thedixitjain Theory and Model Craft (wber-theory-model)
When to trigger
- A reduced-form result needs a model to interpret a mechanism or run a counterfactual
- A referee asks "what is the model that rationalizes this?" or "what does welfare do?"
- You want to extrapolate beyond the estimated policy to an un-tried policy (requires structure)
- The paper has a heavy model but it is decorative — it does not discipline the empirics or the counterfactual
- You are unsure whether WBER expects a formal model here at all
Does WBER even want a model here?
WBER publishes both theoretical and empirical development research, but most accepted papers are empirical, and a formal model is a means, not a merit badge. Add a model only when it earns its place:
- To interpret — convert a reduced-form coefficient into a structural parameter (an elasticity, a friction) policymakers can reason about.
- To extrapolate — answer a counterfactual the data alone cannot (an un-tried transfer size, a national roll-out, a price change).
- To aggregate — move from a partial-equilibrium treatment effect to a general-equilibrium or welfare statement.
If none of these apply, a clean reduced-form evaluation with a clearly stated conceptual framework is the better WBER paper. A decorative model that the empirics ignore is a liability — it invites referee attacks for no payoff.
Disciplining a development model
- Tie every parameter to data. Name what in the developing-country data identifies each parameter (a moment, an elasticity, an experimental treatment effect). "We calibrate to the literature" is weak; "the experimental LATE pins the take-up elasticity" is strong.
- Match the friction to the setting. Development models live or die on the right friction: credit/insurance constraints, missing markets, information asymmetries, enforcement/state-capacity limits, search frictions in informal labor markets. Use the one the data support, not a textbook default.
- Validate out of sample. Show the model reproduces a moment it was not fit to — ideally a treatment effect from the very experiment/reform that motivates the paper.
- Make the counterfactual honest. State which parameters you assume are policy-invariant (Lucas critique) and why; bound the GE channels you cannot fully model.
Reduced-form ↔ structure handoff
| You have... | Add a model only if... | Otherwise |
|---|---|---|
| A clean RCT/quasi-experimental effect | You need to extrapolate to an un-tried policy or aggregate to welfare | Report the effect + a conceptual framework; skip the model |
| A structural estimate | You can tie parameters to data and validate untargeted moments | Reconsider — calibration-in-disguise will be flagged |
| A policy counterfactual claim | The model is identified from credible variation | Do not run the counterfactual on calibrated guesses |
Referee pushback mapped to the model fix
- "What rationalizes this reduced-form pattern?" → Add the minimal model that generates it; do not over-build. Show the model's comparative static matches the sign and rough magnitude you estimate.
- "Your parameters are calibration in disguise." → Tie each parameter to a data moment and report which moment moves which parameter; validate on an untargeted moment.
- "The counterfactual assumes invariance you never defend." → State explicitly which behavioral parameters you treat as policy-invariant and argue why the Lucas critique does not bite here.
- "This is a rich-world model bolted onto a poor-country setting." → Replace the frictionless core with the binding development friction (credit, insurance, information, enforcement) the data reveal.
- "The model adds nothing the regressions don't." → Either give it a job (a counterfactual the data cannot answer) or cut it to a conceptual framework.
Checklist
- The model's job is named: interpret, extrapolate, or aggregate (not decoration)
- Each parameter is tied to identifying variation in the developing-country data
- The friction matches the setting (credit/insurance/information/enforcement/search)
- An untargeted moment or out-of-sample treatment effect validates the model
- Counterfactual states policy-invariance assumptions and bounds GE channels
- Model assumptions are kept separate from policy interpretation in the text
- If no model is warranted, a clear conceptual framework replaces it
Theory-to-policy translation
Whatever its form, the model or framework must end in something a development policymaker can use. A structural elasticity should be reported as "a 10% subsidy raises adoption by X%"; a welfare statement should net out fiscal cost; a counterfactual should name the un-tried policy and its predicted effect with a stated uncertainty range. WBER's value-add over a pure-methods outlet is exactly this last step — the model exists to make a development decision tractable, not to demonstrate technique. If you cannot translate the model's output into a policy magnitude, reconsider whether the model is doing real work.
Anti-patterns
- A decorative model the empirical section never uses or tests
- Calibrating to "the literature" and running a welfare counterfactual on unvalidated parameters
- A first-world frictionless model imposed on a setting defined by missing markets
- Running an un-tried-policy counterfactual without arguing policy-invariance
- Treating a model as a substitute for credible identification rather than a complement
- Reporting model output in abstract parameter units a policymaker cannot use
- Escalating to a full structural model when a conceptual framework would do the job
Worked vignette (illustrative)
A paper estimates a sharp RD effect of a fertilizer subsidy on yields but the policy question is "what subsidy level maximizes welfare net of fiscal cost?" — which the single threshold cannot answer. The WBER-appropriate move: write a small household model with a credit constraint, pin the adoption elasticity to the RD jump and the constraint to observed liquidity heterogeneity, validate by reproducing the (untargeted) take-up gradient across wealth, then trace welfare across subsidy levels. The model earns its place because it answers a counterfactual the RD cannot, and it is disciplined by the same variation that identifies the reduced-form effect.
When a conceptual framework beats a formal model
For many WBER empirical papers, the right "theory" is a tight conceptual framework, not a solved model: a clear statement of the agents, the binding constraint, and the predicted sign of the policy's effect. A framework earns its place when it (a) motivates the empirical specification, (b) makes the mechanism falsifiable, and (c) tells the reader what would not happen if the mechanism were absent. It avoids the trap of a formal model that the data ignore. Use a framework when you need to organize intuition and discipline interpretation; escalate to a formal model only when you must extrapolate or aggregate.
Output format
【Model's job】interpret / extrapolate / aggregate / NONE (framework only)
【Parameters ↔ data】each parameter tied to identifying variation
【Friction】credit / insurance / information / enforcement / search
【Validation】untargeted moment or out-of-sample treatment effect
【Counterfactual】policy-invariance assumptions + GE bounds
【Next step】wber-robustness
Source: brycewang-stanford/Awesome-Journal-Skills → World-Bank-Economic-Review-Skills/skills/wber-theory-model/SKILL.md