Regression Dilution Bias

Regression dilution bias—also termed attenuation bias or errors-in-variables dilution in econometrics and biostatistics—is the systematic underestimation (flattening toward the null value of zero) of the regression slope linking a continuous long-term ("usual") biological exposure (X) to a clinical disease outcome (Y) when the exposure is quantified using an imperfect single-occasion baseline measurement (W) subject to random technical measurement error and short-term within-person physiological fluctuation. Brought to prominence in cardiovascular epidemiology and public health by Stephen MacMahon, Richard Peto, Rory Collins, and Sarah Lewington in 1990, regression dilution bias serves as the continuous-variable counterpart of nondifferential exposure misclassification. It explains why prospective cohort studies relying on a single baseline clinic reading of blood pressure, plasma cholesterol, hemoglobin A_1c, or dietary sodium intake underestimate the true etiological strength of those risk factors by 30% to 65%—and how repeat measurements in a representative subsample enable mathematical correction via the MacMahon–Peto reliability ratio (λ).

Suppose the true biological determinant of long-term vascular or metabolic disease risk in individual i is their usual (long-term average) exposure level X_i (for example, true 10-year mean diastolic blood pressure), which follows a population distribution with mean μ_X and between-person biological variance σ_X². Assume the outcome Y_i (such as log-hazard of stroke or continuous carotid intima-media thickness) depends linearly on true usual exposure X_i with true structural slope β_true:

Y_i = α + β_true X_i + ε_i   (where Cov(X_i, ε_i) = 0)

In a prospective cohort study, investigators cannot directly observe the latent long-term average X_i; instead, at baseline (t = 0), they record a single-occasion measurement W_i0, which equals the true usual level X_i plus a zero-mean within-person random deviation u_i0 (classical additive measurement error):

W_i0 = X_i + u_i0   where E[u_i0] = 0, Var(u_i0) = σ_u², and Cov(X_i, u_i0) = 0

Here, the within-person variance σ_u² captures two combined sources of short-term noise: (1) analytical or instrument error (sphygmomanometer cuff reading variability or laboratory assay coefficient of variation), and (2) true short-term biological fluctuation around the individual's long-term homeostatic set-point (diurnal circadian rhythms, seasonal variation, recent meals, posture, or transient stress).

Because X_i and u_i0 are independent, the total variance of the observed baseline measurements W_i0 across the cohort is inflated relative to the true between-person variance: Var(W_i0) = σ_X² + σ_u². Meanwhile, because random baseline noise u_i0 does not cause future chronic disease (Cov(u_i0, Y_i) = 0), the covariance between the observed baseline reading W_i0 and the outcome Y_i equals the true covariance: Cov(W_i0, Y_i) = β_true σ_X².

When an investigator regresses the disease outcome Y_i on the observed single-occasion baseline measurement W_i0 using ordinary least squares (or fits a Cox proportional hazards model), the expected naive slope β_naive converges in probability to:

β_naive = Cov(W_i0, Y_i) / Var(W_i0) = β_true · [ σ_X² / (σ_X² + σ_u²) ] = λ · β_true

The dimensionless attenuation factor λ = σ_X² / (σ_X² + σ_u²) ∈ (0, 1] is the regression dilution ratio (also termed the reliability ratio or intraclass correlation coefficient [ICC] of a single measurement). To recover the unbiased etiological slope β_true, the naive slope is divided by λ (Rosner, Willett, & Spiegelman, 1989; Frost & Thompson, 2000): β_corrected = β_naive / λ.

In 1990, Stephen MacMahon and Richard Peto illustrated regression dilution using longitudinal blood pressure re-measurements from the Framingham Heart Study:

1. When 5,000 adults were stratified at baseline (t = 0) into 5 quintiles based on a single clinic diastolic blood pressure (DBP) reading W₀, the lowest baseline quintile had a mean DBP of 74 mmHg and the highest baseline quintile had a mean DBP of 104 mmHg—an apparent baseline spread of ΔW₀ = 30.0 mmHg.
2. However, when the same individuals were re-measured 4 years later (t = 4) while keeping them in their original baseline quintile groups, regression to the mean occurred: the bottom baseline group's mean DBP rose to 80 mmHg and the top baseline group's mean DBP fell to 98 mmHg—a true long-term "usual" spread of ΔW₄ = 18.0 mmHg.
3. The nonparametric MacMahon–Peto regression dilution ratio is the ratio of the follow-up range to the initial baseline range across the baseline-defined groups: λ = ΔW_follow-up / ΔW_baseline = 18.0 / 30.0 = 0.60. Consequently, the true etiological slope is 1 / λ = 1 / 0.60 = 1.67 times (67%) steeper than the naive baseline slope.

In 2002, the Prospective Studies Collaboration analyzed 1 million adults (61 cohorts, 12.7 million person-years) with time-dependent regression dilution corrections. Once corrected for regression dilution bias, the log-linear relationship between usual blood pressure and vascular mortality was 60% to 70% steeper than in uncorrected baseline plots, remained log-linear down to 115/75 mmHg, and reconciled observational cohorts with antihypertensive randomized controlled trials.

If an investigator averages k independent replicate measurements (W̄_i = [1/k] ∑_j=1^k W_i,j) per participant at baseline, the within-person error variance shrinks from σ_u² to σ_u² / k, increasing the reliability ratio according to the Spearman–Brown formula:

λ_k = σ_X² / [ σ_X² + (σ_u² / k) ] = (k · λ₁) / [ 1 + (k − 1)λ₁ ]

For example, if a single blood pressure reading has λ₁ = 0.50, averaging k = 4 independent days of readings raises the reliability ratio to λ₄ = 0.80, reducing attenuation from 50% down to 20%.

While dividing β_naive by λ is unbiased in univariable regression with classical independent error, Frost and Thompson (2000) and Davey Smith and Phillips (1996) highlighted two important boundaries:

1. Multivariable Regression Calibration: In multivariable models containing multiple error-prone correlated exposures (W₁, W₂, ..., W_p), within-person errors can bias coefficients in either direction, requiring matrix regression calibration using the full variance–covariance matrix Σ_XΣ_W⁻¹ or Mendelian randomization genetic instruments.
2. Berkson Error vs. Classical Error: When individuals are assigned the group-mean environmental exposure of their geographic district (X_i = W_group + u_i, termed Berkson measurement error), ordinary linear regression yields an unbiased slope (β_naive = β_true) without requiring regression dilution division.

Why must the repeat measurements used to estimate λ be separated by several years rather than taken 5 minutes apart?

Replicate measurements taken 5 minutes apart at the same clinic visit only capture technical cuff or assay noise, missing day-to-day, seasonal, and multi-year biological drift around the person's true 10-year usual exposure. As shown by Clarke et al. (1999), re-measurements separated by the midpoint of cohort follow-up (3 to 5 years apart) capture both short-term fluctuation and medium-term trajectory drift.

Does measurement error in the continuous outcome variable (Y) also cause regression dilution bias?

No. If random additive error e_i occurs in the dependent outcome variable (Y*_i = Y_i + e_i) while the predictor X_i is measured without error, Cov(X_i, Y*_i) / Var(X_i) remains unbiased (β_naive = β_true), although standard errors increase. Regression dilution occurs specifically when random error resides in the independent predictor variable (W_i = X_i + u_i).

1. MacMahon, S., Peto, R., Cutler, J., Collins, R., Sorlie, P., Neaton, J., Abbott, R., Godwin, J., Dyer, A., & Stamler, J. (1990). "Blood pressure, stroke, and coronary heart disease. Part 1, Prolonged differences in blood pressure: prospective observational studies corrected for the regression dilution bias." The Lancet, 335(8692), 765–774. https://doi.org/10.1016/0140-6736(90)90878-990878-9)
2. Clarke, R., Shipley, M., Lewington, S., Youngman, L., Collins, R., Marmot, M., & Peto, R. (1999). "Underestimation of risk associations due to regression dilution in long-term follow-up of prospective studies." American Journal of Epidemiology, 150(4), 341–353. https://doi.org/10.1093/oxfordjournals.aje.a010013
3. Frost, C., & Thompson, S. G. (2000). "Correcting for regression dilution bias: comparison of methods for a single predictor variable." Journal of the Royal Statistical Society: Series A (Statistics in Society), 163(2), 173–189. https://doi.org/10.1111/1467-985X.00164
4. Prospective Studies Collaboration. (2002). "Age-specific relevance of usual blood pressure to vascular mortality: a meta-analysis of individual data for one million adults in 61 prospective studies." The Lancet, 360(9349), 1903–1913. https://doi.org/10.1016/S0140-6736(02)11911-811911-8)