MerQuR+Statistics Glossary

Statistics Glossary → Fixed and random effects

Fixed and random effectsconcept

The distinction between modelling a factor's levels as all the levels of interest (fixed) or as a sample drawn from a wider population (random).

With a fixed effect, a separate parameter is estimated for each level and the conclusions apply only to those levels (e.g. three specific drug doses). With a random effect, the levels are treated as drawn from a distribution (usually normal with mean zero and variance σ²); the interest is not in individual levels but in the variability between them (a variance component), and conclusions generalize to the population the levels came from (e.g. randomly selected schools). Predictions of random effects (BLUPs) shrink towards the overall mean.

In meta-analysis, the fixed-effect model assumes all studies estimate one common true effect, while the random-effects model assumes the true effects vary from study to study (τ² > 0). Panel econometrics uses the terms a little differently: the fixed-effects estimator uses within-unit variation and allows the unit effect to be correlated with the regressors, whereas the random-effects estimator assumes no such correlation and is more efficient. The choice between them is often made with the Hausman test.

An everyday example

An education researcher studies pupil achievement in 30 schools picked at random from all schools in the country. The aim is not to compare these 30 schools one by one but to learn how large the differences between schools are and to generalize to all schools, so school is modelled as a random effect. The 'teaching method' (A or B) in the same study is a fixed effect.

In MerQur

The 'Model' option of Meta-Analysis (Random effects, Fixed effect); 'Group variable (random intercept)' in Linear Mixed Models and GLMM; 'Upper group (P) type' in the Nested Mixed Model and 'Factor A type / Factor B type' (random or fixed) in the Crossed Mixed Model; 'Model' (fixed effects, random effects, pooled OLS …) and 'RE variance estimator' in Panel Data Regression. In Cluster-Robust Cox, the shared gamma frailty form of the 'Model' option adds a random effect shared within clusters.

Analyses where this concept appears

Meta-AnalysisLinear Mixed Models (LMM)Nested Mixed Model (Nested LMM, Multi-Level Hierarchical Mixed Model)Crossed Mixed Model (Crossed LMM, Crossed Random Effects Mixed Model)Generalized Linear Mixed Models (GLMM)Panel Data RegressionDifference-in-Differences (DiD)Variance Components Analysis (VARCOMP)Bayesian ANOVAHierarchical Bayesian RegressionCluster-Robust Cox (Marginal and Frailty Models)

Related concepts

Random slopeVariance componentsNested and crossed random effects