis: G-computation

To assess the robustness of the primary analysis, G-computation²⁶ will be used. G-computation will target the same causal estimand as the primary analysis, corresponding to the per-protocol effect of adherence to the treatment strategies defined over the 12-week grace period, but relying on a model-based standardization approach. An outcome model (Q-model) will be specified using a Cox PH model with PPIs dispensation included as a time-dependent variable²⁷ in order to appropriately address immortal time bias. PPI dispensation will be the only time-varying covariate. Baseline confounders identified via the DAG will be included as fixed covariates in the model. Based on this Q-model, survival outcomes will be predicted under two counterfactual treatment strategies: (i) a hypothetical world in which no patients initiate PPIs during the 12-week grace period and (ii) a hypothetical world in which all patients initiate PPIs within the 12-week grace period according to the observed distribution of initiation times among users. For arms with concomitant PPI dispensation, survival probabilities will be predicted by explicitly defining the corresponding treatment trajectory over time and computing survival functions conditional on this trajectory. Predictions will be obtained by combining the baseline cumulative hazard with interval-specific linear predictors (i.e. one before the start of PPIs and one after). Survival probabilities will be calculated in a piecewise manner as the product of conditional survival contributions across successive time intervals. Predicted survival functions under each counterfactual scenario will be used to estimate adjusted survival curves and RMSTs. Due to the noncollapsibility of the HR,²⁸ we will estimate the average over time of the different time-specific HRs as proposed by Schemper et al.²⁹ In addition to the primary per-protocol estimand, the average treatment effect on the treated (ATT) will be estimated in sensitivity analyses to quantify the effect of PPIs initiation among patients who actually received PPIs.

Subgroup and exploratory analyses

Additional secondary objectives include assessing the impact of concomitant PPI dispensation on OS and rwPFS for each individual TKI or ICI, analyzed separately. Analyses for specific agents will be conducted when sufficient numbers of treated patients are available in the database. Given the recent introduction of cemiplimab as a first-line treatment of NSCLC, and its reimbursement since January 2023, the number of eligible patients receiving this agent in the ESME database may be insufficient to allow a robust individual analysis of the impact of concomitant PPI dispensation for this treatment. We will also carry out a subgroup analysis according to the combination with or without CT.

As an exploratory objective, a third comparison of interest will be conducted to assess the association between PPI exposure and OS and rwPFS in ESME NSCLC patients who were not treated with a TKI or an ICI in the first-line setting, but who met the same eligibility criteria as patients included in the two main target trials. This population will include patients treated with CT with or without bevacizumab and without TKI or ICI. This exploratory analysis is intended to provide a contextual comparison. If the associations observed in TKI- or ICI-treated patient is related to its effect on the absorption or metabolism of anticancer drugs, no or little effect (remember that PPIs are associated with a small increase in cause-specific mortality in the general population—in Veterans and in five large cohorts—and in cancer patients in a Danish cohort30, 31, 32, 33) of PPIs on oncological outcomes should be observed in this ‘control’ comparison. This hypothesis is evaluated within a TTE framework to highlight the absence of substantial impact of concomitant PPI dispensation on OS and rwPFS in patients with advanced or metastatic NSCLC receiving first-line CT (± bevacizumab) but no TKI or ICI.

Sensitivity analysis

Sensitivity analyses will be used to explore unmeasured confounding, in particular, QBA to explore the potential impact of unmeasured confounding using hypothetical confounder scenarios.³⁴ In addition, E values will be calculated to provide a quantitative summary of the minimum strength of association an unmeasured confounder would need to have with both treatment and outcome to fully explain the observed effect.³⁵ E values complement the QBA by providing a transparent, easily interpretable measure of robustness to unmeasured confounding.

We will carry out sensitivity analyses with shorter grace periods of 6 and 8 weeks to assess the robustness of the estimated effect.

Missing data

Missing data in covariates will be summarized, and multiple imputation could be used to impute this data using the missMDA package.^36,37

Sample size

The need for sample size calculations in studies based on observational data remains debated in the methodological literature.38, 39, 40 According to the recommendations of Hernán,³⁸ no formal sample size calculations were carried out for the emulated target trials. Instead, all eligible patients available in the data sources will be included in the analyses. Note that given the size and national coverage of the databases used, the expected sample size is substantially larger than that of the corresponding target trial.

Discussion

This study applies a TTE framework to address a clinically relevant question using RWD, thereby reducing common biases inherent to observational studies, including immortal time bias and selection bias. The use of a CCW approach as the primary analysis allows appropriate alignment of eligibility, treatment strategies, time zero, and follow-up, closely mimicking the design of an RCT. A major strength of this design is that it enables the emulation of a large-scale clinical trial over an extended period of routine clinical practice, with follow-up spanning up to a decade—an approach that would be difficult to achieve in a conventional randomized trial. Secondly, confounding control is strengthened through a priori identification of confounders using DAGs, ensuring transparent and theory-driven covariate selection. Treatment effects will be estimated using both relative and absolute measures, providing complementary and clinically interpretable estimates of the impact of concomitant PPI dispensation. Thirdly, robustness of findings will be explored using complementary causal inference approaches based on G-computation, along with sensitivity analyses including QBA and E values.

Several limitations should be acknowledged. Firstly, despite employing advanced causal inference methods, this remains an observational study, and residual confounding due to unmeasured or imperfectly measured variables cannot be fully ruled out. Pooling of multiple TKIs and ICIs may introduce heterogeneity. Although treatment indication remains well-defined and consistent, differences may arise in prognostic factors and treatment schedules. To address this heterogeneity, key baseline variables will be explicitly included in the adjustment set, including biomarker settings and the calendar period of treatment initiation. Secondly, PPI exposure is inferred solely from dispensation records, with no information on actual intake, treatment duration, or discontinuation, which may result in exposure misclassification and potentially bias or attenuate effect estimates. PPIs’ impact appears to be limited to the period of active treatment, with a progressive decrease of the effect within a few days after PPI discontinuation. However, in routine clinical practice, adherence to PPI is generally high. Indeed, because PPIs show rapid alleviation of symptoms, are well tolerated, and widely reimbursed in France, discontinuations are rare once the treatment is initiated. Therefore, concerns regarding insufficient or minimal exposure are unlikely. In addition, PPIs do not exhibit a dose-response effect on gastric pH, and their effect occurs immediately with TKIs and within a few days with ICIs. Thirdly, outcomes such as rwPFS rely on routinely collected clinical data and may be affected by incomplete or delayed documentation. Finally, subgroup and exploratory analyses may be limited by small sample sizes for newer treatments, such as cemiplimab, potentially limiting precision and statistical power for individual drug-specific estimates.

Conclusion

This study applies a rigorous TTE framework to estimate the causal effects of concomitant PPI dispensation on survival outcomes in patients with advanced NSCLC treated with TKIs or ICIs. By combining CCW analyses with principled confounder identification and complementary causal inference approaches, we aim to generate robust and clinically relevant evidence from nationwide RWD.

Declaration of AI and AI-Assisted Technologies in the Writing Process

During the preparation of this work, the authors used ChatGPT in order to improve the readability of the abstract. After using this tool/service, the authors reviewed and edited the content as needed and take full responsibility for the content of the publication.