Skip to contents

Structured result container for GLM fits

Source: R/model-glm.R
class_glm_object.Rd

An S7 class produced by GLM pipelines. Not constructed manually — use define_model() |> prepare_model(GLM) |> conclude() instead.

Inherits from anova_able, so it participates in anova() directly. Downstream packages can use it as a parent in S7::new_class().

Details

Constructor arguments (populated automatically by GLM):

  • terms: model terms object.

  • df_residual: residual degrees of freedom.

  • deviance: scalar deviance.

  • dispersion: scalar dispersion parameter.

  • family: string naming the error family, e.g. "binomial".

  • link: string naming the link function, e.g. "logit".

  • null_deviance: scalar deviance of the intercept-only model.

  • aic: scalar AIC.

  • logLik: scalar log-likelihood of the fitted model.

  • null_logLik: scalar log-likelihood of the intercept-only model.

  • beta: named numeric vector of coefficient estimates.

  • std_beta: named numeric vector of coefficient standard errors.

  • actual: numeric vector of the original values on the response scale.

  • fitted: numeric vector of fitted values on the response scale.

  • vcov: variance-covariance matrix of the coefficients, e.g. stats::vcov(fit). Required for predict() with interval.

  • x_mat: model matrix stored as a flat numeric vector via as.numeric(stats::model.matrix(fit)). Required for predict().

  • x_levels: factor levels used when fitting, via stats::.getXlevels(fit$terms, stats::model.frame(fit)). Required for predict() on new data with factor predictors.

The following are computed automatically and do not need to be supplied:

  • statistic: per-coefficient test statistics (beta / std_beta).

  • p_value: per-coefficient two-sided p-values. Uses a z-test when family is "binomial" or "poisson" (fixed dispersion), and a t-test against df_residual otherwise (estimated dispersion).

  • coefficients: tibble with columns term, estimate, std_error, statistic, p_value.

  • fit_summary: tibble with columns family, link, null_deviance, deviance, df_residual, aic, n_obs.

predict() arguments

predict() on a class_glm_object accepts:

  • new_data: A data frame of new predictors. NULL (the default) returns fitted values and response-based truth for the training data.

  • type: One of "response" (default, back-transformed through the inverse link) or "link" (linear predictor scale).

  • interval: One of "none" (default) or "confidence". Prediction intervals are not available, since GLMs have no closed-form analogue of OLS prediction error.

  • level: Confidence level for the interval. Default 0.95.

See also

anova_able, GLM

Examples

# Inheriting from class_glm_object in a downstream package:
my_glm = S7::new_class(
    "my_glm",
    parent = class_glm_object
)

# Populating class_glm_object from a fitted glm (as done internally):
fit = glm(am ~ wt + hp, data = mtcars, family = binomial())
s = summary(fit)
fam = fit$family$family

obj = class_glm_object(
    terms = fit$terms,
    df_residual = fit$df.residual,
    deviance = fit$deviance,
    dispersion = if (fam %in% c("binomial", "poisson")) 1 else s$dispersion,
    family = fam,
    link = fit$family$link,
    null_deviance = fit$null.deviance,
    aic = fit$aic,
    beta = coef(s)[, 1],
    std_beta = coef(s)[, 2],
    actual = unname(fit$y),
    fitted = unname(fit$fitted.values),
    vcov = vcov(fit),
    x_mat = as.numeric(model.matrix(fit)),
    x_levels = .getXlevels(fit$terms, model.frame(fit))
)

obj@coefficients
#> # A tibble: 3 × 5
#>   term        estimate std_error statistic p_value
#>   <chr>          <dbl>     <dbl>     <dbl>   <dbl>
#> 1 (Intercept)  18.9       7.44        2.53 0.0113 
#> 2 wt           -8.08      3.07       -2.63 0.00843
#> 3 hp            0.0363    0.0177      2.04 0.0409 
obj@fit_summary
#> # A tibble: 1 × 7
#>   family   link  null_deviance deviance df_residual   aic n_obs
#>   <chr>    <chr>         <dbl>    <dbl>       <int> <dbl> <int>
#> 1 binomial logit          43.2     10.1          29  16.1    32