Psychological and behavioral measures are imperfect proxies for the constructs they are meant to capture. Every observed score reflects some true signal plus random noise — and that noise has consequences. When two variables are each measured with error, the observed correlation between them is systematically pulled toward zero relative to the true correlation between the underlying constructs. This phenomenon is called attenuation, and its magnitude depends on how reliable each measure is.

The attenuation formula, $r_{\text{obs}} \approx r_{\text{true}} \sqrt{\rho_{xx'} \rho_{yy'}}$, shows that even a strong true relationship can appear weak if the measures are noisy. A correlation between two scales with reliability 0.70 each is attenuated to about 70% of its true value before sampling variability is even considered. The correction for attenuation reverses this by dividing the observed correlation by $\sqrt{\rho_{xx'} \rho_{yy'}}$, recovering an estimate of the true relationship — but at the cost of amplified sampling variance, particularly when reliability is low.

Set reliability to 1.0 for both variables to see what a perfect measure would look like. Then reduce reliability and watch the scatterplot cloud widen and the observed correlation shrink — even though the true correlation is unchanged.

A sample of 160 points is drawn each time; use the slider to change the number of observations. Observed values are generated by adding random noise to latent variables with the specified true correlation. Reliability is defined as $\mathrm{var}(T)/\mathrm{var}(O)$; lower reliability means more error.

Three correlations are shown below:
Observed r — the sample correlation between the noisy observed scores.
Expected r — the predicted attenuated correlation: $r_{\text{expected}} = r_{\text{true}} \sqrt{\rho_{xx'} \rho_{yy'}}$. Low reliability pushes this toward zero.
Corrected r — the correction for attenuation applied to the observed r: $r_{\text{corrected}} = \frac{r_{\text{obs}}}{\sqrt{\rho_{xx'} \rho_{yy'}}}$. This attempts to recover the true correlation from the noisy data. With low reliability the correction is violently unstable: at ρ_xx′ = ρ_yy′ = .20 roughly one replication in ten returns a corrected r that exceeds ±1 outright — an impossible value for a correlation, and a signal that the correction has been pushed past what the data can support.

Sampling Distribution Simulation

This simulation draws 1000 independent samples using the settings above and computes the observed r and the corrected r for each. The histograms show how these values are distributed across replications. Vertical lines mark the true correlation (dashed) and the expected attenuated correlation (solid teal).

Convergence Trajectories

Each line is an independent study where participants are added one at a time. The running observed r converges to the attenuated value (solid teal line), while the running corrected r converges to the true correlation (dashed white line) — but with substantially more variability, especially when reliability is low.

Citation

Persson, B. N. (2026). Measurement error and attenuation [Interactive visualization]. https://bjorn-persson.github.io/visualizations/measurement-error/

@misc{Persson2026measurement,
  author = {Björn N. Persson},
  year = {2026},
  title = {Measurement Error and Attenuation},
  note = {Interactive visualization},
  url = {https://bjorn-persson.github.io/visualizations/measurement-error/}}