Jfm Robustness

Robustness Strategy (jfm-robustness)

When to trigger

• The headline holds with the chosen liquidity measure but you have not shown it survives alternatives
• Results may flip with a different sample period, asset universe, or filter rule
• Inference uses plain OLS standard errors on data that are autocorrelated and cross-correlated (panel of stocks over time)
• Microstructure noise (bid-ask bounce, stale quotes, discreteness) could be generating the effect
• A referee will ask "is this the mechanism or the measurement?" and you need each check mapped to a specific threat

The JFM robustness ledger

JFM referees do not want a wall of robustness tables; they want each check tied to a named threat to the microstructure interpretation. Build the ledger threat-first.

Sequencing the checks

1. Measurement robustness first — if the result is an artifact of one liquidity construct, nothing else matters.
2. Inference next — panel microstructure data are doubly correlated; default OLS SEs overstate precision. State the clustering and why.
3. Design robustness — for a market-structure event, placebo dates, alternative control groups, donor sensitivity; honest pre-trends.
4. Mechanism robustness last — show the effect strengthens where the microstructure mechanism predicts (e.g., larger in high-adverse-selection names) — this converts robustness into corroboration.

Keep a one-line rationale per check ("addresses the concern that …"). Park the bulk in the Internet Appendix; keep the load-bearing ones in the main text (see jfm-tables-figures).

The microstructure-noise battery

Bid-ask bounce, price discreteness, and stale quotes can manufacture spurious patterns, so a dedicated noise battery is often expected. Standard moves: show the result is not an artifact of the bid-ask bounce (e.g., using mid-quote rather than transaction prices, or signed measures); test sensitivity to the sampling frequency (5-min vs. 1-min vs. tick) since noise dominates at the finest frequencies; for realized-volatility-based measures, apply a noise-robust estimator; and confirm the effect is not driven by the minimum-tick discreteness alone. Naming this battery explicitly signals to the referee that you know microstructure noise is the field's characteristic confounder.

Worked ledger (illustrative)

Headline: a market-structure change narrows effective spreads by 12 bps. The threat-mapped ledger reads: (1) measure — repeat with quoted and realized spreads and with Amihud; effect 9-14 bps across measures; (2) sample — split by market cap and by sub-period; significant in both halves; (3) noise — show it is not driven by tighter discreteness alone by controlling for the binding-tick fraction; (4) inference — two-way cluster by stock and day, t falls from 6.1 to 3.4 but stays significant; (5) confound — partial out contemporaneous volatility and volume, effect 10 bps; (6) mechanism corroboration — effect is twice as large in high-adverse-selection (high-PIN) names, exactly where theory predicts. Each line names the threat it kills; the corroboration line turns a defensive section into evidence for the mechanism.

Calibrating how much is enough

JFM does not reward a 20-table robustness appendix; it rewards the right checks. The decision rule: include a check if a competent microstructure referee would otherwise doubt the interpretation. Measurement and inference are nearly always load-bearing (keep in main text). Sub-period splits and influential-name drops are usually appendix material. A check that does not map to a named threat should be cut, not kept "to be safe" — orphan tables signal the authors are unsure which concern is real.