Airport Choice in a Multiple Airport Region

8/12/2019 Airport Choice in a Multiple Airport Region

1/11

Transpn. Ru.-A Vol. 21A. No. 6. pp. 439-449. 1987Printed in Great Britain.0191~26U711987 3.00+ .OO0 1987 Pergamon Journals Ltd .

AIRPORT CHOICE IN A MULTIPLEAIRPORT REGION

GREIG HARVEYDepartment of Civil Engineering, Standford University, Stanford, CA 94305, U.S.A.Recei ved 20 M ay 1986; in revi sed form 9 M arch 1987)

Abstrart--This paper reports on an investigation of air traveler behavior in choosing among departureairports in a multiple airport region. Data from a 1980 survey of air passengers in the San FranciscoBay Area were used to study the characteristics of airport choice for local residents. Multinomial logitmodels were developed for business and nonbusiness travelers. The analysis demonstrates that groundaccess time and the frequency of direct air service to the chosen destination can account for a largeportion of the variation in airport usage patterns, but that both access time and frequency have nonlineareffects on airport utility. It also suggests that even multi-stop direct service is strongly preferred toconnecting flights. The results highlight the importance of attention to ground access in planning formultiple airport systems, and the difficulty of predicting airport use without information about market-specific airline schedules.

INTRODUCIIONA number of metropolitan regions have more thanone commercial airport. Where multiple airports ex-ist or are contemplated, it is important to understandhow airlines and air travelers make use of a regionalairport system. Airport utilization results from theinteraction of choices made by airlines (as busi-nesses) and air travelers (as consumers), subject totechnological and institutional constraints. Airlinescheduling in a multiple airport region reflects de-cisions based on anticipated operational efficiencies(such as terminal consolidation and higher load fac-tors) and on improved network connectivity, in lightof consumer preferences. But why do consumers fa-vor one airport over another? What are the char-acteristics of airport location and airline operationsthat affect travelers airport choices?

It has been understood for some time that accesscharacteristics, flight frequency, and service differ-entiation (e.g. pricing) all have an effect on passen-ger distribution among available airports, andnumerous aggregate studies of airport use can befound in the literature. This paper presents the initialresults of an analysis of airport choice within a formalindividual choice framework. Data from the SanFrancisco Bay Area were used to estimate multi-nomial logit models of airport choice as a functionof access time and flight frequency. Particular at-tention was paid to nonlinear effects in the utilityequations. The primary goal of the work was to ex-plore simple hypotheses about time and frequencyin order to build a foundation for later, more detailedand comprehensive studies of air traveler behavior.A secondary goal was to produce information andmodels that could be of use in airport system plan-ning.

CONCEPTUAL FRAMEWORKAn airline trip is considered to result from a set

of choices made by a party of air travelers, where aparty is defined as one or more individuals whojointly plan and execute an air journey. The choicesinclude: whether or not to make an air trip; thedestination of the air trip; time of day to travel;airline; airport; location of departure for airport;fare category; mode of access; and parking option(for auto access modes). Figure 1 illustrates one pos-sible hierarchy of these choices, in which the deci-sions whether, on what days, and where to make atrip are higher level choices (in comparison withchoice of departure airport), and the decisions aboutairline, exact departure time, and access mode arelower level choices.

The choice of departure airport is treated here asa simple process conditioned on the decision to makea trip to a specific destination on a specific date.Airport choice is assumed to depend on personaland airport attributes, including the general char-acteristics of access to each airport. This implies thathigher-level choices have already been determinedby each traveling party, and that lower-level choicesaffect airport choice in general ways only.As Fig. 1 suggests, these structural simplificationsgloss over a number of linkages that may prove im-portant in later research. For example, some trav-elers will have strong preferences for departuretimes, perhaps reflecting desired arrival times attheir destinations. Airport characteristics during thepreferred time intervals will be important to thisgroup, while average daily characteristics may de-termine the behavior of others. Along similar lines,the quality of nonauto access will affect the attrac-tiveness of an airport for some travelers, making

439


2/11

440 G HARVEY

4drpl . . . . . . d>pm

ddl

LevelI

II

III

oaJz-yzz-{oap: . . . . . . oapj

Locatxon of Access DepartureI~lx~~llAccess Mode VIII

.U 0 trrnsi t

dri d ma&r::.

...rt_t5F&erm c2F30+term xFig. 1. A possible decision hierarchy for air travel. Note: Choice of fare category has been omitteddue to probable interactions at multiple levels of the hierarchy.

detailed access characteristics important. In addi-tion, some travelers will develop strong airline pref-erences, which will tend to bias them toward airportsthat are best served by a given airline. Finally, therewill be cases in which travelers choose origin anddestination airports jointly, or perhaps even choosethe destination airport after selecting the origin air-port (i.e. when the destination region has multipleairports as well.)

Exploration of linkages in the choice hierarchy willbe carried out in later stages of the work. The pur-pose here is to see whether the simplest possiblestructural assumption can yield a plausible and ac-curate model of airport choice, and to determinehow the key variables influence airport choice withinthe framework imposed by that assumption.Significant differences in airport choice behaviorarise along at least two dimensions:

(1) Residents vs. nonresidents. Residents of a re-gion have access to more complete information thannonresidents about the available airport options.This information differential may mean that certain

characteristics of airports, such as temporal patternsof access congestion, simply do not influence thedecisions of nonresidents to the same degree as res-idents.

(2) Busi ness vs. nonbusi ness t rav el ers. Businesstravelers may be relatively insensitive to cost becausethey often do not pay their own travel expenses.Also, businessmen often function under tight timeconstraints, making them relatively more sensitiveto travel time and to flight frequency.This necessitates the development of separatemodels for resident business travelers, resident non-business travelers, nonresident nonbusiness travel-ers, and nonresident business travelers. The tworesident categories are analyzed in this paper, whilethe nonresident categories are left for future work.

RELATION TO PREVIOUS WORKDe Neufville has outlined a conceptual model of

the regional airport system that emphasizes the jointrole of air travelers and airlines in determining pat-


3/11

Airport choice in a multiple airport region 441terns of airport use (1976,1984). Invoking data fromprevious studies, he argued that an airports shareof passengers in an air travel market has a nonlinearrelation with the airports share of flights in the mar-ket (1976, pages 69-70); that access distance to anairport affects both airport attractiveness and thedemand for air travel in short-haul markets (1976,pages 63-65); and that high flight concentrations atan airport are attractive enough to outweigh a sub-stantial access disadvantage (1976, pages 67-73).While providing a useful overview of factors inairport competition, deNeufvilles discussion alsohighlights two important data problems: mostinformation about airport competition is highly con-text-specific and thus difficult to generalize; and littleevidence is available about the joint effects of flightfrequency and access time on the behavior of airtravelers.

A number of studies have employed survey dataand aggregate modeling approaches to investigatethe relationships among frequency, access, and air-port market share. In general, however, sample sizeshave been insufficient to support simultaneous dis-aggregation by air travel market and local originzone. One of the few detailed aggregate analyses ofa particular air market was reported by Kanafani, efal (1975), who studied competition among airportsin the heavily traveled San Francisco-Los Angelescorridor. Frequencies and average prices were ob-tained for each of the fifteen routes between the tworegions (using three Bay Area and five Los Angelesairports). Each region was divided into a small num-ber of zones representing ground access origins anddestinations. The distribution of trips over availableroutes for a given origin-destination pair was as-sumed to result from a logit-type allocation process,with a linear utility that included access time (min),frequency (flights per week), and price (dollars).

Separate coefficients were estimated for businessand nonbusiness travelers:

Traveler TypeBusinessNonbusiness

Coefficients oftime frequency fare-.l .003 -.04-.l .002 -.08.

While the fare variable yielded correct signs and theexpected difference between business and nonbusi-ness travelers, its contribution to the explanatorypower of the model was small, likely due to the lackof variation in fares among available routes. Con-versely, access travel time showed strong explana-tory power. Indeed, the magnitude of the travel timecoefficient suggests that access time is the dominantfactor in route choice.

By ignoring ground access, it is possible to deriveuseful insights about frequency from aggregate databased on a much wider range of market conditions.For example, in the context of a study on hubbingat Atlanta, Kanafani and Ghobrial(1985) developeda logit model of air passenger route share betweenP ?(A)21:6 E

city pairs, in which the routing alternatives consistedof nonstop service, direct connection through At-lanta, and indirect connection through Atlanta. Theutility equation included air travel time (min), fre-quency (flights per day), fare (dollars), a dummy foraircraft size (0 for ~30 seats, 1 for ~50 seats, and5 otherwise), and a dummy for quality of connection(1 if there is a change of plane, .5 if not). The es-timated coefficients were:

VariableTravel time (min)Flight frequency (daily)Fare (dollars)Aircraft typeConnection Quality

Coefficient- ,015,239- ,0281.461- 1.557

These values indicate a number of differences fromthe earlier study. The travel time coefficients suggestthat airport access time may be considerably moreonerous than air time (which includes both in-flighttime and ground time at intermediate points). Thefrequency coefficients show an order of magnitude-greater sensitivity (after the earlier coefficients aremultiplied by seven to reflect daily rather thanweekly frequencies). One possible explanation is thedifference in markets treated by the two studies: theAtlanta case offered a wide range of frequencies,while routes in the San Francisco-Los Angeles cor-ridor were densely served. In other words, the dif-ference could be explained by nonlinearities in thefrequency contribution to utility.These aggregate studies provide useful informa-tion about the correlation of flight frequency withair traveler behavior and about the importance ofaccess time in a specific range of market conditions.The purpose of this paper is to amplify and extendthe insights of earlier work by applying discretechoice analysis to the airport choice behavior of asample of individual air travelers. This has the ad-vantage of allowing both markets and access originsto be represented in detail with a relatively smalldata set. By including a wide range of conditions, italso facilitates the exploration of nonlinearities inthe utility function for airport choice.

THE BAY AREA AIRPORT CHOICE SAMPLEAn estimation sample was extracted from air pas-senger survey data collected by the Bay Area Met-ropolitan Transportation Commission in August

1980 (MTC, 1981). The survey was administered tooutbound passengers in the departure lounges of allthree major Bay Area airports-San Francisco In-ternational (SFO), Oakland International (OAK),and San Jose Municipal (SJC). Higher samplingrates were used at SJC and OAK than at SFO. Theresulting choice-based sample required that certainadjustments be made in order to produce unbiasedmodel coefficients, as will be discussed below.The raw survey file from MTC yielded about 2,800observations for residents of the nine-county Bay


4/11

442 G. HARVEYArea. For this study, origin addresses were codedby hand to census tracts. Observations missing crit-ical items (e.g. origin addresses) were removed fromthe estimation file and those remaining were checkedfor inconsistencies. About 1,860 observations sur-vived this culling. A cross-tabulation of the removedobservations did not reveal any strong correlationswith airport choices or with key air travel charac-teristics.

Two types of supplementary data also were as-sembled. Travel characteristics (both peak and off-peak) to each airport from all origin zones wereobtained from the MTC regional model system database. Air service characteristics at each airport wereobtained from the Official Airline Guide (OAG) forthe applicable period in August 1980. For each finalflight destination found in the resident sample, thedata obtained from the OAG included standardcoach (Y Class) fare, number of nonstop flights,number of multi-stop direct flights, number of con-necting flights (explicitly listed in the OAG), numberof commuter flights, and the average flight time ineach category.The MTC survey allowed eight trip purpose re-sponses: convention, business, school, personalemergency, vacation, visiting, military leave, andother. Two categories clearly represent businesstravelers (business and convention) and two clearlyrepresent nonbusiness travelers (vacation and visit-ing). The remaining purposes were included in thenonbusiness subsample. Of the total, 772 observa-tions were business travelers and 1.087 were non-business travelers. This split reflects the timing ofthe survey (during the peak vacation month).

The Bay Area regional airport system is domi-nated by SFO. Of the 27.5 million passengers usingBay Area airports in 1980, 81% were at SFO, 9%were at OAK, and 10% were at SJC. SF0 attracted

passengers from all over the region, while OAK pri-marily served nearby areas of Alameda and ContraCosta Counties, and SJC served Santa Clara Countyand a small portion of southern Alameda County.The survey data do show that both OAK and (to alesser extent) SJC attracted a small but measurablenumber of travelers from all over the region.Survey respondents were asked to state their prin-cipal reason for airport choice: For California Res-idents: Which of the following explains your mainreason for choosing this airport? The answers in-cluded:

1. Chosen by travel agent or office;2. Closest airport to home;3. Closest airport to work;4. Only flight/most convenient flight;5. Easier to get to/from;6. More convenient/cheaper parking;7. Less crowded airport;8. Always use this airport.

Figures 2 and 3 summarize the results, by airportand trip purpose, for Bay Area residents. These re-sults confirm the findings of earlier studies that airtravelers place considerable importance on accesstime and flight frequency in their decisions. For ex-ample, between 74% and 84% of the respondentsin each trip purpose/airport category indicated thataccess (near home, near work, and access) orfrequency was their main criterion for airport selec-tion.Figures 2 and 3 also shed some light on the relativeimportance of flight frequency and access time.Proximity is the dominant factor at OAK and SJC,for both business and nonbusiness travelers. AtSFO, however, about as many travelers list fre-quency as the main criterion as list proximity. One

Agent N Home NWk Flights Access Wrkm9 Crowds Hoblt

(232) (145) (621)Fig. 2. Reason for airport choice (resident business air passengers).


5/11

443irport choice in a multiple airport region70.6

$ 0.5f; 0.4s.E 0.3?II.

0.2

Agmt N HOme NMrk Flights Access Porktng Crowds HabitIZZlSFO 63RY =SJc(357) 13251 (817)

Fig. 3. Reason for airport choice (resident nonbusiness air passengers).would expect this in a situation where travelers aretrading off time and schedule convenience (SF0 hashigher frequencies to most destinations).

MODELS OF AIRPORT CHOICE FOR BAYAREA RESIDENTS

Both the literature on air traveler behavior andtabulations from the Bay Area 1980 survey suggestthat access time and schedule convenience are strongdeterminants of airport choice. Accordingly, atten-tion was focused on these two variables in the de-velopment of model specifications for residentbusiness travelers and resident nonbusiness travel-ers. One group of specifications treated the responseto flight frequency as nonlinear, but represented re-sponse to access travel time as a linear function; asecond group of specifications treated response toboth frequency and travel time as nonlinear. Thebest models of each type are discussed here, alongwith variations in the treatment of variables wheresalient.The basic model structures in all cases is multi-nomial logit with three airport alternatives. Eachalternative has a utility function which includes var-iables for access time and flight frequency, alongwith appropriate alternative-specific constants(which are presented as adjusted to account for thechoice-based sample, mentioned above; see Ben-Akiva and Lerman [1985, p. 2371 for a discussion ofthis procedure). All models were estimated with alogit estimation computer package (Harvey, 1985)utilizing the exogenous sample maximum likelihoodtechnique.Airport choice for resident business travelersTable 1 presents the estimation results for residentbusiness travelers. These can be used to constructthe logit utilities for each alternative. For example,

in model Bl the resident business travelers utilityfor SF0 is:U[SFO] = - .217 - .166*TT[SFO]+3.27*RF[SFO] + .0396*f(DF[SFO]). (1)

Travel time (T7) in this context serves as a proxyfor the bundle of access characteristics at an airport.Since highway modes dominate the airport accessmarket [Harvey, 19861, the appropriate peak or off-peak highway travel times (in minutes) from the re-gional network were used. An increment was addedfor parking time at each airport.In mode1 Bl, the travel time coefficient ( - .166)is large in comparison with typical urban travelchoice models, indicating a strong relationship be-tween proximity and airport choice. This is consis-tent with earlier findings (such as Kanafani, et al .,1975).

Flight frequency appears in the specification intwo different ways. Relative flight frequency RF)is used to capture the information effects (e.g. in-creased advertising) of flight concentration at oneairport. Absolute flight frequency DF) is used as adirect indication of schedule convenience; moreflights to a given destination imply a better matchof flight times with desired departure times.

Relative flight frequency was defined as the frac-tion of direct flights (as listed in the OAG) fromeach airport to the traveling partys destination, in-cluding multi-stop flights that did not require achange of plane. So-called connecting flights (i.e.with a change of plane) were not included in thecount of flights, but were used in the absence ofdirect flights to determine whether a departure air-port was available to the traveler. (By this criterion,Oakland was deemed unavailable to 7% of the busi-ness travelers).


6/11

444 G H RVEYTable 1. Resident business airport choice models

Model Bl Model B2VariableDum(SF0)Dum(OAK)

TTTTs45TT>45,

Overall StatisticsN( SFO)N(OAK)N(SJC)LOLL

Coeff.- ,2171.54- .166

3.27.0396

772155128489-840.4- 710.6- 202.8

Coeff.- .2201.77- .200- .0818.64.0418

(1.57)(7.96)(-13.94)

( ,::;:;(5.17)

- 193.0Rhobar-squared .756 ,767DefinitionsSFO, OAK, and SJC indicate airport alternatives as defined in the text (San Francisco International,Oakland International, and San Jose Municipal).Dum(i)-A constant term in the utility of alternative i.IT-The highway travel time from the traveling partys point of origin to each airport (minutes).7T 5 45-Highway time up to and including the first 45 min of the trip (a cutoff determined em-pirically, as discussed in the text).7T > 45-Highway access time above the first 45 min.RF-The relative frequency of flights to the traveling partys destination at each airport. Connectingand commuter flights are omitted. RF is the number of flights at a given airport divided by the sum offlights at all three airports.W-The daily frequency of flights at each airport to the traveling partys destination. Connectingflights are omitted, as are more than nine daily flights at a given airport (a cutoff determined empirically,as discussed in the text).f(DF)-A parabolic function of DF, constrained to have its maximum at the cutoff point of DPf DF) = 2 9 DF - DF.N(i)-Number of travelers in the sample choosing alternative i.Lo-The sample log-likelihood with all coefficients set to zero.P-The sample log-likelihood with all coefficients set to zero except the alternative specific constants.L*-The sample log-likelihood with the estimated coefficients.Rhobar-squared-tie adjusted likelihood ratio index.

For model Bl, the estimated coefficient of RF was3.27. For typical values of RF, the effect on airportattractiveness appears small in comparison withtravel time. For example, a change from RF = .5to RF = A could be offset by a 2 min decrease inaccess travel time. There are some theoretical rea-sons for omitting RF, including the simultaneity be-tween aggregate shares and relative flight frequency,and the linear relation among RF values for a giventraveler. However, tests of model Bl without RFshowed virtually no effect on the remaining coeffi-cients, suggesting that any specification error due tothis variable does not influence the other coeffi-cients.Direct flight frequency was included in the formof a parabolic variable to capture the decreasingheadway advantage afforded by each additionalflight. The steps leading to the parabolic form areworth describing because they shed light on the na-ture of the relationship between frequency and air-port choice. Initially, DF was included as a singlelinear term without truncation, simply to gauge theaverage effect over the full range of frequency. Its

estimated coefficient was negative. DFthen was bro-ken into three variables based on the number ofintermediate stops-zero, one, and two or more.The resulting coefficients were either negative orindeterminate. To better understand the basis forthis outcome, DF (again aggregated to include alldirect flights) was broken into three linear segments(with varying boundaries). Consistently, the seg-ment coefficients revealed a decreasing marginalutility of frequency that actually became negative atvery high frequencies. Since this could only resultfrom an omitted negative attribute that is stronglycorrelated with high frequency (such as terminalcongestion), it appeared necessary to exclude mar-ginal values of DF above some truncation point.Various truncation values were tried, in conjunctionwith a Box-Tukey transform of DF to find the bestnonlinear form below the truncation point (Hensherand Johnson, 1981, pp., 186-191). The approxi-mately optimal Box-Tukey parameters strongly sug-gested a parabolic form, which was incorporated inBl. With this specification, the optimal truncationpoint was nine flights per day.


7/11

Airport choice in a multiple airport region 445The model Bl function for direct frequency(.0396*[18*DF-OF]) demonstrates the decreasing

marginal utility of flight frequency. It also highlightsthe importance of flight frequency vis-a-vis traveltime. For example, according to this model, a changefrom zero to five direct flights (to a specific desti-nation) would have the same effect on utility as 15.5fewer min of access time (i.e. 2.57/.166 = 15.5).The model thus embodies a strong compensatoryrelation between access time and flight frequency.

The estimated constants (adjusted) are .217 forSF0 and 1.54 for Oakland. There are a number ofpossible reasons for the nonzero constants in thesemodels, and their relative size, including:

(1) Airline fures. Fares have been omitted fromthis analysis because to a specific destination thereis more variation among fare classes on a given flightthan among flights or among airports, and becausenothing is known about the fare actually paid by eachtraveler. At the time of the 1980 survey (before thefull effects of deregulation), Oakland may have hada greater proportion of tourist and charter trafficand thus some price advantage not reflected in themodels.

(2) Avail abil it y of nonauto ground access. SF0and (especially) Oakland are better served by publictransit than San Jose, and SF0 is much better servedby vans, shuttle services, limos, and the like thaneither of the other two. The use of highway data torepresent access time ignores these distinctions.

(3) Congesti on at the air port terminal s. By all ac-counts, some passengers found the terminals at SF0crowded in 1980 (e.g. see Figs. 2 and 3). There isno question that high frequencies are associated withsome kind of disutility (particularly at SFO), but todetermine the source of this disutility (e.g. crowdingwithin the terminals, delays due to airside conges-tion, etc.) would require substantial additional datacollection.

The resulting constants are not unusually large,representing the equivalent of 1.3 and 9.3 min oftravel time savings, respectively. Nevertheless, theywould be an important consideration in any fore-casting application of the model, especially to a newairport (for which the appropriate constant wouldhave to be inferred by analogy to an existing airport).

The standard way of testing a logit model is toperform a likelihood ratio test against a nested spec-ification (i.e. one with a subset of the variables inthe model to be tested). The test measures the ex-planatory power of additional variables. The likeli-hood ratio statistic for model Bl against a modelwith constants only is 1016.0 with three degrees offreedom (the log-likelihood values for this compu-tation are included in Table 1). The correspondingchi-squared (.Ol) value is 11.34. This is simply aformal expression of what is obvious from the esti-mation results: access time and frequency accountfor a large portion of the variation in observed air-

port choices. Indeed, the likelihood ratio index(.756) is large for a model of this type, indicatinghigh correlation between predicted probabilities andobserved choices.

It is instructive to compare the frequency contri-bution to utility in model Bl with the frequencycomponents of the aggregate models discussed inrelation to previous work. Such a comparison is de-picted in Fig. 4 (note that the relative frequency isheld constant here). The results appear quite dif-ferent at first glance, but in fact both of the earliermodels could well be consistent with each other andwith Bl, subject to certain conditions in the esti-mation data sets. Given the relationship shown byBl, a data set dominated by high flight frequencieswould be expected to produce a small linear fre-quency coefficient (such as the upper portion of themodel Bl curve), while a data set reflecting dailyfrequencies between two and twelve might producea linear coefficient close to the Kanafani and Ghob-rial (1985) value.

One feature of model Bl is that an airport withoutdirect major airline service to a destination (but withconnecting or commuter service) has little chance ofbeing chosen by a traveler unless the access timedifferential strongly favors the inferior airport. Thisimplies that connecting flights and commuter flights(either direct or connecting) add little to the attrac-tiveness of an airport as long as an alternative airportwith direct service exists. To further test this prop-osition, two variables representing connecting andcommuter frequencies (as listed in the OAG) wereadded to the model, as fractions of the total numberof flights of all types at an airport (to a specificdestination). This yielded the following coefficients:

VariableConnect frequencyCommuter frequency

Coefficient Estimated t-.909 -1.56-2.41 - 1.42

Other coefficients were virtually identical to thosein Table 1. The likelihood ratio statistic (versusmodel Bl) was 4.6, as compared with a chi-squared. 10) value of 4.61. This suggests that at least a weaknegative effect is present. Thus, it seems likely thatthe competitive situation of an airport without majordirect service is slightly worse than model Bl indi-cates, and that connecting or commuter services ofany type are substantially inferior to direct flights.While model Bls specification for access timeproduces a good t-statistic, and the overall modelperforms well statistically, the linear form used is ahighly simplified representation of access character-istics. This may or may not be a limitation of themodel, depending upon the cognitive processes fol-lowed by air travelers in making access and depar-ture airport decisions. For example, if access modeand airport choices are closely related (i.e. airportchoice is conditioned in part on modal availabilityand quality, and vice versa), then airport choices


8/11

446 G HARVEY

3z. _= 2 nafani and Ghobnol, 1985

1 Kanafani, et al., 1975

0 0 2 4 6 6 10 12 14 16 16Daily Flights

Fig. 4. Frequency components of three utility specifications.may reflect a detailed appreciation of access char-acteristics. On the other hand, if airport choices arebased on only a general impression of modal avail-ability and actual mode choices are settled close tothe time of departure, highway travel time (as anapproximate indicator of access time) may be anadequate representation of the information on accessthat enters the airport choice calculus.

Full resolution of this issue is beyond the scope ofthe current paper. However, a number of relatedquestions can be addressed within the general modelframework presented here. One question derivesfrom an earlier paper on access mode choice(Harvey, 1986), where a strong interaction was iden-tified between length of flight and disutility of accesstravel time. Longer flights were associated withmuch larger (i.e. more negative) access time coef-ficients, possibly due to the increasingly severe con-sequences of missing a flight. If the access mode andairport choices are indeed closely related, then wecan expect to find a similar effect in airport choice.

This proposition was tested by estimating a variantof model Bl having two travel time coefficients: onefor short-haul and one for long-haul. The break pointwas a flight time of two hours for the best flightavailable to a destination. The coefficients were:

Variable Coefficient Estimated tTime, short-haul - .188 - 14.19Time, long-haul - .115 -7.72

The dummy variables were nearly identical to thoseof model Bl, and the frequency coefficients lostabout twenty % of their previous magnitudes. Therho-squared value was .765 and the log likelihoodratio (versus model Bl) was 15.0, compared with achi-squared (.Ol) value of 6.64 for one degree offreedom. While there is clearly some correlationwith frequency (a significant portion of the high fre-quency markets are also short-haul), these results

suggest that access mode choices are not made in asimple joint framework with airport choices.An important question concerns the functionalform of the travel time variable. There are strong

reasons to think that travel time, like frequency,should not enter the utility with a single linear coef-ficient. Air travelers show a willingness to opt forlong access trips-two hours and more-in prefer-ence to use of a local airport with only connectingservice. This is the case, for example, in Bakersfield,where many travelers choose to fly from the LosAngeles International Airport some two hours awayrather than use local connecting flights.

The functional relationship of access time to utilitywas explored through various nonlinear specifica-tions in model B2. Initially, access time was brokeninto a number of segments with separate coefficients.Segment boundaries were varied to maximize thelikelihood function. Estimation results for the besttwo-segment specification are shown in Table 1. Thismodel shows improvement over Bl (according to thelikelihood ratio test), with a strongly diminishingmarginal disutility of travel time.

As a variant of model B2, continuous represen-tations were investigated using a Box-Tukey trans-form. Figure 5 depicts the travel time subfunctionof utility for the best continuous specification (la-beled variant l), along with the linear (Bl) and seg-mented (B2) versions. Statistically, the Box-Tukeyvariant performed slightly worse than the segmentedspecification of model B2.

In general, airport choice for business travel waswell explained by either model Bl or B2 withoutrecourse to personal characteristics. This result mayreflect the fact that business travelers are drawn froma narrow socio-economic range of the population,and for this reason personal characteristics appearto have little effect on airport choice. A number ofmodels were estimated to test for hypothesized ef-fects of personal characteristics such as income and


9/11

Airport choice in a multiple airport region

5 -9--+o--ll--12- Modal B2,Variant 1-13--14--is-- 46-- 17++0 10 20 30 40 60 60 70 60 90

447

Minutes ol AccessTimeFig. 5. Time components of three utility specifications.

auto ownership. None yielded a significant improve-ment over models Bl or B2. A version of model Blwith a separate access time coefficient for those whosaid an agent chose the airport did indicate some-what less sensitivity to access time among this group(- .136 vs. - .180), but neither the t-statistic for thedifference in coefficients nor the likelihood ratio testwas significant.Ai rpor t choi ce for r esident nonbusiness t rav el ers

A parallel analysis was conducted for residentnonbusiness travelers. A comparison of Bl and NBlsuggests major differences between business andnonbusiness travelers. Access time appears to be lessimportant to the nonbusiness traveler, though it stillexerts a strong effect. Direct flight frequency alsoappears to be less important to the nonbusiness trav-eler, though relative frequency has a similar coef-ficient. Experiments with frequency again showed asaturation level of about nine daily flights at an air-port (i.e. the same as for business air travelers). Theconstant terms in NBl and Bl are very close, indi-cating about the same configuration of unexplainedeffects. Overall model statistics indicate that the ex-planatory power of the nonbusiness model is some-what inferior to that of the business model, althoughit still shows good performance.The differences between the business and the non-business travellers responses are consistent with apriori notions about travel behavior. First, since non-business travel uses personal time, the applicablevalue of time would be expected to be lower. Thiswould help to explain lower access travel time andflight frequency coefficients. Second, because non-business travel uses personal funds, cost would beexpected to exert greater influence on airport choice.Hence, the omission of a price variable may reducethe explanatory power of the model. Third, non-business travelers simply may not behave in such anexplicitly rational manner as business travelers. They

may not be as well informed or as experienced, thuslacking knowledge of airport options (e.g. trip fre-quency rates for business travelers are several timeshigher than for nonbusiness travelers); they may be-have by habit, so that an airport used from one originlocation in the past still might be used from lessappropriate locations; or they may respond to ad-vertising and other marketing tools by selecting oth-erwise inferior flight options. Any of thesebehavior patterns would reduce the explanatorypower of a model explicitly incorporating only timeand frequency variables.

Model NB2 repeats the nonlinear access traveltime specification. Unlike model B2, NB2 achievedonly a minor improvement over the linear specifi-cation. A chi-squared test on the difference in log-likelihood between NB2 and NBl indicates signifi-cance at about the .07 level.The effect of income on travel time sensitivity wastested on the nonbusiness sample. Travel time var-iables were created for low ($40K) incomes. The resultsshowed a slightly larger coefficient for high incomethan for the other brackets (which were not signif-icantly different), but a poor likelihood ratio test(versus model NBl). Here, again, it was difficultto demonstrate any relationship between personalcharacteristics and airport choice.One final issue concerns the distinction betweenbusiness and nonbusiness travelers. Differences incoefficient values appear to justify separate treat-ment, but this question can be addressed in a morerigorous way. The basic model (i.e. one with a linearaccess travel time specification as in Bl and NBl)was reestimated on the pooled sample (1,859 ob-servations), yielding a log-likelihood of - 668.8 withfive coefficients. The same specification estimatedon the two samples separately yielded a total log-likelihood of - 658.5 with 10 coefficients (from Ta-bles 1 and 2). The corresponding likelihood ratio


10/11

448 G H RVEYTable 2. Resident nonbusiness airport choice models

Model NBl Model NB2VariableDum(SF0)Dum( OAK)

TTTT=45TT>45RFf DF)

Overall StatisticsObservationsN( SFO)N(OAK)N(SJC)LOLL*Rhobar-squared

Coeff. t- .407 (1.59)1.41 (10.19)- ,120 (-21.61)2.95 (5.43).0176 (3.41)

Coeff.- ,4241.47- ,128- .09163.01.0178

t(1.64)(10.34)

- 18.26)(-6.84)(5.43)(3.42)1087238249600- 1175.- 1072.-455.7 - 454.0.610 ,611

Definitions: All items are defined as in Table 1

statistic is 20.6 with five degrees of freedom. Thiscompares with a chi-squared (.Ol) value of 15.09.The separation of business and nonbusiness travelersappears warranted, although perhaps not as deci-sively as one might expect.

CONCLUSIONThis paper has presented an analysis of airport

choice by residents of the San Francisco Bay Area.The overriding conclusion is that a simple logitmodel based on two variables-airport access timeand flight frequency to the chosen destination-pro-vides a good approximation of airport choice be-havior in the Bay Area sample, especially for busi-ness travelers. The analysis also points to a numberof more specific conclusions:

(1) In most circumstances, the attractiveness ofan airport depends on the specific tradeoff betweentime and frequency faced by an air traveler. TWOextreme cases allow some generalization. Given anadequate number of flights to a destination at eachof two competing airports (nine or more, in thiscase), the closest airport virtually always will be cho-sen (all other things being equal). Conversely, incompetition with a major airport, a facility with littleor no direct service will be bypassed even with sub-stantially shorter access times;(2) Beyond a threshold level, additional directflights to a specific destination do not appear to makean airport more attractive. In the Bay Area data,this threshold is nine flights per day for both businessand nonbusiness travelers. The marginal contribu-tion of each additional direct flight decreases up tonine flights per day. This conclusion does not ruleout secondary effects of frequency on airport at-tractiveness, such as the benefit of more seat de-partures in a market where capacity lags behind

growth in demand. These phenomena simply werenot the focus of the present study;(3) Both connecting and commuter flights appearto be poorly regarded by air travelers who have the

option of selecting a departure airport with directservice;(4) NO effect of nonstop versus multi-stop direct

flights could be discerned. In the choice of airport,the schedule convenience conferred by additionaldirect flights appears to override the time spent onthe ground at interim points;

(5) The marginal disutility of access time appearsto decrease with total time;

(6) The response to access time appears to varywith the length of the flight; short-haul flights havegreater access time sensitivity than long-haul flights;

(7) Airport choice does not appear to take placejointly with access mode choice.

The choice behavior of the nonbusiness travelerdiffers significantly from that of the business trav-eler. Both frequency and access time are less im-portant. In all cases, the above conclusions applyboth to business and to nonbusiness travelers, albeitwith less confidence in the latter case. It is possiblethat other factors may influence nonbusiness airportchoice behavior, perhaps including ticket prices,availability and timing of charter flights, marketingeffects, and other aspects of information availability.It also is possible that business travel decisions arebased on better knowledge and follow a more ex-plicit calculus than nonbusiness decisions.

The results of this study point to a number ofdirections for future work;

(1) extending the analysis to include lower-levelchoices such as access mode in a nested logit frame-work;


11/11

Airport choice in a multiple airport region 449(2) extending the analysis to non-residents. This

would yield a full set of models for building a sim-ulation of a regional airport system;Acknowledgements Thi s work was supported in part bythe National Science Foundation under grant CEE 83-07967.

(3) estimating the same (or similar) models ondata for other area and time periods. This wouldprovide some evidence on the generalizability of thefindings reported here;(4) estimating similar and more sophisticatedmodels on more detailed data, such as a larger sam-ple with adequate representation of costs, chosenfare categories, exact destination locations, and timeof day preferences;(5) designing and conducting special studies of airtravelers to develop a better understanding of theirtravel decision processes, of the nature of their timesensitivity, and of the relationship between regionalairport system characteristics and the demand for airtravel.This is a long-term research agenda, but one with ahigh potential payoff. Air transport is essential topersonal mobility and to productive activity. Whileknowledge about the users of air transportation hasprogressed in the last two decades, there remainsmuch to learn about the behavior of air travelers andabout the effect of this behavior on the structure andevolution of the air transport system.

REFEREN ESBen-Akiva M. and Lerrnan S. R. (1985) Dircrere ChoiceAnalysis: Theory and Application to Travel Demand.MIT-Press, Cambridge, Massachusetts.de Neufville R. (1976) Airport Systems Planning. MITPress, Cambridge, Massachusetts.de Neufville R. (1984) Planning for multiple airports in ametronolitan reaion. Built Envir on. 10. 159-167.Harvey G. (1985)-MicroLOGIT-A muhinomial LOGITestimation program for microcomputer applications. De-partment of Civil Engineering, Stanford University,Stanford, California.Harvey G. (1986) Study of airport access mode choice. J.

Transp. Eng. ll2,525-545.Hensher D. A. and Johnson L. W. (1981) Applied DircreteChoice Modeli ng. Halsted Press, John Wiley & Sons,New York.Kanafani A. and Ghobrial A. A. (1985) Airline hubbing-Some implications for airport economics. Transp. Res.19A, 15-27.Kanafani A., Gosling G. and Taghavi S. (1975) Studies inthe demand for short-haul air transportation. Special Re-port 127. Institute of Transportation and Traffic Engi-neering. Universitv of California. Berkelev. California.Metropolitan Transportation Commission: (1981) SanFrancisco Bay Area 1980 Air Passenger Survey. Met-ropolitanTransportationCommission,Metrocenter,Oak-land, California.

Airport Choice in a Multiple Airport Region

Documents