convData: Prepare data for imputations with BMLCimpute (convData)

In davidevdt/BMLCimpute: BMLCimpute: Bayesian Multilevel Latent Class Models for the Multiple Imputation of Nested Categorical Data

Description

This function takes a categorical dataset as input (categories can be denoted by numbers) and returns a list of objects that will be used by the 'multilevelLCMI' function to perform the imputations.

A package for the multiple imputation of single-level and nested categorical data by means of Bayesian Multilevel Latent Class models.

Usage

convData(dat, GID = NULL, UID = NULL, var2 = NULL)

Arguments

dat	Raw (categorical) data frame with missing data. It can also be a data matrix. The GID and UID arguments, if passed to the function, must be in the first two columns of the dataset.
GID	Group (level-2 unit) indicator (expressed as column number corresponding to the group ID in the dataset). It can be omitted in single-level datasets.
UID	Lower-level unit indicator (expressed as column number corresponding to the unit ID in the dataset). Optional.
var2	Higher-level (group-specific) variables (expressed as a vector of column numbers in the dataset corresponding to the variables measured at the higher levels). Optional.

Details

Convert a raw categorical dataset with missing data into one ready to be imputed with the multilevelLCMI function. In particular, the function will transform factor variables into numeric ones, where numbers denote a different category. A coding list is returned along with the converted dataset.

'BMLCimpute' allows researchers and users of categorical datasets with missing data to perform Multiple Imputation via Bayesian latent class models. Data can be either single- or multi-level. Model estimation and imputations are implemented via a Gibbs sampler run with the Rcpp package interface. The function multilevelLCMI performs the imputations. Prior to the imputation step, data should be processed with the function convData; the resulting list is then passed as input to the multilevelLCMI. Complete datasets are obtained via the compData function.

Value

A convData object, a list containing the following items:

convDat	The converted dataset
codLev1	List containing the new (and original) scores which will be used for the imputations (Level-1 variables).
codLev1	Vector containing the number of categories observed for each variable (Level-1 variables).
nCatLev1	Vector containing the number of categories observed for each variable (Level-1 variables).
codLev2	List containing the new (and original) scores which will be used for the imputations (Level-2 variables).
nCatLev2	List containing the new (and original) scores which will be used for the imputations (Level-2 variables).
GroupIDs	Matrix containing original and new Group ID's.
GID	The column Group ID number (as entered by the user).
UID	The column Unit ID number (as entered by the user).
var2	The column numbers for level-2 variables (as entered in the input).
doVar2	Boolean. Shall the BMLC model impute variables at level-2? (Result of !is.null(var2)).
namesLev1	Vector of variable names (level-1 variables).
namesLev2	Vector of variable names (level-2 variables).
GroupName	Group ID variable name.
CaseName	Unit ID variable name.
caseID	Unit ID vector (re-permuted).
sort_	Vector containing the original permutation of the dataset rows.

Functions

• multilevelLCMI for the imputations and model estimation (internally calls Rcpp code);

• convData for data preparation (preprocessing);

• compData for dataset completion.

Author(s)

D. Vidotto <[e-mail]>

BMLCimpute : Bayesian Multilevel Latent Class Models for the Multiple Imputation of Nested Categorical Data

Examples

## Not run: 

library(BMLCimpute)

# Load data 
data(simul_incomplete)

# Preprocess the Data
cd <- convData(simul_incomplete, GID = 1, UID = 2, var2 = 8:12)

# Model Selection
set.seed(1)
mmLC <- multilevelLCMI( convData = cd, L = 10, K = 10, it1 = 1000, it2 = 3000, it3 = 100,
    it.print = 250, v = 10, I = 0, pri2 = 1 / 10, pri1 = 1 / 15, priresp =  0.01,
    priresp2 = 0.01, random = TRUE, estimates = FALSE, count = TRUE, plot.loglik = FALSE,
    prec = 3, scale = 1.0)

# Select posterior maxima of the number of classes for the imputations 
# (Other alternatives are possible, such as posterior modes or posterior quantiles)
L = max(which(mmLC[[12]] != 0))
K = max(apply(mmLC[[13]], 1, function(x) max(which( x != 0))), na.rm = TRUE)

# Perform 5 imutations on the dataset 
mmLC <- multilevelLCMI( convData = cd, L = L, K = K, it1 = 2000, it2 = 4000, it3 = 100, 
     it.print = 250, v = 10, I = 5, pri2 = 500, pri1 = 50, priresp = 0.01, priresp2 = 0.01,
     random = TRUE, estimates = FALSE, count = TRUE, plot.loglik = TRUE, prec = 4, scale = 1.0)

# Obtain the dataset completed with the first set of imputations (ind = 1)
complete_data = compData( convData = cd, implev1 = mmLC[[1]], implev2 = mmLC[[2]], ind = 1 )


## End(Not run) 

davidevdt/BMLCimpute documentation built on June 5, 2019, 12:36 a.m.