memice

memice is a package with several functions to facilitate the estimation of marginal effects in linear and generalized linear models. It takes a model object (lm() or glm() objects) and a .mids object (a data object containing data multiply imputed by mice::mice()), estimates the marginal effects in each dataset within the .mids object, and pools the marginal effect estimates according to Rubin’s Rules. This package takes inspiration from Daniel Klein’s mimrgns program for Stata.

Installation

You can install the development version of memice like so:

devtools::install_github("jmweinstein/memice")

The recommended workflow for using all memice functions is:

  1. Create your desired model with the dataset that contains missing data
  2. Use mice::mice() to impute data via chained equations and save to an object
  3. Call memice::initialize_memice() to initialize your workspace for using memice (this adds formatting functions to your workspace, which are necessary for memice properly handling different variable types)
  4. Use one of the memice functions, feeding in your model from Step 1 and your .mids object (from mice::mice()) from Step 2

Model estimation is done prior to the use of memice functions to ensure that all quirks with models have been figured out before being fed into any of the memice functions. Things that may cause issues with memice functions pooling marginal effects estimates include non-convergence of generalized linear models or improperly-formatted variables (numbers that should be factors, factors that should be numeric, etc.)- by estimating the model before using the memice functions, the user has the opportunity to identify and remedy these issues beforehand.

Example usage of memice::ame_mice()

The function memice::ame_mice() estimates average marginal effects (AME) in multiply imputed data created by mice::mice().

library(memice)
library(missMethods)


## Create dataset with artificially-induced missingness in 'cyl' variable
data_with_missings <- mtcars |>
  missMethods::delete_MCAR(0.3, "cyl") |>
  dplyr::mutate(cyl = as.factor(cyl))

Create your desired model with the dataset that contains missing data

lm_mod <- lm(formula = mpg ~ wt + hp + cyl + cyl*wt + disp,
             data = data_with_missings)

Create .mids object of imputed data using mice::mice()

imputation <- mice::mice(data = data_with_missings, m = 5, seed = 01701)
#> 
#>  iter imp variable
#>   1   1  cyl
#>   1   2  cyl
#>   1   3  cyl
#>   1   4  cyl
#>   1   5  cyl
#>   2   1  cyl
#>   2   2  cyl
#>   2   3  cyl
#>   2   4  cyl
#>   2   5  cyl
#>   3   1  cyl
#>   3   2  cyl
#>   3   3  cyl
#>   3   4  cyl
#>   3   5  cyl
#>   4   1  cyl
#>   4   2  cyl
#>   4   3  cyl
#>   4   4  cyl
#>   4   5  cyl
#>   5   1  cyl
#>   5   2  cyl
#>   5   3  cyl
#>   5   4  cyl
#>   5   5  cyl

Initialize memice using initialize_memice, then estimate and pool marginal effects across the imputed datasets

memice::initialize_memice()
pooled_mfx <- memice::ame_mice(model = lm_mod, mice_object = imputation)
pooled_mfx
#>   term contrast     estimate  std.error  statistic     p.value
#> 1   wt    dY/dX -3.454409623 1.12719656 -3.0646027 0.002190045
#> 2   hp    dY/dX -0.022379412 0.01171196 -1.9108176 0.056036848
#> 3  cyl    6 - 4 -1.186579675 1.84860872 -0.6418771 0.520971953
#> 4  cyl    8 - 4 -1.611105025 2.64171172 -0.6098716 0.541949282
#> 5 disp    dY/dX -0.002238081 0.01201281 -0.1863078 0.852204214

Acknowledgements

I thank both Dr. Vincent Arel-Bundock and Dr. Noah Greifer for their technical assistance in developing this package. This package takes inspiration from Daniel Klein’s mimrgns program for Stata.