Regression Dilution, explained.
Regression dilution, or attenuation bias, can weaken a regression relationship when a predictor is measured with independent classical error.
Why it happens
True x values are −10 to 10 and y=20+2x. Add independent uniform error with half-width a to x, then fit an ordinary least-squares slope. Compare it with two and the variance-ratio benchmark 2(110/3)/(110/3+a²/3).
Regression dilution, or attenuation bias, can weaken a regression relationship when a predictor is measured with independent classical error.
Read the result
The outcome remains tied to true x while the horizontal coordinates get noisy. Compare the fitted slope with the reference without expecting every finite-sample path to be monotonic.
A worked example
A reliable relationship measured imprecisely
With error width zero, every point lies on y equals twenty plus two x and the fitted slope is exactly two.
Increase the predictor error half-width while leaving the true relationship and outcome values unchanged.
Compare the realised slope and variance-ratio benchmark, then try a new seed to distinguish expected attenuation from a particular dataset.
OPTIONAL DEEPER DETAILGo deeper: inside the model
Inside this model
True x values are −10 to 10 and y=20+2x. Add independent uniform error with half-width a to x, then fit an ordinary least-squares slope. Compare it with two and the variance-ratio benchmark 2(110/3)/(110/3+a²/3).
Where this idea is useful
A practical use
A noisy questionnaire measuring exposure may obscure a relationship that would be clearer with a reliable measurement.
A common misconception
“Measurement noise always just widens uncertainty without changing estimates.”
Classical error in a predictor can also bias a slope toward zero. Other error structures can produce different directions or more complex biases.
What this explanation leaves out
- The benchmark is a population-moment reference, not an exact expectation for this fixed finite design. One seeded slope need not decline monotonically with error. Correlated or systematic error can bias results differently.
Is this the same as regression to the mean?
No. Regression dilution concerns measurement error in a fitted relationship. Regression to the mean concerns expected outcomes after selecting observations partly on extreme noise.
Which measured predictor might be less reliable than the analysis assumes?
Associated thinkers
Associations marked provisional are awaiting source review.
Further reading
Explore the original research or the teaching reference behind this experiment.