MODELLING DETERMINANTS OF TOURISM DEMAND AS A 5-STAGE PROCESS. A DISCRETE CHOICE METHODOLOGICAL APPROACH. Juan L. Eugenio-Martin

Transcription

1 MODELLING DETERMINANTS OF TOURISM DEMAND AS A 5-STAGE PROCESS. A DISCRETE CHOICE METHODOLOGICAL APPROACH Juan L. Eugeno-Martn Departamento de Análss Económco Aplcado Unversdad de Las Palmas de Gran Canara Facultad de Cencas Económcas y Empresarales. Módulo D. C/ Saulo Torón. CP Las Palmas de Gran Canara. Span. Phone: Emal: jlus@empresarales.ulpgc.es Envronment Department Unversty of York Heslngton, York. YO10 5DD. Unted Kngdom. Phone: + 44(0) Emal: jlem100@york.ac.uk Juan Lus Eugeno-Martn s currently undertakng PhD n Economc Analyss at Unversdad de Las Palmas de Gran Canara (Span) and PhD n Envronmental Economcs and Envronmental Management at Unversty of York (UK).

2 MODELLING DETERMINANTS OF TOURISM DEMAND AS A 5-STAGE PROCESS. A DISCRETE CHOICE METHODOLOGICAL APPROACH Abstract In the toursts destnaton choce there are multple factors that affect ther decson. Indvduals or famles wth exactly the same socoeconomc and demographc characterstcs may choose very dfferent destnatons. The paper deals wth ths heterogenety problem recognsng there are taste dfferences among toursts and that fnal destnaton choce s not an ndependent decson, but just last decson of a set of choces that are also determnng t. In ths sense, we argue that toursts face a 5-stage decson process. Frst of all, people have to decde whether or not travel wthn a perod of tme. Second, those who expect to travel need to estmate a budget for toursm expenses. Thrd, gven the budget, they need to determne frequency and length of stay of ther trps. Fourth, once a date and length of the stay s proposed, toursts need to choose whch knd of tourst destnaton wshes to vst. And fnally, among all the avalable destnatons that satsfy tourst s condtons, fnal destnaton and mode of transportaton are chosen. It s the purpose of ths paper to propose a methodologcal framework for modellng each of these stages and ther relatonshp. Keywords: Toursm demand, Outbound toursm, Dscrete choce models, Toursts decsons, Toursm marketng

3 1. INTRODUCTION For many regons toursm has become one of the most sgnfcant economc actvtes n terms of economc growth and employment. World toursm demand s stll growng and new or current destnatons may be developed or extended n order to satsfy such a growth. In ths sense, toursm may be been seen as an opportunty of economc growth for developng regons. However, despte toursm demand s stll growng, toursm supply s also growng n the same fashon and competton among destnatons arses. Destnatons may deal wth competton from a two-level perspectve. On the one hand, from a mcro-level pont of vew, hosptalty sector may change prces and qualty and adapt the servces offered to the preferences of ther potental vstors. On the other hand, from a macro-level pont of vew, local, regonal or natonal authortes may nvest on the development of tourst resorts and promote them. In order to be effcently appled, any of these polces requres a deep knowledge of the characterstcs of potental vstors, ther needs and the nterrelatons wth other compettve destnatons. Therefore, the queston that needs to be answered s why people travel to dfferent places. Ths paper ams to focus n ths way. The man purposes are related wth toursm marketng and toursm plannng decsons from both mcro and macrolevel ponts of vew. More precsely, we try to provde a methodologcal framework whch may analyse the relatve mportance of dfferent attrbutes for toursts destnaton choce; estmate probablty of vstng each knd of destnaton for dfferent knd of toursts; and fnally, generate a tool that lets us smulate changes n the demand under alternatve scenaros. In the toursts destnaton choce there are multple factors that affect ther decson. For nstance, age, labour condtons and ncome are usually condtonng toursts choces. However, what makes modellng destnaton choce even more challengng s the fact that ndvduals or famles wth exactly the same socoeconomc and demographc characterstcs

4 may choose very dfferent destnatons. The paper deals wth ths heterogenety problem recognsng there are taste dfferences among toursts and that fnal destnaton choce s not an ndependent decson, but just last decson of a set of choces that are also determnng t. In ths sense, we argue that toursts face a 5-stage decson process. Frst of all, people have to decde whether or not travel wthn a perod of tme 1. Second, those who expect to travel need to estmate a budget for toursm expenses. Thrd, gven the budget, they need to determne frequency and length of stay of ther trps. Fourth, once a date and length of the stay s proposed, toursts need to choose whch knd of tourst destnaton wshes to vst. And fnally, among all the avalable destnatons that satsfy tourst s condtons, fnal destnaton and mode of transportaton are chosen. It s the purpose of ths paper to propose a methodologcal framework for modellng each of these stages and ther relatonshp. 2. TOURISTS DESTINATION CHOICE PROCESS When studyng toursm demand we can consder two ponts of vew. On the one hand, we can forecast number of toursts that are expected to arrve to a partcular destnaton;.e. we consder nbound toursm. But on the other hand, we can try to understand tourst destnaton choce of the nhabtants of a partcular regon;.e. we analyse outbound toursm. The purpose of ths paper s to provde a methodologcal framework whch may estmate the man determnants of outbound toursm demand. As t s mentoned n the ntroducton of ths paper, we consder that before decdng where to go on holdays, most of the toursts need to make multple decsons. Whle for some people these decsons are perfectly planned, for other people these are mprovsed or hardly planned. Moreover, some people can decde all of them smultaneously or n dfferent stages. Anyway, 1 For our purposes we assume to be a one year perod.

5 ether f t has been planned or not, we argue that most of the people consder conscously or unconscously a process of decsons concernng ther holdays trps. For modellng purposes, we assume that toursts choose fnal destnaton dependng on another four decsons. In ths sense, we consder fve stages: partcpaton decson; toursm budget decson; frequency and length of stay decsons; knd of destnaton decson; and fnal destnaton and mode of transportaton choce. Methodology proposed s a general framework applcable to any place n the world of any sze. For nstance, t may be employed for a country, a regon or a small town or vllage. Obvously, the larger the regon analysed s, the more heterogenety we wll face. The man objectve of the methodology proposed s to deal wth ths heterogenety problem. In order to apply ths methodology we need two dfferent datasets. On one hand, we need mcro data on socoeconomc and demographc characterstcs of a representatve sample of populaton. Ths dataset must also nclude data on toursm trps, as for nstance, places vsted, number of trps, length of stay or expendture on toursm. On the other hand we need data on the attrbutes of choce set. Ths s an objectve dataset and t s easer to obtan. It usually ncludes varables as accommodaton cost ndex, prce ndex, development level or temperature. Concernng the sample we need to comment some ponts. Samplng can be covered from the whole populaton, from on-ste or from a combnaton of both. The man advantage of on-ste samplng s that deeper and wder varablty may be obtaned compared to populaton samplng. Although for a destnaton choce analyss, on-ste samplng mght be more

6 convenent, for our purposes t would be ncomplete because we may loose nformaton concernng the reasons why people decde to travel or not. Another ssue related wth samplng s the perod of tme the study covers. Ths could range from a season, a year or a set of years. The perod chosen depends on the purposes of the analyss. A perod of a season mght be chosen f the regon analysed s remarkably affected by the season and ths effect s relevant. A longer perod than a year s useful f researcher wants to trace tourst s behavour over years. It s very nterestng because t lets to reveal aspects as repetton patterns, rsk averson and toursm budget decson makng. However the nconvenence s the possble mstakes ntervewees may commt especally concernng wth data related wth trps made more than three years ago. A soluton s to generate a panel dataset that trace ndvduals over tme repeatng ntervew for each perod of tme. However, for a general purpose, a perod of one year tme seems to be sensble 2. Next secton deals wth each of the stages, explanng the objectves, varables consdered, methodology proposed for the analyss. Moreover alternatve methodologes are brefly dscussed. 3. STAGE 1: PARTICIPATION DECISION Introducton The frst decson any ndvdual has to make, concerns the choce between travellng or not wthn a perod of tme. Researcher, dependng on the purposes of the analyss, must set a tme nterval. Usually, ths may range from a perod of a year to a perod that contans the whole lfe of the ntervewees. 2 One year perod s also recommended by Morley (1995).

7 Objectve Manly, ths s an nterestng ssue for toursm marketng analysts, because t shows whch the determnants of the decson to travel are. Smlar models can be constructed dependng on socoeconomc or demographc characterstcs of ndvduals. The objectve s to estmate dfferent models attendng to a segmentaton crteron, such that we can compare the results from dfferent models and draw some conclusons. For nstance, we can estmate determnants of partcpaton for dfferent places of resdence. Ths analyss may reveal that resdents of a partcular regon are less nterested n travellng than resdents of other regons due to several aspects, as ncome dfferences or dfferences n especal facltes for recreaton, whch help resdents to enjoy holday s tme n ther own place of resdence. Moreover, we can make segmentaton n the sample attendng to any varable and compare the relatve mportance for each segment of any varable. A common case to be analysed s the effect of the age. We can estmate relatve mportance of partcular problems of each segment. For nstance, we can compare how mportant s the ncome for youth segment wth respect to the others, or how relevant s to have a chld or not for md-age toursts decson or how sgnfcant are health condtons for elderly people s decson. Marketng effort can focus on dfferent segments attendng to ther man determnants for travellng. Man varables Intuton suggests that varables as age, educaton, ncome, labour condtons, characterstcs of the place of resdence and sze and composton of the household or famly, may be sgnfcant when decdng to travel or not. Methodology

8 In order to model partcpaton decson, we consder t s a bnary choce, denoted byt, such that, T = 1 f household or ndvdual decdes to travel and T = 0 otherwse. We want to model probablty that 1 T =,.e. Pr( T = 1). We assume Pr( T = 1) s lnked to a set of exogenous varables, whch may be those already shown above. More precsely, for some approprate functon () j g, Pr( 1) k T = = g α + β SE j j j= 1, where g () 0 1, α denotes a constant, SE denotes j th socoeconomc varable of household or ndvdual and β denotes assocated j parameter to j th socoeconomc varable. Tradtonal lnear probablty model s not recommended to be used to estmate the probablty functon because t would present non normal errors, heteroskedastcty and logcal nconsstency, snce predcton of probabltes may le out of range (0,1). It s well-known that the suggested model for bnary choce estmatons s latent varable model. Ths model consders the exstence of a latent varable T *. Snce ths latent varable s unobserved by the researcher we can consder t s composed by two parts: one observed by the researcher, whch ncludes all the socoeconomc varables and another part that t s unobserved by the researcher and that corresponds to heterogenety reasons among toursts. Thus the model can be represented as: T * = α + β jsej + ε j= 1 k, where ε denotes unobserved part or error term. For our purposes, the latent varable wll work as an ndex functon, such that we wll set T = 1 f * T > 0 and * T = 0 f * T 0. * Let S k = α + β SE, such that j j j= 1 T = S + ε. * Then, Pr( T 1) Pr( S ε 0) Pr( ε S ) 1 Pr( ε S ) 1 F ( S ) = = + > = > = =, where F ε ε denotes cumulatve densty functon of unobserved part. Due to a problem of dentfcaton of

9 locaton and scale of * T, researcher needs to choose a dstrbuton and a value for the varance of ε. The most common approaches assume ε s ndependently and dentcally dstrbuted, ether followng a normal dstrbuton wth zero mean and varance of one, or followng a logstc dstrbuton wth zero mean and varance of π 3 2. If we assume that ε follows the former dstrbuton we are employng the well-known probt model, and f we assume the latter dstrbuton we are employng the also well-known logt model. Any of these dstrbutons can be employed for the partcpaton decson and both present smlar results. Fnally, maxmum lkelhood estmaton s appled to the model n order to estmate parameters of nterest. Under correct specfcaton, these estmates are consstent and asymptotcally normal STAGE 2: TOURISM BUDGET CONSTRAINT Introducton Once a household or ndvdual has decded to travel, they have to decde how much toursm expendture may be. Ths decson depends manly on ncome and preferences of ndvduals. If the analyss s performed wth ncome n absolute terms, thus ths varable s lkely to domnate the estmated regresson. In order to avod a trval result, we can estmate toursm budget as a percentage of ncome,.e. as a rato between toursm expendture and ncome. Ths new formulaton lets us to estmate how much people prefer to dstrbute ther ncome for toursm purposes, or n other words, how much people lke toursm. 3 For a complete exposton of the methodology see Greene (2003).

10 Objectve The man purpose of ths stage s to try to understand the man factors that push dfferent ndvduals to spend part of ther budget n toursm actvtes. Man varables We are nterested n estmatng the relatve mportance that varables as age, ncome, labour condtons, place of resdence and sze of the household or famly possess n toursm budget decsons. Methodology Dependng on the nature of the data we can apply dfferent methodologes. Ideally, both ncome and toursm expendture mght be contnuous varables. However, t s lkely that the questonnares consder dscrete ntervals for these two varables. For the contnuous varables case, we can estmate toursm expendture as dependent on ts own determnants as well as the demand of another goods and servces. A tradtonal approach s to estmate ths toursm expendture as a demand functon that s part of a system of demands whch nclude all other goods and servces. Deaton and Muellbauer (1980a) dealt wth ths ssue developng a model known as Almost Ideal Demand System. Toursm lterature has employed ths approach frequently 4. However, as far as we know, all the studes have used macroeconomc varables rather than mcro data and the nconvenences already commented n prevous sectons apples. Mcro data can be used to estmate ths system of demands, but lack of data has usually been a common problem.

11 Nevertheless, we can assume that the whole set of commodtes can be dvded nto dfferent groups, such that preferences wthn groups can be descrbed ndependently of the quanttes n other groups. Ths assumpton s known as weak separablty 5 of the utlty functon. It s a plausble and common assumpton n the toursm economcs lterature 6. Under ths assumpton we can create a tree of commodtes, where one of the man branches may be a group called entertanment, and toursm may be a further branch lnked to entertanment. Ths structure allows us to analyse toursm expendture allocaton ndependently of expendture levels n any other goods and servces 7. Consequently, under weak separablty assumpton, we can model toursm budget decson dependng on socoeconomc varables already mentoned. However, snce toursm budget s measured as a percentage of total ncome, such that ts value must le between zero and one, lnear probablty model cannot be appled for the same reasons we have already explaned n the partcpaton decson model. Several solutons can be adopted. A tradtonal way s to employ a translog model. Ths model usually works approprately for ndependent functons. However, f model s used for a system of demand functons, t usually fals n testng addtve and homothetc restrctons 8. Another way to model toursm budget s through a double censored regresson model. In ths case, researcher can mpose two censors to the potental freedom gven by default to any endogenous varable n the tradtonal regresson analyss. Obvously the lower censor mght be set on zero and the upper censor on one, such that any estmated value s guaranteed to le between these two values 9. 4 See O Hagan and Harrson (1984), Syropoulos and Snclar (1993), Papatheodorou (1999) or Dvsekera (2003). 5 For a further reference see Deaton and Muellbauer (1980b), pp: See for nstance Rugg (1973). 7 See Hultkrantz (1995), who consders a three-stage budgetng framework. 8 See Bakkal (1991).

12 As we have mentoned, t s lkely that mcro data on ncome or toursm expendture be collected n the ntervew process, as the choce among alternatve predetermned ntervals, rather than an exact number. Unfortunately, ths knd of data collecton s less effcent and therefore loose of nformaton occurs. Anyway, we can estmate the model employng a dscrete approach. For llustratve purposes, say we have ncome varable collected n ten ntervals and toursm expendture n fve ntervals. Therefore, we have ffty dfferent combnatons and potental groups. It s researcher s task to group all these ntervals nto new sensble sets, say for nstance, fve sets. As we have mentoned, we loose nformaton n the process, however, these sets may represent how much people lke toursm. In ths example, these sets may group people, such that those ndvduals or households that belong to the frst set wth percentage of toursm expendture over ncome, between zero and lowest threshold determned by researcher, means these ndvduals do not lke toursm. As long as nterval of the set corresponds to hgher percentage, researcher can defne other labels as do not lke toursm much, lke toursm, lke toursm much and fnally, lke toursm very much. In order to deal wth ths ordered categorcal varable, we can use what s known as an ordered probt model. In the ordered probt model, we have dfferent ordered multnomal outcomes, denoted by j, where j = 1... m. In our example the categores j = { do not lke toursm, lke toursm a bt, lke toursm, lke toursm much and lke toursm very much }. Tradtonal ordered probt model estmate thresholds tself, however, snce researcher may predetermne them, as we have already commented, usng percentage of toursm expendture over ncome, then we can employ a superor regresson known as nterval or grouped data regresson. Ths s no more than a varant n whch the values of the thresholds are known. Because the thresholds are 9 For a complete analyss of ths technque, see Maddala (1983) or Greene (2003).

13 known, the estmates of the parameters are more effcent and t s possble to dentfy the varance of the error term. Smlar justfcaton to the model presented n stage 1 apples n ths case. Agan, we consder a latent varable model B = SEβ + ε, where * * B denotes the latent varable that reflects how much people lke toursm assumng ths s a functon of toursm budget decsons, SE denotes socoeconomc varables correspondng to household or ndvdual and ε denotes unobserved part of the model or error term, whch t s assumed to be normally dstrbuted wth zero mean and untary varance. If we denote the thresholds whch determne the lmts wthn category j les by µ j 1 and µ j. The model assumes ndvdual or household belongs to category j f µ < B µ, j = 1,..., m. Snce * B = SEβ + ε, then substtutng nto the * j 1 j nequalty, we obtan µ < j 1 SEβ + ε µ, and furthermore, j µ j 1 SEβ < ε µ j SEβ. Hence, we can estmate probablty that any ndvdual belongs to any category j as the dfference between two cumulatve densty functons as: Pj ( µ j SEβ ) ( µ j 1 SEβ) =Φ Φ. Moreover, we can obtan how vares ths probablty under a margnal varaton on any socoeconomc varable. As for nstance, to be a year older or to have a baby. We need no more than to dfferentate probablty wth respect to a margnal change on SE : P j SE ( j SE )( ) ( j 1 SE )( ) ( j 1 SE ) ( j SE ) = φ µ β β φ µ β β = β φ µ β φ µ β. These margnal effects wll reveal us how robust enjoyablty of dfferent knds of toursts s wth respect to changes n any of the socoeconomc varables that defne them.

14 5. STAGE 3: FREQUENCY AND LENGTH OF STAY Introducton Once we know that a household s partcpatng n toursm actvtes and that a toursm budget constrant s assgned, we may extend the partcpaton analyss to how often the household travels. Unfortunately, ths analyss s not as straghtforward as prevous stages are. Complexty arses from two smultaneously dependent decsons. Gven a budget constrant, ndvduals decde how often and how long stay n ther trps. These two decsons are dependng on each other because a longer stay may affect frequency of travellng and vce versa. Nevertheless, for smplcty, we may assume that for every travel, each tourst possesses an optmal length of stay. Let me explan the bass for ths assumpton. Usually travellng provdes satsfacton to the travellers. Extra days n the destnatons ncrease satsfacton, however, these ncreases n satsfacton are less and less relevant as number of days n the destnaton ncreases,.e. t does not provde the same satsfacton the frst day of travellng than any of the last days of the travel. Generally speakng, ths satsfacton s the man beneft of the traveller. Unfortunately for the traveller, he or she ncurs n several costs. Costs can be decomposed between fxed and varable costs. Varable costs are the costs ncurred by an extra day of travellng, whle fxed costs are all other necessary costs to travel that are ndependent of the length of stay. For nstance, among the fxed costs we can fnd transportaton costs, travel tme costs or travel plannng costs, and among the varable costs we can fnd accommodaton and other local servces as food, local transportaton and lesure actvtes. It seems obvous that optmal length of stay depends on how large fxed costs are. For nstance, f fxed costs are large, stays of a week or two weeks are expected, whle f fxed

15 costs are not large, stays of a weekend may be long enough. We argue that f ndvdual or household has flexblty to set length of stay, accordng to the balance between satsfacton and costs ncurred they wll determne an optmal length of stay. Hence, gven a destnaton choce, length of stay wll be optmally determned dependng on preferences of the ndvduals. Consequently, n the herarchy of tourst s decson process, t seems plausble that frst, toursts decde how many trps to make n a perod of tme, say a year, and then determne the destnaton to go, where each destnaton s condtoned to an optmal length of stay determned by each ndvdual. Ths assumpton allows us to concentrate on frequency decsons separately. Objectve The purpose of ths analyss s to determne the man factors that contrbute to travel frequency wthn a perod of tme. As n stage 1, we consder a perod of one year. Ths knd of study s of nterest to toursm marketng analysts snce t may reveal some new nformaton about dfferent segments of the market. It s of specal nterest the sgnfcance of varables as age, labour condtons and ncome. Man varables The followng varables are expected to be relevant for the analyss: age, labour condtons, ncome, place of resdence, sze and composton of the household or famly, educaton, health condtons and unobservable varables as rsk averson and propensty to travel. Methodology If we have a look to the frequency of travellng of people, we can see ths follows a dstrbuton that s skewed to the left and contans a large proporton of zeros and ones.

16 Besdes ths dstrbuton, dependent varable s a non-negatve nteger-valued count and therefore we can employ count data methodology to estmate frequency of travellng. A classcal model for count data s Posson process: P( n) n λ e λ =, where n denotes n! number of travels that household or ndvdual makes durng a fxed nterval, say for nstance, a year; P( n ) denotes probablty that household or ndvdual travels n tmes and ( ) exp( SE β ) λ = E n SE =. An mportant feature of the Posson model s that t mposes ( ) ( ) E n SE = Var n SE = λ, whch s known as the equdsperson property. In some cases ths property may be true but n other cases t s volated. If varance s greater than expectaton, we have a case of overdsperson and Posson model wll tend to under-predct the actual frequency of zeros. Mullahy (1997) assocates overdsperson wth the exstence of unobservable heterogenety. He suggests to deal wth overdsperson allowng for a specfcaton whch ncludes an error component whch represents omtted varables or unobserved varables. For ths purpose, negatve bnomal dstrbuton s usually appled. It can be seen as a more general expresson that also contans Posson dstrbuton as a partcular case. The negatve bnomal dstrbuton mposes E( n ) that when ( ) k = λ and Var ( n ) = λ+ aλ, such k = Var n = λ+ aλ, known as negatve bnomal 1 model, when 2 ( ) λ λ k = Var n = + a, known as negatve bnomal 2 model and when ( ) a = Var n = λ we obtan Posson model. An alternatve model s known as zero 0 nflated model, whch employs a mxng specfcaton whch adds extra weght to the probablty of observng a zero. The man nconvenent of these models s that varance s mposed exogenously by the researcher. If ths varance works properly for the model of nterest t wll be suffcent but sometmes excess of zeros may not be assocated wth

17 ncreased dsperson but wth an underlyng tourst s behavour. We refer to unobservable patterns n the behavour of toursts, as fear of flyng or hgh propensty to travel. If researcher beleves ths behavour s relevant enough, then t requres a more complex model for the analyss. In order to deal wth unobservable heterogenety we wll consder brefly two dfferent approaches. On one hand, we can splt populaton accordng to partcpaton decson of stage 1 and assume frequency depends on two separate processes, one followed by those people who decded not to travel and other process that consder all those people who do travel. In the econometrc lterature ths s known as hurdle model. On the other hand, we can splt populaton accordng to how much they lke toursm, already obtaned n stage 2. In ths case, each segment follows an ndependent process 10. The ndependent processes of both of these mechansms are added up n the log-lkelhood fnal functon n order to be estmated by maxmum lkelhood. Ths s the reason why these models are known as mxture models. 6. STAGE 4: KIND OF DESTINATION Introducton In prevous stages, we have studed partcpaton decson, toursm budget decson and frequency of travellng. Once an ndvdual or household have chosen a frequency for travellng, they must decde where to go n ther dfferent trps. Nowadays, for most of the developed countres, the choce set for travellng s qute large. There are multple avalable

18 destnatons. Each of them posses specal features,.e. some partcular characterstcs that make them unque. There are some destnatons that satsfy almost every knd of general need, but maybe do not satsfy the needs of a mnorty of toursts. Once agan, researcher faces a problem of heterogenety. Toursts dffer from what they consder t s an deal destnaton and whch are the needs they want to be satsfed. Wthn ths stage we consder toursts make two smultaneous decsons. One of them concerns the nature of the travel. In ths sense, tourst must decde whch knd of travel wshes. For nstance, they may choose between a famlar, adventure or relaxng trp. The other decson concerns the knd of destnaton n terms of ts physcal attrbutes. In other words, for nstance, toursts may decde f the knd of destnaton they prefer s a mountan, a cty, a countrysde resort or a seasde resort. Objectve Ths stage s a prevous stage to fnal destnaton choce and t helps n defnng more approprately the choce set for dfferent knd of needs that dfferent toursts are demandng. It s mportant to defne properly the choce set because t wll affect effcency of estmates. Man varables Besdes the soco-economc and demographc varables, we must create a tree structure that classfy dfferent knd of destnatons dependng on ther physcal attrbutes and knd of tourst envronment that may possess. 10 See for nstance Deb and Trved (1997)

19 Methodology An llustraton of dfferent crtera to classfy the dfferent knd of destnatons that toursts may wsh to go, s shown below: a) Physcal attrbutes of destnaton a. Seasde b. Countrysde c. Mountan d. Cty In ths case, tree structure posses four branches and researcher needs to determne n whch of these branches every destnaton must be classfed. b) Tourst envronment a. Famlar b. Cultural c. Relax d. Party e. Adventure Smlarly to case a, researcher needs to classfy each potental destnaton accordng to these fve branches. c) Both combned herarchcally In ths case, tree structure has two levels. In a frst stage, toursts mght choose a knd of destnaton, accordng to crteron a (or b) and n a second stage they mght choose a knd of

20 tourst envronment, accordng to crteron b (or a). In any case, we end up wth 20 dfferent knd of tourst destnatons. d) Both combned smultaneously Smlarly to prevous classfcaton, ths proposton do not consder stages but tourst must decde smultaneously among the 20 dfferent knd of toursts destnatons. The choce of any of these alternatve ways to classfy knd of tourst destnaton s relevant for the methodology of fnal destnaton choce. It s possble to model knd of tourst destnaton decson wth a multnomal logt model f we attend crtera a, b, or d. For crteron c, we can model t wth a nested multnomal logt model. Any household, labelled h, may decde to travel not. Once the household has decded to travel, t has to choose the knd of tourst destnaton they wsh, labelled d, among ts choce set. In order to model the knd of destnaton choce, we follow a behavoural model where household chooses the alternatve that provdes hgher level of utlty, denoted by U. In ths sense, household h would choose knd of destnaton d f and only f: Uhd > Uhr r d, where r denotes any other knd of resort. Nevertheless, these utlty levels are unobservable for the researcher. The only aspects that we know are some socoeconomc varables of the household, denoted by SE h, and some attrbutes of the set of destnatons, denoted by A d. V = V SE A d, whch From the nformaton avalable, we can construct a functon (, ) hd h d represents, the utlty that ste s provdes ndvdual. Obvously, ths representatve utlty V hd s an approxmaton to the current utlty U hd. Thus, we can state that utlty can be decomposed as: Uhd = Vhd + ε hd, where ε hd denotes unobserved part of utlty for household h

21 when vsts destnaton d. Thus, the probablty that a household h chooses to travel to knd of destnaton d s: ( ) ( ε ε ) ( ε ε ) P = Pr U > U r d = Pr V + > V + r d = Pr < V V r d hd hd hr hd hd hr hr hr hd hd hr For convenence, we assume that ε hd d extreme value. The advantage of ths assumpton s that the normalsaton requred n any dscrete choce model,.e. εhdr = εhr εhd, mposes that the error dfferences ε hdr s dstrbuted logstcally, whch mples that, after algebrac manpulaton 11, the probablty formula can be obtaned exactly as: P hd Vhd e =. Vhr e r Nevertheless, we can lnk knd of tourst destnaton choce wth fnal destnaton choce through a nested multnomal logt model. Ths s the purpose of last stage. 7. STAGE 5: DESTINATION AND TRANSPORTATION MODE CHOICE Introducton One of the most nterestng decsons of all ths choce process s fnal destnaton choce. As we have remarked along ths paper, researcher needs to deal wth the heterogenety of toursts n order to obtan accurate estmates from the model proposed. Prevous stages provdes addtonal nformaton to be ncluded n the analyss of ths fnal stage that help to obtan more effcent resuts. From stage 2 we can nclude nformaton as how much each ndvdual lkes toursm, from stage 3, f sgnfcant, we can provde nformaton as propensty to travel and stage 4 s very mportant for the methodologcal structure and mnmsaton of heterogenety among toursts. 11 Tran (2003) ncludes the full algebrac manpulaton.

22 We assume that when toursts determne a destnaton to travel, ths decson s lnked to the transportaton mode choce. In nternatonal toursm, plane s the expected transportaton mode and decson about transportaton mode s already taken. However, there are many other destnatons where tourst may choose how to go there. To capture ths effect, we propose to nclude n tourst s choce set, every sensble combnaton between destnaton and transportaton mode. Objectve Ths fnal stage focus n the estmaton of the man determnants of toursts destnaton choce. Ths s the more complex stage n terms of varables that may nfluence toursts. We consder characterstcs of the household or ndvdual and attrbutes of the destnatons. The purpose s to determne the relatve mportance of any of these varables wth respect to the fnal decson. Moreover we want to obtan a framework whch may allow us to smulate how current dstrbuton of toursts would be affected by any change n any varable of nterest. Man varables The man varables that we expect to be relevant are, for llustratve purposes shown n Table 1: (Table 1 should be here. It s located at the end of ths document.)

23 Methodology As we have mentoned n the prevous secton, the methodology proposed lnks stage 4 and 5 employng a nested multnomal logt model 12. Ths s a superor modellng wth respect to the multnomal logt model because t deals more approprately wth unobservable heterogenety. Moreover, multnomal logt model mposes, by constructon, a restrcton known as ndependent of rrelevant alternatves (IIA) property. For our purposes, ths property may cause an nconvenent. More precsely, IIA mples that, gven a change n any attrbute of the alternatves, cross elastctes of the probabltes of choosng any alternatve are exactly the same for every alternatve. If the varables that explan the behavour of toursts are stable over tme ths mplcaton s not problematc as t s the case of the stage 4. However, n stage 5 t s lkely that attrbutes of the destnatons vary qute often. Moreover, f we wsh to smulate changes n these attrbutes, same cross elastctes assumpton would create a bas n the results. Nested multnomal logt deals wth IIA, such that t allows for dfferent cross elastctes between dfferent nests but not wthn the same nest. Generally speakng, ths specfcaton would be flexble enough n order to obtan sensble smulatons. Nevertheless, more flexble alternatve models can be appled. If IIA s stll a relevant problem n the model, researcher may use heteroskedastc extreme value models (HEV). Furthermore, f researcher wshes to apply full flexblty to the estmates, such that, the model can provde partcular estmates for each knd of representatve ndvdual, random parameters model or mxed logt model can be employed. The nconvenence of these models s that probabltes can not be obtaned wth ntegrals whch posses a closed-form but they requre smulaton and consequently estmaton s more complcated.

24 An specfc problem of ths stage that requres specal care s the defnton of the choce set. Three aspects must be taken nto account: crtera to defne dfferent alternatves, flexblty and sze of the alternatve destnatons. Concernng the defnton of dfferent alternatves, we may follow dfferent crtera. We can consder destnatons as a poltcal dvson; a natural dvson, n terms of knd of terrtory or weather; and a dvson accordng to the knd of actvtes that can be practsed n the destnaton. Snce toursm lterature s stll lmted n ths area, we can beneft from outdoor recreaton studes. Parsons and Hauber (1998) showed that for recreatonal fshng trps, 94% of ndvduals choose stes wthn one hour and a half travel tme dstance. Consequently, for ths knd of recreaton spatal boundares for choce set must be establshed n order to optmse the effcency of the estmaton. Moreover, t s nterestng to see Thll (1992), because he consders dfferent approaches to capture the true choce set and Haab and Hcks (2000), who have provded a survey of the way that choce sets have been consdered n recreaton demand models. Researcher mght also thnk about the flexblty of the choce set. Ths can be predetermned and fxed by the researcher or t can be endogenously determned by the model,.e. defnng a partcular choce set for each ndvdual. Endogenety may be explaned, among other varables, through level of nformaton, number of trps or age. General awareness of tourst destnatons t s lkely to wden the choce set. We can consder four levels of awareness: Internatonal, natonal, regonal and local level. The more an ndvdual travels at any of these levels, t s expected that, the wder hs or her choce set wll be. 12 Eymann and Ronnng (1997) apples nested multnomal logt model for the study of German outbound

25 Fnally, we also need to adjust the sze of the destnatons, such that, most of them be homogeneous n sze. Otherwse, we may have a bas wth respect to the effect of the attrbutes. For nstance, despte France and Swtzerland may have n the Alps very smlar attrbutes, France have many more vstors because t s a bgger country. In ths sense, attrbutes of Swtzerland may be undervalued by the estmaton wth respect to France due to the dfferences n the sze of both countres. We need ether to splt up the country nto homogeneous destnatons or adjust the number of arrvals wth respect to the sze of the country (arrvals per km square). 8. CONCLUSIONS The methodology proposed assume that toursts destnaton choce s condtoned to another four decsons. We argue that decson process follows a herarchy wth fve stages: partcpaton decson; toursm budget decson; frequency and length of stay decsons; knd of tourst destnaton decson; and fnal destnaton and mode of transportaton choce. Ths structure responds to the necessty to deal wth the heterogenety among toursts. For the frst stage, the partcpaton decson, we suggest the employment of a probt or logt model. From ths model, we may estmate the probablty of an ndvdual or household to travel and the effects on ths probablty of any change n any socoeconomc varable. In the second stage, the objectve s to determne the man factors that push dfferent ndvduals to spend part of ther budget n toursm actvtes. Rather than estmate toursm toursm demand.

26 expendture n absolute terms we prefer to model percentage of toursm expendture over ncome. Ths transformaton offers more relevant results and lets us estmate how much people lke toursm. Thrd stage analyses the man factors that contrbute to defne travel frequency wthn a perod of tme. In ths stage we dscuss the role of length of stay n the frequency decsons. We conclude that, provded ndvduals or household posses some flexblty to choose the length of stay, then ths wll optmally be chosen dependng on the destnaton chosen and the amount of the fxed costs ncurred. For frequency decsons, the model suggested depends on the dsperson of the frequency. If frequency s equdspersed, Posson process may be appled. However, f frequency s overdspersed, ths may need to be modelled because t would respond to a case of unobservable heterogenety. For ths purpose, negatve bnomal model, zero-nflated model, hurdle model and mxture model are dscussed. In the fourth stage, we propose a tree structure whch may classfy dfferent segments of toursts dependng on ther needs. Two crtera whch defne ths classfcaton are the physcal attrbutes of the destnaton and the tourst envronment wshed. Methodology proposed n ths stage s lnked wth last stage, such that both decsons are modelled wthn a nested multnomal logt. Last stage, corresponds to destnaton and transportaton mode choce. We propose to nclude n tourst s choce set, every sensble combnaton between destnaton and transportaton mode. Specal care needs to be consdered wth the choce set defnton for toursm analyss. In ths sense, we remark three man aspects to be taken nto account: crtera to defne dfferent alternatves, flexblty and sze of the alternatve destnatons.

27 Fnally, we end up wth a complete methodologcal framework that dsaggregates toursts decsons and obtan the determnants for each of these decsons and allows us to smulate how current stuaton may change under alternatve scenaros.

28 References Bakkal, I. (1991). Characterstcs of West German demand for nternatonal toursm n the northern Medterranean regon. Appled Economcs, 23, Deaton, A. and Muellbauer, D. (1980a). An almost deal demand system. Amercan Economc Revew, 70 (3), Deaton, A. and Muellbauer, D. (1980b). Economcs and consumer behavor. Cambrdge: Cambrdge Unversty Press. Deb, P. and Trved, P.K. (1997). Demand for medcal care by the elderly: A fnte mxture approach. Journal of Appled Econometrcs, 12, Dvsekera, S. (2003). A model of demand for nternatonal toursm. Annals of Toursm Research, 30 (1), Eymann, A. and Ronnng, G. (1997). Mcroeconometrc models of toursts destnaton choce. Regonal Scence and Urban Economcs, 27, Greene, W. (2003). Econometrc Analyss. Ffth edton. New Jersey: Prentce Hall. Haab, T.C. and Hcks, R.L. (2000). Choce set consderatons n models of recreaton demand: Hstory and current state of the art. Marne Resource Economcs, 14,

29 Hultkrantz, L. (1995). On determnants of Swedsh recreatonal domestc and outbound travel, Toursm Economcs, 1 (2), Maddala, G. (1983). Lmted dependent and qualtatve varables n econometrcs. New York: Cambrdge Unversty Press. Morley, C.L. (1995). Toursm demand: characterstcs, segmentaton and aggregaton. Toursm Economcs, 1 (4), Mullahy, J. (1997). Heterogenety, excess zeros, and the structure of count data models. Journal of Appled Econometrcs, 12, O Hagan, J.W. and Harrson, M.J. (1984). Market shares of US tourst expendture n Europe: an econometrc analyss. Appled Economcs, 16, Parsons, G.R. and Hauber, A.B. (1998). Spatal boundares and choce set defnton n a random utlty model of recreaton demand. Land Economcs, 74 (1), Papatheodorou, A. (1999). The demand for nternatonal toursm n the Medterranean regon. Appled Economcs, 31, Syropoulos, T.C. and Snclar, M.T. (1993). An econometrc study of toursm demand: the AIDS model of US and European toursm n Medterranean countres. Appled Economcs, 25,

30 Rugg, D. (1973). The choce of journey destnaton: A theoretcal and emprcal analyss. Revew of Economcs and Statstcs, 55, Thll, J. (1992). Choce set formaton for destnaton choce modellng. Progress n Human Geography, 16 (3), Tran, K. (2003). Dscrete choce methods wth smulaton. Cambrdge: Cambrdge Unversty Press.

31 Table 1. Man varables of stage 5: Destnaton and transportaton mode choce Characterstcs of the household Attrbutes of the destnatons Mxed varable Dsposable ncome Budget for toursm expendture Dsposable tme for toursm Labour condtons Frequency of travellng Sze of the household Age of the oldest and youngest members of the household Educaton Place of resdence Sze of communty: rural or cty Rsk averson Party sze of travellers Relatve prces (PPP): Prces and exchange rate Accommodaton cost ndex Weather Safety Crowdng Development and facltes Sze of the country Transportaton cost Travel tme cost (mode of transportaton) Avalable nformaton Language Sutablty of destnaton (loss functon) Marketng n the country of orgn