When using the discriminant method (DISCRIM option in the FCS or MONOTONE statement) in PROC MI, the following message may appear:

ERROR: All generalized squared distances are too big to compute the posterior probabilities with the discriminant method for variable variable. The prior probabilities will be used as the posterior probabilities in the imputation.

One of the steps in the discriminant method involves computing generalized squared distances to obtain the posterior probabilities of an observation belonging to each of the groups. If the distance becomes too large then the probability cannot be computed since it involves using the EXP function. For details, see "Monotone and FCS Discriminant Function Methods" in the Details section of the MI documentation.

Suppose the error occurs with the method option DISCRIM(y1=x1 x2 x3). Specifying the PCOV=FIXED suboption may avoid the error, for example: DISCRIM(y1=x1 x2 x3 / PCOV=FIXED). If not, then since the problem is typically caused by only one or two independent variables, a stepwise process may be needed to find a usable set of variables to use for imputation. Begin with a simple discriminant model — preferably a single independent variable such as DISCRIM(y1=x1). If that is successful, then you can try using additional variables.