Modifier-level treatment effects

Evaluates CAMS modifier-specific treatment effects at the pooled (precision-weighted) information fraction from ifrac(). Vignettes often use published subgroup tables (IV/NIV, age strata, …); the same extractor applies whenever each trial contributes one estimate per modifier level (long format). The modifier column defaults to "level"; pass modifier = on cams() to use another column name. Display headings capitalize the first letter of that column name.

Usage

modifiers(
  fit,
  at = NULL,
  ifrac = NULL,
  type = c("Overall", "Prediction"),
  prob = NULL,
  max_support = NULL,
  summary = TRUE,
  contrast_sep = NULL,
  exponentiate = NULL,
  shrinkage = FALSE,
  contribution
)

Arguments

fit: A interactmeta from cams().
at: Combined-population information fraction $\pi$ at which to evaluate subgroup effects. If NULL (default), uses the pooled reporting IF from ifrac() for fit. A numeric vector is aligned to posterior draws; a scalar (e.g. at = 0.5) is a plug-in at every draw.
ifrac: When at is NULL, passed to ifrac(): "interaction" (default; interaction-weighted pooling of trial-level $\pi_j$) or "overall" for overall-block precision weights.
type: One or both of "Overall" and "Prediction". "Overall" uses fixed effects only; "Prediction" adds a new-trial random effect drawn from the posterior trial SDs at the combined-population IF (effect in a new trial).
prob: Coverage probability for lci / uci when summary = TRUE.
max_support: Optional cap on unweighted posterior rows. Rejected for Bayesian CAMS DIRECT supports (weighted quadrature nodes); omit it.
summary: If TRUE (default), include a summary data frame.
contrast_sep: Rows whose subgroup label contains this separator are excluded when resolving subgroup levels (default "-").
exponentiate: If TRUE, transform posterior draws with exp() before summarizing (e.g. log rate ratios to ratios).
shrinkage: If TRUE, delegate to shrinkage() with type = "modifiers" (trial-specific reconstruction at each trial's $\pi_j$). If FALSE (default), return only Overall / Prediction at the combined-population information fraction.
contribution: Deprecated; use \code{at}.

Value

A interactmeta_subgroup_effects object (list with draws, summary, overall_ifrac, overall_ifrac_source ("fit" or "user"), reference, prob, exponentiate). Draws are indexed by subgroup and type. summary includes estimate (median), mean, sd, lci, and uci. For Bayesian CAMS, lci/uci follow fit$interval.type (default central) on the DIRECT-node mixture.

Details

This is not a marginal treatment effect in the survey-sampling sense: it is the modifier-level linear predictor at the combined-population contribution structure $\widetilde{\pi}$.

For each posterior draw, the linear predictor is evaluated at that draw's combined-population information fraction with the appropriate subgroup fixed effect: $$\mu_A(\pi^\star)=\alpha+\pi^\star(\theta-\gamma),$$ $$\mu_B(\pi^\star)=\alpha+\pi^\star(\theta-\gamma)+\gamma.$$ On the orthogonal CAMS factorization, $\theta$ is the overall-block IF slope and stored b_ifrac is $\beta=\theta-\gamma$. This is not the trial-specific convolution on fit$latent$mu_A[[j]] / fit$latent$mu_B[[j]] ($\mu_{Aj}=m_j-\pi_j g_j$ at that trial's $\pi_j$). A scalar at is a plug-in $\pi^\star$; NULL mixes the ifrac() posterior.

$\alpha$ and $\theta$ share the overall-block support. Population $M^\star=\alpha+\pi^\star\theta$ is formed from that joint, not by convolving the two mixture marginals.

Usage

Arguments

Value

Details

See also