AI Economics Research & Advisory
Independent research partnerships for measuring, testing, and applying AI in economic and policy analysis.
This advisory practice is the collaboration layer of Too Early To Say, Victoria Cholette's independent applied-economics lab. For a policy question, research workflow, or disputed result, each study defines an economic target and an inspectable evidence trail. AI agents handle source assembly, first-pass implementation, and reruns. Victoria defines the estimand, identification strategy, checks, and interpretation.
The work is for policy teams, research organizations, economists, and funders that need to evaluate an AI-assisted workflow or produce a defensible policy analysis. Faster execution creates a verification burden when it produces more outputs than a team can check.
One idea governs each engagement. AI creates research value when the economic question, failure rule, and evidence trail remain visible from design through release.
Ways to work together
Measure AI research work
In Difference-in-Differences in Python, comparing the estimate with the planted 1.60 reveals a clean run whose estimand does not match the economic question.
Agents implement and rerun candidate workflows. Victoria defines the answer key, error taxonomy, and acceptance rule.
Audit a policy claim
The Food Desert Myth reports a 4.7-fold comparison that is now being reconciled across the article, data, code, and public metadata.
Agents trace the claim across source files and public surfaces. Victoria sets the source of record and the limitation included with the result.
Build a verification gate
Instrumental Variables in Python pairs an F-statistic of 2,051 with a causally invalid estimate when exclusion fails.
Agents execute the estimator and diagnostics. Victoria defines the data-generating process, causal claim, and release condition.
What a collaboration produces
- A frozen research question, estimand, and decision context.
- A documented division of labor between agent execution and economist judgment.
- A failure rule, verification plan, and evidence-status record.
- A public or internal package that preserves sources, outputs, limits, and version information.
Propose a collaboration
A useful first note names the policy question, available data, intended decision, and the part of the workflow that needs to be measured or tested.
Email: [email protected]