Tourism and Hospitality Research Volume 4 Number 4

Modelling determinants of tourism demand

as a five-stage process: A discrete choice
methodological approach

Juan L. Eugenio-Martin
Received (in revised form): 8th April, 2003

Environment Department, University of York, Heslington, York YOlO 500, UK

Tel: +44 (0)1904434067; e-mail: jlem100@york.ac.uk
Departamento de Analisis Economico Aplicado, Universidad de Las Palmas de Gran Canaria,
C/Sauio Toron, CP 35017, Spain
Tel: +34928458216; jluis@empresariales.ulpgc.es

Juan Luis Eugenio-Martin has worked as a ists and that final destination choice is not an
lecturer in economics at the Department of independent decision, but the last decision of a
Applied Economics (University of Las set of choices that are determinin<~ it. ln this
PalmasJ, where he is currently finishing sense, it is ar<~ued that tourists face a five-stag«
his PhD dissertation on the determinants decision process. First of all, people have to
of tourism demand. He is also undertaking decide whether or not to travel within a period
a PhD in Environmental Economics and of time. Secondly, those who expect to travel
Management at the Environment Depart- need to estimate a budget for tourism expenses.
ment of the University of York. In 2001, Thirdly, given the budget, they need to deter-
Juan completed a postgraduate pro- mine the frequency and length of stay of their
gramme in econometrics and in 2002, an trip. Fourthly, once a date and the length of
MSc in Economics, both at the University stay are proposed, tourists need to choose which
of York. His research interests include kind of tourist destination to visit. Finally,
tourism economics, environmental eco- from among all the available destinations that
nomics, economic growth and mlcroecono- satisfy a tourist's conditions, final destination
metrics. and mode of transportation are chosen. It is the
purpose of this paper to propose a methodologi-
ABSTRACT cal framework for modelling each of these stages
KEYWORDS: tourism demand, outbound and their relationship.
tourism, discrete choice models, tourists'
decisions, tourism marketing INTRODUCTION
For many regions tourism has become one
In the tourists' destination choice there are mul- of the most significant economic activities
tiple factors involved in their decision. Indivi- in terms of economic growth and employ-
duals or families with exactly the same socio- . ment. World tourism demand is still grow-
economic and demographic characteristics may ing and new or current destinations may
choose very different destinations. The paper be developed or extended in order to Tourism and Hospitality Research.
Vol. 4, No.4. 2003,
deals with this heterogeneity problem, recognis- satisfy such growth. In this sense, tourism pp.341-354
(, Henry Stewart Publications.
ing that there are taste differences among tour- may be seen as an opportunity for the eco- 1467-3584

I
Downloaded from thr.sagepub.com at Mount Royal University on June 6, 2015
Page J41
Modelling determinants of tourism demand as a five-stage process

nomic growth of developing regions. provide a methodological framework that

Despite the fact that tourism demand is still can estimate the main determinants of out-
growing, however, tourism supply is also bound tourism demand.
growing in the same fashion and competi- As mentioned in the introduction of this
tion among destinations increases. Destina- paper, the author considers that before
tions may deal with competition from a deciding where to go on holiday, most
two-level perspective. On the one hand, tourists need to make multiple decisions.
from a micro-level point of view, the hos- While for some people their decisions are
pitality sector may change prices and qual- perfectly planned, for other people these
ity and adapt the services offered to the are improvised or hardly planned. More-
preferences of their potential visitors. On over, some people can make all the deci-
the other hand, from a macro-level point sions simultaneously or in different stages.
of view, local, regional or national authori- Whether planned or not however, it is
ties may invest in the development of tour- argued that most people consider con-
ist resorts and promote them. In order to sciously or unconsciously a process of deci-
be efficiently applied, any of these policies sions concerning their holiday trips. For
requires a deep knowledge of the charac- modelling purposes, it is assumed that tour-
teristics of potential visitors, their needs ists choose a final destination depending on
and the interrelationship with other com- another four decisions. In this sense, five
petitive destinations. Therefore, the ques- stages are considered: participation deci-
tion that needs to be answered is why sion; tourism budget decision; frequency
people travel to different places. The main and the length of stay decisions; the kind
focuses are related to tourism marketing of destination decision; and final destina-
and tourism planning decisions from both tion and the mode of transportation choice.
micro- and macro-level points of view. The methodology proposed is a general
More precisely, this paper tries to provide framework applicable to any place in the
a methodological framework that may world of any size. For instance, it may be
analyse the relative importance of different employed for a country, a region or a
attributes for tourists' destination choice; to small town or village. Obviously, the
estimate the probability of visiting each larger the region analysed, the more het-
kind of destination for different kind of erogeneity will be faced.
tourists; and finally, to generate a tool that In order to apply this methodology two
allows simulation of changes in the different datasets are required. On the one
demand under alternative scenarios. hand, there is a need for micro data on the
socio-economic and demographic charac-
TOURISTS' DESTINATION CHOICE teristics of a representative sample of popu-
PROCESS lation. This dataset must also include data
When studying tourism demand, two on tourism trips; for instance, places vis-
points of view can be considered. On the ited, number of trips, length of stay or
one hand, the number of tourists that are expenditure on tourism. On the other hand
expected to arrive at a particular destina- data are required on the attributes of the
tion can be forecast; ie inbound tourism is choice set. This is an objective dataset and
considered. On the other hand, an attempt is easier to obtain. It usually includes vari-
can be made to understand the tourist des- ables such as accommodation cost index,
tination choice of the inhabitants of a parti- price index, development level or tempera-
cular region; ie outbound tourism is ture.
analysed. The purpose of this paper is to Sampling can be from the whole popu-

Pay" 342 Downloaded from thr.sagepub.com at Mount Royal University on June 6, 2015
 Euqenio-Martin

lation, from on-site or from a combination Usually, this may range from a period of a
of both of these. The main advantage of year to a period that contains the whole
on-site sampling is that deeper and wider life of the interviewee.
variability may be obtained compared with
population sampling. Although for a desti- Objective
/'
nation choice analysis, on-site sampling This is an interesting issue for tourism
might be more convenient, for the pur- marketing analysts, because it shows which
poses of this paper it would be incomplete are the determinants of the decision to
because information might be lost concern- travel. Similar models can be constructed
ing the reasons why people decide whether depending on the socio-economic or
or not to travel. demographic characteristics of the indivi-
Another issue related to sampling is the duals. The objective is to estimate different
period of time the study covers, which models using a segmentation criterion, so
could range from a season to a year or a set that results from different models can be
of years. The period chosen depends on the confirmed and conclusions drawn. For
purposes of the analysis. A period of a instance, it is possible to estimate determi-
season might be chosen if the region ana- nants of participation for different places of
lysed is remarkably affected by seasonality residence. This analysis may reveal that
and this effect is relevant for the purposes residents of a particular region are less
of the study. A longer period than a year is interested in travelling than residents of
useful if the researcher wants to trace tour- other regions due to several aspects such as
ist behaviour over time. This is potentially income differences or differences in speci-
interesting because it would reveal aspects fied facilities for recreation, which may
such as repetition patterns, risk aversion help residents to enjoy leisure time in their
and tourism budget decision making. One own place of residence. Moreover, seg-
inconvenience, however, is the possible mentation in the sample can be made to
mistakes interviewees may make especially relate to any variable, and the relative
concerning data on trips made more than importance for each segment of any vari-
three years ago. A solution is to generate a able can be compared. A common case to
panel dataset that traces individuals over be analysed is the effect of age. The relative
time, repeating the interview for each importance of the particular problems of
period of time. For a general purpose, a each segment can be estimated. For
period of one year seems to be the more instance, comparing how important
appropriate, as recommended by Morley income is for the youth segment with
(1995). respect to other segments; or how relevant
The next section deals with each of the whether or not they have a child is for
stages, explaining the objectives, variables middle-aged tourists' decision; or how sig-
considered and the methodology proposed nificant are health conditions for the deci-
for the analysis, and alternative methodolo- sion of elderly people. Marketing effort can
gies are briefly discussed. focus on different segments and their main
determinants for travelling.
STAGE ONE: PARTICIPATION DECISION
The first decision any individual has to. Meln verlebles
make concerns the choice of whether or Intuition suggests that variables like age,
not to travel within a period of time. The education, income, labour conditions, char-
researcher, depending on the purposes of acteristics of the place of residence and size
the analysis, must set a time interval. and composition of the household or

Downloaded from thr.sagepub.com at Mount Royal University on June 6, 2015

PalJU J4J
Modelling determinants 01 tourism demand as a five-stage process

family, may be significant when deciding Then, Pr( T; = 1) = Pr(Si + 8i> 0) =

whether or not to travel. Pr(8/> Si) = l-Pr(8i~-S;) = l-FE(-S;),

Methodology where FE denotes the cumulative density

In order to model participation decision, it function of the unobserved part. Due to
is considered a binary choice, denoted by, the problem of identification of the loca-
T;, such that T; = 1 if a household or indi- tion and scale of Tt, the researcher needs
vidual decides to travel and T; = 0 other- to choose a distribution and a value for the
wise. To model the probability that T; = 1, variance of e; The most common
ie Pr(T;=l), it is assumed that Pr(T;=l) is approaches assume e, is independently and
linked to a set of exogenous variables, identically distributed, either following a
which may be those already shown above. normal distribution with zero mean and
More precisely, for some appropriate func- variance of one, or following a logistic dis-
tion g(.), Pr(T;=l)=g((X+r.~=lPiSEi;)' tribution with zero mean and variance of
where 0 ~g(.) ~ 1, ex denotes a constant n2 j3. If it is assumed that 8j follows the
term, SEji denotes the lh socio-economic former distribution the equally well-
variable of the household or individual i known probit model is being employed,
and Pi denotes the associated parameter to and if we assume the latter distribution we
the iii socio-economic variable. are employing the also well-known logit
It is not recommended that a traditional model. Any of these distributions can be
linear probability model be used to esti- employed for the participation decision
mate the probability function because it and both present similar results. Finally,
would present non-normal errors, hetero- maximum likelihood estimation is applied
skedasticity and logical inconsistency, since to the model in order to estimate para-
the prediction of probabilities may lie out meters of interest. Under correct specifica-
of range (0,1). It is well known that the tion, these estimates are consistent and
suggested model for binary choice estima- asymptotically normal. For a complex
tions is the latent variable model. This exposition of the methodology see Greene
model considers the existence of a latent (2003).
variable r: Since this latent variable is
unobserved by the researcher it can be con- STAGE TWO: TOURISM BUDGET
sidered to be composed of two parts: one CONSTRAINT
observed by the researcher, which includes Once a household or individual has
all the socio-economic variables and decided to travel, they have to decide,
another part that is unobserved by the consciously or unconsciously, how much
researcher and that corresponds to the their tourism expenditure may be. This
heterogeneous behaviour of the tourists. decision depends mainly on the income
Therefore, the model can be represented and preferences of individuals. If the ana-
as: T;* = (X + r.ki = I PJSE.i; + e.; where 8; de- lysis is performed with income in absolute
notes the unobserved part or error term. terms, this variable is likely to dominate
For the purposes of this paper, the latent the estimated regression. In order to avoid
variable will work as an index function, a trivial result, the tourism budget can be
such that t; = 1 will be set if > 0 and t; estimated as a percentage of income, ie as
t; =0 will be set if t; ~O. a ratio between tourism expenditure and
income. This new formulation allows an
Let S; = IX + r.~ = t{JiSEji, such that estimate of how much people prefer to
r; =S;+8;. distribute their income for tourism pur-

Pd(J" J44 Downloaded from thr.sagepub.com at Mount Royal University on June 6, 2015
 Eugenio Mal lin

poses, or in other words, how much they consider income per capita per year
people like tourism. and expenditure per capita of the country
in the destination. Therefore, the assump-
Objective tion is that all potential tourists possess the
The central purpose of this stage is to try same income and all of them spend the
to understand the main factors that push same percentage of their budget in tourism
different individuals to spend part of their activities. They cannot distinguish if a tour-
budget in tourism activities. ist has travelled once or more times in a
year, or how this may affect their decisions.
Main variables Consequently, it assumes that tourists
It is interesting to estimate the relative travel to all the destinations within the
importance of variables such as age, period of time of the analysis. Further-
income, labour conditions, place of resi- more, they do not consider individual
dence and size of the household or family socio-economic variables such as for
in tourism budget decisions. instance: age, number of children, house-
hold income, labour situation, gender or
Methodology health status that may be relevant to the
Depending on the nature of the data differ- tourist's decision process. On the other
ent methodologies can be applied. Ideally, hand, micro data can be used to estimate
both income and tourism expenditure the system of demands. Lack of data, how-
might be continuous variables. It is likely, ever, has usually been a common problem.
however, that the questionnaires consider Nevertheless, it can be assumed that the
discrete intervals for these two variables. whole set of commodities can be divided
For the continuous variables case, tour- into different groups, such that preferences
ism expenditure can be estimated as depen- within groups can be described indepen-
dent on its own determinants as well as on dently of the quantities in other groups.
the demand for other goods and services. This assumption is known as 'weak separ-
A traditional approach is to estimate this ability' (for a further reference see Deaton
tourism expenditure as a demand function and Muellbauer (1980: 119-136)) of the
that is part of a system of demands which utility function and it is a plausible and
include all other goods and services. common assumption in the tourism eco-
Deaton and Muellbauer (1980a) dealt with nomics literature (see for instance Rugg
this issue, developing a model known as (1973)). Under this assumption a tree of
Almost Ideal Demand System. The tour- commodities can be created, where one of
ism literature has employed this approach the main branches may be a group called
frequently (O'Hagan and Harrison (1984); entertainment, and tourism may be a
Syriopoulos (1993); Papatheodorou (1999); further branch linked to entertainment.
and Divisekera (2003)). Most of the studies, This structure allows analysis of tourism
however, have used macroeconomic vari- expenditure allocation independently of
ables rather than micro data. The main expenditure levels in any other goods and
advantage of macroeconomic variables is services, as done by Hultkrantz (1995),
that the dataset is usually easy to obtain. who considers a three-stage budgeting
Nevertheless these variables are estimates of. framework.
an aggregate measure, and consequently, if Consequently, under weak separability
these are employed in tourism studies, it assumption, the tourism budget decision
has to be taken into account that they are can be modelled depending on the socio-
based on population averages. For instance, economic variables already mentioned.

Downloaded from thr.sagepub.com at Mount Royal University on June 6, 2015

Modelling determinants of tourism demand as a five-stage process

Since the tourism budget is measured as a expenditure over income, between zero
percentage of total income, however, such and the lowest threshold determined by the
that its value must lie between zero and researcher, means these individuals do not
one, a linear probability model cannot be like tourism. As long as the interval of the
applied for the same reasons already set corresponds to a higher percentage, the
explained in the participation decision researcher can define other labels such as
model. Several solutions can be adopted. A 'do not like tourism much', 'like tourism',
traditional way is to employ a translog 'like tourism much' and finally, 'like tour-
model. This model usually works appro- ism very much'. In order to deal with this
priately for independent functions. If the ordered categorical variable, what is known
model is used for a system of demand func- as an ordered probit model can be used.
tions, however, it usually fails in testing In the ordered probit model, there are
additive and homothetic restrictions different ordered multinomial outcomes,
(Bakkal, 1991). Another way to model the denoted by j, where j = 1,..., m. In this
tourism budget is through a double cen- example the categories are j = { do not like
sored regression model. In this case, the tourism, like tourism a bit, like tourism,
researcher can impose two censors to the like tourism a lot and like tourism very
potential freedom given to any endogenous much}. A traditionally ordered probit
variable in the traditional regression analy- model estimates thresholds itself, but since
sis. Obviously the lower censor might be the researcher may predetermine these, as
set at zero and the upper censor at one, already noted, a superior regression known
such that any estimate is guaranteed to lie as interval or grouped data regression can
between these two values (for a complete be employed. This is no more than a var-
analysis of this technique, see Maddala iant in which the values of the thresholds
(1983) or Greene (2003)). are known. Because the thresholds are
As mentioned, it is likely that micro data known, the estimates of the parameters are
on income or tourism expenditure will be more efficient and it is possible to identify
collected in the interview process, as the the variance of the error term.
choice among alternative predetermined Similar justification to the model pre-
intervals, rather than an exact number. sented in stage one applies in this case.
Unfortunately, this kind of data collection Again, a latent variable model Bt = SEi _
is less efficient and therefore loss of infor- fJ + e, is considered, where Bt denotes the
mation can occur. The model can be esti- latent variable that reflects how much
mated employing a discrete approach. For people like tourism, assuming this is a
illustrative purposes, assume the income function of tourism budget decisions, SEi
variable is collected in ten intervals and denotes socio-economic variables corre-
tourism expenditure in five intervals. There sponding to household or individual i and
are 50 different combinations and potential e, denotes the unobserved part of the
groups. It is the researcher's task to regroup model or error term, which is assumed to
all these intervals into new sensible sets, for be normally distributed with zero mean
instance, into five sets. As already men- and unitary variance. The thresholds that
tioned, some information is lost in the pro- determine the range within which category
cess, but these sets may represent how j lies are denoted by JJ.j-1 and JJ.j' The
much people like tourism. In this example, model assumes the individual or household
these sets may group people, such that belongs to category j if JJ.j-1 < e;S JJ.;,
those individuals or households that belong j = 1,..., m. Since Bt = SE;{3 + e.; then substi-
to the first set with a percentage of tourism tuting into the inequality, J.lH < SEi _

Downloaded from thr.sagepub.com at Mount Royal University on June 6, 2015

 Euuemo-Marnn

P+ e,::; JJ.j is obtained, and furthermore, JJ.j-t ler. Extra days in the destinations increase
-SE;P <e;::; JJ.rSE;P. Hence, it can be esti- satisfaction; however, these increases in
mated that the probability that any indivi- satisfaction are less and less relevant as the
dual i belongs to any category j as the number of days in the destination increases,
difference between two cumulative density ie later days do not provide the same satis-
functions as: faction as the first day of travelling. Gener-
ally speaking, this satisfaction is the main
benefit of the traveller. Unfortunately for
the traveller, he or she incurs several costs.
Moreover, the way in which this probabil- Costs can be broken down into fixed and
ity varies under a marginal variation on variable costs. Variable costs are the costs
any socio-economic variable can be incurred by an extra day of travelling,
obtained, for instance, to be a year older or while fixed costs are all the other necessary
to have a baby. It is only necessary to dif- costs of travel that are independent of the
ferentiate probability with respect to a length of stay. For instance, among the
marginal change on SE;: fixed costs are transportation cost, travel
time cost or travel planning cost, and
oPij/oSE; = c/J(JJ.rSE;PH-P) - c/J(JJ.j-t-SE;P) among the variable costs are accommoda-
(- P) = P[c/J(JJ.j-t-SE;P) - c/J(JJ.rSE;P)]· tion cost and other local services such as
food, local transportation and leisure activ-
These marginal effects will reveal how ities. It seems obvious that the optimal
robust the different kind of tourists' enjoy- length of stay depends on how high the
ability is with respect to changes in any of fixed costs are. For instance, if the fixed
the socio-economic variables that define costs are high, stays of a week or two
them. weeks at least are expected, while if fixed
costs are not high, even stays of a weekend
STAGE THREE: FREQUENCY AND could be long enough. It is argued that if
LENGTH OF STAY the individual or household has the flexibil-
Once it is known that a household is parti- ity to set the length of stay, according to
cipating in tourism activities and that a the balance between satisfaction and costs
tourism budget constraint is assigned, the incurred they will determine an optimal
participation analysis can be extended to length of stay. Hence, given a destination
how often the household travels. Unfortu- choice, the length of stay will be optimally
nately, this analysis is not as straightfor- determined depending on the preferences
ward as previous stages. Complexity arises of the individuals. Consequently, in the
from two simultaneously dependent deci- hierarchy of a tourist's decision process, it
sions. Given a budget constraint, indivi- seems plausible that first, tourists decide
duals decide how often to go and how how many trips to make in a period of
long to stay on their trips. These two deci- time, say a year, and then determine the
sions are dependent on each other because destination, where each destination is con-
a longer stay may affect frequency of travel ditioned to an optimal length of stay deter-
and vice versa. mined by each individual. This assumption
Nevertheless, for simplicity, it can be allows the author to concentrate separately
assumed that for every trip, each tourist on frequency of decisions.
possesses an optimal length of stay. The
basis for this assumption is that travelling Objective
usually provides satisfaction to the travel- The purpose of this analysis is to estimate

Downloaded from thr.sagepub.com at Mount Royal University on June 6, 2015

Modelling determinants of tourism demand as a five-stage process

the main factors that contribute to deter- overdispersion by allowing for a specifica-
mine the travel frequency within a period tion that includes an error component
of time. As in stage one, a period of one which represents omitted variables or
year is considered. This kind of study is of unobserved variables. For this purpose,
interest to tourism marketing analysts since negative binomial distribution is usually
it may reveal some new information about applied. It can be seen as a more general
different segments of the market. The expression that also contains Poisson distri-
significance of variables such as age, labour bution as a particular case. The negative
conditions and income are of special inter- binomial distribution imposes E(n;) = A and
est. Var(n;) = A + aA2- k , such that when k = 1:::)
Var(n;) = A + aA, known as the negative
Main variables binomial 1 model, when k = 0:::) Var(nj) =
2
The following variables are expected to be A + aA , known as the negative binomial 2
relevant for the analysis: age, labour condi- model and when a = 0:::) Var(n;) = A the
tions, income, place of residence, size and Poisson model is obtained. An alternative
composition of the household or family, model, known as a zero inflated model,
education, health conditions and unobser- employs a mixing specification which adds
vable variables such as risk aversion and extra weight to the probability of obser-
propensity to travel. ving a zero. The main inconvenience of
these models is that variance is imposed
Methodology exogenously by the researcher. If this var-
The frequency of travel by people follows iance works properly for the model of
a distribution that is skewed to the left and interest it will be sufficient, but sometimes
contains a large proportion of zeros and excess zeros may not be associated with
ones. The dependent variable is a non- increased dispersion but with an underlying
negative integer-valued count and there- tourist behaviour. Unobservable patterns
fore count data methodology can be are referred to in the behaviour of tourists,
employed to estimate frequency of travel. such as fear of flying or a high propensity
A classical model for count data is Pois- to travel. If a researcher believes this
son process: P(n;) = A"'e-)"!n;!, where n, behaviour is relevant enough, then it
denotes number of times that a household requires a more complex model for the
or individual i travels during a fixed inter- analysis.
val, say for instance, a year; P(n;) denotes In order to deal with unobservable
the probability that a household or indivi- heterogeneity two different approaches will
dual j travels n, times and Aj = E(tl;l- be considered. On the one hand, the popu-
SE;} = exp(SE;P). An important feature of lation can be split according to the partici-
the Poisson model is that it imposes pation decision of stage one, assuming that
E(tl;!SE;) = Var(n;ISE;) = A, which is known frequency depends on two separate pro-
as the equidispersion property. In some cesses (one followed by those people who
cases this property may be true but in decided not to travel and another process
other cases it is violated. If the variance is that considers all those people who do
greater than the mean, here is a case of travel). In the econometric literature this is
overdispersion and the Poisson model will known as the hurdle model.
tend to underpredict the current frequency On the other hand, the population can
of zeros. Mullahy (1997) associates overdis- be split according to how much they like
persion with the existence of unobservable tourism, information already obtained in
heterogeneity. He suggests dealing with stage two. In this case, each segment fol-

Downloaded from thr.sagepub.com at Mount Royal University on June 6, 2015

 ELJ[jefliO Mar un

lows an independent process (Deb and Tri- kinds of need that different tourists
vedi (1997». demand. It is important to define properly
The independent processes of both of the choice set because it will affect the effi-
these mechanisms are summed up in the ciency of the estimates.
log-likelihood final function in order to be
estimated by maximum likelihood. This is Main variables
the reason why these models are known as Besides the socio-economic and demo-
mixture models. graphic variables, it is necessary to create a
tree structure that classifies different kinds
STAGE FOUR: KIND OF DESTINATION of destinations depending on their physical
In previous stages, participation decision, attributes and the kind of tourist en viron-
tourism budget decision and frequency of ment they possess.
travelling have been studied. Once an indi-
vidual or household has chosen a frequency Methodology
for travelling, they must decide where to An illustration follows of different criteria
go on their different trips. Nowadays, for to classify the different kinds of destination
most of the developed countries, the choice to which tourists may wish to go:
set for travel is quite large. There are mul-
tiple available destinations. Each of them (a) Physical attributes of destination
possess special features, i.e. some particular - a. Seaside
characteristics that make them unique. - b. Countryside
There are some destinations that satisfy - c. Mountain
almost every kind of general need, but pos- - d. City
sibly do not satisfy the needs of a minority
of tourists. Once again, the researcher faces In this case, the tree structure possesses four
a problem of heterogeneity. Tourists differ branches and the researcher needs to deter-
in what they consider an ideal destination mine in which of these branches each desti-
and which are the needs they want to be nation must be classified.
satisfied.
Within this stage tourists make two (b) Tourist environment
simultaneous decisions. One decision con- - a. Familiar
cerns the nature of the travel. In this - b. Cultural
sense, the tourist must decide which kind - c. Relaxing
of travel is desired; for instance, a familiar, - d. Party
adventurous or relaxing trip. The other - e. Adventure
decision concerns the choice of the kind of
destination they prefer to visit in terms of Similarly to case (a), the researcher needs to
its physical attributes. In other words, for classify each potential destination according
example, tourists may decide if the kind to these five branches.
of destination they prefer is a mountain, a
city, a countryside resort or a seaside (c) Both combined hierarchically
resort. In this case, the tree structure has two
levels. In the first stage, tourists might
Objective choose a kind of destination, according to
This stage is prior to the final destination criterion (a) or (b) and in the second stage
choice and helps in defining more appro- they might choose a kind of tourist envir-
priately the choice set for the different onment, according to criterion (b) or (a).

Downloaded from thr.sagepub.com at Mount Royal University on June 6, 2015

Modelling determinants of tourism demand as a five-stage process

In any case, we end up with 20 different when it visits destination d. Thus, the prob-
kind of tourist destinations. ability that a household h chooses to travel
to kind of destination dis:
(d) Both combined simultaneously
Similarly to previous classification, this PI,d= Pr(Vhd> Vh,'Vr:t=d) = Pr(Vhd+ thd >
proposition does not consider stages but Vii, + Gh, 'V r:t= d) = Pr(GIIT - Ghd < V h,-
the tourist must decide simultaneously Vh,'Vr:t=d)
among the 20 different kinds of tourist des-
tinations. For convenience, it is assumed that Ghd is
The choice of any of these alternative independent and identically distributed
ways of classifying tourist destinations is extreme value. The advantage of this
relevant for the methodology of final desti- assumption is that the normalisation
nation choice. Attending to criteria (a), (b), required in any discrete choice model, ie
or (d), it is possible to model the kind of Gild, = GIlT - Gild, requires that the error differ-
tourist destination decision with a multino- ences Gird' are distributed logistically, imply-
mial logit model. This is shown below. ing that, after algebraic manipulation
Criterion (c), can be modelled with a (Train (2003)), the probability formula can
nested multinomial logit model, as briefly be obtained exactly as:
discussed in the next section.
Any household, labelled h, may decide
to travel or not. Once the household has
decided to travel, it has to choose the kind Nevertheless, kinds of tourist destination
of tourist destination it wishes, labelled d, choice can be linked with final destination
among its choice set. In order to model the choice through a nested multinomial logit
kind of destination choice, a behavioural model. This is the purpose of the last stage.
model is followed where the household
chooses the alternative that provides the STAGE FIVE: DESTINATION AND
higher level of utility, denoted by U. In TRANSPORTATION MODE CHOICE
this sense, household h would choose kind One of the most interesting decisions, from
of destination d if, and only if: the researcher's point of view, is final desti-
VI"I > V",'Vr:t=d, where r denotes any other nation choice. As discussed, the researcher
kind of resort. Nevertheless, these utility needs to deal with the heterogeneity of
levels are unobservable for the researcher. tourists in order to obtain accurate esti-
The only aspects known are some socio- mates from the model proposed. Previous
economic variables of the household, stages provide additional information to be
denoted by SEll, and some attributes of the included in the analysis of this final stage
set of kind of destinations, denoted by Ad. that help to obtain more efficient results.
From the information available, it is pos- From stage two information can be
sible to construct a function V lJd = included as to how much each individual
V(SE", Ad) 'V d, which represents the utility likes tourism; from stage three, if signifi-
that kind of destination d provides to the cant, information can be provided on pro-
individual. Obviously, this representative pensity to travel; and stage four provides
utility VI"I is an approximation to the cur- the methodological structure and minimi-
rent utility VI"I' Therefore, it can be stated sation of heterogeneity among tourists.
that the utility can be decomposed as: It is assumed that when tourists deter-
VI"I = Vl,d + 8hd, where 8hd denotes the mine a destination to visit, this decision is
unobserved part of utility for household h linked to the transportation mode choice.

Downloaded from thr.sagepub.com at Mount Royal University on June 6, 2015

 Euqeruo-Martrn

In long-distance international tourism, the Methodology

plane is the expected. transportation mode As mentioned in the previous section, the
and usually for this case, the decision about methodology proposed links stages four
transportation mode is trivial. There are and five and employs a nested multinomial
many other destinations, however, with logit model as used by Eymann and Ron-
multiple options for travel. To capture this ning (1997) in their study of German out-
effect, every sensible combination between bound tourism demand. This is a superior
destination and transportation mode is model to the multinomial logit model
included in the tourist's choice set. because it deals more appropriately with
unobservable heterogeneity. Moreover, the
Objective multinomial logit model imposes, by con-
This final stage focuses on the estimation of struction, a restriction known as indepen-
the main determinants of tourists' destina- dent of irrelevant alternatives (IIA)
tion choice. This is the more complex stage property. For the purposes of this paper,
in terms of variables that may influence this property may be inconvenient. More
tourists in which the characteristics of the precisely, IIA implies that, given a change
household or individual and attributes of in any attribute of the alternatives, cross
the destinations are considered. The pur- elasticities of the probabilities of choosing
pose is to determine the relative impor- any alternative are exactly the same for
tance of any of these variables with respect every alternative. If the variables that
to the final decision. Moreover, it is desir- explain the behaviour of tourists are stable
able to obtain a framework which may over time this implication is not proble-
allow a simulation of how the current dis- matic as is the case in stage four. In stage
tribution of tourists would be affected by five, however, it is likely that attributes of
any change in any variable of interest. the destinations vary quite often. More-
over, if it were desired to simulate changes
Main variables in these attributes, the same cross elasticities
The main variables expected to be relevant assumption would create a bias in the
are shown in Table 1. results. Nested multinomial logit deals

Table 1: Main variables of stage five: Destination and transportation mode choice

Characteristics of the household Attributes of the destinations Mixed variable

Disposable income Relative prices (PPP): Transportation cost

Budget for tourism expenditure prices and exchange rate Travel time cost
Disposable time for tourism Accommodation cost index (mode of transportation)
Labour conditions Weather Available information
Frequency of travelling Safety Language
Size of the household Crowding Suitability of destination (loss
Age of the oldest and youngest Development and facilities function)
members of the household Size of the country Marketing in the country of
Education origin
Place of residence
Size of community: rural or city
Risk aversion
Party size of travellers

Downloaded from thr.sagepub.com at Mount Royal University on June 6, 2015

Modelling determinants of tourism demand as a five-stage process

with IIA, such that it allows for different set and Haab and Hicks (2000), provide a
cross elasticities between different nests but survey of the way that choice sets have been
not within the same nest. Generally speak- considered in recreation demand models.
ing, this specification would be flexible Researchers might also think about the
enough to obtain sensible simulations. flexibility of the choice set. This can be
Nevertheless, more flexible alternative predetermined and fixed by the researcher
models can be applied. If IIA is still a rele- or it can be endogenously determined by
vant problem in the model, the researcher the model, ie defining a particular choice
may use heteroskedastic extreme value set for each individual. Endogeneity may
models (HEV). Furthermore, if the be explained, among other variables,
researcher wishes to apply full flexibility to through level of information, number of
the estimates, such that the model can pro- trips or age. General awareness of tourist
vide particular estimates for each kind of destinations is likely to widen the choice
representative individual, a random para- set. Four levels of awareness can be consid-
meters model or mixed logit model can be ered: international, national, regional and
employed. The inconvenience of these local. The more an individual travels to
models is that probabilities cannot be places at any of these levels, it is expected
obtained with closed-form expression inte- that the wider his or her choice set will be.
grals but integrals require simulation and Finally, it is also necessary to adjust the
consequently the estimation process is size of the destinations, so that most of
more complicated. them are homogeneous, otherwise there
A specific problem of this stage that may be a bias in the effect of the attributes.
requires special care is the definition of the For instance, while France and Switzerland
choice set. Three aspects must be taken may both have very similar attributes in
into account: criteria to define different the Alps, France has many more visitors
alternatives; flexibility; and size of the alter- because it is a bigger country. In this sense,
native destinations. Concerning the defini- the attributes of Switzerland may be
tion of different alternatives, it is possible undervalued compared to France due to
to follow different criteria. Destinations the differences in the size of both countries.
can be considered as a political division; a It is necessary either to split the country
natural division, in terms of kind of terri- into homogeneous destinations in terms of
tory or weather; and a division according size, or to adjust the number of arrivals
to the kind of activities that can be prac- according to the size of the country (arri-
tised in the destination. vals per square km).
Since the tourism literature is still limited
in this area, benefit can be gained from CONCLUSIONS
examining outdoor recreation studies. For The methodology proposed assumes that
instance, Parsons and Hauber (1998) tourists' destination choice is conditioned
showed that for recreational fishing trips, 94 by another four decisions. It is argued that
per cent of individuals choose sites within a this decision process follows a hierarchy
distance of one hour and a half travel time. with five stages: participation decision;
Consequently, for this kind of recreation tourism budget decision; frequency and
spatial boundaries for choice set must be length of stay decisions; kind of tourist des-
established in order to optimise the effi- tination decision; and final destination and
ciency of the estimation. Moreover, Thill mode of transportation choice. This struc-
(1992) is of interest because he considers dif- ture responds to the necessity to deal with
ferent approaches to capture the true choice the heterogeneity among tourists.

Downloaded from thr.sagepub.com at Mount Royal University on June 6, 2015

 Euqenlo-Martin

For the first stage, the participation deci- sponds to the destination and transporta-
sion, the employment of a probit or logit tion mode choice. It is proposed to include
model is suggested. From this model the in the tourist's choice set, every sensible
probability of whether an individual or combination of transportation mode for
household will travel can be estimated and relevant destinations. Special care needs to
the effects on this probability of any be considered with the choice set definition
change in any socio-economic variable. for tourism analysis. In this sense, three
In the second stage, the objective is to main aspects were taken into account: cri-
determine the main factors that push differ- teria to define different alternatives, flex-
ent individuals to spend part of their ibility and size of the alternative
budget in tourism activities. Rather than destinations.
estimate tourism expenditure in absolute The result is a complete methodological
terms the author prefers to model the per- framework that disaggregates tourists' deci-
centage of tourism expenditure over sions. This framework allows the author to
income. This transformation offers more obtain the determinants for each of these
relevant results and allows an estimate of decisions and to simulate how the current
how much people like tourism. situation may change under alternative
The third stage analyses the main factors scenarios.
that contribute to define travel frequency
within a period of time. In this stage the ACKNOWLEDGMENTS
role of the length of stay in the frequency The author would like to thank Noelia
of decisions is discussed. It is concluded Martin, Peter Simmons and Roberto Leon
that, provided individuals or households for their help and useful comments.
posses some flexibility to choose the length Thanks also go to the participants of the
of stay, then this optimally will be selected conference of Tourism Research 2002:
depending on the destination chosen and Riccardo Scarpa, Casiano Manrique, Gab-
the amount of fixed costs incurred. For fre- riel Eugenio and Sara Gonzalez and to
quency decisions, the model suggested Richard Butler. The usual disclaimer
depends on the dispersion of the frequency. applies.
If frequency is equidispersed, a Poisson
process may be applied. If frequency is REFERENCES
overdispersed, however, this may need to Bakkal, I. (1991) 'Characteristics of West
be modelled because it would respond to a German demand for international tourism
case of unobservable heterogeneity. For in the northern Mediterranean region',
this purpose, a negative binomial model, a Applied Economics, 23. 295-304.
zero-inflated model, a hurdle model and a Deaton, A. and Muellbauer, D. (1980a) 'An
mixture model are discussed. almost ideal demand system' , America"
In the fourth stage, a tree structure is Economic Review, 70, 3. 312-26.
proposed that may classify different seg- Deaton, A. and Muellbauer, D. (1980b) 'Eco-
nomics and consumer behavior', Cam-
ments of tourists depending on their needs.
bridge University Press, Cambridge.
Two criteria which define this classification
Deb, P. and Trivedi, P.K. (1997) 'Demand for
are the physical attributes of the destination medical care by the elderly: A finite mix-
and the kind of tourist environment ture approach', [ouma! of Applied Econo-
wished. The methodology proposed in this metrics, 12. 313-36.
stage is linked with the last stage, such that Divisekera, S. (2003) 'A model of demand for
both decisions are modelled within a nested international tourism', Annals 4 Tourism
multinomial logit. The last stage corre- Research, 30, 1, 31-49.

Downloaded from thr.sagepub.com at Mount Royal University on June 6, 2015

Modelling determinants of tourism demand as a five-stage process

Eymann, A. and Ronning. G. (1997) 'Micro- in

Europe: an econometric analysis'.
econometric models of tourists' destina- Applied Economics, 16.919-31.
tion choice'. Regional Science and Urban Papatheodorou, A. (1999) 'The demand for
Economics, 27. 735-61. international tourism in the Mediterra-
Greene, W. (2003) 'Econometric Analysis', 5th nean region', Applied Economics, 31. 619-
edn, Prentice Hall, New Jersey. 30.
Haab, r.c. and Hicks, R.L. (2000) 'Choice set Parsons. G.R. and Hauber, A.B. (1998) 'Spatial
considerations in models of recreation boundaries and choice set definition in a
demand: History and current state of the random utility model of recreation
art', Marine Resource Economics, 14, 271-81. demand'. Land Economics. 74, 1, 32----48.
Hultkrantz, L. (1995) 'On determinants of Rugg, D. (1973) 'The choice of journey desti-
Swedish recreational domestic and out- nation: A theoretical and empirical analy-
bound travel, 1989-93', Tourism Econom- sis', Review of Economics and Statistics. 55,
ics, 1. 2, 119-45. 64-72.
Maddala, G. (1983) 'Limited dependent and Syriopoulos, T.e. and Sinclair, M.T. (1993)
qualitative variables in econometrics', 'An econometric study of tourism
Cambridge University Press. New York. demand: the AIDS model of US and
Morley, e.L. (1995) 'Tourism demand: charac- European tourism in Mediterranean coun-
teristics, segmentation and aggregation'. tries'. Applied Economics. 25. 1541-52.
Tourism Economics, 1, 4, 315-28. Thill, J. (1992) 'Choice set formation for desti-
Mullahy.]. (1997) 'Heterogeneity. excess zeros. nation choice modelling'. Progress in
and the structure of count data models'. Human Geography. 16, 3, 361-82.
Journal of Applied Econometrics. 12, 337-50. Train. K. (2003) 'Discrete choice methods with
O'Hagan. ].W. and Harrison, M.J. (1984) simulation'. Cambridge University Press.
'Market shares of US tourist expenditure Cambridge.

Downloaded from thr.sagepub.com at Mount Royal University on June 6, 2015