Fredrik Clementz <fredrik.clementz@reflect.se> asked for the TWOSTEP CLUSTER procedure via email.

This is Jason Stover's list of desirable statistical procedures from a few years ago.  In his words:

The list is roughly in order of demand, with the most in-demand listed first (in my opinion). Those not in SPSS or SAS are marked with (*) (my knowledge of these is a few years old; SPSS Inc. or SAS Inc. may have since put these into their flagship programs)

1. Generalized linear models, including logistic regression.

1. Time series models, including ARMA and ARIMA.

1. Linear models with other-than-diagonal error covariance matrices.

(These are referred to as 'linear mixed models', though I think is a misleading term.)

1. Clustering and discriminant analysis

1. (*)Neural networks, including multilayer perceptrons and radial-basis functions

1. (*)Classification and regression trees.

1. Spline smoothers

1. Design of experiment

In addition, NO statistical software I know of has a nice way to combine different models (via boosting or other other recently-developed 'data mining' methods).  If PSPP did this with a semi-painless interface for 'users', it could make PSPP much more popular among the 'user' category.  Most people in that 'user' category are dimly aware of the usefulness of this approach to modeling, but they are learning about it, and will want it in their statistical software.