Simulates trials from a planned design over a grid of sample sizes and effect sizes, runs the light stress test on each (covariance, inference, covariate, weighting and leave-one-subject-out families) and reports, for every cell, the conventional power next to the robustness power: the probability that the pre-specified analysis is significant and classed robust. The deletion search and a delta-adjusted tipping grid are optional and give the expected fragility index and the shift the result can absorb. This is the design-stage use of the stress test: choose the sample size so that a true effect is expected to come out robust rather than only significant, and write the expected verdict rates into the statistical analysis plan before unblinding so that the observed verdict has a pre-registered comparison.
Usage
stress_design(
n_per_arm,
delta,
n_trials = 100,
n_visits = 4,
sigma = 10,
rho = 0.6,
dropout = 0.08,
baseline_beta = 0.5,
covariance = "us",
method = "Satterthwaite",
vcov = "Asymptotic",
alpha = 0.05,
alternatives = c("us", "ar1h", "cs"),
loo = TRUE,
fragility = FALSE,
max_k = 5,
tipping = FALSE,
shifts = seq(0, 2, by = 0.25),
m = 5,
errors = "normal",
t_df = 3,
hetero = 1,
mnar = 0,
interaction = 0,
cores = 1L,
seed = 1,
rule = c("interval", "interval_or_direction")
)
# S3 method for class 'stress_design'
print(x, target = 0.8, ...)Arguments
- n_per_arm
Subjects per arm to evaluate (a vector).
- delta
Treatment effects at the last visit to evaluate (a vector); include 0 to see the null.
- n_trials
Trials per cell.
- n_visits
Number of post-baseline visits.
- sigma
Residual standard deviation at each visit.
- rho
AR(1) correlation between adjacent visits.
- dropout
Baseline per-visit dropout probability.
- baseline_beta
Slope of the response on the baseline covariate.
- covariance, method, vcov
The pre-specified analysis model, as in
stress_spec().- alpha
Two-sided significance level.
- alternatives
Covariance structures offered as alternatives; the pre-specified one is removed from the set.
- loo
Run the leave-one-subject-out family.
- fragility
Run the greedy deletion search for significance up to
max_kremovals (slow). Otherwise a one-subject loss of significance is read from the leave-one-out rows.- max_k
Removals allowed in the deletion search.
- tipping
Run a delta-adjusted tipping grid (imputed active-arm values shifted down) and record the shift at which
pcrossesalpha.- shifts
Shifts for the tipping grid as multiples of
delta(ofsigmawhendeltais 0).- m
Imputations per shift.
- errors
"normal"or"t"(Student t witht_dfdegrees of freedom, scaled to the same standard deviation: heavy tails).- t_df
Degrees of freedom for
errors = "t".- hetero
Multiplier of the residual standard deviation at the last visit relative to the first (1 = homoscedastic); the SD grows linearly.
- mnar
Strength of missing-not-at-random dropout: dropout after a visit depends on the value the subject would have had at the NEXT visit (0 = missing at random, as the default).
- interaction
Treatment effect modification by the baseline covariate: the effect at the last visit is
delta * (1 + interaction * (base - 50) / 10).- cores
Worker processes. Trials are distributed, so each trial's families run on one core.
- seed
Base seed; every trial's seed is derived from it.
- rule
Classification rule for
stress_verdict().- x
A
stress_designobject.- target
Robustness power to reach; the print method names the smallest sample size per arm that reaches it for each non-zero effect.
- ...
Ignored.
Value
A data frame of class stress_design with one row per cell:
n_per_arm, delta, trials, power (proportion significant),
robustness_power (significant and robust), robust_given_sig,
sensitive_given_sig, fragile_given_sig, marginal_given_sig,
k1_given_sig (one deletion loses significance), k_reached_given_sig
and k_median (deletion search, if run), tipping_margin (median shift
that crosses alpha among significant trials, as a multiple of delta
or sigma) and tipping_absorbed (share of significant trials with no
crossing within the grid), and max_shift_se (median largest shift in
pre-specified standard errors). attr(, "trials") holds the per-trial
rows and attr(, "design") the settings. plot.stress_design() draws
power and robustness power against the sample size.
Details
The light family set offers the alternative covariance structures, three
inference methods, the model without the baseline covariate, observed-margin
weighting and, if loo = TRUE, every single-subject deletion. Deletions
dominate the run time at large sample sizes; loo = FALSE gives a fast
first pass whose verdicts omit the deletion family.
See also
stress_calibrate() for the verdict rates of one design at a set
of planned powers; stress_sap() for the analysis-plan text.
Examples
# \donttest{
des <- stress_design(n_per_arm = c(40, 80), delta = c(0, 5), n_trials = 4, loo = FALSE)
des
#> Design stress test: 4 trials per cell, 4 visits, SD 10, AR(1) 0.6, dropout 0.08 per visit; model us / Satterthwaite / Asymptotic; alpha 0.05
#> Families: covariance, inference, covariates, weighting
#> n_per_arm delta trials power robustness_power robust_given_sig
#> 40 0 4 0.00 0.00 NA
#> 80 0 4 0.00 0.00 NA
#> 40 5 4 0.25 0.25 1.00
#> 80 5 4 1.00 0.75 0.75
#> sensitive_given_sig fragile_given_sig marginal_given_sig k1_given_sig
#> NA NA NA NA
#> NA NA NA NA
#> 0.00 0 0 0
#> 0.25 0 0 0
#> max_shift_se
#> 0.467
#> 0.343
#> 0.231
#> 0.539
#> Effect 5: no sample size in the grid reaches robustness power 0.8 (largest 0.75 at n = 80)
# }