Why standardizing pneumonia response alone cannot define treatment effect in trials

A structured framework for judging whether pneumonia has responded to treatment in ventilated patients with hospital-acquired or ventilator-associated pneumonia (HAP/VAP) has drawn an important methodological challenge from a team of researchers writing in the journal Critical Care. The framework, proposed by Haitsma Mulier and colleagues, was published on 22 September 2026 and aims to reduce heterogeneity in how treatment response is assessed in this critically ill population. In a Matters Arising letter published on 2 November 2026, Meiting Li, Mengyan Sun, Wenzhen Zhuo and Xiaoyang Gong argue that standardizing the measurement of pneumonia response does not, by itself, define the treatment effect being estimated in a randomized trial.

The letter acknowledges the strengths of the original work. The framework addresses a genuine source of variability in outcome assessment and appropriately separates pneumonia-specific response from all-cause mortality, which in critically ill patients is often multifactorial. The original authors also recognized that classifying death occurring before treatment success as an indeterminate outcome is not prescriptive, and that its handling may depend on the research question and the estimand. Li and colleagues contend that the implications of this point deserve more explicit specification when the framework is applied in interventional research.

The argument rests on numbers from the reported cohort rather than on abstract reasoning. Of 1,603 pneumonia episodes, 302 had an indeterminate overall treatment outcome, including 240 because of death. Assigning death to treatment failure increased the failure proportion from 22.2 percent to 34.3 percent. In a randomized trial, the letter argues, death before the day-10 assessment is not simply a missing response value: once a patient has died, the subsequent pneumonia-response state is undefined. Moreover, death is a post-randomization event that may itself be affected by treatment.

This creates a methodological trap known as truncation by death. If an analysis is restricted to patients who are alive and classifiable at day 10, the comparison is no longer between the randomized treatment groups as originally constituted. Instead, it contrasts treatment-specific groups of observed survivors. If treatment affects early survival, these groups need not remain comparable despite baseline randomization. The letter cites work describing this setting as truncation by death, in which conditioning on survival can break the causal interpretation that randomization is meant to guarantee.

Regulatory guidance points in the same direction. The International Council for Harmonisation guideline E9(R1) requires that the strategy for handling relevant intercurrent events be specified as part of the estimand, and it distinguishes principal strata from subsets defined by observed post-randomization events. The practical stakes are documented in a systematic review of 255 randomized trials, in which the primary estimand could be determined from reported methods in only 46 percent of trials, and the handling of intercurrent events in only 49 percent. Different approaches to death may therefore correspond to different scientific questions rather than alternative analyses of the same treatment effect.

For interventional use, the letter proposes that the response criteria be paired with an explicit estimand specification. One option is a composite strategy that incorporates death into the definition of the outcome, for example by assigning it the least desirable outcome category. Another is a Desirability of Outcome Ranking (DOOR) framework, which encodes clinically relevant distinctions among treatment response, complications, adverse events, and death within an ordered patient-level outcome. A disease-specific DOOR for hospital-acquired bacterial pneumonia and ventilator-associated bacterial pneumonia (HABP/VABP), incorporating infectious complications, serious adverse events, and mortality, has been developed and applied to two randomized trials. A related endpoint was subsequently evaluated using participant-level data from six registrational antibacterial trials, with death assigned the least desirable category. The response criteria proposed by Haitsma Mulier and colleagues could serve as the response component of an estimand built around such a global outcome.

By contrast, the letter warns against a common misreading. An analysis restricted to patients who actually survive and remain classifiable may be reported descriptively, but it should not be interpreted as preserving the randomized causal contrast when treatment can affect survival. If the scientific question instead concerns patients who would survive under either treatment, a principal-stratum estimand such as the survivor average causal effect may be considered. However, membership in this so-called always-survivor stratum is counterfactual and is not identified by randomization alone without additional assumptions. In other words, the same response criteria can support several distinct treatment-effect definitions, and the choice among them must be made explicitly.

The distinction also matters for interpreting the validation analyses reported with the original framework. Associations between day-10 treatment failure and subsequent mortality, duration of ventilation, and intensive care unit length of stay support the prognostic discrimination of the response classification among assessable patients. But some validation outcomes were themselves evaluated within selected post-baseline populations: ICU length of stay was analyzed among patients still admitted to the index ICU at day 10, and ventilation duration among those still mechanically ventilated at day 10. The reported invariance of results after reclassifying death as treatment failure does not resolve this issue, because the corresponding patient-centered analyses remained restricted to patients alive and at risk at day 10.

These validation analyses are informative about whether the classification distinguishes prognosis among observable patients, the letter states, but they address a different question from whether a survivor-conditioned treatment comparison has the causal interpretation intended in a randomized trial. Prognostic value within a selected population does not establish that a comparison conditioned on survival retains the randomization-based causal contrast. Prospective interventional studies should therefore assess not only the reproducibility of the response criteria, but also how prespecified strategies for death and other outcome-precluding events affect the resulting treatment contrast.

The letter closes on a constructive note. The proposed framework is described as an important step toward reproducible assessment of HAP/VAP response, and its application in interventional research would be strengthened by standardizing the requirement for estimand specification rather than prescribing a single strategy for death. Trial protocols applying the framework should prespecify the target population, treatment conditions, response variable, strategy for death and other outcome-precluding events, and the treatment-effect summary measure. Such specification, the authors argue, would preserve the flexibility emphasized by the original framework while clarifying when trials using identical pneumonia-response criteria do, and do not, target comparable treatment effects.

Subject of Research: Methodological critique of a pneumonia treatment response framework for ventilated patients, arguing that response standardization alone does not define trial treatment estimands.

Article Title: Standardizing pneumonia response does not by itself define the treatment estimand

Article References: Li, M., Sun, M., Zhuo, W., & Gong, X. (2026). Standardizing pneumonia response does not by itself define the treatment estimand. Critical Care, 30(1), Article 558. https://doi.org/10.1186/s13054-026-06444-5

Image Credits: AI Generated

DOI: 10.1186/s13054-026-06444-5

Keywords: pneumonia, estimand, HAP/VAP, randomized trials, truncation by death, intercurrent events, ICH E9(R1), DOOR, treatment response, ventilator-associated pneumonia, causal inference, Critical Care