EDITORIAL| VOLUME 164, ISSUE 4, P816-817, OCTOBER 2023

The Emulated Targeted Trial: A Causal, But Never Casual, Surrogate for Randomized Clinical Trials

Pedro D. Wendel-Garcia, Lieuwe D.J. Bos

Chest 2023; 164: 816-817

DOI: https://doi.org/10.1016/j.chest.2023.06.016

Randomization properly carried out […] relieves the experimenter from the anxiety of considering and estimating the magnitude of the innumerable causes by which his data may be disturbed.R. A. Fisher, 1935¹

Almost 100 years later, the randomized controlled trial remains the highest regarded, and for most medical researchers still “only,” exponent of causal inference for clinical questions. Nevertheless, owing to their nature and/or a lack of resources, a plethora of interventions in critical care medicine remain inadequately addressed through randomized controlled trials. The question arises whether these interventions should remain outside the scope of evidence-based medicine and whether biological plausibility alone is an adequate argument for their use.

Fortunately, nowadays, researchers can use causal theory to relieve the disturbing effects in nonrandomized datasets described by Fisher¹ by formulating their experiment as causal counterfactual structural models between an exposure and an outcome (incorporating the mediating and confounding effects that affect them) and graphically communicating them through directed acyclic graphs.² These causal counterfactual thought processes can then be translated through dedicated statistical methods to data without randomized exposures to extract unbiased effect measures. The last barrier that prevented successful implementation of complex causal models into critical care research was breached with the widespread implementation of digitalization, which has led to the emergence of large clinical databases that contain the full scope of clinical decision-making, including the gray areas of eminence-based decisions that can guide causal research.

The emulated targeted trial represents the end product of all these advances.³ It is based on three fundamental constraints and represents the framework around which to construct a causal study: (1) precisely define the causal question in the form of a theoretic randomized trial; (2) establish the causal framework for the trial and identify mediators and confounders, as well as their time-varying properties; (3) apply the protocol and causal framework to the data by means of a statistical method that adequately respects all predefined constraints.

In this issue of CHEST, Wanis et al⁴ explore a controversial topic in critical care medicine: the decision whether to intubate a critically ill patient early after admission to the ICU or whether to delay intubation. The authors approached the research question by means of emulated targeted trials, which they implemented into a cohort of 5,893 medical and surgical critically ill patients in the Medical Information Mart for Intensive Care-IV database. To address the time-varying property of the intubation process, the authors used augmented inverse probability weighting estimation by means of gradient boosting, which is a doubly robust estimator based on the combination of inverse probability weighting with a g-computation estimator that is less prone to misspecification than single estimators such as matching.⁵ Overall, the authors carefully delineate and argue their methodologic approach throughout the article, which makes the manuscript very “educational” for researchers who are interested in conducting emulated targeted trials themselves. Nevertheless, when judging the clinical question at hand and the translatability of the results into clinical practice, we are faced with some limiting aspects.

First, the decision to intubate and mechanically ventilate a critically ill patient is highly dependent on the underlying disease process. The authors restricted the main analysis to patients with respiratory distress defined by oxygen requirement (Fio₂ ≥ 40%), respiratory rate (≥ 30/min), or use of high-flow oxygen or noninvasive ventilation. Crucially, clinical evaluation of respiratory effort by physical examination could not be included, even though this may be the most important parameter in the decision to intubate.^6,7Furthermore, the data also did not allow for differentiation between common causes of respiratory failure, which spanned from cardiac decompensation to pneumonia and chronic respiratory failure. These clinical pictures respond differently to advanced respiratory support and mechanical ventilation.⁸ In fact, we would never design a trial with such broad inclusion criteria, because the heterogeneity of the patient population is too large for us to expect a causal effect to emerge. When an emulated targeted trial is being designed, the rationale for the targeted intervention must be extremely clear, and the population should be restricted maximally to crystalize out the effect; this is the reason we require such large data sets.³ Second, the number of variables that are used in the adjustment set is clearly limited; many variables that could precipitate the decision to intubate a patient (such as arterial blood gas analyses, vasopressor doses, respiratory efforts) were not considered, potentially omitting important confounders.⁷ Third, the authors missed the opportunity to present physiologic outcomes of the patients, stratified by the intervention. Emulated targeted trials should attempt to underline the hypothetical rationale of the investigated intervention by presenting outcomes that can support the biological plausibility of the results. Finally, the reasons that lead to intubation, in general, are not annotated comprehensibly in medical databases, and operative or diagnostic interventions could strongly bias the investigated intervention.

Overall, the capacity of causal theory to capture highly intricate natural phenomena and uncover complex causal effects is highly regarded in other research fields and has guided profound global policy changes.⁹ Only recently in the medical field could a large emulation initiative show the ability of predefined and well-performed emulated targeted trials to predict closely the results of randomized controlled trials.¹⁰Nevertheless, to date, it remains unclear whether these insights are translatable to the highly variable and confounded setting of critical care. To implement emulated targeted trials as an integral part of evidence-based medicine and to integrate the results into our guidelines and clinical decisions, we thus should first provide proof of concept for their adequacy in the setting of critical illness analogous to the previously mentioned emulation initiative.¹⁰

Finally, to further increase the trust in emulated targeted trials, we must be cautious with their interpretation and implement the same high quality standards we do with randomized controlled trials, possibly in the form of dedicated Equator Network guidelines. These should include, among others, a prospective registration of the full analysis protocol, stringent inclusion/exclusion criteria, exhaustive confounder/mediator consideration, and robustness assessment across multiple large databases. Only if researchers expose themselves to the anxiety of considering all the innumerable causes that could disturb their data and carefully address them can causality be estimated by means of emulated targeted trials and influence our clinical decision-making.