tar-data-analysis

Estimation, Robustness & Data Authenticity (tar-data-analysis)

When to trigger

• The sample is built and it is time to estimate and report
• You are unsure whether your standard errors, fixed effects, or estimator match the design
• Reviewers will probe robustness, alternative measures, or sample-selection screens
• You must assemble the data-authenticity / code-access package TAR requires
• A reviewer says "the result is not robust" or "I cannot tell how the sample was built"

Estimator and inference (large-sample archival core)

• Match the estimator to the design set in tar-methods: OLS with high-dimensional fixed effects (firm, year, industry-year) for panel associations; DiD / staggered-DiD with a modern estimator for adoption shocks; 2SLS for endogenous regressors; RDD for threshold settings; logit/ probit/Poisson/Tobit for limited or count outcomes (e.g., restatement, going-concern, fraud).
• Cluster standard errors at the level of treatment assignment / correlation (firm, or two-way firm-and-year); for few clusters use the wild-cluster bootstrap.
• Report fixed effects explicitly and show how the coefficient moves as you add them — a result that survives tighter fixed effects is more credible than one that does not.
• For accounting-specific measures (discretionary accruals, real earnings management, abnormal audit fees, effective/cash tax rates, disclosure tone), state the construction model and screen, and show the result is not an artifact of the proxy.

Robustness reviewers expect

• Alternative measures of the focal accounting construct (e.g., a second accruals model; cash vs. GAAP ETR; alternative disclosure proxy).
• Alternative samples and screens (drop financials/utilities; winsorize vs. truncate; subperiods).
• Sensitivity of the identifying assumption (pre-trends, placebo dates, alternative instruments, bandwidth choices for RDD).
• Falsification / placebo tests where the effect should be absent.
• Economic magnitude, not just significance — interpret the coefficient in accounting terms.

Data-authenticity & code access (a TAR-specific requirement)

TAR requires authors to enable confirmation of data authenticity, with differentiated rules:

• Publicly available databases (Compustat, CRSP, I/B/E/S, Audit Analytics): provide a precise description of the data and access to the computer code used to process it.
• Data abstracted from public sources (hand-collected from filings, PCAOB reports): provide the abstraction methodology plus code access.
• Privately collected data (proprietary field data, experiments): provide enough detail for reader confidence; corroborating third parties are acceptable.

Code/data sufficiency is part of the submission and acceptance requirements (待核实 whether a named public repository deposit is mandated at acceptance). Build clean, commented scripts from raw extract to every table now — not after the R&R.

Execution bridge (StatsPAI / Stata MCP)

Run the battery, don't just enumerate it. Full map: execution-with-mcp. TAR is archival accounting — DiD around regulation / standard changes, IV, and earnings-based designs; the corporate-causal chain fits directly.

• Many outcomes / specifications: romano_wolf (step-down FWER) or benjamini_hochberg — report the adjusted threshold.
• OVB sensitivity: oster_delta / sensemakr.
• Inference: wild_cluster_bootstrap (few clusters), twoway_cluster / conley; multilevel data → cluster at the right level.
• Re-fit off one handle: audit_result(result_id) lists the missing checks and the exact suggest_function for each.
• Exhibits: etable / did_summary_to_latex from the handle — no retyped numbers.

Keep the decisive checks in the body and the exhaustive battery in the appendix. See the executed chain in the JF execution walkthrough.

Checklist

• Estimator matches the design; FEs reported; SEs clustered at the right level
• Coefficient stability shown across FE/control sets
• Focal accounting measure validated with an alternative construction
• Pre-trends/placebo/falsification tests reported where the design needs them
• Economic magnitude interpreted, not just p-values
• Data-authenticity package assembled per data type (description + processing code / methodology)
• Sample-construction screens documented and reproducible from raw data

Anti-patterns

• Uncluttered p-values, no magnitude — significance without economic interpretation.
• Single proxy for a contested construct (e.g., one accruals model) presented as definitive.
• TWFE on staggered adoption ignoring heterogeneous-treatment-effect bias.
• Robustness theater: many tables that never vary the thing a skeptic doubts.
• Unreproducible sample: screens and merges that no one can rebuild from the raw data.
• No processing code ready, in violation of the data-authenticity policy.

Output format

【Estimator】OLS-HDFE / staggered-DiD / 2SLS / RDD / logit-Poisson ...
【Fixed effects & clustering】... ; SE level ...
【Focal measure】construction + alternative proxy: pass/issues
【Robustness】alt measures / samples / placebo / pre-trends ...
【Economic magnitude】coefficient means ... in accounting terms
【Data authenticity】public-db / abstracted / private — code & description ready? yes/no
【Open issues for reviewers】...
【Next step】tar-contribution-framing