Effect of noninvasive ventilation on tracheal reintubation among patients with hypoxemic respiratory failure following abdominal surgery: a bayesian post-hoc analysis of the NIVAS trial Authors: Arthur Naudet-Lasserre, Joris Pensier, Audrey de Jong, Mathieu Capdevila, Samir Jaber, Nicolas Molinari Published: 2025 ABSTRACT AbstractBackgroundClinicians’ decision-making regarding the use of noninvasive ventilation (NIV) after abdominal surgery requires evaluating the probability of a clinically meaningful benefit. The Bayesian framework may help caregivers interpret the findings of a randomized controlled trial (NIVAS) assessing curative NIV after abdominal surgery by incorporating their prior beliefs, yielding more precise estimates of treatment effects. This study aimed to use a Bayesian framework to estimate posterior probabilities of NIV effect under various prior assumptions, reflecting diverse clinicians’ beliefs.MethodA prospectively registered, post-hoc Bayesian reanalysis of the NIVAS multicenter trial was conducted. The study included 293 patients with acute respiratory failure following abdominal surgery who were randomly assigned to receive either conventional oxygen therapy or NIV. Four statistical priors were defined: minimally informative, skeptical, enthusiastic, and pessimistic, reflecting a range of clinical beliefs. The primary outcome was day-7 reintubation. Secondary outcomes included day-30 mortality. Effect sizes were presented as odds ratios (OR) and absolute risk reduction (ARR) with 95% credible intervals (CrI).ResultsThe minimally informative prior resulted in a posterior median OR for day-7 reintubation of 0.59 (95% CrI 0.37 to 0.95) in favor of NIV. Under the pessimistic prior, the posterior median OR was 0.64 (95% CrI 0.40 to 1.00). The posterior probability that NIV is superior to oxygen therapy varied from 96 to 99% across various priors, from pessimistic to enthusiastic. The probability of benefit beyond an ARR ≥ 5% ranged from 77 to 93%. Regarding day-30 mortality, the posterior median OR was 0.63 (95% CrI 0.32 to 1.31) under minimally informative prior and 0.79 (95% CrI 0.48 to 1.29) under the skeptical prior. The probability of NIV superiority ranged from 82 to 90%.ConclusionThis Bayesian trial re-analysis indicates that NIV has a high probability of achieving a clinically meaningful reduction in day 7 reintubation, even under pessimistic prior beliefs. These results suggest that use of NIV to treat respiratory failure after abdominal surgery is reasonable.Supplementary InformationThe online version contains supplementary material available at 10.1186/s13054-025-05795-9.Keywords: Acute respiratory failure, Noninvasive ventilation, Abdominal surgery, Bayes Background Abdominal surgery, particularly of the upper abdomen, impairs respiratory mechanics through diaphragmatic dysfunction [1] and reduced lung volume. This leads to atelectasis, decreased functional residual capacity, and ventilation-perfusion mismatch, predisposing patients to acute respiratory failure (ARF) [2, 3]. While invasive mechanical ventilation has traditionally been the treatment for ARF [4], it has been associated with increased morbidity, mortality, and healthcare costs [5, 6]. Noninvasive respiratory supports have emerged as an alternative to avoid reintubation [7–9]. A multicenter, randomized controlled trial (RCT), NIVAS [10], concluded that noninvasive ventilation (NIV) was superior to standard oxygen therapy for preventing day-7 reintubation, with a p-value of 0.03. However, traditional p-values provide limited clinical guidance [11] as they ignore prior knowledge about treatment effectiveness and yield only binary “significant” or “non-significant” results. NIV is often perceived as unpleasant [12] for patients, time-consuming for caregivers, and costly [13]. Additionally, it remains a symptomatic treatment with the risk of delaying reintubation. Despite international guidelines [14] supporting its application in treating postoperative ARF [15], its widespread adoption remains limited due to diverging clinical opinions. Bayesian analysis provides a probabilistic framework that updates prior knowledge with new data [16, 17]. In this context of divergent clinicians’ beliefs, submitting the data to different priors [18] could help with the comprehensive examination and adequate propagation of guidelines [19]. Moreover, Bayesian analysis also offers probability estimates of treatment effectiveness rather than binary outcomes, supporting more intuitive clinical interpretation [20]. We aimed to reanalyze the NIVAS trial within a Bayesian framework, assuming various priors that reflect a range of clinician dispositions toward postoperative curative NIV, from enthusiastic support to pessimism. Our findings could either reinforce enthusiasm for NIV or suggest more selective application. Methods Study design and ethical statement We performed a post-hoc Bayesian analysis of the NIVAS trial (Effect of Noninvasive Ventilation on Tracheal Reintubation Among Patients With Hypoxemic Respiratory Failure Following Abdominal Surgery, NCT01971892 [10]). The NIVAS trial was an open-label multicenter randomized controlled trial that evaluated the efficacy of NIV versus standard oxygen therapy in preventing day-7 reintubation in patients with ARF following abdominal surgery. The study protocol was registered online before the data access (https://osf.io/6am3s). Approval for data reuse was obtained from the Scientific and Ethics Committee of Montpellier University Hospital (2025-04-256). The dataset was de-identified before transfer. Written informed consent was collected as part of the NIVAS trial. Re-consent of patients was not required to conduct the reanalysis. Study population The eligible population is thoroughly defined in the original publication [10]. In brief, the study population consisted of patients over 18 years of age who experienced ARF within 7 days of undergoing abdominal surgery under general anesthesia. ARF was defined as a partial oxygen pressure < 60 mm Hg when breathing room air or < 80 mm Hg when breathing 15 L/min of oxygen, or a peripheral oxygen saturation (SpO₂) ≤ 90% when breathing room air, plus either a respiratory rate higher than 30/min or clinical signs indicating intense respiratory muscle work and/or labored breathing. Patients were randomly assigned to receive either oxygen therapy to maintain a SpO₂ of at least 94% or NIV for at least six hours within the first 24 h. Statistical analysis The primary endpoint was day-7 reintubation, and secondary endpoints were day-30 reintubation, day-30 and day-90 mortality, day-7 and day-30 healthcare-associated infections, day-7 and day-30 pneumonia, intensive care unit (ICU) and hospital length of stay to day 30. In Bayesian statistics, “priors” represent our knowledge, beliefs, or reasonable assumptions about treatment effects before seeing the data. The prior mean parameter of our regression model describes the expected treatment effect (as log OR), while the standard deviation (SD) quantifies the degree of uncertainty around this expectation. To assess the feasibility of literature-informed priors, we conducted a systematic literature review identifying all studies published before the NIVAS trial that compared NIV versus oxygen therapy and reported day-7 reintubation (see eMethod 1). This search identified five relevant studies: two RCTs [21, 22] and three observational studies [23–25] as depicted in eFigure 1 and eTable 1. The two RCTs addressed substantially different populations, with reintubation rates in the oxygen groups of 33% and 10%, compared with the 65% rate expected in the NIVAS sample size calculation. While a previous meta-analysis [26] of these trials reported a significant benefit (risk ratio 0.25 [95% CI: 0.08, 0.83]), differences in the studied population precluded using these data for prior specification. The three observational studies examined populations more comparable to the NIVAS trial, with control group reintubation rates from 48% to 64%. However, these retrospective studies carried a serious to critical risk of bias, making them unsuitable for quantitative prior construction (see eFigure 2). While these data suggested modest evidence in favor of NIV, it was still considered a relative contraindication after abdominal surgery [27]. To reflect diverging opinions, we therefore used a minimally informative prior as a reference [16] with a mean set to the absence of effect (log(OR) = 0), and a wide uncertainty (SD = 10): the resulting credible intervals (CrI) closely approximate frequentist confidence intervals. Informative priors were defined following the methodological framework proposed by Zampieri et al. [20]. The skeptical prior was centred on the absence of effect (log OR = 0), but with moderate confidence that extreme benefits or harms are unlikely. This prior is equivalent to a 95% probability that the OR falls between 0.5 and 2 (normal distribution: mean = 0, SD = 0.355). We then defined enthusiastic and pessimistic priors based on the NIVAS protocol’s sample-size assumptions, anticipating an event rate of 65% in the standard oxygen therapy group and 40% in the NIV group (expected OR = 0.35; log OR = −1.02). Thus, the enthusiastic prior follows a normal distribution (mean = −1.02, SD = 0.99), and the pessimistic prior was symmetric about zero (mean = 1.02, SD = 0.99). The SD of 0.99 was calculated to ensure that each informative prior allowed a 15% probability of treatment inferiority (P(log OR >0) = 0.15 for the enthusiastic prior), representing moderate strength of belief while preserving uncertainty. Table 1; Fig. 1 summarize the prior specification. Table 1Characteristics of priors for day-7 reintubationPrior beliefAssumed median of logarithm of ORAssumed SD of logarithm of ORP (OR < 1)Rationale for specifying distribution characteristicsPrior evidence equivalent Minimally informative 01050%Similar to conducting a frequentist analysis Skeptical 00.35550%95% of the probability mass is under the assumption of an OR lying between 0.5 and 2.Equivalent to an RCT enrolling 126 patients, finding 0% RR Enthusiastic −1.020.9985%NIVAS was powered to detect a 25% RD from 65% in the oxygen group to 40% in the NIV group, representing an OR of 0.36, log OR of -1.02Equivalent to a previous RCT enrolling 18 patients finding 25% RR Pessimistic 1.020.9915%Symmetrical about zero of the Enthusiastic, OR of 1/0.36 = 2.8 log OR = 1.02. With SD estimated to allow a 15% chance of benefit from NIV.Equivalent to a previous RCT enrolling 18 patients finding 25%¨Increased riskOR, odds ratio; SD, Standard Deviation; P, probability Characteristics of priors for day-7 reintubation OR, odds ratio; SD, Standard Deviation; P, probability Specification of Prior Distributions for Day-7 Reintubation Treatment Effect. Each curve shows the probability density function of a prior distribution, displayed on the odds ratio scale. Shaded areas below OR = 1 indicate the prior probability that NIV is beneficial; areas above OR = 1 indicate the prior probability of harm For other outcomes, pre-existing knowledge was unavailable; therefore, only the minimally informative and the skeptical priors were employed. We used a Bayesian logistic regression model: P(β|Data) ∝ P(Data|β) · P(β), where β represents the logarithm of the odds ratio (OR) for the treatment effect. The following metrics were reported: (a) median OR, absolute risk reduction (ARR), and posterior distribution (b); Posterior 95% CrI according to the highest density method; (c) probability of effect size beyond 10%, 5% and 2% of ARR. This choice of threshold, along with the corresponding numbers needed to treat, provides intuitive guidance for bedside decision-making beyond traditional statistical significance. In adjunction to the unadjusted model, a multivariable model was fitted to assess the risk of reintubation with greater precision. This model included the adjustment variables from the original NIVAS trial analysis (chronic obstructive pulmonary disease, ischemic heart disease, chronic heart failure, and body mass index >30) as well as more recently identified risk factors [28] (presence of postoperative sepsis). We also included the stratification variables used in the randomization process (age, surgical site, and presence of epidural anesthesia). Prespecified subgroup analyses examined potential treatment effect modification by age (< 60 vs. ≥ 60 years), surgical site (upper vs. lower abdominal), and epidural analgesia use. For each stratification factor, we fitted Bayesian logistic regression models including treatment-by-subgroup interaction terms with minimally informative priors (normal[0, 10]). We calculated the posterior probability that treatment effects differed between subgroups, with Pr(difference) < 0.90 indicating no credible subgroup effect. To assess the robustness of our findings, we conducted sensitivity analyses varying the strength of informative priors and incorporating data-driven priors derived from previous RCTs with different degrees of down-weighting (eMethods 2). We also performed sensitivity analysis with priors on the baseline risk of reintubation (eMethods 2). To formally assess the consistency of findings across prior specifications, we performed a Bayesian meta-analysis of the four sensitivity analyses (eMethods 3). All analyses were performed using R statistical software (version 4.2.1), with the “brms” package [29] implementing No-U-Turn Sampler adaptive Hamiltonian Monte Carlo (4 chains, 1,000 burn-in iterations, 4,000 saved iterations per chain). The complete code for the analysis is available at https://github.com/pollux74/bayesian_reanalysis_NIVAS.git. Role of the funding source There was no funding source for this study. Results All 293 patients from the intention-to-treat analysis described in the primary report were included in the subsequent analysis. Detailed patient characteristics are described in the primary report [10]. Primary outcome: day-7 reintubation As shown in Table 1; Fig. 2B, under the minimally informative prior, NIV reduced the incidence of the primary outcome, with a posterior median OR for day-7 reintubation of 0.59 (95% CrI 0.37 to 0.95), corresponding to an ARR of 12.5% (95% CrI 1.2% to 23.3%). The probability that NIV was superior to standard oxygen therapy was 98%, and the probability of achieving an ARR ≥ 5% was 90%. Under the enthusiastic prior, the posterior median OR was 0.57 (95% CrI 0.36 to 0.91), corresponding to an ARR of 13.1% (95% CrI 2.4% to 23.6%). The pessimistic prior yielded a posterior median OR of 0.64 (95% CrI 0.40 to 1.00) corresponding to an ARR of 10.5% (95% CrI − 0.1% to 21.2%). The skeptical prior produced the most conservative estimates, with a posterior median OR of 0.70 (95% CrI 0.47 to 1.05) corresponding to an ARR of 8.5% (95% CrI − 1.0% to 17.7%). As shown in Table 2; Fig. 2A, the choice of prior had minimal influence on the results, with the probability of benefit from NIV ranging from 96% to 99% and the probability of an ARR ≥ 5% ranging from 77% to 93%. Posterior Distributions of Treatment Effect on Day-7 Reintubation. (A) Posterior distributions across all four priors, expressed as odds ratios (left) and absolute risk reductions (right). Despite different prior assumptions, all posteriors converge toward similar conclusions, indicating that the observed data dominates. (B) Detailed posterior distribution for absolute risk reduction using the minimally informative prior. Green areas indicate the posterior probability that NIV reduces the risk of reintubation compared to standard oxygen therapy, while the red area indicates the probability of harm. Labeled percentages indicate cumulative posterior probabilities: 96.4% for any benefit from NIV, 90.1% for at least 5% absolute risk reduction, and 67.4% for at least 10% absolute risk reduction. The posterior probability that standard oxygen is superior is 1.5% Table 2Posterior probabilities and effect estimate for primary and secondary endpoints under each prior assumptionPriorbeliefPosterior median OR (95% CrI)P(OR < 1)Posterior median ARR, % (95% CrI)P(ARR ≥ 2%)NNT = 50P(ARR ≥ 5%)NNT = 20P(ARR ≥ 10%)NNT = 10 Day-7 reintubation Minimally informative 0.59 (0.37 to 0.95)98%12.5% (1.2% to 23.3%)96%90%67% Enthusiastic 0.57 (0.36 to 0.91)99%13.1% (2.4% to 23.6%)98%93%71% Pessimistic 0.64 (0.40 to 1.00)97%10.5% (−0.1% to 21.2%)94%85%54% Skeptical 0.70 (0.47 to 1.05)96%8.5% (−1.0% to 17.7%)91%77%37% Day-30 mortality Minimally informative 0.63 (0.32 to 1.31)90%5.0% (−2.4% to 12.7%)78%50%10% Skeptical 0.79 (0.48 to 1.29)82%2.5% (−3.3% to 7.8%)57%18%1% Day 7 healthcare-associated infection Minimally informative 0.51 (0.30 to 0.90)99%12.1% (2.4% to 21.6%)98%93%67% Skeptical 0.66 (0.44 to 1.01)97%7.6% (−0.2% to 15.0%)92%74%27% Day-7 pneumonia Minimally informative 0.39 (0.20 to 0.75)100%11.9% (3.7% to 20.4%)99%95%67% Skeptical 0.61 (0.38 to 0.97)98%6.6% (0.3% to 13.1%)92%69%15%OR, odds ratio; CrI, Credible interval; P, probability; ARR, absolute risk reduction; NNT = Number Needed to Treat (of patients to avoid one event) Posterior probabilities and effect estimate for primary and secondary endpoints under each prior assumption OR, odds ratio; CrI, Credible interval; P, probability; ARR, absolute risk reduction; NNT = Number Needed to Treat (of patients to avoid one event) After adjusting for stratification variables and prognostic factors, the treatment effect strengthened: posterior median OR of 0.50 (95% CrI 0.28 to 0.88) and ARR 13.1% (95% CrI 2.5% to 23.5%) under the minimally informative prior (see eTable 2 and eFigure 3). The probability of benefit from NIV ranged from 97% to 99% and the probability of achieving an ARR ≥ 5% from 78% to 94% across the four priors. Secondary outcomes For day-30 mortality, the minimally informative prior resulted in a posterior median OR of 0.63, corresponding to an ARR of 5.0% (95% CrI − 2.4% to 12.7%), as shown in Table 2; Fig. 3. This analysis showed a 90% probability that NIV reduces day-30 mortality compared to standard care, with a 50% probability of achieving a clinically meaningful ARR ≥ 5%. The skeptical prior resulted in a posterior median OR of 0.79 (95% CrI 0.48 to 1.29) and a corresponding median ARR of 2.5% (95% CrI − 3.3% to 7.8%). In the multivariable analysis, adjustment for covariates yielded larger effect estimates favoring NIV, with a posterior median OR of 0.47 (95% CrI 0.18 to 1.19) corresponding to an ARR of 5.1% (95% CrI − 1.1% to 11.4%) under the minimally informative prior. This adjustment increased the probability of benefit from NIV to 95% and increased the probability of achieving an ARR ≥ 5% to 51% (see eTable 3 and eFigure 4). Posterior probability distributions for OR and ARR comparing NIV versus oxygen on day-30 mortality according to each prior Regarding day-90 mortality, the minimally informative prior resulted in a posterior median OR of 0.64 (95% CrI 0.34–1.13), while the skeptical prior yielded an OR of 0.78 (95% CrI 0.49–1.22) as shown in eTable 4 and eFigure 5. The probability that NIV reduces day-90 mortality compared to standard care ranged from 87 to 93%. For day-7 healthcare-associated infections and day-7 pneumonia, both priors indicated a high probability of benefit from NIV, as shown in Table 2 and eFigures 6 and 7. The probability of benefit from NIV ranged from 97% to 99% for day-7 healthcare-associated infections, with a 74% to 93% probability of an effect beyond 5% of ARR. For day-7 pneumonia, the probability of benefit ranged from 98% to 100% with a 69% to 95% probability of an effect beyond 5% of ARR. Results for day-30 reintubation, pneumonia, and healthcare-associated infections are described in eTable 4 and eFigures 8 to 10, and results for ICU and hospital length of stay are described in eTable 5. Subgroup analyses demonstrated consistent treatment effects across all prespecified stratification factors (eFigure 11). The probability that treatment effects differed between subgroups was 0.56 for age, 0.89 for surgical site, and 0.61 for epidural analgesia indicating no credible subgroup effect. Sensitivity analysis Posteriors probabilities for day-7 reintubation demonstrated robustness to prior specifications (see eTable 6). The posterior probability of NIV benefit ranged from 93% to 99% across all pessimistic prior settings, with the lowest probability (93%) observed when using the most conservative prior (5% chance of treatment benefit). Results remained consistent in the sensitivity analysis using data-driven priors from the two previously published RCTs, across all degrees of down-weighting. The Bayesian meta-analysis across all 4 priors yielded a pooled OR of 0.66 (95% CrI 0.52 to 0.85). The choice of prior accounted for only 8.7% of the observed heterogeneity (at mean tau), indicating that our findings are robust to prior specification (eFigure 12). As shown in eTable 7 and eFigures 13 to 15, applying different informative priors on the baseline risk had a negligible effect on effect size estimation, yielding similar estimates. Bayesian model validity was confirmed through chain convergence assessment, as assessed by inspection of traceplots and autocorrelograms for all analyses (see eFigures 16–28). Discussion In this post-hoc Bayesian analysis of the NIVAS multicenter trial evaluating non-invasive ventilation (NIV) for treating acute respiratory failure (ARF) following abdominal surgery, we found a high probability that NIV reduces the day-7 reintubation rate, even when considering clinically relevant thresholds that exceeded the null. This result was consistent across a range of diverging priors, reflecting varying clinicians’ opinions from pessimistic to enthusiastic. Additionally, we observed a high probability that NIV improves secondary outcomes, including day-30 and day-90 mortality, day-7 pneumonia, and healthcare-associated infections. Beyond the original study, our Bayesian analysis provides an enhanced understanding of the findings. First, we used a rigorous approach, including a systematic literature review and bias evaluation of studies published before the trial. This approach contextualizes our findings within the existing evidence base and enables rigorous prior specification for Bayesian analysis. The use of a skeptical prior, which better reflects the typical distribution of treatment effects, and pessimistic priors, which hypothesize oxygen superiority, challenged the original results and enhanced the robustness of our findings. Incorporating priors that constrain extreme effect sizes therefore reinforces the results’ credibility. Second, presenting full posterior distributions rather than point estimates with confidence intervals provides clinicians with comprehensive information about the range of plausible treatment effects and their relative likelihoods, enabling more informed clinical decision-making. Third, while reintubation and mortality are clinically relevant, we calculated probabilities of treatment effects beyond certain thresholds to help caregivers better understand effect magnitude. This approach quantifies the uncertainty around clinically meaningful effect sizes. Considering the lack of new randomized controlled trials since the NIVAS trial, this study is a valuable addition to the existing literature. A large randomized trial involving 4,808 patients across 70 centers worldwide comparing CPAP to standard treatment showed no significant improvement in composite outcomes [12], but used NIV to prevent ARF rather than for treatment. These contrasting results align with the recent network meta-analysis by Pettenuzzo et al. [30], where NIV was found to reduce the risk of reintubation. However, the effect was only achieved when applied as a treatment for ARF or when used in high-risk patients. Among high-risk patients, those who suffer from obesity represent a priority target where NIV is effective for preventing [7] or treating ARF [6]. This reinforces the primary NIVAS findings that NIV should be widely used to treat patients with established ARF, while its preventive use has been demonstrated only in high-risk patients [7]. These divergent findings across similar interventions illustrate why probabilistic approaches to evidence interpretation merit consideration. The current Bayesian analysis emerges within the context of increased criticism [18] regarding the overreliance on p-values and null hypothesis significance testing (NHST), which often reduces complex findings to binary outcomes based on arbitrary thresholds. Given the substantial resources required for randomized controlled trials, maximizing interpretive value from existing data becomes paramount. Bayesian reanalysis has proven valuable in critical care, providing nuanced interpretations that complement traditional analyses. Goligher et al. demonstrated this approach in their EOLIA trial reanalysis [31], providing a nuanced interpretation of the results in the context of a high risk-reward ratio (mortality benefit in patients with severe ARDS). Conversely, Zampieri et al. [32] used Bayesian methods to challenge optimistic interpretations in their STARRT-AKI re-analysis by showing a low probability of reduced mortality. Our study has several limitations. First, we inherit the limitations of the original trial, including limited power to detect mortality effects, and lack of blinding for the somewhat subjective primary endpoint (i.e., day-7 reintubation). We also inherit the lower-than-expected reintubation rate, which may limit generalizability to higher-risk populations. While Bayesian analysis provides a useful probabilistic interpretation, it cannot overcome the fundamental limitations of the original trial design: a moderate sample size (n = 293). However, our use of multiple priors—reflecting diverse clinical beliefs—may facilitate interpretation across different practice contexts and help inform decision-making. Second, this post-hoc analysis carries inherent risks of data-dependent findings. Despite the analysis plan having been pre-registered after trial publication but before data access, it is acknowledged that the report is of an exploratory nature. To mitigate this risk, we limited our analysis to pre-specified outcomes from the original trial and employed a conservative skeptical prior. Third, our method for defining informative priors may be criticized. Although we conducted a systematic review to evaluate prior knowledge, we decided not to use these studies to construct quantitative Bayesian priors because the two RCTs addressed populations that were significantly less severe, and the three observational studies had a high risk of bias. We therefore used assumptions from the NIVAS trial’s sample size calculation. Nonetheless, sample size calculations often involve subjective assumptions and economic/feasibility considerations. However, these assumptions must be credible to initiate a multicenter RCT. To mitigate this risk, we used a minimally informative prior as a reference and employed informative priors that reflect more extreme clinical opinions that may exist in the medical community. We encountered a common situation where uncertain data quality raised questions about the feasibility and relevance of data-driven priors. Furthermore, no universally rigorous method exists for defining priors, and because they can influence results and conclusions, this remains an ongoing challenge. Fourth, reporting multiple outcomes across different priors raises concerns about interpreting the results. To address this, we restrict the analysis of secondary outcomes to the minimally informative prior (reference) and the skeptical (conservative correction). The outcomes reported were pre-specified in the original trial. These secondary analyses remain exploratory and hypothesis-generating, with substantially greater uncertainty than the primary outcome. Conclusions In this post-hoc analysis of the NIVAS trial, NIV consistently reduced the day-7 reintubation rate across various prior distributions, including a pessimistic prior reflecting a reluctant clinician’s perspective, suggesting evidence for its use. There was a high probability of mortality reduction with NIV, though the magnitude of this reduction was imprecise. Supplementary Information Supplementary Material 1 Acknowledgements None. Author contributions ANL, JP, NM, SJ, and ADJ contributed to conception and design of the study, to the acquisition of the data, to the analysis and interpretation of the data, to drafting the submitted article, and to provide final approval of the version to be published. MC contributed to the interpretation of the data, to drafting the submitted article, and to provide final approval of the version to be published. Funding This research received no specific grant from any funding agency in the public, commercial or not-for-profit sectors. Data availability Data will be available to researchers upon reasonable request immediately following publication. Proposals should be directed to the corresponding author. The R code is available at [https://github.com/pollux74/bayesian_reanalysis_NIVAS.git](https:/github.com/pollux74/bayesian_reanalysis_NIVAS.git). Declarations Ethics approval and consent to participate Approval for data reuse was obtained from the Scientific and Ethics Committee of Montpellier University Hospital (number 2025-04-256). Consent for publication Not applicable. Competing interests NM reports receiving consulting fees from Adene, Nomics, Sanofi, Bruxless, ID Solution, Ar-chean Technologies, MSD, BMS, AB science and Simeox SJ reports receiving consulting fees from Drager, Medtronic, Baxter, Fresenius-Xenios, Mindray, and Fisher & Paykel. AD J reports receiving consulting fees from Medtronic, Dräger, Viatris and Fisher & Paykel. No conflicts of interest are reported for other authors. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1 Data Availability Statement Data will be available to researchers upon reasonable request immediately following publication. Proposals should be directed to the corresponding author. The R code is available at [https://github.com/pollux74/bayesian_reanalysis_NIVAS.git](https:/github.com/pollux74/bayesian_reanalysis_NIVAS.git).