Statistical analysis

The clinical and demographic characteristics of the patients were examined for all patients and for those in the awake prone or awake supine groups. Descriptive results are presented as mean with standard deviation or median (interquartile range (IQR)) for quantitative variables and frequencies (percentage) for qualitative variables. Asymmetry and kurtosis were calculated for quantitative variables. Quantitative comparisons were performed with the independent samples t-test; qualitative comparisons were performed with the Chi-squared, Chi-squared of trend or Fisher's exact test. Baseline and post-awake prone positioning S_pO2/F_IO2 ratios were compared with the dependent samples t-test. The PH-Covid19 mortality score was calculated as described in the original model development and validation study [17].

To reduce the risk of bias due to unbalanced groups, propensity score analysis was performed through a logistic regression model adjusted for age, sex, presence of three or more comorbidities, baseline S_pO2/F_IO2, supplemental oxygen device, ICU attention and treatment with systemic steroids, enoxaparin, tocilizumab or ceftriaxone. Patients were matched in a 1:1 ratio according to the nearest-neighbour matching algorithm; changes in density functions are shown in appendix 6 in the supplementary material. All inferential analyses were performed for all patients in the original cohort and for the propensity score-matched cohorts.

Distinct multivariable logistic regression analyses were performed to determine the risk of orotracheal intubation and mortality associated with awake prone positioning. Variables included in the models were selected by the Enter method; adjustment variables were those which had p<0.1 in univariate analyses that have been reported to be associated with higher (or lower) risk for adverse events (age, sex (male), ICU attention, diabetes, systemic arterial hypertension, obesity, heart disease, cancer and chronic kidney disease), pre-prone S_pO2/F_IO2, supplemental oxygen delivery device, ceftriaxone, enoxaparin, tocilizumab, oseltamivir and systemic steroids). A multivariable logistic regression model was subsequently created to determine the risk of intubation among patients who tolerated the awake prone position; the variables included in this model were selected with the Stepwise Forward method, including those with p<0.1 in the final model. Odds ratios with their 95% confidence intervals were calculated. The goodness of fit of the final models was evaluated with the Hosmer–Lemeshow statistic and the discrimination of the model was determined by calculating the area under the curve (AUC). The risks of intubation among awake prone patients according to age and baseline S_pO2/F_IO2 were graphed through the smoothing spline method.

Subanalyses of intubation and mortality risk for patients who had a positive RT-PCR for SARS-CoV-2 (excluding patients in whom RT-PCR was not available but who had a compatible CO-RADS study) were performed in the unmatched and propensity score-matched cohorts through logistic regression models; the size of effect was adjusted for the same variables as the main analyses.

E-values for the lower bound of the confidence intervals were calculated to determine the value at which an unmeasured confounding factor could potentially alter the observed effect of awake prone positioning on the outcomes and drive them to a nonsignificant value [18]. Regression analyses were verified through residual analysis.

To determine the variability of the association between the awake prone position and intubation rates across different centres, multicentre adjustment was performed through generalised estimating equations; the centre with the lowest intubation rate throughout the entire study period was set as the reference. The main effects of every centre and awake prone positioning were calculated in the same model, as well as their interaction within the model.

A systematic search of studies of awake prone positioning was conducted; the search strategy and inclusion criteria for studies are provided in appendix 7 in the supplementary material. Results of eligible studies were summarised alongside the propensity score-matched cohort of APRONOX through a random effects model in a forest and funnel plot of the overall risk of intubation for patients in the awake prone versus awake supine position.

Missing values were not imputed. p<0.05 was used to define bilateral statistical significance. All analyses and graphs were created with SPSS version 21 (IBM, Armonk, NY, USA), R version 3.4.2 (R Foundation for Statistical Computing, Vienna, Austria) and RevMan version 5.3 (Cochrane, London, UK).