You can order your response variable (underwt, normal, overwt, obese). A reasonable model would be a multinomial logistic model, fit with either PROC GENMOD or PROC LOGISTIC, using a cumulative logistic link. See the documentation for examples (Ordinal Logistic Regression and Logistic Modeling with Categorical Predictors in the PROC LOGISTIC documentation, Ordinal Model for Multinomial Data in the PROC GENMOD documentation).
Steve Denham