Statistics Glossary → Best linear unbiased prediction (BLUP)
Best linear unbiased prediction (BLUP)concept
In mixed models, the prediction of each group's random effect (a school, patient or plant family, for instance) obtained by shrinking the group's data toward the overall mean.
In a linear mixed model (y = Xβ + Zb + ε) the fixed effects β are estimated, while the random effects b are random variables rather than parameters, which is why the term 'prediction' is used instead of 'estimation'. The BLUP, developed by Henderson, is the conditional expectation of b given the data. In a simple random-intercept model a group's BLUP is approximately [τ² / (τ² + σ²/nⱼ)] × (ȳⱼ − overall mean). The bracketed term is the group's reliability: if the group is large or the between-group variance τ² is high, the group's own mean is trusted; if the group is small, the prediction is pulled toward the overall mean (shrinkage).
This shrinkage acknowledges that extreme values in small groups may be due to chance, and on average it gives more accurate predictions than the raw group means. In practice the variance components are estimated from the data (usually by REML), so the result is also called the empirical BLUP (EBLUP). BLUPs are useful for ranking and selection, for example breeding values in plant breeding or hospital performance comparisons, but their uncertainty should not be ignored, and in small groups they can shrink real differences as well.
An everyday example
In an exam the overall mean is 70. School A, with 5 pupils, and school B, with 100 pupils, both have a mean of 80. If the between-school variance is 25 and the pupil-level variance is 100, the BLUP-based prediction is about 75.6 for school A and 79.6 for school B. The small school's high mean may be partly due to chance, so it is pulled further toward the overall mean.
In MerQur
In the linear mixed model, 'Residual diagnostics + random-effect predictions (BLUP)' gives BLUP ± conditional SD for each group. Hierarchical Bayesian regression reports group-level BLUPs together with the fixed effects, variance components and ICC; this module is not fully Bayesian and rests on REML estimation.
Analyses where this concept appears
Linear Mixed Models (LMM)Hierarchical Bayesian Regression
Related concepts
Fixed and random effectsRestricted maximum likelihood (REML)