Per-group marginal means for a symbolized model

group_means() returns categorical marginal means for the factors in a symbolized_model, delegating to emmeans::emmeans() under the hood. Each row is one combination of the requested factors' levels with a point estimate, standard error, 95% confidence band, and an excludes_zero indicator.

Use this alongside parameter_interpretation() for any fit that contains a factor: the coefficient table shows contrasts (differences from the reference level); group_means() shows each group's expected response.

Not supported on bivariate Gaussian (biv_gaussian) fits — a marginal mean in a bivariate fit is a joint 2-vector (mu1, mu2), not a scalar, so the emmeans abstraction does not apply. Use drmTMB::predict_parameters() on the fit directly, or refit each response as a univariate Gaussian and call group_means() on each.

Usage

group_means(x, by = NULL, scale = c("response", "link"), ci_method = NULL, ...)

Arguments

x

A symbolized_model whose underlying fit is retained on x$metadata$fit.

by

Optional character vector of factor names to marginalize over. Defaults to all factors in the model.

scale

One of "response" (default) or "link". Controls the scale of the returned estimate and confidence band. For families with identity link on the mean (Gaussian, Student-t), the two are equivalent. For log-link families (Gamma, lognormal, Poisson, nbinom2) "response" reports back-transformed means; "link" reports the log-scale linear predictor. For the logit-link Beta family, "response" reports proportions and "link" reports log-odds.

ci_method

Confidence-interval method. Defaults to the ci_method stored on x$metadata$ci_method so the band matches the symbolize() call. emmeans currently produces asymptotic Wald-style intervals regardless of this argument, but the column is propagated for consistency with parameter_interpretation().

...

Reserved for future use.

Value

A tibble (S3 class symbolizer_group_means) with one row per level combination. Columns: level_combo, one column per factor in by, estimate, std_error, confint_low, confint_high, excludes_zero, ci_method, scale.

See also

Other marginal estimates: group_contrasts(), group_slopes()