corFiml: Find a Full Information Maximum Likelihood (FIML) correlation or covariance matrix from a data matrix with missing data

Description

Makes use of functions adapted from the lavaan package to find FIML covariance/correlation matrices. FIML can be much slower than the normal pairwise deletion option of cor, but provides slightly more precise estimates.

Usage

corFiml(x, covar = FALSE,show=FALSE)

Value

cor: The correlation matrix found using FIML
cov: The covariance matrix found using FIML
fx: The ML fit function

Arguments

x: A data.frame or data matrix
covar: By default, just return the correlation matrix. If covar is TRUE, return a list containing the covariance matrix and the ML fit function.
show: If show=TRUE, then just show the patterns of missingness, but don't do the FIML. Useful for understanding the process of fiml.

Author

Wiliam Revelle

Details

In the presence of missing data, Full Information Maximum Likelihood (FIML) is an alternative to simply using the pairwise correlations. The implementation in the lavaan package for structural equation modeling has been adapted for the simpler case of just finding the correlations or covariances.

The pairwise solution for any pair of variables is insensitive to other variables included in the matrix. On the other hand, the ML solution depends upon the entire set of items being correlated. This will lead to slightly different solutions for different subsets of variables.

The basic FIML algorithm is to find the pairwise ML solution for covariances and means for every pattern of missingness and then to weight the solution by the size of every unique pattern of missingness.

Examples

Run this code

rML <- corFiml(bfi[20:27])
rpw <- cor(bfi[20:27],use="pairwise") 
round(rML - rpw,3)
mp <- corFiml(bfi[20:27],show=TRUE)
mp

Run the code above in your browser using DataLab