restat-identification

Identification & Measurement Strategy (restat-identification)

When to trigger

• A causal claim rests on OLS + controls, or TWFE on staggered timing
• An IV's exclusion restriction or first-stage strength is contestable
• An RD's continuity / manipulation assumptions are not yet defended
• A shift-share / exposure design's exogeneity (shares vs shocks) is unargued
• The outcome or key regressor is measured with error, or you built a new measure/index

The REStat identification-and-measurement bar

REStat is applied econometrics with a measurement tradition, so two things are judged together: the mapping from data to the causal object must be explicit and defended, and the quality of measurement behind every variable must be credible. A clean design on a badly measured construct does not clear the bar; neither does a beautifully measured variable in a hopelessly confounded regression. Report standard errors and modern inference; clustering at the assignment level; address attenuation and other measurement-error biases head-on — REStat referees raise measurement objections sibling journals sometimes wave through.

Branch paths

Branch A: Difference-in-differences / event study

• With staggered adoption, move beyond TWFE (Callaway–Sant'Anna, Sun–Abraham, de Chaisemartin–D'Haultfœuille — the last has REStat-published estimators).
• Show a clean event-study with leads (flat pre-trends) and report a Goodman–Bacon decomposition.
• State the parallel-trends assumption and probe it (pre-trend tests + Rambachan–Roth honest bounds where relevant).

Branch B: Regression discontinuity

• McCrary / Cattaneo–Jansson–Ma density test for manipulation; covariate smoothness at the cutoff.
• Optimal bandwidth + bias-corrected, robust CIs; sensitivity to bandwidth and polynomial order.
• Fuzzy RD: report first stage; defend exclusion of the running variable's other channels.

Branch C: Instrumental variables

• Strong first stage (report effective F / Montiel-Olea–Pflueger); with weak instruments use Anderson–Rubin / weak-IV-robust sets.
• Defend the exclusion restriction in theory, institutions, and falsification tests.
• Shift-share / Bartik: argue exogeneity of shares (Goldsmith-Pinkham–Sorkin–Swift) or of shocks (Borusyak–Hull–Jaravel); report the implied just-identified estimates.

Branch D: Measurement (REStat signature)

• Construct validity: what does the measure actually capture; validate against an external benchmark.
• Measurement error: classical vs non-classical; attenuation correction, validation samples, or bounds.
• New index / data: document construction, sensitivity to choices, and show the applied conclusion is not an artifact of how you measured.

Checklist

• Branch chosen; data-to-object mapping stated in one sentence
• DID: heterogeneity-robust estimator + flat event-study leads + Bacon decomposition
• RD: density test + smoothness + bias-corrected robust CIs + bandwidth sensitivity
• IV: first-stage strength + weak-IV-robust inference + defended exclusion
• Shift-share: exogeneity of shares or shocks argued explicitly
• Measurement: construct validity shown; measurement error addressed (correction / bounds)
• Inference: SEs reported, clustered at the right level; few-cluster issues handled (wild bootstrap)
• The claim never exceeds what identification AND measurement jointly support

Anti-patterns

• TWFE on staggered treatment with no heterogeneity-bias discussion
• An RD with no manipulation test or no bandwidth sensitivity
• A weak first stage reported with conventional t-stats as if robust
• Ignoring attenuation from a noisily measured regressor — a classic REStat referee catch
• A new index presented without validation against any external benchmark
• Conflating "statistically significant" with "credibly identified and well measured"

Worked vignette: a noisily measured regressor (illustrative)

A paper regresses earnings on a survey-reported measure of training hours and finds a small effect. A REStat referee notes the training measure is self-reported and likely error-ridden, biasing the coefficient toward zero. The fix: bring an administrative validation subsample, estimate the reliability ratio (say 0.6, illustrative), and show the attenuation-corrected effect is roughly 1/0.6 larger — turning a "small" effect into an economically meaningful one, with the correction's assumptions stated. Measurement, not just identification, moved the answer.

Output format

【Branch】DID / RD / IV / shift-share / measurement
【Data-to-object mapping】one sentence
【Identification evidence】[event-study+Bacon / density+smoothness / first-stage+AR / shares-or-shocks]
【Measurement evidence】[construct validity / error correction / bounds] — or "n/a, cleanly measured"
【Inference】SEs + clustering level; few-cluster fix if any
【What it does NOT identify】[...]
【Next step】restat-theory-model (or restat-robustness if theory is minimal)