The summary.survfit routine assumed that the times argument was sorted, contrary to the documentation. Pointed out by Torsten Hothorn.
The tmerge function would fail if the time variable was a Date object. It was due to the fact that as.Date(as.numeric(x)) fails when x is a date. (A design flaw in Date, IMHO). There were also flaws when both the first and second data set were not sorted by id; added a more complete test case for this.
An earlier change in dim.survfit had felled the survfit.matrix function: it incorrectly assumed strata when there were none. Unfortunately this didn't generate an error but rather multiple copies of a single curve (and an incomprehensible explanation of this single curve in the vignette). Pointed out by E Lundt.
print.summary.survfitms would complain if only a single time was specified. A case where drop=FALSE was needed.
Create a stronger test suite for summary.survfit, and use it to actually fix the error that 2.39-3 claimed to fix. This uncovered a long-standing inaccuracy with n.risk for in-between time points.
Add a section on monotone splines to the splines vignette.
For multi-state curves, the returned n.event component lost its dimensions if any of the curves had only one observation.
Fix error in summary.survreg. For multiple curves and requested time points at or before the first time point in the data, the values from curve 1 was used for all. Pointed out by T Eigentler.
Fix an unitialized variable in C code, pointed out by Brian Ripley.
Small updates based on feedback from CRAN
Label the output dimnames from pyears with the variable names from the model. This makes it easier to read.
Replace any refrences to model.frame with "stats::model.frame" (all 38 of them). The model.frame function uses non-standard evaluation rules, and holding its hand like this is the only way to ensure that we don't call a user function of the same name.
The Surv function would almost always label the columns of the resulting matrix, and the glmnet function depended on this. It now always labels them per a request from Trevor Hastie.
Add the finegray function and expand the competing risks vignette to document it.
Add a check to the quantile.survfit function for multi-state models; quantiles are not well defined for this case.
Changes to the iteration path and convergence tests for coxph models with (start, stop] data, driven by two user examples that failed. The data sets had serious statistical issues of collinearity and/or outliers such that the final fits are not practically useful, but now the routine finishes gracefully instead of dying. The upshot is much more care about the order in which additions and subtractions of large numbers are done so as to avoid cancellation error.
Fix an error to summary.survfit with the times argument: for intermediate time points it would sometimes choose the wrong value for the number at risk. (Number at risk is a left continuous function.)
Add more graphical arguments to plot.cox.zph in response to a user request.
Remove some of last vestiges of Splus support from the header files for the C code, per a request from R core to remove mention of S.h.
Multiple updates and corrections to the tmerge function, including improvements to the vignette. (As a result of using it in a class where the TA tried out all manner of combinations.)
Update survSplit: it now handles all types of status variables (0/1, TRUE/FALSE, factors), the id and episode arguments are useful for start/stop data, the data retains its original sort order (new observations are inserted rather than put at the end), and the function is illustrated in a vignette.
Add the conf.times argument to plot.survfit. This allows for confidence bars at specified times, which are useful when the plot is crowded.
Survfit changes that are NOT backward compatable!
Change the default for mark.time to FALSE
Change the behavior of xscale so that it matches that of yscale, i.e., it changes only the label and not the underlying scale. Follow on annotations such as legend or locator are in the orignal scale of the data.
For a matrix of curves, e.g. competing risks, print and plot them in column major order rather than row major, so as to match the usual R behavior.
Fix an error in the help page for the cohort argument of survexp, pointed out by Karl Ove Hufthammer.
The anova and logLik functions would fail when given a null model (right hand side of 1 or only an offset). Pointed out by Karl Ove Hufthammer.
The recently added code to generate an error when the same variable appears on both sides of a formula in coxph (a good idea) caused a failure if there was offset statement that contains a '-' sign. Pointed out by Abra Jeffers.
Add more imports to the NAMESPACE file per a request from CRAN
Add a length method for Surv objects. Requested by Max Kuhn. (2015/6/17).
Fix an error in neardate. When both input data sets were unsorted the last match could be wrong.
Change print.coxph to use the printCoefmat routine, which l leads to nicer p-values. Other print routines will follow unless there is an outcry. (But I forced signif.stars=FALSE: my tolerance of bad practice has limits.)
Make those parts of the competing risks vignette which depend on the cmprsk library conditional. Otherwise the build fails for those without the pacakge.
The coxph function could fail converge for a set of very collinear predictors when using (start, stop) data; revealed in a test case sent by G Borstrom. This was due to deficiency in a check for near infinite coefficients, which had already been updated for some but not all cases. (2015/6/3)
Update anova.coxph to use the model.frame.coxph function; the current code had scoping errors if embedded in a function. Add an anova.coxph.penal function to correctly handle models with pspline terms.
Fix an error in the tmerge function. Using the options argument would generate a spurious error.
Pyears could fail on very long formulas due to a deparse() issue.
Add the number of observations used and deleted due to missing to summary.pyears.
Allow the combination of a null coxph model (~1 on the right) and the exact calculation for tied times. No one had ever asked for this before. (2015/3/25)
Shorten the default printout for survfit. The records, n.max and n.start columns are often the same: if so suppress duplicates.
Move the anova.coxphlist function from the survival package to coxme. (2015/3/3)
Change the logLik method for coxph models so that the nobs component is the number of events rather than the number of rows in the data. This is superior for follow on methods such as AIC.
Add a test to the coxexact.c routine for too large a data set; too many tied times could lead to integer overflow. "Fixing" the error is not sensible: the computation for such a data set would take decades. Add some more explanation to the help pages as well.
Fix an error discovered by CRAN, which triggered a core dump for them on a particular manual page (but never for me). The linear predictors from a frailty model contained NA values (incorrect), leading to failure in survConcordance.fit. (2015/2/16).
An error was found in the mgus data set (a progression after death). Now corrected, and added a little more follow-up time for some subjects.
Add error check for infitinte weights or offsets. This in respose to a bug report where someone did this on purpose, trying to mimic cure fractions, and then found that survfit.coxph failed.
Robust variance is not supported for a coxph model with the "exact" approximation. (Rarely requested and a lot of work to add.) Add an error message to clogit(), so users get a more useful notice of the issue rather than a late error from residuals.coxph.
Update the rats data set: it now includes both female and male litters so as to match the documentation.
The term frailty(x) would fail if x were a factor, and not all levels were present. Pointed out by Theodor Balan.
Fix error of "abs" instead of "fabs" in the agfit4.c code; pointed out to me by CRAN.
Replace all instances of the obsolete prmatrix function.
Modify pyears to allow cbind(time, count) as the response, giving a cumulative sum of counts, when the counts per observation may be other than 0/1.
The lines.survfit function was incorrect for data sets that used the start.time option and xscale (it neglected to rescale the start time.)
An increasingly common error is for user to put the time variable on both sides of a coxph equation in the mistaken belief that this is a way to create time-dependent coefficients. Generate a warning message for this case.
Update the basehaz function to a simple alias for "survfit". Prior versions called surfit but then only returned part of the object. Update 2/2015: reverted the change. It turns out that 6 different packages that depend on survival also depended on the old behavior.
Make the default value for the shortlabel argument of strata() more nuanced. If the argument is a single factor, assume that we don't need to prepend the variable name to its levels.
Return the weights vector, if present, as part of the survreg object.
For interval censored points and the symmetric distributions (Gaussian and logistic) response type residuals were incorrect. Silly error: needed (x-mean)/scale not x/scale - mean.
Martingale residuals could be incorrect for the case of model with (start, stop] data and a pspline term. Refactor the code so that all of the possible code paths call the same C routine to do the residuals. Add a new test for this case, and further tests to verify that predict(type='expected') and residuals agree.
Fix bug pointed out by D Dunker: if a model had both tt() and cluster() terms it would fail with a length error.
Fix a rare bug in plot.survfit: if a multistate curve rose and then later fell to exactly the same value, the line would be incorrect.
Add calls to the R_CheckUserInterrupt to several routines, so that long calculations can be interrupted by the user.
The anova.coxph function would fail if the original call had a subset argument. Pointed out by R Fisher. 11May2014
Remove a dependency on the survey package from the adjusted survival curves vignette, at the request of CRAN. (The base + required bundle needs to be capable of a stand-alone build.)
Fix error in calcuation of the y-axis range for survival curve plots whenever the "fun" argument could produce infinite values, e.g., complimentary log-log plots transform 1 to -Inf. Pointed out by Eva Boj del Val. (Add finite=TRUE to range() call).
The plot for competing risk curves could have a spurious segment. (Found within 3 hours of submitting 2.37-5 to CRAN.)
The lines method for survexp objects was defaulting to a step function, restore the documented default of a connected line.
Add a levels method for tcut objects. 14Jan2014
Add vignette on adjusted survival curves.
Add vignette concerning "type 3" tests.
Make the tt() function invisible outside of a coxph formula. There was a complaint about conflicts with another package, and there is not really a good reason to have it be a global name. An R-devel discussion just over 1 year ago showed how to accomplish this.
The modeling routines are set in two parts, e.g., coxph sets up the model and coxph.fit does the work. Export more of the ".fit" routines to make it easier for other packages to build on top of this one.
Updates to the model.matrix and model.frame logic for coxph. A note from F Harrell showed that I was not correctly dealing with the "assign" attribute when there are strata * factor interactions. This led to cleanup in other cases that I had missed but which never had proven fatal. Also added support for tt() terms to the stand alone model.matrix and model.frame functions. (Residuals for tt models are still not available, but this was a necessary first step to that end.) 26Dec13
The Surv function now remembers attributes of the input variables that were passed to it; they are saved as "inputAttributes". This allows the rms package, for instance, to retain labels and units through the call.
Update summary.coxph.penal to produce an object, which in turn has a print method, i.e., make it a "standard" summary function.
Add a logLik method for coxph and survfit objects.
Allow for Inf as the end of the time interval, for interval censored data in the Surv function.
The predict.coxph function would fail if it had both a newdata and a collapse argument. Pointed out by Julian Bothe. 25Sep13
Survexp can now produce expecteds based on a stratified Cox model. Add the 'individual.s' and 'individual.h' options to return indivudual survival and cumulative hazard estimates, respectively. The result of survfit now (sometimes) includes the cumulative hazard. This will be expanded. 29Jul13
Change code in the coxpenal.fit routine: the use of a vector of symbols as arguments to my .C calls was confusing to a new CRAN consistency check. Both the old and new are legal R; but the old was admittedly an unusual construction and it was simpler to change it.
Fix a bug in survfit.coxph pointed out by Chris Andrews, whose root cause was incorrect curve labels when the id option is used. 27Jun13
Add rsurvreg routine.
Change survfit.coxph routine so that it detects whether newdata contains or does not contain strata variables, and acts accordingly. If newdata does containe strata then the output will contain only those data-value and strata combinations specified by the user. Retain strata levels in the coxph routine for use in the survfit routine, to correctly reconstruct strata levels. Warn about curves with interactions. 18Ju13
Add a dim method for survival curves.
For competing risks curves that use the istate option, the plotted curves now start with the correct (initial) prevalence of each state. 22May13
The survreg function failed with the "robust=T" option. Pointed out by Jon Peck. Test case added. 6May13
Kazuki Yoshida pointed out that rep() had no method for Surv objects. This caused the survSplit routine to fail if the data frame contained a Surv object. 3May13
Per a request from Milan Bouchet-Valet fix an issue in survfit that arose when the OutDec option is set to ',': it did not correctly convert times back from character to numeric.
The plot.survfit function now obeys "cex" for the size of the marks used for censored observations.
Subscripting error in predict.coxph for type=expected, se=T, strata in the model, newdata, and multiple strata in the new data set. Pointed out by Chris Andrews. The test program has been tweaked to include multiple strata in newdata.
Minor flaw in [.survfit. If "fit" had multiple curves, and fit$surv was a matrix, and one of those curves had only a single observation time, fit[i,] would collapse columns when "i" selected that curve, though it shouldn't.
Changed all of the .C and .Call statements to make use of "registered native routines", per R-core request. Add file src/init.c
Error in plot.survfit pointed out by K Hoggart – the "+" signs for censored observations were printing one survival time to the left of the proper spot. Eik Vettorazi found another error if mark.time is a vector of numerics. These are the results of merging the code for plot, lines and points due to some discrepancies between them, plus not having any graphical checks in the test suite.
Repair an error in using double subscripts for the survfitms objects.
Add the US population data set, with yearly totals by age and sex for 2000 onward. It is named uspop2, since there is already a "uspop" data set containing decennial totals from 1790 to 1970.
Not all combinations of strata Y/N and CI Y/N worked in the quantile.survfit function, pointed out by Daniel Wallschlaeger (missing a function argument in one if-else combination). Added a new test routine that verifies all paths.
The first example in predict.survreg help file needed to have
I(age^2) instead of
age^2 in the model: R ignores the
second form. (I'm almost sure this worked at one time, perhaps in Splus).
It also needed different plot symbols to actually match the
referenced figure. Pointed out by Evan Newell.
Fix a long-standing problem with cch pointed out by Ornulf Borgan leading to incorrect standard errors. A check in the underlying coxph routines to deal with out of bounds exponents, added in version 2.36-6, interacted badly with the -100 offset used in cch. It only affected models using (start, stop) survival times.
Two bugs were turned up by running tests for all the packages that depend on survival (158 of them).
Add a new multi-state type to the Surv object. Update the survfit routine to work with it. The major change is addition of a proper variance for this case. More functionality is planned.
Remove the fr_colon.R test program. It tests an ability that has been superseded by coxme, on a numerically touchy data set, and it was slow besides. For several other tests that produce warning messages and are supposed to produce said messages, add extra comments to that effect so testers will know it is expected.
The code has had several "if.R" clauses to accomodate Splus vs R differences, which are mostly class vs oldClass. These are now being removed as I encounter them; since our institution no longer uses Splus I can no longer test the clauses' validity.
The fast subsets routine coxexact.fit incorrectly returned the linear predictor vector in the (internal) sorted order rather than data set order. Pointed out by Tatsuki Koyama, affecting the result of a clogit call. 6Nov2012
Jason Law pointed out that the sample data set "rats" is from the paper by Mantel et.al, but the documentation was for a data set from Gail, Santner and Brown. Added the Gail data as rats2 and fixed the documentation for rats.
For predict.coxph with type="terms", use "sample" as the default value for the reference option. For all others the default remains "strata", the current value. Type terms are nearly always passed forward for further manipulation and per strata centering can mess things up: termplot() for instance will no longer show a smooth function if the results are recentered within strata.
Fix bug in summary.aareg, which was unhappy (without cause) if the maxtime option was used for a fit that did not include the dfbeta option. Pointed out by Asa Johannesen.
The coxph fitting functions would report an error for a null model (no X variables) if init was specified as numeric(0) rather than NULL.
Update the description and citation files to use the new "person" function described in the R Journal. Also add the ByteCompile directive per suggestion of R core.
Allow an ordinary vector as the left hand side of survConcordance.
Update anova.coxphlist to reject models with a robust variance.
The survfit function had an undocumented backwards-compatability that allows the newdata argument to be a vector with no names. An example from Damon Krstajic showed that this does not work when the original model has a matrix in the formula. Removed the feature. (This is for survfit.coxph.) Also clarified the code and its documentation about what is found where – environments, formulas, and the arguments of eval, which fixes a problem pointed out by xxx where the result of a Surv call is used in the coxph formula.
Fix an issue in summary.survfit pointed out by Frank Harrell. The strata variable for the output always had its labels in sorted order, even when a factor creating the survival curves was otherwise. (This was due to a call to factor() in the code.) The print routine would then list curves in sorted order, which might well be contrary to the user's wishes. The curves were numerically correct.
Add the anova.coxmelist function to the namespace so that it is visible. If someone has a list of models the first of which was a coxph fit and the list includes coxme fits, then anova.coxph will be the function called by R, and it will call anova.coxmelist.
Fix a bug pointed out by Yi Zhang and Mickael Hartweg. If a coxph model used an offset, then a predicted survival curve that used newdata (and the offset variable of course) would be wrong, e.g. survival values > 1. A simple misplaced parenthesis was the cause. A recent paper by Langholz shows how to get absolute survival from case-control data using an offset, which seems to have suddenly made this feature popular.
Per further interaction with Yi Zhang, a few items were missing from the S3methods in the NAMESPACE file: as.matrix.Surv, model.matrix.coxph, model.matrix.survreg, model.frame.survreg.
A supposedly cosmetic change to coxph in the last release caused formulas with a "." on the right hand side to fail. Fix this and add a case with "." to the test suite.
Add the anova.coxmelist function. This is in the survival package rather than in coxme since "anova(fit1, fit2)" is valid when fit1 is a coxph and fit2 a coxme object, a case which will cause this function to be called by way of anova.coxph.
More work on "predvars" handling for the pspline function, when used in predict calls. Add a new test of this to the suite, and the makepredictcall method to the namespace. Fixes a bug pointed out by C Crowson.
Deprecate the "robust" option of coxph. When there are multiple observations per subject it is almost surely the wrong thing to do, while adding a "cluster(id)" term does the correct thing. When there is only one obs per subject both methods work correctly.
Add documentation of the output structure to the aareg help file.
Change ratetableDate so that it still allows use of chron objects, but doesn't need the chron library. This eliminates a warning messge from the package checks, but is also a reasonable support strategy for a moribund package. (Some of the local users keep datasets for a long long time.)
Fix a bug in summary.survfit for a multiple-strata survival object. If one of the curves had no data after application of the times argument, an output label was the wrong length.
Fix a bug pointed out by Charles Berry: predict for a Cox model which has strata, and the strata is a factor with not all its levels represented in the data. I had a mistake in the subscripting logic: number of groups is not equal to max(as.integer(strata)).
Changes to avoid overflow in the exponent made in 2.36-6 caused failure for one special usage: in case-cohort designs a dummy offset of -100 could be added to some observations. This was being rounded away. The solution is to 1: have coxsafe not truncate small exponents and 2: do not recenter user provided offset values.
Fix bug in survfit.coxph. Due to an indexing error I would sometimes create a huge scratch vector midway through the calculations (size = max value of "id"); the final result was always correct however. Data set provided by Cindy Crowson which had a user id in the billions.
Fix bug pointed out by Nicholas Horton: predictions of type expected, with newdata, from a Cox model without a strata statement would fail with "x not found". A misplaced parenthesis from an earlier update caused it to not recreate the X matrix even though it was needed later. Also add some further information to the predict manual page to clarify an issue with frailty terms.
Fix a bug in the new fast subsets code. The test suite had no examples of strata + lots of tied times, so of course that's the case where I had an indexing error. Add a test case using the clogit function, which exercises this.
Further memory tuning for survexp.
Make survexp more efficient. The X matrix was being modified in several places, leading to multiple copies of the data. When the data set was large this would lead to a memory shortage.
Cause anova.coxph to call anova.coxme when a list of models has both coxph and coxme objects.
Add the quantile.survfit function. This allows a user to extract arbitrary quantiles from a fitted curve (and std err).
Fix an error in predict.coxph. When the model had a strata and the newdata and reference="sample" arguments were used, it would (incorrectly) ask for a strata variable in the new data set.
Incorporate the fast subsets algorithm of Gail et al, when using coxph with the "exact" option. The speed increase is profound though at the cost of some memory. Reflect this in the documentation for the clogit routine. Note that the fast computation is not yet implemented for (start,stop) coxph models.
Change the C routine used by coxph.fit from .C to .Call semantics to improve memory efficiency, in particular fewer copies of the X matrix.
Add scaling to the above routine. This was prompted by a user who had some variables with a 0-1 range and others that were 0 - 10^7, resulting in 0 digits of accuracy in the variance matrix. (Economics data).
Comment out some code sections that are specific to Splus. This reduced the number of "function not found" warnings from R CMD check.
30 Sept 2011: The na.action argument was being ignored in predict.coxph; pointed out by Cindy Crowson.
The log-likelihood for survreg was incorrect when there are case weights in the model. The error is a fixed constant for any given data set, so had no impact on tests or inferences. The error and correction were pointed out by Robert Kusher.
A variable name was incorrect in survpenal.fit. This was in a program path that had never been traversed until Carina Salt used survreg with a psline(..., method='aic') call, leading to a "variable not found" message.
Punctuation error in psline made it impossible for a user to specify the boundary.knots argument. Pointed out by Brandon Stewart.
Add an "id" variable to the output of survobrien.
The survfitCI routine would fail for a curve with only one jump point (a matrix collapsed into a vector).
Fix an error in survfit.coxph when the coxph model has both a strata by covariate interaction and a cluster statement. The cluster term was not dropped from the Terms object as it should have been, led to a spurious "variable not found" error. Pointed out by Eva Bouguen.
If a coxph model with penalized terms (frailty, pspline) also had a redundant covariate, the linear predictor would be returned as NA. Pointed out by Pavel Krivitsky.
Due to a mistake in my script that submits to CRAN, the fix in 2.36-8 below was actually not propogated to the CRAN submission.
Fix an error in the Cauchy example found in the survreg.distributions help page, pointed out by James Price.
Update the coxph.getdata routine to use the model.frame.coxph and model.matrix.coxph methods.
Add the concordance statistic to the printout for penalized models.
Unitialized variable in calcuation of the variance of the concordance. Found on platform cross-checking by Brian Ripley.
Changed testci to use a fixed file of results from cmprsk rather than invoking that package on-the-fly. Suggested by the CRAN maintainers.
Due to changes in R 2.13 default printout, the results of many of the test programs change in trivial way (one more or fewer digits). Update the necessary test/___.Rout.save files. Per the core team's suggestion the dependency for the package is marked as >=2.13.
An example from A Drummond caused iteration failure in coxph: x=c(1,1,1,0,1, rep(0,35)), time=1:40, status=1. The first iteration overshoots the solution and lands on an almost perfectly linear part of the loglik surface, which made the second iteration go to a huge number and exp() overflows. A sanity check routine coxsafe is now invoked on all values of the linear predictor.
1 April: Fix minor bug in survfit. For left censored data where all the left censored are on the very left, it would give a spurious warning message when trying to create a 0 row matrix that it didn't need or use. Pointed out by Steve Su.
31 March 2011: One of the plots in the r_sas test was wrong (it's been a long time since I visually checked these). The error was in predict.survreg; it had not taken into account a change in R2.7.1: the intercept attribute is reset to 1 whenever one subscripts a terms object, leading to incorrect results for a model with "-1" in the formula and a strata(): the intercept returned when removing the strata. I used this opportunity to move most of the logic into model.frame.survreg and model.matrix.survreg functions. Small change to the model.frame.coxph and model.matrix.coxph functions due to a better understanding of xlevels processing.
Round off error issue in survfit: it used both unique(time) and table(time), and the resulting number of unique values is not guarranteed to be the same for times that differ by a tiny amount. Now times are coverted to a factor first. Peter Savicky from the R core team provided a nice discussion of the issue and helped me clarify how best to deal with it. The prior fix of first rounding to 15 digits was good enough for almost every data set – except the one found by a local user just last week.
Round off error in print.survfit pointed out by Micheal Faye. If a survival value was .5 in truth, but .5- eps due to round off the printed median was wrong. But it was ok for .5+eps. Simple if-then logic error.
Re-fix a bug in survfit. It uses both unique and table in various places, which do not round the same; I had added a pre-rounding step to the code. A data set from Fan Chun showed that I didn't round quite enough. But the prior rounding did work for a time of 2 vs (sqrt(2))^2: this bug is very hard to produce. I now use as.numeric(as.character(factor(x))), which induces exactly the same rounding as table, since it is the same compuation path.
Further changes to pspline. The new Boundary.knots argument allows a user to set the boundary knots inside the range of data. Code for extrapolation outside that range was needed, essentially a copy of the code found in ns() for the same issue. Also added a psplineinverse function, which may be useful with certain tt() calls in coxph.
10 Mar 2011: Add the capablilty for time-dependent transformations to coxph, along with a small vignette describing use of the feature. This code is still incompletely incorporated in that the models work but other methods (residuals, predict, etc) are not yet defined.
8 Mar 2011: Expand the survConcordance function. The function
now correctly handles strata and time dependent covariates, and
computes a standard error for the estimate. All computation is based
on a balanced binary tree strucure, which leads to computation in
O(n log(n)) time.
coxph function now adds concordance to its output, and
summary.coxph displays the result.
8 Mar 2011: Add the "reference" option to predict.coxph, a feature and need pointed out by Stephen Bond.
4 Mar 2011: Add a makepredictcall method for pspline(), which in turn required addition of a Boundary.knots argument to the function.
25 Feb 2011: Bug in pyears pointed out by Norm Phillips. If a subject started out with "off table" time, their age was not incremented by that amount as they moved forward to the next "in table" cell of the result. This could lead to using the wrong expected rate from the rate table.
20 Feb 2011: Update survConcordance to correctly handle case weights, time dependent covariates, and strata.
18 Feb 2011: Bug in predict.coxph found by a user (1 day after 36-4!). If the coxph call had a subset and predict used newdata, the subset clause was "remembered" in the newdata construction, which is not appropriate.
17 Feb 2011: Fix to predict.coxph. A small typo that only was exercised if the coxph model had x=T. Discovered via induced error in the rankhazard package. Added lines to the test suite to test for this in the future.
Removed some files from test and src that are no longer needed.
Update the configure script per suggestion from Kurt H.
13 Feb 2011: Add the rmap argument to pyears, as was done for survexp, and update the manual pages and examples. Fix one last bug in predict.coxph (na.action use). Passes all the tests for inclusion on the next R release.
8 Feb 2011: Change the name of the new survfit.coxph.fit routine to survfitcoxph.fit; R was mistaking it for a survfit method. Fix errors in predict.coxph when there is a newdata argument, including adding yet another test program.
1 Feb 2011: Fix bugs in coxph and survreg pointed out by Heinz Tuechler and email@example.com, independently, that were the same wrong line in both programs. With interactions, a non-penalized term could be marked as penalized due to a mismatched vector length, leading to a spurious error message later in the code.
1 Feb 2011: Update survfit.coxph to handle the case of a strata by covariate interaction. All prior releases of the code did this wrong, but it is a very rare case (found by Frank Harrell). Added a new test routine coxsurv4. Also found a bug in [.survfit; for a curve with both strata and multiple columns, as produced by survfit.coxph, it could drop the n.censored item when subscripting. A minor issue was fixed in coxph: when iter=0 the output coefficient vector should be equal to the input even when the variance is singular.
30 Jan 2011: Move the noweb files to a top level directory, out of inst/. They don't need to be copied to binary installs.
22 Jan 2011: Convert the Changelog files to the new inst/NEWS.Rd format.
1 Jan 202011: The match.ratetable would fail when passed a data frame with a character variable. This was pointed out by Heinz Tuechler, who also did most of the legwork to find it. It was triggered by the first few lines of tests/jasa.R (expect <- ....) when options(stringsAsFactors=FALSE) is set.
20 Dec 2010: Add more test cases for survfit.coxph, which led to significant updates in the code.
18 Nov 2010: Add nevent to the coxph output and printout in response to a long standing user request.
14 Dec 2010: Add an as.matrix method for Surv objects.
11 Nov 2010: The prior changes broke 5 packages: the dependencies form a bigger test suite than mine! 1. Survival curve for a coxph model with sparse frailty fit; fixed and added a new test case. 2. survexp could fail if called from within a function due to a scoping error. 3. "Tsiatis" was once a valid type (alias for 'aalen') for survfit.coxph; now removed from the documentation but the code needed to be backwards compatable. The other two conflicts were fixed in the packages that call survival. There are still issues with the rms package which I am working out with Frank H.
27 Oct 2010: Finish corrections and test to the new code. It now passes the checks. The predict.coxph routine now does strata and standard errors correctly, factors propogate through to predictions, and numerous small errors are addressed. Predicted survival curves for a Cox model has been rewritten in noweb and expanded. Change the version number to 2.36-1.
17 Oct 2010: Per a request from Frank Harrell (interaction with his library), survfit.coxph no longer reconstructs the model frame unless it really needs it: in some cases the 'x' and 'y' matrices may be sufficient, and may be saved in the result. Add an argument "mf" to model.matrix.coxph for more efficient interaction when a parent routine has already recovered the model frame. In general, we are trying to make use of model.matrix.coxph in many of the routines, so that the logic contained there (remove cluster() calls, pull out strata, how to handle intercepts) need not be replicated in multiple places.
12 Oct 2010: Fix a bug in the modified lower limits for survfit (Dory & Korn). A logical vector was being inadvertently converted to numeric. Pointed out by Andy Mugglin. A new case was added to the test suite.
15 July 2010: Add a coxph method for the logLik function. This is used by the AIC function and was requested by a user.
29 July 2010: Fix 2 bugs in pyears. The check for a US rate table was off (minor effect on calculations), and there was a call to julian which assumed that the origin argument could be a vector.
21 July 2010: Fix a problem pointed out by a user: calling survfit with almost tied times, e.g., c(2, sqrt(2)^2), could lead to an inconsistent result. Some parts of the code saw these as 2 unique values per the unique() function, some as a single value using the results of table(). We now pre-round the input times to one less decimal digit than the max from .Machine$double.digits. Also added the noweb.R processing function from the coxme package, so that the noweb code can be extracted "on the fly" during installation using commands in the configure and cleanup scripts.
11 July 2010: A rewrite of the majority of the survfit.coxph code. The primary benefits are 1: finally tracked down and eliminated the bug for standard errors of case weights + Cox survival + Efron method; 2: the individual=TRUE and FALSE options now use the same underlying code for curves, before there were some options valid only for one or the other; 3: code was rewritten using noweb with a considerable increase in documentation; 4: during the verification process some errors were found in the test suite and corrected, e.g., a typo in my book led to failure of an all.equal test in book4.R. Similar to the rewrite for survfit several years ago, the new code has far less use of .C to help transparency.
21 May 2010: Fix bug in summary.survfit. For a survival curve from a Cox model with start,stop data, the 'times' argument would generate an error.
24 May 2010: Fix an annoyance in summary.survfit. When the survival data had an event or censor at time 0 and summary is called with a times argument, then my constructed call to approx() would have duplicate x values. The answer was always right, but approx has begun to print a bothersome warning message. A small change to the constructed argument vector avoids it.
7 April 2010: Minor bug pointed out by Fredrik Lundgren. In survfit if the method was KM (default) and error = Tsiatis an error message results. Simple fix: code went down the wrong branch.
24 Feb 2010: Serious bug pointed out by Kevin Buhr. In Surv(time1, time2,stat) if there were i) missing values in time1 and/or time2, ii) illegal value sets with time1 >=time2, and iii) all the instances of ii do not preceed all the instances of i, then the wrong observation (not the illegal) will be thrown out. Repaired, and a new test added. Minor updates to 3 test files: survreg2, testci, ratetable.
8 Feb 2010: Bug pointed out by Heinz Tuechler – if a subscript was dropped from a rate table the 'type' attribute got dropped, e.g. survexp.usr[,1,,].
26 Jan 2010: At the request of Alex Bokov, added the xmax, xscale, and fun arguments to points.survfit.
26 Jan 2010: Fix bug pointed out by Thomas Lumley – with case weights <1 a Cox model with (start, stop) input would inappropriately decide it needed to do step halving to find a solution, eventually failing to converge. It was treating a loglik >0 as an indication of failure, but such values arise for small case weights. Let L(w) be the loglik for a data set where everyone is given a weight of w, then L(w)= wL(1) - d log(w) where d=number of deaths in the data. For small enough w positivity of L(w) is certain.
25 Jan 2010: Fix bug in summary.ratetable pointed out by Heinze Tuechler. Added a call to the function to the test suite as well.
15 Dec 2009: Two users pointed out a bug that crept into survreg() with a cluster statement, when a t(x) fix, but in response I added another test that more formally checks the dfbeta residuals and found a major oversight for the case of multiple strata.
14 Dec 2009: 1.Fix bug in frailty.xxx, if there is a missing value in the levels it gets counted by "length(unique(x))" (frailty is called before NA removal.) 2.SurvfitCI had an incorrect CI with case weights, and 3. in survreg a call to resid instead of residuals.survreg, before the class was attached.
11 Nov 2009: The 'type' argument does not make sense for plot.survfit. (If type='p', should one plot the tops of the step function, the bottoms, or both?). Make it explicitly disallowed in response to an R-help query, rather than the confusing error message that currently arose.
28 Oct 2009: The basehaz function would reorder the labels of the strata factor. Not a bug really, but a "why do this?" Unintended consequence of a character -> factor conversion.
1 Oct 2009: Fix a bug pointed out by Ben Domingue. There was one if-then-else path into step-halving in the frailty.controldf routine that would refer to a non-existent variable. A very rarely followed path, obviously, and with the obvious fix. The mathematics of the update was fine.
30 Sep 2009: For coxph and model.matrix.coxph, re-attach the attributues lost from the X matrix when the intercept is removed, i.e., X <- X[,1]. In particular, some downstream libraries depend on the assign attribute. For predict.coxph remove an earlier edit so that a single variable model + type='terms' returns a matrix, not a vector. This is expected by the termplot() function. It led to a whole lot of changes in the test suite results, though, due to more "matrix" printouts.
4 Sep 2009: Added a model.matrix.coxph and model.frame.coxph methods. The model.matrix.default function ceased to work for coxph models sometime between R 2.9 and 2.9.2 (best guess). This wasn't picked up in the test suite but rather by failure of 3 packages that depend on survival. Also added a test. Update CRAN since this broke other's packages.
20 Aug 2009: One more fix to predict.coxph. It needed to use delete.response(Terms) rather than Terms, so as to not look for (unnecessarily) the response variable when the newdata argment is used. Pointed out by Michael Conklin.
17 Aug 2009: Small bug in survfit.coxph.null pointed out by Frank Harrell. The 'n' component would be missing if the input data included strata, i.e., the initial model had used x=TRUE. He also pointed out the fix.
10 June 2009: Fix an error pointed out by Nick Reich, who was the first to use interval censored data + user defined distribution in survreg, jointly. There was no test case and creating one uncovered several errors (but only for this combination). All the error cases led to catastrophic failure, highlighting the extreme rarity of a user requesting this combination.
2 June 2009: Surv(time1, time2, status, type='interval') would fail for an NA status code. Pointed out by Achim Zeilus.
22 May 2009: Allow single subscripts to rate tables, e.g. survexp[1:10: . Returns a simple vector of values. The str() function does this to print out a short summary. Problem pointed out by Heinz Tuechler.
21 May 2009: Create a test case for factor variables/newdata/predict for coxph and survreg. This led to a set of minor fixes; the code is now in line with the R standard for model functions. One consequence is that model.frame.coxph and model.frame.survreg are no longer needed, so have been removed.
20 May 2009: The manual page for survfit was confusing, since it tries to document both the standard KM (formula method) and the coxph method. I've split them out so that now survfit documents only the basic method and points a user the appropriate specialized page.
1 May 2009: The anova.coxph function was incorrect for models with a strata term. Fixed this, and made chisquare tests the default.
22 April 2009: The coxph code had an override to iter and eps, making both of them more strict for a penalized model. However, the overall default values have changed over time, so that these lines actually decreased accuracy - the opposite of their intent. Removed the lines. Also removed the iter.miss and eps.miss components (on which this check depended) from coxph.control, which makes that function match its documentation.
Issues/decisions in remerging the Mayo and R code: For most of routines, it was easier to start with the Lumley code and add the Therneau fixes. This is because Tom had expanded a lot of partial matches, e.g., fit$coef in the TT code vs fit$coefficients. Routines with substantial changes were, of course, a special case. The most common change is an is.R() construct to choose class vs oldClass.
xtras.R: Move anova.coxph and anova.coxphlist to their own source files. The remainder of the code is R only.
survsum: removed from package
survreg.old: has been removed from the package
survfit.s: Depreciate the "formula with no ~1" option Mayo code for [ allows for reordering curves Separate out the R "basehaz" function as a separate source file
survfit.km.s: The major change of did not get copied into R, so lots of changes. R had "new.time" and Splus 'start.time' for the same argument. Allow them both as synonyms. The output structure also changed: adapt the new one. This is mostly some name changes in the components, removing unneeded redundancies created by a different programmer.
survfit.coxph.s: TMT code finally fixed the "Can't (yet) to case weights" problem. There must have been 10 years been the intent and execution.
survexp.s: Add "bareterms" function from R, which replaces a prior use of terms.inner (in Splus but not R).
survdiff.s: R code had the old (incorrect) expected <- sum(1-offset), since corrected to sum(-log(offset)) .
summary.coxph.s: This was a mess, since Tom and I had independently made the addition of a print.summary.coxph function. Below, TMT means that it was the choice in the Splus code, TL means that it was the choice in R 1. Put the coef=T argument in the print function, not summary (TMT) 2. Change the output's name from coef to coefficients (suggestion of Peter Dalgaard). Also change one column name to Pr(>|z|) for R. 3. Remove last vestiges of a reference to the 'icc' component (TMT) 4. Do not include score, rscore, naive.var in the result (TL) 5. Do include loglik in the result (TMT) 6. Compute the test statistics (loglik, Wald, etc) in the summary function rather than in the print.summary function (TL) 7. Remove the digits option from summary, it belongs in print.summary. (neither)
strata.s: R code added a sep argument, this is ok R changed the character string NA to as.character(NA). Not okay 1. won't work with Splus, 2. This is a label, designed for printing, and so it should be a character string.
residuals.coxph.s: R had added type='partial'. (Which I'm not very partial to, from their statistical properties. But they are legal, and I assume that someone requested them).
print.survfit.s: Rewritten as a part of the general survival rewrite. Created the function 'survmean' which does most of the work, and is shared by print and summary, so that the values from 'print' are now available. Fix the minmin function: min(NULL) gives NA in Splus, which is the right answer for a non-estimable median, but Inf in R. Explicitly deal with this case, and add a bunch of comments. R had the print.rmean option, this has been expanded to a more general rmean option that allows setting the cutoff point. R added a print.n option with 3 choices, my code includes all 3 in the output.
lines.survfit.s: The S version has a new block of code for guessing "firstx" more intellegently when it is missing. (Or, one hopes is is more intellegent!)
coxph.control.s: The R code had tighter tolerances (eps= 1e-9) than Splus (1e-4) and a higher iterationn count (20 vs 10). Set eps to 1e-8 and iter to 15, mostly bending to the world. The tighter iteration is defensible, but I still maintain that a Cox model that takes >10 iterations is not going to finish if you give it 100. The likelihood surface is almost perfectly quadratic near the minimum. (Not true for survreg by the way).
: In Surv, the Mayo code creates NA's out of invalid status values or start,stop pairs, rather than a stop and error message. This is to allow for example coxph(Surv(time1,time2, status).... , subset=(goodlines)) succeed, when "goodlines" is the subset with correct values.
25SepO7: How embarrassing – someone pointed out that I had Dave Harrington's name spelled wrong in the options to survfit.coxph!
9Jul07: In a model with offsets, survreg mistakenly omitted the offset from the returned linear.predictor component.
10May07: Change summary.coxph so that it returns an object of class summary.coxph, and add a print method for that object.
22Jun06: Update match.ratetable, so that more liberal matches are now allowed. For instance, 'F', 'f', 'female', 'fem', 'FEMA', etc are now all considered matches to the dimname "female" in survexp.us.
26Apr06: Fix bug in summary.survfit, pointed out by Bob Treder. With the times option, the value of n.risk would be wrong for "in between" times; e.g., the data had events and/or censoring at times 10, 20,... and we asked for printout at time 15. It should give n.risk at time 20, it was returning the value at time 10. Interestingly, the code had a very careful treatment of this case, along with an example in the comments, and the "the right answer is" part of the comment was wrong! So the code correctly computed an incorrect answer. Added another test case to the test suite, survtest2.
21Apr06: Fix problem in [.survfit, pointed out by Thomas Lumley. If fit <- survfit(Surv(time, status) ~ ph.ecog, lung), then fit[2:1] did not reorder the output correctly. I had never tested putting the subscripts in non-increasing order.
7Feb06: Fix a problem in the coxph iteration (coxfit2.c, coxfit5, agfit3, agfit5, agexact). It will likely never catch anyone again, even if I didn't fix it. In a particular data set, beta overshot and step halving was invoked. During step halving, a loglik happened to occur that was within eps of the prior step's loglik — and the routine decided, erroneously, that it had converged! (A nice quadratic curve, a first guess b1 to the left of the desired max of the curve. The next guess b2 overshot and ends up with a lower loglik, on the right side of the max. Back up to the midpoint of b1 and b2, and this guess, still to the right of the max (still too large) has EXACTLY the same value of y as b1 did, but on the other side of the max from b1. "Last two guesses give the same answer, I'm done" said the routine).
27Sep05: Found and fixed a nasty bug in survfit. When method='fh2' and there were multiple groups I had a subscripting bug, leading to vectors that were supposed to be the same length, but weren't, passed into C. The resulting curves were obviously wrong – survival precipitously drops to zero.
5May05: Add the drop=F arg to one subscripting selection in survfit.coxph. temp <- (matrix(surv$y, ncol=3))[ntime,,drop=F] If you selected only 1 time point (1 row) in the final output, the code would fail. Pointed out by Cindy Crowson.
18Apr05: Bug in survfit.turnbull. The strata variable was not being filled in (number of points per curve). So if multiple curves were generated at once, i.e., with something on the right hand side of ~ in the formula, all the downstream print/plot functions would not work with the result.
8Feb05: Fix small typo in is.ratetable, introduced on 24Nov04: (Today was the first time I added to the standard library, and thus ended up using the non-verbose mode.)
8Feb05: Add the data.frame argument to pyears. This causes the output to contain a dataframe rather than a set of arrays. It is useful for further processing of the data using Poisson regression.
7Feb05: Modified print.ratetable to be more useful. It now tells about the ratetable, rather than printing all of its values.
8Dec04: Fix a small bug in survfit.turnbull. If there are people left censored before the first time point of any other kind (interval, exact, or right censored), the the plotted height of the curve from "rightmost left censoring time" to "leftmost event time", that is the flat tail on the left, was at the wrong height. Added another test to testreg/reliability.s for this.
24Nov04: Change is.ratetable to give longer messages