Two-arm power for the planned human study, priced from the silicon pilot

Analytic two-arm sample sizes with dispersion priors taken from the silicon responses: Likert items use the pilot standard deviation of score; choice items use the pilot share of the modal option; open items are skipped. The priors inherit the panel's calibration status – an uncalibrated pilot prices the design stage, it does not certify effect sizes.

Usage

panel_power(
  responses,
  effect,
  items = NULL,
  focal = NULL,
  alpha = 0.05,
  power = 0.8
)

Arguments

responses: A panel_administer() result (the silicon pilot).
effect: Raw minimum detectable difference between the two arms: scale points for Likert items, a difference in proportions for choice items. A scalar (recycled) or a named vector keyed by item_id.
items: Optional character vector restricting which items to price.
focal: Optional named character vector keyed by item_id, giving the focal response level whose proportion the power calculation should target. For a binary choice item the modal option is a well-defined estimand and focal is optional; for a choice item with three or more options the "modal share" is not a meaningful single proportion, so name the focal response here. Without a focal for a 3+-option item, the modal share is used and a warning is issued.
alpha: Two-sided test size.
power: Target power.

Value

A tibble: item_id, type, dispersion (sd for Likert, the focal or modal share for choice), effect, n_per_arm.

Examples

if (FALSE) { # \dontrun{
set.seed(110)
panel <- panel_from_margins(list(group = c(A = .5, B = .5)), n = 8)
instr <- panel_instrument(list(
  item_likert("lik", "Rate the proposal.", scale = c("low", "mid", "high")),
  item_choice("pick", "Pick one.", c("A", "B"))),
  randomize = character(0))
cfg <- LLMR::llm_config("groq", "openai/gpt-oss-20b")
r <- panel_administer(panel, instr, cfg)
panel_power(r, effect = c(lik = 0.5, pick = 0.2))
} # }