Getting Started with broom.helpers

The broom.helpers package offers a suite of functions that make easy to interact, add information, and manipulate tibbles created with broom::tidy() (and friends).

The suite includes functions to group regression model terms by variable, insert reference and header rows for categorical variables, add variable labels, and more.

As a motivating example, let’s summarize a logistic regression model with a forest plot and in a table.

Our model predicts tumor response using chemotherapy treatment and tumor grade. The data set we’re utilizing has already labelled the columns using the labelled package. The column labels will be carried through to our figure and table.

Forest Plot

To create the figure, we’ll need to add some information to the tidy tibble, i.e. we’ll need to group the terms that belong to the same variable, add the reference row, etc. Parsing this information can be difficult, but the broom.helper package has made it simple.

tidy_forest <-
  model_logit |>
  # perform initial tidying of the model
  tidy_and_attach(exponentiate = TRUE, conf.int = TRUE) |>
  # adding in the reference row for categorical variables
  tidy_add_reference_rows() |>
  # adding a reference value to appear in plot
  tidy_add_estimate_to_reference_rows() |>
  # adding the variable labels
  tidy_add_term_labels() |>
  # removing intercept estimate from model
  tidy_remove_intercept()
tidy_forest
#> # A tibble: 5 × 16
#>   term      variable var_label          var_class var_type var_nlevels contrasts
#>   <chr>     <chr>    <chr>              <chr>     <chr>          <int> <chr>    
#> 1 trtDrug A trt      Chemotherapy Trea… character dichoto…           2 contr.tr…
#> 2 trtDrug B trt      Chemotherapy Trea… character dichoto…           2 contr.tr…
#> 3 gradeI    grade    Grade              factor    categor…           3 contr.tr…
#> 4 gradeII   grade    Grade              factor    categor…           3 contr.tr…
#> 5 gradeIII  grade    Grade              factor    categor…           3 contr.tr…
#> # ℹ 9 more variables: contrasts_type <chr>, reference_row <lgl>, label <chr>,
#> #   estimate <dbl>, std.error <dbl>, statistic <dbl>, p.value <dbl>,
#> #   conf.low <dbl>, conf.high <dbl>

Note: we used tidy_and_attach() instead of broom::tidy(). broom.helpers functions needs a copy of the original model. To avoid passing the model at each step, the easier way is to attach the model as an attribute of the tibble with tidy_attach_model(). tidy_and_attach() is simply a shortcut of model |> broom::tidy() |> tidy_and_attach(model).

We now have a tibble with every piece of information we need to create our forest plot using ggplot2.

tidy_forest |>
  mutate(
    plot_label = paste(var_label, label, sep = ":") |>
      forcats::fct_inorder() |>
      forcats::fct_rev()
  ) |>
  ggplot(aes(x = plot_label, y = estimate, ymin = conf.low, ymax = conf.high, color = variable)) +
  geom_hline(yintercept = 1, linetype = 2) +
  geom_pointrange() +
  coord_flip() +
  theme(legend.position = "none") +
  labs(
    y = "Odds Ratio",
    x = " ",
    title = "Forest Plot using broom.helpers"
  )

Note:: for more advanced and nicely formatted plots of model coefficients, look at ggstats::ggcoef_model() and its dedicated vignette. ggstats::ggcoef_model() internally uses broom.helpers.

Table Summary

In addition to aiding in figure creation, the broom.helpers package can help summarize a model in a table. In the example below, we add header and reference rows, and utilize existing variable labels. Let’s change the labels shown in our summary table as well.

tidy_table <-
  model_logit |>
  # perform initial tidying of the model
  tidy_and_attach(exponentiate = TRUE, conf.int = TRUE) |>
  # adding in the reference row for categorical variables
  tidy_add_reference_rows() |>
  # adding the variable labels
  tidy_add_term_labels() |>
  # add header row
  tidy_add_header_rows() |>
  # removing intercept estimate from model
  tidy_remove_intercept()

# print summary table
options(knitr.kable.NA = "")
tidy_table |>
  # format model estimates
  select(label, estimate, conf.low, conf.high, p.value) |>
  mutate(across(all_of(c("estimate", "conf.low", "conf.high")), style_ratio)) |>
  mutate(across(p.value, style_pvalue)) |>
  print_table()

Note:: for more advanced and nicely formatted tables of model coefficients, look at gtsummary::tbl_regression() and its dedicated vignette. gtsummary::tbl_regression() internally uses broom.helpers.

All-in-one function

There is also a handy wrapper, called tidy_plus_plus(), for the most commonly used tidy_*() functions, and they can be executed with a single line of code:

model_logit |>
  tidy_plus_plus(exponentiate = TRUE)
#> # A tibble: 5 × 18
#>   term      variable var_label          var_class var_type var_nlevels contrasts
#>   <chr>     <chr>    <chr>              <chr>     <chr>          <int> <chr>    
#> 1 trtDrug A trt      Chemotherapy Trea… character dichoto…           2 contr.tr…
#> 2 trtDrug B trt      Chemotherapy Trea… character dichoto…           2 contr.tr…
#> 3 gradeI    grade    Grade              factor    categor…           3 contr.tr…
#> 4 gradeII   grade    Grade              factor    categor…           3 contr.tr…
#> 5 gradeIII  grade    Grade              factor    categor…           3 contr.tr…
#> # ℹ 11 more variables: contrasts_type <chr>, reference_row <lgl>, label <chr>,
#> #   n_obs <dbl>, n_event <dbl>, estimate <dbl>, std.error <dbl>,
#> #   statistic <dbl>, p.value <dbl>, conf.low <dbl>, conf.high <dbl>

model_logit |>
  tidy_plus_plus(exponentiate = TRUE) |>
  print_table()

See the documentation of tidy_plus_plus() for the full list of available options.

Advanced examples

broom.helpers can also handle different contrasts for categorical variables and the use of polynomial terms for continuous variables.

Polynomial terms

When polynomial terms of a continuous variable are defined with stats::poly(), broom.helpers will be able to identify the corresponding variable, and add header rows.

model_poly <- glm(response ~ poly(age, 3) + ttdeath, na.omit(trial), family = binomial)

model_poly |>
  tidy_plus_plus(
    exponentiate = TRUE,
    add_header_rows = TRUE,
    variable_labels = c(age = "Age in years")
  ) |>
  print_table()

You also have an option to relabel polynomial terms. Be aware that, by default, stats::poly() generates orthogonal polynomials. Relabels are more appropriate with raw polynomials.

model_poly2 <- glm(response ~ poly(age, 3, raw = TRUE) + ttdeath, na.omit(trial), family = binomial)

model_poly2 |>
  tidy_plus_plus(
    exponentiate = TRUE,
    add_header_rows = TRUE,
    variable_labels = c(age = "Age in years"),
    relabel_poly = TRUE
  ) |>
  print_table()

Different type of contrasts

By default, categorical variables are coded with a treatment contrasts (see stats::contr.treatment()). With such contrasts, model coefficients correspond to the effect of a modality compared with the reference modality (by default, the first one). tidy_add_reference_rows() allows to add a row for this reference modality and tidy_add_estimate_to_reference_rows() will populate the estimate value of these references rows by 0 (or 1 if exponentiate = TRUE). tidy_add_term_labels() is able to retrieve the label of the factor level associated with a specific model term.

model_1 <- glm(
  response ~ stage + grade * trt,
  gtsummary::trial,
  family = binomial
)

model_1 |>
  tidy_and_attach(exponentiate = TRUE) |>
  tidy_add_reference_rows() |>
  tidy_add_estimate_to_reference_rows(exponentiate = TRUE) |>
  tidy_add_term_labels() |>
  print_table()

Using stats::contr.treatment(), it is possible to defined alternative reference rows. It will be properly managed by broom.helpers.

model_2 <- glm(
  response ~ stage + grade * trt,
  gtsummary::trial,
  family = binomial,
  contrasts = list(
    stage = contr.treatment(4, base = 3),
    grade = contr.treatment(3, base = 2),
    trt = contr.treatment(2, base = 2)
  )
)

model_2 |>
  tidy_and_attach(exponentiate = TRUE) |>
  tidy_add_reference_rows() |>
  tidy_add_estimate_to_reference_rows(exponentiate = TRUE) |>
  tidy_add_term_labels() |>
  print_table()

You can also use sum contrasts (cf. stats::contr.sum()). In that case, each model coefficient corresponds to the difference of that modality with the grand mean. A variable with 4 modalities will be coded with 3 terms. However, a value could be computed (using emmeans::emmeans()) for the last modality, corresponding to the difference of that modality with the grand mean and equal to sum of all other coefficients multiplied by -1. broom.helpers will identify categorical variables coded with sum contrasts and could retrieve an estimate value for the reference term.

model_3 <- glm(
  response ~ stage + grade * trt,
  gtsummary::trial,
  family = binomial,
  contrasts = list(
    stage = contr.sum,
    grade = contr.sum,
    trt = contr.sum
  )
)

model_3 |>
  tidy_and_attach(exponentiate = TRUE) |>
  tidy_add_reference_rows() |>
  tidy_add_estimate_to_reference_rows(exponentiate = TRUE) |>
  tidy_add_term_labels() |>
  print_table()

Other types of contrasts exist, like Helmert (contr.helmert()) or polynomial (contr.poly()). They are more complex as a modality will be coded with a combination of terms. Therefore, for such contrasts, it will not be possible to associate a specific model term with a level of the original factor. broom.helpers will not add a reference term in such case.

model_4 <- glm(
  response ~ stage + grade * trt,
  gtsummary::trial,
  family = binomial,
  contrasts = list(
    stage = contr.poly,
    grade = contr.helmert,
    trt = contr.poly
  )
)

model_4 |>
  tidy_and_attach(exponentiate = TRUE) |>
  tidy_add_reference_rows() |>
  tidy_add_estimate_to_reference_rows(exponentiate = TRUE) |>
  tidy_add_term_labels() |>
  print_table()

Pairwise contrasts of categorical variable

Pairwise contrasts of categorical variables could be computed with tidy_add_pairwise_contrasts().

model_logit <- glm(response ~ age + trt + grade, trial, family = binomial)

model_logit |>
  tidy_and_attach() |>
  tidy_add_pairwise_contrasts() |>
  print_table()


model_logit |>
  tidy_and_attach(exponentiate = TRUE) |>
  tidy_add_pairwise_contrasts() |>
  print_table()


model_logit |>
  tidy_and_attach(exponentiate = TRUE) |>
  tidy_add_pairwise_contrasts(pairwise_reverse = FALSE) |>
  print_table()


model_logit |>
  tidy_and_attach(exponentiate = TRUE) |>
  tidy_add_pairwise_contrasts(keep_model_terms = TRUE) |>
  print_table()

Column Details

Below is a summary of the additional columns that may be added by a broom.helpers function. The table includes the column name, the function that adds the column, and a short description of the information in the column.

Column	Function	Description
contrasts	`tidy_add_contrasts()`	Contrasts used for categorical variables. Require “variable” column. If needed, will automatically apply `tidy_identify_variables()`.
contrasts_type	`tidy_add_contrasts()`	Type of contrasts (“treatment”, “sum”, “poly”, “helmert”, “sdif”, “other” or “no.contrast”). “pairwise is used for pairwise contrasts computed with `tidy_add_pairwise_contrasts()`.
exposure	`tidy_add_n()`	Exposure time (for Poisson and Cox models)
group_by	`tidy_group_by()`	Grouping variable (particularly for multinomial or multi-components models).
header_row	`tidy_add_header_rows()`	Logical indicating if a row is a header row for variables with several terms. Is equal to `NA` for variables who do not have an header row. Require “label” column. If needed, will automatically apply `tidy_add_term_labels()`. It is better to apply `tidy_add_header_rows()` after other `tidy_*` functions
instrumental	`tidy_identify_variables()`	For “fixest” models, indicate if a variable was instrumental.
label	`tidy_add_term_labels()`	String of term labels based on (1) labels provided in `labels` argument if provided; (2) factor levels for categorical variables coded with treatment, SAS or sum contrasts; (3) variable labels when there is only one term per variable; and (4) term name otherwise. Require “variable_label” column. If needed, will automatically apply `tidy_add_variable_labels()`. Require “contrasts” column. If needed, will automatically apply `tidy_add_contrasts()`.
n_event	`tidy_add_n()`	Number of events (for binomial and multinomial logistic models, Poisson and Cox models)
n_ind	`tidy_add_n()`	Number of individuals (for Cox models)
n_obs	`tidy_add_n()`	Number of observations
original_term	`tidy_disambiguate_terms()`, `tidy_multgee()`, `tidy_zeroinfl()` or `tidy_identify_variables()`	Original term before disambiguation. This columns is added only when disambiguation is needed (i.e. for mixed models). Also used for “multgee”, “zeroinfl” and “hurdle” models. For instrumental variables in “fixest” models, the “fit_” prefix is removed, and the original terms is stored in this column.
reference_row	`tidy_add_reference_rows()`	Logical indicating if a row is a reference row for categorical variables using a treatment or a sum contrast. Is equal to `NA` for variables who do not have a reference row. Require “contrasts” column. If needed, will automatically apply `tidy_add_contrasts()`. `tidy_add_reference_rows()` will not populate the label of the reference term. It is therefore better to apply `tidy_add_term_labels()` after `tidy_add_reference_rows()` rather than before.
var_class	`tidy_identify_variables()`	Class of the variable.
var_label	`tidy_add_variable_labels()`	String of variable labels from the model. Columns labelled with the `labelled` package are retained. It is possible to pass a custom label for an interaction term with the `labels` argument. Require “variable” column. If needed, will automatically apply `tidy_identify_variables()`.
var_nlevels	`tidy_identify_variables()`	Number of original levels for categorical variables
var_type	`tidy_identify_variables()`	One of “intercept”, “continuous”, “dichotomous”, “categorical”, “interaction”, “ran_pars” or “ran_vals”
variable	`tidy_identify_variables()`	String of variable names from the model. For categorical variables and polynomial terms defined with `stats::poly()`, terms belonging to the variable are identified.

Note: tidy_add_estimate_to_reference_rows() does not create an additional column; rather, it populates the ‘estimate’ column for reference rows.

Additional attributes

Below is a list of additional attributes that broom.helpers may attached to the results. The table includes the attribute name, the function that adds the attribute, and a short description.

Attribute	Function	Description
coefficients_label	`tidy_add_coefficients_type()`	Coefficients label
coefficients_type	`tidy_add_coefficients_type()`	Type of coefficients
component	`tidy_zeroinfl()`	`component` argument passed to `tidy_zeroinfl()`
conf.level	`tidy_and_attach()`	Level of confidence used for confidence intervals
exponentiate	`tidy_and_attach()`	Indicates if estimates were exponentiated
Exposure	`tidy_add_n()`	Total of exposure time
N_event	`tidy_add_n()`	Total number of events
N_ind	`tidy_add_n()`	Total number of individuals (for Cox models)
N_obs	`tidy_add_n()`	Total number of observations
term_labels	`tidy_add_term_labels()`	Custom term labels passed to `tidy_add_term_labels()`
variable_labels	`tidy_add_variable_labels()`	Custom variable labels passed to `tidy_add_variable_labels()`

Supported models

Model	Notes
`betareg::betareg()`	Use `tidy_parameters()` as `tidy_fun` with `component` argument to control with coefficients to return. `broom::tidy()` does not support the `exponentiate` argument for betareg models, use `tidy_parameters()` instead.
`biglm::bigglm()`
`brms::brm()`	`broom.mixed` package required
`cmprsk::crr()`	Limited support. It is recommended to use `tidycmprsk::crr()` instead.
`fixest::feglm()`	May fail with R <= 4.0.
`fixest::femlm()`	May fail with R <= 4.0.
`fixest::feNmlm()`	May fail with R <= 4.0.
`fixest::feols()`	May fail with R <= 4.0.
`gam::gam()`
`geepack::geeglm()`
`glmmTMB::glmmTMB()`	`broom.mixed` package required
`glmtoolbox::glmgee()`
`lavaan::lavaan()`	Limited support for categorical variables
`lfe::felm()`
`lme4::glmer.nb()`	`broom.mixed` package required
`lme4::glmer()`	`broom.mixed` package required
`lme4::lmer()`	`broom.mixed` package required
`logitr::logitr()`	Requires logitr >= 0.8.0
`MASS::glm.nb()`
`MASS::polr()`
`mgcv::gam()`	Use default tidier `broom::tidy()` for smooth terms only, or `gtsummary::tidy_gam()` to include parametric terms
`mice::mira`	Limited support. If `mod` is a `mira` object, use `tidy_fun = function(x, ...) {mice::pool(x) \|> mice::tidy(...)}`
`mmrm::mmrm()`
`multgee::nomLORgee()`	Use `tidy_multgee()` as `tidy_fun`.
`multgee::ordLORgee()`	Use `tidy_multgee()` as `tidy_fun`.
`nnet::multinom()`
`ordinal::clm()`	Limited support for models with nominal predictors.
`ordinal::clmm()`	Limited support for models with nominal predictors.
`parsnip::model_fit`	Supported as long as the type of model and the engine is supported.
`plm::plm()`
`pscl::hurdle()`	Use `tidy_zeroinfl()` as `tidy_fun`.
`pscl::zeroinfl()`	Use `tidy_zeroinfl()` as `tidy_fun`.
`rstanarm::stan_glm()`	`broom.mixed` package required
`stats::aov()`	Reference rows are not relevant for such models.
`stats::glm()`
`stats::lm()`
`stats::nls()`	Limited support
`survey::svycoxph()`
`survey::svyglm()`
`survey::svyolr()`
`survival::cch()`	Experimental support.
`survival::clogit()`
`survival::coxph()`
`survival::survreg()`
`svyVGAM::svy_vglm()`	Experimental support. It is recommended to use `tidy_svy_vglm()` as `tidy_fun`.
`tidycmprsk::crr()`
`VGAM::vgam()`	Experimental support. It is recommended to use `tidy_vgam()` as `tidy_fun`.
`VGAM::vglm()`	Experimental support. It is recommended to use `tidy_vgam()` as `tidy_fun`.

Note: this list of models has been tested. broom.helpers may or may not work properly or partially with other types of models. Do not hesitate to provide feedback on GitHub.