Per-case posterior decision summary

For each unit and each fixed-effect term that has a corresponding unit-level random effect, returns the unit-specific posterior (b + u) summarized with decision quantities, on the link scale.

Usage

summarize_per_case_effects(
  model,
  case_var = "case",
  threshold = NULL,
  rope = NULL,
  intervention_goal = c("increase", "decrease"),
  ci_level = 0.95
)

Arguments

model

A fitted brmsfit.

case_var

Name of the case-level grouping column in the model (default "case").

threshold

Numeric. A positive magnitude expressed on the improvement-oriented scale (i.e., after any decrease reflection). If supplied, the posterior probability that the effect meets or exceeds this magnitude in the goal direction is reported as prob_meaningful_change. It applies to level terms only: slope terms (trend_*, time_*) get NA, since a per-session slope is meaningful or not only relative to the number of sessions it is projected over.

rope

Numeric length-2. ROPE bounds on the link scale, expressed on the native coefficient scale. For intervention_goal = "decrease" the bounds are reflected together with the draws, so an asymmetric ROPE stays attached to the same parameter region. If supplied, posterior probability inside the ROPE is reported as rope_prob for level terms; slope terms get NA. default_rope() gives the default region.

intervention_goal

"increase" (default) or "decrease". For "decrease", every reported coefficient is sign-flipped so that a positive value denotes improvement (full reflection onto an improvement-oriented scale); pd is unchanged and the ROPE is reflected with the draws. Reversal withdrawal steps (an intervention-role phase followed by a baseline-role phase, e.g. level_B1A2, trend_Intervention1Baseline2) remove the treatment, so they take the opposite reflection: under "increase" they are sign-flipped and under "decrease" they are not. A positive value on every row therefore reads as change in the direction the intervention predicts (improvement at onset, return toward baseline at withdrawal). Intercept and ordinal threshold terms (b_Intercept) and coefficients of distributional parameters (b_sigma_*, b_nu_*, b_phi_*, b_shape_*, b_zi_*, b_hu_*, b_disc_*, ...) are excluded from the table, so only treatment, trend, and moderator slopes on the mean are reported.

ci_level

Credible interval width (default 0.95).

Value

A data frame with columns case (the unit label), term, mean, sd, lower, upper, pd, prob_meaningful_change, rope_prob, plus the response-scale ratio columns of summarize_fixed_effect_decisions() on a log or logit link. Those are exp(b + u), the ratio conditional on that case's random effects, and carry ratio_type = "conditional"; they are not population-averaged ratios (see marginal_contrast()).

Details

The unit is the grouping factor that contains case_var and carries random slopes on fixed-effect terms, the innermost one when several do: with the default structure (1 | case) + (1 + level_AB | case:behavior) the units are case-behavior tiers, labeled "<case>_<behavior>" as brms labels them. Random effects of every enclosing grouping (study, study:case, case) on the same term are added, so the unit posterior is the full b + u_study + u_case + u_tier.

Examples

if (interactive()) {
  dat <- simulate_scd_data()
  fit <- brms::brm(outcome ~ phase + (1 + phase | case),
                   data = dat, refresh = 0)
  summarize_per_case_effects(fit, threshold = 0.5)
}