scale-sensitive-diagnostics
Diagnostics for scale-sensitive transforms and count outcomes: unit rescaling inside ln(1+x) and asinh, count models on the full sample, count versus per-capita specifications, unstable ratios, and placebos in the corrected specification. Invoke from `code-and-data` whenever ln(1+x), asinh, counts, per-capita scaling, or weak denominators carry a target result.
isitcredible · Static checks on file
Installation
Read the source and dependencies before running a command. Copying does not install or execute anything.
npx skills add isitcredible/research_skills --skill scale-sensitive-diagnosticsAbout this skill
name: scale-sensitive-diagnostics
description: >-
Diagnostics for scale-sensitive transforms and count outcomes: unit
rescaling inside ln(1+x) and asinh, count models on the full sample, count
versus per-capita specifications, unstable ratios, and placebos in the
corrected specification. Invoke from code-and-data whenever ln(1+x),
asinh, counts, per-capita scaling, or weak denominators carry a target
result.
Scale-sensitive diagnostics
Establish whether the transform is doing the identifying work. ln(1+x)
and asinh look like harmless ways to keep zeros in a log specification, but
they smuggle in a choice of units, and the estimate can move with that
choice: measure in dollars rather than thousands of dollars and the
"elasticity" changes. The theorem behind that move, what the rescaling
sweep measures, and the alternative estimands (Poisson percentage effect,
explicitly calibrated margins, Lee bounds) are in
../../references/methods/chen_roth_2024_logs_with_zeros.md. The exact
asinh elasticity formulas and the data rules papers cite but violate are in
../../references/methods/bellemare_wichman_2020_ihs_elasticities.md, and
the literature-wide non-robustness record plus the harmonized
alternatives (normalized estimands, quantile regression) are in
../../references/methods/fitzgerald_etal_2026_loglike.md. Once a corrected
specification becomes the honest baseline, hold the paper's placebos to it
too.
Trigger: ln(1+x), asinh, a count outcome, per-capita scaling, a floor, or
a ratio over a weak denominator in the outcome or a load-bearing regressor.
NOT_APPLICABLE when no such transform is load-bearing.
Core checks
- Unit-rescaling sweep inside the transform. Re-estimate with the
variable in different units (×10, ×1000, per-thousand) inside the
ln(1+x)or asinh — always alongside its mandatory companion on the same sample, the extensive-margin regression1[Y>0] ~ D. Interpret movement against its prediction (≈ extensive-margin effect × log a; the headline t converges to the extensive-margin t as the scale factor grows): movement equal to the prediction is the expected behavior of a nonzero extensive margin — the estimand conflates margins — and the prescription is Lee bounds on the intensive-margin effect under monotonicity (tight when the extensive margin is small, judged by its interval, not its p-value), not acceptance of the transform; drop zeros and take the ATE in logs only if the extensive effect is zero for every unit (Chen–Roth Remark 7). Reserve "the transform is doing the work" for movement that changes the substantive answer (calibrate against the crib's Table I magnitudes; the claim's own CI is the tolerance). For a transformed regressor the same sweep applies but the estimand is a units-sensitive semi-elasticity (Aihounton–Henningsen 2021; Thakral–Tô 2023), not the Chen–Roth ATE result — do not cite Chen–Roth for the regressor case; √X rescales cleanly. - asinh–asinh is a third case: the coefficient is not an elasticity. When BOTH sides are log-like, β̂ approximates a log-log elasticity only where both inputs are large — the exact implied elasticity is β̂·[√(y²+1)/y]·[x/√(x²+1)] (Bellemare–Wichman 2020), which converges to β̂ only as x, y → large and diverges near zero, where the elasticity is undefined outright. A paper quoting an asinh–asinh coefficient as "the elasticity" without evaluating the formula has not reported an elasticity at all. Compute it at named evaluation points (mean AND median — the gap is often the finding), check the inputs against BW's own data rules (min nonzero |value| ≥ ~10; ≤ ~1/3 zeros), and run the rescaling sweep as a 2-D grid — both sides are units-sensitive, so the coefficient moves under rescaling of either variable.
- The sweet-spot check (Fitzgerald et al. 2026). The published unit scale is itself a researcher degree of freedom: in their 46-paper reanalysis 38% of estimates sit at a local |t| maximum — rescaling in either direction shrinks significance. So the sweep is a hypothesis test, not just a sensitivity display: report where the published scale sits on the |t|-vs-scale surface, and flag compliance failures with the data rules (99.8% of their estimates violate the min ≥ 10 rule; 32% of transformed outcomes and 38% of exposures are non-positive in over a third of observations). The zeros-tolerant alternative that avoids the transform entirely is the normalized estimand — divide outcomes/treatments by a constant (unit mean, control-group mean, SD) so the answer is a percentage deviation with a scale-invariant coefficient and t-stat; use a dispersion normalizer where the mean can be negative. Quantile regression (Liu–Kaplan 2025) is the consistent intensive-margin route.
- Count model on the full sample. Estimate Poisson PML on the untransformed outcome — consistent for the conditional mean on any non-negative outcome, continuous included, not only counts — zeros included, and compare with the transformed specification and its implicit sample. Report e^β − 1 explicitly as θ_ATE%: a percentage of the mean, levels-weighted and potentially tail-dominated — not an average of individual percentages; θ_Avg% ("what a reader thinks an asinh coefficient is") is not point-identified even on strictly positive support. Run it both ways: where the paper headlines a negative-binomial count model, the load-bearing check is the same spec as full-sample Poisson PML (consistent for the conditional mean under weak assumptions — Gourieroux–Monfort–Trognon — while NB is efficient only under its own variance assumption and heavily down-weights high-count observations). NB significant + full-sample Poisson ≈ 0 (or sign-flipped) = the effect is concentrated in a specific count range — identify which before adjudicating, since Poisson's θ_ATE% is tail-dominated by construction