Port three robustness analyses onto referee-revised paper

[vahid-ahmadi]

Builds on the merged referee revisions (#13). Adds three analyses that the referee pass did not include, with all numbers recomputed on this repo's regenerated population so they are consistent with the paper's base, and written in elasticity-free wording (does not re-introduce the "shrinks" / "model-free" / £100-bin / cubic framings #13 corrected).

Added analyses

1. Value-added robustness of the dominated region (model.tex). Real VAT taxes value added, so a firm remitting net rate τ₀ faces a_i = T*·τ₀/(1−τ₀). On near-threshold firms: ~42% are net input creditors / near-zero remitters (effectively no notch — consistent with the ~43% voluntary-registration share); for the rest the value-added dominated region has a weighted median ~£18,650 (≈£44k for high-net-rate consumer-facing firms), the same order as the £21,250 full-rate width. So the correction scopes out the voluntary registrants rather than collapsing the region. Script: results/task1_value_added_dominated_region.py.
2. Estimator recovery / power test (a_inference.tex, tab:recovery; pointer in bunching.tex). Injecting a known bunching signal into the step-free Placebo-A population, the same estimator recovers ~90% (89.6%) at E_true = 2k/5k/8k. So it is both specific (placebo: no false positive) and sensitive (recovery: has power) — the null on the real data is a property of the data, not the method. Script: analysis/recovery_bunching.py.
3. Fiscal-drag projection (conclusion.tex). A frozen £90k threshold draws ~14% more firms into the dominated region by 2028–29 (~130k→148k); the just-below, bunching-exposed band grows ~9.5% (~41k→45k) — consistent with the OBR fiscal-drag narrative. Script: results/task2_fiscal_drag_projection.py.

Notes

• Numbers were re-run on the regenerated original-model population in this clone (not the other clone's bespoke file, which gave an inconsistent 133k figure and — when mis-based — an OBR-contradicting result). Fiscal drag is correctly based on the 2023-24 turnover distribution aged forward.
• Build: clean (latexmk -lualatex), 36 pp, all refs/citations resolve.
• The deeper v_i / negative-VA model issue remains tracked separately in Fix per-firm VAT liability scaling (value-added → net VAT) + repair bunching reproduction #15 and is not addressed here.

🤖 Generated with Claude Code

[vahid-ahmadi]

Review — looks good, mergeable ✅

Checked the diff against the regenerated population, the scripts, and CI.

Verified

• CI green (ruff / pytest / compile + Quarto), no base conflicts.
• Numbers match their scripts and are reproducible from this repo (with appropriate rounding):
  • value-added: script 41.5% → "~42%"; median £18,650 (exact); top-quartile £43,822 → "~£44k";
  • fiscal drag: 130,481 → 148,403 (+13.7% → "~14%"); just-below +9.5% (41k → 45k);
  • recovery: 89.6% at E_true = 2k/5k/8k.
• No framing regressions: none of the corrected "shrinks" / "model-free" / £100-bin / cubic phrasings reappear; the additions are elasticity-free throughout.
• Fiscal drag is correctly based on the 2023-24 turnover distribution aged forward — not the already-aged £90k vintage (which double-counts and produced an OBR-contradicting result where the just-below band shrank).

One substantive note (non-blocking). In model.tex, the dominated-region subsection now states both that ~137,000 firms are "exposed" to the band and that ~42% of near-threshold firms face no notch (net input creditors). These are measured over different ranges (£85k–£106.25k vs £80k–£90k), so they don't contradict — but a reader could perceive tension. Consider one clause tying them together: the 137k is the exposed population over the band, of which the ~42% net-creditors face no binding notch, so the genuinely-distorted count is smaller. Sharpens the honesty without touching the headline.

Minor

• results/task2_fiscal_drag_projection.py hardcodes the FISCAL_YEARS growth factors that also live in static/model.py — fine for now, but could drift; consider importing them.
• The recovery test's E_recovered is the deficit-differenced signed-excess measure (defined in the appendix); since that's non-standard, keeping the one-line definition prominent is worth it.

Caveat (already documented). This sits on the original model; the deeper v_i / negative-VA issue is tracked in #15 (now draft PR #17). The value-added robustness paragraph here is, in effect, an honest in-paper acknowledgment of part of that same issue — which is a strength, not a problem.

Recommendation: mergeable as-is. The note above is a nice-to-have, not a blocker.