Statistics Glossary → Reference category (comparison level)
Reference category (comparison level)concept
The level against which the other levels are compared when a categorical variable enters a model as dummy variables, and whose own coefficient is fixed at zero.
A categorical variable with k levels enters a regression as k - 1 dummy variables; the level left out is the reference. Each dummy's coefficient shows how that level differs from the reference (a mean difference in linear regression, a log-odds difference in logistic regression, i.e. an odds ratio once exponentiated). The model's intercept corresponds to the expected value in the reference level.
The choice of reference does not change the model's fit, predictions or overall tests; it only changes which comparison the coefficients describe. To keep things interpretable, the reference is usually a natural comparison group (the control group, the most common category, a 'none' level). Choosing a level with very few observations as the reference inflates the standard errors of all coefficients.
An everyday example
The effect of smoking on disease risk is modelled with the levels 'never', 'former' and 'current'. With 'never' as the reference, two odds ratios result: former versus never and current versus never. Had 'current' been the reference, the same model would describe two other comparisons.
In MerQur
'Reference category (factors)' (First level, Last level) in multiple linear and logistic regression; 'Reference class' in multinomial logistic regression (it changes only the comparison the coefficients express, not the model); 'Reference level' in linear mixed models; 'Reference levels' in the weighted regressions and parametric survival; 'ASC reference alternative' in the conditional logit.
Analyses where this concept appears
Multiple Comparison CorrectionMultiple Linear RegressionLogistic RegressionMultinomial Logistic RegressionLinear Mixed Models (LMM)Weighted Linear Regression (Survey Regression)Weighted Logistic Regression (Survey Logistic Regression)Robust Regression (M-estimator)Probit RegressionConditional Logit (Choice Model)Bayesian Linear RegressionHierarchical Bayesian RegressionMediation AnalysisParametric Survival (AFT)Competing RisksTime-dependent Cox RegressionSurvey-PHREG (Weighted Cox)Interval-Censored SurvivalCluster-Robust Cox (Marginal and Frailty Models)Age Standardization and SMR