1 A Motivating Puzzle
A health economist regresses adults' earnings on their childhood family income, reported from memory in a survey decades later. The estimated effect is small and she concludes that childhood circumstances barely matter. A labour economist regresses wages on years of schooling, where "years of schooling" is self-reported and error-prone. He, too, gets a coefficient he suspects is too small. Both are likely victims of the same culprit: measurement error in the explanatory variable, which biases ordinary least squares (OLS) toward zero. Understanding this bias how it arises, which way it points, and how to fix it is essential, because almost no variable we use is measured perfectly [Wooldridge, 2010].
2 The Setup: Classical Measurement Error
Suppose the true relationship we care about is
where X* is the true regressor (true schooling, true childhood income). We do not observe X*. Instead we observe a noisy proxy
where u is measurement error. The classical measurement error assumptions are that the error is mean-zero and uncorrelated with the true value and with the equation error: E[u]=0, Cov(X*,u)=0, Cov(ε,u)=0 In words, the mistakes are pure noise, unrelated to how much schooling someone actually has.
What happens if we run OLS of Y on the mismeasured X? Substituting X*=X-u into (1) gives Y=β₀+β₁X+(ε-β₁u) The new regressor X is correlated with the new error term through u: both contain u. This violates the OLS exogeneity requirement, and the estimator is biased.
3 The Attenuation Result
The probability limit of the OLS slope is
The multiplier λ, called the reliability ratio, is the share of the observed variance in X that reflects genuine signal rather than noise. Because 0<λ<1 whenever there is any error, the estimate is pulled toward zero: this is attenuation bias. Two messages follow:
- The bias is always toward zero (for a single regressor). Noisy measurement makes real effects look weaker than they are never stronger.
- The size of the bias depends on the signal-to-noise ratio. If half the variance in measured schooling is noise (σᵤ²=σₓ⁎²), then λ=0.5 and OLS recovers only half the true effect.
A numerical example. Suppose the true return to a year of schooling is β₁=0.10 (a 10% wage gain). Self-reported schooling has reliability λ=0.85 (15% of its variance is reporting noise). Then OLS converges to 0.10×0.85=0.085; an estimated return of 8.5%, understating the truth by a sixth. If the proxy is much worse, say λ=0.6, the estimate collapses to 0.06, and a researcher unaware of the problem might wrongly conclude that schooling "hardly pays".
4 Error in the Outcome Is Harmless (Mostly)
A common confusion is to fear all measurement error equally. In fact, classical error in the dependent variable is benign for consistency. If we observe Y=Y*+v with v classical noise, the error simply joins the regression disturbance: Y=β₀+β₁X*+(ε+v). As long as v is uncorrelated with X*, OLS remains unbiased and consistent it only becomes less precise (larger standard errors). The asymmetry is worth memorising: noise in X biases the slope; noise in Y merely inflates the variance.
5 When the Simple Story Breaks
Two complications matter in practice.
Non-classical error. If the measurement error is correlated with the truth for example, high earners systematically under-report income, or low-schooling respondents round up- the error is non-classical, and the bias need no longer be toward zero. It can go in either direction or even change sign, so the comforting "attenuation only" result no longer applies.
Multiple regressors. With several explanatory variables, mismeasuring one of them contaminates the coefficients on the others, even if those others are measured perfectly. The bias on the well-measured variables can be positive or negative depending on the correlations among regressors. "Attenuation toward zero" is strictly a single-regressor result; in a multivariate model, mismeasurement spreads bias unpredictably across all the coefficients.
6 Fixing It: The Instrumental Variables Solution
The standard remedy is an instrumental variable (IV)-a variable Z that is correlated with the true X* but uncorrelated with both the equation error ε and the measurement error u [Angrist and Pischke, 2009]. A classic device is a second, independent measurement of the same quantity. If X₁=X*+u₁ and X₂=X*+u₂ with independent errors, then X₂ is a valid instrument for X₁: it tracks the signal X* but its noise u₂ is unrelated to u₁. Two-stage least squares using X₂ to instrument X₁ purges the attenuation and recovers β₁ consistently. This is exactly why studies of the returns to schooling have used a sibling's or co-twin's report of education, or administrative records, to instrument self-reports [Ashenfelter and Krueger, 1994]: the second source breaks the correlation between regressor and error.
7 Common Mistakes and Where to Learn More
- Treating a small coefficient as a small effect. Before declaring "no effect", ask how reliably the regressor is measured. A near-zero estimate from a noisy proxy is consistent with a large true effect.
- Assuming differencing helps. Within-group or first-differenced estimators (twins, panel fixed effects) remove confounding but often amplify attenuation, because differencing strips out the signal variance faster than the noise variance, lowering λ.
- Forgetting the multivariate spillover. One badly measured control can bias the coefficient you care about, so scrutinise the quality of every regressor, not just the treatment.
For deeper treatment, Wooldridge [2010] gives the matrix-algebra derivation, and Angrist and Pischke [2009] connects measurement error to the IV toolkit with applied examples. The practical takeaway is simple: noisy data do not just add scatter they systematically shrink the very effects we are trying to estimate, and recognising attenuation is the first step to correcting for it.
References
- Angrist, J. D., and Pischke, J.-S. (2009). Mostly Harmless Econometrics: An Empiricist' Companion. Princeton University Press.
- Ashenfelter, O., and Krueger, A. (1994). Estimates of the economic return to schooling from a new sample of twins. American Economic Review, 84(5), 1157-1173.
- Wooldridge, J. M. (2010). Econometric Analysis of Cross Section and Panel Data, 2nd ed. MIT Press.