jmcb-robustness

Robustness Strategy (jmcb-robustness)

When to trigger

• The headline (an IRF, an elasticity, a counterfactual welfare number) might flip under nearby choices
• A referee will ask "is this the recursion ordering / lag length / sample window talking?"
• Standard errors look too tight for a bank×time panel or for serially correlated macro data
• The result depends on one regime (a crisis, the ZLB) and you have not shown sub-sample stability
• You have a pile of "robustness" tables but cannot say which threat each one rules out

The JMCB robustness logic

JMCB referees do not reward a wall of additional regressions; they reward checks that are mapped to a named threat to the specific identification. A robustness suite is a list of "the result could be wrong because X — here is the check that rules out X." Given the journal's monetary/banking focus, the recurring threats are: shock contamination, specification dependence (lags, ordering, controls), inference understatement on panels and serially correlated series, and regime/sample instability around crises and policy transitions.

Threat → check map (build yours from this)

| Named threat | Diagnostic / check |

|---|---|

| Shock is contaminated (information effect, anticipation) | Re-identify with info-robust surprises; orthogonalize to forecast revisions; placebo on pre-announcement windows |

| SVAR result is ordering-/restriction-driven | Vary recursive ordering; alternative sign-restriction sets; report the full identified set |

| IRF is lag-length / horizon dependent | Vary VAR lags; local-projection vs. VAR; alternative horizons |

| Panel SEs understated | Two-way (bank and time) clustering; wild-cluster bootstrap with few clusters; Driscoll–Kraay for cross-sectional dependence |

| Result is one-regime (crisis/ZLB) artifact | Split pre/post-2008, exclude crisis, exclude ZLB; state-dependent specification |

| Demand contamination (micro-banking) | Tighter fixed effects (firm×time); single-bank-firm vs. multi-bank-firm subsample |

| Controls are doing the work | Sequentially add controls (Oster-style movement check); show coefficient stability |

| Outliers / measurement | Alternative winsorizing; drop largest institutions; alternative variable definitions |

How to present it

1. Lead with the threats a JMCB referee will actually raise, ordered by how damaging they would be if true.
2. For each, show the headline magnitude next to the baseline so the reader sees stability (or honest movement), not just significance survival.
3. Put the 3–4 load-bearing checks in the main text; relegate the long tail to the online appendix with a pointer (see jmcb-internet-appendix).
4. Where a check does move the result, say so and interpret it — a transparent boundary is more credible than a uniform table of survivors.

Inference deserves its own pass

For JMCB's two dominant data shapes, the default standard errors are usually wrong in a predictable direction:

• Bank/firm panels: a single dimension of clustering understates uncertainty when shocks are common across units within a period. Cluster on both the cross-sectional unit (bank/firm) and time; with few clusters in either dimension, use the wild-cluster bootstrap (Cameron–Gelbach–Miller). If cross-sectional dependence is plausible, report Driscoll–Kraay as a complement.
• Macro time series / local projections: serially correlated errors require HAR/Newey–West or lag-augmentation; for VARs, report bootstrap or bias-corrected bands rather than asymptotic ones at short samples.

State the clustering/inference choice once, prominently, and show the headline survives a reasonable alternative — referees treat a casual one-way-clustered SE as a red flag.

Crisis and regime stability is not optional for long samples

Many JMCB samples straddle the 2008 crisis, the ZLB/QE era, and post-Basel-III regulation. A result that holds only because one of these episodes dominates the variation is fragile. Show the headline in pre/post sub-samples, excluding the crisis window, and — where the mechanism plausibly changes at the bound — in a state-dependent specification (e.g., interacting the shock with a ZLB or high-uncertainty indicator). If the effect genuinely is regime-specific, that is itself a finding; report it as one rather than letting it masquerade as a general result.

Execution bridge (StatsPAI / Stata MCP)

Run the battery, don't just enumerate it. Full map:

[execution-with-mcp](../../../shared-resources/empirical-methods/execution-with-mcp.md). JMCB is monetary/banking — macro time series + bank panels; local projections for the macro lane, DiD/IV for the bank lane.

• Many outcomes / specifications: romano_wolf (step-down FWER) or benjamini_hochberg.
• OVB sensitivity: oster_delta / sensemakr.
• Inference: wild_cluster_bootstrap (few clusters), twoway_cluster / conley.
• Re-fit off one handle: audit_result(result_id) lists missing checks + the exact

suggest_function for each.

• Exhibits: etable / did_summary_to_latex from the handle — no retyped numbers.

Decisive checks in the body, exhaustive battery in the appendix.

[JF execution walkthrough](../../../Journal-of-Finance-Skills/resources/worked-examples/02-execution-walkthrough.md).

Checklist

• [ ] Each robustness check is tied to a named threat to this identification
• [ ] Inference re-examined: clustering dimensions correct; few-cluster and serial-correlation handled
• [ ] Specification dependence shown (ordering/lags/horizon/controls) with magnitudes side by side
• [ ] Sub-sample / regime stability shown (crisis, ZLB, policy transition) if the period spans one
• [ ] Micro-banking: demand-contamination check via tighter fixed effects or single-vs-multi-bank firms
• [ ] Load-bearing checks in main text; long tail in the online appendix with a map
• [ ] Any check that moves the result is reported and interpreted, not hidden

Anti-patterns

• A robustness section that adds controls and reports "still significant" without showing the magnitude
• Twenty appendix tables with no statement of which threat each addresses
• Reporting only the checks that survive and quietly dropping the ones that did not
• Leaving panel SEs one-way clustered when shocks are common across banks in a period
• Claiming generality from a single regime without a crisis/ZLB sub-sample
• Treating statistical-significance survival as the bar when the question is magnitude stability

Don't over-test: a focused suite beats an exhaustive one

A robustness section that runs every permutation signals uncertainty, not rigor. Pick the checks that map to the objections a JMCB referee will actually raise (shock cleanliness, demand contamination, inference, regime stability) and present those in the body with magnitudes side by side. Everything else — alternative winsorization thresholds, dozens of control permutations — goes to the online appendix with a one-line summary in text. The goal is to show the headline is stable where it matters, not to bury the reader.

Worked vignette (illustrative)

An SVAR finds a contractionary monetary shock raises credit spreads. A referee suspects the recursive ordering. The threat-mapped response: re-estimate under three alternative orderings and a sign-restricted scheme, plot the IRFs together, and show the peak spread response stays in a 15–22bp band across all of them (illustrative). One check — using revised instead of real-time data — does shift the peak; the authors report it and argue the real-time version is the policy-relevant one. That honesty reads as strength at JMCB.

Output format

【Journal】Journal of Money, Credit and Banking
【Skill】jmcb-robustness
【Top threats】ranked list of what could make the headline wrong
【Threat → check】each check mapped to the threat it rules out
【Inference fix】clustering dims / few-cluster / serial-correlation handling
【Regime stability】crisis / ZLB / transition sub-samples
【Main vs appendix】load-bearing checks in text; tail mapped to online appendix
【Honest movement】any check that shifts the result + interpretation
【Next skill】jmcb-tables-figures