Correcting economic models when the measurements are wrong

A survey records income, but the recorded figure differs from the truth. If researchers use it to model travel choices, the resulting estimates can be systematically wrong—not merely less precise. In nonlinear models, where relationships need not follow straight lines, correcting this problem can require elaborate statistical machinery.

Kirill S. Evdokimov and Andrei Zeleneev propose a simpler alternative in their arXiv preprint, Simple Estimation of Semiparametric Models with Measurement Errors. Version 5 was posted on November 9, 2026. Their method targets measurement errors that are small or moderate relative to the variation in the underlying variable, rather than errors of arbitrary size.

Correct the equations, not each observation

The approach works within the generalized method of moments, or GMM. Researchers specify mathematical relationships whose averages should equal zero when a model has the correct parameter values. These relationships are called moment conditions. GMM chooses estimates that bring their sample averages as close to zero as possible.

Replacing a true variable with a noisy measurement generally breaks those relationships. Even at the correct parameter values, their averages may no longer be zero. More observations can then produce narrower confidence intervals around a biased estimate.

The authors' Measurement Error Robust Moments estimator, abbreviated MERM, modifies the moment conditions themselves. It does not attempt to recover every person's true income.

The construction uses Taylor expansions: approximations describing how a smooth function changes when its input is perturbed. In the simplest correction, the leading distortion depends on the measurement-error variance and the moment function's curvature, measured by its second derivative. Subtracting an estimated curvature-based term counteracts that distortion.

The error variance need not be known beforehand. MERM estimates correction parameters jointly with the model parameters using standard GMM software. Higher-order corrections use additional derivatives and account for the fact that correction terms can themselves be biased. The procedure avoids estimating an entire unknown probability distribution, a task that becomes difficult when many explanatory variables are involved.

What the guarantees require

The main analysis assumes classical measurement error: an additive error with mean zero that is independent of the true variable and the other variables entering the moment conditions. It does not require normally distributed errors, although it does impose moment and smoothness conditions.

The data must also contain enough information to distinguish the model parameters from the additional correction parameters. Suitable instrumental variables or second measurements can provide that information. An instrument supplies identifying information under specified exclusion restrictions; simply adding any correlated variable is not sufficient.

Under the paper's assumptions, MERM has the familiar square-root-of-sample-size convergence rate and supports conventional GMM confidence intervals and tests. Crucially, the proof uses an approximation in which the noise-to-signal ratio slowly approaches zero as sample size increases. This is a mathematical device for studying moderate errors, not a claim that collecting more survey responses automatically improves measurement quality.

Testing against known answers

The authors assess performance through repeated simulated datasets, where the true answers are known. Their experiments cover several nonlinear regression models and models of choosing among three alternatives.

One experiment draws on the ModeCanada survey of business travelers between Montreal and Toronto. The researchers use 2,769 travelers with train, air and car available to calibrate a choice model. They then generate artificial choices, add independent Gaussian errors to income, and generate an instrument. This is a simulation informed by real data—not direct evidence that the method has recovered actual travelers' true incomes or preferences.

Across 5,000 repetitions, the higher-order correction substantially improves inference. When the error standard deviation is 75% of the true income standard deviation, a conventional test for the air-travel income elasticity falsely rejects its correct simulated value 99.50% of the time. With the fourth-order MERM correction, that rate is 5.84%, much closer to the intended 5%.

A correction with boundaries

More correction is not always better. Higher expansion orders can remove more bias but usually require estimating more parameters, increasing uncertainty. The paper proposes a data-driven choice of correction order and describes sample splitting for valid inference after selection.

The method also requires sufficiently differentiable moment functions, excluding some models such as instrumental-variable quantile regression. When noise overwhelms signal, its approximations become less dependable: additional simulations show deteriorating inference at noise-to-signal ratios of 1.5 and 2.

MERM therefore offers a conditional practical option, not a universal repair for bad data. Its promise is to address consequential measurement bias without the computational burden of reconstructing full distributions—provided the model, identifying information and error magnitude support the correction.