Statistical Analyses
The study population characteristics were described by summary statistics (Table 1). For continuous variables, measures of central tendency (mean/median) and spread (standard deviation/interquartile range) were presented. Mean/standard deviation (SD) were presented for normally distributed continuous variables and median/interquartile range (IQR) for non-normally distributed variables. Normality was checked by conducting the Shapiro-Wilks test and visual inspection of histograms. Frequencies and proportions were presented for categorical variables.
Cox proportional hazards regression models were used to estimate effects of each a priori predictor of home oxygen duration on home oxygen cessation, with simultaneous adjustment for potential confounders in multivariable models. Hazard ratios (HR) with 95% confidence intervals (CI) were calculated. Four Cox-models were built for four different time on oxygen cut-off points: from baseline to three, six, nine, and twelve months CGA respectively. The proportional hazards assumption was assessed by examining the relationship between the scaled Schoenfeld residuals and time on oxygen and the overall fit of the models was tested by examining the Cox-Snell residuals and evidence of non-proportional hazards for oxygen flow was found. Hence, the baseline hazard function was stratified by oxygen flow group and proportional hazards assumption was satisfied by fitting a stratified Cox-regression model. Covariates included for adjustment in multivariable models were GA, birth weight, partial pressure of carbon dioxide (pCO2) at term CGA, invasive ventilation total respiratory support and length of initial NICU hospital stay in days.
We assessed the potential effects of different maternal and infant variables on oxygen flow rates at hospital discharge using multinomial logistic regression. Patients were grouped into three oxygen flow groups based on their oxygen flow requirements at discharge: a) ≤ 125 mL/min; b) 126 mL/min to 250 mL/min; c) 251 mL/min to 750 mL/min. Basic inferential statistical tests were conducted to select variables with significant differences between the three oxygen flow groups in the final multinomial logistic regression model. A Kruskal-Wallis test was used for continuous variables (i.e., birth weight in grams, pCO2 at term CGA in mmHg, total duration of respiratory support in days, length of hospital stay in days and CGA at hospital discharge in weeks). A Chi-Square test was used for categorical variables including GA ≥ 28 weeks and maternal PROM. Odds ratios (OR) with 95% CI were calculated and presented for the final multinomial logistic model. Modelled estimates of the sensitivity, specificity, positive predictive value (PPV) and negative predictive value (NPV) predictive values were derived and presented as Appendix 2.
The Kruskal-Wallis test was used to examine the association between the three oxygen flow groups and continuous post-discharge variables. A univariate linear regression was then conducted to identify the effect of oxygen flow and continuous post-discharge variables. For oxygen flow groups and categorical post-discharge variables, a Fisher’s Exact test and a binomial univariate logistic regression were used