Skip to contents

Pool Model-Level Fit Statistics over Imputations

Source: R/glance_imputations.R
glance_imputations.Rd

Companion of [pool_imputations()] for the one-row-per-model summaries (`broom::glance()`, the `glance` of [cox_single()]): concordance, Brier or integrated Brier score, AIC, degrees of freedom and so on, over the same model fitted to each of `m` imputed data sets.

Usage

glance_imputations(
  x,
  imputation_col = "imputation",
  conf_level = 0.95,
  df_complete = NULL
)

Arguments

x

A list of one-row data frames (or tibbles), or a stacked data frame with a column naming the imputation.

imputation_col

Name of the imputation column when `x` is stacked.

conf_level

Confidence level for statistics that have a standard error.

df_complete

Complete-data degrees of freedom (for example `N - k`) for the Barnard-Rubin (1999) small-sample adjustment of the interval, passed to [pool_imputations()]. `NULL` (default) uses the large-sample Rubin degrees of freedom. Only affects statistics that have a `std.error.<stat>` column.

Value

A tibble with `statistic`, `m`, `estimate`, `between_sd`, `min`, `max`, and, where a standard error was available, `std.error`, `conf.low`, `conf.high` and `fmi` (otherwise `NA`).

Method

Every numeric column is pooled as the mean over the imputations, reported with the between-imputation standard deviation, minimum and maximum. This is the usual summary for fit statistics, which have no Rubin variance of their own. If the column `std.error.<stat>` exists (as `broom::glance()` gives for `concordance`), the statistic is instead pooled with Rubin's rules through [pool_imputations()], giving a standard error, a t interval and the fraction of missing information. Pooled concordance is a mean of C-indices, each computed on its own completed data set; it is not a C-index of a pooled model. Columns named `std.error.*` are not reported as statistics.

See also

[pool_imputations()]

Examples

set.seed(1)
g <- lapply(1:5, function(i) data.frame(concordance = 0.7 + rnorm(1, sd = 0.01),
                                        std.error.concordance = 0.03, AIC = 400 + rnorm(1, sd = 2),
                                        df = 3))
glance_imputations(g)
#> # A tibble: 3 × 10
#>   statistic       m estimate between_sd     min     max std.error conf.low
#>   <chr>       <int>    <dbl>      <dbl>   <dbl>   <dbl>     <dbl>    <dbl>
#> 1 concordance     5    0.700    0.00665   0.692   0.706    0.0309    0.639
#> 2 AIC             5  401.       1.87    398.    403.      NA        NA    
#> 3 df              5    3        0         3       3       NA        NA    
#> # ℹ 2 more variables: conf.high <dbl>, fmi <dbl>