similar to: predict.lm is broken in 1.1.0-patched (2000-August-7) (PR#628)

On 7 Aug 2000, Peter Dalgaard BSA wrote: > ripley@stats.ox.ac.uk writes: > > > predict.lm has been broken by recent changes to the patched branch. > > > > It fails for all singular fits. An example: <snip> > Sometimes the fix for one bug uncovers another. We have > > > coef.aov > function (object, ...) > { > z <- object$coef >

predict.lm has been broken by recent changes to the patched branch. It fails for all singular fits. An example: library(MASS) data(quine) quine.hi <- aov(log(Days + 2.5) ~ .^4, quine) quine.nxt <- update(quine.hi, . ~ . - Eth:Sex:Age:Lrn) predict(quine.nxt) 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 NA NA NA NA NA NA NA NA NA NA NA NA NA

Confirmation from John Maindonald. -- Brian D. Ripley, ripley@stats.ox.ac.uk Professor of Applied Statistics, http://www.stats.ox.ac.uk/~ripley/ University of Oxford, Tel: +44 1865 272861 (self) 1 South Parks Road, +44 1865 272860 (secr) Oxford OX1 3TG, UK Fax: +44 1865 272595 ---------- Forwarded message ---------- Date: Fri, 3

> From: Peter Dalgaard BSA <p.dalgaard at biostat.ku.dk> > Date: 28 Sep 2000 13:58:22 +0200 > > Peter Dalgaard BSA <p.dalgaard at biostat.ku.dk> writes: > > > I think Occam/Ockham himself wrote in Latin. By my failing memory, the > > quote is > > > > "Entia non runt multiplicanda praeter necessitam" > > > > give or take

I now have versions of code that is destined (I believe) for 0.63 which is in a suitable state for comment. The files are at ftp://ftp.stats.ox.ac.uk/pub/R (Our www server is being moved, so may be intermittently down, but this ftp server should be stable.) All are R packages, for the moment for personal use only (no re-distribution). Use with 0.62.3 or 0.63 (although I am aware of some

In looking over the bugs list, it seems to me that most of them are things we can no longer do anything about, because they are on old versions or broken systems or were features or .... There are now so many they obscure the bugs needing attention. I am about the remove the following. If anyone thinks they are still relevant, could you please send a follow-up to R-bugs with details under 1.2.0

R 0.63.3 (any platform) library(lqs) example(lqs) gives Warning in (y - z$residuals) + offset : longer object length is not a multiple of shorter object length repeatedly. The problem is in lm.wfit, which removes cases with zero weights, but does not adjust the offset. The fix is z$fitted.values <- (y - z$residuals) z$weights <- w if (zero.weights) {

I certainly think we should deprecate it and Peter has (probably unintentionally) offered yet another cogent argument as to why (although that could be fixed by including "offset" among the arguments to be handled by model.frame, like subset, weights, &c, and yes, I do know this would break existing code...) For me the main argument against it, though, is that it de-links the offset

> From: tov@phoenix.ece.cmu.edu > Date: Fri, 16 Jun 2000 16:32:36 +0200 (MET DST) > To: r-devel@stat.math.ethz.ch > Subject: [Rd] Missing -lm for tcltk (R 1.1.0, AIX 4.3) (PR#573) > CC: R-bugs@biostat.ku.dk > X-Loop: R-bugs@biostat.ku.dk > > Hi, > > ok, AIX 4.3 insists on being different. I can't compile --with-tcltk > out of the box. I have to add

Dear Martin, I think that your plan makes sense. It's too bad that aov() behaved differently in this respect from lm(), and thus created more work, but it's not be a bad thing that the difference is now explicit and documented. I expect that that other problems like this will surface, particularly with contributed packages (and I know that you're aware that this has already happened

> To: "Heberto Ghezzo" <heberto at MEAKINS.Lan.McGill.CA> > Subject: Re: [R] R-0.63.1 is released > From: Peter Dalgaard BSA <p.dalgaard at biostat.ku.dk> > > "Heberto Ghezzo" <heberto at MEAKINS.Lan.McGill.CA> writes: > > > Sorry for a silly question but. . . > > why the new version is 'bdr0631' ? what is the

> To: Jonathan.Yuen at tvs.slu.se > Cc: "G.Janacek" <G.Janacek at uea.ac.uk>, r-help at stat.math.ethz.ch > Subject: Re: [R] Installing on DEC 4.0b > From: Peter Dalgaard BSA <p.dalgaard at biostat.ku.dk> > Date: 12 Feb 1999 13:25:18 +0100 > > Jonathan.Yuen at tvs.slu.se writes: > > > I don't know how well it is documented, but you need gnu

On Fri, 6 Aug 1999 ihaka@stat.auckland.ac.nz wrote: > On Fri, Aug 06, 1999 at 12:08:05AM +0200, Peter Dalgaard BSA wrote: > > Ross Ihaka <ihaka@stat.auckland.ac.nz> writes: > > > > I played around with this earlier today. There's nothing special about > > dataframes, S does the same with any classed object. I.e. > > > >

> From: Peter Dalgaard BSA <p.dalgaard at biostat.ku.dk> > Date: 01 Sep 2000 09:54:59 +0200 > > Prof Brian D Ripley <ripley at stats.ox.ac.uk> writes: Important omission: specification from Murray Jorgensen The test that I was thinking of basically does an anova on a modified response variable that is the absolute value of the difference between an observation

> Subject: Re: [Rd] options(keep.source = TRUE) -- also for "library(.)" ? > From: Peter Dalgaard BSA <p.dalgaard@biostat.ku.dk> > Date: 27 Apr 2000 14:37:01 +0200 > > Martin Maechler <maechler@stat.math.ethz.ch> writes: > > > Can we [those of us who know how sys.source() works...] > > think of changing this? As it was possible for the base

> Date: Wed, 30 Jun 1999 11:12:24 +0200 (MET DST) > From: Torsten Hothorn <hothorn at amadeus.statistik.uni-dortmund.de> > > yesterday I had a little shock using qr (or lm). having a matrix > > X <- cbind(1,diag(3)) > y <- 1:3 > > the qr.coef returns one NA (because X is singular). So I computed the > Moore-Penrose inverse of X (just from the

On 5 Aug 1999, Peter Dalgaard BSA wrote: > (1) The docs are clearly wrong (or perhaps rather, wishful thinking) Sorry, but always the docs are right and the implementation wrong. That's an axiom of professional computer programming (so I am told by several ex-professional computer programmers.) > (2) Cbind should be generic as in Splus > > (3) For the labels, one needs

Many thanks for your very useful comments and suggestions. Renaud 2006/5/30, Prof Brian Ripley <ripley at stats.ox.ac.uk>: > On Tue, 30 May 2006, Prof Brian Ripley wrote: > > > This is not really a bug. See > > > > http://developer.r-project.org/model-fitting-functions.txt > > > > for how this is handled in other packages. All model-fitting in R used =

anova() on three or more objects behaves inconsistently in R. In R anovalist.lm does a sequential ANOVA using pairwise F tests, ignoring all the other objects, so the larger of the two models provides the denominator. In S anova.lmlist uses the denominator from the largest model (smallest residual df) in the set, as does anova.glmlist in both. I suggest that R's anovalist.lm is wrong (that

> From: Peter Dalgaard BSA <p.dalgaard@biostat.ku.dk> > > Martin Maechler <maechler@stat.math.ethz.ch> writes: > > > Reading the long output of ``cc -flags'', > > I see that more than average optimization is done using > > cc -xO[1-4] > > # And -O is -xO2. If you do want higher speed, you need to use other flags too, and -fast is a