We argued in Section 2 that expenditures on tests should be conditional on the penalty for misclassification errors. Therefore ICET is designed to be sensitive to both the cost of tests and the cost of classification errors. This leads us to the hypothesis that ICET tends to spend more on tests as the penalty for misclassification errors increases. We also expect that the error rate of ICET should decrease as test expenditures increase. These two hypotheses are confirmed in Figure 4. In the plots, the x axis is the value of k and the y axis is (1) the average expenditure on tests, expressed as a percentage of the maximum possible expenditure on tests, T, and (2) the average percent error rate. On average (the sixth plot in Figure 4), test expenditures rise and error rate falls as the penalty for classification errors increases. There are some minor deviations from this trend, since ICET can only guess at the value of a test (in terms of reduced error rate), based on what it sees in the training dataset. The testing dataset may not always support that guess. Note that plots for the other four algorithms, corresponding to the plots for ICET in Figure 4, would be straight horizontal lines, since all four algorithms ignore the cost of misclassification error. They generate the same decision trees for every possible misclassification error cost.