Power curve for a Gaussian graphical model (DPIR, BFDA, or BSDA)

Evaluates power at a grid of sample sizes without searching for n*. Use together with plot() to see how power changes with n. "DPIR" evaluates Pr(DPIR > threshold) both globally and averaged across off-diagonal parameters (unlike design's "pw" target, which is the weakest off-diagonal parameter, not the average); "BFDA" evaluates power and error rate for a single edge; "BSDA" evaluates Pr(measure >= measure_value) over the whole graph.

Usage

# S3 method for class 'ggm_elicited'
power_curve(
  params,
  method = c("DPIR", "BFDA", "BSDA"),
  n,
  target_probability = 0.95,
  edge = NULL,
  rho_quantile = 0.5,
  pow0 = 0.8,
  pow1 = 0.8,
  threshold = NULL,
  measure = c("sen", "spe"),
  measure_value = 0.8,
  control = bsda_control(),
  H = 150L,
  J = 20L,
  ...
)

Arguments

params

A ggm_elicited object, as returned by elicit_prior. It also inherits from bgm_elicited.

method

Which planning method, "DPIR" (default), "BFDA", or "BSDA".

n

Candidate sample sizes to evaluate (numeric vector).

target_probability

DPIR only, as in design. Not used in the grid computation itself (there is no search to stop); kept for the target reference line in plot() and print().

edge

BFDA only: a length-2 integer vector, 1-based indices into the rows/columns of K and G as supplied to ggm_parameters, naming the edge to evaluate. The edge must be present in G. Order does not matter: the pair is sorted internally, since the underlying Bayes factor computation requires m < l. NULL (default) selects via rho_quantile, as in design.

rho_quantile, pow0, pow1

BFDA only, as in design.

threshold

DPIR and BFDA only, as in design. NULL (default) uses 1.0 for DPIR and 10.0 for BFDA.

measure, measure_value

BSDA only: the measure to evaluate ("sen" or "spe") and the target value of that measure, as in design. sen and spe are evaluated by separate calls (they are two different computations, not two outputs of the same one); call power_curve() twice to get both.

control

BSDA only: a list of control parameters, as returned by bsda_control. Default bsda_control().

H, J

Outer and inner Monte Carlo replication counts, as in design.

...

Ignored, present for consistency with the generic power_curve().

Value

A ggm_power_curve object, which also inherits from bgm_power_curve. Its results$table component is a data.frame with one row per n: for DPIR, n, prob_global, prob_pw_mean (the average, across off-diagonal parameters); for BFDA, n, power_h0, fpr_h0, power_h1, fnr_h1; for BSDA, n, power, measure_achieved, halfwidth, se. print and plot methods are available.

See also

Other sample size planning: bsda_control(), design(), design.ggm_elicited(), power_curve(), validate(), validate.ggm_design()

Examples

p <- 3
G <- matrix(0, p, p)
G[1, 2] <- G[2, 1] <- 1
K <- diag(p)
K[1, 2] <- K[2, 1] <- 0.3
ep <- elicit_prior(ggm_parameters(K, G, nu = 10))

# DPIR (the default method)
set.seed(12)
pc <- power_curve(ep, method = "DPIR", n = c(20, 60, 150), threshold = 0.5, H = 30, J = 10)
pc
#> <power_curve>  method: DPIR  family: ggm  prior: gwishart 
#>   criterion: DPIR > 0.500  (target probability = 0.95)
#>    n prob_global prob_pw_mean
#>   20           1            1
#>   60           1            1
#>  150           1            1
plot(pc)


# \donttest{
# BFDA: plans around a single edge
set.seed(12)
pc_bfda <- power_curve(ep, method = "BFDA", edge = c(1, 2), n = c(20, 60, 150), H = 30, J = 10)
plot(pc_bfda)


# BSDA: needs pip and a sparse graph. Here, control() settings are kept light
ep_pip <- elicit_prior(ggm_parameters(K, G, nu = 10, pip = 0.5))
ctrl <- bsda_control(
  fit_iterations = 300, fit_burnin = 150, n_boot = 200,
  gwish_iter = 80
)
set.seed(12)
pc_bsda <- power_curve(ep_pip,
  method = "BSDA", n = c(50, 150, 300), H = 30, J = 1,
  measure = "sen", measure_value = 0.8, control = ctrl
)
plot(pc_bsda)

# }