Calibrating futility spending to conditional power

library(gsDesign)
library(lt)
library(ggplot2)

gsCPFutilitySpending() selects lower spending parameters so that conditional power (CP) at selected interim futility bounds matches requested targets. It provides a way to specify futility in terms of the probability of future efficacy rejection. Each candidate is rebuilt with gsDesign(), including the information needed to retain the design’s target unconditional power. Timing and efficacy spending remain fixed; maximum information need not remain fixed.

The boundary summaries show three perspectives on future efficacy rejection: CP assumes the observed effect at the bound continues, CP H1 assumes the planned alternative effect, and PP averages conditional power over the posterior effect distribution using gsPP(). By default, gsBoundSummary() uses the broad, mildly informative normal prior normalGrid(mu = x$delta / 2, sigma = 10 / sqrt(x$n.fix)) on the standardized effect scale. Its center is halfway between the null (zero) and the planned alternative (x$delta); it is updated using the interim result at the bound. Thus PP incorporates uncertainty about the effect rather than assuming one fixed value. CP H1 and PP provide context in the tables; the examples here primarily calibrate CP. A parallel survival example uses gsPPFutilitySpending() to target PP instead.

A design warning: futility, sample size, and overall power

Targeting high conditional power at a futility boundary can require a much larger maximum sample size than a fixed design with no interims if overall power must remain unchanged. A stringent futility rule stops some trials even when the planned treatment effect is real. Recovering that lost power among trials that continue can be expensive in sample size. An alternative is to accept lower overall power by increasing the design’s beta and recalibrating, rather than insisting on 90% power at any sample-size cost.

An aggressive futility analysis can be viewed as a Phase II proof-of-concept gate embedded in a Phase III confirmatory trial. For perspective, two independent studies with 80% and 90% power have a probability of success in both of 0.80 * 0.90 = 0.72. This is an illustration, not the power formula for an integrated group sequential trial: its interim and final statistics are dependent, and its overall power must be calculated from the joint design. The same multiplication applies if the second probability is explicitly conditional on passing the first gate.

For some development programs, 85% overall power may therefore be a more realistic compromise than 90% with stringent futility. This is a design choice to justify in context, not a universal recommendation. Compare the maximum sample size, expected sample size, and probability of stopping a truly effective treatment. The examples below retain 90% power to make the sample-size cost visible; the binomial and normal examples compare a fixed design, efficacy-only monitoring, and monitoring with calibrated futility.

A useful planning constraint is to limit the maximum sample size to, say, 120% of the fixed-design size at the same alpha, beta, and effect assumptions. Rather than selecting interim CP targets in isolation, seek targets compatible with that inflation budget. This may require lower CP thresholds (less aggressive futility), different interim timing, or reconsideration of overall power. With several interims, assess the targets jointly: the constraint applies to the complete design, not separately to each look. The 20% limit is an illustrative planning choice, not a default constraint enforced by gsCPFutilitySpending(). The sample-size comparisons below show whether the chosen targets meet that illustrative budget; lowering an early CP target does not by itself guarantee inflation below 20%.

As cautious starting points for these examples, avoid CP targets above 0.2 before 50% information and consider up to 0.3 at or after 50%. These are planning heuristics, not universal cutoffs or assurances of acceptable futility. Inspect the effect estimates at the bounds, beta spending, and sample-size inflation before deciding whether even these targets are too stringent.

What is being calibrated?

Each CP target is the probability of crossing any future efficacy bound, from the analysis after the selected interim through the final analysis, conditional on the current statistic equaling the candidate futility bound. Future futility boundaries remain part of the calculation. In a three-analysis design, target_cp = c(.3, .5), i = c(1, 2) targets 0.3 for efficacy at analysis 2 or 3 given the first futility-bound statistic, and 0.5 for efficacy at analysis 3 given the second futility-bound statistic. Each target is a total future probability under its own conditioning value; it is not restricted to the next analysis. Such a two-target fit requires a two-parameter spending family, such as sfLogistic.

With theta = NULL, the default used below, CP assumes the observed effect implied by each candidate bound, lower$bound[i] / sqrt(n.I[i]). Thus we target the CP row in gsBoundSummary(), not CP H1 or PP. A specified future effect can instead be supplied with theta; a scalar is used at all targeted looks. This is a standardized effect on the scale used by gsCP().

The function supports beta-spending futility for test.type 3, 4, 7, and 8. The examples use type 4, fixed information fractions, and Lan-DeMets O’Brien-Fleming efficacy spending (sfu = sfLDOF), without sfupar.

All six calibrators hold x$timing fixed: these are the information fractions x$n.I / x$n.I[x$k] at the analyses. Power-preserving calibration can change the final information and all absolute analysis information levels while keeping those fractions. Fixed-information CA/conditional-POS calibration also keeps the absolute information levels fixed. Spending-function times, stored in x$upper$sTime and x$lower$sTime, are retained separately and can differ from the information fractions, as with calendar-based spending. Fixed information fractions do not by themselves fix calendar analysis dates.

Number of CP targets	Supported spending families
One	sfHSD, sfPower, sfExponential, sfLDOF
Two	sfLogistic, sfBetaDist, sfCauchy, sfNormal, sfExtremeValue, sfExtremeValue2
One or more	sfLinear, with one free cumulative spending proportion per target

The number of targets must match the number of free spending parameters. Targeted looks must be active interim futility analyses; the final analysis is not a calibration target. Multiple-target fitting first tries a joint fit from the reference parameters, with latest-to-earliest initialization as a fallback when needed. A fit is returned only when every residual meets control$cp_tol (default 1e-4).

We independently recompute CP from each returned or reconstructed design, rather than relying only on the calibration metadata:

check_cp <- function(design, target, i = seq_along(target), tol = 1e-4) {
  achieved <- vapply(i, function(j) {
    theta <- design$lower$bound[j] / sqrt(design$n.I[j])
    sum(gsCP(design, i = j, zi = design$lower$bound[j], theta = theta)$upper$prob)
  }, numeric(1))
  stopifnot(max(abs(achieved - target)) <= tol)
  stopifnot(abs(sum(design$upper$prob[, 2]) - (1 - design$beta)) < 2e-5)
  out <- data.frame(
    Analysis = i, Target = target, CP = achieved,
    Residual = achieved - target
  )
  size_label <- if (inherits(design, "gsSurv")) "Events" else if (design$n.fix > 1) {
    "Participants"
  } else "Information"
  out[[size_label]] <- round(design$n.I[i], 1)
  out
}

The comparison tables label cumulative event counts for survival designs and total participants across both groups for binomial and normal designs. These are planning quantities shown to one decimal, without integer conversion or recalibration. Boundary summaries also display the size at each analysis (rounded up for presentation).

One parameter: CP 0.3 at the first interim

Here futility is tested only at IA 1. The second interim retains efficacy testing but has no futility test.

x1 <- gsDesign(
  k = 3, test.type = 4, timing = c(.5, .75),
  sfu = sfLDOF, sfl = sfHSD, sflpar = 1,
  testLower = c(TRUE, FALSE, FALSE)
)
fit1 <- gsCPFutilitySpending(x1, target_cp = .3, i = 1)
fit1$cpFutilitySpending$sflpar
#> [1] 2.430891
check_cp(fit1, .3) |> lt()

Event-driven survival design with gsSurv()

The same information fractions and spending specification can be used in an event-driven survival design. Reuse the fitted parameter from fit1; other survival assumptions use gsSurv() defaults, with equal randomization made explicit. Check CP on the survival design itself, since numerical tolerances can introduce small differences.

surv_design <- gsSurv(
  k = 3, test.type = 4, timing = c(.5, .75),
  sfu = sfLDOF, sfl = sfHSD,
  sflpar = fit1$cpFutilitySpending$sflpar,
  testLower = c(TRUE, FALSE, FALSE), ratio = 1
)
check_cp(surv_design, .3) |> lt()

gsBoundSummary(surv_design, digits = 4, exclude = "B-value") |>
  lt() |> lt_format(columns = c("Efficacy", "Futility"), decimals = 4)

Survival designs can also be calibrated directly: gsCPFutilitySpending(surv_design, target_cp = .3, i = 1). Here the example illustrates transferring a statistical calibration to a survival design. The survival summary reports approximate hazard ratios at the bounds, rather than a generic standardized effect.

This first survival example is deliberately a cautionary one. CP 0.3 after 50% of the planned events gives a futility HR of about 0.8 and spends 7.7 percentage points of the total 10% Type II error budget at that first analysis. Thus it may stop a trial showing approximately a 20% hazard reduction, and uses roughly three quarters of the beta budget there. For a trial planned with 90% overall power, this looks overly aggressive and motivates considering a lower CP target even at 50% information.

Targeting predictive power instead

gsPPFutilitySpending() uses the same default prior as gsBoundSummary(), described in the introduction; neither call requires a prior argument. For comparison with the CP-targeted survival design, keep the same timing, testing indicators, spending families, and survival assumptions. Use the survival reference directly on the event-count scale. The prior’s numeric support must be on that same standardized-effect scale; the normalized n.fix = 1 reference from the opening example is not interchangeable with it.

pp_reference <- surv_design
fit_pp <- gsPPFutilitySpending(pp_reference, target_pp = .3, i = 1)
surv_pp <- fit_pp
# The normalized prior is retained for independent verification with gsPP().
prior <- fit_pp$ppFutilitySpending$prior
achieved_pp <- gsPP(
  surv_pp, i = 1, zi = surv_pp$lower$bound[1],
  theta = prior$z, wgts = prior$wgts
)
stopifnot(abs(achieved_pp - .3) <= 1e-4)
stopifnot(abs(sum(surv_pp$upper$prob[, 2]) - .9) < 2e-5)
gsBoundSummary(surv_pp, digits = 4, exclude = "B-value") |>
  lt() |> lt_format(columns = c("Efficacy", "Futility"), decimals = 4)

The PP row, not the CP or CP H1 row, now attains 0.3 at the first futility bound. The prior stays fixed during fitting, but each candidate’s posterior uses that candidate’s bound and information. The returned design includes the recalculated survival enrollment and event summaries.

survival_comparison <- function(design, label) {
  data.frame(
    Target = label,
    FutilityHR = gsHR(design$lower$bound[1], 1, design),
    BetaSpending = design$lower$spend[1],
    IA1Events = design$n.I[1], FinalEvents = max(design$n.I),
    MaximumN = max(design$N)
  )
}
rbind(
  survival_comparison(surv_design, "CP = 0.3"),
  survival_comparison(surv_pp, "PP = 0.3")
) |> lt() |>
  lt_format(columns = c("FutilityHR", "BetaSpending"), decimals = 4) |>
  lt_format(columns = c("IA1Events", "FinalEvents", "MaximumN"), decimals = 1)

Equal numerical targets for CP and PP need not imply equal futility bounds or required event counts and sample sizes. The earlier cautions about beta spending and sample-size inflation apply to PP targets too; averaging over a prior does not remove these design trade-offs.

Fixed calendar analyses with gsSurvCalendar()

For analyses at months 18, 27, and 36, the information fractions are generally not the calendar fractions, nor the .5, .75 fractions above. Derive them from a reference calendar design before calibrating. This example has a 12-month control median, hazard ratio 0.7, equal randomization, and a fixed 36-month trial with 12 months of minimum follow-up. Enrollment is scaled to power the design; the last enrollment period is extended as needed to fill the 24-month enrollment window.

The first calendar analysis below occurs before 50% of expected information. A CP target of 0.3 also looks too stringent here. We therefore lower it to 0.2 and inspect the resulting bounds; this is a more cautious starting point, not an automatic endorsement of the design.

calendar_args <- list(
  calendarTime = c(18, 27, 36), spending = "information",
  test.type = 4, sfu = sfLDOF, sfl = sfHSD, sflpar = 1,
  testLower = c(TRUE, FALSE, FALSE),
  lambdaC = log(2) / 12, hr = .7, ratio = 1,
  R = 12, minfup = 12
)
calendar_reference <- do.call(gsSurvCalendar, calendar_args)

Calibrate the calendar design directly, retaining its information fractions, spending times, and survival assumptions.

calendar_fit <- gsCPFutilitySpending(calendar_reference, target_cp = .2, i = 1)
calendar_design <- calendar_fit
stopifnot(max(abs(calendar_design$timing - calendar_reference$timing)) < 1e-6)
check_cp(calendar_design, .2) |> lt()

data.frame(
  Analysis = seq_len(calendar_design$k),
  Month = calendar_design$T,
  InformationFraction = calendar_design$timing,
  Events = round(calendar_design$n.I, 1)
) |> lt()

gsBoundSummary(calendar_design, digits = 4, exclude = "B-value") |>
  lt() |> lt_format(columns = c("Efficacy", "Futility"), decimals = 4)

This works because proportional enrollment scaling at fixed calendar times preserves the expected information fractions in this example. We explicitly verify that property and the resulting CP. It is not a general calibration method for designs that change enrollment shape, follow-up, or information fractions during optimization. We use information-based spending here; calendar-based spending would require retaining its distinct spending times as well. Do not simply reuse fit1’s parameter with different information fractions.

Two parameters: a risk-difference design

We now enable futility at both interims. A two-parameter spending function allows us to specify two CP targets. Consider adverse-event probabilities 0.15 on control and 0.10 on experimental treatment, with equal randomization. The benefit is the risk difference p[C] - p[E] = 0.05. nBinomial() gives the fixed-design total sample size; gsDesign() inflates it for sequential monitoring. Alpha is 0.025 one-sided and power is 90% in both calculations. Use sfCauchy for the worked example and compare all six families below. At 40% information, CP 0.3 would also be a stringent first futility target. We instead use CP 0.2 at 40% and CP 0.3 at 75% information. The first target covers efficacy at either the second interim or the final analysis; the second covers efficacy at the final analysis, each conditional on its own fitted futility-bound statistic.

n_binomial <- nBinomial(p1 = .15, p2 = .10, alpha = .025, beta = .1, ratio = 1)
n_binomial
#> [1] 1834.641
x2 <- gsDesign(
  k = 3, test.type = 4, timing = c(.4, .75), sfu = sfLDOF,
  n.fix = n_binomial, delta0 = 0, delta1 = .05, endpoint = "Binomial",
  testLower = c(TRUE, TRUE, FALSE)
)
target2 <- c(.2, .3)
fit2 <- gsCPFutilitySpending(x2, target_cp = target2, i = 1:2, sfl = sfCauchy)
check_cp(fit2, target2) |> lt()

Use the fitted parameter vector directly in a fresh gsDesign() call; do not copy a rounded parameter vector from printed output.

final2 <- gsDesign(
  k = 3, test.type = 4, timing = c(.4, .75),
  n.fix = n_binomial, delta0 = 0, delta1 = .05, endpoint = "Binomial",
  sfu = sfLDOF, sfl = sfCauchy,
  sflpar = fit2$cpFutilitySpending$sflpar,
  testLower = c(TRUE, TRUE, FALSE)
)
check_cp(final2, target2) |> lt()

gsBoundSummary(final2, deltaname = "p[C] - p[E]", Nname = "Participants",
               digits = 4, exclude = "B-value") |>
  lt() |> lt_format(columns = c("Efficacy", "Futility"), decimals = 4)

Fixed sample size and inflation from futility

The fixed-design total is 1834.6 participants across both groups. For comparison, remove futility but keep the same efficacy spending, timing, alpha, and power. The difference between this efficacy-only design and the calibrated design shows the additional maximum sample size associated with the futility rule.

binomial_efficacy_only <- gsDesign(
  k = 3, test.type = 1, timing = c(.4, .75), sfu = sfLDOF,
  n.fix = n_binomial, delta0 = 0, delta1 = .05, endpoint = "Binomial"
)
binomial_sizes <- c(n_binomial, max(binomial_efficacy_only$n.I), max(final2$n.I))
data.frame(
  Design = c("Fixed, no interims", "Efficacy only", "Efficacy and CP-based futility"),
  MaximumTotalN = binomial_sizes,
  InflationPercent = 100 * (binomial_sizes / n_binomial - 1)
) |> lt() |> lt_format(columns = c("MaximumTotalN", "InflationPercent"), decimals = 1)

Here the maximum sample size with calibrated futility is 1.36 times the fixed-design size. The comparison attributes most of the inflation to retaining 90% power in the presence of the futility bounds, rather than to the efficacy monitoring alone. These are unrounded planning sample sizes, not expected enrollment: early stopping can reduce the average number of participants used.

What does delta at the bound mean?

It is the effect estimate corresponding to the boundary Z-statistic, not the alternative effect used to power the trial. delta1 = .05 and delta0 = 0 tell the summary how to translate the canonical effect into a risk difference. The transformation used by gsDelta() is

\[ \widehat\delta_i = \delta_0 + (\delta_1-\delta_0)\frac{Z_i / \sqrt{n.I_i}}{\texttt{x\$delta}}. \]

Here x$delta is the standardized alternative, whereas delta1 is the clinical effect. Thus setting n.fix alone supplies a sample-size scale but does not supply the clinical effect scale. For binomial data these boundary risk differences are normal-approximation translations using design-based information, not exact binomial boundaries.

data.frame(
  Analysis = 1:2,
  Participants = round(final2$n.I[1:2], 1),
  FutilityRiskDifference = gsDelta(final2$lower$bound[1:2], 1:2, final2),
  EfficacyRiskDifference = gsDelta(final2$upper$bound[1:2], 1:2, final2)
) |> lt()

At a futility boundary, the default CP calculation assumes that this observed effect continues for future observations. It need not equal the planned 0.05 risk difference, and the same CP at different looks need not imply the same boundary effect. CP H1 instead uses the planned alternative. All sample sizes here are unrounded; integer conversion can change achieved CP.

Compare all six two-parameter families

The same two targets are feasible for each supported two-parameter family in this example. Function names and function objects are both accepted.

families <- c(
  "sfLogistic", "sfBetaDist", "sfCauchy", "sfNormal",
  "sfExtremeValue", "sfExtremeValue2"
)
fits2 <- setNames(lapply(families, function(family) {
  if (family == "sfCauchy") return(fit2)
  gsCPFutilitySpending(x2, target_cp = target2, i = 1:2, sfl = family)
}), families)
invisible(lapply(fits2, check_cp, target = target2))
spending_curves <- do.call(rbind, lapply(names(fits2), function(family) {
  fit <- fits2[[family]]
  times <- sort(unique(c(seq(0, 1, length.out = 501), .4, .75)))
  data.frame(
    Family = family, Time = times,
    Spending = fit$lower$sf(fit$beta, times, fit$lower$param)$spend
  )
}))

The fitted parameter vectors differ and have different meanings across families. The spending curves are more informative: they meet, up to numerical tolerance, at the two specified interim times, but have different shapes around them. The black points mark the fitted sfCauchy spending at those looks.

spending_plot <- ggplot(spending_curves, aes(Time, Spending, colour = Family)) +
  geom_vline(xintercept = c(.4, .75), linetype = "dashed", colour = "grey65") +
  geom_line(linewidth = .8) +
  geom_point(
    data = subset(spending_curves, Family == "sfCauchy" & Time %in% c(.4, .75)),
    colour = "black", size = 2.2
  ) +
  scale_colour_manual(values = c(
    sfLogistic = "#0072B2", sfBetaDist = "#D55E00", sfCauchy = "#7B3294",
    sfNormal = "#009E73", sfExtremeValue = "#E69F00", sfExtremeValue2 = "#CC79A7"
  )) +
  scale_x_continuous(breaks = c(0, .25, .4, .75, 1)) +
  labs(x = "Information fraction", y = "Cumulative Type II error spending",
       colour = "Spending family") +
  theme_classic(base_size = 12) +
  theme(legend.position = "bottom") +
  guides(colour = guide_legend(nrow = 2, byrow = TRUE))
spending_plot

Cumulative Type II error spending for six calibrated families. Dashed lines mark the interim information fractions; black points mark their common spending, up to calibration tolerance.

All six curves are similar over the range in which these interim analyses are most likely to occur, around and between the planned information fractions. The distinctive feature of sfCauchy is its spending when the first interim occurs substantially earlier than planned: in this example it spends less Type II error there, while smoothly approaching the same spending at the planned looks. That is an attractive shape when considering an early first analysis. All six curves are smooth inside (0, 1); Cauchy is not uniquely smooth. Adding looks or changing their timing requires recalibration if the CP targets are to be retained. The plot describes the spending rule away from the planned looks, not a guarantee of the targeted CP at an unplanned early look.

Piecewise-linear spending: comparing two normal means

Consider equal-sized groups with a planned mean difference of 0.5 and common standard deviation 1.1. nNormal() supplies the fixed-design total sample size for a one-sided alpha of 0.025 and 90% power. This uses a known-variance normal approximation, not a sequential t-test with estimated variance.

With three targets, use sfLinear. Its knot times are fixed at the targeted lower spending times. The free parameters are cumulative proportions of the total lower spending, constrained to be strictly increasing in (0, 1).

The starting proportions below help the numerical search. They are not CP targets and are not fixed final spending values. They may be omitted to use automatically derived starts; suitable explicit starts can speed up convergence.

n_normal <- nNormal(delta1 = .5, sd = 1.1, alpha = .025, beta = .1, ratio = 1)
n_normal
#> [1] 203.4237
x3 <- gsDesign(
  k = 4, test.type = 4, timing = c(.3, .5, .7), sfu = sfLDOF,
  n.fix = n_normal, delta0 = 0, delta1 = .5, endpoint = "Normal",
  testLower = c(TRUE, TRUE, TRUE, FALSE)
)
target3 <- c(.2, .3, .3)
fit3 <- gsCPFutilitySpending(
  x3, target_cp = target3, i = 1:3, sfl = sfLinear,
  control = list(start = c(.85, .95, .98))
)
check_cp(fit3, target3) |> lt()

fit3$cpFutilitySpending$sflpar
#> [1] 0.3000000 0.5000000 0.7000000 0.7667615 0.9288362 0.9671979

For sfLinear, the sflpar vector contains all knot times first, followed by all fitted cumulative spending proportions. There are six entries here but only three free parameters. The endpoint (1, 1) is implicit. These proportions are not the incremental spending amounts and are not CP values.

knots <- data.frame(
  Analysis = 1:3,
  Participants = round(fit3$n.I[1:3], 1),
  Time = x3$lower$sTime[1:3],
  CumulativeProportion = fit3$cpFutilitySpending$free_parameters
)
knots |> lt()

stopifnot(
  all(knots$CumulativeProportion > 0 & knots$CumulativeProportion < 1),
  all(diff(knots$CumulativeProportion) > 0)
)

As with the two-parameter example, verify a newly constructed design with the entire fitted vector, including the knot times:

final3 <- gsDesign(
  k = 4, test.type = 4, timing = c(.3, .5, .7),
  n.fix = n_normal, delta0 = 0, delta1 = .5, endpoint = "Normal",
  sfu = sfLDOF, sfl = sfLinear,
  sflpar = fit3$cpFutilitySpending$sflpar,
  testLower = c(TRUE, TRUE, TRUE, FALSE)
)
check_cp(final3, target3) |> lt()

gsBoundSummary(final3, deltaname = "Mean difference", Nname = "Participants",
               digits = 4, exclude = "B-value") |>
  lt() |> lt_format(columns = c("Efficacy", "Futility"), decimals = 4)

Because delta1 = .5, the effect-at-bound rows are in mean-difference units. They are not standardized mean differences or CP values. Sample sizes are totals across both groups, with their scale supplied by nNormal().

Fixed sample size and inflation from futility

The fixed-design total is 203.4 participants. Again, compare efficacy-only monitoring with the design that must also achieve the specified CP at three futility bounds, while keeping 90% overall power.

normal_efficacy_only <- gsDesign(
  k = 4, test.type = 1, timing = c(.3, .5, .7), sfu = sfLDOF,
  n.fix = n_normal, delta0 = 0, delta1 = .5, endpoint = "Normal"
)
normal_sizes <- c(n_normal, max(normal_efficacy_only$n.I), max(final3$n.I))
data.frame(
  Design = c("Fixed, no interims", "Efficacy only", "Efficacy and CP-based futility"),
  MaximumTotalN = normal_sizes,
  InflationPercent = 100 * (normal_sizes / n_normal - 1)
) |> lt() |> lt_format(columns = c("MaximumTotalN", "InflationPercent"), decimals = 1)

The calibrated design requires a maximum sample size 1.68 times the fixed-design size. As in the binomial example, most of the increase comes from compensating for futility stopping under the alternative, not from efficacy monitoring alone. This illustrates why a lower overall power target can be worth considering with aggressive futility. A different power target requires recalculating nNormal() and the sequential design with the same new beta, then refitting the spending parameters. These numbers are maximum, not expected, sample sizes.

A more conservative sequence of CP targets

The main example uses CP 0.2 at 30% information and CP 0.3 at 50% and 70%. We can also consider lower targets: here CP 0.1, 0.2, and 0.3 at the three futility bounds, retaining the same timing and efficacy spending.

target_varying <- c(.1, .2, .3)
fit_varying <- gsCPFutilitySpending(
  x3, target_cp = target_varying, i = 1:3, sfl = "sfLinear",
  control = list(start = c(.85, .95, .98))
)
check_cp(fit_varying, target_varying) |> lt()

gsBoundSummary(fit_varying, deltaname = "Mean difference", Nname = "Participants",
               digits = 4, exclude = "B-value") |>
  lt() |> lt_format(columns = c("Efficacy", "Futility"), decimals = 4)

Practical limits and diagnostics

• A single early futility analysis may often be sufficient. More futility looks add opportunities to stop a truly effective treatment and need not improve the development strategy.
• Even when several futility looks are planned, fitting a one-parameter spending function to the first interim CP target may be adequate. Keep later desired futility looks enabled in testLower, fit only i = 1, and inspect every resulting bound, its CP, effect estimate, beta spending, and sample-size implications. Later CP values are implied by the fitted spending function, not independently targeted. Use extra parameters only when the resulting later bounds do not meet the design objectives.
• More targets require more free parameters, not just a larger i vector. sfLinear supports additional targeted interim looks, but arbitrary target vectors are not guaranteed feasible under monotone cumulative spending.
• control$cp_tol sets the maximum absolute CP residual; reltol controls internal convergence and does not replace this acceptance check. See ?gsCPFutilitySpending for the complete controls and family-specific limits. For sfLinear, supply only cumulative proportions in control$start, not the complete sflpar; user-supplied lower and upper limits are unsupported.
• Inspect fit$cpFutilitySpending$residual, $information_ratio, and $solver for accuracy, information changes, and convergence diagnostics. Invalid inputs, infeasible targets, and convergence failures have distinct error classes; an unsuccessful solve does not return a calibrated design.
• All examples verify unconditional power as well as CP. In a general reconstruction, retain all reference settings, including alpha, beta, effect assumptions, testing indicators, and any harm-spending inputs. Changing those settings or efficacy spending requires recalibration.
• toInteger() retains the fitted spending specification, but rounding can change achieved CP and does not automatically recalibrate the design.
• All six calibrators accept gsSurv(), gsSurvCalendar(), and gsSurvPower() inputs. For power objects, power-preserving calibration fixes the realized schedule and rescales enrollment; fixed-information CA/conditional-POS calibration retains the event counts and enrollment. See the Survival designs section of ?gsCPFutilitySpending for details. Use gsPPFutilitySpending() for prior-averaged predictive power.