Multivariate analysis. Some guidelines for physicians.

"Mystified" best characterizes the feeling of many physicians as they read the results of a multivariate analysis. Much of the mystery disappears when two items are shown: (1) explicit tabulations revealing how each factor selected by the analysis relates to outcome, both singly and jointly with other factors; and (2) alternative models generated by repeating the analysis after removing the most important variables one by one from the list of factors that was initially analyzed. In this way, one can determine which variables are good stand-ins or substitutes for the best variables in the initial analysis. Showing these substitutes may expose exaggerations in benefits attributed to a particular therapy (ie, elective regional node dissection for clinical stage I malignant melanoma). We have applied these guidelines to examples in the literature, particularly studies of malignant melanoma.