Significant variables were further used to build multinomial logistic regression models by backward stepwise modelling, with age and the fold change of weight always as forced entry variables.
Variable selection was performed using backward stepwise modeling with a p value for variable exclusion of 0.15.
Hazard ratios (HR) were calculated using the Cox proportional hazard model and baseline characteristics were adjusted by using backward stepwise model including covariates with a probability value ≤ 0.20 in the univariate analysis.
Multivariable analysis was conducted using Cox's regression model with backward stepwise model selection of predictors using the Akaike Information Criterion [ 26].
The result from the backward stepwise model was adopted.
Interaction analysis was done by logistic regression backward stepwise model.
Multivariable models used a backward stepwise model-building technique to arrive at a parsimonious model, in which all variables were included in the complete model with backward stepwise elimination of the least significant variable, using p < 0.05 as the cut point.
There was however an exception: arm circumference was selected (with an inverse relation) by forward stepwise Cox model, whereas backward stepwise Cox model selected instead physical activity (also with an inverse relation) pointing to physical fitness as the common descriptor.
Instead of considering a unique final model, as is the case in classical forward, backward, or stepwise model selection procedures, with multi-model selection, it is possible to identify a set of 'top models' that can be ranked and weighted according to information criteria such as AIC.
However, after using both forward and backward stepwise logistic regression model with all the predictors plus age and sex, only sVCAM-1Day1 level (P = 0.009, 0.02), age (P = 0.002, 0.011), and SOFA score (P = 0.007, 0.002) were independently associated with SE.
