Assessing changes in risk factor effect over multiple levels of severity.

Abstract

Use of the logistic function has become standard for regression of proportions onto risk factors, but where multiple levels of response are recorded, several different techniques are common. For most purposes, the "inclusive" method is most appropriate. However, if investigation is focused directly on the pattern of risk factor effects over levels of severity, the inclusive method is too conservative, because its overlapping categories tend to minimize any real differences among risk factor effects at separate levels. Several alternative techniques maximize these estimated differences, but interpretation of their results is much more sensitive to the assumed model than is estimation of category frequencies. In most cases, it is not possible to know precisely what model actually generated the data at hand. Therefore, a procedure is recommended which combines the safety of the inclusive method with an adjuvant technique called the "continuation ratio" method which has greater power to disclose differences between risk factor effects at different levels of response.