posc_emp() estimates a POSC directly from raw data, with no normality assumption: a spline logistic regression on all pairwise differences (or a random subset of max_pairs of them for large samples), set up so that P(0) = 0.5 and P(-delta) = 1 - P(delta), with a person-level bootstrap confidence band.
posc_emp() also computes observed proportions in bins of the difference (bins = number of equal-count bins or a vector of breakpoints). plot() draws them as points with simultaneous (family-wise) confidence intervals, computed by a max-t bootstrap on the empirical-logit scale so they are asymmetric and stay within [0, 1]. posc_bins() returns them as a table.
plot(x, mark = ...) highlights chosen differences: a guide line runs from the x-axis up to the curve, with the probability labelled (add mark_ci = TRUE for its confidence interval). Works for single curves and comparisons.
plot(x, full = TRUE) draws the full sigmoid over negative and positive differences.
posc_comp() accepts any number of curves (named or with labels) and uses a homogeneity test of the correlations when there are more than two.
posc_multi() and posc_lm() gain adjust, which uses the adjusted multiple correlation by default to correct the optimism of in-sample R.
Redesigned plots: by default just the curve and its confidence band, with axis lines, very light gridlines and a probability axis in steps of 0.1. Options add percentages (percent = TRUE), remove gridlines (grid = FALSE), an automatic title and subtitle (title = TRUE, subtitle = TRUE), and for empirical curves the binned proportions (bins = TRUE) and normal-theory curve (reference = TRUE). Comparisons use a colour-blind-safe palette.
Development
The core of the package was written by the author; Claude (Anthropic) assisted with extending, documenting and testing the code in this version, under the author’s direction and review.
Bug fixes
The x-axis of posc_uni(), posc_multi() and posc_lm() plots was stretched by a factor of sqrt(2) relative to the plotted probabilities. Differences are now on the predictor’s own scale.
posc_lm() referred to an object mdl that did not exist and assumed the response was named y; it now works with any lm() model.
Functions no longer change the global warn option.
Breaking changes
Curves are computed for signed differences, and always as the probability of a higher outcome. With a negative relationship the curve falls below 0.5, rather than being relabelled as the probability of a lower outcome.
ci.lvl is now level; colours are set in plot() rather than when the curve is created; posc_multi(outcome_idx =) is now outcome = (the old name still works, with a warning).
The returned object is a "posc" list with a curve data frame (delta, p, lower, upper), replacing predict_matrix and posc_plot_object.
Plots are no longer drawn automatically; call plot() on the result.
