Plotting 1.1 — cross-sectional experiments (RegressionKink, PrePostNEGD, DiD)

[drbenvincent]

Child of #1155 (1.1 declarative plotting migration). Depends on children A and B. Independent of children D and F.

Workflow

Branch from pymc6_and_pymcmarketing1_migration and open the PR against pymc6_and_pymcmarketing1_migration. Not main. See the parent issue for why.

Scope

Three single- or two-panel experiments, in increasing order of difficulty. One commit and one visual gate each.

RegressionKink (_plot, 82 lines)

The cleanest remaining win. Scatter plus one posterior band and line, plus a geom_vline at the kink point, with a title carrying the Bayesian R² and the change-in-gradient CI. Structurally identical to RD, so follow child C's implementation directly.

PrePostNEGD (_plot, 89 lines)

Two facets: pre/post scatter with control and treatment posterior predictive bands on top, treatment-effect posterior below with a zero reference line and credible-interval bounds (replacing az.plot_posterior). #990 reached a full ggplot here. Two things it had to fix afterwards, worth doing right the first time: the scatter series must be labelled with the same names as the posterior bands so points, lines, and bands share one legend with the fill guide suppressed; and plotnine's default panel spacing leaves an oversized vertical gap between the two facets.

DifferenceInDifferences (_plot, 219 lines)

The worked example for component-wise migration. Five components:

Component	Treatment
Raw data scatter	geom_point
Control group band + line	geom_ribbon + geom_line
Treatment group band + line	geom_ribbon + geom_line
Counterfactual, multiple post-treatment periods	geom_ribbon + geom_line
Counterfactual, single post-treatment period	stays matplotlib — continuous-x violin, no plotnine geom
Causal-impact arrow and text	stays matplotlib — reuse _plot_causal_impact_arrow unchanged

Build the shared base once across both counterfactual cases; only the counterfactual component branches. Do not create a plotnine-only parallel path for the multi-period case — that is the guardrail G2 anti-pattern that cost #990 a cleanup commit. Returns (fig, ax).

DiD also needs dtype-aware x-axis handling: the time column can be numeric, boolean, categorical, or datetime, and #990 had to add explicit handling for all four.

Files likely touched

• causalpy/experiments/regression_kink.py, prepostnegd.py, diff_in_diff.py
• Corresponding tests

Acceptance criteria

• G0 inventory posted per experiment before its implementation.
• All three kind values work for all three experiments.
• DiD: the existing violin and arrow snippets are reused unchanged, not rewritten; a test covers each of the four x dtypes.
• PrePostNEGD: one unified legend across points, lines, and bands; facet spacing checked visually.
• Public signatures and (Figure, Axes) returns unchanged.
• Before/after images per experiment per backend, each visually approved before the next is started.
• git diff --stat reported per commit (guardrail G3). A modest net reduction is the realistic expectation for DiD, not a full rewrite.
• prek run --all-files passes; make test-patch-cov passes.

Out of scope

• RegressionDiscontinuity (child C).
• IPW plots — mosaic and ECDF, no plotnine equivalent, out of scope for the whole program.