title: "Tidy, Type-Safe 'prediction()' Methods" output: github_document

The prediction and margins packages are a combined effort to port the functionality of Stata's (closed source) margins command to (open source) R. prediction is focused on one function - prediction() - that provides type-safe methods for generating predictions from fitted regression models. prediction() is an S3 generic, which always return a "data.frame" class object rather than the mix of vectors, lists, etc. that are returned by the predict() methods for various model types. It provides a key piece of underlying infrastructure for the margins package. Users interested in generating marginal (partial) effects, like those generated by Stata's margins, dydx(*) command, should consider using margins() from the sibling project, margins.

In addition to prediction(), this package provides a number of utility functions for generating useful predictions:

• find_data(), an S3 generic with methods that find the data frame used to estimate a regression model. This is a wrapper around get_all_vars() that attempts to locate data as well as modify it according to subset and na.action arguments used in the original modelling call.
• mean_or_mode() and median_or_mode(), which provide a convenient way to compute the data needed for predicted values at means (or at medians), respecting the differences between factor and numeric variables.
• seq_range(), which generates a vector of n values based upon the range of values in a variable
• build_datalist(), which generates a list of data frames from an input data frame and a specified set of replacement at values (mimicking the atlist option of Stata's margins command)

Simple code examples

A major downside of the predict() methods for common modelling classes is that the result is not type-safe. Consider the following simple example:

library("stats")
library("datasets")
x <- lm(mpg ~ cyl * hp + wt, data = mtcars)
class(predict(x))

## [1] "numeric"

class(predict(x, se.fit = TRUE))

## [1] "list"

prediction solves this issue by providing a wrapper around predict(), called prediction(), that always returns a tidy data frame with a very simple print() method:

library("prediction")
(p <- prediction(x))

## Data frame with 32 predictions from
##  lm(formula = mpg ~ cyl * hp + wt, data = mtcars)
## with average prediction: 20.0906

class(p)

## [1] "prediction" "data.frame"

head(p)

##    mpg cyl disp  hp drat    wt  qsec vs am gear carb   fitted se.fitted
## 1 21.0   6  160 110 3.90 2.620 16.46  0  1    4    4 21.90488 0.6927034
## 2 21.0   6  160 110 3.90 2.875 17.02  0  1    4    4 21.10933 0.6266557
## 3 22.8   4  108  93 3.85 2.320 18.61  1  1    4    1 25.64753 0.6652076
## 4 21.4   6  258 110 3.08 3.215 19.44  1  0    3    1 20.04859 0.6041400
## 5 18.7   8  360 175 3.15 3.440 17.02  0  0    3    2 17.25445 0.7436172
## 6 18.1   6  225 105 2.76 3.460 20.22  1  0    3    1 19.53360 0.6436862

The output always contains the original data (i.e., either data found using the find_data() function or passed to the data argument to prediction()). This makes it much simpler to pass predictions to, e.g., further summary or plotting functions.

Additionally the vast majority of methods allow the passing of an at argument, which can be used to obtain predicted values using modified version of data held to specific values:

prediction(x, at = list(hp = seq_range(mtcars$hp, 5)))

## Data frame with 160 predictions from
##  lm(formula = mpg ~ cyl * hp + wt, data = mtcars)
## with average predictions:

##     hp      x
##   52.0 22.605
##  122.8 19.328
##  193.5 16.051
##  264.2 12.774
##  335.0  9.497

This more or less serves as a direct R port of (the subset of functionality of) Stata's margins command that calculates predictive marginal means, etc. For calculation of marginal or partial effects, see the margins package.

Supported model classes

The currently supported model classes are:

• "lm" from stats::lm()
• "glm" from stats::glm(), MASS::glm.nb(), glmx::glmx(), glmx::hetglm(), brglm::brglm()
• "ar" from stats::ar()
• "Arima" from stats::arima()
• "arima0" from stats::arima0()
• "biglm" from biglm::biglm() (including "ffdf" backed models)
• "betareg" from betareg::betareg()
• "bruto" from mda::bruto()
• "clm" from ordinal::clm()
• "coxph" from survival::coxph()
• "crch" from crch::crch()
• "earth" from earth::earth()
• "fda" from mda::fda()
• "Gam" from gam::gam()
• "gausspr" from kernlab::gausspr()
• "gee" from gee::gee()
• "glimML" from aod::betabin(), aod::negbin()
• "glimQL" from aod::quasibin(), aod::quasipois()
• "glmnet" from glmnet::glmnet()
• "gls" from nlme::gls()
• "hurdle" from pscl::hurdle()
• "hxlr" from crch::hxlr()
• "ivreg" from AER::ivreg()
• "knnreg" from caret::knnreg()
• "kqr" from kernlab::kqr()
• "ksvm" from kernlab::ksvm()
• "lda" from MASS:lda()
• "lme" from nlme::lme()
• "loess" from stats::loess()
• "lqs" from MASS::lqs()
• "mars" from mda::mars()
• "mca" from MASS::mca()
• "mclogit" from mclogit::mclogit()
• "mda" from mda::mda()
• "merMod" from lme4::lmer() and lme4::glmer()
• "mnp" from MNP::mnp()
• "naiveBayes" from e1071::naiveBayes()
• "nlme" from nlme::nlme()
• "nls" from stats::nls()
• "nnet" from nnet::nnet(), nnet::multinom()
• "plm" from plm::plm()
• "polr" from MASS::polr()
• "ppr" from stats::ppr()
• "princomp" from stats::princomp()
• "qda" from MASS:qda()
• "rlm" from MASS::rlm()
• "rpart" from rpart::rpart()
• "rq" from quantreg::rq()
• "selection" from sampleSelection::selection()
• "speedglm" from speedglm::speedglm()
• "speedlm" from speedglm::speedlm()
• "survreg" from survival::survreg()
• "svm" from e1071::svm()
• "svyglm" from survey::svyglm()
• "tobit" from AER::tobit()
• "train" from caret::train()
• "truncreg" from truncreg::truncreg()
• "zeroinfl" from pscl::zeroinfl()

Requirements and Installation

The development version of this package can be installed directly from GitHub using remotes:

if (!require("remotes")) {
    install.packages("remotes")
}
remotes::install_github("leeper/prediction")