Anda di halaman 1dari 10
‘Transm Res. B. Vo. 32. No.8, pp. 29-256, 1998 (© 1958 heir Science Lid All gh reserved Pergamon Prima n Great Bean ongt-26t5/98 $19.0 +000 IT: $0191-2615(98)000803-4 ALTERNATIVE NESTED LOGIT MODELS: STRUCTURE, PROPERTIES AND ESTIMATION FRANK S. KOPPELMAN* and CHIEH-HUA WEN Department of Civil Engineering and Transportation Center, Northwestern University, 2615 Sheridan Road, Evanston, Minis 60208, U.S.A, (Received 19 March 1997; in revised form 30 October 1997) [Abstract —Two distinctly diferent nested logit models have been widely used in both research and applica- tions. The diferences, not widely recognized, between these models will substantially influence estimation resulis, behavioral interpretation and policy analysis. The McFadden nested logit model is derived from ‘andor utility theory; the Daly or non-normalized nested logit mode is based on probability relationships ‘and isnot consistent with uty maximization. This paper describes and compares the model structure and ropeties ofthese diferent nested logit models identifying important diferences between the different model ‘Mructures. An empirical application demonstrates that the diferent nested logit models produce dramatically diferent results with respect to nesting structure and the relative importance of ullty components. Thus, the Selection of one or another of the nesed logit models has important consequences for model interpretation ‘and prediction with consequent impacts on policy analysis. The authors prefer the McFadden model because ofits basis in uty theory, intuitively reazonable elasticity relationships and a cleat interpretation of uty unetin parameters actos altematives. © 1998 Elevier Science Ltd, All rights reserved ‘Keywords neste logit, discrete choice, travel demand, intercity ral 1. INTRODUCTION Discrete choice analysis is used to model the choice of one among a set of mutually exclusive alternatives. The multinomial logit (MNL) model (McFadden, 1973), the most widely used dis- crete choice model, is based on principles of utility maximization and has the advantages of simple mathematical structure and ease of estimation. However, it has the property that the relative probabilities of each pair of alternatives are independent of the presence or characteristics of all other alternatives. This property, known as the independence of irrelevant alternatives (IIA), implies that the introduction or improvement of any alternative will have the same proportional impact on the probability of each other alternative. This representation of choice behavior will result in biased estimates and incorrect predictions in cases which violate the IIA property. Extreme examples of non-IIA cases include the addition of a second bus service identical to an existing service, except for a change in vehicle color, to existing competition between a single bus alternative and private rail (the red bus, blue bus problem) or the addition of a classical CD to an existing choice between another classical CD and a jazz CD. Less extreme, but more realistic ‘examples include mode choice between public transit (bus and light rail) and private vehicle (drive alone and shared ride) (Ortuzar, 1983) or route choice among alternatives which include common road segments (Yai et al, 1997). ‘The most widely known relaxation of the MNL model is the nested (NL) or hierarchical (HL) logit model which allows interdependence between the pairs of alternatives in a common group (McFadden, 1978; Ben-Akiva and Lerman, 1985; Borsch-Supan, 19902), The utility maximizing nested logit (UMNL) model is a special case of the generalized extreme value (GEV) model (McFadden, 1978, 1981) which ensures that it is consistent with utility maximization, provided that the logsum parameters are bounded appropriately. An alternative, is the non-normalized nested logit (NNNL) by Daly (1987), which is identical to the McFadden model except that the ‘model does not include the inverse of each logsum parameter in the utility function of the ‘Author for correspondence, Fax: (647) 491-4011; email: Fkoppeiman@awa.edu 289 20 Frank S. Koppelman and Chich-Hua Wen alternatives in the corresponding nest. This apparently small difference has a dramatic impact on the properties and behavior of this model. Most importantly, the NNNL model is not consistent with utility maximization.* Estimation and application of these different NL models will, in gen- eral, result in different utility function parameters, different nesting structures and different travel forecasts; these differences may be substantial enough to impact capital investment and operating decisions. ‘This paper compares these alternative NL models to identify their important differences and presents an empirical example which demonstrates the extent of these differences which are not widely recognized by users of NL models. The remainder of this paper is organized as follows. Section 2 presents the formulation and properties for the alternative NL models. Section 3 describes the estimation structure for both NL models. Section 4 discusses the empirical example and demonstrates the magnitude of the empirical differences in one case. Section 5 provides a summary and implications for model application, 2. ALTERNATIVE NL MODELS 2.1. Model formulation 2.1.1. The utility maximizing nested logit (UMNL) model. McFadden's UMNL model is a member of the GEV family, and is consistent with utility maximization (McFadden, 1978). A GEV model can be derived from any function, G(¥i, Y¥2,.--, Ya), Yis Yas.+-» ¥a20, which is a non-negative, linear homogeneous function‘ which approaches infinity as any Yi, i=, 2,....m does and has mth cross-partial derivatives which are non-negative for odd m and non-positive for even m. If G satisfies these conditions, the probability of choosing alternative iis given by YG, Yao ++ Yo) PGI Tasos) @ ‘where G; is the first derivative of G with respect to ¥;. A two-level UMNL model is obtained from the G function: ¥ a\ GM Yar Me =| = ¥) @ tne, where Ny is the set of alternatives in nest m and jin is an index of the dissimilarity of alte included in nest m.t The UMNL model is consistent with utility maximization if the conditions, 0 < sim 1,8 are satisfied for all Hm. If #m = 1 for all m, the NL collapses to the MNL. To ensure positive ¥;, the transformation ¥; = exp(¥) is adopted, where V; represents the observable components of the utility for each alternative (U; = Vi + 6;,¢= 1, 2,....7). The prob- ability that alternative m is chosen can be written as Py = Prim X Pm @ where etait Pan = aim O) why ‘Except for the special cane when the lognam parameter are equal acros all ns atthe same level ‘McFadden’ orginal formulation requires the function to be linear homogencou.of degree one; however, Ben-Alkiva and ‘Lerman (1985) refer to an unpublished report by Ben-Akiva and Francois (1983) which demonstrates thatthe function can be homogeneous of any degree. ‘Dissimilarity implies the absence of correlation berween the eror distribution of yrs of alternatives within a nest. The ‘correlation between pairs of alternatives equals (1 ~ 13). ‘values of logeum parameters greater than one can be consistent with wility maxisization over a limited portion of the teal space which may be relevant in cases where the range of attribute variables limited during both estization and ‘prediction (B8rsch-Supan 1980a; Kling and Herriges, 1995; Herrges and Kling, 19), Alternative nested logit models a ote Pa o F wit = and Bain So etme © Vr is the utility associated with alternative m in nest mt, Paym is the conditional probability of choosing alternative n conditional on choosing nest m, Py is the marginal probability of choosing nest m of which n is a member; Nw is the set of all alternatives included in nest m, Pm is the logsum variable of nest m* and sm is the logsum or inclusive value (IV) parameter. Thus, Vj, the utility of nest m is equal t© piml'm.* ‘All alternatives may be considered to be in a nest; in the case of a single alternative in a nest, the nest is considered degenerate and the logsum parameter, Hm, will equal one. If only one alternative is available in each nest, the model collapses to the multinomial logit Ras Seve om 2.1.2. The Daly non-normalized nested logit (NNNL) model. Daly's (1987) NNNL model is defined by the following equations: Pa = Prim X Pm ® where ? Prim = mS he oy whe te —— (10) E emt wt Tm =in Soe” ap WM where the notation is the same as that used in the McFadden NL model. The only difference between the models is that the Daly model excludes the inverse of the IV parameters, jim, in the utility function of the elemental alternatives in each nest (as can be seen by comparing eqns (9) and (1D) to eqns (4) and (6)) The NNNL model is equivalent to the UMNL model only when the jm = 14, for all m in both models. In this case, the utility function parameters in both models will be equal up to the scale ‘Tr represents tbe expected vale ofthe maximum ofthe random utilis of alternatives in nes m, ‘Additional variables which apply 1o all nested alternatives may be included in Va however, this can be identically repre- sented by adding these variables to each alternative inthe nes ‘alto, the inclusion of variables common to all alternatives in a nest at the level of the elemental alternatives is not ‘equivalent to incioding these variables atthe level of the net inthe Daly formation, 22 Frank S. Koppelman and Chich-Hua Wen factor, 2. However, when jim # 14, for any m, the NNNL model is neither equivalent to the UMNL model nor consistent with utility maximization.* As a result, the behavioral interpretation of the NNNL model is based solely on the logic of marginal and conditional choice probabilities and the interpretation of the utility parameters across alternatives is obscure. In particular, the utility value of generic variables, such as time or cost, for different alternatives will depend on the TV parameter associated with the nest of which the alternative is a member. While there are con- ditions under which it would be interesting to consider different utility values of some attributes across alternatives (e.g. the value of travel time on a mode may depend on the extent to which the traveler can use his/her in-vehicle time for other purposes), it is always interesting to test such ‘model specifications against the hypothesis that the utility value of time is constrained to be equal across alternatives. This constraint is even more likely to be relevant for the utility value of money spent on travel by each alternative. 2.2. Relationship among alternatives Both nested logit models allow the user to consider the existence of different competitive rela- tionships between groups of alternatives in a common nest. Such differences indicate that the effect of a change in an attribute of an alternative on the probability of that alternative depends on whether itis or is not in a nest and on other alternatives depends on whether they are or are not in a common nest. One way to describe these differences is in terms of direct- and cross-clasticties of the choice probabilities of the same or other alternatives with respect to @ change in the attri- butes of an alternative. The direct- and cross-elasticities for the MNL and the two NL models are shown in Table 1. All of the elasticity terms are proportional to the magnitude of the variable which changes and the parameter associated with that variable. The direct-elastcities are differ- entiated between alternatives which are or are not in a nest and the cross-elasticities are differ- entiated between pairs of alternatives which are or are not in a common nest. The direct- and cross-elasticities of the MNL model depend exclusively on the probability of the mode which is subject to change, Thus, the MNL cross-clastcities are identical across alternatives and produce the commonly observed equal proportional effect of the addition, deletion or change of any alternative on all other alternatives. The direct elasticities for alternatives not in any nest are structurally identical to the MNL direct-elasticties for both NL models. The UMNL model direct-elastcities for alternatives in a nest are greater than the corresponding direct elasticity for alternatives not in a nest; this rela- tionship is consistent with the idea that alternatives in a nest face a more competitive choice con- text than do alternatives not in any nest. However, the NNNL direct-elastcities are smaller for alternatives in a nest than the corresponding direct-elasticity for alternatives which are not in any nest. These NNNL model differences are counter-intuitive in that they imply that similarity of an alternative to other alternatives reduces the effect of a change in that alternative on its own choice probability relative to the case where that alternative is not in any nest. The magnitude of the difference in direct-lasticties between alternatives in or not in a nest increases as the IV para~ meter, Hm, decreases from one to zero for both models. The cross-elasticities between pairs of alternatives for the two NL models are differentiated between pairs which are or are not in the same nest. The cross-clasticities for pairs of alternatives which are not in the same nest are structurally identical to that for the MNL case for the UMNL model. The corresponding cross-elastcities for the NNNL model differ from the MNL model by the scale parameter, Hm. The cross-clasticties between pairs of alternatives in a common nest are greater for both NL models than the corresponding cross-elasticities for alternatives not in. the same nest, This result is consistent with the notion that cross-lasticities between pairs of alter- natives within a nest should be larger than cross-clasticities between pairs of alternatives not in the same nest. In both cases, the cross-elastcities are equal to those for the MNL when jm is equal to one. The NNNL model doesnot satisfy the naceseary condition thatthe addition ofa constant value fo ll elemental aermatves has no eflect onthe choice probublis of the lternatives (Daly and Zachary, 1978; McFadden, 198). ‘The empirical value of the elasiiies wil dif from the MNT model the uti function parameters change asa result of ‘ change in the nesting structure. For this reason, Section 4 includes empirical comparison between the direct-lasticiies for alternatives in of Rot in a nest and the crose-elastities between alternatives which are of are not inthe same nest within each model Alternative nested logit models ca] ‘Table 1. Ditect-and cross-elastictes ofthe MNL, UMNL and NNNL models Moaet Direrelatiy ‘Croseasicty ructre (change in Pd to change in X.) (change in Py do eage in X,) MOL model (IPA — PA, MeFadden ‘and no in sme nest NE model “Pith min nest m rand in est m (OP GL Dt Pale “feet (At Pn| Bt Daly notin nest, ‘and no in sme net nest model eae whinPoPe et ‘and in nest m [= Pa) + (He = I)Paym(l = PBX a [Pa + (1 ~ He) Pai Pe) B%n 3. MODEL ESTIMATION ‘Maximum likelihood techniques are used to estimate parameters of the MNL and both NL ‘models. The log-likelihood for discrete outcomes is of the form L= DY bln Py a2) vo where 5, is 1 if individual g chooses alternative i and 0 otherwise, Py; is the probability that indi- vidual q chooses alternative i. The likelihood function for the two-level NL model is L= LY un Pa => $5 Sgn tt Pen + F tem in Penn) a3) where Sgn is | if individual g chooses nest m and 0 otherwise, Pm is the probability that individual gq chooses nest m, 5,.m is 1 if individual q chooses alternative m in the nest m and 0 otherwise, and enim is the probability that individual g chooses alternative n in the nest m. “The log-likelihood function of the UMNL model is defined by the following equation: reenter ezeEwzal ‘The log-likelihood function of the NNNL model is cali peeldeel 7 Linas aN, eR The difference between these log-likelihood functions is that the log-likelihood of the conditional probability (the last term in each equation) includes the inverse of the scale parameter in the UMNL model but not in the NNNL model. The differences between these equations will result in different estimations for all of the model parameters. The simplification in the NNNL eqn (15), due to the elimination of uw, appears to make the NNNL model less difficult to estimate than the UMNL model. This enables software for the NNNL model to accommodate a large number of levels in the tree structure, an important capability in some contexts. Estimation of both NL models should use full information maximum likelihood to increase estimation efficiency and allow the user to impose constraints on utility function parameters in different nests in the structure (Brownstone and Small, 1989; Hensher, 1991). Commercial soft- ‘ware packages have been developed to estimate either the UMNL model, the NNNL model ot both, These include ALOGIT (Hague Consulting Group, 1988, 1992 and 1995), HieLoW 24 Frank S. Koppelman and Chieh-Hua Wen (Bierlaire, 1995; Bierlaire and Vandevyvere, 1995), HLOGIT (Bérsch-Supan, 19900) and LIM- DEP (Econometric Software, 1996). The model equations included in software documentation should provide the user with a clear statement of the formulations which are estimated by each package. It is particularly important to recognize that software designed to estimate the NNNL model may be used to estimate the UMNL model by inclusion of dummy nodes and links in the NNNL structure and constraining the IV parameters at such nodes appropriately.* This approach, illustrated in the Fig. 1, where the addition of a dummy link (dashed line) and node and imposition of equality constraints on the IV parameters, enables the estimation of the UMNL ‘model. The resultant utility funetion parameters will be scaled by j. The utility function scaling to match the original tree structure can be calculated by multiplication of the parameters by the scale adjustment. The standard errors of the corrected parameter can be obtained by using Cramer's approximation (Ben-Akiva and Lerman, 1985, p.298). This approach can be applied to any tee structure by adding dummy links and logsum parameters so that the product of the logsum parameters between the root and each elemental alternative is identical. 4, EMPIRICAL EXAMPLE ‘The following empirical example demonstrates the differences between the estimation results for the UMNL and NNNL models.' The data used in this study was assembled by VIA rail in 1989 to ‘estimate the demand for high-speed rail in the Toronto-Montreal corridor and to support future decisions on rail service improvements in the corridor (Forinash and Koppelman, 1993). The choice set includes four intercity travel modes: ai, train, bus and car.* 2779 individuals who have all four modes available were selected. The distribution of choices in the data set is: train (463, 16.66%), air (1039, 37.39%), bus (10, 0.36%) and car (1267, 45.59%) ‘The utility function specification includes mode-specific constants, frequency, travel cost, and out- of-vehicle and in-vehicle travel times. The estimation results for the MNL and the two NL 2 3 (w The nested structure for the simple tree (b) The modified nested structure for the simple tree Fig. 1. Estimation of single node UMNL model using software designed to estimate the UMNL model (a) or software ‘designed to estimate the NNNL mode! (b). ‘Software packages differ in their ablity to Incorporate this approach. The user should ensure that he/she can use parti ‘cular software for ths purpose through examination ofthe documentation, esting and/or verification from the software provider ‘The reported models were estimated using Gauss (Aptech Systems, 1993) based programs developed at Northwestern "University's Transportation Center and verified with HLOGTT for the MeFadden model and ALOGIT for Daly mode! ‘Screening of observations without afl tof alteatives does not impact the substantive results reported inthis paper.

Anda mungkin juga menyukai