Compare a between-group difference against the gate uncertainty behind it

Adds two columns to a group comparison table: the typical within-sample uncertainty on the quantity being compared, and the ratio of the observed difference to it.

Usage

annotate_gate_uncertainty(gstats, ufreq, freq = NULL)

Arguments

gstats

a table from stats_group_comparison()

ufreq

the frequencies element of run_gate_uncertainty()

freq

the frequency table the comparison was run on, needed only when gstats compared something other than pct_of_cd45_pos.

Value

gstats with gate_u_pct_points, difference_over_gate_u, total_u_pct_points and difference_over_total_u appended, unchanged if the uncertainty table is missing

Details

READ IT AS A SCREEN, NOT A TEST. The gate uncertainty is partly common to every sample in a run, since they share a panel, a transform and a placement rule, so it cancels to some degree in a difference and the ratio is conservative. A ratio below 1 says the groups differ by less than the typical distance the cut itself can move, which is a reason to look at threshold_uncertainty.csv before interpreting the result, not a p-value.

TWO RATIOS, NOT ONE. difference_over_gate_u compares the difference against gate placement alone and keeps exactly the value it has always had. difference_over_total_u compares it against placement and counting together, and is the stricter of the two. They separate for a rare population, where the cut can be well placed and the frequency still be built on too few events: the first ratio passes and the second does not, and the second is the one to believe.

THE RATIO IS TAKEN IN THE UNITS THE TEST COMPARED. The uncertainty is computed in percentage points of CD45+; when the comparison was made on cells/uL (measure is cells_per_ul), a difference in cells/uL divided by an uncertainty in percentage points is no ratio at all. Each sample's uncertainty is then carried into cells/uL by that sample's own factor, cells_per_ul / pct_of_cd45_pos (its WBC / 100), before the median is taken. The *_pct_points columns stay in percentage points. Without freq, or for a measure with no such factor, the ratios are left NA.