Compare an Original and an Imputed Version of One Domain's Dataset

Bundles the imputation diagnostics for one set of related variables (a "domain") into a list of plain R objects, so a report can loop over the domains and show or assemble the pieces itself. Nothing is rendered.

Usage

compare_domain_datasets(
  orig_df,
  imp_df,
  vars = NULL,
  id_col = NULL,
  time_col = NULL,
  pca = TRUE,
  test = c("t", "wilcoxon")
)

Arguments

orig_df: Data frame before imputation.
imp_df: Data frame after imputation: the same rows as `orig_df`, in the same order.
vars: Character vector of numeric columns to compare. `NULL` (the default) uses the numeric columns present in both data frames, without `id_col` and `time_col`.
id_col, time_col: Optional names of the subject id and time columns. Both are needed for the trajectory pieces.
pca: Logical; also compare PCA loadings?
test: Paired test for the trajectory comparison: `"t"` ([stats::t.test()]) or `"wilcoxon"` ([stats::wilcox.test()]).

Value

A list with elements `ecdf_plots` (ggplot, or `NULL` if `orig_df` has no missing value among `vars`), `ks_tests` (tibble), `corr_diff_table` (tibble), `corr_diff_heatmap` (ggplot, or `NULL` with fewer than two `vars`), `pca_loading_diff` (ggplot or `NULL`), `trajectory_plot` (ggplot or `NULL`) and `trajectory_deltas` (tibble with `variable`, `time`, `mean_orig`, `mean_imp`, `delta`, `n_paired`, `p_value`, `p_adj`, or `NULL`).

Details

The two data frames must describe the same rows in the same order: `imp_df` is `orig_df` after imputation. When `id_col` is given and present in both, the ids (and `time_col`, if given) must agree row by row, and the function stops otherwise.

The pieces are: * **Distributions**: the imputed values against the observed ones, with [compare_imputed_ecdf()] and [test_imputed_distributions()] (Kolmogorov-Smirnov and t tests). Only cells that were missing in `orig_df` count as imputed. * **Correlations**: [correlation_diff()] and [correlation_diff_heatmap()], comparison (`imp_df`) minus baseline (`orig_df`). * **PCA loadings**: with `pca = TRUE`, a [stats::prcomp()] fit (centred and scaled) to the complete cases of `vars` in each data frame, then [pca_loading_diff_heatmap()]. `NULL`, with a warning, if there are fewer than two variables, fewer complete cases than variables plus one, or a variable without variance in either fit. * **Trajectories**, when `id_col` and `time_col` are both given: the mean and standard error of each variable at each time point in each data frame, the difference of the means, and a paired comparison by `id_col` within each time point (`test`).

The paired comparison uses the rows that have a value in both data frames, so for an imputation that leaves the observed values unchanged (the usual case) every paired difference is zero and `p_value` is `NA`; `delta` still shows the shift in the mean that the imputed values cause. The `p_value`s are point-wise, one per variable and time point; `p_adj` adjusts them with Holm's step-down method over all of them.

Examples

set.seed(1)
d <- expand.grid(id = 1:30, time = 1:3)
d$hr <- rnorm(nrow(d), 80, 10)
d$map <- rnorm(nrow(d), 90, 8)
orig <- d
orig$hr[sample(nrow(d), 15)] <- NA
imp <- orig
imp$hr[is.na(imp$hr)] <- mean(imp$hr, na.rm = TRUE)
out <- compare_domain_datasets(orig, imp, id_col = "id", time_col = "time")
out$ks_tests
#> # A tibble: 2 × 8
#>   variable n_observed n_imputed mean_observed mean_imputed ks_statistic
#>   <chr>         <int>     <int>         <dbl>        <dbl>        <dbl>
#> 1 hr               75        15          80.9         80.9         0.52
#> 2 map              90         0          90.1         NA          NA   
#> # ℹ 2 more variables: ks_p_value <dbl>, t_p_value <dbl>
out$trajectory_deltas
#> # A tibble: 6 × 8
#>   variable  time mean_orig mean_imp   delta n_paired p_value p_adj
#>   <chr>    <int>     <dbl>    <dbl>   <dbl>    <int>   <dbl> <dbl>
#> 1 hr           1      81.2     81.1 -0.0451       25      NA    NA
#> 2 hr           2      79.8     79.9  0.186        25      NA    NA
#> 3 hr           3      81.7     81.6 -0.141        25      NA    NA
#> 4 map          1      90.9     90.9  0            30      NA    NA
#> 5 map          2      87.4     87.4  0            30      NA    NA
#> 6 map          3      91.9     91.9  0            30      NA    NA