AI-assisted econometrics

Econometrics is where AI's false confidence is most dangerous. A large language model will explain heteroskedasticity-robust standard errors flawlessly, then in the very next sentence tell you your cross-sectional OLS 'shows that schooling causes higher wages' - and say it with exactly the same calm authority. The deal for this module is simple: AI is an outstanding tutor and a fast pair of hands for drafting and debugging, but the identification strategy and the causal claim are yours to defend - at the seminar, in the referee report, and in front of the National Treasury. Use it to understand and to build. Never use it to decide what your coefficients mean.

The tutor that never sleeps

The single highest-value use of AI for an applied economist is on-demand tutoring. Fixed effects, instrumental variables, difference-in-differences, clustered versus robust standard errors, the Frisch-Waugh-Lovell theorem - you can ask for the intuition, a worked example, and the assumption each method rests on, then keep asking until it clicks. Treat it Socratically: ask 'explain it again as if I have never seen matrix algebra', then 'now show me the exact case where this method fails'. The failure case is the part a textbook buries and the part you actually need before you stake a policy brief on it.

You are a patient graduate econometrics tutor. Explain entity fixed effects to me in three passes:
1. The one-sentence intuition.
2. A worked example using Kenyan county panel data (47 counties, 2013-2023) predicting health spending, where county fixed effects absorb time-invariant county characteristics.
3. The single most important source of confounding that fixed effects CANNOT remove - be specific about what survives the within transformation.
Finish by stating the exact assumption I am relying on when I claim the within-county estimate is causal.

Reading a regression table in plain English

Paste a regression output and ask for a plain-language reading - AI is genuinely good at this, and it is faster than re-deriving what each column means. But bolt on a guardrail, because left to itself it will narrate your table as a causal story. Tell it explicitly to describe associations rather than effects; to state the identifying assumptions each coefficient would need before it could carry a causal reading; and to flag anything that looks read off the table wrongly. You are commissioning a translation, not a verdict.

Below is Stata output from a regression of log monthly wages on years of schooling and controls, estimated on KNBS labour-force cross-section data.

[paste the table]

Explain each coefficient in plain English for a policy audience. Rules:
- Describe these as ASSOCIATIONS, not causal effects. Do not use the words 'causes', 'leads to' or 'raises' about any coefficient.
- For the schooling coefficient, list the identifying assumptions I would need before a causal reading is defensible, and why each is doubtful in a cross-section.
- State what the standard errors are clustered on, if anything, and whether that is appropriate here.
- Flag any coefficient whose sign or magnitude looks suspicious, and say why.

A sounding board for identification: it suggests, you decide

When you are choosing an identification strategy, AI is a useful sounding board. Describe your question, your data and the variation you can exploit, and it will float candidates - panel fixed effects, an instrument, difference-in-differences around a policy change, a regression discontinuity at an eligibility cut-off. That is where its authority ends. Whether an instrument satisfies the exclusion restriction, whether parallel trends is credible for your NHIF roll-out, whether SUTVA survives spillovers between neighbouring counties - those are judgements about the world that only you can defend. The productive move is to make the model argue against itself: it proposes a strategy, then you order it to attack that strategy as a hostile referee.

Act as a hostile seminar discussant and referee. Do not be encouraging; your job is to find every reason this design fails.

My question: does access to mobile money (M-PESA) raise household consumption in rural Kenya?
My strategy: instrument household M-PESA use with distance to the nearest agent, in a 2SLS regression on cross-sectional household data, with county fixed effects and controls.

Attack this identification strategy specifically:
1. State the exclusion restriction in words. Give me three concrete channels through which distance-to-agent could affect consumption OTHER than through M-PESA use.
2. Is agent placement plausibly as-good-as-random, or is it endogenous to local wealth and infrastructure? What does that do to the instrument?
3. If the exclusion restriction fails in the direction you think most likely, which way does my estimate get biased, and roughly how much should I worry?
4. List the diagnostics you would demand (first-stage F-statistic, over-identification tests) and tell me which are NOT available given I have a single instrument.
Do not propose any fixes until you have listed every threat.

Debugging Stata, R and Python

For debugging statistical code, AI is a large accelerant. Give it three things: the exact error message, the shape of your data (variables, types, panel structure), and what you were trying to do. Paste a minimal reproducible example, not your whole 400-line do-file. It will usually spot the mismatch faster than you will. The discipline is verification - read the fix and confirm it does what you meant: that it clusters at the level you intended, absorbs the fixed effects you wanted, and has not silently dropped half your sample through a listwise deletion you never noticed.

# BUG: my clustered SEs come out identical to the default ones - what went wrong?
library(plm)
m <- plm(log_wage ~ educ + exper, data = d, model = "within",
         index = c("id", "year"))
summary(m)   # this prints DEFAULT (non-robust) standard errors

# WHY: the within transformation sweeps out the fixed effects; it does NOT cluster
# the standard errors. plm's summary() still reports classical SEs, so no clustering
# has actually happened. A good AI answer should catch this - a bad one will invent
# a phantom 'cluster = TRUE' argument that plm does not have.

# FIX: request a clustered variance-covariance matrix explicitly, then re-test.
library(lmtest); library(sandwich)
coeftest(m, vcov = vcovHC(m, type = "HC1", cluster = "group"))  # cluster by panel id

Five ways AI fools econometricians

• Correlation dressed as causation. Given any regression, AI will cheerfully narrate coefficients as effects - 'an extra year of schooling raises wages by 12%' - with nothing identifying the verb 'raises'.
• Invented diagnostics. It will confidently cite a test that does not apply to your setting, or reach for a Hausman or Durbin-Wu-Hausman test where it is not identified, complete with a plausible-looking interpretation of the number.
• P-value misstatement. It will tell you p = 0.03 means 'a 3% chance the result is due to chance' or 'a 97% chance the effect is real'. Both are wrong, and it will not warn you.
• Endogeneity ignored. Ask it to interpret an OLS on observational data and it will reach straight for a causal reading, never raising omitted-variable bias, simultaneity or measurement error unless you force it to.
• P-hacked storytelling. Feed it a table with fifteen coefficients and it will assemble a tidy narrative around the three that happen to be significant, quietly discarding the twelve that are not - the exact behaviour that manufactures false findings.

It helps to keep the mechanical facts to hand, because AI garbles them. A t-statistic is simply the coefficient divided by its standard error: t = b / se. As a rough guide, an absolute t above about 1.96 corresponds to significance at the 5% level in a large sample, which is the same thing as a 95% confidence interval that excludes zero. 'Significant' means 'distinguishable from zero given the noise' - it is a statement about precision, not about magnitude, importance or causality. A coefficient can be highly significant and economically trivial, or large and hopelessly imprecise. AI conflates all of these routinely; you must not.