---
title: "Multiple Imputation"
date: "`r Sys.Date()`"
bibliography: "biblio.bib"
link-citations: true
output: rmarkdown::html_vignette
vignette: >
  %\VignetteIndexEntry{Multiple Imputation}
  %\VignetteEngine{knitr::rmarkdown}
  %\VignetteEncoding{UTF-8}
---

```{r, include = FALSE}
knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>",
  warning = FALSE,
  message = FALSE
)
```

## How to use {missRanger} for multiple imputation?

For statistical inference, extra variability introduced by imputation has to be accounted for. This is usually done by multiple imputation.

One of the standard approaches is to impute the dataset multiple times, generating, e.g., 10 or 100 versions of the complete data. Then, the intended analysis (t-test, linear model etc.) is performed with each of the datasets. Their results are then pooled, usually by Rubin's rule [@rubin]: Parameter *estimates* are averaged. Their *variances* are avaraged as well, and corrected upwards by adding the variance of the parameter estimates across imputations.

The package {mice} [@buuren] takes care of this pooling step. The creation of multiple complete data sets can be done by {mice} or also by {missRanger}. In the latter case, in order to keep the variance of imputed values at a more realistic level, we suggest to use predictive mean matching with relatively large `pmm.k` on top of the random forest imputation. 

## Example

```r
library(missRanger)
library(mice)

set.seed(19)

iris_NA <- generateNA(iris, p = c(0, 0.1, 0.1, 0.1, 0.1))

# Generate 20 complete data sets with relatively large pmm.k
filled <- replicate(
  20, 
  missRanger(iris_NA, verbose = 0, num.trees = 100, pmm.k = 10), 
  simplify = FALSE
)
                           
# Run a linear model for each of the completed data sets                          
models <- lapply(filled, function(x) lm(Sepal.Length ~ ., x))

# Pool the results by mice
summary(pooled_fit <- pool(models))

#                term   estimate std.error statistic        df      p.value
# 1       (Intercept)  2.3343548 0.3244342  7.195157  97.08106 1.314353e-10
# 2       Sepal.Width  0.4715273 0.1041384  4.527891  88.55776 1.848669e-05
# 3      Petal.Length  0.7700316 0.0768588 10.018783 122.02953 1.321441e-17
# 4       Petal.Width -0.2506538 0.1739537 -1.440922  88.10220 1.531513e-01
# 5 Speciesversicolor -0.6648375 0.2940828 -2.260715  81.17797 2.645368e-02
# 6  Speciesvirginica -0.9065327 0.4055137 -2.235517  79.87581 2.817491e-02

# Compare with model on original data
summary(lm(Sepal.Length ~ ., data = iris))

# Coefficients:
#                   Estimate Std. Error t value Pr(>|t|)    
# (Intercept)        2.17127    0.27979   7.760 1.43e-12 ***
# Sepal.Width        0.49589    0.08607   5.761 4.87e-08 ***
# Petal.Length       0.82924    0.06853  12.101  < 2e-16 ***
# Petal.Width       -0.31516    0.15120  -2.084  0.03889 *  
# Speciesversicolor -0.72356    0.24017  -3.013  0.00306 ** 
# Speciesvirginica  -1.02350    0.33373  -3.067  0.00258 ** 
# ---
# Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
# 
# Residual standard error: 0.3068 on 144 degrees of freedom
# Multiple R-squared:  0.8673,	Adjusted R-squared:  0.8627 
# F-statistic: 188.3 on 5 and 144 DF,  p-value: < 2.2e-16
```

As expected, inference from multiple imputation seems to be less strong than of the original data without missings.

## References