Statistics Glossary → Independence of irrelevant alternatives (IIA)
Independence of irrelevant alternatives (IIA)concept
The assumption in multinomial and conditional logit models that the odds of choosing one option over another do not depend on which other options are available.
In a multinomial logit model, the odds of options A and B depend only on the characteristics of those two options; adding or removing a third option C leaves them unchanged. This means that a new option draws probability proportionally from all the existing options. When options are very similar to each other, the assumption is unrealistic.
The Hausman-McFadden test estimates the model once with all options and once with one option removed, and compares the remaining coefficients; a clear difference suggests that IIA is violated. These tests have low power and their results depend on which option is dropped. If the assumption fails, models that allow correlation between options, such as nested logit, mixed logit or multinomial probit, can be used.
An everyday example
In McFadden's classic example, half of commuters drive and half take a red bus. When a blue bus identical to the red one is added, the realistic outcome is that bus riders split in two (car 50%, each bus 25%). A model assuming IIA instead keeps the odds between car and buses fixed and gives each option one third, which is not realistic.
In MerQur
'Extra output: IIA test (Hausman-McFadden)' in Multinomial Logistic Regression; the remaining coefficients are expected not to change when an option is removed, and the test is computed as H = d'(V_r − V_f)⁻¹d.
Analyses where this concept appears
Multinomial Logistic Regression