Aejmac Robustness

Robustness Program (aejmac-robustness)

When to trigger

• The headline number rests on one specification, one sample, one lag length, or one grid
• A referee could ask "is this an artifact of [choice]?" and you have no panel of alternatives
• The empirical IRF and the model-implied response are compared but only at the baseline
• A structural/calibrated result has never been re-run under alternative targets

The AEJ: Macro robustness bar

Macro inference is fragile in characteristic ways: short effective samples, structural breaks (Great Moderation, ZLB, COVID), specification forks (lag length, detrending, prior, calibration target), and method dependence (SVAR vs. LP; perturbation vs. global). The AEJ: Macro robustness bar is to show the headline quantity survives the choices a skeptical macro referee would flip, and to be honest where it does not. Robustness is not a graveyard of extra tables — it is a targeted defense of the specific number the paper claims.

A macro robustness program (build the panel)

Empirical (SVAR / LP / narrative)

• Sample splits: pre/post-1984 (Great Moderation), exclude/keep the ZLB period, exclude COVID; report whether the response is stable.
• Specification: lag length, detrending/filtering choice (HP vs. one-sided vs. none), control set, levels vs. differences.
• Method cross-check: if SVAR is baseline, corroborate with LP (and vice versa); agreement is strong evidence.
• Inference: alternative HAC bandwidths / clustering; weak-instrument-robust bands for proxy-VAR/LP-IV.
• Identification variants: alternative orderings / sign sets / instrument constructions.

Quantitative (DSGE / HANK / structural)

• Alternative calibration targets and parameter ranges; show how the headline quantity moves.
• Alternative solution method / accuracy (higher perturbation order, finer grid) where nonlinearity matters.
• Alternative model elements (Taylor-rule coefficients, adjustment costs, market structure) the referee will name.
• Estimation: alternative moments / priors; re-estimate on a subsample.

Cross-cutting

• External validity: another country / dataset / period where the mechanism should also hold.
• Placebo / falsification: a response that should be zero (pre-shock leads; a non-targeted series).

Reporting discipline

• Lead with a one-paragraph summary of what is robust and what is not, then a compact robustness table/figure.
• Keep the baseline number visible in every robustness exhibit so the reader sees the movement.
• Put the bulk in the online appendix; main text carries the decisive checks only.
• A spec-curve / multiverse plot is powerful for empirical macro when many forks exist.

Execution bridge (StatsPAI / Stata MCP)

Run the battery, don't just enumerate it. Full map: execution-with-mcp. AEJ: Macro mixes empirical and structural work — local projections (local_projections / irf) are in StatsPAI, but DSGE / calibration estimation is outside this causal-inference toolchain.

• Many outcomes / specifications: romano_wolf (step-down FWER, accounts for cross-test correlation) or benjamini_hochberg — report the adjusted threshold.
• OVB sensitivity: oster_delta / sensemakr — the confounder strength that would overturn the headline.
• Inference: wild_cluster_bootstrap (few clusters), twoway_cluster / conley.
• Re-fit off one handle: audit_result(result_id) lists the missing checks and the exact suggest_function for each — no guessing the battery.
• Exhibits: etable / did_summary_to_latex from the handle — no retyped numbers.

Keep the decisive checks in the body and the exhaustive (now actually-run) battery in the appendix. See the executed chain in the JF execution walkthrough.

Checklist

• The specific choices a referee would flip are enumerated
• Sample splits across the relevant macro breaks (Great Moderation / ZLB / COVID)
• Specification forks (lags, filtering, controls) tested with baseline shown alongside
• Method cross-check (SVAR↔LP, or perturbation↔global) where both are plausible
• Quantitative: alternative targets/parameters move the headline within a stated range
• Placebo/falsification and at least one external-validity check
• Honest statement of where the result weakens, not just where it holds

Anti-patterns

• A wall of robustness tables that never restate the baseline, so movement is invisible
• Testing only the choices that confirm the result; omitting the obvious adversarial fork
• Ignoring the ZLB/COVID break in a sample that spans it
• Claiming robustness from one alternative specification
• Hiding a fragile headline behind a forest of irrelevant checks
• "Available upon request" instead of an online-appendix robustness section

Worked vignette: is the fiscal multiplier a Great-Moderation artifact? (illustrative)

A paper reports a fiscal multiplier of 1.2 from a proxy-VAR on 1960–2019. A referee suspects it is driven by the volatile pre-1984 period. The robustness program: re-estimate on 1984–2019, exclude the ZLB years, and corroborate with local projections using the same narrative instrument. Suppose the multiplier is 1.2 full sample, 1.0 post-1984, 1.4 at the ZLB, all with overlapping bands, and the LP cross-check agrees within 0.1 — the paper then claims a multiplier "around 1.0–1.4 depending on the monetary regime," which is more credible and more interesting than the single number (illustrative).

Output format

【Headline quantity defended】... (baseline value)
【Empirical robustness】sample splits / specs / method cross-check / inference variants
【Quantitative robustness】alt targets / parameters / solution accuracy
【Placebo + external validity】...
【Where it weakens (honest)】...
【Next step】aejmac-tables-figures