Outcome was defined as the decision to hospitalize or discharge a patient, made by the ER physician. The decision rule for predicting outcome was estimated and assessed using the classification tree methodology described by Rreiman et al. The computations were performed using the “CART* software. This method discriminates between the outcome classes of interest through a series of binary stratifications. It searches for the cutpoint among any of the variables that best separates the data with respect to outcome. The process is repeated until each subgroup reaches a minimum size. The result of the splitting process can be represented by a binary tree, with each terminal node or “leaf” representing a subgroup of the population. As the tree is usually too large to be an effective discriminator, the optimal subtree is estimated using the crossvalidation technique. This method works by dividing the data into ten groups of equal size, building a tree on 90 percent of the data, and then assessing its discriminatory power on the remaining 10 percent of the data.
For descriptive purposes, boxplots were computed and plotted for each of the potentially important predictors, split by presentation or disposition status. Boxplots are a convenient method of displaying the distribution of a variable. The middle bar in the box shows the median, the outer bars show the lower and upper quartiles, and the vertical arms show the 95 percent range. Points outside of this range are possible outliers and are denoted by asterisks.