imposeMissing: Impose MAR, MCAR, planned missingness, or attrition on a data set

Description

Function imposes missing values on a data based on the known missing data types, including MCAR, MAR, planned, and attrition.

Usage

impose(miss, data.mat, pmMCAR = NULL, pmMAR = NULL)
imposeMissing(data.mat, cov = 0, pmMCAR = 0, pmMAR = 0, nforms = 0, 
	itemGroups = list(), twoMethod = 0, prAttr = 0, timePoints = 1, 
	ignoreCols = 0, threshold = 0, logit = "", logical = NULL)

Arguments

miss

Missing data object ('>SimMissing) used as the template for impose missing values

data.mat

Data to impose missing upon. Can be either a matrix or a data frame.

cov

Column indices of a covariate to be used to impose MAR missing, or MAR attrition. Will not be included in any removal procedure. See details.

pmMCAR

Decimal percent of missingness to introduce completely at random on all variables.

pmMAR

Decimal percent of missingness to introduce using the listed covariate as predictor. See details.

nforms

The number of forms for planned missing data designs, not including the shared form.

itemGroups

List of lists of item groupings for planned missing data forms. Unless specified, items will be divided into groups sequentially (e.g. 1-3,4-6,7-9,10-12)

twoMethod

With missing on one variable: vector of (column index, percent missing). Will put a given percent missing on that column in the matrix to simulate a two method planned missing data research design. With missing on two or more variables: list of (column indices, percent missing).

prAttr

Probability (or vector of probabilities) of an entire case being removed due to attrition at a given time point. When a covariate is specified along with this argument, attrition will be predicted by the covariate (MAR attrition). See details.

timePoints

Number of timepoints items were measured over. For longitudinal data, planned missing designs will be implemented within each timepoint. All methods to impose missing values over time assume an equal number of variables at each time point.

ignoreCols

The columns not imposed any missing values for any missing data patterns.

threshold

The threshold of the covariate used to impose missing values. Values on the covariate above this threshold are eligible to be deleted. The default threshold is the mean of the variable.

logit

The script used for imposing missing values by logistic regression. The script is similar to the specification of regression in lavaan such that each line begins with a dependent variable, then '~' is used as regression sign, and the formula of a linear combination of independent variable plus constant, such as y1 ~ 0.5 + 0.2*y2. '#' and '!' can be used as a comment (like lavaan). For the intercept, users may use 'p()' to specify the average proportion of missing, such as y1 ~ p(0.2) + 0.3*y2, which the average missing proportion of y1 is 0.2 and the missing of y1 depends on y2. Users may visualize the missing proportion from the logistic specification by the plotLogitMiss function.

logical

A matrix of logical values (TRUE/FALSE). If a value in the dataset is corresponding to the TRUE in the logical matrix, the value will be missing.

Value

A data matrix with NAs introduced in the way specified by the arguments.

Details

Without specifying any arguments, no missing values will be introduced.

A single covariate is required to specify MAR missing - this covariate can be distributed in any way. This covariate can be either continuous or categorical, as long as it is numerical. If the covariate is categorical, the threshold should be specified to one of the levels.

MAR missingness is specified using the threshold method - any value on the covariate that is above the specified threshold indicates a row eligible for deletion. If the specified total amount of MAR missingness is not possible given the total rows eligible based on the threshold, the function iteratively lowers the threshold until the total percent missing is possible.

Planned missingness is parameterized by the number of forms (n). This is used to divide the cases into n groups. If the column groupings are not specified, a naive method will be used that divides the columns into n+1 equal forms sequentially (1-4,5-9,10-13..), where the first group is the shared form.The first list of column indices given will be used as the shared group. If this is not desired, this list can be left empty.

For attrition, the probability can be specified as a single value or as a vector. For a single value, the probability of attrition will be the same across time points, and affects only cases not previously lost due to attrition. If this argument is a vector, this specifies different probabilities of attrition for each time point. Values will be recycled if this vector is smaller than the specified number of time points.

An MNAR processes can be generated by specifying MAR missingness and then dropping the covariate from the subsequent analysis.

Currently, if MAR missing is imposed along with attrition, both processes will use the same covariate and threshold.

Currently, all types of missingness (MCAR, MAR, planned, and attrition) are imposed independently. This means that specified global values of percent missing will not be additive (10 percent MCAR + 10 percent MAR does not equal 20 percent total missing).

Examples

Run this code

# NOT RUN {
data <- matrix(rep(rnorm(10,1,1),19),ncol=19)
datac <- cbind(data,rnorm(10,0,1),rnorm(10,5,5))

# Imposing Missing with the following arguments produces no missing values
imposeMissing(data)
imposeMissing(data,cov=c(1,2))
imposeMissing(data,pmMCAR=0)
imposeMissing(data,pmMAR=0)
imposeMissing(data,nforms=0)

#Some more usage examples

# No missing at variables 1 and 2
imposeMissing(data,cov=c(1,2),pmMCAR=.1)

# 3-Form design
imposeMissing(data,nforms=3)

# 3-Form design with specified groups of items (XABC)
imposeMissing(data, nforms = 3, itemGroups =
	list(c(1,2,3,4,5), c(6,7,8,9,10), c(11,12,13,14,15), c(16,17,18,19)))

# 3-Form design when variables 20 and 21 are not missing
imposeMissing(datac,cov=c(20,21),nforms=3)

# 2 method design where the expensive measure is on Variable 19
imposeMissing(data,twoMethod=c(19,.8))

# Impose missing data with percent attrition of 0.1 in 5 time points
imposeMissing(datac,cov=21,prAttr=.1,timePoints=5)

# Logistic-regression MAR
colnames(data) <- paste("y", 1:ncol(data), sep="")
script <- 'y1 ~ 0.05 + 0.1*y2 + 0.3*y3
		y4 ~ -2 + 0.1*y4
		y5 ~ -0.5'
imposeMissing(data, logit=script)
# }