Ier Robustness

Robustness Strategy (ier-robustness)

When to trigger

• The headline number is suspected to hinge on a functional-form choice (CES/log/quadratic) or a single calibrated parameter
• A referee asks "is this robust?" and the draft answers with a mechanical appendix of re-runs
• Inference looks fragile — few clusters, serial correlation, or numerically unstable standard errors
• For a structural paper, the counterfactual moves a lot when you re-anchor an externally-set parameter
• You need to know which robustness checks an IER referee will actually demand vs. which are noise

The IER robustness principle: organize by threat, not by list

IER referees reward robustness that is tied to the load-bearing assumption identified in ier-theory-model/ier-identification. A list of twenty re-runs reads as defensive; three checks that each retire a specific threat to the result read as confident. Build the robustness section as a threat map:

The sequence

1. List the result's threats in priority order — the first one or two are the ones tied to the load-bearing assumption. Those go in the main text; the rest go to the appendix.
2. For each main-text threat, run the single most decisive check, not three weak ones.
3. Report robustness as a range or function where a parameter is uncertain ("welfare gain is 3.8–5.1% across the defensible elasticity range") rather than a single point with a footnote.
4. When a check moves the result, say so honestly and bound the movement — IER referees trust authors who report fragility and contain it more than authors who hide it.
5. Keep significance reporting clean (standard errors / confidence sets), not asterisk-driven.

Robustness for theory and structural papers (not just applied)

At a theory-leaning journal, "robustness" is broader than re-running regressions. For a theory paper it means showing the result survives perturbing the load-bearing assumption (the perturbation test from ier-theory-model) — that is the robustness section. For a structural paper it means three distinct things: (a) economic robustness — does the counterfactual survive an alternative preference/technology specification; (b) parameter robustness — does it survive varying the externally-set parameters over a defensible range; and (c) numerical robustness — does it survive re-solving on a finer grid with a different solver. An IER referee distinguishes these and will not accept (b) as a substitute for (c) or (a). Map each explicitly.

Worked example (illustrative)

A quantitative spatial model reports a welfare gain of 4.5% from removing a trade barrier. The threat map flags the externally-set migration elasticity as the prime suspect. Instead of one re-run, the authors report the welfare gain as a function of the elasticity across its literature-supported range — say 3.6% to 5.4% — and note that the qualitative conclusion (positive, economically meaningful) holds throughout. They then re-solve the model on a doubled grid to confirm the number is not a discretization artifact. This reads as confident and bounded; a single "robust to alternative elasticity" footnote would have invited exactly the objection it tried to dodge.

What goes in the main text vs. the online appendix

With a ≤50-page ceiling, the robustness section competes for space with the contribution itself, so be ruthless about placement. The main text holds the one or two checks that retire the threats tied to the load-bearing assumption — these are part of the argument, not an addendum. Everything else (alternative control sets, secondary subsamples, placebo variants) goes to the online appendix, referenced in one sentence. A main text crowded with low-value re-runs signals defensiveness; a main text with two decisive checks and a clear pointer to the appendix signals an author who knows which threats matter.