Skip to contents

Relative treatment effects from a cpaic fit

Source: R/effects.R
relative_effects.Rd

Tidies the relative effects of the fitted model: every treatment versus a chosen reference, or all pairwise comparisons. Effects are reported on the natural scale of the summary measure (e.g. odds ratios) unless backtransf = FALSE.

Usage

relative_effects(
  object,
  reference = NULL,
  all_contrasts = FALSE,
  backtransf = TRUE,
  level = 0.95,
  newdata = NULL,
  estimand = NULL,
  target = NULL,
  weights = NULL,
  measure = NULL,
  baseline_study = NULL,
  times = NULL,
  random_effect = "population",
  ...
)

Arguments

object

A fitted cpaic object (cpaic_bridge, cpaic_maic, cpaic_stc, or cpaic_mlnmr).

reference

Reference treatment. Defaults to the network reference.

all_contrasts

If TRUE, return all pairwise comparisons instead of versus the reference.

backtransf

If TRUE (default) back-transform log-scale measures (OR/RR/HR/...) by exponentiating.

level

Confidence level for the intervals. Default 0.95.

newdata

For cmlnmr() fits: a one-row data frame giving target effect-modifier means. Required when the model has effect modifiers.

estimand

For cmlnmr() fits, either "average_conditional_link" or "marginal". The latter delegates to marginal_effects().

target, weights, measure, baseline_study, times, random_effect

Arguments used only for estimand = "marginal"; see marginal_effects().

...

Unused.

Value

For frequentist and average-conditional outputs, a data frame with columns treatment, comparator, estimate, estimate_link, se_link, lower, upper, scale, and z/p or Bayesian pr_gt0. estimate_link and se_link are on the model link scale. estimand = "marginal" returns the schema documented by marginal_effects(), with canonical estimate_contrast, se_contrast, and contrast_scale columns because natural differences are not model-link quantities.

Details

Relative effects that the component design cannot uniquely identify (their contrast vector lies outside the row space of X = B C) are returned as NA rather than as pseudoinverse or prior-driven artefacts. See estimable_effects().

For cmlnmr() fits, estimand = "average_conditional_link" reports component-additive link-scale contrasts at supplied covariate means: theta_t(x) = C_t' (beta + gamma x). Because this expression is linear in x, evaluating it at E[X] gives the average conditional link-scale effect. Set estimand = "marginal" to delegate to marginal_effects() and standardize treatment-specific outcomes over an explicit target covariate distribution before forming contrasts.

The average-conditional path is narrower than it may look. A one-row newdata profile is not a target covariate distribution, and exponentiating a link-scale contrast does not create a marginal standardized effect. The marginal path therefore requires target instead. In either path, interactions informed only by aggregate arms can reflect ecological rather than within-study effect modification; see estimable_effects_at().