Specification & Data Problemsintermediate

Measurement Error and Proxy Variables

Measurement error in the dependent variable raises the noise and inflates standard errors but, under the classical assumption, leaves OLS unbiased. Classical measurement error in an explanatory variable is more damaging: it causes attenuation bias, pulling the estimated coefficient toward zero. A proxy variable is an observable stand-in for an unobserved factor, such as IQ for innate ability, and including a good proxy reduces omitted variable bias. Like OVB, mismeasurement of a regressor threatens consistency, not just precision.

Why it matters

If the ruler you use to measure the outcome is noisy, your predictions are fuzzier but not systematically off. If instead the ruler is on a key input, the noise blurs the link between that input and the outcome, so the estimated effect looks weaker than it really is, dragged toward zero. Proxies fight a related battle: when you cannot observe ability or quality directly, a decent stand-in soaks up part of what would otherwise be hidden and bias your other coefficients.

Formulas

Classical errors-in-variables (attenuation)

\operatorname{plim}\hat{\beta}_1=\beta_1\,\frac{\sigma_{x^{*}}^2}{\sigma_{x^{*}}^2+\sigma_e^2}

The multiplier is between 0 and 1, so the estimate is biased toward zero. More measurement noise

\sigma_e^2

means stronger attenuation.

Proxy variable regression

y=\beta_0+\beta_1 x_1+\beta_2 x_2+\beta_3\, z+u

z

is a proxy for an unobserved factor correlated with

x_1

. A good proxy reduces the omitted variable bias in

\hat{\beta}_1

Worked examples

Scenario

Self-reported income (the regressor) is recorded with survey error in a savings equation.

Solution

If true income is measured with classical noise, `regress saving income` produces an attenuated slope, biased toward zero, so the marginal propensity to save looks smaller than it is. The remedy is a better income measure or an instrumental variable, since adding controls alone does not undo errors-in-variables attenuation.

NoteAttenuation is a consistency problem: it does not vanish as the sample grows.

Scenario

A wage equation cannot observe innate ability, which is correlated with education.

Solution

Run `regress lwage educ exper IQ`, using `IQ` as a proxy for ability. Because ability raises both schooling and wages, omitting it biases the return to education; a reasonable proxy absorbs part of that effect and brings the educ coefficient closer to its true value, reducing omitted variable bias.

Common mistakes

✗Assuming all measurement error biases coefficients. Classical error in $y$ alone inflates the error variance but keeps OLS unbiased; error in a regressor is what causes attenuation.
✗Thinking attenuation makes effects look larger. Classical errors-in-variables shrinks the coefficient toward zero, understating the true effect.
✗Believing a bigger sample cures measurement error. Attenuation is a consistency problem; the bias persists no matter how large $n$ becomes.
✗Treating any available variable as a valid proxy. A useful proxy must be genuinely related to the unobserved factor and, ideally, redundant once that factor is controlled.

Revision bullets

•Classical error in $y$ : more noise, larger SEs, still unbiased.
•Classical error in $x$ : attenuation bias toward zero.
•Mismeasured regressors threaten consistency, not just precision.
•A proxy stands in for an unobserved factor to reduce OVB.
•Attenuation does not disappear as the sample size grows.

Quick check

Classical measurement error in an explanatory variable causes:

Including IQ as a proxy for unobserved ability in a wage equation is intended to:

Connected topics

MLR model Omitted var bias MLR assumptions Asymptotics Multicollin.

Sources

Wooldridge (2019), §9.4
Wooldridge, Jeffrey M. Introductory Econometrics: A Modern Approach. 7th ed. Cengage, 2019.
Derives classical errors-in-variables attenuation for a mismeasured regressor and the benign case of error in $y$ .
Wooldridge (2019), §9.2
Wooldridge, Jeffrey M. Introductory Econometrics: A Modern Approach. 7th ed. Cengage, 2019.
Develops the proxy variable solution to omitted variables, using IQ for unobserved ability.

How to cite this page

Dr. Phil's Quant Lab. (2026). Measurement Error and Proxy Variables. Derivatives Atlas. https://phucnguyenvan.com/concept/efm-measurement-error-proxies

← Back to the atlas See in the network →