Skills Data Science Advanced Causal Inference for Empirical Studies

Advanced Causal Inference for Empirical Studies

v20260724
smj-data-analysis
A comprehensive guide for addressing endogeneity, selection bias, and reverse causality in empirical research, specifically tailored for top-tier journals. It details advanced econometric methods (DID, IV, PSM, Heckman) and outlines best practices for conducting rigorous robustness checks, ensuring that identified effects are genuinely causal, not merely correlated.
Get Skill
226 downloads
Overview

Data Analysis & Endogeneity (smj-data-analysis)

When to trigger

  • You have a performance regression with no endogeneity / reverse-causality treatment
  • DID, IV, matching, or a selection model is chosen but not yet stress-tested
  • Reviewers will ask "how do you know this is causal and not selection?"
  • You need to plan the mechanism test and the robustness battery

The SMJ endogeneity mandate

Performance regressions with unaddressed endogeneity or reverse causality are the #1 SMJ rejection reason. Treat causal identification as a first-class part of the paper, not a footnote. The reviewer's mental model: firms that make this strategic choice are different in ways that also affect performance. You must close that door explicitly.

SMJ codifies this in Bettis, Gambardella, Helfat & Mitchell (2014), "Quantitative empirical analysis in strategic management," SMJ 35(7): 949–953: acknowledge endogeneity, make a good-faith effort to address it, and avoid data snooping / p-hacking. Report economic magnitudes, not just stars. SMJ will publish well-designed studies that report null results — so do not suppress a theory-relevant null.

Threat → tool map

Threat Primary tools
Self-selection into the strategic choice IV, Heckman selection, PSM/CEM + DID, Rosenbaum bounds
Reverse causality / simultaneity Exogenous shock + DID, lagged + Granger-style tests, IV
Unobserved time-invariant heterogeneity Firm fixed effects (caveat: cannot fix time-varying confounds)
Omitted environmental confound Industry-year FE, region FE, controls, falsification tests
Measurement error in X IV, multiple indicators, sensitivity analysis

Pick from the threat named in smj-methods; usually you will combine FE with one identification tool.

Design-specific execution

DID / natural experiment

  • Test and show parallel pre-trends (event-study plot, not just a claim).
  • If treatment timing is staggered, address heterogeneous-treatment-effect bias (Goodman-Bacon decomposition; Callaway–Sant'Anna or Sun–Abraham estimators) rather than naive two-way FE.
  • Run placebo tests (fake treatment dates; unaffected units) and report effect dynamics.

Instrumental variables

  • Report the first-stage F (weak-instrument concern below conventional thresholds → use weak-IV-robust inference).
  • Make the exclusion argument in prose: why the instrument affects performance only through the strategic choice. Reviewers reject IVs whose exclusion is implausible.
  • Report the reduced form and over-identification tests where applicable.

Matching (PSM / CEM) + DID

  • Report covariate balance before/after; show common support.
  • Matching handles selection on observables only; combine with DID and acknowledge residual selection on unobservables (bounds).

Heckman selection

  • Justify the exclusion restriction in the selection equation (a variable affecting selection but not the outcome). A Heckman with no valid exclusion restriction is identified only off functional form — reviewers know this.

Mechanism & robustness

  • Mechanism test: if you theorized a mediator, test it (prefer evidence beyond a Baron–Kenny mediation regression — e.g., moderation-of-process, subsample variation in the mechanism).
  • Magnitude, not just stars: interpret the key effect in economic terms (e.g., "a 1 SD increase in X → a Y% change in performance"); SMJ cares about meaningful effects.
  • Robustness battery (report, do not bury): alternative DVs; alternative samples (drop dominant industries/years); alternative estimators; clustering choices; prior performance; survivorship. Where feasible, show sensitivity to the identifying assumption (partial-identification / bounding).
  • Inference: cluster standard errors at the level of treatment assignment (often firm); justify the choice.

Execution bridge (StatsPAI / Stata MCP)

Run the battery, don't just enumerate it. Full map: execution-with-mcp. SMJ is strategy — firm-level panels where strategic choices are endogenous; foreground IV / DiD identification and the endogeneity-of-strategy objection.

  • 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

  • The identifying threat is stated and the matching tool is deployed
  • Reverse causality is addressed by design, not by lags alone
  • Firm (and industry-year) fixed effects included where appropriate
  • DID: parallel-trends evidence shown; staggered-timing bias addressed; placebos run
  • IV: first-stage strength reported; exclusion argued in prose; reduced form shown
  • Matching: balance + common support reported; unobservable selection acknowledged
  • Heckman: valid exclusion restriction, not functional-form identification
  • Mechanism tested, not just asserted
  • Robustness across DV, sample, estimator, and clustering reported
  • Standard errors clustered at the assignment level

Anti-patterns

  • Cross-sectional correlations interpreted causally — an instant credibility loss at SMJ
  • "We include fixed effects" treated as a complete endogeneity defense (FE miss time-varying confounds)
  • IV with an exclusion restriction no reviewer would believe
  • Heckman or PSM run mechanically with no defensible exclusion / balance
  • Ignoring that firms self-select into the very strategic choice being studied
  • Staggered DID with naive two-way FE and no heterogeneity correction
  • A wall of robustness tables that never confronts the central threat
  • Specification hunting until p < 0.05 (data snooping); reporting stars with no economic magnitude; suppressing a theory-relevant null — all discouraged by SMJ

Output format

【Identifying threat】selection | reverse causality | unobserved heterogeneity | omitted confound
【Estimator】FE + [IV | DID | matching | Heckman | ...]
【Identification evidence】[parallel trends / first-stage F / balance / exclusion argument]
【Placebo / falsification】[done?]
【Mechanism test】[what + result]
【Robustness】[DV alt, sample alt, estimator alt, clustering]
【Residual threat acknowledged】...
【Economic magnitude reported】yes / add
【Nulls reported honestly (no p-hacking)】yes
【Next step】smj-contribution-framing

Templates & resources

Info
Category Data Science
Name smj-data-analysis
Version v20260724
Size 7.8KB
Updated At 2026-07-29
Language