revacc-data-analysis

Data Analysis & Robustness (revacc-data-analysis)

When to trigger

• Data are built and it is time to estimate and report
• You are unsure how to cluster standard errors for your accounting panel
• Referees will probe endogeneity, construct measurement, or the channel
• You must document the Compustat/CRSP/I/B/E/S/audit-data provenance behind the sample
• An analytical paper needs a stylized empirical illustration of its comparative statics

Get the standard errors right (a RAST signature check)

Empirical-accounting referees scrutinize inference. Default to clustering by firm, and consider two-way clustering by firm and year (Petersen) when both cross-sectional and time-series dependence are present. With few clusters (e.g., a state- or country-level policy), use the wild-cluster bootstrap rather than asymptotic cluster-robust SEs. Match the clustering to the source of correlated shocks implied by your design, and report the choice explicitly — an unjustified SE choice is a fast credibility hit at a journal that often decides in one round.

Execute the identification, don't just assert it

• DiD: report pre-trends and use heterogeneity-robust estimators for staggered timing (Callaway–Sant'Anna / Sun–Abraham), not naive two-way FE.
• RD: report the optimal bandwidth, robust bias-corrected estimates, a manipulation (density) test, and covariate balance at the cutoff.
• IV/2SLS: report the first stage and instrument strength (e.g., F-statistic) and defend the exclusion in words.
• Event studies: report abnormal returns with a defensible benchmark and window, and confront confounding events — central for value-relevance and information-content claims.

Measure accounting constructs credibly

Use measures with precedent in prior RAST/JAR/JAE/TAR work (discretionary accruals, accruals quality, earnings persistence/smoothness, disclosure indices, comparability, audit-quality proxies, bid-ask spread / PIN for information asymmetry). Show the proxy behaves sensibly (validation, correlation with established measures) and test sensitivity to alternative proxies — proxy fragility is one of the most common RAST rejection reasons. For analyst/forecasting work, document I/B/E/S handling (actuals definition, stale-forecast screens, splits adjustments).

The robustness battery referees expect

• Alternative specifications: controls in/out, alternative fixed effects, alternative key-construct measures.
• Subsamples and falsification/placebo tests (effect absent where theory says it should be).
• Sensitivity of identification assumptions (alternative instruments, donut RD, bounds).
• Cross-sectional partitions that confirm the predicted channel (the conditional predictions from revacc-theory-development).
• Sample-construction, winsorization, and screen choices documented and varied.

Provenance is a deliverable, not a courtesy

RAST does not run JAE's mandatory archive or JAR's posted package as the headline, but referees and the editor still expect a credible, reconstructable sample. Keep top-to-bottom runnable scripts that regenerate every table from raw extracts; document screens, vintages, and access dates for each source; respect database terms of use. If the work entered through the RAST Conference path, keep the version history clean for the conference-issue timeline. Confirm current data/code expectations on the official page (待核实; 检索于 2026-06).

Checklist

• SE clustering matches the design (firm / firm-and-year; wild bootstrap if few clusters), stated explicitly
• Identification executed with diagnostics (pre-trends / bandwidth / first-stage / balance)
• Modern DiD estimators used for staggered treatment timing
• Key construct validated; results robust to alternative proxies
• Robustness, falsification/placebo, and channel partitions reported
• Winsorization/screens documented and varied; I/B/E/S handling documented where relevant
• Provenance trail (sources, vintages, screens) reconstructable; scripts runnable end-to-end

Anti-patterns

• White/robust SEs on panel data ignoring within-firm correlation.
• Naive two-way-FE DiD with staggered adoption.
• One proxy, no validation for a contested accounting construct.
• Significance fishing across windows, bandwidths, or specifications.
• Confounded event windows in value-relevance/information-content tests.
• Black-box sample: screens and vintages undocumented, results unreproducible.

Output format

【Estimator & SEs】model; clustering (firm / firm×year / wild bootstrap) + justification
【Identification executed】diagnostics reported (pre-trends/bandwidth/first-stage/balance)
【Construct measurement】proxy + validation + alt-proxy robustness
【Robustness/falsification】[...]
【Channel partitions】conditional predictions confirmed? [...]
【Provenance】sources/vintages/screens documented; scripts runnable
【Open issues for referees】[...]
【Next skill】revacc-contribution-framing