Compare unit-level and cell-level effect estimates

Puts the two aggregation levels of a screen side by side and quantifies how much narrower the cell-level intervals are. A ratio well above one is pseudo-replication rather than precision: cells within a unit are not independent, so the cell-level interval understates the uncertainty of the very same estimate.

Usage

cr_compare_levels(unit_effects, cell_effects, by = NULL, estimate = "cohens_d")

Arguments

unit_effects

Effect grid computed with unit set, from cr_effect_grid().

cell_effects

Effect grid computed on the same data without unit.

by

Join keys. NULL (default) uses every non-numeric column the two grids share, which is the stratifying columns plus contrast.

estimate

Name of the effect size column to compare.

Value

A tibble with the join keys, estimate_unit, estimate_cell, width_unit, width_cell and ratio (width_unit / width_cell). The median ratio is attached as the median_ratio attribute.

See also

cr_effect_grid().

Other effect sizes: cr_effect_grid(), cr_effect_size()

Examples

set.seed(1)
cells <- data.frame(
  compound = rep(c("CompoundA", "CompoundB"), each = 240),
  arm = rep(rep(c("reference", "interval_short"), each = 120), 2),
  unit_id = rep(paste0("u", 1:8), each = 60),
  log2_fc = stats::rnorm(480)
)
cells$log2_fc <- cells$log2_fc +
  ifelse(cells$compound == "CompoundA" & cells$arm == "interval_short",
         -1, 0)
u <- cr_effect_grid(cells, "log2_fc", "arm", "reference",
                    by = "compound", unit = "unit_id")
k <- cr_effect_grid(cells, "log2_fc", "arm", "reference",
                    by = "compound", min_n = 10)
cmp <- cr_compare_levels(u, k)
cmp
#> # A tibble: 0 × 7
#> # ℹ 7 variables: compound <chr>, contrast <chr>, estimate_unit <dbl>,
#> #   estimate_cell <dbl>, width_unit <dbl>, width_cell <dbl>, ratio <dbl>
attr(cmp, "median_ratio")
#> [1] NA