Quickstart: Prediction-Powered Power & Sample Size

What this vignette covers

This is the basic-use guide for pppower. It shows the minimum calls you need for power, sample size, paired designs, and 2x2 contingency-table planning. See the “Detailed Dive” and “2x2 Calculator” tabs on the pkgdown site for derivations, regression contrasts, and odds-ratio / relative-risk setups.

library(pppower)

What it does (analogy to `pwr`)

Drop-in power / sample-size calculators that leverage predictions on unlabeled data (PPI++).
Same spirit as pwr::pwr.t.test, but variance comes from predictions + residuals.
Works for means, paired differences, regression contrasts, and 2x2 contingency tables.

1. Power for a prediction-powered mean test

power_ppi_mean(
  delta    = 0.2,   # effect size
  N        = 5000,  # unlabeled size
  n        = 200,   # labeled size
  var_f    = 0.6,   # Var(f(X))
  var_res  = 0.4,   # Var(Y - f(X))
  cov_y_f  = 0.6    # Cov(Y, f)
)
#> [1] 0.9915412

2. Required labeled size for target power

power_ppi_mean(
  delta    = 0.2,
  N        = 5000,
  n        = NULL,
  power    = 0.80,
  sigma_y2 = 1.0,
  sigma_f2 = 0.6,
  cov_y_f  = 0.6
)
#> [1] 81

You can also supply prediction metrics (MSE, $R^2$ , sensitivity/specificity) instead of raw moments; see the “Detailed Dive” vignette.

3. Paired designs in one line

power_ppi_paired(
  delta    = 0.3,   # mean difference
  N        = 1000,  # unlabeled pairs
  n        = 50,    # labeled pairs
  sigma_D2 = 1.0,
  rho_D    = 0.7
)
#> [1] 0.8276133

4. 2x2 designs in one line

power_ppi_2x2(
  p_exp          = 0.40,
  p_ctrl         = 0.20,
  N              = 1000,
  power          = 0.80,
  sens           = 0.85,
  spec           = 0.85,
  effect_measure = "OR"
)
#> [1] 108

Where next?

Detailed Dive: formulas, metrics-to-moments mapping, regression contrasts.
2x2 Calculator: odds-ratio and relative-risk planning from contingency tables.