The parameterization has been changed slightly from the original.
The interaction term BT is highly correlated with both Trt and log.Base4.
This induces a strong cross-correlation in the Markov chain between
the corresponding parameters (alpha.BT and alpha.Trt, alpha.BT and
alpha.Base), which in turn leads to very poor mixing.

We can remove the cross-correlation by using, in place of BT, the residual
of BT, after a linear regression on Trt and log.Base4. The R script file
make-BT.R creates the necessary data.

NB By changing the interaction term, we also change the definition of the
main effect terms for log.Base4 and Trt. Accordingly we use different
parameter names alpha.Base0 and alpha.Trt0.  A parameter transformation
is required to get the original parameters alpha.Base and alpha.Trt
that are comparable with the original Breslow and Clayton (1993) results.
