Statistics Glossary → Prior class probabilities (discriminant analysis)
Prior class probabilities (discriminant analysis)concept
In classification, the assumed probability that an observation belongs to each class before the data are seen; these probabilities also shape the classification decision and its threshold.
Discriminant analysis assigns an observation, by Bayes' rule, to the class with the highest posterior probability; the posterior is proportional to the class's prior probability times the density of the observation in that class. With equal priors (1/K for each class), the decision rests only on how close the observation is to the class centres. Using the class proportions in the sample makes common classes easier to assign. If the population proportions are known (for example the prevalence of a disease), supplying them as user-defined priors gives more realistic results.
Priors do not change the direction of the discriminant functions, but they shift the decision boundaries between classes, so the classification table and the accuracy rate are sensitive to them. If the sample proportions differ from those in the population (as in a case-control design that samples patients and healthy people half and half), using the sample proportions as priors is misleading. If missing a rare class is costly, the priors or misclassification costs can be adjusted accordingly. Decision trees can use prior probabilities in a similar way.
An everyday example
A mushroom forager judges whether a find is poisonous from its cap colour and stem length. Half the mushrooms in her sample are poisonous, but only 5% of those in the region are. With 50%-50% priors she would mistake many edible mushrooms for poisonous ones; with 5%-95% priors her guesses fit the real situation in the region much better.
In MerQur
In Discriminant Analysis, the 'Prior probabilities' option: Equal (1/K; default), Class proportions or User-defined (with the 'Prior values' field; normalized to 1). The priors change the classification boundaries, the posterior probabilities and the Fisher classification constant, but not the canonical solution. In Decision Tree, the 'Imbalanced classes: balanced weights' option uses equal prior probabilities (prior = 1/K).
Analyses where this concept appears
Decision Tree (CART)Discriminant Analysis (LDA / QDA)