Nested Logit Model
Drawbacks of MNL
multinominal logit
MNL may not work well in either of the following cases
due to its IIA property:
Independence of Irrelevant Alternatives
When alternatives are not independent
i.e., when there are groups of alternatives which are more similar
than others, such as public transport modes versus the private
vehicles
When there are taste variations among individuals
i.e., perceptions of individuals vary with their socio economic status
In such cases we require random coefficient models rather than
mean value models as the MNL
Nested logit or hierarchical logit model addresses to a
limited extent the first problem
Nested Logit Model
Subsets of alternatives
[AI(q)] which are similar are
grouped in hierarchies or
nests.
Each nest in turn is
considered as a composite
alternative (NI) which
competes with the other
alternatives [A(q) AI(q)]
available to the individual.
AS(q) = A(q) -AI(q) + NI
NI
AI(q)
Nested Logit Model
First estimate an MNL for the AI(q) alternatives of the lower nest,
taking care of omitting all those variables (z) which take the same
value for this subset of options.
The utility of the composite alternative has two components:
One that consists of the expected maximum utility (EMU) of the lower
nest options, and
Another which considers the vector z of attributes which are common to
all members of the nest
EMU has the following expression:
EMU = log j exp(Wj)
Where, Wj is the utility of alternative Aj in the nest
Therefore, the composite utility of the nest is:
VI = EMU + z
Where, and are parameters to be estimated
Nested Logit Model
At the higher nest, an MNL consisting of all
composite alternatives representing lower
hierarchies and alternatives which are nonnested at that level is estimated.
The probability that individual q selects option
AjAI(q) is computed as the product of the
marginal probability of choosing the composite
alternative NI (in the higher nest) and the
conditional probability of choosing option Aj (in
the lower nest).
Estimation of NL
Sequential estimation method
The method outlined is sequential method of
estimation of NL
Simultaneous estimation method
This is also termed as full information method of
estimation.
A single maximum likelihood function is formulated
and maximised.
ALOGIT estimates NL models using the efficient
simultaneous estimation method.
NESTED LOGIT MODEL IN A TRINOMIAL
MODAL CHOICE SITUATION
Consider a situation
involving choice among
car, bus and metro.
Assuming that bus and
metro are correlated,
they will be considered
Car
at the lower nest. The
composite alternative of
this nest is PT.
PT
Bus
Metro
The lower public transport nest would be modelled by a simple
binary logit model of the form
P ( M / PT ) =
exp(WM )
exp(WM ) + exp(WB )
and
P ( B / PT ) = 1 P ( M / PT )
Where the utilities W contain only
those elements which are not
common to both modes (i.e., the
cost of travel would not enter if
both the modes charged the same
fare)
PT
Car
Bus
Metro
Another binary logit model at the higher nest between car and
the composite alternative PT is modelled as:
P (C ) =
exp(VC )
exp(VC ) + exp(VPT )
and
P ( PT ) = 1 P (C )
Where Vc incorporates all the attributes of the car option, i.e., it
has exactly the same form as in a MNL
The public transport utility is given by
VPT = EMU + k Z k
k
where
EMU = ln[exp(WB ) + exp(WM )]
and the summation over k considers all the common elements
Z that were taken out to estimate the binary logit model at the
lower nest
The modelled choice probabilities
of each option are given by
PC = P (C )
PB = P ( B / PT ) P ( PT )
PM = P ( M / PT ) P ( PT )
PT
Car
Bu
s
Metro
Consistency of Structural Parameter
The structural parameter should satisfy the condition:
0 < 1
If < 0, an increase in the utility of an alternative in the nest, which
should increase the value of EMU, would actually diminish the
probability of selecting the nest
If = 0, such an increase would not effect the nests probability of
being selected, as EMU would not effect the choice between car
and PT.
If >1, an increase in the utility of an alternative in the nest would tend to
increase not only its selection probability but also those of the rest of the
options in the nest.
If = 1, the model becomes equivalent to MNL.
Variable Selection Process
Decision
Sign
Correct sign
Significance
Policy
Other
Significant
Include
Include
Not significant
Include
May reject
Big problem
Reject
Problem
Reject
Significant
Wrong sign
Not significant