Revacc methods

Research Design & Identification (revacc-methods)

When to trigger

• Your treatment (a disclosure, a standard adoption, an audit/tax regime) may be endogenous
• Adoption is staggered across firms/years and you need a defensible DiD
• You have an association and a referee will ask "is this causal or just correlation?"
• You are designing an experiment to isolate a channel archival data cannot separate
• You are building an analytical model and must fix primitives, timing, and the equilibrium concept

Identification for archival accounting at RAST

Accounting treatments — disclosure choices, conservatism, auditor selection, tax positions, standard adoption — are rarely random, so RAST referees expect a credible identification strategy, not kitchen-sink controls. Pick the design that breaks the endogeneity for your accounting setting.

Identification threat / setting	Design
Regulation / standard adoption with a clean date (e.g., a reporting mandate)	Difference-in-differences; event study around the adoption date
Staggered adoption across firms/states/countries	Staggered DiD with modern estimators (avoid naive TWFE bias)
Endogenous accounting/auditor/tax choice	Instrumental variables / 2SLS with a defensible exclusion
A threshold rule (covenant, index inclusion, size or regulatory cutoff)	Regression discontinuity
Selection on observables	Matching (PSM/entropy) as a complement, not the main claim
A plausibly exogenous shock to the information environment	Natural experiment; pre-trends shown
Information content of an accounting signal	Short-window event study with a clean benchmark and confound check

State the estimating equation, the unit and level, the fixed effects (firm, year, industry-year), and the identifying variation explicitly. The design section must make a skeptic believe the variation is as-good-as-random conditional on controls. RAST's first-round-decision culture means a weak design is more likely to draw a reject than a "fix it in revision."

Measurement design is part of identification

For contested accounting constructs (discretionary accruals, earnings quality, disclosure indices, audit quality, information asymmetry), the proxy choice is a design decision. Pre-commit a primary measure with precedent and plan the alternative proxies you will use to show the result is not proxy-driven (carried out in revacc-data-analysis). A clean design on a fragile proxy still fails.

If the lane is analytical

• Fix the information structure, players, timing, and payoffs before solving; state the equilibrium concept (PBE, sequential, etc.).
• Show the model is the minimal structure that generates the accounting result; defend each assumption as load-bearing.
• Plan the comparative statics that become testable or normative accounting implications.

If the lane is experimental

• Manipulate the focal accounting construct with realistic stimuli and a fit-for-purpose pool (investors, auditors, managers); IRB documentation is expected.
• Pre-register where feasible; include manipulation and attention checks; power the design for the interaction, not just the main effect.

Design hygiene

• Show parallel pre-trends for any DiD and report dynamic (event-time) effects.
• Defend the exclusion restriction for any IV in words — relevance alone is not enough.
• Pre-commit the main specification; relegate alternatives to robustness.
• Plan the data provenance trail now (Compustat/CRSP/I/B/E/S/audit-data vintages and screens).

Execution bridge (StatsPAI / Stata MCP)

For the empirical / causal lane, estimate and audit rather than only specify. Full map: execution-with-mcp. RAST is empirical accounting; emphasize identification of disclosure / governance effects and the multiple-testing haircut for mined associations.

• detect_design → recommend → fit with as_handle=true → audit_result to enumerate the checks the design owes.
• Panel / staggered DiD: callaway_santanna / sun_abraham + bacon_decomposition
  • honest_did_from_result. IV: effective_f_test + anderson_rubin_ci. RDD: rdrobust + mccrary_test.
• Experiments: randomization-based inference and romano_wolf for the many-outcome family-wise correction reviewers expect.

Match the toolchain to the reviewer pool, and report the effect size the venue wants. A run end-to-end (synthetic data, real returns) is in the JF execution walkthrough.

Checklist

• The identifying variation (shock/setting/threshold) is named and defended
• Estimating equation, unit, level, fixed effects, and clustering plan are stated
• DiD shows pre-trends and dynamic effects; staggered designs use a modern estimator
• IV exclusion restriction is argued, not asserted; matching is a complement, not the claim
• Primary construct proxy pre-committed; alternative proxies planned
• Analytical models fix primitives/timing and the solution concept before solving
• Experiments have IRB, realistic stimuli, manipulation/attention checks, adequate power

Anti-patterns

• Kitchen-sink controls standing in for identification ("we control for everything").
• TWFE on staggered adoption without addressing heterogeneous-treatment-effect bias.
• IV by convenience: an instrument correlated with the outcome directly.
• Matching as causal proof when selection is on unobservables.
• Proxy fragility: a single contested construct measure with no alternative planned.
• Non-minimal model: primitives a referee can strip without losing the result.

Output format

【Lane】archival / analytical / experimental
【Setting & identifying variation】...
【Design】DiD / staggered-DiD / IV / RDD / event study / experiment / model
【Spec】equation; unit/level; fixed effects; clustering plan
【Construct】primary proxy + alternatives planned
【Identification defense】pre-trends / exclusion / discontinuity / randomization ...
【Data provenance】sources + vintages + screens noted
【Next skill】revacc-data-analysis