lmac: Available Cases Method for Missing Data

In matloff/regtools: Regression and Classification Tools

Available Cases Method for Missing Data

Description

Various estimators that handle missing data via the Available Cases Method

Usage

lmac(xy,nboot=0) 
makeNA(m,probna)
NAsTo0s(x)
ZerosToNAs(x,replaceVal=0)
## S3 method for class 'lmac'
coef(object,...)
## S3 method for class 'lmac'
vcov(object,...)
pcac(indata,scale=FALSE) 
loglinac(x,margin) 
tbltofakedf(tbl)

Arguments

replaceVal	Value to be replaced by NA.
xy	Matrix or data frame, X values in the first columns, Y in the last column.
indata	Matrix or data frame.
x	Matrix or data frame, one column per variable.
nboot	If positive, number of bootstrap samples to take.
probna	Probability that an element will be NA.
scale	If TRUE, call cor instead of cov.
tbl	An R table.
m	Number of synthetic NAs to insert.
object	Output from lmac.
...	Needed for consistency with generic function. Not used.
margin	A list of vectors specifying the model, as in loglin.

Details

The Available Cases (AC) approach applies to statistical methods that depend only on products of k of the variables, so that cases having non-NA values for those k variables can be used, as opposed to using only cases that are fully intact in all variables, the Complete Cases (CC) approach. In the case of linear regression, for instance, the estimated coefficients depend only on covariances between the variables (both predictors and response). This approach assumes thst the cases with missing values have the same distribution as the intact cases.

The lmac function forms OLS estimates as with lm, but applying AC, in contrast to lm, which uses the CC method.

The pcac function is an AC substitute for prcomp. The data is centered, corresponding to a fixed value of center = TRUE in prcomp. It is also scaled if scale is TRUE, corresponding scale = TRUE in prcomp. Due to AC, there is a small chance of negative eigenvalues, in which case stop will be called.

The loglinac function is an AC substitute for loglin. The latter takes tables as input, but loglinac takes the raw data. If you have just the table, use tbltofakedf to regenerate a usable data frame.

The makeNA function is used to insert random NA values into data, for testing purposes.

Value

For lmac, an object of class lmac, with components

• coefficients, as with lm; accessible directly or by calling coef, as with lm

• fitted.values, as with lm

• residuals, as with lm

• r2, (unadjusted) R-squared

• cov, for nboot > 0 the estimated covariance matrix of the vector of estimated regression coefficients; accessible directly or by calling vcov, as with lm

For pcac, an R list, with components

• sdev, as with prcomp

• rotation, as with prcomp

For loglinac, an R list, with components

• param, estimated coefficients, as in loglin

• fit, estimated expected call counts, as in loglin

Author(s)

Norm Matloff

Examples

n <- 25000
w <- matrix(rnorm(2*n),ncol=2)  # x and epsilon
x <- w[,1]
y <- x + w[,2]
# insert some missing values
nmiss <- round(0.1*n)
x[sample(1:n,nmiss)] <- NA
nmiss <- round(0.2*n)
y[sample(1:n,nmiss)] <- NA
acout <- lmac(cbind(x,y))
coef(acout)  # should be near pop. values 0 and 1

matloff/regtools documentation built on Jan. 22, 2025, 9:07 p.m.