--- title: "Installing Supporting Packages" author: "Noah Greifer" date: "`r Sys.Date()`" output: html_vignette: df_print: kable toc: true vignette: > %\VignetteIndexEntry{Installing Supporting Packages} %\VignetteEngine{knitr::rmarkdown_notangle} %\VignetteEncoding{UTF-8} bibliography: references.bib link-citations: true --- ```{r setup, cache = FALSE, include=FALSE, eval = TRUE} knitr::opts_chunk$set(eval = FALSE) ``` `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 a few packages, which are automatically installed along with `WeightIt`, so you don't need to worry about installing them separately; these are not listed here. 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 `multi.method = "weightit"`, the default, `weightit()` uses internal code. ### `missing = "saem"` When missing data is present and `missing = "saem"` is supplied, the `misaem` package is required. To install `misaem` from CRAN, run ```{r} 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](https://github.com/julierennes/misaem) using the following code: ```{r} remotes::install_github("julierennes/misaem") ``` ### Binary and multi-category treatments with `link = "br.logit"` For binary and multi-category treatments, when `link` is supplied as `"br.logit"` or another link beginning with `"br."`, the `brglm2` package is required. To install `brglm2` from CRAN, run ```{r} install.packages("brglm2") ``` If `brglm2` is not on CRAN, or if you want to install the development version from source, you can do so from the developer, Ioannis Kosmidis's, [GitHub repo](https://github.com/ikosmidis/brglm2) using the following code: ```{r} remotes::install_github("ikosmidis/brglm2") ``` `brglm2` requires compilation, which means you may need additional software installed on your computer to install it from source. ### Multi-category treatments with `multi.method = "mclogit"` For multi-category treatments, when `multi.method = "mclogit"`, the `mclogit` package is required for multinomial logistic regression. To install `mclogit` from CRAN, run ```{r} 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](https://github.com/melff/mclogit/) using the following code: ```{r} remotes::install_github("melff/mclogit") ``` ### Multi-category treatments with `multi.method = "mnp"` For multi-category treatments, when `multi.method = "mnp"`, the `MNP` package is required for Bayesian multinomial probit regression. To install `MNP` from CRAN, run ```{r} install.packages("MNP") ``` If `MNP` 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](https://github.com/kosukeimai/MNP) using the following code: ```{r} remotes::install_github("kosukeimai/MNP") ``` `MNP` requires compilation, which means you may need additional software installed on your computer to install it from source. ## 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 ```{r} 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](https://github.com/gbm-developers/gbm) using the following code: ```{r} 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"`) For `method = "cbps"`, `WeightIt` uses code written for `WeightIt`, so no additional packages need to be installed to use CBPS. ## Nonparametric Covariate Balancing Propensity Score weighting (`method = "npcbps"`) For `method = "npcbps"`, `WeightIt` uses the R package `CBPS` to perform nonparametric covariate balancing propensity score weighting. To install `CBPS` from CRAN, run ```{r} 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](https://github.com/kosukeimai/CBPS) using the following code: ```{r} 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. ## Inverse probability tilting (`method = "ipt"`) `WeightIt` uses the R package `rootSolve` to perform the root finding required for inverse probability tilting. To install `rootSolve` from CRAN, run ```{r} install.packages("rootSolve") ``` ## 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 ```{r} 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](https://github.com/ngreifer/optweight) using the following code: ```{r} 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 ```{r} 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](https://github.com/ecpolley/SuperLearner) using the following code: ```{r} 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](https://cran.r-project.org/package=SuperLearner) to see which packages might be used with `SuperLearner`. There are additional functions for use with `SuperLearner` in the `SuperLearnerExtra` [repository](https://github.com/ecpolley/SuperLearnerExtra). 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 ```{r} 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 ```{r} 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](https://github.com/vdorie/dbarts) using the following code: ```{r} 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 ```{r} 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](https://osqp.org/docs/get_started/r.html), though it is a bit more involved than other installations from source.