Skip to contents

Changelog

Source: NEWS.md

frmtmb.survival 0.2.2

  • The package code does not use the global assignment operator <<-, which rOpenSci’s pkgcheck refuses. bhaz()’s argument reader passes the arguments it has read from one step to the next. The results do not change.

frmtmb.survival 0.2.1

Other changes

  • frm_compat() by nl: cox and royston_parmar are “works”. A body with a level of its own reaches the linear maximum.

frmtmb.survival 0.2.0

Needs frmtmb 0.75.0, whose frmtmb_register_aterm(option =) this package calls when it loads.

New features

  • brms’s bhaz() addition term. time | cens(c) + bhaz(df = 6) ~ x with family = cox() is the model of cox(df = 6): the two fits have identical() log likelihoods, estimates and covariance matrices. bhaz() takes brms 2.23.0’s arguments df, degree, intercept, knots and Boundary.knots, and sets the argument of the same name of cox(). Its basis equals the Zbhaz and Zcbhaz of brms’s standata() to 6.7e-16 of their size or less (the largest difference over the largest entry). At each of 1000 posterior draws of three brms fits, the log likelihood equals the sum of brms’s log_lik() to 1.5e-16 of its size.

  • cox(knots =, Boundary.knots =) set the internal and the boundary knots of the baseline basis, which only bhaz() could set in brms. With knots, df is length(knots) + degree + intercept and need not be given.

Where bhaz() differs from brms

  • An argument that bhaz() and cox() both set must have one value. bhaz(df = 6) with family = cox(df = 5) stops and names the two. df and knots that disagree stop too, in either place: brms ignores df when knots is given. The basis is checked after the two are merged, so cox(df = 3) with bhaz(intercept = FALSE) fits.

  • bhaz(gr = g), brms’s stratified Cox model with one baseline per level of g, stops. The message says that cox() has one baseline. It tells you to fit each stratum as a model of its own, or to add g to the formula. An unnamed first argument, bhaz(g), is gr as in brms; bhaz(6), which brms reads as one stratum named 6, stops and suggests bhaz(df = 6).

  • A shortened argument name that brms reads is read the same way: bhaz(d = 6) is df = 6, kn = is knots, deg = is degree. A name that brms drops without a message stops and names the argument that was probably meant: bhaz(int = FALSE) and bhaz(inter = FALSE) (brms keeps the intercept), bhaz(Bound = c(0, 30)) (brms keeps the default boundary knots). Other names brms does not document, such as bhaz(foo = 1), stop too. warn.outside is taken and has no effect, as in brms.

  • Boundary knots that do not cover every observed time stop. brms extrapolates the basis past them. With Boundary.knots = c(0.5, 10) on times from 0.0044 to 14.4, 62 rows lie below the lower knot and 1 above. At equal weights, brms’s cumulative baseline hazard reaches -1.08 there, so the survivor function exceeds one. The hazard is negative for some weights.

Other changes

  • cox(bhaz = list(df = 6)), the spelling of brms before 2.22.0, stops with the spellings to use instead. Before, it stopped with “unused argument”. brms 2.23.0’s cox() has no bhaz argument either.

  • A df below degree + intercept is still refused when the model is fitted, and the message now gives the right minimum without an intercept: “df must be at least degree (3) without an intercept”.

frmtmb.survival 0.1.0

The first release. It holds the survival families that were in two other packages, so that each has one home. Needs frmtmb 0.74.0, whose frmtmb_family() takes simplex_pars = and aterm_data =.

Breaking changes

  • cox() and cox_baseline() moved here from frmtmb. Load this package with library(frmtmb.survival), which attaches frmtmb as well. family = "cox" and brmsfamily("cox") reach this package’s cox() once it is loaded, and without it stop with a message that names it. A cox() fit is identical() to one from frmtmb 0.73.0, in its log likelihood, estimates and standard errors.

  • royston_parmar() and rp_floored() moved here from frmtmb.spline, which is retired, with the article “How to fit a flexible parametric survival model”. A fit is identical() to one from frmtmb.spline 0.10.2. The curve functions that read a fit, frm_curve(), frm_curve_deriv() and frm_curve_feature(), are in frmtmb now. They still refuse a fit whose cumulative hazard is not monotone, through the family’s post$curve_check hook.

  • The refusals of these functions carry the class frmtmb_survival_error. They carried frmtmb_error (cox()) and frmtmb_spline_error (royston_parmar()) before. All of them still carry frmtmb_error.

  • frmtmb.spline’s NEWS, the history of royston_parmar() before the move, is the article “History of frmtmb.spline” on this package’s site.