Omnibus across the score projections

Refits the same model once per projection, substituting each into the column named score_col, and combines the p-values with the aggregated Cauchy test. ACAT is valid under arbitrary dependence, so hedging across projections costs no multiplicity - unlike a Bonferroni over arms, which is what the old min(1, 2*min(p)) gate amounted to.

Usage

dmsa_model_omnibus(
  scores,
  formula,
  data,
  term,
  block = NULL,
  score_col = "S",
  B = 1999,
  seed = NULL
)

Arguments

scores

data.frame from dmsa_scores().

formula

Formula referring to the score by score_col.

data, block, B, seed

As in dmsa_model().

term

Term to test.

score_col

Column name the formula uses for the score.

Value

list: p_by_flavour, the omnibus p, and the fits.

Examples

set.seed(1)
n <- 80
M <- matrix(rnorm(n * 6), n, 6)
al <- dmsa_align(data.frame(cpg = paste0("cg", 1:6), d = c(1, 1, -1, 1, -1, 1),
                            p_plus = rep(.9, 6)),
                 genes = rep("OXTR", 6), level = "gene")
sc <- dmsa_scores(M, al)
d <- data.frame(age = rnorm(n), cID = rep(seq_len(n / 2), each = 2))
d$y <- 0.6 * sc$aligned + rnorm(n)
## one p per projection, combined by ACAT - hedging costs no multiplicity
o <- dmsa_model_omnibus(sc, y ~ S + age, d, term = "S", block = d$cID,
                        B = 99, seed = 1)
round(c(o$p_by_flavour, omnibus = o$p), 3)
#> aligned    mean     pc1   blind omnibus 
#>   0.010   0.010   0.330   0.250   0.019