README.md

In zrmacc/MNMix: Missingness Aware Gaussian Mixture Models

Missingness Aware Gaussian Mixture Models

Zachary McCaw Updated: 2021-12-14

This package performs estimation and inference for Gaussian Mixture Models (GMMs) where the input data may contain missing values. Rather than imputing missing values before fitting the GMM, this package uses an extended EM algorithm to obtain the true maximum likelihood estimates of all model parameters given the observed data. In particular MGMM performs the following tasks:

• Maximum likelihood estimation of cluster means, covariances, and proportions.
• Calculation of cluster membership probabilities and maximum a posteriori classification of the input vectors.
• Deterministic completion of the input data, by imputing missing elements to their posterior means, and stochastic completion of the input data, by drawing missing elements from the fitted GMM.

The method is detailed in Fitting Gaussian mixture models on incomplete data.

Main Functions

• FitGMM estimates model parameters, performs classification and imputation.
• rGMM simulates observations from a GMM, potentially with missingness.
• ChooseK provides guidance on choosing the number of clusters.
• GenImputation performs stochastic imputation for multiple imputation-based inference.

Compact Example

set.seed(101)
library(MGMM)

# Parameter settings.
mean_list <- list(
  c(1, 1),
  c(-1, -1)
)
cov_list <- list(
  matrix(c(1, -0.5, -0.5, 1), nrow = 2),
  matrix(c(1, 0.5, 0.5, 1), nrow = 2)
)

# Generate data.
data <- rGMM(
  n = 1e3, 
  d = 2, 
  k = 2, 
  miss = 0.1, 
  means = mean_list, 
  covs = cov_list
)

# Original data.
head(data)

##           y1          y2
## 1  1.6512855  2.60621938
## 2 -0.5721069          NA
## 2 -2.0045376 -2.31888263
## 2 -0.6229388 -1.51543968
## 1  2.0258413  0.06921658
## 2 -1.3476380 -1.51915826

# Choose cluster number.
choose_k <- ChooseK(
  data,
  k0 = 2,
  k1 = 4,
  boot = 10,
  maxit = 10,
  eps = 1e-4,
  report = TRUE
)

## Cluster size 2 complete. 11 fit(s) succeeded.
## Cluster size 3 complete. 11 fit(s) succeeded.
## Cluster size 4 complete. 11 fit(s) succeeded.

# Cluster number recommendations. 
show(choose_k$Choices)

##   Metric k_opt   Metric_opt k_1se   Metric_1se
## 1    BIC     4 2285.4269195     2 2340.4441587
## 2    CHI     4    4.5081110     4    4.5081110
## 3    DBI     2    0.7818011     2    0.7818011
## 4    SIL     2    0.4785951     2    0.4785951

# Estimation.
fit <- FitGMM(
  data,
  k = 2,
  maxit = 10
)

## Objective increment:  11.1 
## Objective increment:  2.39 
## Objective increment:  1.57 
## Objective increment:  1.08 
## Objective increment:  0.91 
## Objective increment:  0.745 
## Objective increment:  0.617 
## Objective increment:  0.507 
## Objective increment:  0.416 
## Objective increment:  0.34 
## 10 update(s) performed without reaching tolerance limit.

# Estimated means. 
show(fit@Means)

## [[1]]
##        y1        y2 
## -1.056071 -1.070412 
## 
## [[2]]
##        y1        y2 
## 0.9437473 0.9513797

# Estimated covariances. 
show(fit@Covariances)

## [[1]]
##           y1        y2
## y1 0.9447484 0.5201638
## y2 0.5201638 0.9611714
## 
## [[2]]
##            y1         y2
## y1  0.9973684 -0.4489898
## y2 -0.4489898  0.9728258

# Cluster assignments. 
head(fit@Assignments)

##   Assignments      Entropy
## 1           2 7.957793e-02
## 2           1 8.426200e-01
## 2           1 8.629837e-07
## 2           1 7.790894e-03
## 1           2 8.451183e-02
## 2           1 4.521627e-04

# Deterministic imputation.
head(fit@Completed)

##           y1          y2
## 1  1.6512855  2.60621938
## 2 -0.5721069 -0.14379539
## 2 -2.0045376 -2.31888263
## 2 -0.6229388 -1.51543968
## 1  2.0258413  0.06921658
## 2 -1.3476380 -1.51915826

# Stochastic imputation.
imp <- GenImputation(fit)
head(imp)

##           y1          y2
## 1  1.6512855  2.60621938
## 2 -0.5721069  0.88613853
## 2 -2.0045376 -2.31888263
## 2 -0.6229388 -1.51543968
## 1  2.0258413  0.06921658
## 2 -1.3476380 -1.51915826

Vignette

Additional examples and details may be found here.

zrmacc/MNMix documentation built on April 30, 2023, 6:37 a.m.