Detailed Description

Statistical methods Continuous data are presented as mean and standard deviation (SD) if it follows a normal distribution assessed using histograms and Q-Q plots; otherwise, as median and interquartile range (IQR). Categorical data are presented as numbers and percentages. Missing data will not be imputed. Patients from a specific analysis were excluded if the data for the related variable are missing. The first objective is to investigate the relative effectiveness of different NIVs, including high-flow nasal cannula (HFNC), bilevel positive airway pressure (BiPAP), and continuous positive airway pressure (CPAP). Patients will be divided into three groups per the type of NIV they first received. Patients who received invasive mechanical ventilation (IMV) before NIV will be excluded. Patients who received two or three different types of NIV will also be excluded in this analysis. Patients who received IMV after the use of NIV are eligible. The secondary objective is to investigate the impact of IMV on mortality via comparison with patients who received NIV only. The role of lung protective ventilation in patients who received IMV will also be investigated. The characteristics of the ventilator settings that are associated with an improved outcome will be explored. The primary outcome is in-hospital mortality and patients will be followed until hospital discharge. Patients are considered alive if they were discharged alive from the hospital or are still hospitalized at the closure of data extraction. For the first objective, the rate of respiratory support escalation from NIV to IMV will also be analyzed (as a secondary outcome measure in this analysis). Patients will be balanced using propensity score matching. The propensity score model will include demographic characteristics, comorbidities, the pandemic phase, severity of acute illness (24 hours before the targeted respiratory support), laboratory results (24 hours before the targeted respiratory support), and vital signs (24 hours before the targeted respiratory support). The balance between matched pairs will be assessed using a standardized 10% difference and calculated using the method described by Yang and Dalton. A stratified Cox proportional-hazards model will be used to analyze the matched pairs. Additionally, survival will be estimated using the product-limit Kaplan-Meier estimator, and the log-rank statistic will be used to compare survival curves. The backup statistical analysis plan is as follows. The univariate Cox proportional-hazard models to screen for potential factors associated with lower mortality will be performed. A multivariable Cox proportional-hazards model to estimate independent associations between respiratory supports and mortality will be performed. The confounders included in the multivariable analysis are as follows: 1) known risk factors for mortality (age, sex, and hypertension); 2) the severity of the acute illness 24 hours before the targeted respiratory support (Sequential Organ Failure Assessment score and Glasgow Coma Scale score); 3) the various phases during the first pandemic year, including the first phase (February 1, 2020, to May 31, 2020), the second phase (June 1, 2020, to August 31, 2020), the third phase (September 1, 2020, to November 30, 2020), and the fourth phase (December 1, 2020, to last date of data extraction); 4) the demographics and comorbidities with a P-value < 0.25 in the univariate analysis; and 5) the laboratory results and vital signs 24 hours before the targeted respiratory support that have a P-value < 0.25 in the univariate analysis. All treatments considered to be part of COVID-19 management will be included in the multivariable analysis for confounding control. To avoid collinearity, only one variable will be included if two variables have an absolute Pearson's or Spearman's rank correlation coefficient greater than 0.5. Variables with more than 10% missing data will also be excluded. Multiple testing will be corrected using the Bonferroni method to reduce the chance of type I error at a two-sided 0.05 alpha level, considering the hypotheses for all of the COVID-19-related respiratory supports/treatments as a family. The association between exposures and mortality will be estimated using hazard ratios (HRs) and reported with 95% confidence intervals (CIs). To account for clustering within hospitals, robust sandwich estimators to compute standard errors for the HRs will be used. The proportional hazards assumption will be assessed using Schoenfeld residuals. With a two-tailed hypothesis test, the significance level for each general hypothesis is 0.05. All analyses will be performed in R software (version 3.5.3, R Foundation for Statistical Computing).