Jimf Robustness

Robustness Strategy (jimf-robustness)

When to trigger

• The main coefficient is in hand but a referee will ask whether it survives sample, period, and measurement choices
• The result might be driven by one crisis episode, one dominant country (the US), or one regime
• Inference is OLS-clustered-one-way and the panel has serial and cross-sectional correlation
• You are assembling a robustness section and want it to answer threats, not pad column count

The JIMF robustness logic: every check answers a named threat

A JIMF robustness section is persuasive when each exhibit is tied to an international-finance threat to inference, not when it is a wall of specifications. Build the section as a threat → check map. The threats that JIMF referees raise most are: a single global episode driving everything; US-centrism; regime dependence; a fragile measurement choice; and inference that ignores cross-country dependence.

How to sequence and present it

1. Lead with the threat the referee will raise first — usually "this is just the GFC" or "this is just the dollar." Answer it in the main text, not the appendix.
2. Show point-estimate stability, not just significance. A coefficient that wanders from 0.6 to 0.1 across measures is fragile even if each stays "significant." Plot the estimate across specifications (a specification curve or coefficient-stability figure reads well at JIMF).
3. Use time fixed effects as both a control and a diagnostic. Adding time FE that absorb all common global shocks and seeing the within-time (cross-country) coefficient survive is the cleanest answer to "it's the global cycle."
4. Match inference to the panel. With ~20–60 countries and serially correlated errors, one-way clustering understates SEs; report Driscoll–Kraay or two-way clustering, and wild-cluster bootstrap when the cluster count is small.
5. Demote the rest to the online appendix with a map (see jimf-internet-appendix).

Execution bridge (StatsPAI / Stata MCP)

Run the battery, don't just enumerate it. Full map: execution-with-mcp. JIMF is international macro-finance; cross-country panels + asset pricing — identification plus factor/Newey-West inference.

• 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

• Each robustness exhibit names the threat it answers
• "It's just the GFC/COVID" answered by dropping the episode and by rolling samples
• "It's just the dollar/Fed" answered by dropping the US, using NEER, or a non-US center
• Regime/openness dependence tested (IRR classification; Chinn–Ito / AREAER openness)
• Headline measure swapped for the leading alternative; point estimate stable, not just significant
• Inference accounts for cross-country and serial dependence (Driscoll–Kraay / two-way / wild bootstrap)
• Influential single country ruled out (jackknife by country)
• The two reflex checks (drop-US, drop-episode) are in the main text, not buried in the appendix
• Time fixed effects shown to leave within-time identification intact (or absence justified)

Anti-patterns

• A robustness section that is 15 specifications with no statement of which threat each addresses
• Reporting only that significance survives while the point estimate halves across measures
• Never dropping the US or the GFC episode — the two checks every JIMF referee wants
• One-way clustering on a 30-country panel with obvious serial and cross-sectional correlation
• Treating "we added more controls" as robustness when the threat is a global confounder (which time FE, not controls, address)
• Burying the most important robustness (regime split, drop-US) in the appendix where the editor won't see it