One model, block-permutation calibrated, for any term

The upper level of DMSA 2.0. Fits formula once, then permutes the reduced-model residuals within permutation blocks of equal size (Freedman-Lane) to obtain an exact p-value for one named coefficient - including an interaction. dmsa_moderate is the special case of this for score * moderator with a declared dmsa_design(); use that when you want the design contract enforced and simple slopes reported, and this when the question is an arbitrary term.

Usage

dmsa_model(
  formula,
  data,
  term,
  block = NULL,
  B = 1999,
  seed = NULL,
  nulls = FALSE
)

Arguments

formula

Model formula.

data

data.frame.

term

Name of the coefficient to test, exactly as model.matrix names it (e.g. "S:ACE").

block

Block labels, one per row of data - the family or cluster identifier. Rows are permuted only within blocks of EQUAL SIZE, because concatenating a random order of unequal-size families is not a valid relabelling.

B

Permutations.

seed

Optional integer.

nulls

Optional list of pre-computed permutation null matrices, as returned by an earlier call on the same frame. Supplying it reuses the shared permutation stream instead of drawing a new one, which is what keeps lens, engine and level results jointly calibrated.

Value

An object of class dmsa_model.

Examples

set.seed(1)
n <- 80
d <- data.frame(S = rnorm(n), ACE = rnorm(n),
                cID = rep(seq_len(n / 2), each = 2))   # couples
d$y <- 0.5 * d$S * d$ACE + rnorm(n)
## rows are relabelled only within equal-size families, so the couple
## dependence is preserved under the null
dmsa_model(y ~ S * ACE, d, term = "S:ACE", block = d$cID, B = 99, seed = 1)
#> DMSA one-model test of 'S:ACE'
#>   n = 80 in 40 blocks
#>   b = +0.4358 (SE 0.1200)   t = +3.63   partial r = +0.385
#>   p = 0.0100  (block permutation, B = 99)   parametric p = 0.0005075