uncertainUCDP (version 0.7.6)

models_rubin_rules: Pool regression coefficients and standard errors via Rubin's rules

Description

Pools coefficient estimates and standard errors from multiple fitted models (typically one per dataset) using Rubin's rules, returning a `coeftest`-like matrix with pooled estimates, SEs, t-statistics, and p-values. Can be used when when pooling mulitple models with different draws from the UCDP uncertainty model.

Usage

models_rubin_rules(list, df = NULL, cluster = NULL)

Value

A numeric matrix with the same layout as `lmtest::coeftest()`, containing pooled coefficients, pooled standard errors, t-statistics, and two-sided p-values. The returned object uses the first element's `coeftest` matrix as a template (including row/column names and attributes).

Arguments

list

A list of fitted model objects (e.g., `lm`, `glm`, etc.) for which `lmtest::coeftest()` methods exist.

df

Optional. Degrees of freedom for t-distribution when computing p-values. If `NULL`, uses the degrees of freedom from the first model's `coeftest` output, i.e. the df from the original model.

cluster

Optional. A clustering variable passed to `sandwich::vcovCL()`. Typically a one-sided formula specficying clusters, e.g., `~ cluster_id`

Details

If `cluster` is provided, cluster-robust covariance matrices are computed for each model via `sandwich::vcovCL()` and passed to `lmtest::coeftest()`.

The pooling follows the usual within/between decomposition: within-imputation variance is the mean of squared SEs, and between-imputation variance is the sample variance of point estimates across imputations. The total variance is \(W + (1 + 1/m)B\) where \(m\) is the number of models.

Examples

library(foreach)
library(dplyr)
# Draw 100 datasets from UCDP uncertainty model and return as a list
data('ucdpged')
ucdp_sb <- ucdpged %>% filter(type_of_violence == 1)
list_of_models <- foreach(i = 1:100) %do%{
 lm(ged_sb_draw ~ year + region, data =
ucdp_sb %>% mutate(ged_sb_draw = runcertainUCDP(n = n(), fatalities = best, tov = 'sb')))
}
#' # Pool model coefficients and standard errors via Rubin's rules
pooled_results <- models_rubin_rules(list_of_models)