FastSurvival is built for speed, but a fast estimator is only useful
if it returns the same answer as the established implementation. This
vignette checks the numerical agreement between each FastSurvival
function and a reference, on a real clinical-trial dataset. We use the
gbsg data from the survival package, the German Breast
Cancer Study Group cohort of 686 patients, with recurrence-free survival
time rfstime (in days), the event indicator
status, and the treatment indicator hormon (0
= no hormonal therapy, 1 = hormonal therapy). Throughout we take the
no-hormone arm as the control (control = 0) and report
one-sided tests of treatment benefit (side = 1).
The references are: the survival package for the Kaplan-Meier
estimate, the log-rank test, the Cox hazard ratio, milestone survival,
and the median survival comparison; survRM2 for the restricted mean
survival time; survAH for the average hazard with survival weight; nph
for the weighted log-rank test and the max-combo test; and nphRCT for
the robust modestly-weighted test. The window mean survival time, which
has no package counterpart on CRAN, is checked against a Kaplan-Meier
integral computed from survfit(). The weighted Kaplan-Meier
test is checked against the restricted mean survival time identity and
against its reduction to the two-sample z test without censoring (Pepe
and Fleming, 1989). The same comparisons form the basis of the automated
test suite shipped with the package.
library(survival)
# German Breast Cancer Study Group cohort
str(gbsg[, c("rfstime", "status", "hormon")])
#> 'data.frame': 686 obs. of 3 variables:
#> $ rfstime: int 1838 403 1603 177 1855 842 293 42 564 1093 ...
#> $ status : int 0 1 0 0 0 1 1 0 1 1 ...
#> $ hormon : int 0 0 0 0 1 0 0 1 1 0 ...
table(hormon = gbsg$hormon)
#> hormon
#> 0 1
#> 440 246survfit_fast() evaluates the Kaplan-Meier estimate at a
single time point. We compare against summary(survfit(...))
at the same time (1000 days).
# survfit_fast assumes time-sorted input; the C++ core groups tied times in the
# survival::survfit risk-set convention, so the data only need ordering by time.
ord <- order(gbsg$rfstime)
fast <- survfit_fast(gbsg$rfstime[ord], gbsg$status[ord],
t_eval = 1000, conf.type = "log-log")
fit <- survfit(Surv(rfstime, status) ~ 1, data = gbsg)
ref <- summary(fit, times = 1000)
data.frame(
quantity = c("survival", "std.err"),
fast = c(unclass(fast)["surv"], unclass(fast)["std.err"]),
survival = c(ref$surv, ref$std.err),
row.names = NULL
)
#> quantity fast survival
#> 1 survival 0.65775040 0.65775040
#> 2 std.err 0.01905146 0.01905146survdiff_fast() with side = 1 returns the
signed Z-score of the log-rank test (Mantel, 1966). Its square is the
chi-square statistic from survdiff().
survdiff_fast() also computes Fleming-Harrington G(rho,
gamma) weighted log-rank tests through the weight = "fh"
argument. The nph
package provides logrank.test() with the same family, so we
compare the chi-square statistic (the squared one-sided Z) across
several weight choices.
fh_grid <- data.frame(rho = c(0, 1, 0, 1), gamma = c(1, 0, 0, 1))
do.call(rbind, lapply(seq_len(nrow(fh_grid)), function(i) {
r <- fh_grid$rho[i]
g <- fh_grid$gamma[i]
fast <- as.numeric(survdiff_fast(gbsg$rfstime, gbsg$status, gbsg$hormon,
control = 0, side = 1,
weight = "fh", rho = r, gamma = g))
nph_chisq <- nph::logrank.test(gbsg$rfstime, gbsg$status, gbsg$hormon,
rho = r, gamma = g)$test$Chisq
data.frame(rho = r, gamma = g, fast = fast^2, nph = nph_chisq)
}))
#> rho gamma fast nph
#> 1 0 1 5.110660 5.110660
#> 2 1 0 8.713791 8.713791
#> 3 0 0 8.564781 8.564781
#> 4 1 1 5.881310 5.881310The two implementations agree to numerical precision across the
weight family, including the ordinary log-rank test recovered at
rho = 0, gamma = 0.
coxph_fast() returns the maximum partial likelihood
estimate of the Cox model (Cox, 1972), found by Newton-Raphson iteration
with the starting value, convergence criterion, and maximum number of
iterations of coxph(), with the Efron (1977) handling of
tied event times by default and the Breslow (1974) handling on request.
We compare the log hazard ratio and its standard error with
coxph() for each handling, without and with stratification
by tumor grade (a separate baseline hazard in each stratum). The
reference uses timefix = FALSE, so that, as in
coxph_fast(), only exactly equal times are tied.
cox_ctrl <- coxph.control(timefix = FALSE)
cox_rows <- do.call(rbind, lapply(c("efron", "breslow"), function(tie) {
do.call(rbind, lapply(c(FALSE, TRUE), function(strat) {
fast <- coxph_fast(gbsg$rfstime, gbsg$status, gbsg$hormon, control = 0,
strata = if (strat) gbsg$grade, ties = tie)
ref <- if (strat) {
coxph(Surv(rfstime, status) ~ hormon + strata(grade), data = gbsg,
ties = tie, control = cox_ctrl)
} else {
coxph(Surv(rfstime, status) ~ hormon, data = gbsg, ties = tie,
control = cox_ctrl)
}
data.frame(ties = tie, strata = strat,
coef_fast = unname(fast["coef"]),
coef_cox = unname(coef(ref)),
se_fast = unname(fast["se(coef)"]),
se_cox = sqrt(ref$var[1, 1]))
}))
}))
cox_rows
#> ties strata coef_fast coef_cox se_fast se_cox
#> 1 efron FALSE -0.3640099 -0.3640099 0.1250446 0.1250446
#> 2 efron TRUE -0.3397231 -0.3397231 0.1255153 0.1255153
#> 3 breslow FALSE -0.3638988 -0.3638988 0.1250441 0.1250441
#> 4 breslow TRUE -0.3396679 -0.3396679 0.1255131 0.1255131
cox_diff <- max(abs(cox_rows$coef_fast - cox_rows$coef_cox),
abs(cox_rows$se_fast - cox_rows$se_cox))The estimates and standard errors agree to rounding error; the largest absolute difference is 2.2e-15.
rmst_fast() integrates the Kaplan-Meier survival curve
up to a horizon (Royston and Parmar, 2013). We compare the two-group
RMST difference against survRM2::rmst2() at 1000 days.
library(survRM2)
tau <- 1000
fast_rmst <- rmst_fast(gbsg$rfstime, gbsg$status, gbsg$hormon,
control = 0, tau = tau, side = 1)
ref_rmst <- rmst2(time = gbsg$rfstime, status = gbsg$status,
arm = gbsg$hormon, tau = tau)
c(fast = unclass(fast_rmst)["diff"],
survRM2 = ref_rmst$unadjusted.result[1, 1])
#> fast.diff survRM2
#> 51.14447 51.14447wmst_fast() integrates the Kaplan-Meier curve between a
lower and an upper window limit, generalizing the restricted mean
survival time (Paukner and Chappell, 2021). CRAN has no dedicated WMST
package, so we validate the per-group windowed area directly against a
Kaplan-Meier step-function integral computed from survfit()
over the same window. A full comparison against the
survWMST package, which is distributed on GitHub, is
provided in tools/compare_wmst_survwmst.R of the package’s
GitHub repository.
tau1 <- 200
tau2 <- 1000
fast_wm <- wmst_fast(gbsg$rfstime, gbsg$status, gbsg$hormon,
control = 0, side = 1, tau1 = tau1, tau2 = tau2)
# Integral of the Kaplan-Meier step function over [lo, hi] for one group.
km_window_area <- function(gi, lo, hi) {
sf <- survfit(Surv(rfstime, status) ~ 1, data = gbsg[gbsg$hormon == gi, ])
tk <- c(0, sf$time)
sk <- c(1, sf$surv)
brk <- sort(unique(c(lo, hi, sf$time[sf$time > lo & sf$time < hi])))
area <- 0
for (b in seq_len(length(brk) - 1L)) {
u <- brk[b]
su <- sk[max(which(tk <= u))]
area <- area + su * (brk[b + 1L] - u)
}
area
}
data.frame(
group = c("control", "treatment"),
fast = unclass(fast_wm)[c("wmst.control", "wmst.treatment")],
survfit = c(km_window_area(0, tau1, tau2), km_window_area(1, tau1, tau2)),
row.names = NULL
)
#> group fast survfit
#> 1 control 628.0064 628.0064
#> 2 treatment 678.3393 678.3393wkm_fast() computes the weighted Kaplan-Meier
(Pepe-Fleming) test, the weighted integral of the difference between the
two Kaplan-Meier curves. With the constant weight the weighted
difference reduces exactly to the difference in restricted mean survival
time over the observed range, which gives a clean internal check against
rmst_fast() evaluated at the largest observed time.
tmax <- max(gbsg$rfstime)
fast_wk <- wkm_fast(gbsg$rfstime, gbsg$status, gbsg$hormon,
control = 0, side = 1, weight = "constant")
# tau = tmax extends past the shorter group's follow-up on purpose, to cover
# the whole observed range as the constant-weight test does; rmst_fast()
# warns about this, so the warning is suppressed here.
ref_rm <- suppressWarnings(
rmst_fast(gbsg$rfstime, gbsg$status, gbsg$hormon,
control = 0, tau = tmax, side = 1))
c(wkm.constant = unclass(fast_wk)["wdiff"], rmst = unclass(ref_rm)["diff"])
#> wkm.constant.wdiff rmst.diff
#> 256.7207 256.7207The variance is the pooled estimator of Pepe and Fleming (1989, p. 501). Without censoring the Pepe-Fleming weight is 1, the weighted difference is the difference in mean survival times, and this variance estimate is the variance of the pooled sample, so the test is the two-sample z test. We check this on the recurrence times of the patients with a recurrence, treated as uncensored data.
rec <- gbsg$status == 1
x_rec <- gbsg$rfstime[rec]
g_rec <- gbsg$hormon[rec]
fast_unc <- wkm_fast(x_rec, rep(1L, length(x_rec)), g_rec,
control = 0, side = 2)
# Two-sample z test with the variance of the pooled sample (divisor n)
v_rec <- mean((x_rec - mean(x_rec))^2)
se_rec <- sqrt(v_rec * (1 / sum(g_rec == 1) + 1 / sum(g_rec == 0)))
z_rec <- (mean(x_rec[g_rec == 1]) - mean(x_rec[g_rec == 0])) / se_rec
c(wkm.z = unname(fast_unc["z"]), z.test = z_rec)
#> wkm.z z.test
#> 1.80783 1.80783The Overall.test() function of the
ComparisonSurv package also reports a weighted Kaplan-Meier
test. Its values differ from those of wkm_fast(), because
it evaluates the censoring weight on each interval at the preceding
observed time and ends the integral at the last time at which the
Kaplan-Meier estimates of survival and censoring of both groups are
positive.
milestone_fast() compares Kaplan-Meier survival between
two groups at a milestone timepoint, here with the complementary log-log
intervals and the MOVER interval of the difference of Tang (2021). The
per-group survival probabilities match those from
survfit().
tstar <- 1000
fast_ms <- milestone_fast(gbsg$rfstime, gbsg$status, gbsg$hormon,
control = 0, tau = tstar, method = "loglog",
side = 1)
fit_g <- survfit(Surv(rfstime, status) ~ hormon, data = gbsg)
ref_g <- summary(fit_g, times = tstar)
data.frame(
group = c("control", "treatment"),
fast = fast_ms$surv[c("control", "treatment")],
survival = ref_g$surv,
row.names = NULL
)
#> group fast survival
#> 1 control 0.6208762 0.6208762
#> 2 treatment 0.7230082 0.7230082medsurv_fast() estimates the Kaplan-Meier median
survival time and, for two groups, their difference. The point estimate
is the Kaplan-Meier median (the first time at which the product-limit
curve reaches 0.5), the same convention as survfit(), so
the per-group medians match those reported by survfit()
exactly.
fast_med <- medsurv_fast(gbsg$rfstime, gbsg$status, gbsg$hormon,
control = 0, side = 1)
fit_med <- survfit(Surv(rfstime, status) ~ hormon, data = gbsg)
med_ref <- summary(fit_med)$table[, "median"]
data.frame(
group = c("control", "treatment"),
fast = unclass(fast_med)[c("median.control", "median.treatment")],
survival = c(med_ref[1], med_ref[2]),
row.names = NULL
)
#> group fast survival
#> 1 control 1528 1528
#> 2 treatment 2018 2018We use survfit() as the reference here, rather than
nph::nphparams(), because the two define the median on
different survival curves: survfit() and
medsurv_fast() use the Kaplan-Meier (product-limit) median,
whereas nph::nphparams() reads the median off the
Nelson-Aalen curve, S(t) = exp(-H(t)), which lies above the Kaplan-Meier
curve and so can cross 0.5 at a later time on tied data. The standard
error of medsurv_fast(), from Greenwood’s variance and a
kernel estimate of the hazard at the median (see
?medsurv_fast), is checked against values computed
independently in the package test suite.
ahsw_fast() computes the average hazard with survival
weight of Uno and Horiguchi. We compare the per-group average hazard
against survAH::ah2().
library(survAH)
tau <- 1000
fast_ah <- ahsw_fast(gbsg$rfstime, gbsg$status, gbsg$hormon,
control = 0, tau = tau, side = 1)
ref_ah <- ah2(time = gbsg$rfstime, status = gbsg$status,
arm = gbsg$hormon, tau = tau)
data.frame(
quantity = c("AH (control)", "AH (treatment)"),
fast = unclass(fast_ah)[c("ah.ctrl", "ah.trt")],
survAH = c(ref_ah$ah["AH (arm0)", "Est."],
ref_ah$ah["AH (arm1)", "Est."]),
row.names = NULL
)
#> quantity fast survAH
#> 1 AH (control) 0.0004585857 0.0004585857
#> 2 AH (treatment) 0.0003155277 0.0003155277The average hazard is the ratio of the cumulative event probability
to the restricted mean survival time, so for the gbsg data,
where follow-up is measured in days, the values are on the order of
1e-04 per day in both groups.
ahr_fast() computes the Kalbfleisch-Prentice average
hazard ratio between two groups over a restricted interval. The
reference implementation is ahrKM() from the AHR package,
but that package has been archived on CRAN, so here we check the point
estimates against a direct survival-based computation: each group’s
Kaplan-Meier curve from survfit(), integrated to form the
group shares of the total hazard.
tau <- 1000
fast_ahr <- ahr_fast(gbsg$rfstime, gbsg$status, gbsg$hormon,
control = 0, tau = tau, side = 1)
g <- gbsg$hormon
ev <- c(gbsg$rfstime[g == 0 & gbsg$status == 1],
gbsg$rfstime[g == 1 & gbsg$status == 1])
grid <- sort(unique(c(0, ev[ev <= tau], tau)))
km_on_grid <- function(gi) {
sf <- survfit(Surv(rfstime, status) ~ 1, data = gbsg[g == gi, ])
approxfun(sf$time, sf$surv, method = "constant",
yleft = 1, rule = 2, f = 0)(grid)
}
S0 <- km_on_grid(0)
S1 <- km_on_grid(1)
m <- length(grid)
dS0 <- S0 - c(1, S0[-m])
GL <- S0[m] * S1[m]
ref_theta_ctrl <- -sum(S1 * dS0) / (1 - GL)
data.frame(
quantity = c("theta (control)", "theta (treatment)", "AHR"),
fast = c(fast_ahr$theta[[1]], fast_ahr$theta[[2]], fast_ahr$ahr),
reference = c(ref_theta_ctrl, 1 - ref_theta_ctrl,
(1 - ref_theta_ctrl) / ref_theta_ctrl),
row.names = NULL
)
#> quantity fast reference
#> 1 theta (control) 0.5974108 0.5974108
#> 2 theta (treatment) 0.4025892 0.4025892
#> 3 AHR 0.6738899 0.6738899The point estimates match the survival-based reference to numerical
precision. The average hazard ratio was additionally cross-checked
against the archived AHR package, the reference implementation used by
Dormuth et al. (2024): the two group shares, the average hazard ratio,
the variances, and both test statistics agree to within about 1e-14.
That external comparison is reproducible with the
tools/compare_ahr_ahrKM.R script of the package’s GitHub
repository.
maxcombo_fast() computes the max-combo test, the most
extreme of a set of Fleming-Harrington weighted log-rank statistics
(Karrison, 2016). The nph package provides
logrank.maxtest(), whose default weight set is FH(0, 0),
FH(0, 1), and FH(1, 0). We request the same three weights from
maxcombo_fast() and compare the component Z-scores. The
individual Z-scores are kept in the z attribute, signed so
that a negative value indicates benefit for the treatment group under
the package convention. Because nph orients the contrast in
the opposite direction, the signs are mirrored, so we compare absolute
values.
rho <- c(0, 0, 1)
gamma <- c(0, 1, 0)
fast_mc <- maxcombo_fast(gbsg$rfstime, gbsg$status, gbsg$hormon,
control = 0, side = 1, rho = rho, gamma = gamma)
nph_mc <- nph::logrank.maxtest(gbsg$rfstime, gbsg$status, gbsg$hormon)
data.frame(
weight = c("FH(0,0)", "FH(0,1)", "FH(1,0)"),
fast = abs(attr(fast_mc, "z")),
nph = abs(nph_mc$tests$z),
row.names = NULL
)
#> weight fast nph
#> 1 FH(0,0) 2.926565 2.926565
#> 2 FH(0,1) 2.260677 2.260677
#> 3 FH(1,0) 2.951913 2.951913The component statistics agree in absolute value. The final max-combo
statistic is the largest of these, which for the one-sided test is the
most extreme component Z. The p-values are not compared here, because
nph::logrank.maxtest() reports two-sided p-values and the
call above is one-sided.
rmw_fast() computes the robust modestly-weighted (rMW)
test of Magirr and Öhrn, the maximum of the standard log-rank statistic
and a single modestly-weighted log-rank statistic. Each component is a
weighted log-rank test, so we check the two component Z-scores against
nphRCT, the
implementation used by the method’s authors. The modestly-weighted
component uses the survival-threshold parameterization, so
s_star = 1 recovers the standard log-rank test and
s_star = 0.5 gives the modestly-weighted component. As with
the max-combo test the package orients a negative Z toward treatment
benefit while nphRCT uses the opposite sign, so we compare
absolute values.
gbsg_df <- data.frame(
time = gbsg$rfstime,
event = gbsg$status,
arm = factor(ifelse(gbsg$hormon == 0, "control", "experimental"),
levels = c("control", "experimental"))
)
fit_rmw <- rmw_fast(gbsg_df$time, gbsg_df$event, gbsg_df$arm,
control = "control", side = 1, s_star = 0.5)
z_lr_nph <- nphRCT::wlrt(Surv(time, event) ~ arm, data = gbsg_df,
method = "mw", s_star = 1)$z
z_mw_nph <- nphRCT::wlrt(Surv(time, event) ~ arm, data = gbsg_df,
method = "mw", s_star = 0.5)$z
data.frame(
component = c("log-rank (s_star = 1)", "modestly-weighted (s_star = 0.5)"),
fast = abs(attr(fit_rmw, "z")),
nphRCT = abs(c(z_lr_nph, z_mw_nph)),
row.names = NULL
)
#> component fast nphRCT
#> 1 log-rank (s_star = 1) 2.926565 2.926565
#> 2 modestly-weighted (s_star = 0.5) 2.773749 2.773749The component statistics agree in absolute value. The null
correlation of the two components, reported in the corr
attribute, is checked against an independent implementation in the
package test suite. The combined statistic and one-sided p-value then
follow from the bivariate normal distribution of the two components,
evaluated with mvtnorm::pmvnorm.
The simulation layer determines the calendar time of each analysis
with cutoff_fast(), which combines a target number of
events, a planned calendar time, a maximum calendar time, a minimum time
after the previous look, and a minimum follow-up after a given number of
enrolled subjects. The same rule is implemented one simulated trial at a
time by get_analysis_date() of the simtrial package.
We compare the two on 50 simulated trials for a look at “150 events, but
not before month 18 and not before 6 months after the 250th enrolled
subject, and in any case by month 30”.
sim_c <- simdata_fast(nsim = 50, n = c(150, 150), a.time = c(0, 12),
a.rate = 300 / 12, e.median = list(12, 16),
d.hazard = 0.01, seed = 2026)
cut_c <- sim_c |>
cutoff_fast(event.looks = 150, time.looks = 18, max.time = 30,
min.enrolled = 250, min.followup = 6)
ref_c <- vapply(seq_len(50), function(s) {
d <- sim_c[sim_c$sim == s, ]
x <- data.frame(enroll_time = d$accrual_time,
cte = d$accrual_time + d$tte,
fail = d$event, stratum = "All")
simtrial::get_analysis_date(x, planned_calendar_time = 18,
target_event_overall = 150,
max_extension_for_target_event = 30,
min_n_overall = 250, min_followup = 6)
}, numeric(1))
c(max_abs_difference = max(abs(cut_c[, 1] - ref_c)))
#> max_abs_difference
#> 0The cutoffs agree exactly. The two functions differ only when an
event target is never met in the simulated data, which
cutoff_fast() reports as an unreached look rather than as
the time of the last observed event.
switch_fast() changes only the outcomes after each
subject’s switch. For an exponential survival time with hazard
lambda, a switch at time s after accrual that
multiplies the remaining time by f gives the survival
function exp(-lambda t) for t <= s and
exp(-lambda s - lambda (t - s) / f) for
t > s. We let every control subject still alive at
calendar month 12 switch with f = 2, so that
s = 12 - a for a subject accrued at time a,
and compare the empirical survival of the control group with the average
of this survival function over the accrual times.
lam <- log(2) / 12
sw <- simdata_fast(nsim = 200, n = c(250, 250), a.time = c(0, 6),
a.prop = 1, e.hazard = list(lam, lam / 1.5),
seed = 2027) |>
switch_fast(group = 1, when = "cutoff", cutoff = rep(12, 200),
aft.factor = 2)
ctl <- sw$group == 1
s_sw <- 12 - sw$accrual_time[ctl]
t_grid <- c(6, 12, 18, 24, 36)
data.frame(
t = t_grid,
empirical = round(sapply(t_grid, function(t) mean(sw$surv_time[ctl] > t)), 4),
analytic = round(sapply(t_grid, function(t) {
mean(ifelse(t <= s_sw, exp(-lam * t),
exp(-lam * s_sw - lam * (t - s_sw) / 2)))
}), 4)
)
#> t empirical analytic
#> 1 6 0.7052 0.7071
#> 2 12 0.5456 0.5460
#> 3 18 0.4594 0.4591
#> 4 24 0.3860 0.3861
#> 5 36 0.2728 0.2730The empirical survival matches the analytic survival function to Monte Carlo error (the standard error is at most about 0.002 with 50,000 control subjects). The vignette on treatment switching also checks the switching model with the rank-preserving structural failure time estimator of the rpsftm package.
Across all functions the FastSurvival results reproduce the reference
values on the gbsg data. The point estimates and test
statistics agree to numerical precision for the Kaplan-Meier, log-rank,
weighted log-rank, RMST, milestone, median, and average-hazard
quantities; the window mean survival time matches a Kaplan-Meier
integral and the weighted Kaplan-Meier statistic matches the RMST
identity and the two-sample z test without censoring; the average hazard
ratio matches a survival-based reference; and the Cox hazard ratio and
its standard error, with or without strata and with either handling of
tied event times, agree with coxph() to rounding error. In
the simulation layer, the analysis cutoffs agree with simtrial and the
switched survival times follow their analytic distribution. This
agreement is verified continuously by the package test suite.
Kaplan, E. L., & Meier, P. (1958). Nonparametric estimation from incomplete observations. Journal of the American Statistical Association, 53(282), 457-481.
Mantel, N. (1966). Evaluation of survival data and two new rank order statistics arising in its consideration. Cancer Chemotherapy Reports, 50(3), 163-170.
Fleming, T. R., & Harrington, D. P. (1991). Counting Processes and Survival Analysis. New York: John Wiley & Sons.
Cox, D. R. (1972). Regression models and life-tables. Journal of the Royal Statistical Society. Series B (Methodological), 34(2), 187-202.
Breslow, N. (1974). Covariance analysis of censored survival data. Biometrics, 30(1), 89-99.
Efron, B. (1977). The efficiency of Cox’s likelihood function for censored data. Journal of the American Statistical Association, 72(359), 557-565.
Royston, P., & Parmar, M. K. B. (2013). Restricted mean survival time: an alternative to the hazard ratio for the design and analysis of randomized trials with a time-to-event outcome. BMC Medical Research Methodology, 13, 152.
Paukner, M., & Chappell, R. (2021). Window mean survival time. Statistics in Medicine, 40(25), 5521-5533.
Pepe, M. S., & Fleming, T. R. (1989). Weighted Kaplan-Meier statistics: a class of distance tests for censored survival data. Biometrics, 45(2), 497-507.
Pepe, M. S., & Fleming, T. R. (1991). Weighted Kaplan-Meier statistics: large sample and optimality considerations. Journal of the Royal Statistical Society. Series B (Methodological), 53(2), 341-352.
Tang, Y. (2021). Some new confidence intervals for Kaplan-Meier based estimators from one and two sample survival data. Statistics in Medicine, 40(23), 4961-4976.
Uno, H., Claggett, B., Tian, L., et al. (2014). Moving beyond the hazard ratio in quantifying the between-group difference in survival analysis. Journal of Clinical Oncology, 32(22), 2380-2385.
Uno, H., & Horiguchi, M. (2023). Ratio and difference of average hazard with survival weight: new measures to quantify survival benefit of new therapy. Statistics in Medicine, 42(7), 936-952.
Kalbfleisch, J. D., & Prentice, R. L. (1981). Estimation of the average hazard ratio. Biometrika, 68(1), 105-112.
Dormuth, I., Pauly, M., Rauch, G., & Herrmann, C. (2024). Sample size calculation under nonproportional hazards using average hazard ratios. Biometrical Journal, 66(6), e202300271.
Karrison, T. G. (2016). Versatile tests for comparing survival curves based on weighted log-rank statistics. The Stata Journal, 16(3), 678-690.
Magirr, D., & Burman, C.-F. (2019). Modestly weighted logrank tests. Statistics in Medicine, 38(20), 3782-3790.
Magirr, D., & Öhrn, F. (2026). Robust modestly weighted log-rank tests. Pharmaceutical Statistics, 25(1), e70066.