Predicted probabilities for nested logit model may be incorrect when input data is not sorted by ID and CHOICE variables in PROC MDC


For a given ID value, the associated nested logit predicted probabilities in the OUT= data set in PROC MDC are incorrect if

(1) the decision variable has missing values for the particular ID value and if

(2) a level 2 (or higher) utility function variable is not constant across the primitive alternatives below each of the nodes at that level and if

(3) the observations for the particular ID value are not sorted by the CHOICE variable.

To circumvent the problem, sort the input data by the ID variable and the CHOICE variable prior to estimating the nested logit model.