Estimation of Random Utility Models in R: The

mlogit Package

Yves Croissant
Université de la Réunion

Abstract
mlogit is a package for R which enables the estimation of random utility models
with individual and/or alternative specific variables. The main extensions of the
basic multinomial model (heteroscedastic, nested and random parameter models)
are implemented.

Keywords: discrete choice models, maximum likelihood estimation, R, econometrics.

Introduction
Random utility models is the reference approach in economics when one wants to analyze
the choice by a decision maker of one among a set of mutually exclusive alternatives.
Since the seminal works of Daniel Mc Fadden (McFadden 1974) who won the Nobel prize
in economics “for his development of theory and methods for analyzing discrete choice”,
a large amount of theoretical and empirical literature have been developed in this field.1
These models rely on the hypothesis that the decision maker is able to rank the different
alternatives by an order of preference represented by a utility function. For a choice of
an alternative l among a set of J alternatives, each alternative is therefore characterized
by an utility value Ul and the alternative l is chosen if and only if Ul > Uj ∀ j 6= l.
These models are called random utility models because the researcher is unable to mea-
sure the whole level of utility, but only part of it. Therefore, the utility for alternative l is
written as: Ul = Vl +l where Vl is a function of some observable covariates and unknown
parameters to be estimated, and l a random deviate which contains all the unobserved
determinants of the utility. Alternative l is therefore chosen if j < (Vl − Vj ) + l ∀ j 6= l
and this choice can be written in probabilistic terms:

P(1 < Vl − V1 + l , 2 < Vl − V2 + l , . . . , J < Vl − VJ + l )

Different hypothesis on the distribution of  lead to different flavors of random utility

1
For an extensive presentation of this literature, see Train (2003) ; the theoretical parts of this paper
draw heavily on Kenneth Train’s book.
2 Estimation of Random Utility Models in R: The mlogit Package

models. Early developments of these models were based on the hypothesis of identically
and independent errors following a Gumbell distribution.2 Much more general models
have since been proposed, based on much less restrictive distribution hypothesis, and
often estimated using simulations.
The first version of mlogit was posted in 2008, it was the first R package allowing the
estimation of random utility models. Since then, other package have emerged (see Sarrias
and Daziano 2017, page 4 for a survey of revelant R pakages). mlogit still provides the
widests set of estimators for random utility models and, moreover, its syntax has been
adopted by other R packages, especially by gmnl (Sarrias and Daziano 2017) and mnlogit
(Hasan, Wang, and Mahani 2016) which, respectively, implements advanced mixed logit
models and estimates efficiently multinomial logit models on large data sets.
The article is organized as follow. Section 1 explains how the usual formula-data and
testing interface can be extended in order to describes in a very natural way the model
to be estimated. Section 2 describe the landmark multinomial logit model. Section 3
and 4 present two important extensions of this basic model: section 3 presents models
that relax the iid Gumbell hypothesis and section 4 introduces slope heterogeneity by
considering some parameters as random. Section 5 concludes.

1. Data management, model description and testing

The formula-data interface is a critical advantage of the R software. It provides a prac-
tical way to describe the model to be estimated and to store data. However, the usual
interface is not flexible enough to deal correctly with random utility models. Therefore,
mlogit introduce tools to construct richer data.frames and formulas.

1.1. Data management

mlogit is loaded using:

library("mlogit")

It comes with several data sets that we’ll use to illustrate the features of the library.
Data sets used for multinomial logit estimation concern some individuals, that make one
or a sequential choice of one alternative among a set of mutually exclusive alternatives.
The determinants of these choices are covariates that can depends on the alternative and
the choice situation, only on the alternative or only on the choice situation.
Such data have therefore a specific structure that can be characterized by three indexes:
the alternative, the choice situation and the individual. These three indexes will be
denoted alt, chid and id. Note that the distinction between chid and id is only
relevant if we have repeated observations for the same individual.
2
This distribution has the distinctive advantage that it leads to a probability which can be written
has an integral which has a closed form.
 Yves Croissant 3

Data sets can have two different shapes: a wide shape (one row for each choice situation)
or a long shape (one row for each alternative and, therefore, as many rows as there are
alternatives for each choice situation).
mlogit deals with both format. It provides a mlogit.data function that take as first argu-
ment a data.frame and returns a data.frame in “long” format with some supplementary
information about the structure of the data.

Wide format
Train3 is an example of a wide data set:

data("Train", package = "mlogit")

head(Train, 3)

## id choiceid choice price_A time_A change_A comfort_A price_B

## 1 1 1 A 2400 150 0 1 4000
## 2 1 2 A 2400 150 0 1 3200
## 3 1 3 A 2400 115 0 1 4000
## time_B change_B comfort_B
## 1 150 0 1
## 2 130 0 1
## 3 115 0 0

This data set contains data about a stated preference survey in Netherlands. Each
individual has responded to several (up to 16) scenarios. For every scenario, two train
trips are proposed to the user, with different combinations of four attributes: price (the
price in cents of guilders), time (travel time in minutes), change (the number of changes)
and comfort (the class of comfort, 0, 1 or 2, 0 being the most comfortable class).
This “wide” format is suitable to store individual specific variables. Otherwise, it is
cumbersome for alternative specific variables because there are as many columns for
such variables that there are alternatives.
For such a wide data set, the shape argument of mlogit.data is mandatory, as it default
value is "long". The alternative specific variables are indicated with the varying ar-
gument which is a numeric vector that indicates their position in the data frame. This
argument is then passed to reshape that coerced the original data.frame in “long” for-
mat. Further arguments may be passed to reshape. For example, as the names of the
variables are of the form price_A, one must add sep = ’_’ (the default value being ".").
The choice argument is also mandatory because the response has to be transformed in
a logical value in the long format. To take the panel dimension into account, one has to
add an argument id.var which is the name of the individual index.

3
used by Ben-Akiva, Bolduc, and Bradley (1993) and Meijer and Rouwendal (2006).
4 Estimation of Random Utility Models in R: The mlogit Package

Tr <- mlogit.data(Train, shape = "wide", choice = "choice",

varying = 4:11, sep = "_",
alt.levels = c("A", "B"), id.var = "id",
opposite = c("price", "comfort", "time", "change"))

Note the use of the opposite argument for the 4 covariates: we expect negative coeffi-
cients for all of these, taking the opposite of the covariates will lead to expected positive
coefficients. We next convert price and time in more meaningful unities, hours and euros
(1 guilder was 2.20371 euros):

Tr$price <- Tr$price / 100 * 2.20371

Tr$time <- Tr$time / 60

head(Tr, 3)

## id choiceid choice alt price time change comfort chid

## 1.A 1 1 TRUE A -52.88904 -2.5 0 -1 1
## 1.B 1 1 FALSE B -88.14840 -2.5 0 -1 1
## 2.A 1 2 TRUE A -52.88904 -2.5 0 -1 2

An index attribute is added to the data, which contains the two relevant indexes: chid
is the choice situation index and alt the alternative index. As a id.var is provided,
the index contains a third column, the individual index. This attribute is a data.frame
that can be extracted using the index function, which returns it:this data.frame.

head(index(Tr), 3)

## chid alt id
## 1.A 1 A 1
## 1.B 1 B 1
## 2.A 2 A 1

Long format
ModeCanada,4 is an example of a data set in long format. It presents the choice of
individuals for a transport mode for the Montreal-Toronto corridor:

4
Used in particular by Forinash and Koppleman (1993), Bhat (1995), Koppelman and Wen (1998)
and Koppelman and Wen (2000).
 Yves Croissant 5

data("ModeCanada", package = "mlogit")

head(ModeCanada)

## case alt choice dist cost ivt ovt freq income urban noalt
## 1 1 train 0 83 28.25 50 66 4 45 0 2
## 2 1 car 1 83 15.77 61 0 0 45 0 2
## 3 2 train 0 83 28.25 50 66 4 25 0 2
## 4 2 car 1 83 15.77 61 0 0 25 0 2
## 5 3 train 0 83 28.25 50 66 4 70 0 2
## 6 3 car 1 83 15.77 61 0 0 70 0 2

There are four transport modes (air, train, bus and car) and most of the variable are
alternative specific (cost for monetary cost, ivt for in vehicle time, ovt for out of vehicle
time, freq for frequency). The only individual specific variables are dist (the distance
of the trip), income (household income), urban (a dummy for trips which have a large
city at the origin or the destination) and noalt the number of available alternatives. The
advantage of this shape is that there are much fewer columns than in the wide format,
the caveat being that values of dist, income and urban are repeated four times.
For data in “long” format, the shape and the choice arguments are no more mandatory.
To replicate published results (latter in the text), we’ll use only a subset of the choice
situations, namely those for which the 4 alternatives are available. This can be done
using the subset function with the subset argument set to noalt == 4. This can also
be done within mlogit.data, using the subset argument.
The information about the structure of the data can be explicitly indicated using choice
situations and alternative indexes (respectively case and alt in this data set) or, in
part, guessed by the mlogit.data function. Here, after subsetting, we have 2779 choice
situations with 4 alternatives, and the rows are ordered first by choice situation and then
by alternative (train, air, bus and car in this order).
The first way to read correctly this data frame is to ignore completely the two index
variables. In this case, the only supplementary argument to provide is the alt.levels
argument which is a character vector that contains the name of the alternatives in their
order of appearance:

MC <- mlogit.data(ModeCanada, subset = noalt == 4,

alt.levels = c("train", "air", "bus", "car"))

Note that this can only be used if the data set is “balanced”, which means than the same
set of alternatives is available for all choice situations. It is also possible to provide an
argument alt.var which indicates the name of the variable that contains the alternatives
6 Estimation of Random Utility Models in R: The mlogit Package

MC <- mlogit.data(ModeCanada , subset = noalt == 4,

alt.var = "alt")

The name of the variable that contains the information about the choice situations can
be indicated using the chid.var argument:

MC <- mlogit.data(ModeCanada, subset = noalt == 4,

chid.var = "case",
alt.levels = c("train", "air", "bus", "car"))

Both alternative and choice situation variable can also be provided:

MC <- mlogit.data(ModeCanada, subset = noalt == 4,

chid.var = "case", alt.var = "alt")

and dropped from the data frame using the drop.index argument:

MC <- mlogit.data(ModeCanada, subset = noalt == 4,

chid.var = "case", alt.var = "alt",
drop.index = TRUE)
head(MC)

## choice dist cost ivt ovt freq income urban noalt

## 109.train 0 377 58.25 215 74 4 45 0 4
## 109.air 1 377 142.80 56 85 9 45 0 4
## 109.bus 0 377 27.52 301 63 8 45 0 4
## 109.car 0 377 71.63 262 0 0 45 0 4
## 110.train 0 377 58.25 215 74 4 70 0 4
## 110.air 1 377 142.80 56 85 9 70 0 4

1.2. Model description

Standard formulas are not very practical to describe random utility models, as these
models may use different sets of covariates. Working with random utility models, one
has to consider at most five sets of covariates:

• alternative and individual specific covariates xij with generic coefficients β,

• individual specific covariates zi with alternative specific coefficients γj ,

• alternative and individual specific covariates wij with alternative specific coeffi-
cients δj ,
 Yves Croissant 7

• alternative specific covariates tj with a generic coefficient ν,

• individual specific covariates vi that influence the variance of the errors.

Ignoring for the moment the 5th set of covariates, the observable part of the utility index
for alternative j is:

Vij = αj + βxij + γj zi + δj wij + νtj

As the absolute value of utility is irrelevant, only utility differences are useful to modelize
the choice for one alternative. For two alternatives j and k, we obtain:

Vij − Vik = (αj − αk ) + β(xij − xik ) + (γj − γk )zi + (δj wij − δk wik ) + ν(tj − tk )

It is clear from the previous expression that coefficients of individual specific variables
(the intercept being one of those) should be alternative specific, otherwise they would
disappear in the differentiation. Moreover, only differences of these coefficients are rel-
evant and may be identified. For example, with three alternatives 1, 2 and 3, the three
coefficients γ1 , γ2 , γ3 associated to an individual specific variable cannot be identified,
but only two linear combinations of them. Therefore, one has to make a choice of nor-
malization and the simplest one is just to set γ1 = 0.
Coefficients for alternative and individual specific variables may (or may not) be alter-
native specific. For example, transport time is alternative specific, but 10 mn in public
transport may not have the same impact on utility than 10 mn in a car. In this case,
alternative specific coefficients are relevant. Monetary cost is also alternative specific,
but in this case, one can consider than 1$ is 1$ whatever it is spent for the use of a car
or in public transports. In this case, a generic coefficient is relevant.
The treatment of alternative specific variables don’t differ much from the alternative and
individual specific variables with a generic coefficient. However, if some of these variables
are introduced, the ν parameter can only be estimated in a model without intercepts to
avoid perfect multicolinearity.
Individual-related heteroscedasticity (see Swait and Louviere 1993) can be addressed by
writing the utility of choosing j for individual i: Uij = Vij + σi ij , where  has a variance
that doesn’t depend on i and j and σi is a parametric function of some individual-specific
covariates. As the overall scale of utility is irrelevant, the utility can also be writen as:
Uij∗ = Uij /σi = Vij /σi + ij , i.e. with homoscedastic errors. if Vij is a linear combination
of covariates, the associated coefficients are then divided by σi .
A logit model with only individual specific variables is sometimes called a multinomial
logit model, one with only alternative specific variables a conditional logit model and one
with both kind of variables a mixed logit model. This is seriously misleading: conditional
logit model is also a logit model for longitudinal data in the statistical literature and
mixed logit is one of the names of a logit model with random parameters. Therefore,
8 Estimation of Random Utility Models in R: The mlogit Package

in what follows, we’ll use the name multinomial logit model for the model we’ve just
described whatever the nature of the explanatory variables used.
mlogit package provides objects of class mFormula which are built upon Formula objects
provided by the Formula package.5 The Formula package provides richer formulas,
which accept multiple responses (a feature not used here) and multiple set of covariates.
It has in particular specific model.frame and model.matrix methods which can be used
with one or several sets of covariates.
To illustrate the use of mFormula objects, let’s use again the ModeCanada data set and
consider three sets of covariates that will be indicated in a three-part formula:

• cost (monetary cost) is an alternative specific covariate with a generic coefficient

(part 1),
• income and urban are individual specific covariates (part 2),
• ivt (in vehicle travel time) is alternative specific and alternative specific coefficients
are expected (part 3).

f <- mFormula(choice ~ cost | income + urban | ivt)

Some parts of the formula may be omitted when there is no ambiguity. For example,
the following sets of formulas are identical:

f2 <- mFormula(choice ~ cost + ivt | income + urban)

f2 <- mFormula(choice ~ cost + ivt | income + urban | 0)

f3 <- mFormula(choice ~ 0 | income | 0)

f3 <- mFormula(choice ~ 0 | income)

f4 <- mFormula(choice ~ cost + ivt)

f4 <- mFormula(choice ~ cost + ivt | 1)
f4 <- mFormula(choice ~ cost + ivt | 1 | 0)

By default, an intercept is added to the model, it can be removed by using + 0 or - 1

in the second part.

f5 <- mFormula(choice ~ cost | income + 0 | ivt)

f5 <- mFormula(choice ~ cost | income - 1 | ivt)

model.frame and model.matrix methods are provided for mFormula objects. The latter
is of particular interest, as illustrated in the following example:

5
See Zeileis and Croissant (2010) for a description of the Formula package.
 Yves Croissant 9

f <- mFormula(choice ~ cost | income | ivt)

head(model.matrix(f, MC), 4)

## air:(intercept) bus:(intercept) car:(intercept) cost

## 109.train 0 0 0 58.25
## 109.air 1 0 0 142.80
## 109.bus 0 1 0 27.52
## 109.car 0 0 1 71.63
## air:income bus:income car:income train:ivt air:ivt
## 109.train 0 0 0 215 0
## 109.air 45 0 0 0 56
## 109.bus 0 45 0 0 0
## 109.car 0 0 45 0 0
## bus:ivt car:ivt
## 109.train 0 0
## 109.air 0 0
## 109.bus 301 0
## 109.car 0 262

The model matrix contains J − 1 columns for every individual specific variables (income
and the intercept), which means that the coefficient associated to the first alternative
(air) is set to 0. It contains only one column for cost because we want a generic coefficient
for this variable. It contains J columns for ivt, because it is an alternative specific variable
for which we want alternative specific coefficients.

1.3. Testing
As for all models estimated by maximum likelihood, three testing procedures may be
applied to test hypothesis about models fitted using mlogit. The set of hypothesis tested
defines two models: the unconstrained model that doesn’t take these hypothesis into
account and the constrained model that impose these hypothesis.
This in turns define three principles of tests: the Wald test, based only on the uncon-
strained model, the Lagrange multiplier test (or score test), based only on the constrained
model and the likelihood ratio test, based on the comparison of both models.
Two of these three tests are implemented in the lmtest package (Zeileis and Hothorn
2002): waldtest and lrtest. The Wald test is also implemented as linearHypothesis in package
car (Fox and Weisberg 2010), with a fairly different syntax. We provide special methods
of waldtest and lrtest for mlogit objects and we also provide a function for the
Lagrange multiplier (or score) test called scoretest.
We’ll see later that the score test is especially useful for mlogit objects when one is
interested in extending the basic multinomial logit model because, in this case, the
unconstrained model may be difficult to estimate. For the presentation of further tests,
we provide a convenient statpval function which extract the statistic and the p-value
10 Estimation of Random Utility Models in R: The mlogit Package

from the objects returned by the testing function, which can be either of class anova or
htest.

statpval <- function(x){

if (inherits(x, "anova"))
result <- as.matrix(x)[2, c("Chisq", "Pr(>Chisq)")]
if (inherits(x, "htest")) result <- c(x$statistic, x$p.value)
names(result) <- c("stat", "p-value")
round(result, 3)
}

2. Random utility model and the multinomial logit model

2.1. Random utility model

Remind from the introduction that, in a random utility model, the probability of choosing
an alternative l among a set of mutually exclusive ones is:

P(1 < Vl − V1 + l , 2 < Vl − V2 + l , . . . , K < Vl − VJ + l )

Denoting F−l the cumulative density function of all the s except l , this probability is:

(Pl | l ) = F−l (Vl − V1 + l , . . . , Vl − VJ + l ) (1)

Note that this probability is conditional on the value of l . The unconditional probability
(which depends only on β and on the value of the observed explanatory variables) is
obtained by integrating out the conditional probability using the marginal density of l ,
denoted fl :
Z
Pl = F−l (Vl − V1 + l , . . . , Vl − VJ ) + l )fl (l )dl (2)

The conditional probability is an integral of dimension J − 1 and the computation of the

unconditional probability adds on more dimension of integration.

2.2. The distribution of the error terms

The multinomial logit model (McFadden 1974) is a special case of the model developed
in the previous section. It is based on three hypothesis
The first hypothesis is the independence of the errors. In this case, the univariate
distribution of the errors can be used, which leads to the following conditional and
unconditional probabilities:
 Yves Croissant 11

Y Z Y
(Pl | l ) = Fj (Vl − Vj + l ) and Pl = Fj (Vl − Vj + l ) fl (l ) dl (3)
j6=l j6=l

which means that the conditional probability is the product of J −1 univariate cumulative
density functions and the evaluation of only a one-dimensional integral is required to
compute the unconditional probability.
The second hypothesis is that each  follows a Gumbel distribution, whose density and
probability functions are respectively:

z−µ Z +∞ z−µ
1 z−µ − −
f (z) = e− θ e−e θ and F (z) = f (t)dt = e−e θ
(4)
θ −∞

where µ is the location parameter and θ the scale parameter. The first two moments of
the Gumbel distribution are E(z) = µ + θγ, where γ is the Euler-Mascheroni constant
2
(≈ 0.577) and V(z) = π6 θ2 . The mean of j s is not identified if Vj contains an intercept.
We can then, without loss of generality suppose that µj = 0, ∀j. Moreover, the overall
scale of utility is not identified. Therefore, only J − 1 scale parameters may be identified,
and a natural choice of normalization is to impose that one of the θj is equal to 1.
The last hypothesis is that the errors are identically distributed. As the location param-
eter is not identified for any error term, this hypothesis is essentially an homoscedasticity
hypothesis, which means that the scale parameter of the Gumbel distribution is the same
for all the alternatives. As one of them has been previously set to 1, we can therefore
suppose that, without loss of generality, θj = 1, ∀j ∈ 1 . . . J. The conditional and
unconditional probabilities (4) then further simplify to:
Z +∞ Y
−e−(Vl −V j+l ) −(Vl −V j+t) −t
e−e e−t e−e dt
Y
(Pl | l ) = e and Pl = (5)
j6=l −∞ j6=l

The probabilities have then very simple, closed forms, which correspond to the logit
transformation of the deterministic part of the utility.

eVl
Pl = PJ Vj
(6)
j=1 e

2.3. IIA property

Vl
If we consider the probabilities of choice for two alternatives l and m, we have Pl = Pe
j
eVj
eVm
and Pm = P . The ratio of these two probabilities is:
j
eVj

Pl eVl
= Vm = eVl −Vm
Pm e
12 Estimation of Random Utility Models in R: The mlogit Package

This probability ratio for the two alternatives depends only on the characteristics of
these two alternatives and not on those of other alternatives. This is called the IIA
property (for independence of irrelevant alternatives).
Consider for example inter urban trips between two towns, with 3 modes available. The
initial situation is presented in table 1, which indicates that the initial market shares
are 30% for car, 10% for plane and 60% for train. Consider now that a low-cost airline
company enters the market, so that the price of plane ticket decreases from 150 to 100
euros.

price time share s1 s2

car 50 4 30% 29% 25%
plane 150 1 10% 25% 25%
train 80 2 60% 46% 50%

Table 1: Market shares

We consider in the table two scenarios of evolution of the market shares (columns s1 and
s2). Both indicates that the market share of the plane increases and that the market
shares of the other two modes decrease. For the first scenario (column s1), almost all the
increase of plane market share is due to former train users who now take the plane. The
car market share decreases very slightly (from 30 to 29%). This situation may be realistic
if car trips are mainly family trips, as train and plane trips mainly trips for professional
purpose. This situation cannot be predicted by a multinomial logit model, because of the
IIA property; the ratio of the probabilities of choosing train and car should be identical
before and after the decrease of plane price ticket. This is the case in column s2, where
the relative decrease of the market shares for car and train are identical.
IIA relies on the hypothesis that the errors are identical and independent. It is not
a problem by itself and may even be considered as a useful feature for a well specified
model. However, this hypothesis may be in practice violated, especially if some important
variables are omitted.

2.4. Interpretation
In a linear model, the coefficients can be directly considered as marginal effects of the
explanatory variables on the explained variable. This is not the case for the multinomial
logit model. However, meaningful results can be obtained using relevant transformations
of the coefficients.

Marginal effects
The marginal effects are the derivatives of the probabilities with respect to the covariates,
 Yves Croissant 13

which can be be individual-specific (zi ) or alternative specific (xij ):

 
∂Pil X
= Pil βl − Pij βj 
∂zi j


∂Pil
= γPil (1 − Pil )



∂xil
∂Pil
= −γPil Pik



∂xik

• For an individual specific variable, the sign of the marginal effect is not necessarily
the
 sign of the coefficient. Actually, the sign of the marginal effect is given by
βl − j Pij βj , which is positive if the coefficient for alternative l is greater than
P

a weighted average of the coefficients for all the alternatives, the weights being
the probabilities of choosing the alternatives. In this case, the sign of the marginal
effect can be established with no ambiguity only for the alternatives with the lowest
and the greatest coefficients.

• For an alternative-specific variable, the sign of the coefficient can be directly in-
terpreted. The marginal effect is obtained by multiplying the coefficient by the
product of two probabilities which is at most 0.25. The rule of thumb is therefore
to divide the coefficient by 4 in order to have an upper bound of the marginal
effect.

Note that the last equation can be rewriten: dPdilx/Pil = −γPik .

ik

Therefore, when a characteristic of alternative k changes, the relative change of the

probabilities for every alternatives except k is the same, which is a consequence of the
IIA property.

Marginal rates of substitution

Coefficients are marginal utilities, which cannot be interpreted. However, ratios of coef-
ficients are marginal rates of substitution. For example, if the observable part of utility
is: V = βo + β1 x1 + βx2 + βx3 , join variations of x1 and x2 which ensure the same level
of utility are such that: dV = β1 dx1 + β2 dx2 = 0 so that:

dx2 β1
− |dV =0 =
dx1 β2

For example, if x2 is transport cost (in $), x1 transport time (in hours), β1 = 1.5 and
β2 = 0.2, ββ21 = 30 is the marginal rate of substitution of time in terms of $ and the value
of 30 means that to reduce the travel time of one hour, the individual is willing to pay
at most 30$ more. Stated more simply, time value is 30$ per hour.
14 Estimation of Random Utility Models in R: The mlogit Package

Consumer’s surplus
Consumer’s surplus has a very simple expression for multinomial logit models, which was
first derived by Small and Rosen (1981). The level of utility attained by an individual
is Uj = Vj + j , j being the alternative chosen. The expected utility, from the searcher’s
point of view is then: E(maxj Uj ), where the expectation is taken over the values of all
the error terms. Its expression is simply, up to an additive unknown constant, the log of
the denominator of the logit probabilities and is often called the “log-sum”:

J
X
E(U ) = ln eVj + C
j=1

If the marginal utility of income (α) is known and constant, the expected surplus is
simply E(U ) .
α

2.5. Application
Random utility models are fitted using the mlogit function. Basically, only two argu-
ments are mandatory, formula and data, if a mlogit.data object (and not an ordinary
data.frame) is provided.

ModeCanada
We first use the ModeCanada data set, which was already coerced to a mlogit.data
object (called MC) in the previous section. The same model can then be estimated using
as data argument a mlogit.data object:

ml.MC1 <- mlogit(choice ~ cost + freq + ovt | income | ivt, MC)

or a data.frame. In this latter case, further arguments that will be passed to mlogit.data
should be indicated:

ml.MC1b <- mlogit(choice ~ cost + freq + ovt | income | ivt, ModeCanada,

subset = noalt == 4, alt.var = "alt", chid.var = "case")

mlogit provides two further usefull arguments:

• reflevel indicates which alternative is the “reference” alternative, i.e. the one for
which the coefficients of individual specific covariates are 0,

• alt.subset indicates a subset of alternatives on which the estimation has to be

performed; in this case, only the lines that correspond to the selected alternatives
are used and all the choice situations where not selected alternatives has been
chosen are removed.
 Yves Croissant 15

We estimate the model on the subset of three alternatives (we exclude bus whose market
share is negligible in our sample) and we set car as the reference alternative. Moreover,
we use a total transport time variable computed as the sum of the in and the out of
vehicule time variables.

MC$time <- with(MC, ivt + ovt)

ml.MC1 <- mlogit(choice ~ cost + freq | income | time, MC,
alt.subset = c("car", "train", "air"),
reflevel = "car")

The main results of the model are computed and displayed using the summary method:

summary(ml.MC1)

##
## Call:
## mlogit(formula = choice ~ cost + freq | income | time, data = MC,
## alt.subset = c("car", "train", "air"), reflevel = "car",
## method = "nr", print.level = 0)
##
## Frequencies of alternatives:
## car train air
## 0.45757 0.16721 0.37523
##
## nr method
## 6 iterations, 0h:0m:0s
## g'(-H)^-1g = 6.94E-06
## successive function values within tolerance limits
##
## Coefficients :
## Estimate Std. Error z-value Pr(>|z|)
## train:(intercept) -0.97034440 0.26513065 -3.6599 0.0002523 ***
## air:(intercept) -1.89856552 0.68414300 -2.7751 0.0055185 **
## cost -0.02849715 0.00655909 -4.3447 1.395e-05 ***
## freq 0.07402902 0.00473270 15.6420 < 2.2e-16 ***
## train:income -0.00646892 0.00310366 -2.0843 0.0371342 *
## air:income 0.02824632 0.00365435 7.7295 1.088e-14 ***
## car:time -0.01402405 0.00138047 -10.1589 < 2.2e-16 ***
## train:time -0.01096877 0.00081834 -13.4036 < 2.2e-16 ***
## air:time -0.01755120 0.00399181 -4.3968 1.099e-05 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
16 Estimation of Random Utility Models in R: The mlogit Package

## Log-Likelihood: -1951.3
## McFadden R^2: 0.31221
## Likelihood ratio test : chisq = 1771.6 (p.value = < 2.22e-16)

The frequencies of the different alternatives in the sample are first indicated. Next, some
information about the optimization are displayed: the Newton-Ralphson method (with
analytic gradient and hessian) is used, as it is the most efficient method for this simple
model for which the log-likelihood function is concave. Note that very few iterations
and computing time are required to estimate this model. Follows the usual table of
coefficients and some goodness of fit measures: the value of the log-likelihood function,
which is compared to the value when only intercepts are introduced, which leads to the
computation of the McFadden R2 and to the likelihood ratio test.
The fitted method can be used either to obtain the probability of actual choices (type
= "outcome") or the probabilities for all the alternatives (type = "probabilities").

head(fitted(ml.MC1, type = "outcome"))

## 109 110 111 112 113 114

## 0.1909475 0.3399941 0.1470527 0.3399941 0.3399941 0.2440011

head(fitted(ml.MC1, type = "probabilities"), 4)

## car train air

## 109 0.4206404 0.3884120 0.1909475
## 110 0.3696476 0.2903582 0.3399941
## 111 0.4296769 0.4232704 0.1470527
## 112 0.3696476 0.2903582 0.3399941

Note that the log-likelihood is the sum of the log of the fitted outcome probabilities
and that, as the model contains intercepts, the average fitted probabilities for every
alternative equals the market shares of the alternatives in the sample.

sum(log(fitted(ml.MC1, type = "outcome")))

## [1] -1951.344

logLik(ml.MC1)

## 'log Lik.' -1951.344 (df=9)

apply(fitted(ml.MC1, type = "probabilities"), 2, mean)

## car train air

## 0.4575659 0.1672084 0.3752257
 Yves Croissant 17

Predictions can be made using the predict method. If no data is provided, predictions
are made for the average of the sample on which the estimation as been performed.

predict(ml.MC1)

## car train air

## 0.5066362 0.2116876 0.2816761

Assume, for example, that we wish to predict the effect of a reduction of train transport
time of 20%. We first create a new data.frame simply by multiplying train transport
time by 0.8 and then using the predict method with this new data.frame

NMC <- MC
NMC[index(NMC)$alt == "train", "time"] <-
0.8 * NMC[index(NMC)$alt == "train", "time"]
Oprob <- fitted(ml.MC1, type = "probabilities")
Nprob <- predict(ml.MC1, newdata = NMC)
rbind(old = apply(Oprob, 2, mean),
new = apply(Nprob, 2, mean))

## car train air

## old 0.4575659 0.1672084 0.3752257
## new 0.4044736 0.2635801 0.3319462

If, for the first individuals in the sample, we compute the ratio of the probabilities of
the air and the car mode, we obtain:

head(Nprob[, "air"] / Nprob[, "car"])

## 109 110 111 112 113 114

## 0.4539448 0.9197791 0.3422401 0.9197791 0.9197791 0.6021092

head(Oprob[, "air"] / Oprob[, "car"])

## 109 110 111 112 113 114

## 0.4539448 0.9197791 0.3422401 0.9197791 0.9197791 0.6021092

which is an illustration of the IIA property. If train time changes, it changes the proba-
bilities of choosing air and car, but not their ratio.
We next compute the surplus for individuals of the sample induced by train time reduc-
tion. This requires the computation of the log-sum term (also called inclusive value or
inclusive utility) for every choice situation, which writes:
18 Estimation of Random Utility Models in R: The mlogit Package

J
X >x
ivi = ln eβ ij

j=1

For this purpose, we use the logsum function, which works on a vector of coefficients
and a model.matrix. The basic use of logsum consists on providing as unique argument
(called coef) a mlogit object. In this case, the model.matrix and the coef are extracted
from the same model.

ivbefore <- logsum(ml.MC1)

To compute the log-sum after train time reduction, we must provide a model.matrix
which is not the one corresponding to the fitted model. This can be done using the X
argument which is a matrix or an object from which a model.matrix can be extracted.
This can also be done by filling the data (a data.frame or an object from which a
data.frame can be extracted using a model.frame method), and eventually the formula
argument (a formula or an object for which the formula method can be applied). If no
formula is provided but if data is a mlogit.data object, the formula is extracted from
it.

ivafter <- logsum(ml.MC1, data = NMC)

Surplus variation is then computed as the difference of the log-sums divided by the
opposite of the cost coefficient which can be interpreted as the marginal utility of income:

surplus <- - (ivafter - ivbefore) / coef(ml.MC1)["cost"]

summary(surplus)

## Min. 1st Qu. Median Mean 3rd Qu. Max.

## 0.5852 2.8439 3.8998 4.6971 5.8437 31.3912

Consumer’s surplus variation range from 0.6 to 31 Canadian $, with a median value of
about 4$.
Marginal effects are computed using the effects method. By default, they are computed at
the sample mean, but a data argument can be provided. The variation of the probability
and of the covariate can be either absolute or relative. This is indicated with the type
argument which is a combination of two a (as absolute) and r (as relative). For example,
type = "ar" means that what is measured is an absolute variation of the probability
for a relative variation of the covariate.

effects(ml.MC1, covariate = "income", type = "ar")

## car train air

## -0.1822177 -0.1509079 0.3331256
 Yves Croissant 19

The results indicate that, if income has doubled, the probability of choosing air increases
by 33 points of percentage, as the probabilities of choosing car and train decrease by
18 and 15 points of percentage.
For an alternative specific covariate, a matrix of marginal effects is displayed.

effects(ml.MC1, covariate = "cost", type = "rr")

## car train air

## car -0.9131273 0.9376923 0.9376923
## train 0.3358005 -1.2505014 0.3358005
## air 1.2316679 1.2316679 -3.1409703

The cell in the lth row and the cth column indicates the change of the probability of
choosing alternative c when the cost of alternative l changes. As type = "rr", elasticities
are computed. For example, a 10% change of train cost increases the probabilities of
choosing car and air by 3.36%. Note that the relative changes of the probabilities of
choosing one of these two modes are equal, which is a consequence of the IIA property.
Finally, in order to compute travel time valuation, we divide the coefficients of travel
times (in minutes) by the coefficient of monetary cost (in $).

coef(ml.MC1)[grep("time", names(coef(ml.MC1)))] /
coef(ml.MC1)["cost"] * 60

## car:time train:time air:time

## 29.52728 23.09447 36.95360

The value of travel time ranges from 23 for train to 37 Canadian $ per hour for plane.

NOx
The second example is a data set used by Fowlie (2010), called NOx. She analyzed the
effect of an emissions trading program (the NOx budget program which seeks to reduce
the emission of nitrogen oxides) on the behavior of producers. More precisely, coal
electricity plant managers may adopt one out of fifteen different technologies in order to
comply to the emissions defined by the program. Some of them require high investment
(the capital cost is kcost) and are very efficient to reduce emissions, some other require
much less investment but are less efficient and the operating cost (denoted vcost) is then
higher because pollution permits must be purchased to offset emissions exceeding their
allocation.
The focus of the paper is on the effects of the regulatory environment on manager’s
behavior. Some firms are deregulated, whereas other are either regulated or public. Rate
of returns is applied for regulated firms, which means that they perceive a “fair” rate of
return on their investment. Public firms also enjoy significant cost of capital advantages.
20 Estimation of Random Utility Models in R: The mlogit Package

Therefore, the main hypothesis of the paper is that public and regulated firms will
adopt much more capitalistic intensive technologies than deregulated and public ones,
which means that the coefficient of capital cost should take a higher negative value for
deregulated firms. Capital cost is interacted with the age of the plant (measured as a
deviation from the sample mean age), as firms should weight capital costs more heavily
for older plants, as they have less time to recover these costs.
Multinomial logit models are estimated for the three subsamples defined by the regula-
tory environment. The 15 technologies are not available for every plants, the sample is
therefore restricted to available technologies, using the available covariate. Three tech-
nology dummies are introduced: post for post-combustion polution control technology,
cm for combustion modification technology and lnb for low NOx burners technology.
A last model is estimated for the whole sample, but the parameters are allowed to be
proportional to each other. The scedasticity function is described in the fourth part of
the formula, it contains here only one covariate, env. Note also that for the pooling
model, the author has introduced a specific capital cost coefficient for deregulated firms6

data("NOx", package = "mlogit")

NOx$kdereg <- with(NOx, kcost * (env == "deregulated"))

NOxml <- mlogit.data(NOx, chid.var = "chid",

alt.var = "alt", id.var = "id")
ml.pub <- mlogit(choice ~ post + cm + lnb + vcost +
kcost + kcost:age | - 1,
subset = available & env == "public", data = NOxml)
ml.reg <- update(ml.pub, subset = available & env == "regulated")
ml.dereg <- update(ml.pub, subset = available & env == "deregulated")
ml.pool <- mlogit(choice ~ post + cm + lnb + vcost + kcost +
kcost:age + kdereg | - 1 | 0 | env,
subset = available == 1, data = NOxml, method = "bhhh")
library("texreg")
screenreg(list(Public = ml.pub, Deregulated = ml.dereg,
Regulated = ml.reg, pooled = ml.pool))

##
## =====================================================================
## Public Deregulated Regulated pooled
## ---------------------------------------------------------------------
## post -5.71 *** -1.50 *** -2.67 *** -2.31 ***
## (1.02) (0.22) (0.28) (0.21)
6
Note the use of the method argument, set to bhhh. mlogit use its own optimisation functions, but
borrows its syntax from package maxLik (Toomet and Henningsen 2010). The default method is bfgs,
except for the basic model, for which it is nr. As the default algorithm failed to converged, we use here
bhhh.
 Yves Croissant 21

## cm -4.43 *** -1.54 *** -1.91 ***

-2.06 ***
## (0.56) (0.19) (0.18)
(0.16)
## lnb -3.96 *** -1.55 *** -2.21 ***
-2.03 ***
## (0.59) (0.22) (0.22)
(0.17)
## vcost -1.56 *** -0.19 *** -0.28 ***
-0.31 ***
## (0.36) (0.06) (0.06)
(0.04)
## kcost 0.04 -0.06 ** 0.01
0.01
## (0.11) (0.02) (0.03)
(0.02)
## kcost:age -0.08 -0.04 ** -0.02 *
-0.02 ***
## (0.04) (0.01) (0.01)
(0.01)
## kdereg -0.07 ***
## (0.01)
## sig.envderegulated 0.32 **
## (0.12)
## sig.envpublic -0.33 ***
## (0.08)
## ---------------------------------------------------------------------
## AIC 168.92 690.15 731.48 1634.22
## Log Likelihood -78.46 -339.07 -359.74 -808.11
## Num. obs. 113 227 292 632
## =====================================================================
## *** p < 0.001, ** p < 0.01, * p < 0.05

Coefficients are very different on the different sub-samples defined by the regulatory en-
vironment. Note in particular that the capital cost coefficient is positive and insignificant
for public and regulated firms, as it is significantly negative for deregulated firms. Errors
seems to have significant larger variance for deregulated firms and lower ones for public
firms compared to regulated firms. The hypothesis that the coefficients (except the kcost
one) are identical up to a multiplicative scalar can be performed using a likelihood ratio
test:

stat <- 2 * (logLik(ml.dereg) + logLik(ml.reg) +

logLik(ml.pub) - logLik(ml.pool))
stat

## 'log Lik.' 61.6718 (df=6)

pchisq(stat, df = 9, lower.tail = FALSE)

## 'log Lik.' 6.377283e-10 (df=6)

The hypothesis is strongly rejected.

22 Estimation of Random Utility Models in R: The mlogit Package

3. Logit models relaxing the iid hypothesis

In the previous section, we assumed that the error terms were iid (identically and in-
dependently distributed), i.e. uncorrelated and homoscedastic. Extensions of the basic
multinomial logit model have been proposed by relaxing one of these two hypothesis
while maintaining the hypothesis of Gumbell distribution.

3.1. The heteroskedastic logit model

The heteroskedastic logit model was proposed by Bhat (1995). The probability that
Ul > Uj is:

(Vl −Vj +l )

−
θj
−e
P (j < Vl − Vj + l ) = e

which implies the following conditional and unconditional probabilities

(Vl −Vj +l )

−
θj
−e
Y
(Pl | l ) = e (7)
j6=l

(Vl −Vj +t)

 
Z +∞ Y − t
e−e
θj 1 − θt −e− θl
Pl =  e le dt
−∞ j6=l θl
 Vl −Vj −θl ln t  (8)
Z +∞ −
θj
P
− e  e−t dt
= e j6=l
0

There is no closed form for this integral but it can be efficiently computed using a Gauss
quadrature method, and more precisely the Gauss-Laguerre quadrature method.

3.2. The nested logit model

The nested logit model was first proposed by McFadden (1978). It is a generalization
of the multinomial logit model that is based on the idea that some alternatives may
be joined in several groups (called nests). The error terms may then present some
correlation in the same nest, whereas error terms of different nests are still uncorrelated.
Denoting m = 1 . . . M the nests and Bm the set of alternatives belonging to nest m, the
multivariate distribution of the error terms is:
  λm 
M
e−j /λm  
X X
exp −
  
m=1 j∈Bm

The marginal distributions of the s are still univariate extreme value, but there is now
some correlation within nests. 1 − λm is a measure of the correlation, i.e. λm = 1 implies
 Yves Croissant 23

no correlation. In the special case where λm = 1 ∀m, the errors are iid Gumbel errors
and the nested logit model reduce to the multinomial logit model. It can then be shown
that the probability of choosing alternative j that belongs to nest l is:
P λl −1
eVj /λl k∈Bl eVk /λl
Pj = P
λm
M
eVk /λm
P
m=1 k∈Bm

and that this model is a random utility model if all the λ parameters are in the 0 − 1
interval.7
Let us now write the deterministic part of the utility of alternative j as the sum of two
terms: the first one (Zj ) being specific to the alternative and the second one (Wl ) to the
nest it belongs to:

V j = Zj + Wl

We can then rewrite the probabilities as follow:

P λl
(Zj +Wl )/λl k∈Bl
e(Zk +Wl )/λl
Pj = P e × PM
λ m
e(Zk +Wl )/λl
P
k∈Bl m=1 k∈Bm
e(Zk +Wm )/λm
 P λl
Zj /λl eWl /λl k∈Bl
eZk /λl
= P e × PM
λ m
eZk /λl
P
k∈Bl m=1
eWm /λm k∈Bm
eZk /λm

Then denote Il = ln k∈Bl eZk /λl which is often called the log-sum, the inclusive value
P

or the inclusive utility.8 We then can write the probability of choosing alternative j as:

eZj /λl eWl +λl Il

Pj = P Zk /λl
× M Wm +λm Im
k∈Bl e
P
m=1 e

The first term Pj|l is the conditional probability of choosing alternative j if the nest l
is chosen. It is often referred as the lower model. The second term Pl is the marginal
probability of choosing the nest l and is referred as the upper model. Wm + λm Im can be
interpreted as the expected utility of choosing the best alternative in m, Wm being the
expected utility of choosing an alternative in this nest (whatever this alternative is) and
λm Im being the expected extra utility gained by being able to choose the best alternative
in the nest. The inclusive values link the two models. It is then straightforward to show
that IIA applies within nests, but not for two alternatives in different nests.
7
A slightly different version of the nested logit model (Daly 1987) is often used, but is not compatible
with the random utility maximization hypothesis. Its difference with the previous expression is that the
deterministic parts of the utility for each alternative is not divided by the nest elasticity. The differences
between the two versions have been discussed in Koppelman and Wen (1998), Heiss (2002) and Hensher
and Greene (2002).
8
We’ve already encountered this expression in section 2.4.3.
24 Estimation of Random Utility Models in R: The mlogit Package

A consistent but inefficient way of estimating the nested logit model is to estimate
separately its two components. The coefficients of the lower model are first estimated,
which enables the computation of the inclusive values Im . The coefficients of the upper
model are then estimated, using Im as covariates. Maximizing directly the likelihood
function of the nested model leads to a more efficient estimator.

3.3. Applications

ModeCanada
Bhat (1995) estimated the heteroscedastic logit model on the ModeCanada data set. Us-
ing mlogit, the heteroscedastic logit model is obtained by setting the heterosc argument
to TRUE:

ml.MC <- mlogit(choice ~ freq + cost + ivt + ovt | urban + income, MC,
reflevel = 'car', alt.subset = c("car", "train", "air"))
hl.MC <- mlogit(choice ~ freq + cost + ivt + ovt | urban + income, MC,
reflevel = 'car', alt.subset = c("car", "train", "air"),
heterosc = TRUE)
coef(summary(hl.MC))[11:12, ]

## Estimate Std. Error z-value Pr(>|z|)

## sp.train 1.2371829 0.1104610 11.200182 0.000000e+00
## sp.air 0.5403239 0.1118353 4.831425 1.355592e-06

The variance of the error terms of train and air are respectively higher and lower than
the variance of the error term of car (set to 1). Note that the z-values and p-values
of the output are not particularly meaningful, as the hypothesis that the coefficient is
zero (and not one) is tested. The homoscedascticity hypothesis can be tested using any
of the three tests. A particular convenient syntax is provided in this case. For the
likelihood ratio and the wald test, one can use only the fitted heteroscedastic model as
argument. In this case, it is guessed that the hypothesis that the user wants to test is
the homoscedasticity hypothesis.

lr.heter <- lrtest(hl.MC, ml.MC)

wd.heter <- waldtest(hl.MC, heterosc = FALSE)

or, more simply:

lrtest(hl.MC)
waldtest(hl.MC)

The Wald test can also be computed using the linearHypothesis function from the car
package :
 Yves Croissant 25

library("car")
lh.heter <- linearHypothesis(hl.MC, c('sp.air = 1', 'sp.train = 1'))

For the score test, we provide the constrained model as argument, which is the stan-
dard multinomial logit model and the supplementary argument which defines the un-
constrained model, which is in this case heterosc = TRUE.

sc.heter <- scoretest(ml.MC, heterosc = TRUE)

sapply(list(wald = wd.heter, lh = lh.heter,

score = sc.heter, lr = lr.heter), statpval)

## wald lh score lr
## stat 25.196 25.196 9.488 6.888
## p-value 0.000 0.000 0.009 0.032

The homoscedasticity hypothesis is strongly rejected using the Wald test, but only a the
1 and 5% level for, respectively, the score and the likelihood ratio tests.

JapaneseFDI
Head and Mayer (2004) analyzed the choice of one of the 57 European regions belonging
to 9 countries by Japenese firms to implement a new production unit.

data("JapaneseFDI", package = "mlogit")

jfdi <- mlogit.data(JapaneseFDI, chid.var = "firm", alt.var = "region",
group.var = "country")

Note that we’ve used an extra argument to mlogit.data called group.var which indicates
the grouping variable, which will be used later to define easily the nests. There are
two sets of covariates: the wage rate wage, the unemployment rate unemp, a dummy
indicating that the region is eligible to European funds elig and the area area are observed
at the regional level and are therefore relevant for the estimation of the lower model,
whereas the social cotisation rate scrate and the corporate tax rate ctaxrate are observed
at the country level and are therefore suitable for the upper model.
We first estimate a multinomial logit model:

ml.fdi <- mlogit(choice ~ log(wage) + unemp + elig + log(area) +

scrate + ctaxrate | 0, data = jfdi)

Note that, as the covariates are only alternative specific, the intercepts are not identified
and therefore have been removed. We next estimate the lower model, which analyses
26 Estimation of Random Utility Models in R: The mlogit Package

the choice of a region within a given country. Therefore, for each choice situation, we
estimate the choice of a region on the subset of regions of the country which has been
chosen. Moreover, observations concerning Portugal and Ireland are removed as these
two countries are mono-region.

lm.fdi <- mlogit(choice ~ log(wage) + unemp + elig + log(area) | 0,

data = jfdi, subset = country == choice.c &
! country %in% c("PT", "IE"))

We next use the fitted lower model in order to compute the inclusive value, at the country
level:
X >x
ivig = ln eβ ij

j∈Bg

where Bg is the set of regions for country g. When a grouping variable is provided in the
mlogit.data function, inclusive values are by default computed for every group g (global
inclusive values are obtained by setting the type argument to "global"). By default,
output is set to "chid" and the results is a vector (if type = "global") or a matrix (if
type = "region") with row number equal to the number of choice situations. If output
is set to "obs", a vector of length equal to the number of lines of the data in long format
is returned. The following code indicates different ways to use the logsum function:

lmformula <- formula(lm.fdi)

head(logsum(ml.fdi, data = jfdi, formula = lmformula, type = "group"), 2)

## BE DE ES FR IE IT NL
## 3 3.595818 5.415838 3.593702 5.153709 1.933707 5.051387 4.077845
## 4 4.113243 5.765190 4.445012 5.383095 1.960462 5.687569 4.490379
## PT UK
## 3 2.702028 4.900622
## 4 3.200124 5.378561

head(logsum(ml.fdi, data = jfdi, formula = lmformula, type = "global"))

## 3 4 5 7 8 9
## 6.736116 7.182139 7.121855 7.084245 7.133368 7.133368

head(logsum(ml.fdi, data = jfdi, formula = lmformula, output = "obs"))

## [1] 3.595818 3.595818 3.595818 3.595818 5.415838 5.415838

head(logsum(ml.fdi, data = jfdi, formula = lmformula,

type = "global", output = "obs"))

## [1] 6.736116 6.736116 6.736116 6.736116 6.736116 6.736116

 Yves Croissant 27

To add the inclusive values in the original data.frame, we use output = "obs" and the
type argument can be omitted as its default value is "group":

JapaneseFDI$iv <- logsum(lm.fdi, data = jfdi, formula = lmformula,

output = "obs")

We next select the relevant variables for the estimation of the upper model, select unique
lines in order to keep only one observation for every choice situation / country combina-
tion and finally we coerce the response (choice.c) to a logical for the chosen country.

JapaneseFDI.c <- subset(JapaneseFDI,

select = c("firm", "country", "choice.c",
"scrate", "ctaxrate", "iv"))
JapaneseFDI.c <- unique(JapaneseFDI.c)
JapaneseFDI.c$choice.c <- with(JapaneseFDI.c, choice.c == country)

Finally, we estimate the upper model, using the previously computed inclusive value as
a covariate.

jfdi.c <- mlogit.data(JapaneseFDI.c, choice = "choice.c",

alt.var = "country", chid.var = "firm",
shape = "long")
um.fdi <- mlogit(choice.c ~ scrate + ctaxrate + iv | 0, data = jfdi.c)

If one wants to obtain different iv coefficients for different countries, the iv covariate
should be introduced in the 3th part of the formula and the coefficients for the two
mono-region countries (Ireland and Portugal) should be set to 1, using the constPar
argument.

um2.fdi <- mlogit(choice.c ~ scrate + ctaxrate | 0 | iv, data = jfdi.c,

constPar = c("iv:PT" = 1, "iv:IE" = 1))

We next estimate the full-information maximum likelihood nested model. It is obtained

by adding a nests argument to the mlogit function. This should be a named list of
alternatives (here regions), the names being the nests (here the countries). More simply,
if a group variable has been indicated while using mlogit.data, nests can be a boolean.
Two flavors of nested models can be estimated, using the un.nest.el argument which is
a boolean. If TRUE, one imposes that the coefficient associated with the inclusive utility
is the same for every nest, which means that the degree of correlation inside each nest
is the same. If FALSE, a different coefficient is estimated for every nest.
28 Estimation of Random Utility Models in R: The mlogit Package

mult. logit lower model upper model upper model nested logit nested logit
log(wage) 0.47 1.21∗ 0.46 0.77∗∗
(0.25) (0.48) (0.25) (0.25)
unemp −8.90∗∗∗ −9.77∗∗∗ −7.62∗∗∗ −6.95∗∗
(1.69) (2.39) (1.60) (2.28)
elig −0.25 −0.89∗∗ −0.34 −0.18
(0.21) (0.33) (0.20) (0.18)
log(area) 0.31∗∗∗ 0.29∗∗∗ 0.29∗∗∗ 0.15∗∗
(0.05) (0.06) (0.05) (0.05)
scrate −2.26∗∗∗ −2.49∗∗∗ 0.15 −2.44∗∗∗ −0.88
(0.38) (0.38) (1.72) (0.38) (0.86)
ctaxrate −4.82∗∗∗ −3.69∗∗∗ −1.78 −4.13∗∗∗ −2.35
(0.59) (0.60) (1.32) (0.66) (1.25)
iv 0.66∗∗∗ 0.85∗∗∗
(0.06) (0.08)
iv:BE 0.72∗∗∗ 0.48∗
(0.08) (0.19)
iv:DE 0.72∗∗∗ 0.52∗∗
(0.08) (0.17)
iv:ES 0.86∗∗∗ 0.77∗∗∗
(0.05) (0.20)
iv:FR 0.75∗∗∗ 0.67∗∗∗
(0.04) (0.09)
iv:IT 0.62∗∗∗ 0.25∗∗
(0.05) (0.08)
iv:NL 0.69∗∗∗ 0.18
(0.06) (0.10)
iv:UK 0.87∗∗∗ 0.86∗∗∗
(0.06) (0.10)
AIC 3469.13 1738.04 1746.58 1713.86 3467.96 3437.01
Log Likelihood -1728.57 -865.02 -870.29 -845.93 -1726.98 -1703.50
Num. obs. 452 421 452 452 452 452
∗∗∗
p < 0.001, ∗∗ p < 0.01, ∗ p < 0.05

Table 2: Nested logit models for the choice by Japanese firms of a european region

nl.fdi <- mlogit(choice ~ log(wage) + unemp + elig + log(area) +

scrate + ctaxrate | 0, data = jfdi,
nests = TRUE, un.nest.el = TRUE)
nl2.fdi <- update(nl.fdi, un.nest.el = FALSE,
constPar = c('iv:PT' = 1, 'iv:IE' = 1))

The results of the fitted models are presented in table 2 using the texreg package.
For the nested logit models, two tests are of particular interest:

• the test of no nests, which means that all the nest elasticities are equal to 1,
• the test of unique nest elasticities, which means that all the nest elasticities are
equal to each other.
 Yves Croissant 29

For the test of no nests, the nested model is provided as the unique argument for the lrtes
and the waldtest function. For the scoretest, the constrained model (i.e. the multinomial
logit model) is provided as the first argument and the second argument is nests, which
describes the nesting structure that one wants to test.

lr.nest <- lrtest(nl2.fdi)

wd.nest <- waldtest(nl2.fdi)
sc.nest <- scoretest(ml.fdi, nests = TRUE,
constPar = c('iv:PT' = 1, 'iv:IE' = 1))

The Wald test can also be performed using the linearHypothesis function:

lh.nest <- linearHypothesis(nl2.fdi,

c("iv:BE = 1", "iv:DE = 1", "iv:ES = 1",
"iv:FR = 1", "iv:IT = 1", "iv:NL = 1",
"iv:UK = 1"))

sapply(list(wald = wd.nest, lh = lh.nest,

score = sc.nest, lr = lr.nest), statpval)

## wald lh score lr
## stat 208.407 208.407 60.28 50.122
## p-value 0.000 0.000 0.00 0.000

The three tests reject the null hypothesis of no correlation. We next test the hypothesis
that all the nest elasticities are equal.

lr.unest <- lrtest(nl2.fdi, nl.fdi)

wd.unest <- waldtest(nl2.fdi, un.nest.el = TRUE)
sc.unest <- scoretest(ml.fdi, nests = TRUE, un.nest.el = FALSE,
constPar = c('iv:IE' = 1, 'iv:PT' = 1))
lh.unest <- linearHypothesis(nl2.fdi,
c("iv:BE = iv:DE", "iv:BE = iv:ES",
"iv:BE = iv:FR", "iv:BE = iv:IT",
"iv:BE = iv:NL", "iv:BE = iv:UK"))

sapply(list(wald = wd.unest, lh = lh.unest,

score = sc.unest, lr = lr.unest), statpval)

## wald lh score lr
## stat 73.535 73.535 60.28 46.954
## p-value 0.000 0.000 0.00 0.000
30 Estimation of Random Utility Models in R: The mlogit Package

Once again, the three tests strongly reject the hypothesis.

4. The random parameters (or mixed) logit model

4.1. Derivation of the model

A mixed logit model or random parameters logit model is a logit model for which the
parameters are assumed to vary from one individual to another. It is therefore a model
that takes the heterogeneity of the population into account.

The probabilities
For the standard logit model, the probability that individual i choose alternative j is:
0
eβ xil
Pil = P β 0 x
je
ij

Suppose now that the coefficients are individual-specific. The probabilities are then:

0
eβi xil
Pil = P β 0 x
je i
ij

The mixed logit model consists on considering the βi ’s as random draws from a dis-
tribution whose parameters are estimated. The probability that individual i choose
alternative l, for a given value of βi is:

0
eβi xil
Pil | βi = P β 0 x
je i
ij

To get the unconditional probability, we have to integrate out this conditional probability,
using the density function of β. Suppose that Vil = α + βi xil , i.e. there is only one
individual-specific coefficient and that the density of βi is f (β, θ), θ being the vector of
the parameters of the distribution of β. The unconditional probability is then:

0
eβi xil
Z Z
Pil = E(Pil | βi ) = (Pil | β)f (β, θ)dβ = P β 0 xij f (β, θ)dβ
β β je i

which is a one-dimensional integral that can be efficiently estimated by quadrature meth-

ods. If Vil = βi> xil where βi is a vector of length K and f (β, θ) is the joint density of
the K individual-specific coefficients, the unconditional probability is:
Z Z Z
Pil = E(Pil | βi ) = ... (Pil | β)f (β, θ)dβ1 dβ2 . . . dβK
β1 β2 βK
 Yves Croissant 31

This is a K-dimensional integral which cannot easily be estimated by quadrature meth-

ods. The only practical method is then to use simulations. More precisely, R draws of the
parameters are taken from the distribution of β, the probability is computed for every
draw and the unconditional probability, which is the expected value of the conditional
probabilities is estimated by the average of the R probabilities.

Individual parameters
The expected value of a random coefficient (E(β)) is simply estimated by the mean of
the R draws on its distribution : β̄ = R
P
r=1 βr . Individual parameters are obtained by
first computing the probabilities of the observed choice of i for every value of βr :
0
yij eβr xij
P
j
Pir = P 0
j eβr xij

where yij is a dummy equal to one if i has chosen alternative j. The expected value of
the parameter for an individual is then estimated by using these probabilities to weight
the R β values:
P
Pir βr
β̂i = Pr
r Pir

Panel data
If there are repeated observations for the same individuals, this longitudinal dimension of
the data can be taken into account in the mixed logit model, assuming that the random
parameters of individual i are the same for all his choice situations. Denoting yitl a
dummy equal to 1 if i choose alternative l for the tth choice situation, the probability of
the observed choice is:

yitj eβi xitl

P
j
Y
Pit = P βi xitj
j je

The joint probability for the T observations of individual i is then:

yitj eβi xitj

P
j
YY
Pi = P βi xitj
t j je
P
and the log-likelihood is simply i ln Pi .

4.2. Application
The random parameter logit model is estimated by providing a rpar argument to mlogit.
This argument is a named vector, the names being the random coefficients and the
32 Estimation of Random Utility Models in R: The mlogit Package

values the name of the law of distribution. Currently, the normal ("n"), log-normal
("ln"), zero-censored normal ("cn"), uniform ("u") and triangular ("t") distributions
are available. For these distributions, two parameters are estimated which are, for normal
related distributions, the mean and the standard-deviation of the underlying normal
distribution and for the uniform and triangular distribution, the mean and the half
range of the distribution. For these last two distributions, zero-bounded variants are
also provided ("zbt" and "zbu"). These two distributions are defined by only one
parameter (the mean) and their definition domain varies from 0 to twice the mean.
It’s often the case that we are willing to impose that the distribution of a random
parameter takes only positive or negative values. For example, the price coefficient should
be negative for every individual. In this case, "zbt" and "zbu" can be used. The use of
"ln" and "cn" can also be relevant but, in this case, if only negative values are expected,
one should consider the distribution of the opposite of the random price coefficient.
This can easily be done using the opposite argument of mlogit.data. For example,
if opposite = c("price", "time") is used, mlogit.data returns the opposite of the
two variables, so that the corresponding coefficients should now be positive.
R is the number of draws, halton indicates whether halton draws (see Train 2003, chapter
9) should be used (NA indicates that default halton draws are used), panel is a boolean
which indicates if one wishes to use the panel data version of the log-likelihood.
Correlations between random parameters can be introduced only for normal-related dis-
tributed random parameters, using the correlation argument. If TRUE, all the normal-
related random parameters are correlated. The correlation argument can also be a
character vector indicating the random parameters that one wishes to be correlated.

Train
We first use the Train data set, previously coerced to a mlogit.data object called Tr. We
first estimate the multinomial model: both alternatives being virtual train trips, it is
relevant to use only generic coefficients and to remove the intercept:

Train.ml <- mlogit(choice ~ price + time + change + comfort | - 1, Tr)

coef(summary(Train.ml))

## Estimate Std. Error z-value Pr(>|z|)

## price 0.06735804 0.003393252 19.850585 0.000000e+00
## time 1.72055142 0.160351702 10.729861 0.000000e+00
## change 0.32634094 0.059489152 5.485722 4.117843e-08
## comfort 0.94572555 0.064945464 14.561842 0.000000e+00

All the coefficients are highly significant and have the predicted positive sign (remind
than an increase in the variable comfort implies using a less comfortable class). The
coefficients can’t be directly interpreted, but dividing them by the price coefficient, we
get monetary values:
 Yves Croissant 33

coef(Train.ml)[- 1] / coef(Train.ml)[1]

## time change comfort

## 25.54337 4.84487 14.04028

We obtain the value of 26 euros for an hour of traveling, 5 euros for a change and 14
euros to travel in a more comfortable class.
We then estimate a model with three random parameters, time, change and comfort. We
first estimate the uncorrelated mixed logit model:

Train.mxlu <- mlogit(choice ~ price + time + change + comfort | - 1, Tr,

panel = TRUE, rpar = c(time = "n", change = "n",
comfort = "n"), R = 100,
correlation = FALSE, halton = NA, method = "bhhh")
names(coef(Train.mxlu))

## [1] "price" "time" "change" "comfort"

## [5] "sd.time" "sd.change" "sd.comfort"

Compared to the multinomial logit model, there are now three more coefficients which
are the standard deviations of the distribution of the three random parameters. The
correlated model is obtained by setting the correlation argument to TRUE.

Train.mxlc <- update(Train.mxlu, correlation = TRUE)

names(coef(Train.mxlc))

## [1] "price" "time" "change"

## [4] "comfort" "time.time" "time.change"
## [7] "time.comfort" "change.change" "change.comfort"
## [10] "comfort.comfort"

There are now 6 parameters which are the elements of the Choleski decomposition of
the covariance matrix of the three random parameters.
The summary method supplies the usual table of coefficients, and also some statistics
about the random parameters. Random parameters may be extracted using the func-
tion rpar which take as first argument a mlogit object, as second argument par the
parameter(s) to be extracted and as third argument norm the coefficient (if any) that
should be used for normalization. This is usually the coefficient of the price (taken as a
non random parameter), so that the effects can be interpreted as monetary values. This
function returns a rpar object, and several methods/functions (summary, mean, med for
the median and stdev for the standard deviation) are provided to describe it:
34 Estimation of Random Utility Models in R: The mlogit Package

time.value <- rpar(Train.mxlc, "time", norm = "price")

summary(time.value)

## Min. 1st Qu. Median Mean 3rd Qu. Max.

## -Inf 8.753119 33.367588 33.367588 57.982056 Inf

In case of correlated random parameters further functions are provided to analyze the
correlation of the coefficients:

cor.mlogit(Train.mxlc)

## time change comfort

## time 1.00000000 -0.02956296 0.3695645
## change -0.02956296 1.00000000 0.2489270
## comfort 0.36956453 0.24892701 1.0000000

cov.mlogit(Train.mxlc)

## time change comfort

## time 28.6460389 -0.2787999 5.557933
## change -0.2787999 3.1047367 1.232467
## comfort 5.5579334 1.2324667 7.895535

stdev(Train.mxlc)

## time change comfort

## 5.352199 1.762026 2.809899

As the change attribute seems to be weakly correlated with the two other random
parameters, the correlation can be restricted to the time and comfort attributes by
filling the correlation argument with a character vector.

Train.mxlc2 <- update(Train.mxlc, correlation = c("time", "comfort"))

cor.mlogit(Train.mxlc2)

## time comfort
## time 1.0000000 0.3909467
## comfort 0.3909467 1.0000000

The presence of random coefficients and their correlation can be investigated using any
of the three tests. Actually, three nested models can be considered, a model with no
random effects, a model with random but uncorrelated effects and a model with random
and correlated effects. We first present the three tests of no correlated random effects:
 Yves Croissant 35

lr.mxc <- lrtest(Train.mxlc, Train.ml)

wd.mxc <- waldtest(Train.mxlc)
lh.mxc <- linearHypothesis(Train.mxlc,
c("time.time = 0", "time.change = 0",
"time.comfort = 0", "change.change = 0",
"change.comfort = 0", "comfort.comfort = 0"))
sc.mxc <- scoretest(Train.ml,
rpar = c(time = "n", change = "n", comfort = "n"),
R = 100, correlation = TRUE, halton = NA,
panel = TRUE)
sapply(list(wald = wd.mxc, lh = lh.mxc,
score = sc.mxc, lr = lr.mxc), statpval)

## wald lh score lr
## stat 288.287 288.287 208.765 388.057
## p-value 0.000 0.000 0.000 0.000

The hypothesis of no correlated random parameters is strongly rejected. We then present

the three tests of no correlation, the existence of random parameters being maintained.

lr.corr <- lrtest(Train.mxlc, Train.mxlu)

lh.corr <- linearHypothesis(Train.mxlc, c("time.change = 0",
"time.comfort = 0",
"change.comfort = 0"))
wd.corr <- waldtest(Train.mxlc, correlation = FALSE)
sc.corr <- scoretest(Train.mxlu, correlation = TRUE)
sapply(list(wald = wd.corr, lh = lh.corr,
score = sc.corr, lr = lr.corr), statpval)

## wald lh score lr
## stat 103.195 103.195 10.483 42.621
## p-value 0.000 0.000 0.015 0.000

The hypothesis of no correlation is strongly reject with the Wald and the likelihood ratio
test, only at the 1% level for the score test.

RiskyTransport
The second example is a study by León and Miguel (2017) who consider a mode-choice
model for transit from Freetown’s airport (Sierra-Leone) to downtown. Four alternatives
are available: ferry, helicopter, water-taxi and hovercraft. A striking characteristic of
their study is that all these alternatives experienced fatal accidents in recent years, so
that the fatality risk is non-negligible and differs much from an alternative to another.
36 Estimation of Random Utility Models in R: The mlogit Package

For example, the probabilities of dying using the water taxi and the helicopter are
respectively of 2.55 and 18.41 out of 100,000 trips. This feature enables the authors
to estimate the value of a statistical life. For an individual i, the utility of choosing
alternative j is:

Uij = βil (1 − pj ) + βic (cj + wi tj ) + ij

where pj is the probability of dying while using alternative j, cj and tj the monetary
cost and the transport time of alternative j and wi the wage rate of individual i (which
is supposed to be his valuation of transportation time). Cij = cj + wi tj is therefore
the individual specific generalized cost for alternative j. βil and βic are the (individual
specific) marginal utility of surviving and of expense. The value of the statistical life
(VSL) is then defined by:

βil ∆Cij
VSLi = − =
βc ∆(1 − pj )

The two covariates of interest are cost (the generalized cost in $PPP) and risk (mortality
per 100,000 trips). The risk variable being purely alternative specific, intercepts for the
alternatives cannot therefore be estimated. To avoid endogeneity problems, the authors
introduce as covariates marks the individuals gave to 5 attributes of the alternatives:
comfort, noise level, crowdedness, convenience and transfer location. We first estimate
a multinomial logit model.

data("RiskyTransport", package = "mlogit")

RT <- mlogit.data(RiskyTransport, shape = "long", choice = "choice",
chid.var = "chid", alt.var = "mode", id.var = "id")
ml.rt <- mlogit(choice ~ cost + risk + seats + noise + crowdness +
convloc + clientele | 0, data = RT, weights = weight)

Note the use of the weights argument in order to set weights to the observations, as done
in the original study. The ratio of the coefficients of risk and of cost is 9.84 (hundred of
thousands of $), which means that the estimated value of the statistical life is a bit less
than one million $. We next consider a mixed logit model. The coefficients of cost and
risk are assumed to be random, following a zero-bounded triangular distribution.

mx.rt <- mlogit(choice ~ cost + risk + seats + noise + crowdness +

convloc + clientele | 0, data = RT, weights = weight,
rpar = c(cost = 'zbt', risk = 'zbt'),
R = 100, halton = NA, panel = TRUE)
 Yves Croissant 37

screenreg(list('mult.logit' = ml.rt, 'mixed logit' = mx.rt), digits = 4)

##
## ==============================================
## mult.logit mixed logit
## ----------------------------------------------
## cost -0.0095 *** -0.0187 ***
## (0.0011) (0.0013)
## risk -0.0939 *** -0.1030 ***
## (0.0110) (0.0159)
## seats 0.1517 0.1085
## (0.2443) (0.2333)
## noise -0.0290 0.1423
## (0.2654) (0.2289)
## crowdness -0.9186 *** -0.7157 **
## (0.2445) (0.2225)
## convloc -0.3772 -0.1497
## (0.2016) (0.1971)
## clientele -0.2567 -0.3314
## (0.2651) (0.2541)
## ----------------------------------------------
## AIC 3250.7472 3177.2501
## Log Likelihood -1618.3736 -1581.6251
## Num. obs. 1793 1793
## ==============================================
## *** p < 0.001, ** p < 0.01, * p < 0.05

Individual-level parameters can be extracted using the fitted method, with the type
argument set to parameters.

indpar <- fitted(mx.rt, type = "parameters")

We can then compute the VSL for every individual and analyse their distribution, using
quantiles and plotting the empirical density of VSL for African and non-African travelers
(as done in León and Miguel 2017, table 4, p.219 and figure 5, p.223).

indpar$VSL <- with(indpar, risk / cost * 100)

quantile(indpar$VSL, c(0.025, 0.975))

## 2.5% 97.5%
## 432.4132 1054.3365

mean(indpar$VSL)
38 Estimation of Random Utility Models in R: The mlogit Package

## [1] 608.9317

library("ggplot2")
indpar <- merge(unique(subset(RT, select = c("id", "african"))), indpar)
ggplot(indpar) + geom_density(aes(x = VSL, linetype = african)) +
scale_x_continuous(limits = c(200, 1200))

0.004

african
density

no
yes

0.002

0.000

250 500 750 1000 1250

VSL

5. Conclusions
mlogit estimates a large set of random utility models with a unified and friendly interface.
Some of these models haven’t been presented in this article, namely the rank-ordered
logit model, the overlapping nested logit model, the paired combinatorial logit model
and the multinomial probit model. Moreover, it provides usefull functions and methods
which compute and return useful results, like predicted probabilities, inclusive values,
marginal effects and consumer surplus.

References
 Yves Croissant 39

Ben-Akiva M, Bolduc D, Bradley M (1993). “Estimation of travel choice models with

randomly distributed values of time.” Transportation Research Record, 1413, 88–97.

Bhat C (1995). “A heterocedastic extreme value model of intercity travel mode choice.”
Transportation Research B, 29(6), 471–483.

Daly A (1987). “Estimating ‘tree’ logit models.” Transportation Research B, pp. 251–267.

Forinash CV, Koppleman FS (1993). “Application and interpretation of nested logit

models and intercity mode choice.” Transportation Record, 1413, 98–106.

Fowlie M (2010). “Emissions Trading, Electricity Restructuring, and Investment in Pol-

lution Abatement.” American Economic Review, 100(3), 837–69. doi:10.1257/aer.
100.3.837. URL http://www.aeaweb.org/articles?id=10.1257/aer.100.3.837.

Fox J, Weisberg S (2010). An R Companion to Applied Regression. Second edi-

tion. Sage, Thousand Oaks CA. URL http://socserv.socsci.mcmaster.ca/jfox/
Books/Companion.

Hasan A, Wang Z, Mahani A (2016). “Fast Estimation of Multinomial Logit Models: R

Package mnlogit.” Journal of Statistical Software, Articles, 75(3), 1–24. ISSN 1548-
7660. doi:10.18637/jss.v075.i03. URL https://www.jstatsoft.org/v075/i03.

Head K, Mayer T (2004). “Market Potential and the Location of Japanese Investment in
the European Union.” The Review of Economics and Statistics, 86(4), 959–972. doi:
10.1162/0034653043125257. https://doi.org/10.1162/0034653043125257, URL
https://doi.org/10.1162/0034653043125257.

Heiss F (2002). “Structural choice analysis with nested logit models.” The Stata Journal,
2(3), 227–252.

Hensher DA, Greene WH (2002). “Specification and estimation of the nested logit model:
alternative normalisations.” Transportation Research Part B, 36, 1–17.

Koppelman FS, Wen CH (1998). “Alternative nested logit models: structure, properties
and estimation.” Transportation Research B, 32(5), 289–298.

Koppelman FS, Wen CH (2000). “The paired combinatorial logit model: properties,
estimation and application.” Transportation Research B, 34, 75–89.

León G, Miguel E (2017). “Risky Transportation Choices and the Value of a Statisti-
cal Life.” American Economic Journal: Applied Economics, 9(1), 202–28. doi:10.
1257/app.20160140. URL http://www.aeaweb.org/articles?id=10.1257/app.
20160140.

McFadden D (1974). “The measurment of urban travel demand.” Journal of public

economics, 3, 303–328.
40 Estimation of Random Utility Models in R: The mlogit Package

McFadden D (1978). “Spatial interaction theory and planning models.” In A Karlqvist

(ed.), Modeling the choice of residential location, pp. 75–96. North-Holland, Amster-
dam.

Meijer E, Rouwendal J (2006). “Measuring welfare effects in models with random co-
efficients.” Journal of Applied Econometrics, 21(2), 227–244. doi:10.1002/jae.
841. https://onlinelibrary.wiley.com/doi/pdf/10.1002/jae.841, URL https:
//onlinelibrary.wiley.com/doi/abs/10.1002/jae.841.

Sarrias M, Daziano R (2017). “Multinomial Logit Models with Continuous and Discrete
Individual Heterogeneity in R: The gmnl Package.” Journal of Statistical Software,
Articles, 79(2), 1–46. ISSN 1548-7660. doi:10.18637/jss.v079.i02. URL https:
//www.jstatsoft.org/v079/i02.

Small KA, Rosen HS (1981). “Applied welfare economics with discrete choice models.”
Econometrica, 49, 105–130.

Swait J, Louviere J (1993). “The role of the scale parameter in the estimation and use
of multinomial logit models.” Journal of Marketing Research, 30.

Toomet O, Henningsen A (2010). maxLik: Maximum Likelihood Estimation. R package

version 0.8-0, URL http://CRAN.R-project.org/package=maxLik.

Train KE (2003). Discrete choice methods with simulation. Cambridge University Press,
Cambridge, UK.

Zeileis A, Croissant Y (2010). “Extended Model Formulas in R: Multiple Parts and

Multiple Responses.” Journal of Statistical Software, 34(1), 1–13. URL http://www.
jstatsoft.org/v34/i01/.

Zeileis A, Hothorn T (2002). “Diagnostic Checking in Regression Relationships.” R

News, 2(3), 7–10. URL http://CRAN.R-project.org/doc/Rnews/.

Affiliation:
Yves Croissant
Faculté de Droit et d’Economie
Université de la Réunion
15, avenue René Cassin
BP 7151
F-97715 Saint-Denis Messag Cedex 9
Telephone: +33/262/938446
E-mail: [email protected]