Nested Logit (NL) is an extension of the Multinomial Logit (MNL) model that groups similar alternatives into nests, relaxing the restrictive Independence of Irrelevant Alternatives (IIA) assumption. It is the correct choice whenever a market contains alternatives that are closer substitutes for each other than for the rest of the menu.
The problem with standard MNL
Standard MNL assumes error terms across alternatives are independent — which implies that if a new alternative is added, every existing alternative loses share in the same proportion. The textbook failure case is the red-bus / blue-bus problem:
A commuter chooses between car (50%) and red bus (50%). Adding a blue bus — identical to the red bus except colour — should steal share only from the red bus. Standard MNL instead predicts a 33/33/33 split, because the model can’t tell that red and blue buses are the same thing with a different coat of paint.
How Nested Logit fixes it
NL groups the two buses into a “bus” nest and models the choice in two steps: first between car and bus, then within the bus nest between red and blue. The choice probability becomes:
$$P(j) = \underbrace{\frac{e^{V_j / \lambda_k}}{\sum_{i \in B_k} e^{V_i / \lambda_k}}}_{\text{within-nest}} \cdot \underbrace{\frac{(\sum_{i \in B_k} e^{V_i / \lambda_k})^{\lambda_k}}{\sum_m (\sum_{i \in B_m} e^{V_i / \lambda_m})^{\lambda_m}}}_{\text{between-nest}}$$
where $V_j$ is deterministic utility and $\lambda_k \in (0, 1]$ is the nest coefficient. When $\lambda_k = 1$, NL collapses to MNL. When $\lambda_k < 1$, alternatives within the nest are correlated — which is exactly the red-bus/blue-bus structure.
When to use Nested Logit over CBC
Branded vs private-label — private-label options are near-identical substitutes
Transport mode choice — public transport options correlate
Insurance tiers — “Gold Plan A” and “Gold Plan B” compete with each other more than with “Silver”
Menu-based studies — several items serve the same job-to-be-done
Diagnostic test
Fit both MNL and NL and run a likelihood ratio test on the nest coefficients. If $\lambda_k$ is reliably below 1, the nest structure matters and NL is the honest model.