The modelsummary package for R produces beautiful, customizable, publication-ready tables to summarize statistical models. Results from several models are presented side-by-side, with uncertainty estimates in parentheses (or brackets) underneath coefficient estimates. Tables can be saved to HTML, LaTeX and RTF (MS Word-ready) formats, or they can be fed to a dynamic report pipeline like knitr or Sweave.
- Sales pitch
- Installation
- A simple example
- Customizing your tables
- A complex example
- Other useful features
- Alternative summary table packages for R
Here are a few benefits of modelsummary over some alternative packages:
- Customizability
- Tables are endlessly customizable, thanks to the power of the
gtpackage. In this README, you will find tables with colored cells, weird text, spanning column labels, row groups, titles and subtitles, global footnotes, cell-specific footnotes, significance stars, etc. This only scratches the surface of possibilities. For more, see gt.rstudio.com and the Power Users section of this README.
- Tables are endlessly customizable, thanks to the power of the
- Flexibility
- Tables can be saved to html, rtf, jpeg, png, pdf, or LaTeX files. (Coming soon: TXT/ASCII, and more.)
- Integration
modelsummaryis extremely well integrated with RStudio. When you typemsummary(models), the summary table immediately appears in the Viewer window.
- Transparency, replicability, and automation
- By combining
knitrandmodelsummary, you can easily produce beautiful, replicable, and automated documents and reports. Click here for details.
- By combining
- Community
modelsummarydoes not try to do everything. It leverages the incredible work of theRcommunity by building on top of the popularbroompackage. Thanks to thebroomteam,modelsummaryalready supports dozens of model types out of the box. Most importantly, asbroomandgtimprove,modelsummaryalso improves.
- Reliability
modelsummaryis developed using a suite of unit tests. It (probably) won't break.
- Simplicity
- By using the
broomandgtpackages for key operations,modelsummaryhas a massively simplified codebase. This should improve long term code maintainability, and allow contributors to participate through GitHub.
- By using the
CFITCRS!
At the modelsummary factory, we are serious about customizability. Are your bored of regression tables with good ol' "Intercept"? If so, we have a solution for you:
You can install modelsummary from CRAN:
install.packages('modelsummary')If you want the very latest version, install it from Github:
library(remotes)
remotes::install_github('vincentarelbundock/modelsummary')modelsummary relies heavily on the gt package, which is not available on CRAN yet. You can install it from github:
remotes::install_github('rstudio/gt')Make sure you also install tidyverse, as modelsummary depends on a lot of its packages (e.g., stringr, dplyr, tidyr, purrr):
install.packages('tidyverse')Load packages and download some data from the RDatasets repository. Then, estimate 5 different models and store them in a named list. The name of each model in that list will be used as a column label:
library(gt)
library(MASS)
library(modelsummary)
url <- 'https://vincentarelbundock.github.io/Rdatasets/csv/HistData/Guerry.csv'
dat <- read.csv(url)
dat$Clergy <- ifelse(dat$Clergy > 40, 1, 0) # binary variable for logit model
models <- list()
models[['OLS 1']] <- lm(Literacy ~ Crime_prop + Infants, dat)
models[['NBin 1']] <- glm.nb(Literacy ~ Crime_prop + Donations, dat)
models[['OLS 2']] <- lm(Desertion ~ Crime_prop + Infants, dat)
models[['NBin 2']] <- glm.nb(Desertion ~ Crime_prop + Donations, dat)
models[['Logit 1']] <- glm(Clergy ~ Crime_prop + Infants, dat, family = binomial())Produce a simple table:
msummary(models)Of course, modelsummary can also summarize single models:
mod <- lm(Clergy ~ Crime_prop, data = dat)
msummary(mod)modelsummary prints an uncertainty estimate in parentheses below the corresponding coefficient estimate. The statistic argument must be a string which is equal to conf.int or to one of the columns produced by the broom::tidy function. When using conf.int, users can specify a confidence level with the conf_level argument.
msummary(models, statistic = 'std.error')
msummary(models, statistic = 'p.value')
msummary(models, statistic = 'statistic')
msummary(models, statistic = 'conf.int', conf_level = .99)Display the uncertainty estimate next to the coefficient instead of below it:
msummary(models, statistic_vertical = FALSE)You can override the uncertainty estimates in a number of ways. First, you can specify a function that produces variance-covariance matrices:
library(sandwich)
msummary(models, statistic_override = vcovHC, statistic = 'p.value')You can supply a list of functions of the same length as your model list:
msummary(models,
statistic_override = list(vcov, vcovHC, vcovHAC, vcovHC, vcov))You can supply a list of named variance-covariance matrices:
vcov_matrices <- lapply(models, vcovHC)
msummary(models, statistic_override = vcov_matrices)You can supply a list of named vectors:
custom_stats <- list(`OLS 1` = c(`(Intercept)` = 2, Crime_prop = 3, Infants = 4),
`NBin 1` = c(`(Intercept)` = 3, Crime_prop = -5, Donations = 3),
`OLS 2` = c(`(Intercept)` = 7, Crime_prop = -6, Infants = 9),
`NBin 2` = c(`(Intercept)` = 4, Crime_prop = -7, Donations = -9),
`Logit 1` = c(`(Intercept)` = 1, Crime_prop = -5, Infants = -2))
msummary(models, statistic_override = custom_stats)You can also display several different uncertainty estimates below the coefficient estimates. For example,
msummary(models, statistic = c('std.error', 'p.value', 'conf.int'))Will produce something like this:
You can add titles and subtitles to your table as follows:
msummary(models,
title = 'This is a title for my table.',
subtitle = 'And this is the subtitle.')Add notes to the bottom of your table:
msummary(models,
notes = list('Text of the first note.',
'Text of the second note.'))Add numbered footnotes to a column, a row, or a cell:
msummary(models) %>%
tab_footnote(
footnote = md("This is a **very** important model, so we are pointing it out in a column-specific footnote."),
locations = cells_column_labels(columns = vars(`OLS 1`))) %>%
tab_footnote(
footnote = "This is the variable of interest.",
locations = cells_body(columns = 1, rows = 3)) %>%
tab_footnote(
footnote = "Most important model + most important variable = most important estimate.",
locations = cells_body(columns = vars(`OLS 1`), rows = 3))The coef_map argument is a named vector which allows users to rename, reorder, and subset coefficient estimates. Values of this vector correspond to the "clean" variable name. Names of this vector correspond to the "raw" variable name. The table will be sorted in the order in which terms are presented in coef_map. Coefficients which are not included in coef_map will be excluded from the table.
cm <- c('Crime_prop' = 'Crime / Population',
'Donations' = 'Donations',
'(Intercept)' = 'Constant')
msummary(models, coef_map = cm)An alternative mechanism to subset coefficients is to use the coef_omit argument. This string is a regular expression which will be fed to stringr::str_detect to detect the variable names which should be excluded from the table.
msummary(models, coef_omit = 'Intercept|Donation')gof_omit is a regular expression which will be fed to stringr::str_detect to detect the names of the statistics which should be excluded from the table.
msummary(models, gof_omit = 'DF|Deviance')A more powerful mechanism is to supply a data.frame (or tibble) through the gof_map argument. This data.frame must include 4 columns:
raw: a string with the name of a column produced bybroom::glance(model).clean: a string with the "clean" name of the statistic you want to appear in your final table.fmt: a string which will be used to round/format the string in question (e.g.,"%.3f"). This follows the same standards as thefmtargument in?modelsummary.omit:TRUEif you want the statistic to be omitted from your final table.
You can see an example of a valid data frame by typing modelsummary::gof_map. This is the default data.frame that modelsummary uses to subset and reorder goodness-of-fit statistics. As you can see, omit == TRUE for quite a number of statistics. You can include setting omit == FALSE:
gm <- modelsummary::gof_map
gm$omit <- FALSE
msummary(models, gof_map = gm)The goodness-of-fit statistics will be printed in the table in the same order as in the gof_map data.frame.
Notice the subtle difference between coef_map and gof_map. coef_map works as a "white list": any coefficient not explicitly entered will be omitted from the table. gof_map works as a "black list": statistics need to be explicitly marked for omission.
Create spanning labels to group models (columns):
msummary(models) %>%
tab_spanner(label = 'Literacy', columns = c('OLS 1', 'NBin 1')) %>%
tab_spanner(label = 'Desertion', columns = c('OLS 2', 'NBin 2')) %>%
tab_spanner(label = 'Clergy', columns = 'Logit 1')Some people like to add "stars" to their model summary tables to mark statistical significance. The stars argument can take three types of input:
NULLomits any stars or special marks (default)TRUEuses these default values:* p < 0.1, ** p < 0.05, *** p < 0.01- Named numeric vector for custom stars.
msummary(models)
msummary(models, stars = TRUE)
msummary(models, stars = c('+' = .1, '*' = .01)) Whenever stars != NULL, modelsummary adds a note at the bottom of the table automatically. If you would like to omit this note, just use the stars_note argument:
msummary(models, stars = TRUE, stars_note = FALSE) If you want to create your own stars description, you can add custom notes with the notes argument.
The fmt argument defines how numeric values are rounded and presented in the table. This argument follows the sprintf C-library standard. For example,
%.3fwill keep 3 digits after the decimal point, including trailing zeros.%.5fwill keep 5 digits after the decimal point, including trailing zeros.- Changing the
ffor anewill use the exponential decimal representation.
Most users will just modify the 3 in %.3f, but this is a very powerful system, and all users are encouraged to read the details: ?sprintf
msummary(models, fmt = '%.7f')The power of the gt package makes modelsummary tables endlessly customizable. For instance, we can color columns and cells, and present values in bold or italics:
msummary(models) %>%
tab_style(style = cell_fill(color = "lightcyan"),
locations = cells_body(columns = vars(`OLS 1`))) %>%
tab_style(style = cell_fill(color = "#F9E3D6"),
locations = cells_data(columns = vars(`NBin 2`), rows = 2:6)) %>%
tab_style(style = cell_text(weight = "bold"),
locations = cells_body(columns = vars(`OLS 1`))) %>%
tab_style(style = cell_text(style = "italic"),
locations = cells_data(columns = vars(`NBin 2`), rows = 2:6))
Thanks to gt, modelsummary accepts markdown indications for emphasis and more:
msummary(models,
title = md('This is a **bolded series of words.**'),
notes = list(md('And an *emphasized note*.')))This will produce a table with extra large variable names.
msummary(models) %>%
tab_style(style = cell_text(size = 'x-large'),
locations = cells_body(columns = 1)) Note that gt's tab_style function is more developed for HTML output than for RTF or LaTeX, so some styling options may not be availble yet. The gt package is under heavy development, so feel free to file an issue on github if you have a special request, and stay tuned for more!
Use the add_rows argument to add rows manually to the bottom of the table.
row1 <- c('Custom row 1', 'a', 'b', 'c', 'd', 'e')
row2 <- c('Custom row 2', 5:1)
msummary(models, add_rows = list(row1, row2))Use the add_rows argument to specify where the custom rows should be displayed in the bottom panel. For example, this prints custom rows after the coefficients, but at first position in the goodness of fit measures:
msummary(models, add_rows = list(row1, row2), add_rows_location = 0)This prints custom rows after the 2nd GOF statistic:
msummary(models, add_rows = list(row1, row2), add_rows_location = 2)Insert images in your tables using the gt::text_transform and gt::local_image functions.
msummary(models) %>%
text_transform(
locations = cells_body(columns = 1, rows = 1),
fn = function(x) {web_image(url = "https://raw.githubusercontent.com/vincentarelbundock/modelsummary/master/examples/squirrel.png", height = 120)}
)This is the code I used to generate the "complex" table posted at the top of this README.
cm <- c('Crime_prop' = 'Crime / Population',
'Donations' = 'Donations',
'Infants' = 'Infants',
'(Intercept)' = 'Constant')
msummary(models,
coef_map = cm,
stars = TRUE,
gof_omit = "Deviance",
title = 'Summarizing 5 statistical models using the `modelsummary` package for `R`.',
subtitle = 'Models estimated using the Guerry dataset.',
notes = c('First custom note to contain text.',
'Second custom note with different content.')) %>%
# add spanning labels
tab_spanner(label = 'Literacy', columns = c('OLS 1', 'NBin 1')) %>%
tab_spanner(label = 'Desertion', columns = c('OLS 2', 'NBin 2')) %>%
tab_spanner(label = 'Clergy', columns = 'Logit 1') %>%
# footnotes
tab_footnote(
footnote = md("This is a **very** important model, so we are pointing it out in a column-specific footnote."),
locations = cells_column_labels(columns = vars(`OLS 1`))) %>%
tab_footnote(
footnote = "This is the variable of interest.",
locations = cells_body(rows =5)) %>%
tab_footnote(
footnote = "Most important model + most important variable = most important estimate.",
locations = cells_body(columns = vars(`OLS 1`), rows = 5)) %>%
# color and bold
tab_style(style = cell_text(color = "red", weight = "bold"),
locations = cells_body(columns = vars(`OLS 1`), rows = 5))To save a table to file, use the filename argument. modelsummary guesses the output format based on the filename extension. The supported extensions are: .tex, .rtf, .html (ASCII/Text tables coming soon).
msummary(models, filename = 'table.tex')
msummary(models, filename = 'table.rtf')
msummary(models, filename = 'table.html')
msummary(models, filename = 'table.jpeg')
msummary(models, filename = 'table.png')If filename is not specified, modelsummary returns a gt object which can be further customized and rendered by the gtsave function from the gt package. RStudio renders the html version of this object automatically.
Warning: When creating complex tables by chaining multiple gt functions with the %>% pipe operator, the filename argument will not work. The problem is that modelsummary is trying to write-to-file immediately at the main msummary() call, before the rest of the functions in the chain are executed. In that case, it is better to use gt::gtsave explicitly at the very end of your chain. For example,
msummary(models) %>%
tab_spanner(label = 'Literacy', columns = c('OLS 1', 'NBin 1')) %>%
tab_spanner(label = 'Desertion', columns = c('OLS 2', 'NBin 2')) %>%
tab_spanner(label = 'Clergy', columns = 'Logit 1') %>%
gtsave('table.tex')You can use modelsummary to produce LaTeX tables and to create dynamic documents with knitr. When knitting in html format, adding a msummary(models) call to a code chunk should work out of the box.
When creating LaTeX tables to generate PDF document, things are slightly different. Indeed, the gt output functionality for LaTeX is still in development and it is somewhat limited. To avoid common sources of compilation errors, and to allow users to use \label{}, modelsummary includes two convencience function:
clean_latexreturns a LaTeX table as a string.knit_latexreturns an object of typeas_is, which can be used directly by theknitrpackage.
For instance, this code will produce a LaTeX table in as a string object:
msummary(models, title = 'Model summary') %>%
clean_latex(label = 'tab:example')This code will produce a table inside a knitr PDF document:
msummary(models, title = 'Model summary') %>%
knit_latex(label = 'tab:example')Please note that the tables produced by modelsummary require the following LaTeX packages to compile: caption, longtable, booktabs. You will need to include those in your header or preamble if you want documents to compile properly.
Here are two minimal working examples of markdown files which can be converted to HTML or PDF using the knitr package. Just open one the .Rmd files in RStudio and click the "Knit" button:
My goal is to deprecate the clean_latex and knit_latex functions when gt LaTeX export features improve.
modelsummary relies on two functions from the broom package to extract model information: tidy and glance. If broom doesn't support the type of model you are trying to summarize, modelsummary won't support it out of the box. Thankfully, it is extremely easy to add support for most models using custom methods.
For example, models produced by the MCMCglmm package are not currently supported by broom. To add support, you simply need to create a tidy and a glance method:
# load packages and data
library(modelsummary)
library(MCMCglmm)
data(PlodiaPO)
# add custom functions to extract estimates (tidy) and goodness-of-fit (glance) information
tidy.MCMCglmm <- function(object, ...) {
s <- summary(object, ...)
ret <- tibble::tibble(term = row.names(s$solutions),
estimate = s$solutions[, 1],
conf.low = s$solutions[, 2],
conf.high = s$solutions[, 3])
ret
}
glance.MCMCglmm <- function(object, ...) {
ret <- tibble::tibble(dic = object$DIC,
n = nrow(object$X))
ret
}
# estimate a simple model
model <- MCMCglmm(PO ~ 1 + plate, random = ~ FSfamily, data = PlodiaPO, verbose=FALSE, pr=TRUE)
# summarize the model
msummary(model, statistic = 'conf.int')Two important things to note. First, the methods are named tidy.MCMCglmm and glance.MCMCglmm because the model object I am trying to summarize is of class MCMCglmm. You can find the class of a model by running: class(model).
Second, in the example above, we used the statistic = 'conf.int' argument. This is because the tidy method produces conf.low and conf.high columns. In most cases, users will define std.error column in their custom tidy methods, so the statistic argument will need to be adjusted.
If you create new tidy and glance methods, please consider contributing them to broom so that the rest of the community can benefit from your work: https://github.com/tidymodels/broom
modelsummary can pool and display analyses on several datasets imputed using the mice package. For example:
library(mice)
# Create a new dataset with missing values
url <- 'https://vincentarelbundock.github.io/Rdatasets/csv/HistData/Guerry.csv'
tmp <- read.csv(url)[, c('Clergy', 'Donations', 'Literacy')]
tmp$Clergy[sample(1:nrow(tmp), 3)] <- NA
tmp$Donations[sample(1:nrow(tmp), 3)] <- NA
tmp$Literacy[sample(1:nrow(tmp), 3)] <- NA
# Impute dataset 5 times
tmp <- mice(tmp, m = 5, printFlag = FALSE, seed = 1024)
# Estimate models
mod <- list()
mod[[1]] <- with(tmp, lm(Clergy ~ Donations))
mod[[2]] <- with(tmp, lm(Clergy ~ Donations + Literacy))
# Summarize
msummary(mod, statistic = 't')
msummary(mod, statistic = 'ubar')The statistic argument can take any column name in the tidy data frame obtained by:
generics::tidy(mod[[1]])The gt package allows a bunch more customization and styling. Power users can use modelsummary's extract function to produce a tibble which can easily be fed into gt.
> modelsummary::extract(models)
# A tibble: 21 x 8
group term statistic `OLS 1` `NBin 1` `OLS 2` `NBin 2` `Logit 1`
<chr> <chr> <chr> <chr> <chr> <chr> <chr> <chr>
1 estimates (Intercept) estimate 64.114 4.218 57.331 4.384 1.006
2 estimates (Intercept) statistic (5.247) (0.144) (8.315) (0.233) (0.710)
3 estimates Crime_prop estimate -0.002 -0.000 -0.002 -0.000 -0.000
4 estimates Crime_prop statistic (0.001) (0.000) (0.001) (0.000) (0.000)
5 estimates Infants estimate -0.001 "" 0.000 "" -0.000
6 estimates Infants statistic (0.000) "" (0.000) "" (0.000)
7 estimates Donations estimate "" -0.000 "" -0.000 ""
8 estimates Donations statistic "" (0.000) "" (0.000) ""
9 gof R2 "" 0.237 "" 0.073 "" ""
10 gof Adj.R2 "" 0.218 "" 0.051 "" ""
# … with 11 more rowsThere are several excellent alternative summary table packages for R: