library(knitr) opts_chunk$set(warning = FALSE, message = FALSE)
The broom package takes the messy output of built-in functions in R, such as
t.test, and turns them into tidy tibbles.
The concept of "tidy data", as introduced by Hadley Wickham, offers a powerful framework for data manipulation and analysis. That paper makes a convincing statement of the problem this package tries to solve (emphasis mine):
While model inputs usually require tidy inputs, such attention to detail doesn't carry over to model outputs. Outputs such as predictions and estimated coefficients aren't always tidy. This makes it more difficult to combine results from multiple models. For example, in R, the default representation of model coefficients is not tidy because it does not have an explicit variable that records the variable name for each estimate, they are instead recorded as row names. In R, row names must be unique, so combining coefficients from many models (e.g., from bootstrap resamples, or subgroups) requires workarounds to avoid losing important information. This knocks you out of the flow of analysis and makes it harder to combine the results from multiple models. I'm not currently aware of any packages that resolve this problem.
broom is an attempt to bridge the gap from untidy outputs of predictions and estimations to the tidy data we want to work with. It centers around three S3 methods, each of which take common objects produced by R statistical functions (
nls, etc) and convert them into a tibble. broom is particularly designed to work with Hadley's dplyr package (see the broom+dplyr vignette for more).
broom should be distinguished from packages like reshape2 and tidyr, which rearrange and reshape data frames into different forms. Those packages perform critical tasks in tidy data analysis but focus on manipulating data frames in one specific format into another. In contrast, broom is designed to take format that is not in a tabular data format (sometimes not anywhere close) and convert it to a tidy tibble.
Tidying model outputs is not an exact science, and it's based on a judgment of the kinds of values a data scientist typically wants out of a tidy analysis (for instance, estimates, test statistics, and p-values). You may lose some of the information in the original object that you wanted, or keep more information than you need. If you think the tidy output for a model should be changed, or if you're missing a tidying function for an S3 class that you'd like, I strongly encourage you to open an issue or a pull request.
This package provides three S3 methods that do three distinct kinds of tidying.
tidy: constructs a tibble that summarizes the model's statistical findings. This includes coefficients and p-values for each term in a regression, per-cluster information in clustering applications, or per-test information for
augment: add columns to the original data that was modeled. This includes predictions, residuals, and cluster assignments.
glance: construct a concise one-row summary of the model. This typically contains values such as R^2, adjusted R^2, and residual standard error that are computed once for the entire model.
Note that some classes may have only one or two of these methods defined.
Consider as an illustrative example a linear fit on the built-in
lmfit <- lm(mpg ~ wt, mtcars) lmfit summary(lmfit)
This summary output is useful enough if you just want to read it. However, converting it to tabular data that contains all the same information, so that you can combine it with other models or do further analysis, is not trivial. You have to do
coef(summary(lmfit)) to get a matrix of coefficients, the terms are still stored in row names, and the column names are inconsistent with other packages (e.g.
Pr(>|t|) compared to
Instead, you can use the
tidy function, from the broom package, on the fit:
This gives you a tabular data representation. Note that the row names have been moved into a column called
term, and the column names are simple and consistent (and can be accessed using
Instead of viewing the coefficients, you might be interested in the fitted values and residuals for each of the original points in the regression. For this, use
augment, which augments the original data with information from the model:
Note that each of the new columns begins with a
. (to avoid overwriting any of the original columns).
Finally, several summary statistics are computed for the entire regression, such as R^2 and the F-statistic. These can be accessed with the
This distinction between the
glance functions is explored in a different context in the k-means vignette.
These functions apply equally well to the output from
glmfit <- glm(am ~ wt, mtcars, family = "binomial") tidy(glmfit) augment(glmfit) glance(glmfit)
Note that the statistics computed by
glance are different for
glm objects than for
lm (e.g. deviance rather than R^2):
These functions also work on other fits, such as nonlinear models (
nlsfit <- nls(mpg ~ k / wt + b, mtcars, start = list(k = 1, b = 0)) tidy(nlsfit) augment(nlsfit, mtcars) glance(nlsfit)
tidy function can also be applied to
htest objects, such as those output by popular built-in functions like
tt <- t.test(wt ~ am, mtcars) tidy(tt)
Some cases might have fewer columns (for example, no confidence interval):
wt <- wilcox.test(wt ~ am, mtcars) tidy(wt)
tidy output is already only one row,
glance returns the same output:
augment method is defined only for chi-squared tests, since there is no meaningful sense, for other tests, in which a hypothesis test produces output about each initial data point.
chit <- chisq.test(xtabs(Freq ~ Sex + Class, data = as.data.frame(Titanic))) tidy(chit) augment(chit)
In order to maintain consistency, we attempt to follow some conventions regarding the structure of returned data.
glancefunctions is always a tibble.
PValue?" every time). The examples below are not all the possible column names, nor will all tidy output contain all or even any of these columns.
tidyoutput typically represents some well-defined concept, such as one term in a regression, one test, or one cluster/class. This meaning varies across models but is usually self-evident. The one thing each row cannot represent is a point in the initial data (for that, use the
term"" the term in a regression or model that is being estimated.
p.value: this spelling was chosen (over common alternatives such as
pval) to be consistent with functions in R's built-in
statistica test statistic, usually the one used to compute the p-value. Combining these across many sub-groups is a reliable way to perform (e.g.) bootstrap hypothesis testing
conf.lowthe low end of a confidence interval on the
conf.highthe high end of a confidence interval on the
dfdegrees of freedom
augment(model, data)adds columns to the original data.
dataargument is missing,
augmentattempts to reconstruct the data from the model (note that this may not always be possible, and usually won't contain columns not used in the model).
augmentoutput matches the corresponding row in the original data.
augmentturns them into a column called
.to avoid overwriting columns in the original data.
.fitted: the predicted values, on the same scale as the data.
.resid: residuals: the actual y values minus the fitted values
.cluster: cluster assignments
glancealways returns a one-row tibble.
glance(NULL)returns an empty tibble.
glmglance output does not need to contain a field for
family, since that is decided by the user calling
glmrather than the modeling function itself.
r.squaredthe fraction of variance explained by the model
adj.r.squared$R^2$ adjusted based on the degrees of freedom
sigmathe square root of the estimated variance of the residuals
Any scripts or data that you put into this service are public.
Add the following code to your website.
For more information on customizing the embed code, read Embedding Snippets.