DRM.jl

Appearance

Simulation plot grammar ​

Status — Implemented

Mirrors drmTMB's Simulation plot grammar. DRM.jl keeps the base package plotting-dependency-free: the helpers below return the numbers a plot needs — grids, deviances, correlations. Optional drawing is provided by DRMMakieExt (Makie + AlgebraOfGraphics weakdeps) via drm_figure / thin plot_* aliases. Default CI does not load Makie or render figures — only the method-less stub is gated.

There are two ingredients: simulate (generate replicate data from a fit) and the plot-data providers (turn a fit into plottable grids).

Simulating from a fit ​

simulate(fit; rng) draws a parametric replicate of the response from the fitted model — the same machinery the parametric bootstrap (bootstrap_ci) uses:

julia
y_rep = simulate(fit)            # one replicate on the response scale

Re-fitting many replicates and summarising the spread of the estimates is the recovery study (ADEMP) pattern used throughout test/ — simulate from known coefficients, fit, and check the truth is recovered. That loop is exactly what a simulation plot visualises.

Plot-data providers ​

All three return a plain NamedTuple of numbers; you supply the backend.

profile_curve(fit, k; npoints = 41, span = 3.0, level = 0.95) ​

The 1-D profile-likelihood curve for coefficient index k. At each grid value of θ[k] the nuisance parameters are re-optimised, so it is a genuine likelihood-ratio profile, not a quadratic/Wald approximation. Returns (x, deviance, estimate, cutoff, k, param, coef, level):

  • x — grid spanning θ̂[k] ± span·se[k], including the MLE exactly;

  • deviance — 2(ℓ̂ − ℓ_profile) at each x;

  • cutoff — the χ²₁(level) reference line the profile interval inverts.

julia
pc = profile_curve(fit, 2)
# plot(pc.x, pc.deviance); hline!([pc.cutoff]); vline!([pc.estimate])

parameter_surface(fit, k1, k2; npoints = 25, span = 3.0) ​

The 2-D profile-deviance surface over two coefficients — the data behind plot_parameter_surface. At each node the remaining parameters are profiled out, so z is the true profile deviance, 0 at the MLE. Returns (x, y, z, k1, k2) with z an npoints × npoints matrix; χ²₂ contours give joint confidence regions.

julia
ps = parameter_surface(fit, 1, 2)
# contour(ps.x, ps.y, ps.z')

corpairs_data(fit) ​

The estimated-coefficient correlation summary behind plot_corpairs — the off-diagonal correlations of the coefficient covariance, ready for a pairs / heatmap panel.

The Confidence Eye ​

Florence's Confidence Eye contract: a pale compatibility region, a darker outline, and a hollow point estimate. With Makie + AlgebraOfGraphics loaded, drm_figure(profile_curve(...); kind = :profile) (or plot_profile) draws that Eye on the profile interval. The Documenter figure gallery may also render Eyes from confint rows — that is docs/gallery evidence, not default-CI evidence.

Because every provider returns numbers (not figures), the base package stays free of a plotting dependency; drawing is opt-in via DRMMakieExt.