mi.eval: Multiple-imputation evaluation

View source: R/mult-imp.R

mi.evalR Documentation

Multiple-imputation evaluation

Description

Evaluation of an expression across multiply imputed data sets.

Usage

mi.eval(EXPR, robust, cluster, coef., vcov., df.=NULL,
parallel=FALSE, lazy=NULL, ...)

Arguments

EXPR

An R expression to evaluate. This expression must contain a data argument that specifies a list containing the imputed data sets or a an object of class amelia, mids, or imputationList.

robust

Whether to use Huber-White robust standard errors. The default is TRUE if cluster is specified and FALSE otherwise.

cluster

A vector specifying clusters for the purpose of computing clustered robust standard errors. This can be a variable inside the imputed data set. If unspecified, standard errors are not clustered. If specified, robust cannot be FALSE.

coef.

The function used to get a numeric vector of coefficient estimates when evaluated on an object returned from evaluating EXPR for each data set. The default is to use coef.

vcov.

The function that returns a numeric matrix giving the variance-covariance matrix when evaluated on an object returned from evaluating EXPR for each data set. The default is to use vcovCR if cluster is specified, vcovHC if robust=TRUE and cluster is not specified, and vcov otherwise.

df.

Either the degrees of freedom for each model or a function that calculates degrees of freedom on an object returned from evaluating EXPR for each data set. The default value of NULL uses the minimum result of applying df.residual it returns a numeric value when applied to the object returned by EXPR and this object is not of class glm and Inf otherwise.

parallel

A logical indicating whether to evaluate EXPR across data sets in parallel using mclapply. Otherwise, evaluation is done serially using lapply. NULL means to use parallel evaluation if and only if the 'parallel' package can be loaded and getOption("mc.cores", detectCores()-1L) is greater than 1.

lazy

A logical indicating whether to use lazy evaluation to avoid copying all imputed data sets into memory. When the data argument to EXPR generates the multiply imputed data set, this is generally a bad idea because it means redoing the imputation multiple times. The default value of NULL means to use lazy evaluation if and only if the data argument to EXPR is a name.

...

Any additional arguments to be passed to lapply or mclapply when evaluating EXPR across data sets.

Details

This function evaluates a R command for each of several multiply imputed data sets and combines results across data sets into a single set of estimates. This is similar to the functionality provided by with.mids but also works with multiply-imputed data sets generated by other packages like 'Amelia' as well as those from 'mice'.

For generating formatted tables of regression coefficients, the outputted objects should be compatible with the 'texreg' package. When used with lm, glm or a few other types of models, these objects are also compatible with the 'stargazer' package.

Value

An object of class mi.estimates containing the coefficient estimates, variance-covariance matrix, and related information.

See Also

See Also summary.mi.estimates, with.mids, amelia, mice, coef, vcov and df.residual.

Examples

if (require("Amelia")) {
    ## Load data
    data(africa)
    africa$civlib <- factor(round(africa$civlib*6), ordered=TRUE)

    ## Estimate a linear model using imputed data sets
    model0 <- lm(trade ~ log(gdp_pc), data=africa, subset=year==1973)
    summary(model0)

    ## Impute using Amelia    
    a.out <- amelia(x = africa, cs = "country", ts = "year",
                    logs = "gdp_pc", ord="civlib")

    ## Estimate a linear model using imputed data sets
    model1 <- mi.eval(lm(trade ~ log(gdp_pc), data=a.out, subset=year==1973))

    ## Show estimates
    model1
    coef(model1)

    ## Show summary information
    summary(model1)

    if (require("MASS")) {
        ## Estimate an ordered logit model
        model2 <- mi.eval(polr(civlib ~ log(gdp_pc) + log(population),
                               data=a.out))
        summary(model2)

        ## Also show thresholds by including thresholds with coefficients
        model3 <- mi.eval(polr(civlib ~ log(gdp_pc) + log(population),
                               data=a.out),
                          coef=function(x) c(x$coefficients, x$zeta))
        summary(model2)
    }
}

bucky documentation built on March 26, 2022, 1:12 a.m.