Statistical analysis
Visualization techniques (Q-Q plots and histograms) and Kolmogorov-Smirnov tests were applied to assess the distribution of continuous variables. Mean \(\pm\) standard deviation (SD) and median and 25th and 75th percentiles were reported for continuous variables that followed normal and non-normal distributions, respectively. Demographic and clinical characteristics of DPI and MDI device users were compared using parametric and nonparametric tests (Mann-Whitney U test, Pearson Chi-Square, or Fisher’s Exact tests). According to the literature and consulting with experts in the team, confounding factors were drawn in the directed acyclic graph (DAG, Figure 2) and, if applicable, considered for adjustment in further analyses. Bacterial microbiome composition and global diversity (α and β diversity using phyloseq, 26and vegan 27 R-packages) were compared between DPI and MDI groups. The richness and the Shannon index were used to assess the differences in α diversity. The Wilcoxon rank-sum test evaluated the significant difference in α diversity. The Bray-Curtis and weighted UniFrac distance measures were used to compare the two groups’ β diversity with p-values calculated by PERMANOVA models after adjusting for multiple covariates (as defined in the DAG, Figure 2). A p-value compared between the two groups using the Analysis of Compositions of Microbiomes with the Bias Correction (ANCOM-BC) method by ANCOMBC R-package, 28 including an internal normalization. Taxons present in \(\geq\) 5% of samples, 20,29 and covariates, as defined in the DAG (Figure 2), were included in the ANCOM-BC differential abundance model. Multiple testing was corrected using the Benjamini–Hochberg method, with 0.05 as a significant cut-off if applicable. R (version 4.2.2, 2022-10-31) and R-studio (version 2023.03.1+446) were used for the analyses and data visualizations.