ISSUES:

* gamm4 isn't handling NA's properly with formulae like y~factor(z).
  When inserting z into dataframe, it fails to drop.

* gamm4 (and indeed gamm) will fail if the fixed effects are not identifiable. 
  This can happen, quite easily. e.g.  s(x,by=fac1) + s(x,by=fac2) means that
  the columns x:fac1 and x:fac2 are not independent.

* The computation of the covariance matrix of the response/pseudodata 
  is very resource intensive if the random effects are in only a few groups. 
  When there are only few random effects an alternative computation would 
  be much better.

0.1-6

* added return of R factor from QR of WX.

* cov matrix computation modified to use LAPACK and pivoting properly.

* some changes to replace direct slot acces with getME, but incomplete (y, pWt
  and var still direct).

0.1-5

* gamm4 return object now of class "gamm4"

0.1-4

* upgrade to gamm4.setup to make smooth to r.e. conversion object oriented
  and hence cleaner.

* gamm4 can now deal with "sf" class smooth factor interactions. This gives 
  an efficient way to handle subject specific random smooths.

0.1-3

* Incorrect pivoting of covariance matrix of data/pseudodata could lead 
  to incorrect covariance matrix for coefficients and incorrect EDF 
  computation. Pivoting now corrected, and results tested. 

0.1-2

* Bad bug fix: I'd failed to track an internal lme4 change, so that 
  gamm4 had stopped extracting random effect variances correctly. 
  This meant that gamm4 standard errors were typically too low. 
  Fixed and checks added to test suite to detect this sort of problem.  

0.1-1

* Allow for centering of smooth model matrix columns, when there is an 
  intercept, but columns are not centered by constraint.

0.1-0

* Upgraded to use `t2' type tensor product smooths

* bug fix so that s(...,fx=TRUE) works

* workaround in gamm4 so that g/lmer handles offset properly.

0.0-4

* covariance matrix calculation was still not robust enough. Improved 
further.

0.0-3

* solving for the coefficient covariance matrix could fail under heavy 
  smoothing --- now made more robust. 

* `gamm4' can now be supplied with prior weights. 

* The `cbind(success,failure)' form for a binomial response now works 
  properly.

* help file has been updated for mgcv_1.6-2, and to avoid running to 
  many slow `gamm' calls in checking.

0.0-2

* gamm4 now returns a `scale.estimated' field in its `gam' object
  part.