108 Chapter 5 Statistical analysis Statistical analysis was performed in the R statistical framework (Version 4.2.0, R Core Team 2020. R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria); the package NLME (Linear and Nonlinear Mixed Effects Models) was used for linear mixed model analyses, and the packages ggeffects with ggplot2 for visual purposes. After log10 transformation of biomarker values, regression coefficients with 95% confidence intervals were backtransformed. Mixed model analyses were used to compare host response differences between cases and controls at baseline, host response biomarkers at the time of ICU-acquired pneumonia diagnosis relative to controls, and host response trajectories, i.e. the change in biomarkers over time from prior to ICU-acquired pneumonia to the day of ICU-acquired pneumonia. The compound symmetry structure was chosen as variancecovariance matrix which formally outperformed the goodness of fit indices (i.e. -2*log likelihood ratio test and subtracting Akaike’s information criterion) of other common chosen correlation structures (i.e. autoregressive process of order 1, and unstructured correlation matrix). For 16 patients in the case group in whom the sample drawn on day 7 after inclusion was taken on the same day as pneumonia was diagnosed, both samples were treated as event sample in the mixed model analyses. In cases in whom the day 7 sample was taken after the event (i.e., in these patients ICU-acquired pneumonia was diagnosed prior to day 7; eTable 5) biomarker trajectories across three time points were included in the mixed model without controls. Differences between time points were analyzed by contrast dummy coding of the time variable and by changing the reference category. In another analysis only including cases, linear regression was used with logtransformed biomarkers at baseline as outcome and time to ICU-acquired pneumonia in days as predictor. Time cut-offs to onset of ICU-acquired pneumonia (0-5 days, 6-9 days, >10 days) were chosen to maintain a balanced number of patients in each group (eTable 6). Nonlinearity was assessed by using restricted cubic splines with 3 predefined internal knots at day 3, 6, and 10. Expanding the number of knots did not improve the model fit of Akaike’s information criterion. All results are presented as counts (percentages) for categorical variables, median and interquartile ranges (IQR, 25th and 75th percentiles) for nonparametric quantitative variables, and mean ± standard deviation of the mean (SD) for parametric quantitative variables. Histograms, density plots and Q-Q plots were examined to assess data distribution. Continuous nonparametric data were analyzed using a Wilcoxon signed rank sum- or Kruskal-Wallis test; categorical data were analyzed using a Fisher exact test; continuous parametric data were analyzed using a Student t test. Group differences of the biomarker results were corrected for multiple testing using the BenjaminiHochberg method. Moderation effects were not corrected for multiple testing. P values below 0.05 were considered statistically significant.
RkJQdWJsaXNoZXIy MTk4NDMw