Test an aligned tone score against a non-Gaussian outcome

The subject-level arm of DMSA for outcomes that are not continuous: binary (logistic), ordered-categorical (proportional odds), unordered-categorical (multinomial), or counts (Poisson / negative binomial). Optionally moderated.

Usage

dmsa_outcome(
  data,
  outcome,
  score,
  family = c("gaussian", "binomial", "ordinal", "multinomial", "poisson"),
  moderator = NULL,
  covariates = character(),
  block = NULL,
  center = TRUE,
  B = 999L,
  seed = NULL
)

Arguments

data

data.frame with one row per subject.

outcome

Name of the outcome column. For "binomial" it may be 0/1, logical, or a two-level factor; for "ordinal" an ordered factor or an integer-coded severity; for "multinomial" a factor.

score

Name of the tone-score column (from dmsa_score()).

family

One of "gaussian", "binomial", "ordinal", "multinomial", "poisson".

moderator

Optional moderator column; when given, the tested term is score:moderator.

covariates

Character vector of covariate columns.

block

Optional column naming the exchangeable unit for permutation (families, repeated measures).

center

Center the moderator (default TRUE).

B

Permutations. 0 skips them and returns model-based inference only.

seed

Optional integer.

Value

An object of class dmsa_outcome.

Details

Direction. A binary, count or ordinal outcome yields ONE coefficient for the tone score, so DMSA's directional claim survives intact - the sign of the log-odds (or log-rate, or latent-scale slope) is the expression-aligned direction. A multinomial outcome yields K-1 coefficients and there is no single sign: the test becomes an omnibus "tone predicts category" and the per-category log-odds are reported separately. If the categories are in any sense ordered, family = "ordinal" keeps the direction and is strictly preferable.

Examples

set.seed(1)
n <- 120
d <- data.frame(S = rnorm(n), cv = rnorm(n),
                cid = rep(seq_len(n / 2), each = 2))
d$y <- rbinom(n, 1, plogis(0.9 * d$S))
## a binary outcome yields ONE coefficient, so the sign of the log-odds
## is still the expression-aligned direction
dmsa_outcome(d, outcome = "y", score = "S", family = "binomial",
             covariates = "cv", block = "cid", B = 99, seed = 2)
#> DMSA subject-level test (binomial)
#>   y ~ S   n = 120
#>   log-odds = +0.7094  se = 0.2377  |z| = 2.98  p(model) = 0.002845
#>   exp(coef) = 2.033
#>   permutation p = 0.01  (B = 99, predictor-side residual permutation)