FastSurvival

• The results of the estimation and testing functions have the additional class "fastsurv". Its as.data.frame() method returns the estimates and tests of any result, including the lists returned by milestone_fast() and ahr_fast(), as rows of a data frame with the columns term, estimate, std.error, conf.low, conf.high, statistic, and p.value, so that the results of a loop can be stacked with rbind(). Its coef() and confint() methods return the estimates and their confidence limits. Ratios are reported on the log scale. The results are otherwise unchanged. Code that compares class() of a result with a single class name should use inherits() instead, and as.data.frame() and data.frame() of a result, and functions that call them such as write.csv(), now return the table described in ?as.data.frame.fastsurv; with R 4.3.0 or later they returned the elements of a numeric result as a one-column data frame, which as.data.frame(unclass(x)) still gives.
• plot.kmcurve_fast() can print the numbers at risk below the survival panel (risk_table and risk_times) and draw the pointwise confidence limits as a shaded band (conf.style = "band"). as.data.frame() returns the Kaplan-Meier curves of a kmcurve_fast object as a data frame, with the numbers at risk, events, and censored times at each observed time.
• simsummary_fast() results have a plot() method, which draws the cumulative probabilities of stopping for efficacy and for futility by look, or the stopping probabilities at each look, for every arm and population.

Bug fixes

• print.simsummary_fast() failed with an error for a selection of the columns of a simsummary_fast() result, which keeps the class but not the boundary settings, and for a selection of rows without any look row. Such subsets are now printed as ordinary data frames. A selection of rows that keeps only some of the looks, or rows in a different order, paired the boundaries and information fractions with the wrong looks or failed; it is now also printed as an ordinary data frame, and the report is printed only when every block keeps all its looks in their original order. Results combined by rbind(), which keep the boundary settings of the first result only, are printed as an ordinary data frame when the rows of a block differ from those of the first result (for example rows of a population summarized with other boundaries, even under the same population name), instead of pairing them with the boundaries of the first. The boundary columns are labeled Efficacy Bound and Futility Bound, because the statistic need not be a Z-score.
• The median survival time of medsurv_fast() and of the "medsurv" statistic of analysis_fast() was infinite when the Kaplan-Meier estimate of a group stayed at 0.5 up to an infinite observed time (subjects with tte = Inf at a look that is not reached), and with method = "nph" the test then gave z = Inf and p = 0. The midpoint of such a stretch is not defined, and the median and the statistics that depend on it are now NA. The median printed by print.kmcurve_fast() follows the same rule.
• simdata_fast() truncated a group size that is a whole number up to rounding error, such as 90 * 0.7 (62.99999999999999), to the integer below, so the group had one subject fewer, and with fixed.alloc = TRUE and subgroups the C++ kernel wrote past the end of a buffer. The sizes are now rounded, and the kernels check that the fixed subgroup counts add up to the group size. The data simulated for such sizes change. The nsim argument of simdata_fast() and the stream argument of simdata_fast() and switch_fast(), which were truncated in the same way, are now rounded too.
• The modestly-weighted log-rank test (weight = "mwlrt" in survdiff_fast() and analysis_fast()) capped the weights at 1, which gives the ordinary log-rank test, when the pooled Kaplan-Meier estimate reached 0 before t_star (within a stratum for the stratified test). The weights 1 / S(t-) are now not capped in this case, as the definition 1 / max(S(t-), S(t_star-)) and nphRCT imply.
• analysis_fast() and cutoff_fast() accepted negative event times, which the analysis kernel sorted after all positive times, and analysis_fast() placed a negative zero (-0) last, unlike the stand-alone functions. Negative times are now an error, and -0 is treated as 0.
• An event indicator given as a factor with the levels "0" and "1" passed the checks of most functions and was then replaced by its integer codes 1 and 2, which exchanges events and censored observations (for example in milestone_fast(), ahr_fast(), analysis_fast(), and cutoff_fast()). Factor and character event indicators are now an error. medsurv_fast(), wmst_fast(), and wkm_fast() now also reject non-numeric and negative times, which they sorted as text or accepted. Times of class difftime, which is.numeric() does not count as numeric, are rejected by these functions and in the tte column of analysis_fast() and cutoff_fast(); convert them with as.numeric() in the intended unit.
• wmst_fast() returned infinite or NaN values when tau2 was not supplied and each group had an infinite observed time. tau2 must now be finite, and an error asks for it when its default is infinite.
• simdata_fast() with the illness-death model and without prevalence did not check switch.prop (a percentage such as 40 made every subject switch) or the length of per-group lists (a third element was ignored). Both are now checked as with prevalence. A multi-arm n with the illness-death model now gives the error message intended for it.
• gen_scenario_fast() kept a shared e.hazard when a scenario gave e.median (and likewise for the other hazard and median pairs), so the printed and plotted scenario used the shared hazard, and simdata_fast() failed with the merged arguments. A scenario value now replaces the other member of the pair in shared.
• analysis_fast() reported NA for the per-group average hazards of "ahsw" whenever the contrasts were not defined, for example for a group without events up to tau. Finite average hazards are now reported, as in ahsw_fast().
• analysis_fast() combined several strata columns through pasted labels, so values containing "." could merge two strata. The columns are now combined through their integer codes.
• Clearer errors: analysis_fast() for an empty mc.rho, a conf.level that is NA or not a single value, and a by.subgroup that is not TRUE or FALSE, and it names event.looks or time.looks in the message about non-positive looks; simdata_fast() names the breakpoint argument (d.time, h01.time, and so on) in the messages about piecewise hazards and requires a whole-number nsim; kmcurve_fast() rejects missing group values.
• The weighted Kaplan-Meier test (wkm_fast() and the "wkm" statistic of analysis_fast()) returned NaN or an infinite weighted difference when an observed time was infinite, for example for a subject without a finite event or dropout time (a cure fraction without dropout) at a look of analysis_fast() that is not reached. The integral over the observed range is not defined in this case, and all the statistics are now NA.

Documentation

• maxcombo_fast() and analysis_fast() now describe the precision of the max-combo p-value correctly. The weights of G(0,1) and G(1,0) add up to the weight of G(0,0), so the correlation matrix of the default weights is singular, and the quasi-Monte-Carlo GenzBretz integral converges slowly: with the default maxpts the error of the p-value is of the order of 1e-4 rather than abseps. The statement that abseps can be loosened with negligible effect is removed.
• The description of mc.alpha in analysis_fast(), the README, and three vignettes is corrected. The decisions that the Bonferroni bounds settle are those of the exact p-values. The other p-values are integrated as without mc.alpha, but the GenzBretz integration uses random numbers, so near the level a decision can differ from that of a run without mc.alpha, as it can between two runs with different seeds. The statement that the decisions are the same as with every p-value integrated is removed.
• The speed-comparison vignette now gives the ratio of the time per simulated trial with simtrial or TrialSimulator to that with FastSurvival as about two orders of magnitude, because the ratio varies between runs, and states that the one-second timing of 10,000 trials comes from a separate benchmark with a delayed treatment effect.
• The page ranges of Cox (1972) and Gehan (1965) follow the publishers' records (187-202 and 203-224), and the mrct-regional-consistency vignette states the result of Homma (2024) for binary outcomes more precisely. The group-sequential-design vignette attributes its planning assumptions to the trial protocol.
• The reference to an unpublished manuscript on the Pike-Halley Estimator is removed. The documentation of coxph_fast() describes the computation.
• The documentation of coxph_fast() and the README describe the correction as a Halley correction expanded to second order, and no longer cite Berry, Kitchin, and Mock (1991) for the statement that the error of the Pike anchor does not vanish with the sample size (their simulation has a fixed sample size). The README no longer quotes a result on the pharmacoSmoking data that has no source in the package, and describes survfit_fast() as more than an order of magnitude faster than survfit() plus summary().
• milestone_fast() defines its three test statistics. The "mover" statistic is obtained by inverting the interval at conf.level, so the statistic and its p-value depend on conf.level (a code comment said the opposite), and the "loglog" statistic is negative when treatment is better. The print methods of coxph_fast(), rmst_fast(), milestone_fast(), medsurv_fast(), ahr_fast(), wkm_fast(), and wmst_fast() label the p-value of a one-sided test Pr(<z) or Pr(>z), and their documentation no longer calls the p-value two-sided.
• analysis_fast() states that no warning is given when tau, t.eval, or wmst.tau2 exceeds the follow-up of a group at a look (the stand-alone functions warn), that conf.level applies to every confidence interval, that maxcombo.stat and rmw.stat depend on side, that all events at the calendar time of the target event are included, and when the stratified test reduces to the ordinary one within a subgroup.
• simdata_fast() and switch_fast() describe the random numbers used when seed is not supplied, which set.seed() does not control. The package now requires dqrng 0.4.0 or later, whose default generator is the Xoroshiro128++ generator described in the documentation. simdata_fast() also states the rounding rule of fixed.alloc = TRUE and cites Fleischer, Gaschler-Markefski, and Bluhmki (2009) for the maximal-independence model.
• wkm_fast() explains that the constant and square-root weights can be biased when follow-up differs between the groups, because the curve of the group with shorter follow-up is carried forward.
• Vignettes: compare-logrank-rmst says that the sample size comes from the Lachin and Foulkes method of gsDesign::nSurv() (not the Schoenfeld formula); group-sequential-design says that the trial planned to enroll 482 patients, gives the section of the protocol, and states the type I error for overall survival; correlated-pfs-os-gsd requires a median overall survival longer than the median progression-free survival; investigate-freidlin-and-korn attributes the parameters of its strong null scenario to Fig. A1 of Freidlin and Korn (2019), which the simtrial documentation also uses; and the computing times are described as they were measured.
• Smaller corrections: the variances of the two group shares of ahr_fast() are equal algebraically (code comments); the example of a zero variance in survdiff_fast() and rmw_fast(); "orthant" instead of "half-space" for the one-sided max-combo and rMW regions; the left-continuous form of the modestly-weighted weight; "Kalbfleisch and Prentice" and the spelling "Öhrn"; the print-edition ISBN of Collett (2014) in DESCRIPTION; the issue numbers of Gehan (1965) and Tarone and Ware (1977); and the formatting of several references.

Tests

• New tests cover the formula interface of every function (compared with the vector interface), the as.data.frame(), coef(), and confint() methods, the numbers at risk and the Kaplan-Meier table of a kmcurve_fast object (compared with summary() of a survival::survfit() object and with a hand calculation), and the plot methods.
• The one-sided max-combo p-values with three and four weights, whose correlation matrices are singular, are compared with an independent numerical integral that uses this structure.
• New tests cover the printing of subsets of simsummary_fast() results and the weighted Kaplan-Meier test with infinite times. The tests of print.simsummary_fast() are moved to their own file.
• A test checks that the switching proportion, the hazard after switching, and the dropout hazard of the illness-death model of simdata_fast() can differ by subgroup.
• New tests cover each bug fix above. The comparisons of the max-combo p-values with the pure-R reference set the seed and use an absolute tolerance, so that they no longer depend on the state left by earlier tests. The rMW p-value is compared with an independent one-dimensional integral. The data generated without stream are compared with data printed by FastSurvival 1.0.0, before stream existed. A test no longer turns an error of medsurv_fast() into a skip, and the error expectations give the expected messages.