worlddev-identification

Identification & Inference Strategy (worlddev-identification)

When to trigger

• A causal claim rests on OLS + controls, or TWFE on staggered timing
• An RCT or quasi-experiment exists but the estimand and external relevance are unstated
• A qualitative paper makes causal-sounding claims with no explicit inferential logic
• A mixed-methods paper runs strands in parallel without saying how they jointly warrant the claim
• You are unsure whether the evidence clears WD's bar for the kind of claim being made

WD's pluralist standard

WD does not impose a single identification template. It imposes a single demand: the route from evidence to claim must be explicit, defended, and matched to the claim's strength. A paper that says "is associated with" needs descriptive credibility; a paper that says "causes" needs a design that earns it; a paper that says "this mechanism explains why" needs a logic — quantitative or interpretive — that traces the mechanism. The cardinal sin at WD is a claim that outruns its warrant, in either direction (over-claiming causality from correlation, or burying a genuine causal design under hedging). Pick the branch that fits your claim and make the warrant transparent.

Branch A — Quantitative causal design

• DID / event study: with staggered adoption, move past TWFE (Callaway–Sant'Anna, Sun–Abraham, de Chaisemartin–D'Haultfœuille); show clean pre-trends; report a Goodman-Bacon decomposition where relevant.
• IV: strong first stage; with weak instruments use Anderson–Rubin / weak-IV-robust sets; defend exclusion in institutions and theory, not just algebra — WD referees probe whether the instrument is plausible in that development context.
• RDD: density (McCrary / Cattaneo–Jansson–Ma) test; optimal bandwidth + robustness; covariate smoothness; bias-corrected CIs.
• RCT: explicit estimand; randomization balance; attrition (Lee bounds if differential); multiple-hypothesis adjustment; and — distinctively for WD — external relevance: why this site/sample teaches something beyond itself.
• Inference clustered at the assignment/treatment level; address few-cluster issues (wild-cluster bootstrap). Spatial correlation matters in development data — consider Conley SEs.

Branch B — Qualitative inference

Qualitative is not second-class at WD. The bar is transparency of inference, not sample size.

• Name the logic: process tracing, comparative case logic (most-similar / most-different), congruence testing, or interpretive thematic analysis — say which.
• Case selection justified by the logic, not by convenience; explain why these cases can bear the claim.
• Evidence trail: how data were generated (interviews, archives, observation), coding approach, and how rival explanations were tested and rejected (not merely mentioned).
• Trustworthiness: triangulation across sources, attention to negative cases, reflexivity about the researcher's position.

Branch C — Mixed-methods integration

The contribution of a mixed-methods paper is the integration, not the coexistence of two strands. State the design (sequential explanatory, sequential exploratory, convergent) and what each strand does that the other cannot: e.g., the quant estimates how much, the qual explains why and for whom. A paper where the qualitative section is decorative will be read as a quantitative paper with an unexamined appendix.

Referee pushback mapped to the fix

• "OLS with controls is not identification." → Find a design (DID/IV/RDD) or downgrade the claim to descriptive and own it.
• "Staggered TWFE is biased here." → Re-estimate with Callaway–Sant'Anna / Sun–Abraham; show flat leads.
• "Why these three cases?" → Justify selection by the inferential logic and show how rivals were ruled out.
• "The RCT is clean but local." → Argue external relevance: mechanism, scope conditions, transferable lessons.
• "The qual and quant don't talk to each other." → Make the integration explicit; show the joint inference.

Execution bridge (StatsPAI / Stata MCP)

Estimate and audit the design, don't only describe it. Full map: execution-with-mcp. World Development is multidisciplinary development studies; the chain serves its quantitative-causal lane, mixed-methods work uses its own standards.

• detect_design → recommend → fit with as_handle=true → audit_result.
• Observational causal claims: 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 + romano_wolf for many-outcome control.
• Sensitivity: oster_delta / sensemakr for observational claims.

Report the magnitude in interpretable units; route the full battery to the appendix. A run end-to-end (synthetic data, real returns) is in the JF execution walkthrough.

Checklist

• Branch chosen; the evidence-to-claim route stated in one sentence
• Quant: design-appropriate diagnostics (pre-trends / density / first-stage / balance); modern estimator where TWFE would bias; clustering correct
• Qual: inferential logic named; case selection justified; rivals tested, not just listed
• Mixed: integration design stated; each strand's distinct job is explicit
• External relevance argued for single-site/single-case work
• Inference reported with SEs / coverage (no significance asterisks); the claim never exceeds the warrant

Anti-patterns

• "Causes" from OLS + controls with no design
• Treating a qualitative paper as needing "more N" rather than a clearer inferential logic
• A mixed-methods paper that runs strands in parallel and never integrates them
• An RCT presented with no discussion of who is left out or whether it travels
• Significance asterisks instead of standard errors / confidence intervals

Output format

【Journal】World Development (WD)
【Skill】worlddev-identification
【Branch】quant-causal / qualitative / mixed
【Evidence→claim route】one sentence
【Warrant evidence】design diagnostics / inferential logic + rival tests / integration design
【External relevance】why it travels beyond the site/case
【What it does NOT warrant】the claim it cannot support
【Source status】verified URL / 待核实 / not asserted
【Next skill】worlddev-theory-model