Zero-safe centred log-ratio of each sample's population composition

WHY ZEROS NEED HANDLING: log(0) is -Inf, and a genuinely absent population (0 cells scored) is common for rare subsets in a small sample. Dropping those rows would drop exactly the populations most likely to differ between cohorts.

Usage

clr_frequencies(
  freq,
  value_col = "pct_of_cd45_pos",
  populations = NULL,
  delta_frac = 0.65
)

Arguments

freq

population_frequencies table

value_col

the compositional column (percentages)

populations

optional restriction to a closed set; defaults to every population present, the right closure when they partition CD45+, as the table from exclusive_label_frequencies() does

delta_frac

Fraction of the smallest non-zero part used to replace zeros. Default 0.65.

Value

freq plus a clr column, or NULL

Details

WHY MULTIPLICATIVE REPLACEMENT: the standard treatment for rounded zeros in compositional data (Martin-Fernandez et al. 2003). Each zero takes a small delta and the non-zero parts are scaled down to preserve the total, so the result is still a composition. Delta defaults to 65% of the smallest non-zero part observed, the usual choice when no detection limit is declared.

WHY THE GEOMETRIC MEAN IS PER SAMPLE: the constraint is per sample – one sample's populations sum to 100, not one population's samples. Centring across samples instead would be a different, and wrong, operation.