Installing Supporting Packages

Noah Greifer

2023-05-22

WeightIt is a wrapper for several other packages that aid in estimating balancing weights. In many ways, this is WeigthIt’s strength, because it is easy to try out several weighting methods without having to learn a completely new syntax for each one. One weakness of this is that when one of these packages is not available (e.g., on CRAN), the method that relies on that package cannot be used.

This document explains how to install each package WeightIt uses. You do not need to install every single one; you only need the one you want to use. For example, the miseam package provides support for logistic regression with missing data, but if you have no missing data or you don’t want to use the approach implemented in miseam, you don’t need to install it. WeightIt strongly depends on four packages, which are automatically installed along with WeightIt, so you don’t need to worry about installing them separately; these are cobalt, ggplot2, crayon, and backports.

Below we note each method (by name and by the input to the method argument of weightit()) and how to install the required packages either from CRAN or otherwise when the CRAN version is not available. In many cases, this involves installing the package from the author’s GitHub repository, which requires either the remotes or devtools package, which both contain the function install_github().

Propensity score weighting using GLMs (method = "glm")

Several options are available for estimating propensity score weights using GLMs depending on the treatment type and other features of the desired model. For binary treatments, weightit() uses stats::glm() by default, and for continuous treatments, weightit() uses stats::lm() by default, so no additional packages are required. For multi-category treatments with use.mlogit = FALSE, weightit() uses stats::glm().

Binary and continuous treatments with missing = "saem"

For binary and continuous treatments, when missing data is present and missing = "saem" is supplied, the misaem package is required. To install misaem from CRAN, run

install.packages("misaem")

If misaem is not on CRAN, or if you want to install the development version from source, you can do so from the developer’s GitHub repo using the following code:

remotes::install_github("julierennes/misaem")

Multicategory treatments with use.mlogit = TRUE

For multicategory treatments, when use.mlogit = TRUE (the default), the mlogit package is required for multinomial logistic regression. To install mlogit from CRAN, run

install.packages("mlogit")

If mlogit is not on CRAN, or if you want to install the development version from source, you can do so from the developer, Yves Croissant’s, R-Forge repo using the following code:

install.packages("mlogit", repos = "http://R-Forge.R-project.org")

Multicategory treatments with use.mclogit = TRUE

For multicategory treatments, when use.mclogit = TRUE, the mclogit package is required for multinomial logistic regression. To install mclogit from CRAN, run

install.packages("mclogit")

If mclogit is not on CRAN, or if you want to install the development version from source, you can do so from the developer, Martin Elff’s, GitHub repo using the following code:

remotes::install_github("melff/mclogit")

Ordinal multicategory treatments

For multicategory ordinal treatments (i.e., where the treatment variable is an ordered factor), the MASS package is required unless link = "br.logit" (in which case the brglm2 package is needed; see above). MASS is a core package and can always be installed from CRAN using the code below:

install.packages("MASS")

Propensity Score weighting using GBM (method = "gbm")

WeightIt uses the R package gbm to estimate propensity score weights using GBM. It does not rely on the twang package at all. To install gbm from CRAN, run

install.packages("gbm")

If gbm is not on CRAN, or if you want to install the development version from source, you can do so from the developer’s GitHub repo using the following code:

remotes::install_github("gbm-developers/gbm")

gbm requires compilation, which means you may need additional software installed on your computer to install it from source.

Covariate Balancing Propensity Score weighting (method = "cbps" and method = "npcbps")

WeightIt uses the R package CBPS to perform (nonparametric) covariate balancing propensity score weighting. To install CBPS from CRAN, run

install.packages("CBPS")

If CBPS is not on CRAN, or if you want to install the development version from source, you can do so from the developer, Kosuke Imai’s, GitHub repo using the following code:

remotes::install_github("kosukeimai/CBPS")

Entropy balancing (method = "ebal")

WeightIt uses code written for WeightIt, so no additional packages need to be installed to use entropy balancing.

Empirical balancing calibration weighting (method = "ebcw")

Empirical balancing calibration weighting is no longer available with WeightIt. Use entropy balancing, which in most cases is mathematically identical.

Optimization-based weighting (method = "optweight")

WeightIt uses the R package optweight to perform optimization-based weighting. To install optweight from CRAN, run

install.packages("optweight")

If optweight is not on CRAN, or if you want to install the development version from source, you can do so from the developer, Noah Greifer’s (my), GitHub repo using the following code:

remotes::install_github("ngreifer/optweight")

optweight depends on the osqp package, which requires compilation, which means you may need additional software installed on your computer to install it from source.

Propensity score weighting using SuperLearner (method = "super")

WeightIt uses the R package SuperLearner to estimate propensity score weights using SuperLearner. To install SuperLearner from CRAN, run

install.packages("SuperLearner")

If SuperLearner is not on CRAN, or if you want to install the development version from source, you can do so from the developer, Eric Polley’s, GitHub repo using the following code:

remotes::install_github("ecpolley/SuperLearner")

SuperLearner itself is a wrapper for many other packages. The whole point of using SuperLearner is to include many different machine learning algorithms to combine them into a well-fitting stacked model. These algorithms exist in many different R packages, which each need to be installed to use them. See the Suggested packages on the SuperLearner CRAN page to see what packages might be used with SuperLearner.

There are additional functions for use with SuperLearner in the SuperLearnerExtra repository. To read these into your R session to be used with method = "super", use source() on the raw text file URL. For example, to read in the code for SL.dbarts, run

source("https://raw.githubusercontent.com/ecpolley/SuperLearnerExtra/master/SL/SL.dbarts.R")

Propensity score weighting using BART (method = "bart")

WeightIt uses the R package dbarts to estimate propensity score weights using BART. To install dbarts from CRAN, run

install.packages("dbarts")

If dbarts is not on CRAN, or if you want to install the development version from source, you can do so from the developer, Vincent Dorie’s, GitHub repo using the following code:

remotes::install_github("vdorie/dbarts")

dbarts requires compilation, which means you may need additional software installed on your computer to install it from source.

Energy Balancing (method = "energy")

WeightIt uses the R package osqp to perform the optimization required for energy balancing. To install osqp from CRAN, run

install.packages("osqp")

If osqp is not on CRAN, or if you want to install the development version from source, you can do so from the developer’s site using the instructions given here, though it is a bit more involved than other installations from source.