Job market paper · with Alexander Kwon · November 2026 draft

From One LATE to Another

Machine Learning and the External Validity of IV Estimates

Read the paper (PDF)

The question

An instrumental variable (IV) study estimates a local average treatment effect (LATE) in one population, which we call the sample population. We want to predict the LATE in another population, the target population, where we only observe people's characteristics. The prediction error is the prediction minus the true value. Zero means a perfect prediction.

I2SLS
Interacted two-stage least squares, the standard approach. It controls for characteristics with linear terms.
DML
Double/debiased machine learning. It uses machine learning to estimate how characteristics relate to the outcome, the treatment and the instrument, so these relationships do not have to be linear.

Notice that both methods still model the treatment effect as linear in characteristics.

Where the prediction error comes from

We split the prediction error into four terms that add up to the total. The terms can have opposite signs, so they can partly cancel.

Term I · Sample errorThe estimate is already off in the sample population itself.
Term II · Slope errorThe estimated slope of the effect is off. Because the target population has different characteristics, this error carries into the prediction.
Term III · Curve errorThe true effect is nonlinear, but the model is linear. This matters when the two populations differ.
Term IV · Weighting errorThe weights that tell us who in the target population responds to the instrument are estimated with error.

Two simulations

We simulate the data, so we know the true LATE and can measure each method's error.

Simple designPeople differ in one characteristic. We set each relationship ourselves, so we can make it nonlinear on purpose. Each setting is repeated 500 times with 10,000 people.
Card designCalibrated to Card (1995), a classic study of the return to schooling that uses living near a college as the instrument. Each setting is repeated 300 times with 18,000 people.

What DML fixes

Characteristics relate to the outcome, the treatment and the instrument. I2SLS controls for these relationships with linear terms, while DML estimates them flexibly. When they are nonlinear, the linear terms in I2SLS can distort the estimated effect. The error shows up in Term I and, when the target population differs from the sample population, in Term II. This figure varies the instrument and outcome relationships and the difference between the sample and target populations.

  • This part uses the simple design, where people differ in one characteristic, X.
  • Difference between sample and target populations is how different the target population's mix of X is from the sample population's.
  • Spread of X is how much people differ in X.
  • Instrument nonlinearity is how nonlinear the chance of getting the instrument is in X. That chance follows an S-shaped (logistic) curve.
  • Outcome nonlinearity is how nonlinear the direct link from X to the outcome is, apart from the treatment effect. The true effect is the same at both levels.

In this figure the true effect is linear, so Term III is zero.

Difference between sample and target populations
Spread of X
Instrument nonlinearity
Outcome nonlinearity

I2SLSDMLA bar above zero means the prediction is too high, and a bar below zero means it is too low. The vertical scale changes with the spread of X and the difference between the sample and target populations.
Show the numbers

Numbers come from Table 2 of the paper (simple design). The Card note is from Table 15. Hover over or tap a bar to see its Monte Carlo standard error.

What DML does not fix

Two sources of error remain for both methods. This part uses the Card design. Pick a tab, change the setting, and watch how each term moves.

  • The treatment effect can be nonlinear in characteristics. Both methods model it as linear, so DML's flexibility does not help with this error.
  • Treatment effect nonlinearity is how nonlinear the true effect is. The average effect is the same at both levels.

I2SLSDMLA bar above zero means the prediction is too high, and a bar below zero means it is too low.
Show the numbers

Numbers come from Table 3 of the paper (Card design). Hover over or tap a bar to see its Monte Carlo standard error.

I built this page with Claude Code, an AI coding assistant, from the November 13, 2026 draft. A script read every number from the paper's tables, and each row was checked against the totals in the paper.