Imfer robustness

Robustness Suite (imfer-robustness)

When to trigger

• The main estimate is in hand and you must show it is not an artifact of one specification
• A referee asks "is this driven by a few countries / a particular crisis episode / the sample window?"
• The result depends on a sample-selection rule, a country grouping, or a deflator/exchange-rate convention
• You suspect a global common shock or one influential economy is doing the work
• You want to pre-empt specification-search and country-cherry-picking concerns

The IMFER robustness bar

IMFER referees probe whether the headline number is stable, honestly inferred, and not the product of country-composition or specification search — and, distinctively, whether it would still guide policy under reasonable alternatives. Robustness here is a targeted set of checks, each tied to a specific threat to an international-macro design, not a wall of regressions. The persuasive object is that the point estimate barely moves when you drop the obvious country, switch the window, or change the inference.

Threat to the result	The check that answers it
One country / region drives it	leave-one-country-out; drop the dominant economy or region; jackknife
A single crisis episode drives it	drop the GFC / euro-crisis / COVID window; alternative event definitions
Global common shock confounds it	add the global financial cycle factor / US-shock control; time effects
Sample / coverage selection	balanced vs. unbalanced panel; alternative country-inclusion rules; advanced-vs-EM split
Specification search	specification curve; pre-registered or primary spec; stepwise controls
Measurement convention	alternative deflators, USD vs. local currency, gross vs. net flows
Inference too narrow	Driscoll–Kraay (cross-sectional dependence); country clustering; wild-cluster bootstrap (few countries)
Omitted confounders	Oster δ / coefficient-stability bounds

Inference under cross-country dependence (the common slip)

Standard errors are where IMFER robustness most often fails, because country panels violate the assumptions behind default clustering. Flows, spreads, and prices are cross-sectionally dependent (a global shock hits everyone), so country-clustered SEs alone understate uncertainty — use Driscoll–Kraay or a two-way (country and time) cluster. The country dimension is usually small (30–60 economies), so cluster-robust asymptotics are unreliable — report a wild-cluster bootstrap. Serial correlation within country is the norm — do not assume i.i.d. errors. State which of these you address and why; a referee who suspects the SEs are too tight will discount the whole result.

Robustness craft

1. Lock the primary specification first. Everything else perturbs it; do not present five co-equal panels and let the referee guess the preferred one.
2. Lead with the country-composition checks. Leave-one-country-out and drop-the-dominant-economy are the first things an IMFER referee runs in their head — show them in the main paper.
3. Stress the global common shock. Demonstrate the effect is not the global financial cycle masquerading as your policy variable.
4. Match inference to cross-country structure. Driscoll–Kraay or two-way clustering, and wild-cluster bootstrap when the country count is small — wrong SEs are the most common IMFER robustness failure.
5. Report stability, not just significance. Show the point estimate range across checks; a check that moves it is information — bound the implication for policy rather than hiding it.

Execution bridge (StatsPAI / Stata MCP)

Run the battery, don't just enumerate it. Full map: execution-with-mcp. IMFER is international macro/finance; cross-country panels with confounded policy — emphasize identification and clustering.

• 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.

Checklist

• Primary specification declared before perturbations
• Leave-one-country-out / drop-dominant-economy shown (composition not driving the result)
• Crisis-episode sensitivity (drop GFC / euro / COVID) reported
• Global common shock / global financial cycle controlled
• Sample-selection alternatives (balanced/unbalanced; AE/EM split; coverage rules) tested
• Measurement conventions (currency, deflator, gross/net) varied
• Inference correct for cross-sectional dependence and few-country count; SEs not asterisks
• Point-estimate range across checks reported; any check that moves it explained

Anti-patterns

• A 20-column robustness table with no map from check to threat ("kitchen-sink robustness")
• Never running leave-one-country-out when one economy obviously dominates the panel
• Ignoring the global financial cycle / US-shock common factor
• Reporting only that significance survives while the point estimate wanders
• Country clustering with very few countries and no wild-cluster bootstrap
• Hiding the window or sample rule that breaks the result

Worked vignette (illustrative)

A panel estimate of the spillover from US tightening to EM inflows is −0.8% of GDP (s.e. 0.2). The suite: (i) leave-one-country-out keeps it in [−0.9, −0.7] with no single economy decisive; (ii) dropping the GFC window leaves it at −0.7; (iii) adding the global financial cycle factor barely shifts it; (iv) the AE/EM split shows the effect concentrated in EMs as the mechanism predicts; (v) Driscoll–Kraay SEs (cross-country dependence) keep the CI away from zero; (vi) gross vs. net flows give the same sign and magnitude. The point estimate barely moves — the IMFER target — and the one check that softens it (excluding commodity exporters) is reported with its policy reading.

Referee pushback mapped to the robustness fix

• "This is driven by a few countries." → Show leave-one-country-out and drop-the-dominant-economy in the main paper; report the estimate range.
• "This is the GFC / euro crisis, not your variable." → Drop the crisis window and add the global-cycle control; show the result holds.
• "Did you cluster correctly?" → Use Driscoll–Kraay for cross-country dependence; with few countries report a wild-cluster bootstrap.
• "Different currency / flow convention would change this." → Re-run in USD and local currency, gross and net; show the same sign and magnitude.

The composition check IMFER referees run first

Because IMFER panels are small in the country dimension and dominated by a few large economies, the single most predictable referee move is "drop the obvious country and show me it survives." Build this into the paper, not the appendix: report a leave-one-country-out distribution (or a coefficient plot with each country excluded), and if one economy is decisive, say so and bound the implication. A panel where China, the US, or one euro-area crisis country silently drives the headline is the most common IMFER robustness failure — and the easiest to pre-empt.

Output format

【Journal】IMF Economic Review
【Skill】imfer-robustness
【Primary spec】declared? [Y/N] — estimate: ___ (s.e. ___)
【Composition】leave-one-country-out / drop-dominant range: [___, ___]
【Episode / window】drop-crisis result: ___
【Common shock】global-financial-cycle control result: ___
【Sample / measurement】AE-EM split, currency, gross/net: ___
【Inference】Driscoll–Kraay / wild-cluster (few countries): ___
【Estimate stability】range across checks: [___, ___]; checks that move it: ___
【Next skill】imfer-tables-figures