Power Analysis for Between-Subject Discounting Designs

The Monte Carlo runs below are precomputed (the full set of power_discounting() / find_n_discounting() calls in this vignette takes several minutes, well past CRAN’s vignette-build budget), so the outputs shown come from real runs; copy the chunks into your session to reproduce them.

Why simulation-based power?

Before running a study, a researcher planning a delay-discounting experiment needs an answer to: “How many subjects do I need to detect a group difference in discounting of this size at 80% power?” Closed-form power formulas do not exist for mixed-effects discounting models with scale-location-truncated beta (SLT-beta) observation models, and Wald asymptotics can be unreliable at realistic study sizes. power_discounting() therefore estimates power by Monte Carlo simulation:

Simulate a two-group indifference-point dataset with [simulate_dd_ip()] under assumed population parameters, a hypothesized delta_k (the condition shift on natural-log k), and the design.
Refit it with the package’s TMB mixed-effects fitter, [fit_dd_tmb()]. The simulation and refit always use the same discounting function and observation family, so the tested contrast is the simulated effect.
Extract the Wald standard error of the log-k condition contrast from the TMB sdreport.
Record both a p-value verdict and a confidence-interval verdict (does the Wald CI exclude 0). Both are referred to a t distribution with n_subjects - 2 degrees of freedom (the design’s two-sample df) rather than the asymptotic normal, which is anticonservative at study-relevant N. This df is an empirical calibration device, not a df derived from the mixed model (the TMB fit has no exact t sampling theory); df = Inf recovers the z-test. The two verdicts use the same standard error and reference distribution, so they always agree; both report formats are returned.
Repeat n_sim times. Power is the proportion of usable fits (converged, positive-definite Hessian, finite SE) that reject, reported with a Wilson confidence interval because it is itself an estimate from finitely many replicates.

The design is between-subject: n_subjects is the total N, split across the two conditions round-robin (use even numbers for equal groups). This mirrors beezdemand::power_demand(), which models a within-subject condition. The asymmetry matches what each package’s simulator and fitter already represent.

Estimating power at a fixed sample size

Suppose pilot data suggest the comparison group discounts about twice as steeply (delta_k = log(2) ≈ 0.69) and subjects vary with sigma_u = 0.6 around their group’s log k:

res <- power_discounting(
  n_subjects = 60,
  effect = list(delta_k = log(2)),
  design = list(
    delays = c(7, 30, 180, 365, 730, 1460, 2920),
    sigma_u = 0.6,  # subject SD on log k
    phi = 10        # SLT-beta precision
  ),
  n_sim = 40,       # small for a fast vignette; see guidance below
  seed = 1234,
  verbose = FALSE
)
res
#> Monte Carlo power analysis (beezdiscounting)
#>   Target: k:conditionC2 (delta_k = 0.6931), two-sided alpha = 0.05, t reference (df = 58)
#>   n_subjects = 60 (between-subject, 2 conditions), n_sim = 40 (converged 40, usable 40)
#>   Power (CI-exclusion): 0.925 [95% MC CI 0.801, 0.974]
#>   p-value hit rate:     0.925

Convergence diagnostics are part of the output. Replicates whose refit fails are excluded from the power denominator and counted in n_converged / n_used rather than treated as “no effect detected”.

res$n_converged
#> [1] 40
res$n_used
#> [1] 40
table(res$replicates$status)
#> 
#> ok 
#> 40

Searching for a sample size

find_n_discounting() wraps the engine in a bisection search over total N for the smallest sample whose estimated power reaches the target. The search accounts for Monte Carlo noise: at each candidate N it adds replicates until the Wilson interval falls clearly above or below the target (up to n_sim_max), and it re-evaluates the selected N and its lower neighbor (or the lower bound itself, when the target is already met there) before reporting. Bisection assumes power is monotone in N; because each N is judged from independent replicates, a fluctuation at a lower N can hide a crossing the search never revisits. Evaluated N that contradict monotonicity demote the status to uncertain, but never-visited N cannot be checked, so treat the result as an estimate and confirm it with a large n_sim at the chosen N.

search <- find_n_discounting(
  target_power = 0.8,
  effect = list(delta_k = log(3)),
  n_range = c(6, 60),
  n_sim = 30,       # small for a fast vignette
  seed = 5678,
  verbose = FALSE
)
search
#> Sample-size search (Monte Carlo power)
#>   Target power 0.80 for delta_k = 1.099 at alpha = 0.05
#>   Estimated minimum n_subjects = 13 (status: uncertain)
#>   This is an estimated minimum under Monte Carlo uncertainty;
#>   at least one decision fell back to a point estimate, so the true
#>   power at this N may be below the target.
#>   Validate it with an independent power_discounting() run at this N
#>   (large n_sim, new seed) before reporting.
#> 
#>   Evaluations:
#>  n_subjects n_sim_total n_used usable_fraction     power  ci_lower  ci_upper
#>          60          30     30               1 1.0000000 0.8864866 1.0000000
#>           6          30     30               1 0.5666667 0.3919731 0.7262251
#>          33          30     30               1 1.0000000 0.8864866 1.0000000
#>          19          60     60               1 0.9666667 0.8863623 0.9908107
#>          12         120    120               1 0.7750000 0.6924310 0.8405085
#>          15          60     60               1 0.9500000 0.8629948 0.9828505
#>          13         120    120               1 0.8333333 0.7565472 0.8894400
#>          13         120    120               1 0.8833333 0.8136665 0.9292188
#>          12         120    120               1 0.8333333 0.7565472 0.8894400
#>         decision
#>            above
#>            below
#>            above
#>            above
#>  ambiguous_below
#>            above
#>  ambiguous_above
#>            above
#>  ambiguous_above

The result is an estimated minimum under Monte Carlo uncertainty (the status field says whether the confirmation pass was conclusive). For a defensible grant number, rerun power_discounting() at the returned N with a large n_sim and report that estimate with its Monte Carlo interval.

Sensitivity to the assumed effect size

Power claims are conditional on the assumed effect and heterogeneity. A small sweep makes that dependence visible:

deltas <- c(log(1.5), log(2), log(3))
sweep <- vapply(deltas, function(d) {
  power_discounting(
    n_subjects = 60, effect = list(delta_k = d),
    n_sim = 30, seed = 42, verbose = FALSE
  )$power
}, numeric(1))
data.frame(k_ratio = exp(deltas), power = sweep)
#>   k_ratio     power
#> 1     1.5 0.6333333
#> 2     2.0 1.0000000
#> 3     3.0 1.0000000

How many replicates do you need?

The Monte Carlo standard error of a power estimate near p is sqrt(p (1 - p) / n_sim):

`n_sim`	MC SE near power = 0.8	Use
100	0.040	quick interactive exploration
500	0.018	default; serious exploration
2000	0.009	grant applications, preregistrations

Always report the Wilson interval (power_mc_ci) alongside the point estimate.

Validity and Limitations

Validity checks. The package test suite (tests/testthat/test-dd-power.R) verifies, with preregistered seeds and tolerance bands fixed before the tests were first run:

Type I error calibration (the primary check): with delta_k = 0, the empirical false-positive rate at nominal alpha = .05 over 1,200 replicates must fall in [0.03, 0.07] (a band of 3.18 binomial standard errors that excludes both half and 1.5 times the nominal rate). n_sim = 1200 was computed from that tolerance (9 * .05 * .95 / .02^2 ≈ 1069). A second null check runs at N = 60. The t(n - 2) reference distribution these checks calibrate was adopted after the sibling beezdemand calibration battery showed the asymptotic z-test is anticonservative at study-relevant sample sizes.
Convergence handling: a configuration that reliably produces non-convergence confirms that failed replicates are excluded from the power denominator and reported in n_converged / n_used rather than counted as misses.
Closed-form benchmark: with family = "gaussian", tiny residual error, and many delays, each subject’s log k is recovered nearly exactly and the design reduces to a two-sample comparison of log k; the Monte Carlo estimate must match pwr::pwr.t.test() within a preregistered tolerance of 0.10.
Monotonicity: power increases with n_subjects and delta_k and decreases with sigma_u, within Monte Carlo slack.
Reproducibility: identical seed gives identical results, asserted exactly.

These checks validate the default configuration (equation = "mazur", family = "sltb", random_effects = k ~ 1), where the refit model exactly matches the data-generating process.

The find_n_* search statuses ("confirmed", "uncertain", "unresolved") describe a heuristic Monte Carlo decision rule (repeated looks at ordinary Wilson intervals across candidate N) rather than a formal sequential testing procedure with a guaranteed error rate. When the confirmation pass contradicts the search, the function returns n = NA rather than an unsupported number.

Explicitly out of scope in v1 (deferred to future work; none of these is approximated):

Effects on the curvature exponent s (Green-Myerson / Rachlin) or the SLT-beta precision phi; the two-parameter equations are excluded entirely.
Power for derived measures (ED50, AUC, indifference points at a delay).
Random effects beyond the single log-k intercept; correlated k/phi or k/s random effects. (random_effects accepts such formulas, but the simulator still generates only the log-k intercept, so the refit is over-specified relative to the data-generating process; treat such a run as a robustness probe rather than a power estimate under those random effects.)
Arbitrary user-supplied designs beyond the package’s between-subject simulator (two conditions, round-robin allocation, all subjects at all delays).
Any graphical or interactive interface.