ecj-robustness

Robustness & Alternative Explanations (ecj-robustness)

When to trigger

• The headline result is one regression with one set of choices
• You have not ruled out the obvious competing economic explanations
• A structural result has not been shown to survive perturbing key assumptions
• You suspect a referee will say "this is fragile" or "this is mechanism A, not your mechanism B"

The EJ logic of robustness

At EJ, robustness is not a ritual table of "still significant." It is an argument that the economic interpretation survives, and that rival mechanisms are ruled out — and because the interpretation is what makes the result broadly interesting, fragility there is fatal. A demanding general-interest referee thinks adversarially: which alternative economic story produces the same coefficient, and how do you exclude it? Over-reliance on a single specification is an explicit anti-pattern. And because acceptance triggers the EJ Data Editor rerunning your code against the Zenodo deposit (RES policy, DCAS-endorsed; see ecj-replication-package), every robustness number must come from code that actually executes and reproduces — fragility you papered over will surface in verification. Distinguish three jobs:

1. Specification robustness — the number is not an artifact of arbitrary choices.
2. Mechanism discrimination — your channel, not a competing one, drives it.
3. External / structural validity — the result generalizes (it must, for EJ's broad-interest bar) / the model's conclusions are not knife-edge.

For a short paper (AER:Insights-style), keep the headline robustness in the tight exhibit budget and push the rest to an online appendix — but the leading rival mechanism must still be addressed.

What to run

Specification robustness

• Vary controls (parsimonious → saturated); show coefficient stability and use Oster (2019) δ / bounds for selection on unobservables.
• Alternative functional forms, sample windows, and exclusion of influential subsamples.
• Alternative standard-error structures (clustering level, wild bootstrap with few clusters).
• Inference robustness: randomization inference or permutation tests where design allows.

Mechanism discrimination (the broad-interest part)

• Name the 2–3 alternative economic mechanisms that could generate the same reduced-form sign.
• For each, design a test that the alternatives fail and your mechanism passes (heterogeneity only your channel predicts, an auxiliary outcome, a dose-response the rival cannot explain).
• Triangulate: a second data source, a second identification strategy, or a structural-vs-reduced-form cross-check.

External validity / structural papers

• Argue (or bound) how far the estimate generalizes — central to EJ's broad-relevance claim.
• Sensitivity of estimates and counterfactuals to identifying assumptions and to fixed/calibrated parameters.
• Untargeted-moment fit; over-identification evidence; alternative model specifications that nest or rival the baseline.

Checklist

• Coefficient stability across control sets shown; Oster-style selection bound reported
• Sample-window / outlier / subsample sensitivity reported
• Inference robust to clustering choice / few clusters
• The 2–3 rival economic mechanisms are named and tested against
• At least one triangulation (second data source, design, or structural cross-check)
• External validity / generalizability argued or bounded (EJ broad-interest bar)
• Structural results shown not to be knife-edge in key assumptions
• Robustness lives in the appendix/online appendix, with main text stating the punchline

Anti-patterns

• A wall of "still significant" tables that never address why the effect is your mechanism
• Treating robustness as cosmetic while the headline rival explanation goes untested
• Reporting only specifications that work; hiding the fragile ones (a referee will ask, and the EJ Data Editor reruns the code)
• Selection-on-unobservables waved away with "we control for X" and no bound
• Asserting the result generalizes without any external-validity evidence
• Burying so many checks in the main text that the economic story is lost (use the online appendix)

Output format

【Headline result】coefficient + interpretation
【Spec robustness】[controls, windows, SEs, Oster δ, ...]
【Rival mechanisms】1... 2... — test that discriminates each
【Triangulation】second source / design / structural cross-check
【External validity】how far it generalizes (or bound)
【Residual fragility】honest statement of what is not bulletproof
【Next】ecj-tables-figures