Statistical analysis
Continuous variables were shown as mean ± standard deviation (SD) or
median (interquartile range [IQR]), whichever was appropriate.
Categorical variables were shown as numbers (percentage). To identify
factors independently associated with escalated care, we used logistic
regression and estimated odds ratios (ORs) and adjusted for potential
confounding variables.
Initially, the regression model only included variables associated with
escalated care with values of p <0.2 or that changed the
effect estimate by over 10% after their inclusion. A multivariate
logistic regression model was performed with a backward elimination
method, used with a p value of 0.05 as the limit value for the model
entry. The variable selection and modeling processes were made following
the recommendations of Greenland(12). The method of Sullivan et al. was
used to generate the risk scores and estimate the risks observed with
each score (13). To assess discrimination, area-under-the-curve (AUC)
was estimated with a 95% confidence interval and plotted using AUC-ROC
plots(14). To correct prediction probabilities for over-optimistic
predictions and to evaluate the calibration, the model was analyzed by
comparing the predicted probability to the observed probability of
mortality and examined with a calibration plot and calibration slope
with 95% CI. Calibration plots (STATA function: pmcalplot) displayed
observed risk by deciles of the predicted risk and examined risk at the
individual level using Locally Weighted Scatterplot Smoothing
algorithms(15). We also calculated the Hosmer – Lemeshow goodness of
fit test as well as calibration curves between the predicted
probabilities and the observed data. To correct sampling bias of
variance parameters and to evaluate the internal validity of the model,
repeated curved validation ”tenfold cross-validation” were used,
comparing the area under the ROC curve obtained in the repetitions with
that observed in the model (15, 16); as well as with the bootstrapping
technique with which both the adjusted oppressiveness and area values
were estimated. All statistical tests were two-tailed, and the
significance level used was p < 0.05. The data were
analyzed with Statistical Package Stata 15.0 (Stata Corporation, College
Station, TX).