worlddev-robustness

Robustness & Trustworthiness (worlddev-robustness)

When to trigger

• The headline result moves under plausible alternative specifications
• A referee suspects the finding is driven by one region, one wave, or one measurement choice
• Development data are messy (recall error, measurement in informal economies, attrition) and this is unaddressed
• A qualitative finding rests on a few vivid quotes with no account of disconfirming evidence
• The robustness section is a mechanical dump of appendix tables organized by table, not by threat

Organize by threat, not by table

The single biggest WD robustness failure is a wall of appendix tables with no logic. A WD referee — often from a different discipline than the author — wants to see that you identified the threats to your specific claim and addressed each one. Structure the robustness work as a short list of named threats, each with the check that retires it and a one-line verdict. For each threat: what would break the claim, what test isolates it, what the test shows.

Quantitative threat map

| Threat | Check |

|--------|-------|

| Specification dependence | Add/drop controls in a disciplined sequence (Oster-style δ/bounds); specification curve if the literature is unsettled |

| Sample / outlier dependence | Drop influential units, regions, or waves; leave-one-out; trim |

| Measurement error (acute in development data) | Alternative measures; validation against an independent source; bounds |

| Inference fragility | Cluster at the right level; few-cluster wild bootstrap; spatial (Conley) SEs; randomization inference for RCTs |

| Selection / attrition | Lee bounds; selection models; characterize who exits |

| Multiple hypotheses | Romano–Wolf or sharpened q-values across the family of outcomes |

| Mechanism vs. confound | Show the proposed mechanism's footprint; rule out the leading alternative explicitly |

Run the checks the threat justifies — not the full menu. A paper that reports forty robustness tables but never addresses the obvious confound has gold-plated the wrong corner.

Qualitative trustworthiness map

Robustness for qualitative WD work is trustworthiness, and it is judged, not waived:

• Triangulation: corroborate key claims across data sources or informant types.
• Negative-case analysis: actively present and account for evidence that cuts against the argument — its absence is a red flag.
• Member checking / saturation: where appropriate, evidence that interpretations were checked and categories stabilized.
• Audit trail: enough on coding and analysis that another researcher could follow the inference.
• Reflexivity: acknowledge how the researcher's position shaped access and interpretation.

Mixed-methods

Show the strands converge or that divergence is informative. When quant and qual disagree, that tension is data — explain it rather than hiding the weaker strand.

Development-specific traps WD referees catch

• Treating survey measures from informal/subsistence settings as if measured with the precision of administrative data
• Ignoring spatial autocorrelation in geographically clustered development data
• Pooling heterogeneous countries/regions and reporting one average that masks the policy-relevant variation
• Generalizing from one program/site without scope conditions
• Reporting the robust result but not the fragile one a skeptic would run

Execution bridge (StatsPAI / Stata MCP)

Run the battery, don't just enumerate it. Full map:

[execution-with-mcp](../../../shared-resources/empirical-methods/execution-with-mcp.md). World Development is multidisciplinary development studies; the chain serves its quantitative-causal lane, mixed-methods work uses its own standards.

• Many outcomes / specifications: romano_wolf (step-down FWER) or benjamini_hochberg.
• OVB sensitivity: oster_delta / sensemakr.
• Inference: wild_cluster_bootstrap (few clusters), twoway_cluster / conley.
• Re-fit off one handle: audit_result(result_id) lists missing checks + the exact

suggest_function for each.

• Exhibits: etable / did_summary_to_latex from the handle — no retyped numbers.

Decisive checks in the body, exhaustive battery in the appendix.

[JF execution walkthrough](../../../Journal-of-Finance-Skills/resources/worked-examples/02-execution-walkthrough.md).

Checklist

• [ ] Robustness organized by named threat, each with check + one-line verdict
• [ ] Inference matched to the design (clustering level, few-cluster, spatial, randomization)
• [ ] Measurement error addressed where development data warrant it
• [ ] The leading alternative explanation is ruled out, not merely mentioned
• [ ] Qual: triangulation + negative cases + audit trail present
• [ ] Heterogeneity that matters for policy is shown, not averaged away
• [ ] No significance asterisks; effect sizes and uncertainty reported in real units

Anti-patterns

• A robustness appendix sorted by table number with no threat logic
• Forty checks for a non-threat, zero for the obvious confound
• Burying a fragile headline result and reporting only the survivor specifications
• Qualitative work that quotes only confirming voices and never the disconfirming ones
• Hiding quant/qual divergence in a mixed paper instead of explaining it

Output format

【Journal】World Development (WD)
【Skill】worlddev-robustness
【Verdict】robust / fragile / mixed
【Threats addressed】[threat → check → verdict] for each
【Leading alternative】how it is ruled out
【Qual trustworthiness】triangulation / negative cases / audit trail (if applicable)
【Policy-relevant heterogeneity】shown / hidden
【Source status】verified URL / 待核实 / not asserted
【Next skill】worlddev-tables-figures