Fast, memory efficient Multiple Imputation by Chained Equations (MICE) with random forests. It can impute categorical and numeric data without much setup, and has an array of diagnostic plots available.
This document contains a thorough walkthrough of the package, benchmarks, and an introduction to multiple imputation. More information on MICE can be found in Stef van Buuren’s excellent online book, which you can find here.
You can download the latest stable version from CRAN:
install.packages("miceRanger")You can also download the latest development version from this repository:
library(devtools)
devtools::install_github("FarrellDay/miceRanger")For more information about updates, please see NEWS.md.
In these examples we will be looking at a simple example of multiple imputation. We need to load the packages, and define the data:
require(miceRanger)
set.seed(1)
# Load data
data(iris)
# Ampute the data. iris contains no missing values by default.
ampIris <- amputeData(iris,perc=0.25)
head(ampIris,10)## Sepal.Length Sepal.Width Petal.Length Petal.Width Species
## 1: 5.1 3.5 NA 0.2 <NA>
## 2: 4.9 3.0 1.4 0.2 setosa
## 3: 4.7 3.2 1.3 0.2 setosa
## 4: 4.6 3.1 1.5 0.2 setosa
## 5: 5.0 3.6 1.4 0.2 setosa
## 6: 5.4 3.9 1.7 0.4 setosa
## 7: NA 3.4 1.4 0.3 setosa
## 8: 5.0 3.4 1.5 0.2 setosa
## 9: 4.4 2.9 1.4 0.2 <NA>
## 10: 4.9 3.1 1.5 0.1 setosa
# Perform mice, return 6 datasets.
seqTime <- system.time(
miceObj <- miceRanger(
ampIris
, m=6
, returnModels = TRUE
, verbose=FALSE
)
)
miceObj## Class: miceDefs
## Datasets: 6
## Iterations: 5
## Total Seconds: 3
## Imputed Cols: 5
## Estimated Time per Additional Iteration is 1 Seconds
## Estimated Time per Additional Dataset is 0 Seconds
##
## For additional metrics, see the different plotting functions.
Printing the miceDefs object will tell you some high level information, including how long the process took, and how long it estimates adding more datasets/iterations will take.
Running in parallel is usually not necessary. By default, ranger will
use all available cores, and data.tables assignment by reference is
already lightning fast. However, in certain cases, we can still save
some time by sending each dataset imputation to a different R back end.
To do this, we need to set up a core cluster and use parallel = TRUE.
This causes the dataset to be copied for each back end, which may eat
up your RAM. If the process is memory constrained, this can cause the
parallel implementation to actually take more time than the sequential
implementation.
library(doParallel)
# Set up back ends.
cl <- makeCluster(2)
registerDoParallel(cl)
# Perform mice
parTime <- system.time(
miceObjPar <- miceRanger(
ampIris
, m=6
, parallel = TRUE
, verbose = FALSE
)
)
stopCluster(cl)
registerDoSEQ()Let’s take a look at the time we saved running in parallel:
perc <- round(1-parTime[[3]]/seqTime[[3]],2)*100
print(paste0("The parallel process ran ",perc,"% faster using 2 R back ends."))## [1] "The parallel process ran 7% faster using 2 R back ends."
We did not save that much time by running in parallel. ranger already
makes full use of our CPU. Running in parallel will save you time if you
are using a high meanMatchCandidates, or if you are working with large
data and use a low num.trees. See
benchmarks/
for more details.
If you plot your data and notice that you need to may need to run more iterations, or you would like more datasets for your analysis, you can use the following functions:
miceObj <- addIterations(miceObj,iters=2,verbose=FALSE)
miceObj <- addDatasets(miceObj,datasets=1,verbose=FALSE)It is possible to customize our imputation procedure by variable. By
passing a named list to vars, you can specify the predictors for each
variable to impute. You can also select which variables should be
imputed using mean matching, as well as the mean matching candidates, by
passing a named vector to valueSelector and meanMatchCandidates,
respectively:
v <- list(
Sepal.Width = c("Sepal.Length","Petal.Width","Species")
, Sepal.Length = c("Sepal.Width","Petal.Width")
, Species = c("Sepal.Width")
)
vs <- c(
Sepal.Width = "meanMatch"
, Sepal.Length = "value"
, Species = "meanMatch"
)
mmc <- c(
Sepal.Width = 4
, Species = 10
)
miceObjCustom <- miceRanger(
ampIris
, vars = v
, valueSelector = vs
, meanMatchCandidates = mmc
, verbose=FALSE
)Multiple Imputation can take a long time. If you wish to impute a
dataset using the MICE algorithm, but don’t have time to train new
models, it is possible to impute new datasets using a miceDefs object.
The impute function uses the random forests returned by miceRanger
to perform multiple imputation without updating the random forest at
each iteration:
newDat <- amputeData(iris)
newImputed <- impute(newDat,miceObj,verbose=FALSE)All of the imputation parameters (valueSelector, vars, etc) will be
carried over from the original miceDefs object. When mean matching,
the candidate values are pulled from the original dataset. This method
returns results just as good as re-running the data through MICE in
benchmarking:
In the chart above, a dataset with 15 variables (a-j numeric, k-p
categorical) and 51200 rows was imputed using miceRanger. A different
dataset with the same dimensions, but different data, was then imputed
using the models created with miceRanger. See the
benchmarks/
folder for scripts and more information on this chart. See ?impute for
more details on the function.
miceRanger comes with an array of diagnostic plots that tell you how
valid the imputations may be, how they are distributed, which variables
were used to impute other variables, and so on.
We can take a look at the imputed distributions compared to the original distribution for each variable:
plotDistributions(miceObj,vars='allNumeric')
The red line is the density of the original, nonmissing data. The smaller, black lines are the density of the imputed values in each of the datasets. If these don’t match up, it’s not a problem, however it may tell you that your data was not Missing Completely at Random (MCAR).
We are probably interested in knowing how our values between datasets
converged over the iterations. The plotCorrelations function shows you
a boxplot of the correlations between imputed values in every
combination of datasets, at each iteration:
plotCorrelations(miceObj,vars='allNumeric')
Different correlation measures can be plotted by specifying
factCorrMetric and numbCorrMetric. See ?plotCorrelations for more
details.
Sometimes, if the missing data locations are correlated with higher or lower values, we need to run multiple iterations for the process to converge to the true theoretical mean (given the information that exists in the dataset). We can see if the imputed data converged, or if we need to run more iterations:
plotVarConvergence(miceObj,vars='allNumeric')
It doesn’t look like this dataset had a convergence issue. We wouldn’t expect one, since we amputed the data above completely at random for each variable. When plotting categorical variables, the center and dispersion metrics plotted are the percent of the mode and the entropy, respectively.
Random Forests give us a cheap way to determine model error without cross validation. Each model returns the OOB accuracy for classification, and r-squared for regression. We can see how these converged as the iterations progress:
plotModelError(miceObj,vars='allNumeric')
It looks like the variables were imputed with a reasonable degree of accuracy. That spike after the first iteration was due to the nature of how the missing values are filled in before the models are run.
Now let’s plot the variable importance for each imputed variable. The top axis contains the variable that was used to impute the variable on the left axis.
plotVarImportance(miceObj)
The variable importance metric used is returned by ranger when
importance = 'impurity'. Due to large possible variances in the
returned value, the data plotted here has been 0-1 scaled within each
imputed variable. Use display = 'Absolute' to show unscaled variable
importance.
We are probably interested in how “certain” we were of our imputations.
We can get a feel for the variance experienced for each imputed value
between the datasets by using plotImputationVariance() function:
plotImputationVariance(miceObj,ncol=2,widths=c(5,3))
When plotting categorical data, the distribution of the number of unique
imputed levels is compared to the theoretical distribution of unique
levels, given they were drawn randomly. You can see that most of the
imputed values only had 1 imputed value across our 8 datasets, which
means that the imputation process was fairly ‘certain’ of that imputed
class. According to the graph, most of our samples would have had 3
different samples drawn, if they were drawn randomly for each dataset
sample.
When plotting the variance of numeric features, the standard deviation
of the imputed values is calculated for each sample. This is then
compared to the total population standard deviation. Percentage of the
samples with a SD below the population SD is shaded in the densities
above, and the Quantile is shown in the title. The iris dataset tends
to be full of correlation, so all of our imputations had a SD lower than
the population SD, however this will not always be the case.
To return the imputed data simply use the completeData function:
dataList <- completeData(miceObj)
head(dataList[[1]],10)## Sepal.Length Sepal.Width Petal.Length Petal.Width Species
## 1: 5.1 3.5 1.3 0.2 setosa
## 2: 4.9 3.0 1.4 0.2 setosa
## 3: 4.7 3.2 1.3 0.2 setosa
## 4: 4.6 3.1 1.5 0.2 setosa
## 5: 5.0 3.6 1.4 0.2 setosa
## 6: 5.4 3.9 1.7 0.4 setosa
## 7: 5.1 3.4 1.4 0.3 setosa
## 8: 5.0 3.4 1.5 0.2 setosa
## 9: 4.4 2.9 1.4 0.2 setosa
## 10: 4.9 3.1 1.5 0.1 setosa
We can see how the imputed data compares to the original data before it was amputed:
It looks like most of our variables were imputed with a high degree of
accuracy. Sepal.Width had a relatively poor Spearman correlation,
however we expected this when we saw the results from plotModelError()
above.
Scripts and more details on benchmarks can be found in
benchmarks/.
Using artificial data, the time and performance of miceRanger, mice
(method = "rf") and missForest were recorded. parlmice was used to run
mice in parallel, and a parallel back end was set up for missForest.
All runs used 5 cores. miceRangerPar refers to miceRanger being run with
parallel = TRUE.
Multiple Imputation by Chained Equations ‘fills in’ (imputes) missing data in a dataset through an iterative series of predictive models. In each iteration, each specified variable in the dataset is imputed using the other variables in the dataset. These iterations should be run until it appears that convergence has been met.
This process is continued until all specified variables have been imputed. Additional iterations can be run if it appears that the average imputed values have not converged, although no more than 5 iterations are usually necessary.
MICE is particularly useful if missing values are associated with the target variable in a way that introduces leakage. For instance, let’s say you wanted to model customer retention at the time of sign up. A certain variable is collected at sign up or 1 month after sign up. The absence of that variable is a data leak, since it tells you that the customer did not retain for 1 month.
Information is often collected at different stages of a ‘funnel’. MICE can be used to make educated guesses about the characteristics of entities at different points in a funnel.
MICE can be used to impute missing values, however it is important to keep in mind that these imputed values are a prediction. Creating multiple datasets with different imputed values allows you to do two types of inference:
- Imputed Value Distribution: A profile can be built for each imputed value, allowing you to make statements about the likely distribution of that value.
- Model Prediction Distribution: With multiple datasets, you can build multiple models and create a distribution of predictions for each sample. Those samples with imputed values which were not able to be imputed with much confidence would have a larger variance in their predictions.
miceRanger can make use of a procedure called predictive mean matching
(PMM) to select which values are imputed. PMM involves selecting a
datapoint from the original, nonmissing data which has a predicted value
close to the predicted value of the missing sample. The closest N
(meanMatchCandidates parameter in miceRanger()) values are chosen as
candidates, from which a value is chosen at random. This can be
specified on a column-by-column basis in miceRanger. Going into more
detail from our example above, we see how this works in practice:
This method is very useful if you have a variable which needs imputing which has any of the following characteristics:
- Multimodal
- Integer
- Skewed
As an example, let’s construct a dataset with some of the above characteristics:
# random uniform variable
nrws <- 1000
dat <- data.table(Uniform_Variable = runif(nrws))
# slightly bimodal variable correlated with Uniform_Variable
dat$Close_Bimodal_Variable <- sapply(
dat$Uniform_Variable
, function(x) sample(c(rnorm(1,-2),rnorm(1,2)),prob=c(x,1-x),size=1)
) + dat$Uniform_Variable
# very bimodal variable correlated with Uniform_Variable
dat$Far_Bimodal_Variable <- sapply(
dat$Uniform_Variable
, function(x) sample(c(rnorm(1,-3),rnorm(1,3)),prob=c(x,1-x),size=1)
)
# Highly skewed variable correlated with Uniform_Variable
dat$Skewed_Variable <- exp((dat$Uniform_Variable*runif(nrws)*3)) + runif(nrws)*3
# Integer variable correlated with Close_Bimodal_Variable and Uniform_Variable
dat$Integer_Variable <- round(dat$Uniform_Variable + dat$Close_Bimodal_Variable/3 + runif(nrws)*2)
# Ampute the data.
ampDat <- amputeData(dat,0.2)
# Plot the original data
plot(dat)
We can see how our variables are distributed and correlated in the graph above. Now let’s run our imputation process twice, once using mean matching, and once using the model prediction.
mrMeanMatch <- miceRanger(ampDat,valueSelector = "meanMatch",verbose=FALSE)
mrModelOutput <- miceRanger(ampDat,valueSelector = "value",verbose=FALSE)Let’s look at the effect on the different variables.
The affect of mean matching on our imputations is immediately apparent. If we were only looking at model error, we may be inclined to use the Prediction Value, since it has a higher OOB R-Squared. However, we are left with imputations that do not match our original distribution, and therefore, do not behave like our original data.
We see a similar occurance in the skewed variable - the distribution of the values imputed with the Prediction Value are shifted towards the mean.
The most obvious variable affected by mean matching was our integer
variable - using valueSelector = 'value' allows interpolation in the
numeric variables. Using mean matching has allowed us to keep the
distribution and distinct values of the original data, without
sacrificing accuracy.