Move the anova.coxmelist function from the survival package to coxme.
Add logLik methods for coxme and lmekin.
The lmekin function ignored the x=TRUE and y=TRUE arguments. Simply an oversight.
Modify the gwas example in the lmekin vignette to use 2*kmat; rescaling kinship matrices is no longer automatic.
Remove the warning message for "diagonal is not constant" when a variance matrix is rescaled. It now almost always arises for a kinship matrix where there is inbreeding, and in that case it is not an error but does cause user confusion. Earlier in coxme development, when users built their own matrices by hand, it often corresponded to a mistake and so was useful.
Change the output for lmekin. Let the variance of the random effects be A = sB where s= variance of the residuals, the internal calculations are done in terms of B. In the prior code the print routine translated from B to A and the vcoef component was in terms of B. Now the translation happens at the end of lmekin. This is more handy for those who extract and use the components.
When a variance matrix is supplied as the varlist argument, the random effects are always produced in the same order as the rows/columns of said matrix. The labels attached to the vector, however, were in the order of the variables in the data set. Pointed up in an example from M deAndrade.
Expose the grid search values used to find initial values for the variance parameters, as part of the coxme.control and lme.control functions.
Add a summary.coxme method. It is currently identical to print, but some users were getting confused by the results printed out by the default summary function.
A serious bug in lmekin was pointed out by Paola Sebastiani. It occurs whenever the variance matrix of the random effects is not diagonal: in one location I had an upper diagonal matrix U when I needed to have the lower diagonal form U'. This affects all the genetics models. A test case with non-diagonal variance has been added. Unfortunately, all of the prior test cases were designed to validate agains lme and thus had diagonal penalty matrices. Also address a slowness problem by speeding up the conversion between bdsmatrix and matrix objects.
Change the default in coxmeMlist to rescale=FALSE. For kinship matrices, which have a diagonal of 1/2, automatic rescaling was a handy feature. For other types of matrices, e.g., a smoothing spline example currently in development, it causes much confusion.
Add an anova method.
Add a genetic example to the lmekin vignette.
Add dimnames to the result of vcov, per the suggestion of Joshua Wiley
Fix an indexing error, when there were multiple random terms and the first of them had only sparse intercept terms, then entries for the second were incorrectly placed in the combined matrix. This is a serious bug, I can only assume that people were not using multiple random terms enough to have encountered it.
Allow for different integration methods in the refine.n code. This is a research question, and the option may disappear once I answer it.
Further refinement to the laplace and coxme vignettes.
Remove a Makefile from the vignettes directory that was causing errors at CRAN. Why it didn't lead to errors here is unknown.
Create vignettes for the laplace approximation and lmekin. In the process changed the way in which the refine.n option works. It is now more robust.
Fix two bugs in the return values, caused by incorrect mapping of the b coefficients into the final output, when there were multiple random terms in the model. (The fit was correct. One bug the returned ranef(fit) to be incorrectly ordered, the other led to NA in the linear.predictor.)
Fix a bug noted by Lawrence Hunsicker: if a model had no fixed covariates or only an offset, the gridsearch for starting values would always declare failure. One variable was mistakenly NA.
Add a new version of lmekin to the package. This also marks the start of a conversion to use of the Matrix library.
Change the definition of ranef and add VarCorr so as to match the usage in lme and lmer. This will break some old code, but will save much heartache in the long term. Add a vcov method as well.
Further interaction with Gabor Mezaros and his huge data set showed that I didn't get the iteration counts quite right (see second bullet for 2.1-0). As a positive, we now have cross checks and agreement with the Survival Kit software for a substantial example, a major increase in our assurance of the code's correctness. The change in iteration paths causes small changes in some tests, all statistically inconsequential.
Fix a bug pointed out by Goran Brostrom: using (start, stop) data the program could give a memory fault. A 1 character typo was using the wrong ending variable in a for loop, if the model had random slopes. Added two more tests to the suite to catch this.
Updates to the vignette per suggestions from Marianne (co-author).
Add further checks on user-written coxme variance functions. The goal was to give a more informative error message for common oversights.
Update the manual pages for ranef and fixef to remove a warning message from the package checks.
Add a small vignette that shows the use of user-constructed variance methods.
Printout change: the chisq, AIC, BIC printed when thinking of the model as a penalized calculation should use the full, unpenalized value of the likelihood. I had been using loglik - penalty.
The default number of iterations was made more intelligent. This was driven by a query from Gabor Meszaros; his large model required 10 iterations for an ordinary coxph fit, so the default inner.iter value of 4 was totally inadequate leading to a nonsense iteration path. Now the program uses 1 + the number of iterations used in the no-frailty model or 4, whichever is greater. A large fraction of coxph models require 3 iterations, so there will be no change for most runs.
Per a bug report from Sophie Armitage, there was an indexing error in the code when using multiple random effects terms, in the map for the X matrix. This has prompted a new release.
Allow for multiple starting estimates. It became clear that a single default starting element that would always work was not realistic. The default grid is now (.02, .1, .4, .8)^2 for variances and (0, .3) for correlations; all elements in the grid are tested and the best is used as a starting value for optim. This changed the iteration path for several of the tests, of course.
Change the default starting value of variances from .2 to .1. The final answer will be closer to .1 the majority of the time, and I had a data set where .2 led to wrong iteration. This area needs more investigation.
Fix error in the fitting routine; if there were 3 or more crossed terms the indexing was wrong, leading to no penalty for terms after the second.
Allow for NULL as an initial value. This occured for a model with 3 terms (x1|grp) + (x2|grp) + (x3|grp), and I wanted to give values for x2 and x3 only. You can't use names since they all are the same, instead I now use vfixed=list(NULL, .1, .1).