Pool SUEST systems across multiple imputations

suest_mi() combines suest() results fitted in each of several imputed datasets, applying Rubin’s rules to the full set of coefficients and their joint covariance matrix. Supply a list of compatible suest() results, the mice::mira object returned by with() when it fits a SUEST system, or the result list from mitools::with.imputationList().

Usage

suest_mi(fits)

# S3 method for class 'suest_mi'
coef(object, ...)

# S3 method for class 'suest_mi'
vcov(object, ...)

# S3 method for class 'suest_mi'
nobs(object, ...)

# S3 method for class 'suest_mi'
summary(object, conf.level = 0.95, ...)

# S3 method for class 'suest_mi'
print(x, ...)

Arguments

fits

A list containing at least two compatible "suest_model" objects, one per imputation; a "mira" object containing those fits; or a result list returned by mitools::with.imputationList().

object, x

A "suest_mi" object.

…

Additional arguments; currently ignored.

conf.level

Confidence level for summary.suest_mi().

Value

A "suest_mi" object containing pooled coefficients, total, within-imputation, and between-imputation covariance matrices, Rubin large-sample degrees of freedom, and the original SUEST fits.

Details

The pooled covariance is Ubar + (1 + 1 / m) B, where Ubar is the mean within-imputation joint covariance, B is the between-imputation covariance of the complete joint coefficient vector, and m is the number of imputations. Because each within-imputation input is a complete SUEST system, cross-model covariance is retained in both components.

A "mira" object from mice::with() is read through its public analyses component. The list returned by mice::getfit() is accepted as well. The mice package is not required when pooling an ordinary list.

Results from mitools::with.imputationList() use the same list-based route, with their originating call retained in the returned object. Legacy "imputationResultList" wrappers are accepted as well. The mitools package is not required for an already-created result list.

Marginal effects

Predictions, slopes, comparisons, and hypothesis tests from marginaleffects are estimated within each imputation, using that imputation’s joint SUEST covariance, and then pooled with Rubin’s rules, as Stata’s mimrgns does. Cross-model differences, such as hypothesis = difference ~ revpairwise, are pooled the same way. This works on the "mira" object from with(imp, suest(...)) and equally on a "suest_mi" object, so the two give identical results:

analyses <- with(imp, suest(glm(y ~ x, family = binomial),
                            glm(y ~ x + z, family = binomial)))
marginaleffects::avg_comparisons(analyses, variables = "x",
  hypothesis = difference ~ revpairwise)
marginaleffects::avg_comparisons(suest_mi(analyses), variables = "x",
  hypothesis = difference ~ revpairwise)

Without newdata, each imputation’s completed data are used when the imputations came from mice. For a plain list or mitools results, each imputation’s own estimation sample is used instead, and marginaleffects warns that it could not recover the original data from a mids object; that warning does not apply to suest() results. mice warns “Large sample assumed.” because SUEST, like Stata’s suest, reports large-sample (z) statistics; the pooled degrees of freedom are then Rubin’s large-sample values, as in Stata’s mi estimate.

Back to top