Summary

Using the GLM Multivariate procedure, you have performed a multivariate analysis of variance on the patient lengths of stay and treatment costs, using the surgical procedure performed and thrombolytic administered as fixed factors. Your initial model indicated that the final treatment costs for reteplase and alteplase are not significantly different from those for streptokinase. However, that model violated the equal variances assumption. The spread vs. level plot showed that a log-transformation of Treatment costs might be appropriate, so the model was re-run, replacing Treatment costs with Log-cost as a dependent variable. This second model passed Levene's test, but now showed a significant difference in the final costs for thrombolytics. The new difference in costs translates to an extra 550 to 860 dollars for the "average" MI patient, so further study of the cost-effectiveness of the new drugs is necessary.

What happened? The differences in Treatment costs in the original model fall in the range of 550 to 860 dollars, but that model did not find the difference to be significant. Why should it matter now? Since Treatment costs is a positive-valued variable, its distribution is probably right-skew, so it is likely that there are patients who incurred unusually high costs, thus inflating the error variation in the first model. By log-transforming Treatment costs, the influence of these high-cost patients is reduced. In this case, it was enough to make the differences in costs to be statistically significant.

Once satisfied with Log-cost as a dependent variable, you should fit a "final" model without the interaction term, because it has not contributed to either of the first two models.