Debian Codegraph @

ChangeLog

raw
1 ** denotes quite substantial/important changes
2 *** denotes really big changes
3
4 Liable to change in future release:
5 - method="GCV.Cp" as default in gam. Default will
6 change to "REML". (1.8-24, June 2018)
7
8 Currently deprecated and liable to be removed:
9 - gam.fit (1.9-0, July 2023)
10
11 Issues:
12
13 * openblas 0.3.x x<7 is not thread safe if itself compiled for single thread
14 use and then called from multiple threads (unlike the reference BLAS, say).
15 0.2.20 appears to be OK. For 0.3.x x>6 make USE_THREAD=0 USE_LOCKING=1
16 to make openblas ensures thread safety.
17
18 * t2 in bam(...,discrete=TRUE) - not treated as tensor products at
19 present, and reparameterization needs checking (also for bam).
20
21 * bam(...,discrete=TRUE) has no option to use the identical
22 discretizaton to that used in fitting when predicting
23
24 1.9-4
25
26 *** scasm function provided for shape constrained additive smooth models, via
27 linear and quadratic programming, with smoothness by extended Fellner-Schall.
28
29 * Censored Box-Cox Gaussian, bcg, family introduced.
30
31 * plot.gam now allows plotting of derivatives of 1-D smooth terms, to give
32 direct analogue of GLM regression coefficients, also some re-arrangement to
33 allow use of bootstrap samples for uncertainty bands.
34
35 * predict.gam now allows a 'bs' field of simulated replicate parameter vectors,
36 and direct computation of CIs at different levels.
37
38 * New scheme 2 for default 1-D plots.
39
40 * bam now returns approx REML/ML for the model itself, rather than for the
41 working linearized model (other smoothness criteria still reported for
42 working model only).
43
44 * Improved scale estimation initialization for extended families in bam: can
45 make a noticable difference with 'tw' family in close to Poisson situation.
46
47 * 4-D plotting code axis error fix (y axis labels were being recycled from
48 x axis labels).
49
50 * Long ignored argument 'shade' removed from 'plot.gam'.
51
52 * cnorm/clog single vector response fix.
53
54 * cnorm and cpois improved numerical stability.
55
56 * Fix of one letter typo in gam.outer call to efsud that caused offsets to be
57 ignored. Thanks to Brian Montgomery for reporting the problem.
58
59 * nat.param now warns is model matrix not full rank.
60
61 * sf and sw soap films given new classes to avoid clas=h with sf package
62 - thanks Gavin Simpson.
63
64 * Slight change to gammals scale link to make interpretation more
65 straightforward.
66
67 * gumbls link change and derivative correction.
68
69 * Added option to inSide to allow nmes for x and y to be provided.
70
71 * gam.vcomp modified to have a print method, rather than print itself - thanks
72 to Gavin Simpson.
73
74 * long deprecated function full.score removed.
75
76 1.9-3
77
78 * fixes a read-one-beyond-allocated memory error (in a redundant check with no
79 consequences)
80
81 1.9-2
82
83 ** 'NCV' added as a method option for bam(...,discrete=TRUE). More efficient
84 than gam(...,method='NCV'), but much less efficient than default REML.
85
86 * added basic linear programming routine, lp, mainly to find initial values
87 under linear constraint.
88
89 * default margins changed for plot.gam(...,pages=1)
90
91 * 'clog' censored logistic family added for AFT models. Written by Chris Shen.
92
93 * 'cpois' censored Poisson family added.
94
95 * improved cov matrix estimation under NCV used with correlated errors.
96
97 * names in NCV 'nei' structure modified to better match paper and documentation
98 notation.
99
100 * estimate.theta improved for semi-definite case (e.g. infinite DoF in scat).
101
102 * extended families now give a useful error if an illegal link is supplied.
103
104 * pcls now returns indices of active set constraints as attribute of return
105 vector.
106
107 * fix bam to always return a square QR R factor of the model matrix
108 (previously it was possible for it not to be square for a rank deficient
109 model matrix, which could cause a failure in summary). Thanks to Amy Hughes
110
111 * Fixed bam/bam.update to pick up weights correctly, rather than failing.
112 Thanks to Max Moldovan. Also adjusted to allow correct behaviour if model
113 frame has been deleted.
114
115 * Added sparse matrix Cholesky routine, mchol, a wrapper for
116 Matrix::Cholesky, giving the same functionality as Matrix::chol prior
117 to the decision for this not to return the pivot sequence.
118
119 * 'mroot' modified so that svd option uses an efficient symmetric eigen routine
120 to compute SVD (argument is positive semi-definite after all).
121
122 * fixed gam.vcomp to avoid failure with extended families and bam. Thanks to
123 Nathan LaSpina for reporting.
124
125 * modification of discrete bam prediction method function to better deal with
126 extra factor levels in prediction. Slightly differs from predict.gam, in
127 setting effect to zero, rather than NA.
128
129 * fix to NCV handling of dropped NA points - previously could end up with
130 different numbers of predict and drop neighbourhoods. Thanks to Nicole
131 Augustin.
132
133 * XWXd fix for row col interchange bug introduced with long vector support in
134 1.9-1.
135
136 * fix of gumbls initializtion code bug. Initial parameters were so bad that
137 fit could fail.
138
139 * zoom in bfgs modified so that start values always the same to avoid
140 occasional step failures from repeatedly changing initial values.
141
142 * gevlss improved xi initialization.
143
144 * gam(...,discrete=TRUE) no longer ignores starting coefficient values (thanks
145 to Matteo Fasiolo).
146
147 * Sl.setup and ldetS changes to return totla penalty matrices directly
148 (avoiding a redundant crossproduct)
149
150 * Fix to Pearson residuals in gammals family - thanks to gavin Simpson and
151 Arthur Fendrich
152
153 * ldetS slight efficiency improvement by direct matrix diagonal modification
154 in C code, as suggested by Matteo Fasiolo.
155
156 1.9-1
157
158 * Revised bam discrete code to allow indexing vectors to be long vectors,
159 thereby allowing much larger datasets to be used. n limit is still
160 .Machine$integer.max.
161
162 * Revised C code underlying mvn family to allow larger problems, via use of
163 R long vectors.
164
165 * Revised gfam to work correctly with bam(...,discrete=FALSE). Subsetting
166 in bgam.fit was incorrect for gfam. Thanks to Dave Miller for reporting it.
167
168 * Correction of ziplss residual computation - incorrect probability of non
169 zero used. Thanks to Xiao Liang.
170
171 * correction of ginla integration weights (correcting typo from Rue et al.
172 2009). Thanks to Paul Van Dam-Bates.
173
174 * C code modification to remove some new compiler warnings.
175
176 1.9-0
177
178 *** NCV smoothing parameter estimation now available for most models. See ?NCV.
179
180 *** 'gfam' allows reponse variables to be from several different families.
181
182 * Restriction that number of coefficients must be fewer than number of data
183 removed.
184
185 * Fix to tensor product constructor to allow matrix factor arguments to be
186 handled correctly (previously failed). Thanks to Dave Miller.
187
188 * removed argument 'pers' from 'plot.gam' (deprecated since 1.8-23, Nov 2017).
189
190 * removed "nlm.fd" optimizer (deprecated since 1.8-19, Sept 2017).
191
192 * removed discrete method prediction from discrete bam objects fitted
193 before 1.8-32 (deprecated since then).
194
195 * 'multinom' variance fix for K>=3 cases. A counter was initialized in the wrong
196 place. Thanks to Max Goplerud for finding and reporting this.
197
198 * NCV variance estimation improvements + na handling.
199
200 * 'in.out' argument added for 'bam'.
201
202 * slight internal change to mgcv:::get.var to enable user defined functions
203 in arguments of smooths to work.
204
205 * smooth2random tweak to guarantee same eigenvector sign everytime for same
206 problem (needed by some developers who don't want to carry the original
207 transform around).
208
209 * Some unregistered methods fixes.
210
211 * tests directory removed, as check time was tripping up every CRAN submission.
212
213 * coxph memory error fix - could make bfgs fitting fail.
214
215
216 1.8-42
217
218 * One remaining old style C declaration fixed.
219
220 1.8-41
221
222 ** 'cnorm' family added for left, right, interval or un censored Gaussian
223 data. Useful for log normal Accelerated Failure Time models, Tobit
224 regression, rounded data etc.
225
226 ** 'sz' factor smooth interaction class added for implementing models with
227 main effect smooths and difference smooths for levels of a factor.
228 See ?factor.smooth.
229
230 ** 'NCV' smoothing parameter method added, but still experimental.
231
232 * replacment of some old (K&R) style C function declarations.
233
234 * mono.con corrected for cases with upper bounds. Thanks to Sean Wu.
235
236 * plot.gam(...,seWithMean=TRUE) modified to only include mean uncertainty
237 for the linear predictor of which the smooth is a part, when there are
238 multiple linear predictors. Thanks to Gavin Simpson.
239
240 * modifications of sparce matrix coercions, to avoid deprecated direct
241 coercions. as(as(as(a, "dMatrix"), "generalMatrix"), "XsparseMatrix") in
242 place of as(a,"dgXMatrix") where 'X' is 'C' or 'T'. Actually this requires
243 Matrix 1.4-2 to work (will be added to dependencies in future).
244
245 * vis.gam now deals properly with models with more than one linear predictor.
246
247 * slight change in bgam.fitd to check scale parameter estimate converged when
248 using bam(...,discrete=TRUE), otherwise scale could be wrong for all fixed
249 smoothing parameters.
250
251 * predict.gam modified so that 'terms' and 'exclude' control all terms, smooth
252 or parametric, in the same way. Including the "(Intercept)" term.
253
254 * Warnings from 'model.matrix' suppressed in 'terms2tensor' called by e.g.
255 predict.bamd. There is a warning if any extra contrasts are supplied to
256 model matrix that do not relate to a term in the model (which contradicts
257 the documentation). Doc vs code bug report also filed.
258
259 * Fix of broken rank deficiency handling in gam.fit5. Thanks to Cesko Voeten.
260
261 * trind.generator modified to allow return of index functions in place
262 of index arrays.
263
264 * summary.gam (recov/reTest) modified to deal with 'fs' smooths fitted using
265 'gamm'.
266
267 * gam.fit4 convergence testing improved and bug fix in computation of dVkk
268 matrix used to check for converged 'infinite' smoothing parameters in bfgs.
269
270 1.8-40
271
272 * Small gam.fit5 convergence change to reduce chance of repeating step
273 failures.
274
275 * ExtractData fix avoiding assumption that 'xt' argument to a smooth is always
276 a list (could cause setup failure with e.g. fs smooth). Thanks Keith Woolner.
277
278 * Minor changes to mat.c and tprs.c to allow DEFS = -DSTRICT_R_HEADERS build.
279
280 1.8-39
281
282 * gam.fit5 convergence logic tightened up, to avoid pointlessly lengthy
283 iterations when can't meet convergence tolerance. Faster, more reliable,
284 but also generates more warnings when tolerances not met.
285
286 * gam.fit3 convergence test modified to more reliable version.
287
288 * Some modification of warning handling to only print inner optimization
289 warnings if they occur at final call to inner optimizer.
290
291 * newton and bfgs optimizers reset inner loop tolerance to 1e-2 of outer
292 loop tolerance if it is larger than this, to avoid inner loop being too
293 inaccurate for outer tolerance.
294
295 * minor R 4.2.0 C compatibility change.
296
297 1.8-38
298
299 * uniquecombs fix to reduce memory footprint for text data.
300
301 1.8-37
302
303 * minor update for clang compatibility.
304
305 * PredictMat fix of bug in which dimensions could be dropped for 1 row matrix
306 as part of constraint handling (leading to disaster in predict.bamd).
307 Thanks to Shawn Ligocki.
308
309
310 1.8-36
311
312 * 'fs' smooth construction modified to (approximately) orthogonalise
313 smoothing penalty penalized and unpenalized bases. This makes the
314 assumption that the associated variance components are independent
315 more natural. Thanks to Matteo Fasiolo and Harald Bayaan for showing
316 how the original construction could be problematic.
317
318 * correction to 'fs' smooth handling in gamm to allow correlation structures
319 to be used in the same model without cuasing an error.
320
321 * AIC calculation modified for non-Gaussian families with a scale parameter,
322 so that the model estimated scale parameter is used, rather than the
323 deviance estimator employed with glm exponential families. This matters for
324 avoiding substantial bias with very low mean count data (particulalry 'tw'
325 and 'Tweedie' families).
326
327 * ziP family response scale se fix for case when b>0 (thanks to Mark Donoghoe)
328
329 * prediction on response scale could fail for models fitted by
330 bam(...,discrete=TRUE) - fixed.
331
332 * bug fix to multi-model anova.gam code to deal properly with extended family
333 scale parameters (still not the recommended way of testing!)
334
335 1.8-35
336
337 * Fix to bug in Fletcher scale parameter estimate with weighted data (e.g.
338 quasibinomial with n>1). Gaussian case not affected. Thanks to John
339 Maindonald.
340
341 * Italian translation updated thanks to Daniele Medri
342
343 * German translation updated thanks to Detlef Steuer
344
345 * French translation updated thanks to Philippe Grosjean
346
347 * gumbls initialization improved.
348
349 * Modification so that smooths can insist on not being reparameterized when
350 called using default bam methods, by having a 'repara' element set to FALSE
351 in the object returned by the smooth constructor.
352
353 * Modification so that smooths can have the summation convention for matrix
354 arguments turned off, by having an 'xt$sumConv' element set to FALSE in the
355 'xt' argument of 's'.
356
357 1.8-34
358
359 * 'gam.mh' proposal df typo fix and help file '%*%' -> '\%*\*' fix. Thanks
360 Len Thomas.
361
362 * various random number generation calls in rd functions in families could
363 fail if simulating one point - fixed.
364
365 * gp smooths can now be specified to be strictly stationary (see ?gp.smooth)
366
367 * fix to predict.bamd bug in handling fixed effect contrasts (wrong contrast
368 could be used for prediction, if non-default used in fitting). Thanks to
369 Jalal Al-Tamimi.
370
371 * predict.bamd modified to catch case where all prediction data is NA
372
373 * predict.gam modified to allow NA to be a factor level.
374
375 1.8-33
376
377 * 'inline' -> 'static inline' error fix - caused installation failure on some
378 platforms.
379
380 1.8-32
381
382 ** bam(...,discrete=TRUE) now uses discretization on the parametric model
383 components as well as the smooths.
384
385 * 'gumbls' family added for Gumbel location scale models.
386
387 * 'shash' location scale and shape family added.
388
389 * psum.chisq function added to compute c.d.f. of weighted sums of chi-squared
390 r.v.s using method of Davies, 1980. This is now used for computing p-values
391 in summary.gam in place of the Liu et al, 2009 approximation.
392
393 * score and Schoenfeld residuals added for cox.ph to facilitate PH assumption
394 checking (see ?cox.ph).
395
396 * added 'gam.mh' for posterior dampling from models fitted with 'gam'.
397
398 * 'ocat' upgraded to not ignore weights.
399
400 * added multivariate t and Gaussian densities and multivariate t generation.
401
402 * fix to predict.bamd to correct handling of `terms' and `exclude' arguments.
403
404 * A recent lme change broke gamm offset handling. Fixed.
405 Thanks Aaron Benjamin Shev.
406
407 * dgesvd LAPACK calls replaced by faster dgesdd, as some versions of MKL
408 BLAS/LAPACK have broken dgesvd.
409
410 * Fix to derivative w.r.t. scale parameter calculation for extended families
411 ('tw' is the only one at present) when using 'ML' smoothing parameter
412 estimation. Could lead to step failure with 'tw'. Thanks Kevin Hawkshaw.
413
414 * Fix to prediction/fake formula environment that could cause failure of
415 gam/bam called from a function. Thanks to Duncan Murdoch.
416
417 * Fix to smoothing parameter uncertainy correction for general families.
418
419 * Much cleaner design for the discretization of covariates and associated
420 discretization index matching in bam (predict.bamd and discrete.mf).
421
422 * Fix to jagam logic for identifying separable penalties which led to
423 wrong partioning of penalties for 'fs' smooths used with jagam. Thanks to
424 Chris Jackson for spotting the problem.
425
426 * Internal Sl list handling changes.
427
428 * tensor.prod.model.matrix now also works with sparse matrices of class
429 "dgCMatrix".
430
431 1.8-31
432
433 * fix in initalization in gammPQL
434
435 * fix of some C routines of type void in place of SEXP called by .Call.
436
437 1.8-30
438
439 * anova.gam now uses GLRT for multi-model comparisons in both extended and
440 general family cases.
441
442 * Fix to bug in bam(...,discrete=TRUE) offset handling introduced in 1.8-29,
443 which corrupted offset (and generated numerous warnings). Also fixes a
444 less obvious bug introduced at the same time in predict.gam which could get
445 the offset wrong when predicting with such models. Thanks to Brian
446 Montgomery and Sean Wilson.
447
448 * Fix to problem in pen.reg (used to initialize location scale models in
449 particular), which could lead to initialization failure for lightly
450 penalized models. Thanks Matteo Fasiolo.
451
452 * Fix to predict.bamd handling of 'terms' and 'exclude' for models fit by
453 bam(...,discrete=TRUE).
454
455 * Work around in predict.gam for a spurious model.matrix warning when a
456 contrasts.arg relates to a variable in data not required by object.
457
458 * Fix to gammals family which had 'd2link' etc defined with argument 'eta'
459 instead of the intended 'mu'. Thanks to Jim Stagge.
460
461 * 'in.out' modified to allow boundaries specified exactly as for a soap film
462 smoother.
463
464 * soap film smoother constructor modified to check knots are within boundary
465 on entry and drop those that are not, with a warning. Also halts if it
466 detects that basis setup is catastrophically ill-conditioned.
467
468 1.8-29
469
470 * gammPQL modified to use standard GLM IRLS initilization (rather than
471 glmmPQL method) to improve convergence of `gamm' fits.
472
473 * bam(...,discrete=TRUE) now drops rownames from parametric model matrix
474 components, to save substantial memory.
475
476 * All BLAS/LAPACK calls from C now explicitly pass hidden string length
477 arguments to avoid breakage by recent gfortran optimizations (stack
478 corruption causing BLAS/LAPACK to return error code).
479
480 * predict.gam bug fix - parameteric interaction terms could be dropped for
481 type="terms" if there were no smooths. knock on was that they were also
482 dropped for all bam(...,discrete=TRUE) fits. (Thanks Justin Davis.)
483
484 * bam(...,discrete=TRUE) indexing bug in setup meant that models containing
485 smooths with matrix arguments and other smooths with factor by variable
486 would fail at the setup stage.
487
488 * gam.fit4 initial divergence bug fix.
489
490 * Gamma location-scale family 'gammals' added. See ?gammals.
491
492 * row-wise Kronecker product operator %.% added for convenience.
493
494 * changes to general families to allow return of first deriv of penalized
495 Hessian component more easily.
496
497 * ocat bug fix. Response scale prediction was wrong - it ignored the estimated
498 thresholds. Thanks to Fabian Scheipl.
499
500 * bam deviance could be wrongly returned, leading to 100% explained
501 deviance. Fixed.
502
503 1.8-28
504
505 * fix of obscure sp naming bug.
506
507 * changed some contour default colours from green to blue (they overlay
508 heatmaps, so green was not clever).
509
510 * Tweedie likelihood evaluation code made slightly more robust - for a model
511 with machine zero scale parameter estimate it could segfault, as series
512 maximum location could then overflow integer storage. Fixed + upper limit
513 imposed on series length (warning if it's not enough).
514
515 1.8-27
516
517 ** Added routine 'ginla' for fully Bayesian inference based on an integrated
518 nested Laplace approximation (INLA) approach. See ?ginla.
519
520 * Tweedie location scale family added: 'twlss'.
521
522 * gam.fit5 modified to distinguish more carefully between +ve semi definite
523 and +ve definite. Previously could fail, claiming indefiniteness when it
524 should not have. Affects general families.
525
526 * bam was ignoring supplied scale parameter in extended family cases - fixed.
527
528 * work around in list formula handling for reformulate sometimes setting
529 response to a name in place of a call.
530
531 * preinitialize in general families is now a function, not an expression. See
532 cox.ph for an example.
533
534 * added routine cholup for rank one modification of Cholesky factor.
535
536 * two stage gam/bam fitting now allows 'sp' to be modified.
537
538 * predict.gam could fail with type="response" for families requiring the
539 response to be provided in this case (e.g. cox.ph). Fixed.
540
541 * sp.vcov defaults to extracting edge corrected log sp cov matrix, if
542 gam(...,gam.control(edge.control=TRUE)) used for fitting.
543
544 * gam(...,gam.control(edge.correct=TRUE)) could go into infinite loop if
545 sp was effectively zero. Corrected.
546
547
548 1.8-26
549
550 * LINPACK dependency removed.
551
552 * Added service routine choldrop to down date a Cholesky factor on row/col
553 deletion.
554
555 * liu2 had a check messed up when vectorized. Fix to stop vector being
556 checked for equality to zero.
557
558 1.8-25
559
560 ** bam(...,discrete=TRUE) methods improved. Cross products now usually faster
561 (can be much faster) and code can now make better use of optimised BLAS.
562
563 * fix to 'fs' smooth.construct method and smooth2random method, to allow
564 constructor to be called without a "gamm" atribute set on the smooth spec
565 but still get a sensible result from smooth2random (albeit never using
566 sarse matrices). Useful for other packages using constructors and
567 smooth2random, for 'fs' smooths.
568
569 * The mrf smooth constructor contained an obsolete hack in which the term
570 dimension was set to 2 to avoid plotting when used as a te marginal. This
571 messed up side constraints for terms where a mrf smooth was a main effect
572 and te marginal. Fixed.
573
574 * extract.lme.cov/2 documentation modified to cover NA handling properly, and
575 routines modified to not require data to be supplied.
576
577 * fix of efsudr bug whereby extended families with no extra parameters to
578 estimate could give incorrect results when using optimer="efs" in 'gam'.
579
580 * negbin() corrected - it was declaring the log link to be canonical, leading
581 to poor convergence and slight misfit.
582
583 * predict.bam(...,discrete=TRUE) now handles na.action properly, rather than
584 always dropping NAs.
585
586 * Fix of very obscure bug in which very poor model of small dataset could
587 end up with fewer `good' data than coefs, breaking an assumption of C code
588 and segfaulting.
589
590 * fix of null deviance computation bug introduced with extended families in
591 bam - null deviance was wrong for non-default methods.
592
593 * liu2 modified to deal with random effects estimated to be exactly 0, so
594 that summary.gam does not fail in this case.
595
596 1.8-24
597
598 * Extended Fellner Schall optimizer now avaialable for all families with 'gam'
599 using gam(...,optimizer="efs").
600
601 * Change to default behaviour of plot.gam when 'seWithMean=TRUE', and of
602 predict.gam when 'type="iterms"'. The extra uncertainty added to CIs or
603 standard errors now reflects the uncertainty in the mean in all other model
604 terms, not just the uncertanity in the mean of the fixed effects as before.
605 See ?plot.gam and ?predict.gam (including for how to get the old behaviour).
606
607 * 're' smooths can now accept matrix arguments: see ?linear.functional.terms.
608
609 * cox.ph now allows an offset to be provided.
610
611 * Fix in smoothCon for bug in case in which only a single coefficient is
612 involved in a sum-to-zero constraint. Could cause failure in e.g. t2 with
613 cc marginal.
614
615 * Model terms s, te etc are now always evaluated in mgcv workspace explicitly
616 to avoid masking problems in obscure circumstances.
617
618 * 'mrf' smooth documentation modified to make it clearer how to specify 'nb',
619 and code modified so that it is now possible to specify the neighbour
620 structure using names rather than indices.
621
622 * 'bfgs' fix to handle extended families.
623
624 * plot.gam modified to only prompt (via devAskNewPage) for a new page after
625 the first page is used up.
626
627 * export 'k.check'.
628
629 * Fix to 'Rrank'. Previously a matrix R with more columns than rows could
630 cause a segfault.
631
632 * Fix to non-finite likelihood handling in gam.fit5.
633
634 * Fix in bgam.fitd to step reduce under indefinite deviance and to ensure
635 penalty evaluation is round off negative proof.
636
637 * newton slighty modified to avoid (small) chance of all sp's being dropped
638 for indef likelihood.
639
640 1.8-23
641
642 * default plot methods added for smooths of 3 and 4 variables.
643
644 * The `gamma' control parameter for gam and bam can now be used with RE/ML
645 smoothness selection, not just GCV/AIC. Essentially smoothing parameters
646 are chosen as if the sample size was n/gamma instead of n.
647
648 * The "bs" basis now allows multiple penalties of different orders on the same
649 spline. e.g. s(x,bs="bs",m=c(3,2,0)). See ?b.spline.
650
651 * bam(...,discrete=TRUE) can now use the smooth 'id' mechanism to link
652 smoothing parameters, but note the method constraint that the linked bases
653 are not forced to be identical in this case (unlike other fitting methods).
654
655 * summary.gam now allows random effects tests to be skipped (in some large
656 models the test is costly and uninteresting).
657
658 * 'interpret.gam0' modified so that masked 's', 'te', etc from other packages
659 does not cause failure.
660
661 * coxph fix of prediction bug introduced with stratified model (thanks
662 Giampiero Marra)
663
664 * bam(...,discrete=TRUE) fix to handle nested matrix arguments to smooths.
665
666 * bam(...,discrete=TRUE) fix to by variable handling with fs and re smooths
667 which could fail during re-representation as tensor smooths (for
668 discretization purposes).
669
670 * bam extended family extension had introduced a bug in null deviance
671 computation for Gaussian additive case when using methods other than fREML
672 or GCV.Cp. Fixed.
673
674 * bam(...,discrete=TRUE) now defaults to discrete=FALSE if there are no
675 smooths, rather than failing.
676
677 * bam was reporting wrong family under some smoothing parameter selection
678 methods (not default).
679
680 * null deviance computation improved for extended families. Previous version
681 used an approximation valid for most families, and corrected the rest - now
682 replaced with exact computations for all cases.
683
684 * scat initialization tweaked to avoid -ve def problems at start.
685
686 * paraPen handling in bam was broken - fixed.
687
688 * slight adjustment to sp initialization for extended families - use observed
689 information in weights if possible.
690
691 1.8-22
692
693 * Fix of bug whereby testing for OpenMP and nthreads>1 in bam, would fail if
694 OpenMP was missing.
695
696 1.8-21
697
698 * When functions were added to families within mgcv some very large
699 environments could end up attached to those functions, for no good reason.
700 The problem originated from the dispatch of the generic 'fix.family.link'
701 and then propagated via fix.family.var and fix.family.ls. This is now avoided,
702 resulting in smaller gam objects on disk and lower R memory usage.
703 Thanks to Niels Richard Hansen for uncovering this.
704
705 * Another bug fix for prediction from discrete fit bam models with an offset,
706 this time when there were more than 50000 data. Also fix to bam fitting when
707 the number of data was an integer multiple of the chunk size + 1.
708
709 * check.term was missing a 'stop' so that some unhandled nesting structures
710 in bam(...,discrete=TRUE) failed with an unhelpful error, instead of a
711 helpful one. Fixed.
712
713 1.8-20
714
715 * bam(,discrete=TRUE) could produce garbage with ti(x,z,k=c(6,5),mc=c(T,F))
716 because tensor re-ordering for efficiency failed to re-order mc (this is
717 a *very* specialist bug!). Thanks to Fabian Scheipl.
718
719 * plot(...,residuals=TRUE) weighted the working residuals by the sqrt working
720 weights divided by the mean sqrt working weight. The standardization by the
721 mean sqrt weight was non standard and has been removed.
722
723 * Fix to bad bug in bam(...,discrete=TRUE) offset handling, and predict.bamd
724 modified to avoid failure predicting with offset. Thanks to Paul Shearer.
725
726 * fix of typo in bgam.fit, which caused failure of extended families
727 when dataset larger than chunk size. Thanks Martijn Wieling.
728
729 * bam(...,discrete=TRUE)/bgam.fitd modified to use fisher weights with
730 extended families if rho!=0.
731
732 1.8-19
733
734 ** bam() now accepts extended families (i.e. nb, tw, ocat etc)
735
736 * cox.ph now allows stratification (i.e. baseline hazard can differ between
737 groups).
738
739 * Example code for matched case control in ?cox.ph was just plain wrong. Now
740 fixed. Thanks to Patrick Farrell.
741
742 * bam(...,discrete=TRUE) tolerance for judging whether smoothing parameters
743 are on boundary was far too low, so that sps could become so large that
744 numerical instability set in. Fixed. Thanks to Paul Rosenfield.
745
746 * p.type!=0 removed in summary.gam (previously deprecated)
747
748 * single penalty tensor product smooths removed (previously deprecated).
749
750 * gam(...,optimizer="perf") deprecated.
751
752 * extra divergence check added to bam gam default gam fitting (similar to
753 discrete method).
754
755 * preinitialize and postproc components of extended families are now functions,
756 not expressions.
757
758 * coefficient divergence check was missing in bam(...,discrete=TRUE) release
759 code - now fixed.
760
761 * gaulss family link derivatives modified to avoid overflow. Thanks to
762 Kristen Beck for reporting the problem.
763
764 * redundant 'n' argument removed from extended family 'ls' functions.
765
766 * Convergence checking can step fail earlier in fast.REML.fit. If trial step
767 is no improvement and equal to previous best (to within a tolerance), then
768 terminate with step failure after a few step halvings if situation persists.
769 Thanks to Zheyuan Li for reporting problem.
770
771 1.8-18
772
773 * Tweak to 'newton' to further reduce chance of false convergence at
774 indefinite point.
775
776 * Fix to bam.update to deal with NAs in response.
777
778 * 'scat' family now takes a 'min.df' argument which defaults to 3. Could
779 otherwise occasionally have indefinite LAML problems as df headed towards 2.
780
781 * Fix to `gam.fit4' where in rare circumstances the PIRLS iteration could
782 finish at an indefinite point, spoiling implicit differentiation.
783
784 * `gam.check' modified to fix a couple of issues with `gamm' fitted models, and
785 to warn that interpretability is reduced for such models.
786
787 * `qq.gam' default method slight modification to default generation of reference
788 quantiles. In theory previous method could cause a problem if enough
789 residuals were exactly equal.
790
791 * Fix to `plot.mrf.smooth' to deal with models with by variables.
792
793 * `plot.gam' fix to handling of plot limits when using 'trans' (from 1.8-16
794 'trans' could be applied twice).
795
796 * `plot.gam' argument 'rug' now defaults to 'NULL' corresponding to 'rug=TRUE'
797 if the number of data is <= 10000 and 'rug=FALSE' otherwise.
798
799 * bam(...,discrete=TRUE) could fail if NAs in the smooth terms caused data rows
800 to be dropped which led to parametric term factors having unused levels
801 (which were then not being dropped). Fixed (in discrete.mf).
802
803 * bam(...,discrete=TRUE,nthreads=n) now warns if n>1 and openMP is not
804 available on the platform being used.
805
806 * Sl.addS modified to use C code for some otherwise very slow matrix
807 subset and addition ops which could become rate limiting for
808 bam(...,discrete=TRUE).
809
810 * Parallel solves in Sl.iftChol can speed up bam(...,discrete=TRUE) with
811 large numbers of smoothing/variance parameters.
812
813 * 'gamm' now warns if called with extended families.
814
815 * disasterous 'te' in place of 'ti' typo in ?smooth.terms fixed thanks to
816 John McKinlay.
817
818 * Some `internal' functions exported to facilitate quantile gam methods
819 in separate package.
820
821 * Minor fix in gam.fit5 - 1 by 1 matrix coerced to scalar, to prevent failure
822 in some circumstances.
823
824
825 1.8-17
826
827 * Export gamlss.etamu, gamlss.gH and trind.generator to facilitate user
828 addition of new location-scale families.
829
830 * Re-ordering of initialization in gam.fit4 to avoid possible failure of
831 dev.resids call before initialization.
832
833 * trap in fast.REML.fit for situation in which all smoothing parameters
834 satisfy conditions for indefinite convergence on entry, with an
835 immediate warning that this probably indicates iteration divergence (of bam).
836
837 * "bs" basis modified to allow easier control of the interval over which the
838 spline penalty applies which in turn allows more sensible control of
839 extrapolation behaviour, when this is unavoidable.
840
841 * Fix in uniquecombs - revised faster code (from 1.8-13) could occasionally
842 generate false matches between different input combinations for integer
843 variables or factors. Thanks to Rohan Sadler for reporting the issue that
844 uncovered this.
845
846 * A very bad initial model for uninformative data could lead to a negative
847 fletcher estimate of the scale parameter and optimizer failure - fixed.
848
849 * "fREML" allowed in sp.vcov so that it works with bam fitted models.
850
851 * 2 occurances of 'return' replaced by (correct) return().
852
853 1.8-16
854
855 * slightly improved intial value heuristics for overlapping penalties in
856 general family case.
857
858 * 'ocat' checks that response is numeric.
859
860 * plot.gam(...,scale=-1) now changes scale according to 'trans' and 'shift'.
861
862 * newton optimizer made slightly more cautious: contracts step if reduction
863 in true objective too far different from reduction predicted by quadratic
864 approximation underlying Newton step. Also leaves parameters unchanged
865 in Newton step while their grad is less than 1% of max grad.
866
867 * Fix to Fisher weight computation in gam.fit4. Previously a weight could
868 (rarely) evaluate as negative machine prec instead of zero, get passed to
869 gdi2 in C code, generate a NaN when square rooted, resulting in a NaN passed
870 to the LAPACK dgeqp3 routine, which then hung in a non-interuptable way.
871
872 * Fix of 'sp' argument handling with multiple formulae. Allocation to terms
873 could be incorrect.
874
875 * Option 'edge.correct' added to 'gam.control' to allow better correction
876 of edge of smoothing parameter space effects with 'gam' when RE/ML used.
877
878
879 * Fix to setting of penalty rank in smooth.construct.mrf.smooth.spec.
880 Previously this was wrong, which could cause failure with gamm if the
881 penalty was rank deficient. Thanks Paul Buerkner.
882
883 * Fix to Vb.corr call from gam.fit3.post.proc to ensure that sp not
884 dropped (wrongly treated as scale estimate) when P-REML or P-ML used.
885 Could cause failure depending on BLAS. Thanks Matteo Fasiolo.
886
887 * Fix in gam.outer that caused failure with "efs" optimizer and fixed sps.
888
889 * Fix to `get.var' to drop matrix attributes of 1 column matrix variables.
890
891 * Extra argument added to `uniquecombs' to allow result to have same row
892 ordering regardless of input data ordering. Now used by smooth constructors
893 that subsample unique covariate values during basis setup to ensure
894 invariance to data re-ordering.
895
896 * Correction of scaling error in spherical correlation structure GP smooth.
897
898 * qf and rd functions for binomial family fixed for zero n case.
899
900 1.8-15
901
902 * Fix of survival function prediction in cox.ph family. Code used expression
903 (8.8.5) in Klein and Moeschberger (2003), which is missing a term. Correct
904 expression is, e.g., (10) from Andersen, Weis Bentzon and Klein (1996)
905 Scandinavian Journal of Statistics.
906
907 * Added help file 'cox.pht' for Cox PH regression with time dependent
908 covariates.
909
910 * fix of potential seg fault in gdi.c:get_bSb if single smooth model
911 rank deficient (insufficient workspace allocated).
912
913 * gam.fit5 modified to step half if trial penalized likelihood is infinite.
914
915 * Fix so that bam works properly with drop.intercept=TRUE.
916
917 1.8-14
918
919 * bug fix to smoothCon that could generate NAs in model matrix when using bam
920 with numeric by variables. The problem was introduced as part of the
921 bam(...,discrete=TRUE) coding.
922
923 1.8-13
924
925 * Added help file ?one.se.rule on the `one standard error rule' for obtaining
926 smoother models.
927
928 * bam(...,discrete=TRUE) no longer complains about more coefficients than data.
929
930 * 's', 'te', 'ti' and 't2' modified to allow user to specify that the smooth
931 should pass through zero at a specified point. See ?identifiability.
932
933 * anova.gam modified to use more appropriate reference degrees of freedom
934 for multiple model call, where possible. Also fixed to allow multiple
935 formulae models and to use -2*logLik in place of `deviance' for
936 general.family models.
937
938 * offsets allowed with multinomial, ziplss and gaulss families.
939
940 * gevlss family implementing generalized extreme value location, scale and
941 shape models.
942
943 * Faster code used in 'uniquecombs'. Speeds up discretization step in
944 'bam(...,discrete=TRUE)'. Could still be improved for multi-column case.
945
946 * modification to 'smoothCon' to allow resetting of smooth supplied
947 constraints - enables fix of bug in bam handling of 't2' terms, where
948 parameterization of penalty and model matrix did not previously match
949 properly.
950
951 * clarification of `exclude' argument to predict.gam in help files.
952
953 * modification to 'plot.gam' etc, so that 'ylim' is no longer shifted by
954 'shift'.
955
956 * ylim and ... handling improved for 'fs' plot method (thanks Dave Miller)
957
958 * gam.check now recognises RStudio and plots appropriately.
959
960 * bam(...,sparse=TRUE) removed - not efficient, because of unavoidability
961 of dense off diagonal terms in X'X or equivalent. Deprecated since 1.8-5.
962
963 * tweak to initial.sp/g to avoid infinite loop in s.p. initialization, in
964 rather unusual circumstances. Thanks to Mark Bravington.
965
966 * bam and gam have `drop.intercept' argument to force the parametric terms not
967 to include a constant in their span, even when there are factor variables.
968
969 * Fix in Vb.corr (2nd order edf correction) for fixed smoothing parameter case.
970
971 * added 'all.vars1' to enable terms like s(x$y) in model formulae.
972
973 * modification to gam.fit4 to ignore 'start' if it is immediately worse than
974 'null.coef'.
975
976 * cSplineDes can now accept a 'derivs' argument.
977
978 * added drop.intercept handling for multiple formulae (mod by Matteo Fasiolo).
979
980 * 'gam.side' fix to avoid duplication of smooth model matrices to be tested
981 against, when checking for numerical degeneracy. Problem could occasionally
982 cause a failure (especially with bam), when the total matrix to be tested
983 against ended upo with more columns than rows.
984
985 * 4 occurances of as.name("model.frame") changed to quote(stats::model.frame)
986
987 * fix in predict.bamd discrete prediction code to be a bit more relaxed about
988 use of as.factor, etc in formulae.
989
990 * fix in predict.gam handling of 'na.action' to avoid trying to get type of
991 na.action from name of na.action function - this was fragile to nesting
992 and could cause predict.bam to fail in the presence of NA's.
993
994 * fix of gam.outer so that general families (e.g. cox.ph) can have all
995 their smoothing parameters supplied without then ignoring the penalties!
996
997 * fix in multiple formulae handling of fixed smoothing parameters.
998
999 * Fix of bug in zlim handling in vis.gam perspective plot with standard
1000 errors. Thanks Petra Kuhnert.
1001
1002 * probit link added to 'jagam' (suggested by Kenneth Knoblauch).
1003
1004 * 'Sl.' routines revised to allow operation with non-linearly parameterized
1005 smoothers.
1006
1007 * bug fix in Hessian computation in gam.fit5 - leading diagonal of Hessian of
1008 log|Hp| could be wrong where Hp is penalized Hessian.
1009
1010 * better use of crossprod in gamlss.gH
1011
1012 1.8-12
1013
1014 ** "bs" B-spline smoothing basis. B-splines with derivative based penalties
1015 of various orders.
1016
1017 * 'gamm' now uses a fixed scale parameter in PQL estimation for Poisson and
1018 binomial data via the `sigma' option in lmeControl.
1019
1020 * bam null deviance computation was wrong with prior weights (including
1021 binomial other than binary case), and returned deviance was wrong for
1022 non-binary binomial. Fixed (did not affect estimation).
1023
1024 * improvements to "bfgs" optimizer to better deal with `infinite' smoothing
1025 parameters.
1026
1027 * changed scheme=3 in default 2-D plotting to grey scale version of
1028 scheme=2.
1029
1030 * 'trichol' and 'bandchol' added for banded Cholesky decompositions, plus
1031 'sdiag' functions added for extracting and setting matrix sub- and
1032 super- diagonals.
1033
1034 * p-spline constructor and Predict.matrix.pspline.smooth now allow set
1035 up of SCOP-spline monotonic smoothers, and derivatives of smooths. Not
1036 used in modelling functions yet.
1037
1038 * s(...,bs="re") now allows known precision matrix structures to be defined
1039 using the `xt' argument of 's' see ?smooth.construct.re.smooth.spec for
1040 details and example.
1041
1042 * negbin() with a grid search for `theta' is no longer supported - use
1043 'nb' instead.
1044
1045 * bug fix to bam aic computation with AR rho correction.
1046
1047 1.8-11
1048
1049 * bam(...,discrete=TRUE) can now handle matrix arguments to smooths (and hence
1050 linear functional terms).
1051
1052 * bam(...,discrete=TRUE) bug fix in fixed sp handling.
1053
1054 * bam(...,discrete = TRUE) db.drho reparameterization fix, fixing nonsensical
1055 edf2. Also bam edf2 limited to maximum of edf1.
1056
1057 * smoothCon rescaling of S changed to use efficient matrix norm in place of
1058 relatively slow computation involving model matrix crossproduct.
1059
1060 * bam aic corrected for AR model if present.
1061
1062 * Added select=TRUE argument to 'bam'.
1063
1064 * Several discrete prediction fixes including improved thread safety.
1065
1066 * bam/gam name gcv.ubre field by "method".
1067
1068 * gam.side modified so that if a smooth has 'side.constrain==FALSE' it is
1069 neither constrained, nor used in the computation of constraints for other
1070 terms (the latter part being new). Very limited impact!
1071
1072 * No longer checks if SUPPORT_OPENMP defined in Rconfig.h, but only if _OPENMP
1073 defined. No change in actual behaviour.
1074
1075 1.8-10
1076
1077 ** 'multinom' family implemented for multinomial logistic regression.
1078
1079 * predict.bam now defaults to using efficient discrete prediction methods
1080 for models fit using discrete covariate methods (bam(...,discrete=TRUE)).
1081
1082 * with bam(...,discrete=TRUE) terms like s(a,b,bs="re") had wrong p-value
1083 computation applied, as a result of being treated as tensor product terms.
1084 Fixed.
1085
1086 * minor tweak to soap basis setup to avoid rounding error leading to 'approx'
1087 occasionally producing NA's with fixed boundaries.
1088
1089 * misc.c:rwMatrix made thread safe (had been using R_chk_calloc, which isn't).
1090
1091 * some upgrading for 64bit addressing.
1092
1093 * uniquecombs now preserves contrasts on factors.
1094
1095 * variable summary tweak so that 1 column matrices in parametric model are
1096 treated as regular numeric variables.
1097
1098 1.8-9
1099
1100 * C level fix in bam(...,discrete=TRUE) code. Some memory was mistakenly
1101 allocated via 'calloc' rather than 'R_chk_calloc', but was then freed
1102 via 'R_chk_free'. This could cause R to halt on some platforms.
1103
1104 1.8-8
1105
1106 ** New "gp" smooth class (see ?gp.smooth) implemeting the Matern
1107 covariance based Gaussian process model of Kamman and Wand (2003), and a
1108 variety of other simple GP smoothers.
1109
1110 * some smooth plot methods now accept 'colors' and 'contour.col' argument to
1111 set color palette in image plots and contour line colors.
1112
1113 * predict.gam and predict.bam now accept an 'exclude' argument allowing
1114 terms (e.g. random effects) to be zeroed for prediction. For efficiency,
1115 smooth terms not in 'terms' or in 'exclude' are no longer evaluated, and
1116 are instead set to zero or not returned. See ?predict.gam.
1117
1118 * ocat saturated likelihood definition changed to zero, leading to better
1119 comprability of deviance between model fits (thanks to Herwig Friedl).
1120
1121 * null.deviance calculation for extended families modified to make more sense
1122 when `mu' is the mean of a latent variable, rather than the response itself.
1123
1124 * bam now returns standarized residuals 'std.rsd' if `rho!=0'.
1125
1126 * bam(...,discrete=TRUE) can now handle 'fs' terms.
1127
1128 * bam(...,discrete=TRUE) now accepts 'by' variables. Thanks to Zheyaun Li
1129 for debugging on this.
1130
1131 * bam now works with drop.unused.levels == TRUE when random effects should
1132 have more levels than those that exist in data. (Thanks Alec Leh)
1133
1134 * bam chunk.size logic error fix - error could be triggered if chunk.size
1135 reset automaticlly to be larger than data size.
1136
1137 * uniqucombs can now accept a data frame with some or all factor columns,
1138 as well as purely numeric marices.
1139
1140 * discrete.mf modified to avoid discretizing a covariate more than once,
1141 and to halt if a model requires the same covariate to be discretized
1142 two different ways (e.g. s(x) + s(x,z)). This affects only
1143 bam(...,discrete=TRUE).
1144
1145 * Some changes to ziP and ziplss families to improve numerical robustness,
1146 and to ziP help file to suggest appropriate checking. Thanks to Keren
1147 Raiter, for reporting problems.
1148
1149 * numerical robustness of extended gam methods (gam.fit4) improved for cases
1150 with many zero or near zero iterative weights. Handling of zero weights
1151 modified to avoid divide by (near) zero problems. Also tests for poor
1152 scaling of sqrt(abs(w))*z and substitutes computations based on w*z if
1153 detected. Also 'newton' routine now step halves if REML score not finite!
1154
1155 * Sl.setup (used by bam) modification to allow more efficient handling of terms
1156 with multiple diagonal penalties with no non-zero elements in common, but
1157 possibly with non zero elements `interleaved' between penalties.
1158
1159 1.8-7
1160
1161 ** 'gam' default scale parameter changed to modified Pearson estimator
1162 developed by Fletcher 2012 Biometrika 99(1), 230-237. See ?gam.scale.
1163
1164 ** 'bam' now has a 'discrete' argument to allow discretization of covariates
1165 for more efficient computation, with substantially more parallelization
1166 (via 'nthreads'). Still somewhat experimental.
1167
1168 * Slightly more accurate smoothing parameter uncertainty correction. Changes
1169 edf2 used for AIC (under RE/ML), and hence may change AIC values.
1170
1171 * jagam prior variance on fixed effects is now set with reference to data and
1172 model during initialization step.
1173
1174 * bam could lose offset for small datasets in gaussian additive case. fixed.
1175
1176 * gam.side now setup to include penalties in computations if fewer data than
1177 coefs (an exceedingly specialist topic).
1178
1179 * p-value computation for smooth terms modified to avoid an ambiguity in the
1180 choice of test statistic that could lead to p-value changing somewhat between
1181 platforms.
1182
1183 * gamm now warns if attempt is made to use extended family.
1184
1185 * step fail logic improved for "fREML" optimization in 'bam'.
1186
1187 * fix of openMP error in mgcv_pbsi, which could cause a problem in
1188 multi-threaded bam computation (failure to declare a variable as private).
1189
1190 * Smoothing parameter uncertainty corrected AIC calculations had an
1191 indexing problem in Sl.postproc, which could result in failure of bam with
1192 linked smooths.
1193
1194 * mroot patched for fact that chol(...,pivot=TRUE) does not operate as
1195 documented on rank deficient matrices: trailing block of triangular factor
1196 has to be zeroed for pivoted crossprod of factor to equal original matrix.
1197
1198 * bam(...,sparse=TRUE) deprecated as no examples found where it is really
1199 worthwhile (but let me know if this is a problem).
1200
1201 * marginal model matrices in tensor product smooths now stored in
1202 re-parameterized form, if re-parameterization happened (shouldn't change
1203 anything!).
1204
1205 * initial.spg could fail if response vector had dim attributes and extended
1206 family used. fixed.
1207
1208 1.8-6
1209
1210 * Generalization of list formula handling to allow linear predictors to
1211 share terms. e.g. gam(list(y1~s(x),y2~s(z),1+2~s(v)+w-1),family=mvn(d=2))
1212
1213 * New German translation thanks to Detlef Steuer.
1214
1215 * plot.gam now silently returns a list of plotting data, to help advanced
1216 users (Fabian Scheipl) to produce customized plot.
1217
1218 * bam can now set up an object suitable for fitting, but not actually do
1219 the fit, following a suggestion by Fabian Scheipl. See arguments 'fit'
1220 and 'G'.
1221
1222 1.8-5
1223
1224 * Korean translation added thanks to Chel Hee Lee.
1225
1226 * scale parameter handling in edf in logLik.gam made consistent with glm
1227 (affects AIC).
1228
1229 * 'bam', 'gam' and 'gamm' modified to often produce smaller files when models
1230 saved (and never to produce absurdly large files). Achieved by setting
1231 environment of formula, terms etc to .GlobalEnv. Previously 'save' could
1232 save entire contents of environment of formula/terms with fitted model
1233 object. Note that change can cause failure in user written functions calling
1234 gam/bam and then 'predict' without supplying all prediction variables
1235 (fix obvious).
1236
1237 * A help file 'single.index' supplied illustrating how single index models
1238 can be estimated in mgcv.
1239
1240 * predict.gam now only creates a "constant" attribute if the model has one.
1241
1242 * gam.fit4 convergence testing of coefs modified to more robust test of
1243 gradients of penalized dev w.r.t. params, rather than change in params,
1244 which can fail under rank deficiency.
1245
1246 * mgcv_qrqy was not thread safe. Not noticeable on many platforms as all
1247 threads did exactly the same thing to the same matrix, but very noticeable
1248 on Windows. Thread safe mgcv_qrqy0 added and used in any parallel sections.
1249
1250 * Allow openMP support if compiler supports it and provides pre-defined macro
1251 _OPENMP, even if SUPPORT_OPENMP undefined. (Allows multi-threading on
1252 Windows, for example.)
1253
1254 * 'eps' is now an argument to 'betar' allowing some control on how to
1255 handle response values too close to 0 or 1. Help file expanded to
1256 emphasise the problems with using beta regression with 0s and 1s in
1257 the data.
1258
1259 * fix of bug in multi-formula contrast handling, causing failure of prediction
1260 in some cases.
1261
1262 * ziP and ziplss now check for non-integer (or binary) responses and produce
1263 an error message if these are found. Previously this was not trapped and
1264 could lead to a segfault.
1265
1266 1.8-4
1267
1268 ** JAGS/BUGS support added, enabling auto-generation of code and data
1269 required to used mgcv type GAMs with JAGS. Useful for complex random
1270 effects structures, for example.
1271
1272 * smoothCon failed if selection penalties requested, but term was unpenalized.
1273 Now fixed (no selection penalties on unpenalized terms.)
1274
1275 * gam.check would fail for tensor product smooths with by variables - fixed.
1276
1277 * predict.gam would fail when predicting for more data than the blocksize
1278 but selecting only some terms. Fixed thanks to Scott Kostyshak.
1279
1280 * smoothCon now has an argument `diagonal.penalty' allowing single penalty
1281 smooths to be re-parameterized in order to diagonalize the penalty matrix.
1282 PredictMat is modified to apply the same reparameterization, making it
1283 user transparent. Facilitates the setup of smooths for export to other
1284 packages.
1285
1286 * predict.bam now exported in response to a request from another
1287 package maintainer.
1288
1289 * 1.8 allows some prediction tasks for some families (e.g. cox.ph) to
1290 require response variables to be supplied. NAs in these then messed up
1291 prediction when they were not needed (e.g. if response variables with
1292 NAs were provided to predict.gam for a simple exponential family GAM).
1293 Response NAs now passed to the family specific prediction code, restoring
1294 the previous behaviour for most models. Thanks Casper Wilestofte Berg.
1295
1296 * backend parallel QR code used by gam modified to use a pivoted block
1297 algorithm.
1298
1299 * nthreads argument added to 'bam' to allow for parallel computation
1300 for computations in the main process (serial on any cluster nodes).
1301 e.g. QR based combination of results from cluster nodes is now
1302 parallel.
1303
1304 * fREML computation now partly in parallel (controlled by 'nthreads'
1305 argument to 'bam')
1306
1307 * slanczos now accepts an nt argument allowing parallel computation of
1308 main O(n^2) step.
1309
1310 * fix to newton logic problem, which could cause an attempt to use 'score2'
1311 before definition.
1312
1313 * fix to fREML code which could cause matrix square root to lose dimensions
1314 and cause an error.
1315
1316 * initial.sp could perform very poorly for very low basis dimensions - could
1317 set initial sp to effective infinity.
1318
1319 1.8-3
1320
1321 * Fix of two illegal read/write bugs with extended family models with no
1322 smooths. (Thanks to Julian Faraway for reporting beta regr problem).
1323
1324 * bam now checks that chunk.size > number of parameters and resets the
1325 chunk.size if not.
1326
1327 * Examples of use of smoothCon and PredictMat for setting up bases
1328 for use outside mgcv (and then predicting) added to ?smoothCon.
1329
1330 1.8-2
1331
1332 * For exponential family gams, fitted by outer iteration, a warning is now
1333 generated if the Pearson scale parameter estimate is more than 4 times
1334 a robust estimate. This may indicate an unstable Pearson estimate.
1335
1336 * 'gam.control' now has an option 'scale.est' to allow selection of the
1337 estimator to use for the scale parameter in exponential family GAMs.
1338 See ?gam.scale. Thanks to Trevor Davies for providing a clear unstable
1339 Pearson estimate example.
1340
1341 * drop.unused.levels argument added to gam, bam and gamm to allow
1342 "mrf" (and "re") terms to have unobserved factor levels.
1343
1344 * "mrf" constructor modified to deal properly with regions that contain no
1345 observations.
1346
1347 * "fs" smooths are no longer eligible to have side conditions set, since
1348 they are fully penalized terms and hence always identifiable (in theory).
1349
1350 * predict.bam was not declared as a method in NAMESPACE - fixed
1351
1352 * predict.bam modified to strip down object to save memory (especially in
1353 parallel).
1354
1355 * predict.gam now has block.size=NULL as default. This implies a block
1356 size of 1000 when newdata supplied, and use of a single block if no
1357 new data was supplied.
1358
1359 * some messages were not printing correctly after a change in
1360 message handling to facilitate easier translation. Now fixed.
1361
1362 1.8-1
1363
1364 * bam modified so that choleski based fitting works properly with rank
1365 deficient model matrix (without regularization).
1366
1367 * fix of 1.8-0 bug - gam prior weights mishandled in computation of cov matrix,
1368 resulting in incorrect variance estimates (even without prior weights
1369 specified). Thanks Fabian Scheipl.
1370
1371 1.8-0
1372
1373 *** Cox Proportional Hazard family 'cox.ph' added as example of general
1374 penalized likelihood families now useable with 'gam'.
1375
1376 *** 'ocat', 'tw', 'nb', 'betar', 'ziP' and 'scat' families added for
1377 ordered categorical data, Tweedie with estimation of 'p', negative binomial
1378 with (fast) estimation of 'theta', beta regression for proportions, simple
1379 zero inflated Poisson regression and heavy tailed regression with scaled t
1380 distribution. These are all examples of 'extended families' now useable
1381 with 'gam'.
1382
1383 *** 'gaulss' and 'ziplss' families, implementing models with multiple linear
1384 predictors. For gaulss there is a linear predictor for the Gaussian mean
1385 and another for the standard deviation. For ziplss there is a linear
1386 predictor controlling `presence' and another controlling
1387 the Poisson parameter, given presence.
1388
1389 *** 'mvn' family for multivariate normal additive models.
1390
1391 ** AIC computation changed for bam and gam models estimated by REML/ML
1392 to account for smoothing parameter uncertainty in degrees of freedom
1393 term.
1394
1395 * With REML/ML smoothness selection in gam/bam an extra covariance matrix 'Vc'
1396 is now computed which allows for smoothing parameter uncertainty. See
1397 the 'unconditional' arguments to 'predict.gam' and 'plot.gam' to use this.
1398
1399 * 'gam.vcomp' bug fix. Computed intervals for families with fixed scale
1400 parameter were too wide.
1401
1402 * gam now defaults to the Pearson estimator of the scale parameter to avoid
1403 poor scale estimates in the quasipoisson case with low counts (and possibly
1404 elsewhere). Gaussian, Poisson and binomial inference invariant to change.
1405 Thanks to Greg Dropkin, for reporting the issue.
1406
1407 * Polish translation added thanks to Lukasz Daniel.
1408
1409 * gam.fit3 now forces eta and mu to be consistent with coef and valid on
1410 return (previously could happen that if step halving was used in final
1411 iteration then eta or mu could be invalid, e.g. when using identity link
1412 with non-negative data)
1413
1414 * gam.fit3 now bases its convergence criteria on grad deviance w.r.t. model
1415 coefs, rather than changes in model coefs. This prevents problems when
1416 there is rank deficiency but different coefs get dropped at different
1417 iterations. Thanks to Kristynn Sullivan.
1418
1419 * If mgcv is not on the search path then interpret.gam now tries to
1420 evaluate in namespace of mgcv with environment of formula as enclosing
1421 environment, if evaluation in the environment of the formula fails.
1422
1423 * bug fix to sos plotting method so that it now works with 'by' variables.
1424
1425 * 'plot.gam' now weights partial residuals by *normalized* square root
1426 iterative weights so that the average weight is 1 and the residuals
1427 should have constant variance if all is ok.
1428
1429 * 'pcls' now reports if the initial point is not feasible.
1430
1431 * 'print.gam' and 'summary.gam' now report the rank of the model if it is
1432 rank deficient. 'gam.check' reports the model rank whenever it is
1433 available.
1434
1435 * fix of bug in 'k.check' called by 'gam.check' that gave an error for
1436 smooths with by variables.
1437
1438 * predict.gam now checks that factors in newdata do not contain more
1439 levels than those used in fitting.
1440
1441 * predict.gam could fail for type "terms" with no intercept - fixed.
1442
1443 * 'bfgs' now uses a finite difference approximation for the initial inverse
1444 Hessian.
1445
1446 1.7-29
1447
1448 * Single character change to Makevars file so that openMP multi-threading
1449 actually works.
1450
1451 1.7-28
1452
1453 * exclude.too.far updated to use kd-tree instead of inefficient search for
1454 neighbours. This can make plot.gam *much* faster for large datasets.
1455
1456 * Change in smoothCon, so that sweep and drop constraints (default for bam
1457 for efficiency reasons) are no longer allowed with by variables and matrix
1458 arguments (could lead to confusing results with factor by variables in bam).
1459
1460 * 'ti' terms now allow control of which marginals to constrain, via 'mc'.
1461 Allows e.g. y ~ ti(x) + ti(x,z,mc=c(0,1)) - for experts only!
1462
1463 * tensor.prod.model.matrix re-written to call C code. Around 5-10 times
1464 faster than old version for large data sets.
1465
1466 * re-write of mini.mf function used by bam to generate a reduced size
1467 model frame for model setup. New version ensures that all factor levels
1468 are present in reduced frame, and avoids production of unrealistic
1469 combinations of variables in multi-dimensional smooths which could occur
1470 with old version.
1471
1472 * bam models could fail if a penalty matrix was 1 by 1, or if multiple
1473 penalties on a smooth were in fact seperable into single penalties.
1474 Fixed. Thanks to Martijn weiling for reporting.
1475
1476 * Constant in tps basis computation was different to published version
1477 for odd dimensions - makes no difference to fit, but annoying if you
1478 are trying to test a re-implementation. Thanks to Weijie Cai at SAS.
1479
1480 * prediction for "cc" and "cp" classes is now cyclic - values outside the
1481 range of knots are wrapped back into the interval.
1482
1483 * ldTweedie now returns derivatives w.r.t. a transform of p as well as
1484 w.r.t log of scale parameter phi.
1485
1486 * gamm can now handle 'varComb' variance functions (thanks Sven Neulinger
1487 for reporting that it didn't).
1488
1489 * fix of a bug which could cause bam to seg fault for a model with no smooths
1490 (insufficient storage allocated in C in this case). Thanks Martijn Weiling.
1491
1492 1.7-27
1493
1494 * Further multi-threading in gam fits - final two leading order matrix
1495 operations parallelized using openMP.
1496
1497 * Export of smooth.construct.t2.smooth.spec and Predict.matrix.t2.smooth,
1498 and Rrank.
1499
1500 * Fix of of missing [,,drop=FALSE] in predict.gam that could cause problems
1501 with single row prediction when 'terms' supplied (thanks Yang Yang).
1502
1503 1.7-26
1504
1505 * Namespace fixes.
1506
1507 1.7-25
1508
1509 * code added to allow openMP based multi-threading in gam fits (see
1510 ?gam.control and ?"mgcv-parallel").
1511
1512 * bam now allows AR1 error model to be split blockwise. See argument
1513 'AR.start'.
1514
1515 * magic.post.proc made more efficient (one of two O(np^2) steps removed).
1516
1517 * var.summary now coerces character to factor.
1518
1519 * bugs fixed whereby etastart etc were not passed to initial.spg and
1520 get.null.coefs. Thanks to Gavin Simpson.
1521
1522 * reformulate removed from predict.gam to avoid (slow) repeated parser
1523 calls.
1524
1525 * gaussian(link="log") initialization fixed so that negative data
1526 does not make it fail, via fix.family patching function.
1527
1528 * bug fix in plot method for "fs" basis - ignored any side conditions.
1529 Thanks to Martijn Weiling and Jacolien van Rij.
1530
1531 * gamm now checks whether smooths nested in factors have illegal side
1532 conditions, and halts if so (re-ordering formula can help).
1533
1534 * anova.glmlist no longer called.
1535
1536 * Compiled code now uses R_chck_calloc and R_chk_free for memory management
1537 to avoid the possibility of unfriendly exit on running out of memory.
1538
1539 * fix in gam.side which would fail with unpenalized interactions in the
1540 presence of main effects.
1541
1542 1.7-24
1543
1544 * Examples pruned in negbin, smooth.construct.ad.smooth.spec and bam help
1545 files to reduce CRAN checking load.
1546
1547 * gam.side now warns if only repeated 1-D smooths of the same variable are
1548 encountered, but does not halt.
1549
1550 * Bug fix in C code for "cr" basis, that could cause a memory violation during
1551 prediction, when an extrapolation was immediately followed by a prediction
1552 that lay exactly on the upper boundary knot. Thanks to Keith Woolner for
1553 reporting this.
1554
1555 * Fix for bug in fast REML code that could cause bam to fail with ti/te only
1556 models. Thanks to Martijn Wieling.
1557
1558 * Fix of bug in extract.lme.cov2, which could cause gamm to fail when
1559 a correlation structure was nested inside a grouping factor finer than
1560 the finest random effect grouping factor.
1561
1562 * Fix for an interesting feature of lme that getGroups applied to the
1563 corStruct that is part of the fitted lme object returns groups in
1564 sorted order, not data frame order, and without an index from one order
1565 to the other. (Oddly, the same corStruct Initialized outside lme has its
1566 groups in data frame order.) This feature could cause gamm to fail,
1567 complaining that the grouping factors for the correlation did not appear
1568 to be nested inside the grouping structure of the random effects. A
1569 bunch of ordering sensitivity tests have been added to the mgcv test suite.
1570 Thanks to Dave Miller for reporting the bug.
1571
1572 1.7-23
1573
1574 *** Fix of severe bug introduced with R 2.15.2 LAPACK change. The shipped
1575 version of dsyevr can fail to produce orthogonal eigenvectors when
1576 uplo='U' (upper triangle of symmetric matrix used), as opposed to 'L'.
1577 This led to a substantial number of gam smoothing parameter estimation
1578 convergence failures, as the key stabilizing re-parameterization was
1579 substantially degraded. The issue did not affect gaussian additive models
1580 with GCV model selection. Other models could fail to converge any further
1581 as soon as any smoothing parameter became `large', as happens when a
1582 smooth is estimated as a straight line. check.gam reported the lack of full
1583 convergence, but the issue could also generate complete fit failures.
1584 Picked up late as full test suite had only been run on R > 2.15.1 with an
1585 external LAPACK.
1586
1587 ** 'ti' smooth specification introduced, which provides a much better (and
1588 very simple) way of allowing nested models based on 'te' type tensor
1589 product smooths. 'ti' terms are used to set up smooth interactions
1590 excluding main effects (so ti(x,z) is like x:z while te(x,z) is more
1591 like x*z, although the analogy is not exact).
1592
1593 * summary.gam now uses a more efficient approach to p-value computation
1594 for smooths, using the factor R from the QR factorization of the weighted
1595 model matrix produced during fitting. This is a weighted version of the
1596 Wood (2013) statistic used previously - simulations in that paper
1597 essentially unchanged by the change.
1598
1599 * summary.gam now deals gracefully with terms such as "fs" smooths
1600 estimated using gamm, for which p-values can not be computed. (thanks to
1601 Gavin Simpson).
1602
1603 * gam.check/qq.gam now uses a normal QQ-plot when the model has been fitted
1604 using gamm or gamm4, since qq.gam cannot compute corrext quantiles in
1605 the presence of random effects in these cases.
1606
1607 * gamm could fail with fixed smooths while assembling total
1608 penalty matrix, by attempting to access non-existent penalty
1609 matrix. (Thanks Ainars Aunins for reporting this.)
1610
1611 * stripped rownames from model matrix, eta, linear predictor etc. Saves
1612 memory and time.
1613
1614 * plot.soap.film could switch axis ranges. Fixed.
1615
1616 * plot.mgcv.smooth now sets smooth plot range on basis of xlim and
1617 ylim if present.
1618
1619 * formXtViX documentation fixed + return matrix labels.
1620
1621 * fixDependence related negative index failures for completely confounded
1622 terms - now fixed.
1623
1624 * sos smooth model matrix re-scaled for better conditioning.
1625
1626 * sos plot method could produce NaNs by a rounding error in argument to
1627 acos - fixed.
1628
1629 1.7-22
1630
1631 * Predict.matrix.pspline.smooth now allows prediction outside range of knots,
1632 and uses linear extrapolation in this case.
1633
1634 * missing drop=FALSE in reTest called by summary.gam caused 1-D random effect
1635 p-value computation to fail. Fixed (thanks Silje Skår).
1636
1637 1.7-21
1638
1639 ** soap film smoother class added. See ?soap
1640
1641 * Polish translation added thanks to Lukasz Daniel.
1642
1643 * mgcv/po/R-mgcv.pot up-dated.
1644
1645 * plot methods for smooths modified slightly to allow methods to return
1646 plot data directly, without a prediction matrix.
1647
1648 1.7-20
1649
1650 * '...' now passed to termplot by plot.gam (thanks Andreas Eckner).
1651
1652 * fix to null deviance computation for binomial when n>1, matrix response
1653 used and an offset is present. (Thanks to Tim Miller)
1654
1655 * Some pruning of unused code from recov and reTest.
1656
1657 * recov modified to stop it returning a numerically non-symmetric Ve, and
1658 causing occasional failures of summary.gam with "re" terms.
1659
1660 * MRF smooth bug. Region ordering could become confused under some
1661 circumstances due to incorrect setting of factor levels. Corrected
1662 thanks to detailed bug report from Andreas Bender.
1663
1664 * polys.plot colour/grey scale bug. Could ask for colour 0 from colour
1665 scheme, and therefore fail. Fixed.
1666
1667 1.7-19
1668
1669 ** summary.gam and anova.gam now use an improved p-value computation for
1670 smooth terms with a zero dimensional penalty null space (including
1671 random effects). The new scheme has been tested by full replication
1672 of the simulation study in Scheipl (2008,CSDA) to compare it to the best
1673 method therein. In these tests it is at least as powerful as the best
1674 method given there, and usually indistinguishable, but it gives slightly
1675 too low null p-values when smoothing parameters are very poorly identified.
1676 Note that the new p-values can not be computed from old fitted gam objects.
1677 Thanks to Martijn Wieling for pointing out how bad the p-values for regular
1678 smooths could be with random effects.
1679
1680 * t2 terms now take an argument `ord' that allows orders of interaction to
1681 be selected.
1682
1683 * "tp" smooths can now drop the null space from their construction via
1684 a vector m argument, to allow testing against polynomials in the null space.
1685
1686 * Fix of vicious little bug in gamm tensor product handling that could have
1687 a te term pick up the wrong model matrix and fail.
1688
1689 * bam now resets method="fREML" to "REML" if there are no free smoothing
1690 parameters, since there is no advantage to the "fREML" optimizer in this
1691 case, and it assumes there is at least one free smoothing parameter.
1692
1693 * print.gam modified to print effective degrees of freedom more prettily,
1694
1695 * testStat bug fix. qr was called with default arguments, which includes
1696 tol=1e-7...
1697
1698 * bam now correctly returns fitting weights (rather than prior) in weights
1699 field.
1700
1701 1.7-18
1702
1703 * Embarrassingly, the adjusted r^2 computation in summary.gam was wrong
1704 for models with prior weights. Now fixed, thanks to Antony Unwin.
1705
1706 * bam(...,method="fREML") could give incorrect edfs for "re" terms as a
1707 result of a matrix indexing error in Sl.initial.repara. Now fixed.
1708 Thanks to Martijn Wieling for reporting this.
1709
1710 * summary.gam had freq=TRUE set as default in 1.7-17. This gave better
1711 p-values for paraPen terms, but spoiled p-values for fixed effects in the
1712 presence of "re" terms (a rather more common setup). Default now reset to
1713 freq=FALSE.
1714
1715 * bam(...,method="fREML") made fully compatible with gam.vcomp.
1716
1717 * bam and negbin examples speeded up
1718
1719 * predict.gam could fail for models of the form y~1 when newdata are supplied.
1720 (Could make some model averaging methods fail). Fixed.
1721
1722 * plot.gam had an overzealous check for availibility of variance estimates,
1723 which could make rank deficient models fail to plot CIs. fixed.
1724
1725 1.7-17
1726
1727 ** p-values for terms with no un-penalized components were poor. The theory on
1728 which the p-value computation for other terms is based shows why this is,
1729 and allows fixes to be made. These are now implemented.
1730
1731 * summary p value bug fix --- smooths with no null space had a bug in
1732 lower tail of p-value computation, yielding far too low values. Fixed.
1733
1734 * bam now outputs frequentist cov matrix Ve and alternative effective degrees
1735 of freedom edf1, in all cases.
1736
1737 * smoothCon now adjusts null.space.dim on constraint absorption.
1738
1739 * Prediction with matrix arguments (i.e. for models using summation
1740 convention) could be very memory hungry. This in turn meant that
1741 bam could run out of memory when fitting models with such terms.
1742 The problem was memory inefficient handling of duplicate evaluations.
1743 Now fixed by modification of PredictMat
1744
1745 * bam could fail if the response vector was of class matrix. fixed.
1746
1747 * reduced rank mrf smooths with supplied penalty could use the incorrect
1748 penalty rank when computing the reduced rank basis and fail. fixed
1749 thanks to Fabian Scheipl.
1750
1751 * a cr basis efficiency change could lead to old fitted model objects causing
1752 segfaults when used with current mgcv version. This is now caught.
1753
1754 1.7-16
1755
1756 * There was an unitialized variable bug in the 1.7-14 re-written "cr" basis
1757 code for the case k=3. Fixed.
1758
1759 * gam.check modified slightly so that k test only applied to smooths of
1760 numeric variables, not factors.
1761
1762 1.7-15
1763
1764 * Several packages had documentation linking to the 'mgcv' function
1765 help page (now removed), when a link to the package was meant. An alias
1766 has been added to mgcv-package.Rd to fix/correct these links.
1767
1768 1.7-14
1769
1770 ** predict.bam now added as a wrapper for predict.gam, allowing parallel
1771 computation
1772
1773 ** bam now has method="fREML" option which uses faster REML optimizer:
1774 can make a big difference on parameter rich models.
1775
1776 * bam can now use a cross product and Choleski based method to accumulate
1777 the required model matrix factorization. Faster, but less stable than
1778 the QR based default.
1779
1780 * bam can now obtain starting values using a random sub sample of the data.
1781 Useful for seriously large datasets.
1782
1783 * check of adequacy of basis dimensions added to gam.check
1784
1785 * magic can now deal with model matrices with more columns than rows.
1786
1787 * p-value reference distribution approximations improved.
1788
1789 * bam returns objects of class "bam" inheriting from "gam"
1790
1791 * bam now uses newdata.guaranteed=TRUE option when predicting as part
1792 of model matrix decomposition accumulation. Speeds things up.
1793
1794 * More efficient `sweep and drop' centering constraints added as default for
1795 bam. Constaint null space unchanged, but computation is faster.
1796
1797 * Underlying "cr" basis code re-written for greater efficiency.
1798
1799 * routine mgcv removed, it now being many years since there has been any
1800 reason to use it. C source code heavily pruned as a result.
1801
1802 * coefficient name generation moved from estimate.gam to gam.setup.
1803
1804 * smooth2random.tensor.smooth had a bug that could produce a nonsensical
1805 penalty null space rank and an error, in some cases (e.g. "cc" basis)
1806 causing te terms to fail in gamm. Fixed.
1807
1808 * minor change to te constructor. Any unpenalized margin now has
1809 corresponding penalty rank dropped along with penalty.
1810
1811 * Code for handling sp's fixed at exactly zero was badly thought out, and
1812 could easily fail. fixed.
1813
1814 * TPRS prediction code made more efficient, partly by use of BLAS. Large
1815 dataset setup also made more efficient using BLAS.
1816
1817 * smooth.construct.tensor.smooth.spec now handles marginals with factor
1818 arguments properly (there was a knot generation bug in this case)
1819
1820 * bam now uses LAPACK version of qr, for model matrix QR, since it's
1821 faster and uses BLAS.
1822
1823 1.7-13
1824
1825 ** The Lanczos routine in mat.c was using a stupidly inefficient check for
1826 convergence of the largest magnitude eigenvectors. This resulted in
1827 far too many Lanczos steps being used in setting up thin plate regression
1828 splines, and a noticeable speed penalty. This is now fixed, with many thanks
1829 David Shavlik for reporting the slow down.
1830
1831 * Namespace modified to import from methods. Dependency on stats and graphics
1832 made explicit.
1833
1834 * "re" smooths are no longer subject to side constraint under nesting (since
1835 this is almost always un-necessary and undesirable, and often unexpected).
1836
1837 * side.con modified to allow smooths to be excluded and to allow side
1838 constraint computation to take account of penalties (unused at present).
1839
1840 1.7-12
1841
1842 * bam can now compute the leading order QR decomposition on a cluster
1843 set up using the parallel package.
1844
1845 * Default k for "tp" and "ds" modified so that it doesn't exceed 100 +
1846 the null space dimension (to avoid complaints from users smoothing in
1847 quite alot of dimensions). Also default sub-sample size reduced to 2000.
1848
1849 * Greater use of BLAS routines in the underlying method code. In particular
1850 all leading order operations count steps for gam fitting now use BLAS.
1851 You'll need R to be using a rather fancy BLAS to see much difference,
1852 however.
1853
1854 * Amusingly, some highly tuned blas libraries can result in lapack not always
1855 giving identical eigenvalues when called twice with the same matrix. The
1856 `newton' optimizer had assumed this wouldn't happen: not any more.
1857
1858 * Now byte compiled by default. Turn this off in DESCRIPTION if it interferes
1859 with debugging.
1860
1861 * summary.gam p-value computation options modified (default remains the
1862 same).
1863
1864 * summary.gam default p-value computation made more computationally
1865 efficient.
1866
1867 * gamm and bam could fail under some options for specifying binomial models.
1868 Now fixed.
1869
1870 1.7-11
1871
1872 * smoothCon bug fix to avoid NA labels for matrix arguments when
1873 no by variable provided.
1874
1875 * modification to p-value computation in summary.gam: `alpha' argument
1876 removed (was set to zero anyway); computation now deals with possibility
1877 of rank deficiency computing psuedo-inverse of cov matrix for statistic.
1878 Previously p-value computation could fail for random effect smooths with
1879 large datasets, when a random effect has many levels. Also for large data
1880 sets test statistic is now based on randomly sampling max(1000,np*2) model
1881 matrix rows, where np is number of model coefficients (random number
1882 generator state unchanged by this), previous sample size was 3000.
1883
1884 * plot.mrf.smooth modified to allow passing '...' argument.
1885
1886 * 'negbin' modified to avoid spurious warnings on initialization call.
1887
1888 1.7-10
1889
1890 * fix stupid bug in 1.7-9 that lost term labels in plot.gam.
1891
1892 1.7-9
1893
1894 * rather lovely plot method added for splines on the sphere.
1895
1896 * plot.gam modified to allow 'scheme' to be specified for plots, to easily
1897 select different plot looks.
1898
1899 * schemes added for default smooth plotting method, modified for mrfs and
1900 factor-smooth interactions.
1901
1902 * mgcv function deprected, since magic and gam are much better (let me know
1903 if this is really a problem).
1904
1905 1.7-8
1906
1907 * gamm.setup fix. Bug introduced in 1.7-7 whereby gamm with no smooths would
1908 fail.
1909
1910 * gamm gives returned object a class "gamm"
1911
1912
1913 1.7-7
1914
1915 * "fs" smooth factor interaction class introduced, for smooth factor
1916 interactions where smoothing parameters are same at each factor level.
1917 Very efficient with gamm, so good for e.g. individual subject smooths.
1918
1919 * qq.gam default method modified for increased power.
1920
1921 * "re" terms now allowed as tensor product marginals.
1922
1923 * log saturated likelihoods modified w.r.t. weight handling, so that weights
1924 are treated as modifying the scale parameter, when scale parameter is free.
1925 i.e. obs specific scale parameter is overall scale parameter divided by
1926 obs weight. This ensures that when the scale parameter is free, RE/ML based
1927 inference is invariant to multiplicative rescaling of weights.
1928
1929 * te and t2 now accept lists for 'm'. This allows more flexibility with
1930 marginals that can have vector 'm' arguments (Duchon splines, P splines).
1931
1932 * minor mroot fix/gam.reparam fix. Could declare symmetric matrix
1933 not symmetric and halt gam fit.
1934
1935 * argument sparse added to bam to allow exploitation of sparsity in fitting,
1936 but results disappointing.
1937
1938 * "mrf" now evaluates rank of penalty null space numerically (previously
1939 assumed it was always one, which it need not be with e.g. a supplied
1940 penalty).
1941
1942 * gam.side now corrects the penalty rank in smooth objects that have
1943 been constrained, to account for the constraint. Avoids some nested
1944 model failures.
1945
1946 * gamm and gamm.setup code restructured to allow smooths nested in factors
1947 and for cleaner object oriented converion of smooths to random effects.
1948
1949 * gam.fit3 bug. Could fail on immediate divergence as null.eta was matrix.
1950
1951 * slanczos bug fixes --- could segfault if k negative. Could also fail to
1952 return correct values when k small and kl < 0 (due to a convergence
1953 testing bug, now fixed)
1954
1955 * gamm bug --- could fail if only smooth was a fixed one, by looking for
1956 non-existent sp vector. fixed.
1957
1958 * 'cc' Predict.matrix bug fix - prediction failed for single points.
1959
1960 * summary.gam failed for single coefficient random effects. fixed.
1961
1962 * gam returns rV, where t(rV)%*%rV*scale is Bayesian cov matrix.
1963
1964 1.7-6
1965
1966 ** factor `by' variable handling extended: if a by variable is an
1967 ordered factor then the first level is treated as a reference level
1968 and smooths are only generated for the other levels. This is useful
1969 for avoiding identifiability issues in complex models with factor by
1970 variables.
1971
1972 * bam bug fix. aic was reported incorrectly (too low).
1973
1974 1.7-5
1975
1976 * gam.fit3 modified to converge more reliably with links that don't guarantee
1977 feasible mu (e.g poisson(link="identity")). One vulnerability removed + a
1978 new approach taken, which restarts the iteration from null model
1979 coefficients if the original start values lead to an infinite deviance.
1980
1981 * Duchon spline bug fix (could fail to create model matrix if
1982 number of data was one greater than number of unique data).
1983
1984 * fix so that 'main' is not ignored by plot.gam (got broken in 1.7-0
1985 object orientation of smooth plotting)
1986
1987 * Duchon spline constructor now catches k > number of data errors.
1988
1989 * fix of a gamm bug whereby a model with no smooths would fail after
1990 fitting because of a missing smoothing parameter vector.
1991
1992 * fix to bug introduced to gam/bam in 1.7-3, whereby '...' were passed to
1993 gam.control, instead of passing on to fitting routines.
1994
1995 * fix of some compiler warnings in matrix.c
1996
1997 * fix to indexing bug in monotonic additive model example in ?pcls.
1998
1999 1.7-4
2000
2001 * Fix for single letter typo bug in C code called by slanczos, could
2002 actually segfault on matrices of less than 10 by 10.
2003
2004 * matrix.c:Rlanczos memory error fix in convergence testing of -ve
2005 eigenvalues.
2006
2007 * Catch for min.sp vector all zeroes, which could cause an ungraceful
2008 failure.
2009
2010 1.7-3
2011
2012 ** "ds" (Duchon splines) smooth class added. See ?Duchon.spline
2013
2014 ** "sos" (spline on the sphere) smooth class added. See ?Spherical.Spline.
2015
2016 * Extended quasi-likelihood used with RE/ML smoothness selection and
2017 quasi families.
2018
2019 * random subsampling code in bam, sos and tp smooths modified a little, so
2020 that .Random.seed is set if it doesn't exist.
2021
2022 * `control' argument changed for gam/bam/gamm to a simple list, which is
2023 then passed to gam.control (or lmeControl), to match `glm'.
2024
2025 * Efficiency of Lanczos iteration code improved, by restructuring, and
2026 calling LAPACK for the eigen decompostion of the working tri-diagonal
2027 matrix.
2028
2029 * Slight modification to `t2' marginal reparameterization, so that `main
2030 effects' can be extracted more easily, if required.
2031
2032
2033 1.7-2
2034
2035 * `polys.plot' now exported, to facilitate plotting of results for
2036 models involving mrf terms.
2037
2038 * bug fix in plot.gam --- too.far had stopped working in 1.7-0.
2039
2040 1.7-1
2041
2042 * post fitting constraint modification would fail if model matrix was
2043 rank deficient until penalized. This was an issue when mixing new t2
2044 terms with "re" type random effects. Fixed.
2045
2046 * plot.mrf.smooth bug fix. There was an implicit assumption that the
2047 `polys' list was ordered in the same way as the levels of the covariate
2048 of the smooth. fixed.
2049
2050 * gam.side intercept detection could occasionally fail. Improved.
2051
2052 * concurvity would fail if model matrix contained NA's. Fixed.
2053
2054 1.7-0
2055
2056 ** `t2' alternative tensor product smooths added. These can be used with
2057 gamm4.
2058
2059 ** "mrf" smooth class added (at the suggestion of Thomas Kneib).
2060 Implements smoothing over discrete geographic districts using a
2061 Markov random field penalty. See ?mrf
2062
2063 * qq.gam added to allow better checking of distribution of residuals.
2064
2065 * gam.check modified to use qq.gam for QQ plots of deviance residuals.
2066 Also, it now works with gam(*, na.action = "na.replace") and NAs.
2067
2068 * `concurvity' function added to provide simple concurvity measures.
2069
2070 * plot.gam automatic layout modified to be a bit more sensible (i.e.
2071 to recognise that most screens are landscape, and that usually
2072 squarish plots are wanted).
2073
2074 * Plot method added for mrf smooths.
2075
2076 * in.out function added to test whether points are interior to
2077 a region defined by a set of polygons. Useful when working with
2078 MRFs.
2079
2080 * `plot.gam' restructured so that smooths are plotted by smooth specific
2081 plot methods.
2082
2083 * Plot method added for "random.effect" smooth class.
2084
2085 * `pen.edf' function added to extract EDF associated with each penalty.
2086 Useful with t2 smooths.
2087
2088 * Facilty provided to allow different identifiability constraints to be
2089 used for fitting and prediction. This allows t2 smooths to be fitted
2090 with a constraint that allows fitting by gamm4, but still perform
2091 inference with the componentwise optimal sum to zero constraints.
2092
2093 * mgcv-FAQ.Rd added.
2094
2095 * paraPen works properly with `gam.vcomp' and full.sp names returned
2096 correctly.
2097
2098 * bam (and bam.update) can now employ an AR1 error model in the
2099 guassian-identity case.
2100
2101 * bam.update modified for faster updates (initial scale parameter
2102 estimate now supplied in RE/ML case)
2103
2104 * Absorption of identifiability constraints modified to allow
2105 constraints that only affect some parameters to leave rest of
2106 parameters completely unchanged.
2107
2108 * rTweedie added for quick simulation of Tweedie random deviates
2109 when 1<p<2.
2110
2111 * smooth.terms help file fixed so cyclic p spline identifies as "cp"
2112 and not "cs"!
2113
2114 * bug fix in `gamm' so that binomial response can be provided as 2 column
2115 matrix, in standard `glm' way.
2116
2117 1.6-2
2118
2119 ** Random effect support for `gam' improved with the addition of
2120 a random effect "smooth" class, and a function for extracting
2121 variance components. See ?random.effects and links for details.
2122
2123 * smooths now contain extra elements: S.scale records the scale factor
2124 used in any linear rescaling of a penalty matrix; plot.me indicates
2125 whether `plot.gam' should attempt to plot the term; te.ok indicates
2126 whether the smooth is a suitable marginal for a tensor product.
2127
2128 * Fix in `gamm.setup' -- models with no fixed effects (except smooths)
2129 could fail to fit properly, because of an indexing error (caused odd
2130 results with indicator by variables)
2131
2132 * help files have had various misuses of `itemize' fixed.
2133
2134 * initialization could fail if response was actually a 1D array. fixed.
2135
2136 * New function `bam.update' allows efficient updating of very large
2137 strictly additive models fitted by `bam' when additional data become
2138 available.
2139
2140 * gam now warns if RE/ML used with quasi families.
2141
2142 * gam.check now accepts graphics parameters.
2143
2144 * fixed problem in welcome message that messed up ESS.
2145
2146 1.6-1
2147
2148 * Bug in cSplineDes caused bases to be non-cyclic unless first
2149 knot was at 0. This also affected the "cp" smoother class. Fixed.
2150
2151 * null.deviance calculation was wrong for case with offset and
2152 weights. Fixed.
2153
2154 * Built in strictly 1D smoothers now give an informative error message if
2155 an attempt is made to use them for multidimensional smoothing.
2156
2157 * gam.check generated a spurious error when applied to a model with no
2158 estimated smoothing parameters. Fixed.
2159
2160 1.6-0
2161
2162 *** Routine `bam' added for fitting GAMs to very large datasets.
2163 See ?bam
2164
2165 ** p-value method tweaked again. Reference DoF for testing now
2166 defaults to the alternative EDF estimate (based on 2F - FF where
2167 F = (X'WX+S)^{-1}X'WX). `magic.post.proc' and `gam.fit3.post.proc'
2168 changed to provide this. p-values still a bit too small, but only
2169 slightly so, if `method="ML"' is the smoothness selector.
2170
2171 * bad bug in `get.null.coef' could cause fit failure as a result of
2172 initial null coefs predicting infinite deviance.
2173
2174 * REML/ML convergence could be response scale sensitive, because of
2175 innapropriate convergence testing scaling in newton and bfgs -
2176 fixed.
2177
2178 * Slight fix to REML (not ML) score calculation in gam.fit3 -
2179 Mp/2*log(2*pi*scale) was missing from REML score, where Mp is
2180 total null space dimension for model.
2181
2182 * `summary.gam' bug fix: REML/ML models were always treated as if
2183 scale parameter had been estimated. gamObject should now contain
2184 `scale.estimated' indicating whether or not scale estimated
2185
2186 * some modifications to smoothCon and gam.setup to allow smooth
2187 constructors to return Matrix style sparse model matrices and
2188 penalty matrices.
2189
2190 * fixed misplaced bracket in print.mgcv.version, called on attachment.
2191
2192 * added utility function `ls.list' to give memory usage for elements
2193 of a list.
2194
2195 * added function `rig' to generate inverse Gaussian random deviates.
2196
2197 1.5-6
2198
2199 * "ts" and "cs" modified so that zero eigen values of penalty
2200 matrix are reset to 10% of smallest strictly positive eigen
2201 value, rather than 1%. This seems to lead to more reliable
2202 performance.
2203
2204 * `bfgs' simplified and improved so that it now checks the Wolfe
2205 conditions are met at each step. No longer uses any Newton steps,
2206 so if it's used with gam.control(outerPIsteps=0) then it's
2207 first derivative only for smoothing parameter optimization.
2208
2209 * `outerPIsteps' now defaults to zero in `gam.control'.
2210
2211 * New routine `initial.spg' gets jth initial sp to equalize
2212 Frobenious norm of S_j and cols of sqrt(W)X which it penalizes,
2213 where W are initial fisher weights. This removes the need for a
2214 performance iteration step to get starting values (so
2215 outerPIsteps=0 in gam.control can now bypass PI completely).
2216
2217 * fscale set from get.null.coef (facilitates cleaner initialization).
2218
2219 * large data set rare event logistic regression example added to
2220 ?gam.
2221
2222 * For p-value calculation for smooths, summary.gam subsamples rows of
2223 the model matrix if it has more than 3000 rows. This speeds things
2224 up for large datasets.
2225
2226 * minor bug fix in `gamm' so that intercept gets correct name, if
2227 it's the only non-smooth fixed effect.
2228
2229 * .pot files updated, German translation added, thanks to Detlef Steuer.
2230
2231 * `in.out' was not working from 1.5 --- fixed.
2232
2233 * loglik.gam now ups parameter count for Tweedie by one to account for
2234 scale estimation.
2235
2236 * There was a bug in the calculation of the Bayesian cov matrix, when the
2237 scale parameter was known: it was always using an estimated scale
2238 parameter. Makes no statistically meaningful difference for a model
2239 that fits properly, of course.
2240
2241 * Some junk removed from gam object.
2242
2243 * summary.gam pseudoinversion made slightly more efficient.
2244
2245 * adaptive smooth constructor is a bit more careful about the ranks
2246 of the penalties.
2247
2248 * 2d adaptive smoother bug fix --- part of penalty was missing due
2249 to complete line error.
2250
2251 * `smoothCon' and `PredictMat' modified so that sparse smooths can
2252 optionally have sparse centering constraints applied.
2253
2254 * `gamm' fix: prediction and visualization from `x$gam' where x is a
2255 fitted `gamm' object should not require the random effects to be
2256 provided. Now it doesn't.
2257
2258 * minor bug fix: a model with no penalties except a fixed one would fail
2259 with an index error.
2260
2261 * `te' terms are now only subject to centering constraints if all
2262 marginals would usually have a centering constraint.
2263
2264 * `te' no longer resets multi-dimensional marginals to "tp", unless
2265 they have been set to "cr", "cs", "ps" or "cp". This allows tensor
2266 products with user supplied smooths.
2267
2268 * Example of obtaining derivatives of a smooth (with CIs) added to
2269 `predict.gam' help file.
2270
2271 * `newdata.guaranteed' argument to predict.gam didn't work. fixed.
2272
2273
2274 1.5-5
2275
2276 * `gamm.setup' made an assumption about basis dimensions which was not
2277 true for tensor products involving the "cc" basis. This is now fixed.
2278
2279 1.5-4
2280
2281 * smooth.construct.tensor.smooth.spec modified, so that
2282 re-parameterization in terms of function values is only if it's
2283 stable, and by default the parameters are function values with
2284 even spacing. Otherwise it was possible for tensor products of
2285 p-splines to fail.
2286
2287 1.5-3
2288
2289 * `gam' now attempts to coerce `data' to a data frame, if it is not
2290 already a list or a data frame, provided that it is already an object
2291 that model.frame can deal with. This is to support an undocumented
2292 feature of versions prior to 1.5-2 that `data' could actually be
2293 something other than a list or data frame.
2294
2295 * An offset of type "array" could cause gam.fit3 to fail. fixed.
2296
2297 * `variable.summary' bug fixed, (it caused gam(y~1) to fail).
2298
2299 1.5-2
2300
2301 * Several exported functions had no usage entries in the help files.
2302 Everything exported does now.
2303
2304 * `vis.gam' had a bunch of bugs (which could make it fail altogether)
2305 as a result of trying to set default conditioning values from the gam
2306 object model frame. `gam' and `gamm' now obtain summary statistics of
2307 the predictor variables, stored in `var.summary' in the gam object,
2308 which `vis.gam' now uses. As a result `vis.gam' `view' and `cond'
2309 arguments should now contain original variable names, not model frame
2310 term names.
2311
2312 * `data' argument of `gam' no longer stored in the `gam' object, by
2313 default to save memory (can restore this --- see `gam.control').
2314
2315 * `summary.gam' failed under na.exclude. Fixed.
2316
2317 * `mroot' failed on 1 by 1 matrices, Fixed.
2318
2319 1.5-1
2320
2321 * The stability of the fitting methods had become substantially greater
2322 than the stability of the edf calculations after fitting. So it was
2323 possible to fit very poor models, and then have non-sensical effective
2324 degrees of freedom reported. This has been fixed by using much more stable
2325 expressions for edf calculation, when outer iteration smoothness
2326 selection is employed. (Changes are to gam.fit3, gam.outer and a new
2327 routine gam.fit3.post.proc).
2328
2329 * edfs in version 1.5-0 were calculated using newton, rather than fisher
2330 weights, in the matrix F=(X'WX+S)^{-1}X'WX, the diagonal of which gives
2331 the edf array. The problem with this is that it is possible for X'WX
2332 not to be +ve definite, and then degrees of freedom can be non-sensical.
2333 Fisher weights are always used now (although the original problem is
2334 exceedingly hard to generate an example of).
2335
2336 * The summation convention code could be *very* memory intensive for cases
2337 in which the matrix arguments of a smooth feature many repeated values.
2338 Code now fixed to make much more efficient use of any repeated rows in
2339 matrix arguments. This enables much larger signal regression problems to
2340 be tackled.
2341
2342 * Some help file fixes.
2343
2344 1.5-0
2345
2346 *** Efficient and general REML/ML smoothing selection implemented.
2347 Smoothness selection criterion and numerical optimizer are now
2348 selected using arguments `method' and `optimizer' of `gam', and
2349 `gam.method' has been removed.
2350
2351 *** To further enhance stability and efficiency, Fisher scoring is now
2352 only used for canonical links, when it corresponds to full Newton.
2353 With non-canonical links PIRLS is based on full Newton.
2354
2355 ** Derivative iteration as in Wood (2008) has been replaced by a direct
2356 implicit function method (which costs no more given Newton based
2357 PIRLS).
2358
2359 ** An option `select' has been added to `gam' to allow terms to be
2360 completely removed from a model by smoothness selection.
2361
2362 ** The shrinkage smoothers "cs" and "ts" have been modified
2363 substantially. The Wood (2006, 4.1.6) proposal of adding a
2364 small multiple of the identity matrix to the penalty matrix
2365 is flawed in that it tends to corrupt small eigen values
2366 of the penalty matrix for large (dimension) penalty matrices.
2367 It is much better to set the zero eigenvalues of the penalty matrix
2368 to a small proportion of the smallest +ve eigenvalue, and to
2369 use the matrix with the resulting eigen-decomposition as the
2370 penalty. This is now done. Thanks to Roman Torgovitsky for
2371 reporting the original problem.
2372
2373 ** Tweedie family added (including `ldTweedie' function to evaluate
2374 log Tweedie densities for powers in (1,2]).
2375
2376 * "ps" "cp" and "cc" smooths can now be supplied with 2 knots to be
2377 treated as `endpoints' of the smooths (full set of knots can still
2378 be supplied as before).
2379
2380 * The `newton' optimizer was dropping terms when their gradient was
2381 below the convergence threshold (and allowing re-entry). This
2382 promotes zig-zagging unless the terms are independent. Now only
2383 drops terms if gradient and second derivative are very small
2384 (so obective is really flat).
2385
2386 * The adaptive smoothing "ad" class has been greatly simplified and 2D
2387 penalty improved. Much faster as a result, and 2D adaptive actually
2388 quite good.
2389
2390 * gam.fit3 now checks that the initial PIRLS step produces an
2391 improvement in penalized deviance relative to a null model. If not
2392 then step halving towards the null model parameter is employed.
2393 The null model is as close to constant predicted values as the model
2394 structure allows (it is estimated up front in estimate.gam, to save
2395 computation).
2396
2397 * `gam.side' now takes account of whether the model has an intercept
2398 (or the model model matrix column corresponding to an intercept is
2399 in the column space of the model matrix of the parametric model
2400 components).
2401
2402 * The smoothing parameter array returned by `gam' now includes names
2403 for the smoothing parameters.
2404
2405 * s and te check that `id' is a single element.
2406
2407 * By default, partial residuals are no longer plotted for smooths with
2408 `by' variables since they are usually meaningless here (they can
2409 be re-instated by argument `by.resids').
2410
2411 * `min.sp' processing modified to work with `paraPen' argument to
2412 `gam'.
2413
2414 * vcov.gam defaults to Bayesian covariance matrix.
2415
2416 * indexing error in `parametricPenalty' corrected.
2417
2418 * plot.gam modified so that page change behaviour is like plot.glm
2419
2420 * negbin family upgraded to work with (RE)ML.
2421
2422 * `sp.vcov' function added to extract covariance matrix of log
2423 smoothing parameters from (RE)ML based fits.
2424
2425 * `power' links now handled by default fitting methods (i.e. gam.fit3)
2426
2427 * `magic.post.proc' now expects weights, not sqrt(weights) as the
2428 `w' argument (unless `w' is a matrix).
2429
2430 * p-values tweaked again, for slightly better performance with smooths of
2431 several variables. Still not quite right.
2432
2433 * record of intial sp's is now carried in `smooth' objects in field
2434 `sp'.
2435
2436 * ?linear.functional.term error fix.
2437
2438 * memory leak in magic.c:magic fixed --- all fixed smoothing
2439 parameters lead to 2 arrays being left unfreed.
2440
2441 * various .Rd file fixes.
2442
2443 1.4-2
2444
2445 * Some minor .Rd file fixes
2446
2447 1.4-1
2448
2449 * `Predict.matrix2' was not in NAMESPACE: fixed.
2450
2451 * term specific offsets handled properly w.r.t. `by' variables in
2452 `smoothCon' (a rather specialized topic!)
2453
2454 * minor doc bug fix for `smooth.construct'.
2455
2456 1.4-0
2457
2458 *** Model terms can now include linear functionals of smooths, by
2459 supplying matrix arguments and matrix `by' variables to model
2460 smooth terms. This allows, for example, a model to depend on the
2461 integral of a smooth function, or its derivative, or for models to
2462 depend on functional predictors. See ?linear.functional.terms. Main
2463 code changes are in `smoothCon' and `predMat'.
2464
2465 ** Smooth terms can now be linked in order that they have the same
2466 smoothing parameters (and, by default, bases). Linkage is specified
2467 using the `id' argument to `s' or `te'. Terms with the same `id' value
2468 will have the same smoothing parameter(s).
2469
2470 ** `by' variables can now be factor variables. Also smooth terms with a
2471 `by' variable are only subject to a sum-to-zero constraint if it
2472 is needed for identifiability.
2473
2474 ** Argument `paraPen' of `gam' allows (multiple) penalization of
2475 parametric model terms. This allows `gam' to fit any model that
2476 can be expressed as a penalized GLM.
2477
2478 ** p-values returned by `summary.gam' now default to a Bayesian
2479 approximation which gives (substantially) better frequentist behaviour
2480 than the old method.
2481
2482 ** The 2 standard error bands for smooths shown by `plot.gam' can now
2483 include the uncertainty about the overall mean, by default. Such
2484 intervals have better coverage probability (of their target of
2485 inference) than intervals for centred smooths. Argument
2486 `type="iterms"' to `predict.gam' will return such standard errors.
2487
2488 ** An adaptive smoother class has been added, for smoothing with respect
2489 to one or two variables: invoked with `s(...,bs="ad",...).
2490
2491 ** `gamm' now supports nested smooth terms, and uses the same, constraint
2492 absorbed, parameterization as `gam'.
2493
2494 ** `s' and `te' terms accept an `sp' argument setting the term specific
2495 smoothing parameters (and over-riding argument `sp' of `gam'). Ignored
2496 by `gamm'.
2497
2498 ** Negative binomial handling changed. `negbin' family added: adapted from
2499 MASS to work with gam outer iteration fitting. `gam.negbin' fitting
2500 routine added in order to enable use of `negbin' with outer iteration.
2501 See ?negbin for details. MASS families no longer supported.
2502 `nb.theta.mult' removed from `gam.control'.
2503
2504 ** The Eilers and Marx style p-spline class is now one of the default
2505 smoothing classes, rather than just being an example of how to set up
2506 a class in the help file. cyclic versions are also available.
2507
2508 ** `smoothCon' now handles `by' variables and centering constraints
2509 automatically, removing the need for smooth constructors to do so.
2510 `PredMat' handles `by' variables automatically. Users can over-ride
2511 this behaviour when adding smooth classes, if needed - see
2512 documentation.
2513
2514 ** The interface for adding user defined smooths has been simplified,
2515 but this may mean that some user defined classes which worked before
2516 no longer work: see ?user.defined.smooths
2517
2518 * `smooth.construct' methods are now expected to set default values
2519 for the penalty order `p.order' and the basis dimension `bs.dim'
2520 if none are supplied. They should also sanity check supplied
2521 values. Previously this was done by `s', but this put unhelpful
2522 restrictions on new smooth classes.
2523
2524 * `smooth.construct' now expects to recieve `data' and `knots' arguments
2525 with names corresponding exactly to `object$term'. In addition `data'
2526 should contain only what is required by `object$term' + a final column
2527 containing a `by' variable, if present. Predict.mat expects the same of
2528 its `data' argument. wrapper functions smooth.construct2 and
2529 Predict.mat2 will accept a data frame containing any number of variables
2530 -- all that is required is that `object$terms' can be evaluated using
2531 it. These functions handle repeat rows in matrix arguments efficiently.
2532
2533 * bug fix in `plot.gam' -- no longer requires to hit return if `select'
2534 used (ever).
2535
2536 * bug fix in `fixDependence' --- a completely dependent `X2' would not
2537 be detected, since the first element of R2 would be zero: used first
2538 element of R1 to set scale instead.
2539
2540 * `gam.fit' passes corrected n to `magic' so that `n' used in gcv/ubre
2541 does not include obs with zero prior weight. `gam.fit3' already doing
2542 this...
2543
2544 * `magic' and `gam.fit3' now allow log smoothing parameters to be a
2545 linear transformation of a smaller set of underlying smoothing parameters.
2546
2547 * `mgcv' based fitting has been removed as an option in `gam', as has
2548 Pearson based GCV. In consequence `am' argument removed from
2549 `gam.method' and `globit' removed from `gam.control'.
2550
2551 * `get.var' now coerces matrix values to numeric vectors, to facilitate
2552 the handling of linear functionals of smooths.
2553
2554 * `gam.fit2' has been removed, since gam.fit3 is simply better.
2555
2556 * The default optimizer for the generalized case has been made slightly
2557 more efficient (derivative free evaluation of GCV/AIC has been
2558 improved). The upshot is that the default is now faster than performance
2559 iteration in almost all cases (while still being more reliable).
2560
2561 * the `absorb.cons' option has been removed from `gam.control'.
2562
2563 * `fix.family.link' and `fix.family.var' bug fix --- only return
2564 family unmodified if all required derivative functions are present.
2565
2566 * `smoothCon' now returns a list of smooth objects to facilitate factor
2567 `by' variables.
2568
2569 * `smoothCon' makes smooth object labels more informative, if there are
2570 `by' variables... this also makes default plots more informative.
2571
2572 * `plot.gam' indicates `by' variables in labels
2573
2574 * `gamm.setup' modified to call `gam.setup' for most of the setup, leaving
2575 just the re-parameterization step to do.
2576
2577 * `gamm' modified to allow constraint absorption (same as `gam')
2578
2579 * `gamm' bug fixed whereby "cc" smooths would get the wrong null space
2580 dimension (effect was small, but noticeable, in practice e.g. Cairo
2581 temperature example from chapter 6 of Wood, 2006, book).
2582
2583 * print methods now return first argument invisibly as they should.
2584
2585 * code for (very) old style summary removed.
2586
2587 * `gam.fit3' now traps derivative iteration divergence, and suggests
2588 tightening the convergence tolerance `epsilon' in `gam.control'.
2589 Divergence can happen for ill-conditioned models if the PIRLS has
2590 not converged sufficiently.
2591
2592 * gamm.Rd updated to reflect change to gammPQL in 1.3-28.
2593
2594 1.3-31
2595
2596 * There was a most annoying warning generated by R 2.7.0 every time `gam'
2597 was used. Now there isn't.
2598
2599 1.3-30
2600
2601 * change to DESCRIPTION file.
2602
2603 1.3-29
2604
2605 * `magic' could segfault if supplied with many constraints and relatively
2606 high rank penalties, so that after constriant the penalty matrix
2607 square roots had more columns than rows (never happened in additive
2608 model case, but can happen in more general settings). Fixed.
2609
2610 * `gamm' now silently drops grouping factors within the correlation
2611 structure formulae that duplicate random effects grouping factors
2612 (which automatically act as grouping factors on the correlation
2613 structures anyway).
2614
2615 * Some replacement of dubious `as.matrix' calls with use of `,drop=FALSE]'
2616 in gamm.r
2617
2618 1.3-28
2619
2620 ** `gamm' modified to call a routine `gammPQL' in place of MASS::glmmPQL.
2621 This avoids some duplication, and facilitates maintainance.
2622
2623 * Bug fix in `formXtViX' where matrix dimensions got dropped when
2624 subsetting thereby messing up variance calculations for gamm fits in
2625 which some group sizes were 1.
2626
2627 1.3-27
2628
2629 ** Fix of nasty bug in large dataset handling with "tp" basis (introduced
2630 in 1.3-26). Subsampling code was re-seeding RNG instead of intended
2631 behaviour of saving RNG state and restoring it. Fixed and tested.
2632
2633 1.3-26
2634
2635 * modification to `gam' so that GCV/UBRE scores reported with all fixed
2636 smoothing parameters are consistent with equivalent under s.p.
2637 estimation.
2638
2639 * gam.fit3 modified to test for convergence of coefficients as well
2640 as penalized deviance, otherwise in extreme cases the derivative
2641 iterations can diverge.
2642
2643 * modifications of gam.setup, predict.gam and plot.gam to allow smooths
2644 to contribute an offset term to the model (offset is returned from
2645 smooth.construct or Predict.matrix as an "offset" attribute of
2646 model/prediction matrix). This is useful for smooths which have known
2647 boundary conditions of some sort.
2648
2649 * PredictMat can now handle NAs in a returned prediction matrix.
2650
2651 * vis.gam can handle NA's in predictions.
2652
2653 ** Modification of large dataset handling for "tp" and "ts" bases. If
2654 there are more that 3000 unique covariate combinations for a tprs then
2655 3000 combinations are randomly sub-sampled, and used as the initial
2656 knots for tprs basis construction. The same random number seed is used
2657 every time, (R's RNG state is unaltered by this). Control of this is
2658 usually via the `max.knots' (default 3000) and `seed' (default 1)
2659 elements of the `xt' argument of `s'. In consequence, `max.tprs.knots'
2660 has been removed from `gam.control'.
2661
2662 * Modification of `s' and `te' to allow an extra argument `xt' which can
2663 contain extra information to pass to the basis constructors for smooths.
2664
2665 * removal of `full.call' from smooth.spec objects - it wasn't used
2666 anywhere any more, and is a pain to maintain.
2667
2668 * removal of `full.formula' from the `gam' object - it is no longer used
2669 anywhere and requires alot of code to construct.
2670
2671 1.3-25
2672
2673 * A bug in `null.space.dimension' caused prediction to fail for `s' terms
2674 of 4 or more variables, unless the `m' argument was supplied explicitly
2675 (and was large enough for the number of variables). Fixed.
2676
2677 1.3-24
2678
2679 * summary.gam modified so that it behaves correctly if fitting routines
2680 detect and deal with rank deficiency in parameteric part of a model.
2681
2682 * spring cleaning of help files.
2683
2684 * gam.check modified to report more useful convergence diagnostics.
2685
2686 ** `model.matrix.gam' added.
2687
2688 ** "cr" basis constructor modified to use the same centering conditions
2689 as other bases (sum to zero over covariates, rather than parameters
2690 sum to zero). This makes centred confidence intervals for smooths, of
2691 the sort used in plot.gam, behave in a similar way for all bases. With
2692 the old "cr" centering constraint there could be high negative
2693 correlation between coefficients of a centered smooth and the intercept:
2694 this could make centred "cr" smooth CIs wider than CIs for other bases
2695 (not really wrong, but disconcerting).
2696
2697 1.3-23
2698
2699 * step size correction bug fixed in gam.fit3. `Perfect' convergence could
2700 cause the divergence control loop to fail: the divergence control loop
2701 was asking for near strict decrease in the penalized deviance, which
2702 could be numerically impossible to achieve if the algorithm had actually
2703 converged completely.... fixed.
2704
2705 * minor doc bug fixes.
2706
2707
2708 1.3-22
2709
2710 * Cheap but unneccesary code added to gdi.c and magic.c to stop
2711 inappropriate uninitialized variable warnings from some compilers.
2712
2713 ** Bad bug in gam.fit3 fixed. Prior weights of zero were not handled
2714 correctly - prior weight vector should have been subsetted before
2715 gdi call, but this didn't happen. Result was infinite derivatives
2716 and fit failure. fixed.
2717
2718 * Related bug in gam.fit3: dropped observations not handled correctly
2719 in deviance calculation, which can result in inappropriate step
2720 halving. fixed.
2721
2722 * inner loop 3 in gam.fit2 and gam.fit3 modified so that step halving
2723 continues until penalized deviance is at worst non-increasing.
2724
2725 * stupid bug in summary.gam, p-value calc. fixed.
2726
2727 1.3-21
2728
2729 * minor bug in gam.fit() - edf array not passed to `mgcv.find.theta'
2730 if method "perf.magic" used - so wrong EDF used for theta estimation
2731 with neagative binomial. fixed.
2732
2733 * Theta estimate added to family object of fitted gam if negative binomial
2734 used...
2735
2736 * extract.lme.cov(2) modified to allow use with single level grouping
2737 factors (not really sure when this is useful)
2738
2739 * bug in gam4objective called when using gam.method(outer="nlm") - never
2740 used GCV.
2741
2742 * fixed bug in `newton' whereby immediate convergence actually caused
2743 routine to fail.
2744
2745 * modified `smoothCon' and `predictMat' so that `qrc' attribute always
2746 created if constraint absorption used, even if there are no constraints.
2747 This attribute can then be used to test that there are no unabsorbed
2748 constraints (e.g. in `gam.outer').
2749
2750 1.3-20
2751
2752 * Bad bug in `newton' - step halving set up so that step *never*
2753 accepted (it still beat all previous methods in simulations)
2754
2755 * Minor bug in `newton' step limiting of Newton steps reduced step
2756 to max component 1, rather than `maxNstep'.
2757
2758 * Some documentation fixes
2759
2760 1.3-19
2761
2762 *** SUBSTANTIAL CHANGE: Improved outer iteration added via gdi.c coupled
2763 with gam.fit3. Exact first and second derivatives of GACV/GCV/UBRE/AIC
2764 are now available via new iteration methods. These improve the
2765 speed and reliability of fitting in the *generalized* additive model
2766 case.
2767
2768 * numerous changes to NAMESPACE and gamm related functions to pass
2769 codetools checks.
2770
2771 ** gam.method() modified to allow GACV as an option for outer GCV
2772 model selection.
2773
2774 * magic.c::mgcv_mmult modified so that all inner loop calculations are
2775 optimal (i.e. inner loop pointers increments are all 1).
2776
2777 * `smooth.construct' functions for "cc" and "cr" smooths now increase `k'
2778 to the minimum possible value (and warn), if it's too low.
2779
2780 ** `gam' modified to allow passing of `mustart' etc to gam.fit and
2781 gam.fit2, properly
2782
2783 * `gam' modified to fix a bug whereby fitting in two steps using argument
2784 `G' could fail when some sp's are to be estimated and some fixed.
2785
2786 ** an argument `in.out' added to `gam' to allow user initialization of
2787 smoothing parameters when using `outer' iteration in the generalized
2788 case. This can speed up analyses that rely on several refits of the same
2789 model.
2790
2791 1.3-18
2792
2793 * gamm modifed so that weights dealt with properly if lme type varFunc
2794 used. This is only possible in the non-generalized case, as gamm.Rd
2795 now clarifies.
2796
2797 * slight modification to s() to add `width.cutoff=500' to `deparse'
2798
2799 * by variables not handled properly in p-spline example in
2800 smooth.construct.Rd - fixed.
2801
2802 * bug fix in summary.gam.Rd example (pmax -> pmin)
2803
2804 * color example added to plot.gam.Rd
2805
2806 * bug fix in `smooth.construct.tensor.smooth.spec' - class "cyclic.smooth"
2807 marginals no longer re-parameterized.
2808
2809 * `te' documentation modified to mention that marginal reparameterization
2810 can destabilize tensor products.
2811
2812 1.3-17
2813
2814 * print.summary.gam prints estimated ranks more prettily (thanks Martin
2815 Maechler)
2816
2817 ** `fix.family.link' can now handle the `cauchit' link, and also appends a
2818 third derivative of link function to the family (not yet used).
2819
2820 * `fix.family.var' now adds a second derivative of the link function to
2821 the family (not yet used).
2822
2823 ** `magic' modified to (i) accept an argument `rss.extra' which is added
2824 to the RSS(squared norm) term in the GCV/UBRE or scale calculation; (ii)
2825 accept argument `n.score' (defaults to number of data), the number to
2826 use in place of the number of data in the GCV/UBRE calculation.
2827 These are useful for dealing with very large data sets using
2828 pseudo-model approaches.
2829
2830 * `trans' and `shift' arguments added to `plot.gam': allows, e.g. single
2831 smooth models to be easily plotted on uncentred response scale.
2832
2833 * Some .Rd bug fixes.
2834
2835 ** Addition of choose.k.Rd helpfile, including example code for diagnosing
2836 overly restrictive choice of smoothing basis dimension `k'.
2837
2838 1.3-16
2839
2840 * bug fix in predict.gam documentation + example of how to predict from a
2841 `gam' outside `R'.
2842
2843 1.3-15
2844
2845 * chol(A,pivot=TRUE) now (R 2.3.0) generates a warning if `A' is not +ve
2846 definite. `mroot' modified to supress this (since it only calls
2847 `chol(A,pivot=TRUE)' because `A' is usually +ve semi-definite).
2848
2849 1.3-14
2850
2851 * mat.c:mgcv_symeig modified to allow selection of the LAPACK routine
2852 actually used: dsyevd is the routine used previously, and seems very
2853 reliable. dsyevr is the faster, smaller more modern version, which it
2854 seems possible to break... rest of code still calls dsyevd.
2855
2856 * Symbol registration added (thanks largely to Brian Ripley). Version
2857 depends on R >= 2.3.0
2858
2859 1.3-13
2860
2861 * some doc changes
2862
2863 ** The p-values for smooth terms had too low power sometimes. Modified
2864 testing procedure so that testing rank is at most
2865 ceiling(2*edf.for.term). This gives quite close to uniform p-value
2866 distributions when the null is true, in simulations, without excessive
2867 inflation of the p-values, relative to parametetric equivalents when
2868 it is not. Still not really satisfactory.
2869
2870 1.3-12
2871
2872 * vis.gam could fail if the original model formula contained functions of
2873 covariates, since vis.gam calls predict.gam with a newdata argument
2874 based on the *model frame* of the model object. predict.gam now
2875 recognises that this has happened and doesn't fail if newdata is a model
2876 frame which contains, e.g. log(x) rather than x itself. offset handling
2877 simplified as a result.
2878
2879 * prediction from te smooths could fail because of a bug in handling the
2880 list of re-parameterization matrices for 1-D terms in
2881 Predict.matrix.tensor.smooth. Fixed. (tensor product docs also updated)
2882
2883 * gamm did not handle s(...,fx=TRUE) terms properly, due to several
2884 failures to count s(...,fx=FALSE) terms properly if there were fixed
2885 terms present. Now fixed.
2886
2887 * In the gaussian additive mixed model case `gamm' now allows "ML" or
2888 "REML" to be selected (and is slightly more self consistent in
2889 handling the results of the two alternatives).
2890
2891 1.3-11
2892
2893 * added package doc file
2894
2895 * added French error message support (thanks to Philippe Grosjean), and
2896 error message quotation characters (thanks to Brian Ripley.)
2897
2898 1.3-10
2899
2900 * a `constant' attribute has been added to the object returned by
2901 predict.gam(...,type="terms"), although what is returned is still not an
2902 exact match to what `predict.lm' would do.
2903
2904 ** na.action handling made closer to glm/lm functions. In particular,
2905 default for predict.gam is now to pad predictions with NA's as opposed
2906 to dropping rows of newdata containing NA's.
2907
2908 * interpret.gam had a bug caused by a glitch in the terms.object
2909 documentation (R <=2.2.0). Formulae such as y ~ a + b:a + s(x) could
2910 cause failure. This was because attr(tf,"specials") is documented as
2911 returning indices of specials in `terms'. It doesn't, it indexes
2912 specials in the variables dimension of the attr(tf,"factors") table:
2913 latter now used to translate.
2914
2915 * `by' variable use could fail unreasonably if a `by' variable was not of
2916 mode `numeric': now coerced to numeric at appropriate times in smooth
2917 constructors.
2918
2919 1.3-9
2920
2921 * constants multiplying TPRS basis functions were `unconventional' for d
2922 odd in function eta() in tprs.c. The constants are immaterial if you are
2923 using gam, gamm etc, but matter if you are trying to get out the
2924 explicit representation of a TPRS term yourself (e.g. to differentiate
2925 a smooth exactly).
2926
2927 1.3-8
2928
2929 * get.var() now checks that result is numeric or factor (avoids
2930 occasional problems with variable names that are functions - e.g `t')
2931
2932 * fix.family.var and fix.family.link now pass through unaltered any family
2933 already containing the extra derivative functions. Usually, to make a
2934 family work with gam.fit2 it is only necessary to add a dvar function.
2935
2936 * defaults modified so that when using outer iteration, several performance
2937 iteration steps are now used for initialization of smoothing parameters
2938 etc. The number is controlled by gam.control(outerPIsteps). This tends
2939 to lead to better starting values, especially with binary data. gam,
2940 gam.fit and gam.control are modified.
2941
2942 * initial.sp modified to allow a more expensive intialization method, but
2943 this is not currently used by gam.
2944
2945 * minor documentation changes (e.g. removal of full stops from titles)
2946
2947 1.3-7
2948
2949 * change to `pcls' example to account for model matrix rescaling changing
2950 smoothing parameter sizes.
2951
2952 * `gamm' `control' argument set to use "L-BFGS-B" method if `lme' is using
2953 `optim' (only does this if `nlminb' not present). Consequently `mgcv' now
2954 depends on nlme_3.1-64 or above.
2955
2956 * improvement of the algorithm in `initial.sp'. Previously it was possible
2957 for very low rank smoothers (e.g. k=3) to cause the initialization to
2958 fail, because of poor handling of unpenalized parameters.
2959
2960 1.3-6
2961
2962 * pdIdnot class changed so that parameters are variances not standard
2963 deviations - this makes for greater consistency with pdTens class, and
2964 means that limits on notLog2 parameterization should mean the same thing
2965 for both classes.
2966
2967 ** niterEM set to 0 in lme calls. This is because EM steps in lme are not
2968 set up to deal properly with user defined pdMat classes (latter
2969 confirmed by DB).
2970
2971 1.3-5
2972
2973 ** Improvements to anova and summary functions by Henric Nilsson
2974 incorporated. Functions are now closer to glm equivalents, and
2975 printing is more informative. See ?anova.gam and ?summary.gam.
2976
2977 * nlme 3.1-62 changed the optimizer underlying lme, so that indefintie
2978 likelihoods cause problems. See ?logExp2 for the workaround.
2979 - niterEM now reset to 25, since parameterization prevents parameters
2980 wandering to +/- infinity (this is important as starting values for
2981 Newton steps are now more critical, since reparameterization
2982 introduces new local minima).
2983
2984 ** smoothCon modified to rescale penalty coefficient matrices to have
2985 similar `size' to X'X for each term. This is to try and ensure that
2986 gamm is reasonably scale invariant in its behaviour, given the
2987 logExp2 re-parameterization.
2988
2989 * magic dropped dimensions of an array inapproporiately - fixed.
2990
2991 * gam now checks that model does not have more coefficients than data.
2992
2993 1.3-4
2994
2995 * inst/CITATION file added. Some .Rd fixes
2996
2997 30/6/2005 1.3-3
2998
2999 * te() smooths were not always estimated correctly by gamm(): invariance
3000 lost and different results to equivalent s() smooths. The problem seems
3001 to lie in a sensitivity of lme() estimation to the absolute size of the
3002 `S' attribute matrices of a pdTens class pdMat object: the problem did
3003 not occur at the last revision of the pdTens class, and there are no
3004 changes logged for nlme that could have caused it, so I guess it's down
3005 to a change in something that lme calls in the base distribution.
3006 To avoid the problem, smooth.construct.tensor.smooth.spec has been
3007 modified to scale all marginal penalty matrices so that they have
3008 largest singular value 1.
3009
3010 * Changes to GLMs in R 2.1.1 mean that if the response is an array, gam
3011 could fail, due to failure of terms like w * X when w is and array
3012 rather than a vector. Code modified accordingly.
3013
3014 * Outer iteration now suppresses some warnings, until the final fitted
3015 model is obtained, in order to avoid printing warnings that actually
3016 don't apply to the final fit.
3017
3018 * Version number reporting made (hopefully) more robust.
3019
3020 * pdconstruct.pdTens removed absolute lower limit on coef - replaced with
3021 relative lower limit.
3022
3023 * moved tensor product constraint construction to BEFORE by variable
3024 stuff in smooth.construct.tensor.smooth.spec.
3025
3026 1.3-1
3027
3028 * vcov had been left out of namespace - fixed.
3029
3030 * cr and cc smooths now trap the case in which the incorrect number of
3031 knots are supplied to them.
3032
3033 * `s(.)' in a formula could cause a segfault, it get's trapped now,
3034 hopefully it will be handled nicely at some point in the future. Thanks
3035 Martin Maechler.
3036
3037 * wrong n reported in summary.gam() in the generalized case - fixed.
3038 Thanks YK Chau.
3039
3040 1.3-0
3041
3042 *** The GCV/UBRE score used in the generalized case when fitting by
3043 outer iteration (the default) in version 1.2 was based on the Pearson
3044 statistic. It is prone to serious undersmoothing, particularly of binary
3045 data. The default is now to use a GCV/UBRE score based on the deviance:
3046 this performs much better, while still maintaining the enhanced
3047 numerical convergence performance of outer iteration.
3048
3049 * The Pearson based scores are still available as an option (see
3050 ?gam.method)
3051
3052 * For the known scale parameter case the default UBRE score is now
3053 just a linearly rescaled AIC criterion.
3054
3055 1.2-6
3056
3057 * Two bugs in smooth.sconstruct.tensor.smooth.spec: (i) incorrect
3058 testing of class of smooth before re-parameterizing, so that cr smooths
3059 were re-parameterized, when there is no need to; (ii) knots used in
3060 re-parameterization were based on quantiles of the relevant marginal
3061 covariate, which meant that repeated knots could be generated: now uses
3062 quantiles of unique covariate values.
3063
3064 * Thanks to Henric Nilsson a bug in the documentation of magic.post.proc has
3065 been fixed.
3066
3067 1.2-5
3068
3069 ** Bug fix in gam.fit2: prior weights not subsetted for non-informative
3070 data in GCV/UBRE calculation. Also plot.gam modified to allow for
3071 consequent NA working residuals. Thanks to B. Stollenwerk for reporting
3072 this bug.
3073
3074 ** vcov.gam written by Henric Nilsson included... see ?vcov.gam
3075
3076 * Some minor documentation fixes.
3077
3078 * Some tweaking of tolerances for outer iteration (was too lax).
3079
3080 ** Modification of the way predict.gam picks up variables.
3081 (complication is that it should behave like other predict functions, but
3082 warn if an incomplete prediction data frame is supplied -since latter
3083 violates what white book says).
3084
3085 1.2-2
3086
3087 *** An alternative approach to GCV/UBRE optimization in the
3088 *generalized* additive model case has been implemented. It leads to more
3089 reliable convergence for models with concurvity problems, but is slower
3090 than the old default `performance iteration'. Basically the GAM IRLS
3091 process is iterated to convergence for each trial set of smoothing
3092 parameters, and the derivatives of the GCV/UBRE score w.r.t. smoothing
3093 parameters are calculated explicitly as part of the IRLS iteration. This
3094 means that the GCV/UBRE optimization is now `outer' to the IRLS
3095 iteration, rather than being performed on each working model of the IRLS
3096 iteration. The faster `performance iteration' is still available as an
3097 option. As a side effect, when using outer iteration, it is not possible
3098 to find smoothing parameters that marginally improve on the GCV/UBRE
3099 scores of the estimated ones by hand tuning: this improves the logical
3100 self consistency of using GCV/UBRE scores for model selection purposes.
3101
3102 * To facilitate the expanded list of fitting methods, `gam' now has a
3103 `method' argument requiring a 3 item list, specifying which method to
3104 use for additive models, which for generalized additive models and if using
3105 outer iteration, which optimization routine to use. See ?gam.method for
3106 details. `gam.control' has also been modified accordingly.
3107
3108 *** By default all smoothing bases are now automatically
3109 re-parameterized to absorb centering constraints on smooths into the
3110 basis. This makes everything more modular, and is usually user
3111 transparent. See ?gam.control to get the old behaviour.
3112
3113 ** Tensor product smooths (te) now use a reparameterization of the
3114 marginal smoothing bases, which ensures that the penalties of a tensor
3115 product smooth retain the interpretation, in terms of function shape, of
3116 the marginal penalties from which they are induced. In practice this
3117 almost always improves MSE performance (at least for smooth underlying
3118 functions.) See ?te to turn this off.
3119
3120 *** P-values reported by anova.gam and summary.gam are now based on
3121 strictly frequentist calculations. This means that they are much better
3122 justified theoretically, and are interpretable as ordinary frequentist
3123 p-values. They are still conditional on smoothing parameters, however,
3124 and are hence underestimates when smoothing parameters have been
3125 estimated.
3126
3127 ** Identifiability side conditions modified to work with all smooths
3128 (including user defined). Now works by identifying possible dependencies
3129 symbolically, but dealing with the resulting degeneracies numerically.
3130 This allows full ANOVA decompositions of functions using tensor product
3131 smooths, for example.
3132
3133 * summary.gam modified to deal with prior weights in adjusted r^2
3134 calculation.
3135
3136 ** `gam' object now contains `Ve' the frequentist covariance matrix of
3137 the paremeter estimators, which is useful for p-value calculation. see
3138 ?gamObject and ?magic.post.proc for details.
3139
3140 * Now depends on R >=2.0.0
3141
3142 * Default residual plots modified in `gam.check'
3143
3144 ** Added `cooks.distance.gam' function.
3145
3146 * Bug whereby te smooths ignored `by' variables is now fixed.
3147
3148 1.1-6
3149
3150 * Smoothing parameter initialization method changed in magic, to allow
3151 better initialization of te() terms. This affects default gam fits.
3152
3153 * gamm and extract.lme.cov2 modified to work correctly when the
3154 correlation structure applies to a finer grouping than the random
3155 effects. (Example of this added to gamm help file)
3156
3157 * modifications of pdTens class. pdFactor.pdTens now returns a vector,
3158 not a matrix in accordance with documentation (in nlme 3.1-52). Factors
3159 are now always of form A=B'B (previously, could be A=BB') in accordance
3160 with documentation (nlme 3.1-52). pdConstruct.pdTens now tests whether
3161 initializing matrix is proportional to r.e. cov matrix or its inverse
3162 and initializes appropriately. gamm fitting with te() class tested
3163 extensively with modifications and nlme 3.1-52, and lme fits with pdTens
3164 class tested against equivalent fits made using re-parameterization and
3165 pdIdent class. In particular for gamm testing : model fits with single
3166 argument te() terms now match their equivalent models using s() terms;
3167 models fitted using gam() and gamm() match if gam() is called with the
3168 gamm() estimated smoothing parameters.
3169
3170 * modifications of gamm() for compatibility with nlme 3.1-52: in
3171 particular a work around to allow everything to work correctly with a
3172 constructed formula object in lme call.
3173
3174 * some modifications of plot.gam to allow greater control of
3175 appearance of plots of smooths of 2 variables.
3176
3177 * added argument `offset' to gam for further compatibility with
3178 glm/lm.
3179
3180 * change to safe prediction for parameteric terms had a bug in offset
3181 handling (offset not picked up if no newdata supplied, since model frame
3182 not created in this case). Fixed. (thanks to Jim Young for this) 1.1-5
3183
3184 * predict.gam had a further bug introduced with parametric safe
3185 prediction. Fixed by using a formula only containing the actual variable
3186 names when collecting data for prediction (i.e. no terms like
3187 `offset(x)')
3188
3189 1.1-5
3190
3191 * partial argument matching made col.shade be matched by col passed in
3192 ..in plot.gam, taking away user control of colors. 1.1-5
3193
3194 * 2d smooth plotting in plot.gam modified.
3195
3196 * plot.gam could fail with residuals=TRUE due to incorrect counting in
3197 the code allowing use of termplot. plot.gam failed to prompt before a
3198 newpage if there was only one smooth. gam and gamm .Rd files updated
3199 slightly.
3200
3201 1.1-3
3202
3203 * extract.lme.cov2 could fail for random effect group sizes of 1
3204 because submatrices with only a row or column lose their dimensions, and
3205 because single number calls to diag() result in an identity matrix.
3206
3207 1.1-2
3208
3209 * Some model formulae constructed in interpret.gam and used in
3210 facilitating safe prediction for parametric terms had the wrong
3211 environment - this could cause gam to fail to find data when e.g. lm,
3212 would find it. (thanks Thomas Maiwald)
3213
3214 * Some items were missing from the NAMESPACE file. (thanks Kurt
3215 Hornik)
3216
3217 * A very simple formula.gam function added, purely to facilitate
3218 better printing of anova method results under R 2.0.0.
3219
3220 1.1-1
3221
3222 * Due, no doubt, to gross moral turpitude on the part of the author,
3223 gamm() calculated the complete estimated covariance matrix of the
3224 response data explicitly, despite the fact that this matrix is usually rather
3225 sparse. For large datasets this could easily require more memory than
3226 was available, and huge computational expense to find the choleski
3227 decomposition of the matrix. This has now been rectified: when the
3228 covariance matrix has diagonal or block diagonal structure, then this is
3229 exploited.
3230
3231 * Better examples have been added to gamm().
3232
3233 * Some documentation bugs were fixed.
3234
3235 1.1-0
3236
3237 Main changes are as follows. Note that `gam' object has been modified, so
3238 old objects will not always work with version 1.1 functions.
3239
3240 ** Two new smooth classes "cs" and "ts": these are like "cr" and "tp"
3241 but can be penalized all the way down to zero degrees of freedom to
3242 allow fully automatic model selection (more self consistent than having a
3243 step.gam function).
3244
3245 * The gam object expanded to allow inheritance from type lm and type
3246 glm, although QR related components of glm and lm are not available
3247 because of the difference in fitting method between glm/lm and gam.
3248
3249 ** An anova method for gam objects has been added, for *approximate*
3250 hypothesis testing with GAMs.
3251
3252 ** logLik.gam added (logLik.glm with df's fixed): enables AIC() to be
3253 used with gam objects.
3254
3255 ** plot.gam modified to allow plotting of order 1 parametric terms via
3256 call to termplot.
3257
3258 * Thanks to Henric Nilsson option `shade' added to plot.gam
3259
3260 * predict.gam modified to allow safe prediction of parametric model
3261 components (such as poly() terms).
3262
3263 * predict.gam type="terms" now works like predict.glm for parametric
3264 components. (also some enhancements to facilitate calling from
3265 termplot())
3266
3267 * Range of smoothing parameter estimation iteration methods expanded
3268 to help with non-convergent cases --- see ?gam.convergence
3269
3270 * monotonic smoothing examples modified in light of above changes.
3271
3272 * gamm modified to allow offset terms.
3273
3274 * gamm bug fixed whereby terms in a model formula could get lost if
3275 there were too many of them.
3276
3277 * gamm object modified in light of changes to gam object.
3278
3279 1.0-7
3280
3281 * Allows a model frame to be passed as `newdata' to predict.gam: it
3282 must contain all the terms in the gam objects model frame, `model'.
3283
3284 * vis.gam() now passes a model frame to predict.gam and should be more
3285 robust as a result. `view' and `cond' must contain names from
3286 `names(x$model)' where x is the gam object.
3287
3288 1.0-6/5/4
3289
3290 * partial residuals modified to be IRLS residuals, weighted by IRLS
3291 weights. This is a much better reflecton of the influence of residuals
3292 than the raw IRLS residuals used before.
3293
3294 * gamm summary sorted out by using NextMethod to get around fact that
3295 summary.pdMat can't be called directly (not in nlme namespace exports).
3296
3297 * niterPQL and verbosePQL arguments added to gamm to allow more
3298 control of PQL iteration.
3299
3300 * backquote=TRUE added when deparsing to allow non-standard names.
3301 (thanks: Brian Ripley)
3302
3303 * bug in gam corrected: now gives correct null deviance when an offset
3304 is present. (thanks: Louise Burt)
3305
3306 * bug in smooth.construct.tp.smooth.spec corrected: k=2 caused a
3307 segfault as the C code was reseting k to 3 (actually null space
3308 dimension +1), and not enough space was being allocated in R to handle
3309 the resultng returned objects. k reset in R code, with warning. (Thanks:
3310 Jari Oksanen)
3311
3312 * predict.gam() now has "standard" data searching using a model frame
3313 based on a fake formula produced from full.formula in the fitted object.
3314 However it also warns if newdata is present but incomplete. This means
3315 that if newdata does not meet White book specifications, you get a
3316 warning, but the function behaves like predict.lm etc. predict.gam had
3317 been segfaulting if variables were missing from newdata (Thanks: Andy
3318 Liaw and BR)
3319
3320 * contour option added to vis.gam
3321
3322 * te smooths can be forced to use only a single penalty (theoretical
3323 interest only - not recommended for practical use)
3324
3325 1.0-3
3326
3327 * Fixes bugs in handling graphics parameters in plot.gam()
3328
3329 * Adds option of partial residuals to plot.gam()
3330
3331 1.0-2/1
3332
3333 * Fixes a bug in evaluating variables of smooths, knots and by-variables.
3334
3335 1.0-0
3336
3337 *** Tensor product smooths - any bases available via s() terms in a gam
3338 formula can be used as the basis for tensor product smooths of multiple
3339 covariates. A separate wiggliness penalty and smoothing parameter is
3340 associated with each `marginal' basis.
3341
3342 ** Cyclic smoothers: penalized cubic regression splines which have the
3343 same value and first two derivatives at their first and last knots.
3344
3345 *** An object oriented approach to handling smooth terms which allows
3346 the user to add their own smooths. Smooth terms are constructed using
3347 smooth.construct method functions, while predictions from individual
3348 smooth terms are handled by predict.matrix method functions.
3349
3350 ** p-splines implemented as the illustrative example for the above in
3351 the help files.
3352
3353 *** A generalized additive mixed model function gamm() with estimation
3354 via lme() in the normal-identity case and glmmPQL() otherwise. The main
3355 aim of the function is to allow a defensible way of modelling correlated
3356 error structures while using a GAM.
3357
3358 * The gam object itself has changed to facilitate the above. Most
3359 information pertaining to smooth terms is now stored in a list of smooth
3360 objects, whose classes depend on the bases used. The objects are not
3361 back compatible, and neither are the new method functions. This has been done
3362 in an attempt to minimize the scope for bugs, given the amount of time
3363 available for maintenance.
3364
3365 ** s() no longer supports old stlye (version <0.6) specification of
3366 smooths (e.g. s(x,10|f)). This is in order to reduce the scope for
3367 problems with user defined smooth classes.
3368
3369 * The mgcv() function now has an argument list more similar to magic().
3370
3371 * Function GAMsetup() has been removed.
3372
3373 * I've made a general attempt to make the R code a bit less like a
3374 simultaneous translation from C.
3375
3376 0.9-5/4/3/2/1
3377
3378 * Mixtures of fixed degree of freedom and estimated degree of freedom
3379 smooths did not work correctly with the perf.iter=FALSE option. Fixed.
3380
3381 * fx=TRUE not handled correctly by fit.method="magic": fixed.
3382
3383 * some fixes to GAMsetup and gam documentation.
3384
3385 * call re-instated to the fitted gam object to allow updating
3386
3387 * -Wall and -pedantic removed from Makevars as they are gcc specific.
3388
3389 * isolated call to Stop() replaced by call to stop()!
3390
3391 0.9-0
3392
3393 *** There is a new underlying smoothing parameter selection method,
3394 based on pivoted QR decomposition and SVD methods implemented in LAPACK.
3395 The method is more stable than the Wood (2000) method and allows the
3396 user to fix some smoothing parameters while estimating others,
3397 regularize the GAM fit in non-convergent cases and put lower bounds on
3398 smoothing parameters. The new method can deal with rank deficient
3399 problems, for example if there is a lack of identifiability between the
3400 parametric and smooth parts of the model. See ?magic for fuller details.
3401 The old method is still available, but gam() defaults to the new method.
3402
3403 * Note that the new method calls LAPACK routines directly, which means
3404 that the package now depends on external linear algebra libraries,
3405 rather than relying entirely on my linear algebra routines. This is a
3406 good thing in terms of numerical robustness and speed, but does mean
3407 that to install the package from source you need a BLAS library installed
3408 and accesible to the linker. If you sucessfully installed R by building
3409 from source then you should have no problem: you have everything already
3410 installed, but occasionally users may have to install ATLAS in order to
3411 install from source.
3412
3413 * Negative binomial GAMs now use the families supplied by the MASS library
3414 and employ a fast integrated GCV based method for estiamting the
3415 negative binomial parameter. See ?gam.neg.bin for details. The new
3416 method seems to converge slightly more often than the old method, and
3417 does so more quickly.
3418
3419 * persp.gam() has been replaced by a new routine vis.gam() which is
3420 prettier, simpler and deals better with factor covariates and at all
3421 with `by' variables.
3422
3423 * NA's can now be handled properly in a manner consistent with lm()
3424 and glm() [thanks to Brian Ripley for pointing me in the right direction
3425 here] and there is some internal tidying of GAM so that it's behavious
3426 is more similar to glm() and lm().
3427
3428 * Users can now choose to `polish' gam model fits by adding an nlm()
3429 based optimization after the usual Gu (2002) style `power iteration' to
3430 find smoothing parameters. This second stage will typically result in a
3431 slightly lower final GCV/UBRE score than the defualt method, but is much
3432 slower. See ?gam.control for more information.
3433
3434 * The option to add a ridge penalty to the GAM fitting objective has been
3435 added to help deal with some convergence issues that occur when the
3436 linear predictor is essentially un-identifiable. see ?gam.control.
3437
3438 0.8-7
3439
3440 * There was a bug in the calculation of identifiability side conditions
3441 that could lead to over constraint of smooths using `by' variables in
3442 models with mixtures of smooths of different numbers of variables. This
3443 has been fixed.
3444
3445 0.8-6
3446
3447 * Fixes a bug which occured with user supplied smoothing parameters, in
3448 which the weight vector was omitted from part of the influence (hat)
3449 matrix calculation. This could result in non-sensical variance
3450 estimates.
3451
3452 * Stronger consistency checks introduced on estimated degrees of freedom.
3453
3454 0.8-5
3455
3456 * mgcv was using Machine() which is deprecated from R 1.6.0, this
3457 version uses .Machine instead.
3458
3459 0.8-4
3460
3461 * There was a memory bug which could occur with the "cr" basis, in
3462 which un-allocated memory was written to in the tps_g() routine in the
3463 compiled C code - this occured when that routine was asked to clean up
3464 its memory, when there was nothing to clean up. Thanks to Luke Tierney for
3465 finding this problem and locating it to tps_g()!
3466
3467 * A very minor memory leak which occured when knots are used to start
3468 a tps basis was fixed.
3469
3470 0.8-3
3471
3472 * Elements on leading diagonal of Hat/Influence matrix are now
3473 returned in gam object.
3474
3475 * Over-zealous error trap introduced at 0.8-2, caused failure with
3476 smoothless models.
3477
3478 0.8-2
3479
3480 * User can now supply smoothing parameters for all smooth terms (can't
3481 have a mixture of supplied and estimated smoothing parameters). Feature
3482 is useful if e.g. GCV/UBRE fails to produce sensible estimates.
3483
3484 * svd() replaced by La.svd() in summary.gam().
3485
3486 * a bug in the Lanczos iteration code meant that smooths behaved
3487 poorly if the smooth had exactly one less degree of freedom than the
3488 number of data (the wrong eigenvectors were retained in this case) -
3489 this was a rather rare bug in practice!
3490
3491 * pcls() was not using sensible tolerances and svdroot() was using
3492 tolerances incorrectly, leading to problems with pcls(), now fixed.
3493
3494 * prior weights were missing from the pearson residuals.
3495
3496 * Faulty by variable documentation fixed (have lost name of person who
3497 let me know this, but thanks!)
3498
3499 * Scale factor removed from Pearson residual calculation for
3500 consistancy with a higher proportion of authors.
3501
3502 * The proportion deviance explained has been added to summary.gam() as
3503 a better measure than r-squared in most cases.
3504
3505 * Routine SANtest() has been removed (obsolete).
3506
3507 * A bug in the select option of plot.gam has been fixed.
3508
3509 0.8-1
3510
3511 * The GCV/UBRE score can develop phantom minima for some models: these
3512 are minima in the score for the IRLS problem which suggest large
3513 parameter changes, but which disappear if those large changes are
3514 actually made. This problem occurs in some logistic regression models.
3515 To aid convergence in such cases, gam.fit now switches to a cautious
3516 mgcv optimization method if convergence has not been obtained in a user
3517 defined number of iterations. The cautious mode selects the local
3518 minimum of the GCV/UBRE closest to the previous minimum if multiple
3519 minima are present. See gam.control for details about controlling
3520 iterations.
3521
3522 * Option trace in gam.control now prints and plots more useful
3523 information for diagnosing convergence problems.
3524
3525 * The one explicit formation of an inverse in the underlying multiple
3526 GCV optimization has been replaced with something more stable (and
3527 quicker).
3528
3529 * A bug in the calculation of side conditions has been fixed - this
3530 caused a failure with models having parametric terms and terms like:
3531 s(x)+s(z)+s(z,x).
3532
3533 * A bug whereby predict.gam simply failed to pick up offset terms has
3534 been fixed.
3535
3536 * gam() now drops unused levels in factors.
3537
3538 * A bug in the conversion of svd convergence criteria between version
3539 0.7-2 and 0.8-0 has been fixed.
3540
3541 * Memory leaks have been removed from the C code (thanks to the superb
3542 dmalloc library).
3543
3544 * A bug that caused an undignified exit when 1-d smoothing with full
3545 splines in 0.8-0 has been fixed.
3546
3547 0.8-0
3548
3549 * There was a problem on some platforms resulting from the default
3550 compiler optimizations used by R. Specifically: floating point registers
3551 can be used to store local variables. If the register is larger than a
3552 double (as is the case for Intel 486 and up), this means that:
3553 double a,b;
3554 a=b;
3555 if (a==b)
3556 can evaluate as FALSE. The mgcv source code assumed that this could
3557 never happen (it wouldn't under strict ieee fp compliance, for example).
3558 As a result, for some models using the package compiled using some
3559 compiler versions, the one dimensional "overall" smoothing parameter
3560 search could fail, resulting in convergence failure, or undersmoothing.
3561 The Windows version from CRAN was OK, but versions installed under Linux
3562 could have problems. Version 0.8 does not make the problematic
3563 assumption.
3564
3565 * The search for the optimal overall smoothing parameter has been
3566 improved, providing better protection against local minima in the
3567 GCV/UBRE score.
3568
3569 * Extra GCV/UBRE diagnostics are provided, along with a function
3570 gam.check() for checking them.
3571
3572 * It is now possible for the user to supply "knots" to be used when
3573 producing the t.p.r.s. basis, or for the cubic regression spline basis.
3574 This makes it feasible to work with very large datasets using the
3575 of the data. It also provides a mechanism for obtaining purely "knot
3576 based" thin plate regression splines.
3577
3578 * A new mechanism is provided for allowing a smooth term to be
3579 multiplied by a covariate within the model. Such "by" variables allow
3580 smooths to be conditional on factors, for example.
3581
3582 * Formulae such as y~s(x)+s(z)+s(x,z) can now be used.
3583
3584 * The package now reports the UBRE score of a fitted model if UBRE was
3585 used for smoothing parameter selection, and the GCV score otherwise.
3586
3587 * A new help page gam.models has been added.
3588
3589 * A bug whereby offsets in model formulae only worked if they were at
3590 the end of the formulae has been fixed.
3591
3592 * A bug whereby weights could not be supplied in the model data frame
3593 has been fixed.
3594
3595 * gam.fit has been upgraded using the R 1.5.0 version of glm.fit
3596
3597 * An error in the documentaion of xp in the gam object has been fixed,
3598 in addition to numerous other changes to the documentation.
3599
3600 * The scoping rules employed by gam() have been brought into line with
3601 lm() and glm by searching for variables in the environment of the model
3602 formula rather than in the environment from which gam() was called -
3603 usually these are the same, but not always.
3604
3605 * A bug in persp.gam() has been fixed, whereby slice information had
3606 to be supplied in a particular order.
3607
3608 * All compiled code calls now specify package mgcv to avoid any
3609 possibility of calling the wrong function.
3610
3611 * All examples now set the random number generator seed to facilitate
3612 cross platform comparisons.
3613
3614 0.7-2
3615
3616 * T and F changed to TRUE and FALSE in code and examples.
3617
3618 * Minor predict.gam error fixed (didn't get correct fitted values if
3619 called without new data and model contained multi-dimensional smooths).
3620
3621 0.7-1
3622
3623 * There was a somewhat over-zealous warning message in the single
3624 smoothing parameter selection code - gave a warning everytime that GCV
3625 suggested a smoothing parameter at the boundary of the search interval -
3626 even if this GCV function was also flat. Fixed.
3627
3628 * The search range for 1-d smoothing parameter selection was too wide
3629 - it was possible to give so little weight to the data that numerical
3630 problems caused all parameters to be estimates as zero (along with the
3631 edf for the term!). The range has been narrowed to something more sensible
3632 [above warning should still be triggered if it is ever too narrow - but
3633 this should not be possible].
3634
3635 * summary.gam() documentation extended a bit. p-values for smooths are
3636 slightly improved, and an example included that shows the user how to
3637 check them!
3638
3639 0.7-0
3640
3641 * The underlying multiple GCV/UBRE optimization method has been
3642 considereably strengthened, as follows:
3643 o First and second guess starting values for the relative
3644 smoothing parameters have been improved.
3645 o Steepest descent is used if either: i) the Hessian of the
3646 objective is not positive definite, or (ii) Steps in the Newton direction
3647 fails to improve the GCV/UBRE score after 4 step halvings (since in
3648 this case the quadratic model is clearly poor).
3649 o Newton steps are rescaled so that the largest step component
3650 (in log relative smoothing parameters) is of size 5 if any step
3651 components are >5. This avoids very large Newton steps that can occur
3652 in flat regions of the objective.
3653 o All steepest descent steps are initially scaled so that their
3654 longest component is 1, this avoids long steps into flat regions of
3655 the objective.
3656 o MGCV Convergence diagnostics are returned from routines mgcv
3657 and gam.
3658 o In gam.fit() smoothing parameters are re-auto-initialized
3659 during IRLS if they have become so far apart that some are likely to
3660 be in flat parts of the GCV/UBRE score.
3661 o A bug whereby poor second guesses at relative smoothing
3662 parameters could lead to acceptance of the first guess at these
3663 parameters has been removed.
3664 o The user is warned if the initial smoothing parameter guesses
3665 are not improved upon (can happen legitmately if all s.p.s should be
3666 very high or very low.)
3667
3668 The end result of these changes is to make fits from gam much more
3669 reliable (particularly when using the tprs basis available from version
3670 0.6).
3671
3672 * A summary.gam and associated print function are provided. These
3673 provide approximate p-values for all model terms.
3674
3675 * plot.gam now provides a mechanism for selecting single plots, and
3676 allows jittering of rug plots.
3677
3678 * A bug that prevented models with no smooth terms from being fitted
3679 has been removed.
3680
3681 * A scoping bug in gam.setup has been fixed.
3682
3683 * A bug preventing certain mixtures of the bases to be used has been
3684 fixed.
3685
3686 * The neg.bin family has been renamed neg.binom to avoid masking a
3687 function in the MASS library.
3688
3689 0.6-2
3690 revisions from 0.6.1
3691
3692 * Relatively important fix in low level numerics. Under some circumstances
3693 the Lanczos routines used to find the thin plate regression spline basis
3694 could fail to converge or give wrong answers (many thanks to Charles
3695 Paxton for spotting this). The problem was with an insufficiently stable
3696 inverse iteration scheme used to find eigenvectors as part of the
3697 Lanczos scheme. The scheme had been used because it was very fast:
3698 unfortuantely stabilizing it is as computationally costly as simply
3699 accumulating eigen-vectors with the eigen-values - hence the latter has
3700 now been done. Some further examples also added.
3701
3702 0.6-1
3703
3704 * Junk files removed from src directory.
3705
3706 * 3 C++ style comments removed from tprs.c.
3707
3708 0.6-0
3709
3710 * Multi-dimesional smoothing is now available, using "thin plate
3711 regression splines" (MS submitted). These are based on optimal
3712 approximations to the thin-plate splines.
3713
3714 * gam formula syntax upgraded (see ?s ). Old syntax still works, with
3715 the exception that if no df specified then the tprs basis is always used
3716 by default.
3717
3718 * plot.gam can now deal with two dimensional smooth terms as well as
3719 one dimensional smooths.
3720
3721 * persp.gam added to allow user to visualize slices through a gam
3722 [Mike Lonergan]
3723
3724 * negative binomial family added [Mike Lonergan] - not quite as robust
3725 as rest of families though [can have convergence problems].
3726
3727 * predict.gam now has an option to return the matrix mapping the
3728 parameters to the linear predictor at the supplied covariate values.
3729
3730 * Variance calculation has been made more robust.
3731
3732 * Routine pcls added, for penalized, linearly constrained optimization
3733 (e.g. monotonic splines).
3734
3735 * Residual method provided (there was a bug in the default - Thanks
3736 Carmen Fernandez).
3737
3738 * The cubic regression spline basis behaved wrongly when extrapolating
3739 [thanks Sharon Hedley]. This is now fixed.
3740
3741 * Tests included to check that there are enough unique covariate
3742 combinations to support the users choise of smoothing basis dimension.
3743
3744 * Internal storage improved so that large numbers of zeroes are no
3745 longer stored in arrays of matrices.
3746
3747 * Some method argument lists brought into line with the R default
3748 versions.
3749
3750 0.5
3751
3752
3753 * There was a bug in gam.fit(). The square roots of the correct iterative
3754 weights were being used in place of the weights: the bug was
3755 apparent because the sum of fitted values didn't always equal the sum of
3756 the response data when using the canonical link (which it should as a
3757 result of X'f=X'y when canonical link used and unpenalized). The bug has
3758 been corrected, and the correction tested. This problem did not affect
3759 (unweighted) additive models, only generalized additive models.
3760
3761 * There was a bug that caused a crash in the compiled code when there were
3762 more than 8000 datapoints to fit. This has been fixed.
3763
3764 * The package now reports its version number when loaded into R.
3765
3766 * predict.gam() now returns predictions for the original covariate values
3767 (used to fit the model) when called without new data.
3768
3769 * predict.gam() now allows type="response" as an argument - returning
3770 predictions on the scale of the response variable.
3771
3772 * plot.gam() no-longer defaults to automatic page layout, use argument
3773 pages=1 to get the old default behaviour.
3774
3775 * A bug that could cause a crash with the model formula y~s(x)-1 has been
3776 fixed.
3777
3778 * Yet more sloppy practices are now allowed for naming variables in model
3779 formulae. e.g. d$y ~ s(d$x) now works, although its not recommended.
3780
3781 * The GCV score is now reported by print.gam() (whether or not GCV was
3782 actually used - it isn't the default for Poisson or binomial).
3783
3784 * plot.gam() modified to avoid prompting for input when not used
3785 interactively.
3786
3787 0.4
3788
3789 * Transformations allowed on lhs of gam formulae .
3790
3791 * Argument order same as Splus gam.
3792
3793 * Search for data now designed to be like lm() , so you can now be quite
3794 sloppy about where your data are.
3795
3796 * The above mean that Venables and Ripley examples can be run without
3797 having to read the documentation for gam() so carefully!
3798
3799 * A bug in the standard error calculations for parametric terms in
3800 predict.gam() is fixed.
3801
3802 * A serious bug in the handling of factors was fixed - it was previously
3803 possible to obtain a rank deficient design matrix when using factors,
3804 despite having specified an identifiable model.
3805
3806 * Some glitches when dealing with formulae containing offset() and/or I()
3807 have been fixed.
3808
3809 * Fitting defaults can now be altered using gam.control when calling gam()
3810
3811 0.3-3
3812
3813 * Documentation updated, including removal of wrong information about
3814 constraints and mgcv . Also some readability changes in code and no
3815 smooths are now allowed.
3816
3817 0.3-2/1
3818
3819 * Allows all ways of specifying a family that glm() allows (previously
3820 family=poisson or family="poisson" would fail). Some more documentation
3821 fixes.
3822
3823 * 0.2 lost the end of long formulae (because of a difference in the way
3824 that R and Splus deal with formulae). This is now fixed.
3825
3826 * A minor error that meant that QT() failed under some versions of Windows
3827 is now fixed.
3828
3829 * All package functions now have help(). Also the help files have been
3830 more carefully checked - version 0.2 actually contained no information
3831 on how to write a GAM formula as a result of a single missing '}' in the
3832 help file!
3833
3834 0.2
3835
3836 * Fixed d.f. regression splines allowed as part of gam() model
3837 specification.
3838
3839 * Bug in knot placement algorithm fixed (caused crash with df close to
3840 number of data).
3841
3842 * Replicate covariate values dealt with properly in gam()!
3843
3844 * Data search method in gam() revised - now looks in frame from which
3845 gam() called.
3846
3847 * plot.gam() can now deal with missing variance estimates gracefully.
3848
3849 * Low (1,2) d.f. smooths dealt with gracefully by gam() - no longer cause
3850 freeze or crash.
3851
3852 * Confidence intervals simulation tested for normal(identity),
3853 poisson(log), binomial(logit) and gamma(log) cases. Average coverage
3854 probabilities from 0.89 to 0.97 term by term, 0.93 to 0.96 "across the
3855 model", for nominal 0.95.
3856
3857 * R documentation updated and tidied.