Consumer Store Choice and Choice Set Definition Author(s): A ...

Consumer Store Choice and Choice Set Definition Author(s): A. Stewart Fotheringham Reviewed work(s): Source: Marketing Science, Vol. 7, No. 3 (Summer, 1988), pp. 299-310 Published by: INFORMS Stable URL: http://www.jstor.org/stable/183719 . Accessed: 23/08/2012 10:08 Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at . http://www.jstor.org/page/info/about/policies/terms.jsp

. JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about JSTOR, please contact [email protected].

.

INFORMS is collaborating with JSTOR to digitize, preserve and extend access to Marketing Science.

http://www.jstor.org

MARKETING SCIENCE Vol. 7, No. 3, Summer 1988 Printed in U.S.A.

Note

.

CONSUMER STORE CHOICE AND CHOICE SET DEFINITION A. STEWART FOTHERINGHAM UWIST, Cardiff Consumer store choice results from a process whereby information on various alternatives is evaluated by the consumer prior to the selection of one of these alternatives. In the application of store choice models it is often assumed that the information-processing strategy underlying store choice is a simultaneous one in which all possible alternatives are evaluated by an individual. A competing assumption, increasingly recognized in aspatial choice, is that individuals initially evaluate clusters of alternatives and then only evaluate alternatives within a chosen cluster. The nested logit model is a well-known formulation for capturing this type of choice process. This paper adds to the above understanding of consumer spatial choice in the following ways: (i) It describes an alternative to the nested logit model, known as a competing destinations model, which can be used to model hierarchical spatial choice processes. While this model has been described previously in the geography and planning literature, this paper contains a novel derivation of it from within the random utility maximization framework. (ii) A general choice formulation is derived from which the logit, nested logit, and competing destinations models can be obtained. This allows the behavioral assumptions embedded in each model to be seen clearly and it allows a comparison of the nested logit and competing destinations formulations. The latter is shown to be preferable in most spatial choice situations. In particular, the latter can be used when restricted choice sets are fuzzy. (iii) The errors that arise when spatial choice results from a hierarchical process and when a simultaneous choice structure is assumed are demonstrated and are related to the presence of consumer competition/agglomeration effects between stores which are thought to be important in store choice. (Random Utility; Information Processing; Hierarchical Choice)

1.

Consumer Choice Processes

The investigation of two types of choices that consumers regularly make dominates the marketing literature: the choice of particular brands and the choice of particular stores. These choices are generally examined independently and indeed possess different properties (brand choice is aspatial whereas store choice is spatial), yet the decision processes leading to the two choices are very similar and can be investigated with 299 0732-2399/88/0703/0299$01.25 Copyright ? 1988, The Institute of Management Sciences/Operations Research Society of America

300

A. STEWART FOTHERINGHAM

mathematical models of the same general form. One similarity, for example, concerns the extent to which choice depends on a hierarchical evaluation of alternatives and to what extent it depends on a simultaneous evaluation. However, while debate on this issue is prevalent in the brand choice literature, it is relatively absent from the literature on store choice. This paper is an attempt to correct this situation. Traditionally in consumer choice modelling it has been assumed that individuals, whether they are selecting a brand or a store, evaluate each alternative in terms of the utility or benefit to be derived from selecting that alternative and then select the alternative yielding maximum utility. The assumption that consumers undertake such a task for a large number of alternatives is increasingly being questioned. An alternative assumption is that choice results from a hierarchical or sequential decision process whereby a cluster of similar alternatives is first selected and then a specific alternative is chosen from within this cluster. In this way, the individual increases the efficiency of his/her decision-making process by avoiding having to evaluate all possible alternatives. As an example of the difference in these two conflicting hypotheses regarding consumer choice, consider the selection of a clothes store by an individual living in a large city. It is unlikely that a consumer has the ability or time to evaluate all such stores within the city. More likely, the consumer will make an initial choice of a cluster of stores, a shopping district perhaps or a mall, and then select a store or stores, from within this cluster. The store choice problem has an added complexity, however, not usually found in brand choice, in that it may not always be possible to define the clusters of alternatives perceived by individuals and, hence, there may be uncertainty as to cluster membership. This complexity has led to the development of a new type of hierarchical choice model, the competing destinations model, which is the subject of subsequent discussion. The remainder of the paper proceeds as follows. After a brief review of the evidence for hierarchical consumer choice, a general choice model is described which can accommodate a selection process from restricted sets of alternatives. Three particular choice models are then derived from this general formulation based on different assumptions about consumers' decision processes. The first, a logit model, is based on the assumption that individuals evaluate all alternatives prior to making a selection; the second, a nested logit model, is based on the assumption that individuals evaluate alternatives hierarchically and there is no uncertainty regarding cluster membership; and the third, a competing destinations model, is based on the assumption of hierarchical evaluation but where there is uncertainty regarding cluster membership. Finally, the errors that arise when a simultaneous store choice process is incorrectly assumed are identified. 2. Rationale for Hierarchical Consumer Choice Kahn et al. (1985) outline a number of reasons why consumer choice can be expected to result from a hierarchical, rather than a simultaneous, evaluation of alternatives. The main rationale for hierarchical choice is that the capacity of humans to process large amounts of information is limited and a simplifying procedure is needed in order to reach a decision in situations where there are large numbers of alternatives. Even if it were possible for an individual to process information on every alternative, however, it may be that time pressures and other distractions can motivate a decision-maker to employ a short-cut evaluation process. Empirical evidence exists to suggest that consumers often make brand choices in a hierarchical manner. Amongst others, Alexis et al. (1968) examine the selection of women's clothing outlets as a hierarchical decision process. Rao and Sabavala (1981) and Moore et al. (1985) investigate the purchase of soft drinks as the result of a hierarchical evaluation process whereby consumers initially make decisions such as

301

CONSUMER STORE CHOICE AND CHOICE SET DEFINITION

cola versus noncola drinks, diet versus nondiet drinks, and caffeine-free versus regular drinks and Dubin (1986) examines the hierarchical choice of space and water heat systems. In research with a slightly different emphasis, Guadagni (1983) used scannerpanel data for ground coffee sales in Kansas City to suggest that consumers first choose a brand of coffee and then decide whether to make a purchase. Relatively little research though exists on hierarchical destination choice although some initial work has begun to appear in other disciplines such as geography (Fotheringham 1983, 1986) and regional science (Borgers and Timmermans 1987). Fotheringham provides evidence that a hierarchical process exists in nonretailing choice situations such as migration and airline travel while Borgers and Timmermans provide some initial evidence from shopping behavior in the Netherlands that consumers' choice of retail outlets results from a hierarchical decision process. Similar evidence is provided by Recker and Schuler (1981) in an investigation of supermarket choice in Bloomington, Indiana. 3. Derivation of a General Spatial Choice Model Consider an individual located at place i who is of socioeconomic type k (a middleaged, high-income female, for example) and who is faced with choosing a retail outlet j from a set J containing J such outlets, each of which will generate a certain level of utility. A measure of the utility, Uijk'is given by: L

Uijk = z

aiklf(Xij)

+ 1ijk

(1)

where Xij,is the level of destination attribute 1, fis the functional relationship between Xij and Uik, aikl describes the relationship betweenf(Xij) and Uijk,and ijk is a random error component. Assume for convenience only that individuals are homogeneous or that aikl does not vary across socioeconomic groups so that the subscript k can be omitted from equation (1). Define pij as the probability that individual i selects alternative j'. A choice rule is established so that 1' A/'=

if

U,i>Ui

(jEJ,j+J'),

.

(2)

otherwise.

That is, the individual selects the destination yielding him/her maximum benefit with probability equal to one. Other destinations have a zero probability of being selected. Since individuals' exact utilities are unknown, the probability of' being selected is Using the definition that

Pij = p[Ui" > U,(j E J, j :J')].

(3)

L

Vij -

s aif(Xijl)

(4)

it is well known (McFadden 1974) that r++

Pii = =-oo

g (ij

= x)

j j

VW-Vij,+x

Jy=-~o

g (i

= y)dy

dx

(5)

where g( ) represents a probability density function. Consider, however, that the individual does not evaluate all J alternatives but instead only evaluates M of these alternatives where M E J and from either the point of view of the modeller or the individual making the choice, there is uncertainty about the composition of M. That is, the set M can be considered as 'fuzzy' (Zadeh 1965). Then equation (3) should be replaced by


302

pij = p[Pi (' E M). Ui, > pi (j E M). Ui (j E J, j j')]

(6)

where the utility associated with a particular alternative is weighted by the probability of that alternative being in the set M that the individual at i considers. By following the derivation of equation (5) but replacing equation (3) with equation (6), a more general choice model is obtained whose formula is 00 voi,-vii+x 'E )S J P+i g (ij = y)' dy- dx. (7) g ('ij = x) I Pi (j E M)=-o Pij = Pi (' E M)f J X=-OG

j j$j

Clearly, if pi (j E M) = 1 for all j, that is, the individual evaluates all possible alternatives, then equation (7) is equivalent to equation (5), the traditional model. Equation (7), however, provides more flexibility in modelling choice processes through the inclusion of the uncertainty surrounding the composition of the set M. In a situation where individuals are selecting a retail outlet to patronize, this uncertainty may arise for one of several reasons that are described in ?5. Three specific forms of the general choice model in equation (7) are now derived in order to clarify the conditions under which use of each of these models is appropriate. One form, the logit model, assumes that the choice process by which a consumer reaches a selection of a particular store consists of a simultaneous evaluation and comparison of all stores; the second, a nested logit model, consists of a hierarchical evaluation where there is no uncertainty regarding membership of the set M; and the third, a competing destinations model, consists of a hierarchical evaluation where there is uncertainty regarding membership of the set M. 4. Simultaneous Store Evaluation: The Logit Model Assume in equation (7) that the tijs are independently and identically distributed with a Type I extreme value distribution (Fisher and Tippett 1928) so that g (ij = x) = exp[-exp(-x)].

(8)

Further assume that pi (j E M) = 1 for all j. Then McFadden (1974) has shown that, under such circumstances, exp(Vj) Pi'j= Zj exp(V0)

VjEJ,

(9)

which is termed the multinomial logit model and which has been applied extensively to store choice problems (inter alia, Gautschi 1981; Stanley and Sewell 1976; Fotheringham 1987). However, the above derivation highlights a major problem with its application in retailing and marketing. It is well known that models whose general form is given by equation (9) contain what is known as the Independence from Irrelevant Alternatives (IIA) property (McFadden 1974). This property states that the likelihood of selecting one alternative over a second is independent of any other alternative so that if a new alternative is added to the set of existing outlets, models having the general form of equation (9) would predict the new alternative from each existing outlet in proportion to its original share. Classic violations of this behavior in reality include the red bus/blue bus problem in mode choice (Domencich and McFadden 1975) and the choice of soft drinks in brand choice (Moore et al. 1985). Violations of the IIA property in spatial choice models are equally obvious. The location of an outlet with respect to its competitors will affect the probability of a consumer selecting that outlet. If agglomeration forces are present, an individual is more likely to select an outlet in close proximity to other outlets (perhaps for comparison shopping), ceteris paribus. Alternatively, if

303


competition forces are present, an individual is more likely to select an outlet in relative isolation from its competitors, ceteris paribus. 5. Hierarchical Store Evaluation with Known Choice Set Membership: The Nested Logit Model The nested logit can be derived from equation (7) by assuming that in equation (7) the ijis are independently and identically distributed with a Type I extreme value distribution and that pi (j E M) = 1 for all j E M and Pi (j E M) = 0 for all j N M. That is, assume the membership of the set M from which the consumer selects a particular retail outlet is known with certainty. Such might be the case, for example, when an individual is shopping for clothes and first selects a particular mall. Once a mall has been chosen, the membership of the set M, that is the set of clothing stores within the selected mall, is known with certainty. It then follows from the derivation of equation (9) that exp( V,) jEMexp(Vi)

Pio =

Vj

1o

M,

(10)

V?j'M,

which is equivalent to the application of a nested logit model at the lower tier of the choice hierarchy (Sobel 1981). Where well defined clusters of alternatives exist, such as cola versus noncola drinks, or caffeinated versus decaffeinated coffees in brand choice modelling, and shopping malls in spatial choice modelling, the nested logit formulation can be applied with relative confidence. However, several problems exist with its application to choice problems where clusters of alternatives are not well defined but instead are perceived by consumers as fuzzy spatial units with ill-defined boundaries. Such is likely to be the case, for example, in grocery shopping and in nongrocery shopping where malls do not exist and where there is a continuous, albeit uneven, distribution of retail outlets. Under these circumstances, the following problems arise in the application of the nested logit model: (1) Unlike aspatial hierarchical choice situations such as brand selection where the clusters of alternatives are readily identified based on the presence or absence of a discrete variable (such as diet versus regular soft drinks), the clusters of retail outlets perceived by individuals are separated by a continuous variable, space. Hence, the divisions between clusters, and consequently, the composition of clusters, are much less clear in a spatial choice context than in an aspatial context; (2) The imposition of an artificial, discrete choice set on a continuous surface can also lead to the situation in the application of the nested logit model whereby a retail outlet near the border of one cluster, and an outlet near the border of an adjacent cluster are not considered substitutes for one another yet they are located very close together; (3) Similarly, outlets within the same cluster, no matter how far apart they are located, are considered to be equal substitutes for each other. Consider, for example, the following three choices: A B C *

*

0

Suppose these three choices represent different soft drinks and A is substitutable with B and B is substitutable with C: it then follows that A is substitutable with C. That is, there is transitivity between the choices. Now, however, suppose that A, B, and C are retail


304

outlets within a geographic region and that A and B and B and C are substitutable pairs. Because A and C are located much further apart from one another, the relationships between A and B and between B and C do not imply that A and C are substitutes for one another. That is, there is no guarantee of transitivity between alternatives in the same cluster although this is assumed in the application of the nested logit model; and (4) The set of outlets forming each cluster is assumed to be constant across individuals yet such consistency is extremely unlikely in a spatial context due to locational variations in the perception of space. The set of shopping alternatives identified by an individual in one location is likely to be different from the set identified by an individual in another location. For the above reasons, serious problems can exist with transfering the nested logit model from an aspatial choice context where it is easily operationalized to a spatial choice context where it appears to have severe limitations. We now consider a hierarchical choice model which does not have these limitations. 6. Hierarchical Store Evaluation with Unknown Choice Set Membership: The Competing Destinations Model Assume in equation (7) that the uijsare independently and identically distributed with a Type I extreme value distribution but make no assumption regarding the values of the pi (j E M) terms. Then, following McFadden's (1974) derivation of the logit model, the choice model that is derived is: exp(Vij)- pi (j' E M)

Pi' - Zj exp(Vij) .i (j E M)1 where the exponential of the measurable component of each alternative's utility is weighted by the probability of that alternative being evaluated by the consumer. Thus, when the consumer selects an alternative from a restricted choice set, pi (j E M) measures the probability ofj being in that restricted set. Two special cases of equation (11) arise. If consumers do make choices from a restricted choice set but membership of this choice set is known with certainty, equation (11) is equivalent to the part of the nested logit model represented in equation (10). If consumers do not make choices hierarchically but evaluate all alternatives, then pi (j E M) = 1 for all j and equation (11) is equivalent to the logit formulation in equation (9). Notice that since pi (j E M) can be represented as li (j E M) divided by some constant with respect to j, where li (j E M) is the likelihood that an individual at i perceives j being in M, equation (11) can be rewritten as: Pij'

exp( Vij,) li (j' E M) Zj exp(Vi) i (j E M)

(12)

Two approaches to measuring li (j E M) presently exist. One approach considers that the likelihood of a particular alternative being in the restricted choice set is a function of the dissimilarity of that alternative to all others. The rationale for this approach is that the degree to which an alternative possesses distinctive properties affects its chances of being included in M. Whether it is affected positively or negatively is an empirical question. Several formulations have been suggested to measure an alternative's dissimilarity to other alternatives. Batsell (1981), for example, suggests the following: li(j'EM)=

exp(_- I1 W

1Jk

C k0klX'k -Xik

/

(13)

305


where Xjk is the kth attribute of alternative j and Ok is a parameter reflecting the contribution of dissimilarity on the kth attribute to j"s overall dissimilarity. Meyer and Eagle (1982) provide a slightly different measure where: li (j' EM)=

X1 , ]0.51 j-

J-

I I

(14)

Here rjj is the correlation coefficient between j' and j across their attributes. As in Batsell's formulation, this difference measure is averaged across all alternatives. Borgers and Timmermans (1987) suggest a similar formulation where differences on one attribute are averaged across all alternatives and then the product of these average differences is taken over all attributes: li (j'E M) = nI [J

1

IXjk-Xjk ]

(15)

The other approach to measuring li (j E M) is most applicable to spatial choice and recognizes that outlets in close geographic proximity are more likely to be substitutes for one another than are outlets located at greater distances from each other. Hence, the location of an outlet with respect to all other outlets affects its chances of being included in the restricted choice set of an individual. For example, an individual may be less likely to include a retail outlet that is spatially isolated in his/her choice set than an outlet which is relatively close to other outlets and for which it can act as a substitute. Two closely related formulations exist to measure li (j E M) in this way. Fotheringham (1983) suggests using a sum of weighted distances from one alternative to all others where the weight is the size of each alternative. That is, i (j'EM)=

(-

wj/dj)

(16)

where represents the weight of outlet and d represents the distance between j' and where wj represents the weight of outlet j and djj represents the distance between j' and j. This is a traditional measure of relative location within disciplines concerned with the role of space in human activities and is sometimes referred to as potential accessibility (Hansen 1959). Large values of the variable indicate outlets that are in close proximity to other outlets; low values indicate outlets that are spatially isolated. For simplicity, all stores are considered to be of the same type so we discuss only comparative shopping and multipurpose shopping is effectively ignored. However, it would be a simple matter to incorporate multipurpose shopping into this model by defining a set of relative location variables measuring the relative location of a store to particular types of stores or land uses. An alternative to equation (16) has been proposed by Borgers and Timmermans (1987) who suggest a simpler average distance measure: l (j' EM) =( -

djj

(17)

Use of either equation (16) or (17) to measure li (j E M) produces what Fotheringham has termed a competing destinations model (Fotheringham 1983, 1986). Such a model no longer contains the IIA property and the property Huber et al. (1982) term regularity. That is, in the logit formulation of equation (11) it is impossible to increase the probability of selecting an existing alternative by adding a new alternative to the choice set [the numerator of equation (9) remains constant while the denominator increases]. In reality the addition of a new retail outlet to a shopping mall, for example, can increase the market share of existing stores in the mall. The competing destinations

306


model can model such behavior if the increase in li (j E M) due to the addition of the new outlet is greater than the increase in the denominator of equation (12). The competing destinations model is equivalent to the logit formulation whenever the parameter 0 in equation (16) is equal to 0 so that pi (j E M) = 1 for all j. That is, whenever consumers evaluate all alternatives. Batsell and Polking (1985) suggest that modelling hierarchical choice with equation (12) has several advantages over other methods such as the nested logit procedure. These include the following: (1) No choice structure has to be specified a priori and alternatives do not have to be assigned to clusters prior to model calibration; (2) The model is simple to calibrate and can be calibrated with existing gravity or logit software such as SIMODEL (Williams and Fotheringham 1984) or by least squares techniques using Nakanishi and Cooper's transformation (Nakanishi and Cooper 1974); (3) The model structure provides insight into the choice process and into competition between outlets through the parameter 0 in the formulation for 1i (j E M). For instance, when the model incorporating equation (16) is calibrated and an estimate of 0 obtained, if 0 > 0, outlets will increase their market share by locating in close proximity to other outlets and agglomeration forces are said to exist (one of the reasons for the success of shopping malls). Conversely, if 0 < 0, outlets will increase their market share by being isolated from their competitors (as supermarkets are) and competition forces are said to exist. More information on these processes in a retailing context is provided elsewhere (Fotheringham 1985); and (4) The simple nature of the model's structure makes it convenient to work with analytically as Fotheringham and Knudsen (1986a, b) have done in analyzing the spatial dynamics of retailing systems with catastrophe theory. 7. Errors in Assuming Simultaneous Evaluation of Alternatives When the Evaluation Process Is Hierarchical There are two competing hypotheses regarding the process by which a consumer evaluates alternatives: one assumes a simultaneous evaluation whereby the spatial clustering of outlets is ignored and all possible alternatives are evaluated and compared; the other assumes a hierarchical process where consumers first evaluate clusters of alternatives and only from within a chosen cluster do they evaluate individual alternatives. As discussed above, in the latter process the definition of the individual's choice hierarchy may be known (leading to a nested logit modeling approach) or it may be fuzzy (leading to a competing destinations modeling approach). In either case, the errors that arise when a simultaneous choice process is incorrectly assumed can be identified from the retailing system described in Figure 1. Assume for the sake of exposition only that the observable utility a consumer would receive from selecting a particularoutlet (its perceived attraction to the consumer) is constant across the outlets. That is, (18) vi = Vi Vj. Let n, represent the number of outlets in cluster s and Visbe the perceived attractiveness of cluster s. Consider the two evaluation processes described above. 'Even in shopping malls, there is some evidence to suggest that competition forces exist. Miller and Lerman (1981) found that as the number of women's clothing stores increased within a mall, the average expenditure/store decreased.

307


0

-.cluster

1

s

*.--outlet

j

i*j

I

Ir

FIGURE1. A Simple Retailing System.

Simultaneous Evaluation of all Alternatives If consumers evaluate only individual outlets without regard to their spatial clustering, the attractiveness of a spatial cluster of outlets is merely the sum of the attractions of its individual outlets. That is, Vis= z Vi (19) jEs

which on substituting equation (18) implies that

Vis= nsVi.

(20)

Consider now what happens to the perceived attractiveness of the cluster as ns increases. The rate of increase in this attractiveness is aVisans = Vi

(21)

which is a constant and is represented by the slope a in Figure 2. That is, regardless of the original size of a cluster, the addition of a new outlet to the cluster will increase the perceived attractiveness by Vi, the perceived attractiveness of the new outlet. In many circumstances it would seem unreasonable to expect such a relationship to occur in reality. For example, the addition of one store to a cluster of 100 stores probably does not add as much to the perceived attractiveness of the cluster as does the addition of the same store to a cluster of 25. Alternatively, a cluster of 25 stores may be perceived as being more than 25 times as attractive as an individual store. In either case, such relationships cannot be modelled with the logit model where the relationship between perceived attractiveness and size is described by slope a. (b)

Perceived

(a)

Agglomeration forces among I outlets

attractiveness of cluster

Logit model assumption

forces

Competition among Small

Small

clusters

FIGURE2. Cluster.

outlets

Large

o0 Namber

Large

Number

clusters

=1

outlets

of

outlets

in cluster

Relationships Between the Perceived Attractiveness of an Outlet Cluster and the Size of the

308


Hierarchical Evaluation of Alternatives If consumers do first evaluate clusters of outlets, the perceived attractiveness of a cluster can, according to equation (4), be described by Vis= af(

Vi)

(22)

jEs

wheref( ) is some functional form to be determined theoretically or empirically and a is a parameter relating Vi1to Vij. In what follows it is necessary to select a specific functional form and, for convenience, assume that the perceived attractiveness of a cluster and the sum of its constituents are related through a power function. That is, Vis= (2 Vi)a.

(23)

jEs

Clearly equation (23) is a more general statement of the relationship between these two variables than is equation (19) in which a is assumed to equal one. Consider now what happens to Visas the size of the cluster increases. Noting equation (18), equation (23) can be rewritten as Vis= naV'

(24)

Vis/ns = an- ' V

(25)

so that which depends critically on a as shown in Figure 2. When 0 < a < 1, the relationship produces slope c. The rate at which the perceived attractiveness of a cluster continues to increase as outlets are added decreases so that some sort of mental discounting takes place. In such a situation there are competition forces among outlets since the addition of a new outlet to the cluster will decrease the market share of existing outlets in the cluster. When a > 1, the relationship produces slope b where the rate at which the perceived attractiveness of a cluster increases at an increasing rate as outlets are added to the cluster. In such a situation, agglomeration forces exist among outlets since the addition of a new outlet to the cluster may increase the market share of existing outlets. When a = 1, the relationship between perceived attractiveness and cluster size that is assumed in the logit model occurs (slope a) and this situation is hence equivalent to that found when all alternatives are evaluated simultaneously. To see the errors that would arise when the logit model is applied to situations where individuals evaluate alternatives hierarchically and when a : 1, compare the slopes in Figure 2. When a > 1 and agglomeration forces exist, the logit model would overpredict the patronage of outlets in small clusters and underpredict the patronage of outlets in large clusters. When 0 < a < 1 and competition forces exist, the logit model would underpredict the patronage of outlets in small clusters and overpredict the patronage of outlets in large clusters. Clearly, these errors have implications for forecasting the performance of a new store. If agglomeration forces are present, the logit model will underpredict the probability of consumers patronizing a new outlet that is to be located relatively close to existing outlets and will overpredict the probability of consumers patronizing an outlet to be located in relative isolation from existing outlets. The converse holds when competition forces are present (Fotheringham 1985). 8.

Summary and Conclusions

Traditional models of consumers' store choice such as the gravity and MCI models share a common multinomial logit framework based on the assumption that consumers evaluate all possible alternatives prior to making a selection. Due to a combination of factors such as lack of information and limits on time and ability to process


309

information,consumersareunlikelyto behavein sucha mannerwhenconfrontedwith large number of stores. It seems unreasonable,for instance, to expect consumersto evaluate all of the alternativesfor groceryshopping in New York City or, for that matter, in a city the size of Gainesville,Florida.The errorsin using the traditional multinomiallogit frameworkto model consumers'storechoices resultingfrom a hierarchicalevaluationprocessaredescribedand areshownto haveimportantimplications for predictingconsumerstorechoice. It is hoped that the resultsin this paper encouragean increasedawarenessof the possibleexistenceof hierarchicaldecision-makingin the choice of retailoutletsand in the existence of fuzzy spatialchoice sets. Evidencefor both can be obtainedby calibratingthe competingdestinationsmodel and findingvaluesof 0 significantlydifferent from0. Some othertypesof evidencethatcan be usedto identifythe misuseof the logit formulationare describedby Fotheringham(1981).2 Acknowledgements. The author would like to thank the referees and the Area Editor for their useful comments on earlier versions of this paper. 2

This paper was received in July 1985 and has been with the author 14 months for 3 revisions.

References Alexis, M., G. Haines and L. Simon (1968), "Consumer Information Processing: The Case of Women's Clothing," in Marketing and the New Science of Planning, R. L. King (ed.), American Marketing Association: Chicago Batsell, R. R. (1981), "A Multiattribute Extension of the Luce Model which Simultaneously Scales Utility and Substitutability," Working Paper, J. H. Jones Graduate School of Administration, Rice University. and J. C. Polking (1985), "A New Class of Market Share Models," Marketing Science, 4, 177-198. Borgers, A. and H. Timmermans (1987), "Choice Model Specification, Substitution and Spatial Structure Effects: A Simulation Experiment," forthcoming in Regional Science and Urban Economics. Dubin, J. A. (1986), "A Nested Logit Model of Space and Water Heat System Choice," Marketing Science, 5, 112-124. Fisher, R. A. and L. H. C. Tippett (1928), "Limiting Forms of the Frequency Distribution of the Largest or Smallest Member of a Sample," Proceedings of the Cambridge Philosophical Society, 24, 180-190. Fotheringham, A. S. (1981), "Spatial Structure and Distance-Decay Parameters,"Annals of the Association of American Geographers, 71, 425-436. (1983), "A New Set of Spatial Interaction Models: The Theory of Competing Destinations," Environment and Planning A, 15, 15-36. (1985), "Spatial Competition and Agglomeration in Urban Modelling," Environment and Planning A, 17, 213-230. (1986), "Modelling Hierarchical Destination Choice," Environment and Planning A, 18, 401-418. (1987), "Market Share Analysis Techniques: A Review and Illustration of Current U.S. Practice," in Store Choice, Store Location and Market Analysis, N. Wrigley (Ed.), at press by Routledge, Keagan and Paul. and D. C. Knudsen (1986a), "Modelling Discontinuous Change in Retailing Systems: Extensions of the Harris-Wilson Framework with Results from a Simulated Urban Retailing System," Geographical Analysis, 18, 295-312. and (1986b), "Modelling Discontinuous Change in the Spatial Pattern of Retail Outlets: A Methodology," pp. 273-292 in Transformations Through Space and Time: An Analysis of Nonlinear Structures, Bifurcation Points and Auto regressive Dependencies, D. A. Griffith and R. P. Haining (eds.), NATO ASI Series, Mortinus Nijhoff: Boston. Gautschi, D. A. (1981), "Specification of Patronage Models for Retail Center Choice," Journal of Marketing Research, 18, 162-181. Guadagni, P. M. (1983), "A Nested Logit Model of Product Choice and Purchase Incidence," in Advances and Practices of Marketing Science, F. S. Zufryden (Ed.), The Institute of Management Sciences: Providence, R.I. Hansen, W. G., "How Accessibility Shapes Land Use," Journal of the American Institute of Planners, 25, 73-76. Huber, J., J. W. Payne and C. Pluto (1982), "Adding a Symmetrically Dominated Alternative: Violations of Regularity and the Similarity Hypothesis," Journal of Consumer Research 9, 90-98.

310


Kahn, B., W. L. Moore and R. Glazer (1985), "An Experiment in Constrained Choice," Working Paper, Graduate School of Management, UCLA. McFadden, D. (1974), "Conditional Logit Analysis of Qualitative Choice Behavior," pp. 105-142 in Frontiers in Econometrics, P. Zarembka (ed.), Academic Press: New York. Meyer, R. J. and T. C. Eagle (1982), "Context-Induced Parameter Instability in a Disaggregate-Stochastic Model of Store Choice," Journal of Marketing Research, 19, 62-71. Miller, E. and S. R. Lerman (1981), "Disaggregate Modelling and Decisions of Retail Firms: A Case Study of Clothing Retailers," Environment and Planning A, 13, 729-746. Moore, W. L., E. A. Pessemier and D. R. Lehmann (1985), "Hierarchical Representations of Market Structures and Choice Processes via Preference Trees," Working Paper, Department of Marketing, University of Utah. Nakanishi, M. and L. G. Cooper (1974), "Parameter Estimation for a Multiplicative Competitive Interaction Model-Least Squares Approach," Journal of Marketing Research, 11, 303-311. Rao, V. R. and D. J. Sabavala (1981), "Inference of Hierarchical Choice Processes from Panel Data," Journal of Consumer Research, 8, 85-96. Recker, W. W. and H. J. Schuler (1981), "Destination Choice and Processing Spatial Information: Some Empirical Tests with Alternative Constructs," Economic Geography, 57, 373-383. Sobel, K. L. (1981), "Travel Demand Forecasting by Using the Nested Multinomial Logit Model," Transportation Research Record, 755, 48-55. Stanley, T. J. and M. A. Sewall (1976), "Image Inputs to a Probabilistic Model: Predicting Retail Potential," Journal of Marketing, 40, 48-53. Williams, P. A. and A. S. Fotheringham (1984), The Calibration of Spatial Interaction Models by Maximum Likelihood Estimation with Program SIMODEL, Geographic Monograph Series, 7, Department of Geography, Indiana University. Zadeh, L. A. (1965), "Fuzzy Sets," Information and Control, 8, 338-353.