qje-robustness

Robustness & Online Appendix (qje-robustness)

When to trigger

• The main result exists but the robustness section / online appendix is thin
• You want to pre-empt the "what about X?" referee report
• You are unsure which checks are load-bearing vs. padding
• The appendix is a pile of tables with no organizing logic

QJE expectation: an exhaustive, organized appendix

QJE imposes no hard page limit and the modern QJE empirical paper carries a very extensive online appendix — the Chetty–Hendren neighborhoods papers (QJE 2018) and Chetty–Hendren–Kline–Saez (QJE 2014) are the genre standard, with online appendices dwarfing the printed text. A short robustness section signals an unfinished paper. But volume is not the goal — anticipation is. The appendix should answer, in advance, every serious threat a smart referee would raise, organized so the handling Editor can find each answer fast. Note: at submission everything ships in one PDF (no separate appendix file at the initial stage), so the appendix must be cleanly sectioned within that document. Lead with the checks that defend the identifying assumption, then measurement, then specification.

Robustness priority ladder (defend in this order)

1. Identification threats first. The checks that defend exogeneity: pre-trends, placebo timing/outcomes, falsification where the channel is absent, alternative control groups, donor-pool / synthetic checks. These are not optional.
2. Selection & sample. Attrition, sample-definition sensitivity, alternative inclusion windows, trimming/winsorizing, outlier robustness.
3. Measurement. Alternative outcome/treatment definitions, alternative data sources, measurement-error bounds.
4. Specification. Functional form, fixed-effects structure, controls in/out (and a discussion that the estimate is not control-sensitive), clustering alternatives.
5. Inference. Wild-cluster bootstrap with few clusters, randomization inference, multiple-hypothesis corrections across outcomes/subgroups.
6. Magnitude & external validity. Benchmark the effect size against the literature; show where the result does and does not extend.

What goes in the body vs. the appendix

Execution bridge (StatsPAI / Stata MCP)

Run the robustness battery, don't just enumerate it. Full map: shared-resources/empirical-methods/execution-with-mcp.md. QJE instantiation:

• Many outcomes / specifications: romano_wolf (step-down, FWER, accounts for cross-test correlation) or benjamini_hochberg — report the adjusted threshold, not a grid of naive stars.
• OVB sensitivity: oster_delta / sensemakr — quantify the confounder strength that would overturn the headline.
• Inference: wild_cluster_bootstrap with few clusters; twoway_cluster / conley where the dependence structure demands it.
• Re-fit off one handle: audit_result(result_id) lists the missing checks and the exact suggest_function to run for each — no guessing the battery.
• Emit appendix-ready exhibits with etable / did_summary_to_latex.

Keep the decisive checks in the body and the exhaustive (now actually-run) battery in the appendix; preserve the scripts for qje-replication-package.

Checklist

• Every identification threat has a corresponding check, named and reported
• Pre-trends / placebo / falsification evidence is in the body, not buried
• Estimate shown to be insensitive to reasonable specification choices
• Few-cluster / multiple-hypothesis inference handled
• Appendix is organized by threat category, not dumped chronologically
• Appendix is cleanly sectioned within the single submission PDF
• Each appendix result is referenced and one-line-summarized in the main text
• Magnitudes benchmarked against prior estimates in the literature

Anti-patterns

• A two-paragraph robustness section in a paper claiming a causal effect (QJE's no-page-limit norm leaves no excuse)
• Appendix tables with no narrative linking them to specific threats
• Reporting only the checks that pass; hiding the fragile specification
• "Results are robust to a battery of checks (see Appendix)" with no specifics in text
• Padding the appendix with redundant specifications instead of addressing real threats

Output format

【Identification checks】[pre-trends, placebo, falsification, alt controls, ...]
【Selection/sample checks】[...]
【Measurement checks】[...]
【Specification checks】[...]
【Inference checks】[wild-cluster, RI, MHT, ...]
【Body vs. appendix split】what stays in text
【Gaps to close】[...]
【Next step】qje-tables-figures