- Email: [email protected]
Airport size and travel time
Journal of Air Transport Management 32 (2013) 17e23
Contents lists available at SciVerse ScienceDirect
Journal of Air Transport Management journal homepage: www.elsevier.com/locate/jairtraman
Airport size and travel time Tim Hazledine*, Rory Bunker University of Auckland, Auckland, New Zealand
a b s t r a c t JEL classification: L93 R41 Keywords: Airport size Travel time Scale diseconomies
Econometric modelling of the scheduled duration of 2010 flights between 57 origin and 375 destination airports in the year 2009 supports hypotheses (a) that airlines will incorporate realistic predictions of aircraft time on the ground into their published schedules, and (b) that this time will depend positively on airport size, as well as other factors. That is, larger airports generate time diseconomies of scale. A corollary is that actual lateness of flights is not related to airport size. The value of additional time is significant compared with airports’ operating revenues and costs of slot congestion at large airports. Ó 2013 Published by Elsevier Ltd.
1. Introduction Viewed as a part of the urban landscape, perhaps the most striking feature of airports is just how big they are. London’s Heathrow airport, for example, covers twelve square kilometres e more than the city’s largest open area, Richmond Park. Whereas wide open spaces in a park are part of their point, for a commercial transport operation such as an airport, space, in itself, is a nuisance e ground to be covered by people or aircraft at a cost in time, fuel and other resources which of itself contributes nothing to the travellers’ goal of getting from where they do not want to be to where they do as quickly and conveniently as possible. It is a largely unavoidable nuisance, of course, since aircraft need flat space to maneuver in, and even airport terminal buildings cannot sensibly be extended in three dimensions to more than about three floor levels. Nevertheless, given (i) that in general the average distance between two randomly selected points on a twodimensional space increases with the area of the space; (ii) that there has been no empirical investigation of the implications of this for airports, and (iii) airport size is often a private or public investment decision variable (as, for example, when possible expansion is mooted), then it seems reasonable to explore the matter further. We do this by testing the following hypothesis. Airlines do not like running late (or early), because their customers do not like it, and because it creates further operational problems down the line. Accordingly, airlines will tend to incorporate realistic forecasts of the actual time taken by a particular flight in their published schedule
* Corresponding author. Tel.: þ64 93767141. E-mail address: [email protected] (T. Hazledine). 0969-6997/$ e see front matter Ó 2013 Published by Elsevier Ltd. http://dx.doi.org/10.1016/j.jairtraman.2013.06.003
for this. This forecast will be the sum of three elements: expected time on the ground at the origin airport; expected flight time, and expected time on the ground at the destination airport. 1Then, time on the ground at both ends of the flight will be positively related to the size of the airports, because of longer taxiing distances and possibly other factors such as the need to build in a precautionary margin of time because of a higher probability of “lateness” on the part of passengers and/or ground handling operations. We test this hypothesis by estimating econometric models of scheduled flight duration on a cross section of data from a sample of flights and airports around the world, incorporating a measure of airport size as well as other controls (such as, obviously, flight distance) as regressors, and breaking down flight duration into ground and air time. These models are the main outputs of the paper. We also test a corollary of the hypothesis: airport size should not explain late departures, because if it systematically did so, then airlines would be failing to use this information to adjust their schedules to compensate. More informally, we look for evidence that airport size is related to passenger time spent in the terminal; and that size tends to be related to duration and distance of the passenger’s trip to the airport. The normative implications of our results of course depend on the proposition that “time is money”, which places our work in the context of the quite large literature on airport productivity or efficiency. We do not attempt a full survey of this literature in the paper, but will note its salient features. Most studies focus on what could be called “hard” measures of productivity, relating measures
1 By time on the ground we mean time from when the aircraft’s brakes are released for push-off from the gate to time of lift-off from the runway, and equivalently for the arrival.
18
T. Hazledine, R. Bunker / Journal of Air Transport Management 32 (2013) 17e23
of physical output such as passengers handled or aircraft movements to inputs such as terminal area, number of employees, and number of runways. These studies are of two general types. One type seeks to identify the technological frontier, assumed to represent efficient operations, and then calculates the relative (in) efficiency of actual airports as their distance from the frontier (e.g. Gillen and Lall, 1997). The second type of study focuses on the characteristics of airport technologies by means of econometric estimation of production or cost functions e an approach which naturally lends itself to the identification of scale (dis)economies in the production of airport outputs. A recent example which includes a comprehensive literature survey is Martin et al. (2011). These authors find that returns to scale are increasing throughout the size range of their sample of Spanish airports, a result which they say contradicts the ‘standard view’ in the literature that returns to scale are exhausted at fairly low airport sizes.2 A smaller literature explores the “soft” efficiency dimension of use of time: in particular, delays to flights due to airport congestion or other factors. Here it is usually the behaviour or characteristics of airlines, not airports, that is focused on e for example, a tendency for airlines to create congestion by cramming their flights into particular time slots. Santos and Robin (2010), Rupp (2009) and Mayer and Sinai (2003a) link flight delays to use of hub-and-spoke route structures, and to the interesting issue of whether airlines with significant market shares at an airport have an incentive to partially internalize the congestion they cause. The latter has implications for the effectiveness of congestion pricing of airport slots, explored in a number of papers, including Daniel and Harback (2009). We note here that delay e meaning flights departing and/ or arriving later than scheduled e is not necessarily linked to congestion, since the latter can be predicted and allowed for in scheduling. Finally, in an unpublished working paper, Mayer and Sinai (2003b) do empirically model scheduled flight time (i.e., the same variable we will model), and find, with US data, that an airline which has a “hub” at a particular airport is likely to schedule more time for its flights originating at that airport, though the effects are not large (less than a minute, on average). Section 2 describes the data and sources, Section 3 presents the econometric and other empirical results, and Section 4 discusses normative implications of our findings that airport size matters for all three components of a trip: journeys to/from airports; moving through airport terminals, and time in the aircraft on the ground. 2. Database A key source of data was the website www.flightstats.com, on which is recorded information on just about every commercial flight taking off and landing at just about every commercial airport in the world every day of the year. The information includes: operating airline; flight number; equipment type; scheduled and actual departure time; scheduled and (in most but not all cases) actual arrival time at destination. Only direct (non-stop) flights are recorded, and there are no data on numbers of passengers, so the basic unit is the aircraft movement, not the passenger’s journey, which of course may involve stop-overs and changes of plane. The flightstats data are e at least in their public record e ephemeral, being on display for just five dates: two days before; flight date and two days after (at which times any divergences between scheduled and actual arrival times can be observed). We and our research assistants observed the flight data manually,
2 Chang et al. (2013) use a combination of frontier and econometric analysis of the productivity of 41 Chinese airports, and find that airports serving larger cities (populations greater than 2 million) tend to be relatively more efficient.
printing out pages from the website and then recording these on a spreadsheet. This is a quite time-consuming process which eventually yielded data on 2010 actual flights, over various dates in 2009 and 2011. The flight data were supplemented with data culled from search engines, websites and other sources on airport3 and aircraft characteristics. Although we did not undertake any formal randomization of the observation process, we did work with the research assistants to get a good spread of flights: across days of the week; time of day; length of flight; type of aircraft, and size of origin and destination airports. There is some regional bias in the choice of the 57 origin or departure airports: 51% of these are in the United States; 23% in Europe; 16% in Canada, and 10% in Australia or New Zealand. This means that nearly all flights to small airports are also within these regions, which in turn means that English-language information from websites and Wikipedia would likely be available for the small airports, few of which have basic data such as annual numbers of passengers recorded on standard databases. There are 366 different destination airports in the sample, including all of the origin airports. A key concept to quantify is the size or scale of an airport. We are interested in economic size, and will measure this in terms of physical output e the quantities of goods and services handled. A quite standard measure of output in the airport productivity literature is the number of “Workload Units” (WLUs), where one WLU is either one passenger4 handled (either departing or arriving) or 100 kg of freight loaded on or off aircraft. For smaller airports, information on freight handled is often not available from websites or annual reports, and so we will simply use total numbers of passengers as our measure of the functional size of each airport. For the 27 largest US airports, freight handled, in hundreds of kilogrammes, is on average 16% of the number of passengers handled. We do check the correlation between total passengers and two other size-related variables: total commercial aircraft movements,5 and size of the airport terminal in square metres. Ordinary Least Squares results using the EViews 7.2 regression package are shown in Tables 1 and 2. The Table 1 regression has the log of total annual passengers handled by an airport (PAX) explained by the log of total commercial aircraft movements and a dummy variable equal to one if the airport is in North America. The correlation between passengers and movements is very strong and the coefficient (greater than one) implies that, overall, larger airports tend to have larger aircraft (more seats) using them,6 which is unsurprising. The large negative coefficient on the dummy variable tells us, at least in our sample of airports, average aircraft size is smaller in North America, which in turn may reflect a greater tendency there to use air travel, with relatively small aircraft, for shorter trips that in Europe might be carried out by rail. In any case, the regression suggests that passenger numbers are a more generally comparable indicator of the output of an airport than in aircraft movements. The Table 2 model explores the link between output and one of the major airport inputs: terminal size in square metres. Larger flows of passengers do seem to require a larger terminal, though
3 For larger airports, data on passenger numbers are available from the Airports Council International (ACI) World Airport Traffic Report. 4 These are passengers on scheduled or commercial charter flights. This excludes travellers on private aircraft and corporate jets. A “passenger” is someone either taking off or landing at an airport. 5 That is, excluding general aviation movements (e.g. corporate aircraft). A “movement” is an aircraft either taking off or landing at an airport (i.e., a round trip counts as two movements). 6 Dividing number of passengers by number of movements gives the average number of seats utilized per aircraft.
T. Hazledine, R. Bunker / Journal of Air Transport Management 32 (2013) 17e23 Table 1 Number of passengers and aircraft movements. Dependent variable: log(PAX) Coefficient
t-Statistic
Constant Log(MOVEMENT) NA
1.565 1.253 0.586
5.3 49.2 9.7
R-squared (R2) Adj. e R-squared (R2) Observations
0.924 0.924 214
Table 2 Terminal size and number of passengers. Dependent variable: log(TERMINAL) Coefficient
t-Statistic
Constant log(PAX)
2.461 0.863
2.1 12.3
R-squared (R2) Adj. R-squared (R2) Observations
0.605 0.601 101
with a coefficient somewhat less than one, consistent with there being economies of scale in the processes of handling passengers.7 So, measures of passengers using an airport are quite closely related to two other plausible metrics for airport size. A valuable advantage of the passenger numbers variable is its availability e only 214 of the 366 airports in our sample had accessible data on aircraft movements, and for only 101 could we find a measure of terminal building size. The variables used in the main part of the paper are defined as follows: DURATION: Scheduled duration in hours of the flight as reported to flightstats.com, where duration is measured from the scheduled time of pushback from the origin airport gate to the time at which the aircraft brakes are scheduled to be applied at the destination. The flightstats data are in local time, so that time-zone conversions have to be made, for most flights. DIST: Great circle distance between origin and destination airports, in kilometres. PAXO; PAXD: Total number of commercial passengers handled annually (i.e., arrivals plus departures) by origin and destination airport, usually measured for 2009. So, these measures exclude “general aviation” traffic (private planes, charters). PROP: Dummy variable ¼ 1 if aircraft is a turboprop (i.e., not jetpowered). SPEED: Manufacturer’s reported economical cruising speed of the aircraft type; km/h. SEATS: total number of seats (all classes) on the aircraft as usually configured by the airline. LCC: Dummy ¼ 1 if operating airline is a “Low-cost Carrier”. HUBO; HUBD: Dummy ¼ 1 if the origin or destination airport is designated as a “hub” for the operating airline (US only) EAST: Dummy ¼ 1 if the net direction of the flight is easterly. xxORIGIN: Dummy ¼ 1 if the flight originates in region xx. We also culled data from airport websites for all the originating airports, giving estimates of typical travel times and distance
7 Another major input is land. We do not have readily available data on the land holdings of airports, which in any case will be affected by the incidence of “land banking” e holding currently unutilized land to meet expected future expansion needs.
19
between airport and city centre (TIME CITY, DIST CITY). The descriptive characteristics of the data are shown in Table 3. Table 3 tells that the average flight in our database was just over 3 h in scheduled duration, and 2000 km in distance. The average cruise speed of the operating aircraft was 792 km/h, and this aircraft had 133 seats and was in 88% of cases a jet. The operating airline was a low-cost carrier for one flight in five. Table 3 also shows descriptive statistics for two variables measuring the ground journey time from origin airport to the central area of the nearest city (TIMECITY) by the generally quickest mode of transport, and the distance of this trip (DISTCITY). These data were culled from airport and other websites. 3. Results We will first show results for our main variable of interest, scheduled flight duration, then for the other dimensions of the airport size issue. Econometric models are estimated using the EViews 7.2 package. 3.1. Scheduled flight duration First we model total scheduled duration, then break this down to ground and air time. For total flight duration we use a log model to capture non-linearities, shown in Table 4. The least squares model for the log of scheduled flight duration appears to be highly successful. Log(DIST) is hugely significant, with a coefficient well below one, which is likely to be because of on-theground flight time, and, perhaps, time to and from cruising altitude. The size of this coefficient, however, results in the model seriously under-predicting duration of long-haul flights (greater than seven hours). Despite this, the overall goodness of fit is very high for a cross-sectional regression model, with an R2 of 0.97. Both the origin and destination airport size measures (PAXO, PAXD) are significant with positive coefficients, in support of the main hypothesis of this paper. We will return to this. Then we have a number of successful control variables. Bigger aircraft (SEATS) apparently require slightly more generous flight scheduling. Faster aircraft (SPEED) get there more quickly, naturally, and Low-cost carriers (LCC dummy) spend (it seems) less time on the ground. Even controlling for aircraft speed, turboprop aircrafts’ flights (PROP dummy) take longer, which is a surprise. The prevailing winds around the Earth are such that Easterly flights are faster (EAST dummy). Perhaps surprisingly, a (non-stop) flight on an airline which has a major Hub (in the US) takes less time, other things equal.8 Of the regional dummies, only that for New Zealandoriginating flights (NZORIGIN) was significant, with a negative coefficient, perhaps reflecting the more relaxed security procedures for domestic flights within NZ.9 The regression model shown in Table 4 is encouraging in that it shows we have found a set of variables significantly related to scheduled flight duration, including the airport size measures that are our main point of interest.10 However the single-equation specification is awkward, given that determinants of the two
8 Of course one-stop journeys using the hub-and-spoke system are likely to take longer than direct flights. 9 Turboprop flights within NZ have no security checkpoints, and even for jet flights check-in on Air New Zealand does not close until 20 min before scheduled departure time. 10 The regression model of Table 4 was arrived at after a limited amount of experimentation with different sets of regressors, including the other regional dummies, and unsuccessful attempts to identify a time-of-day effect on flight duration, with “peak” period departures and/or arrivals expected to be scheduled to take longer.
20
T. Hazledine, R. Bunker / Journal of Air Transport Management 32 (2013) 17e23
Table 3 Descriptive statistics for airports and flights.
Mean Min Max
DURATION h
DIST km
PAXO number
PAXD number
PROP dummy
SPEED km/h
SEATS number
LCC dummy
TIME CITY min
DIST CITY km
3.1 0.4 15.9
2011 73 13,552
32,765,473 194,864 84,846,639
19,984,785 3354 84,846,639
0.12 0 1
792 440 907
133 8 517
0.20 0 1
27 6 60
19 4 45
Table 4 Modelling total scheduled flight duration. Dependent variable: log(DURATION) Coefficient
t-Statistic
Constant log(DIST) log(PAXO) log(PAXD) SEATS SPEED LCC PROP EAST HUBO HUBD NZORIGIN
4.487 0.686 0.037 0.018 0.000 0.001 0.058 0.103 0.054 0.016 0.044 0.066
51.4 172.3 12.0 9.9 8.4 7.1 9.0 4.4 10.4 1.9 4.3 5.3
R-squared (R2) Adj. R-squared (R2) Observations
0.972 0.972 2009
additive components of flight duration e ground time and flying time e are likely to be quite different. For example, the log specification generates a constant elasticity response of duration to airport size, implying a smaller absolute effect on shorter flights, which may not be true. Accordingly, we now break down total scheduled duration into those two components, and model them separately. We do not have any direct information on the size of the two components in each airline’s flight time scheduling computations, but if we could estimate a simple linear model of duration, using only determinants of in-the-air time (flying time) as regressors, then we could reasonably interpret the intercept of this model as an estimate of average ground time. So, we posit that the major determinant of flying time will be distance to be travelled divided by aircraft speed, and estimate the models shown in Table 5. The first model shows distance divided by aircraft type cruise speed is a hugely successful predictor of duration, with a coefficient very close to one. Slightly “tweaking” this variable based on Table 4 results, allowing for a five percent direction speed bonus for Easterly flights further improves the fit. In both variants the intercept is around 0.63e0.65, which we interpret as an average total ground time of nearly two thirds of an hour, or a bit less than 20 min in each airport, which seems reasonable. Accordingly, we take the predicted duration from the second Table 5 model, subtract 0.64 from this, and then subtract the result from actual scheduled duration to get an estimate of
(scheduled) time on the ground, multiplied by 60 to turn it into minutes, for easier interpretation. This variable is then dependent in a linear model bringing in all the putative ground timedeterminants as regressors. In this model we include quadratic terms for the two airport size variables, to allow for possible nonlinear effects. Table 6 shows three estimated versions of this model: first with the full sample, then with flights longer than six hours removed, and then with fixed effects for the origin airport. Not surprisingly, since we are now trying to explain variation in a calculated variable that is the difference between two variables, overall explanatory power is much lower than in the previous models, and we do not find so many strongly significant regressors. Eliminating long-haul flights, for which predicted ground time has quite a large variance, including some negative numbers, improves the model. The truncated-sample model tells us that LCCs spend about three minutes less on the ground11 than legacy carriers; prop planes are nearly 7 min quicker (perhaps due to quicker taxis to/ from smaller specialised terminals); there is no origin-hub effect, but flights headed for an airline’s hub as destination save about 4 min on the ground; New Zealand and US origin flights are about two and four minutes faster and slower, respectively. As regards airport size impacts, the quadratic effects on groundtime of origin and destination airport size are concave, and positive up to total annual millions of passenger numbers of around 57 and 58, for origin and destination airports respectively.12 In terms of the size of these effects, a doubling of both origin and destination airport sizes from the sample mean of around 20 million passengers/year to 40 million would increase predicted on-the-ground time for each flight by about 6 min, with one third of this at the origin airport and two thirds at the destination. We will discuss in Section 4 the normative significant of these numbers. The third model in Table 6 has fixed effects for the origin airport. Estimation of this model requires dropping any variables which are constant for each airport, so both the regional origin dummies and total origin airport passenger numbers are excluded, with their effects now subsumed in the fixed effects. The jump in the R2 over the second model tells us that (not surprisingly) there are some airport-related explanatory factors which we have not measured specifically in this study, although the coefficients of the included variables are reasonably stable, with the exception of the LCC dummy, of which the coefficient falls away in size and significance, suggesting that the airports used by low-cost carriers tend to have some characteristics that differ from other airports.13 In what follows we will work with the second Table 6 model, which gives us estimates of both airport size effects.
Table 5 Flying time models. Dependent variable DURATION Coefficient
t-Statistic
Coefficient
t-Statistic
Constant DIST/SPEED
0.647 1.018
73.6 411.6
0.634 1.045
81.4 464.1
R-squared Adj. R-squared Observations
0.988 0.988 2010
0.991 0.991 2010
11 Note that this is additional to the well-known “turn-around” time advantage of LCCs, whereby time between flights is reduced by less time spent restocking galleys or handling baggage or passengers in transit, as well as less need to bunch arrival and departure times of flights to fit hub connecting flights. 12 The positive range covers most of the sample. Only three of the 56 origin airports (LAX, ATL, LHR) processed more than 57 million passengers in 2009. 13 In the US and Europe, but not in Australia, Canada or New Zealand, low-cost carriers often use smaller, less congested “secondary” airports, these often being sited quite far from the city that is the ostensible origin or destination of the flight.
T. Hazledine, R. Bunker / Journal of Air Transport Management 32 (2013) 17e23
21
Table 6 Ground-time models. Dependent variable: GROUNDTIME*60 Flights <6 h
Full sample
Origin fixed effects
Coefficient
t-Statistic
Coefficient
t-Statistic
Coefficient
t-Statistic
C LCC PROP HUBO HUBD NZORIGIN USORIGIN PAXO PAXO^2 PAXD PAXD^2
31.106 2.661 6.868 0.487 1.064 5.208 6.449 0.090 0.001 0.316 0.003
26.4 3.1 6.1 0.4 0.8 3.3 7.5 1.7 1.6 6.3 4.4
29.408 2.964 6.962 0.145 4.064 1.999 4.200 0.199 0.002 0.426 0.004
32.5 4.6 8.2 0.2 3.7 1.6 6.2 4.8 4.0 10.7 6.6
34.731 0.767 6.487 0.146 2.601
72.4 1.2 8.5 0.2 2.7
0.380 0.003
10.5 6.0
R-squared Adj. R-squared Observations
0.137 0.132 2009
3.2. Is lateness ‘scheduled’? We have found that airlines tend to build into their schedules extra time on the ground at larger airports, but whether this is enough time needs to be considered. If not, then we could expect to find some relationship between late departures and airport size. For official statistical and reporting purposes, “late” is defined as a flight pushing off from the gate 15 or more minutes after the scheduled departure time.14 Of the 1934 flights in our sample for which actual departure times were noted on flightstats.com, just over 14% were officially late. A Logit model predicting the probability of late departure (DLATE ¼ 1) found no significance in the overall airport size variable (PAXO). If this is replaced by dummy variables for airport sizes in ranges of 10 million passengers/year, then only the dummy for the largest airports (greater than 50 million passengers) showed any significance. This model is shown in Table 7. That is, there is apparently some tendency for airlines to be optimistic about delays at the very large airports (i.e., to not fully allow for the likelihood of delays in their scheduling), but not otherwise. Our calculated scheduled ground-time variable has no effect on actual lateness, consistent with airlines being fully aware of the determinants of time on the ground at each airport (excepting perhaps the very largest). The probability of late departure is higher in the late afternoon/early evening peak period, presumably because of the cumulative effect of delays earlier in the day.15 3.3. Passenger ground time Finally, we move from what happens after the cabin door is closed on an aircraft to what happens before that event, from the passengers’ perspective. Specifically, we look at time getting to (or from) the airport, and time spent in the airport terminal. Our empirical strategy here in both cases is to use readily available proxies rather than attempt to construct comprehensive data. So, to estimate passengers’ ground travel time accurately would require either surveys of actual travellers, or elaborate
14 In the US, the Bureau of Transportation Statistics (BTS) defines “departure” as “the instance when the pilot releases the aircraft parking brake after passengers have loaded and aircraft doors have been closed”. Arrival time is defined analogously (www.transtats.bts.gov/glossary.asp). 15 PMPEAK takes the value 1 for flights scheduled to depart after 3pm and before 8pm.
0.239 0.235 1787
0.450 0.430 1787
manipulations of (census) data on the distribution of population in the airport’s catchment and distances (and mode of travel) to the airport from each location. This is a major data gathering exercise even for a single airport.16 However, there is a plausible single number proxy for airport accessibility, which is simply the distance of the airport from the Central Business District of the major city it serves. This figure can invariably be quickly culled from airport or other websites, and this fact alone suggests that airports know that distance from downtown is a useful piece of information for travellers. Our hypothesis is that the economics of land prices will tend to push larger airports to locate further away from the CBD, which we test for our group of 57 origin airports, with results in Table 8. There is indeed a moderately robust relationship, with an elasticity of distance with respect to airport size of 0.165. If we use city population instead of airport size as the regressor, we get an almost identical result, which of course reflects the tendency of larger cities to build larger airports. Actual time spent getting to the airport will depend, as well as on distance, on the transport infrastructure e whether there is a large capacity freeway, and/or whether there is direct train service between the CBD and the airport. We can get rather approximate estimates from airport websites of the likely minimum journey time from city centres, and we model this in Table 9. Here, size of airport is still significant, and there is some evidence that trains are faster than buses or taxis. However, these two variables are not independent: if we do a simple Probit model of the probability of a direct airport-CBD rail link with airport size as the regressor, there is indeed a significant positive relationship. This is presumably because the high fixed costs of rail service tend to need a large airport with many passengers to justify them. So, the Table 9 regression model implies that a doubling of airport size goes with a 12% longer travel time to the airport e perhaps two or three minutes onto the average travel time, which is 23 minutes e unless the larger airport now has train service, in which case the net effect is a small decrease in the predicted ground travel time. Secondly, we explore passenger time spent in the airport. Personal experience has shown that the larger terminals of larger airports result in longer walks to and from gates, but this has to be documented systematically. As with travel to/from airports, generating accurate data would require a substantial amount of work for each case, perhaps entailing customer surveys and/or time
16 Ishii et al. (2009) make good use of a survey of passengers travelling in the San Francisco/Oakland Bay area.
22
T. Hazledine, R. Bunker / Journal of Air Transport Management 32 (2013) 17e23
Table 7 Modelling late departures.
Table 9 Time to airport from city centre. Dependent variable: DLATE
Constant PAXTD50 HUBO GROUNDTIME NZORIGIN PMPEAK McFadden R-squared Observations Procedure:
Dependent variable: log(TIMECITY)
Coefficient
z-Statistic
1.853 0.481 0.308 0.021 0.520 0.375
10.0 2.7 1.5 0.1 1.6 2.4
Coefficient
t-Statistic
Constant LOG(PAXTOTO) TRAIN
1.083 0.126 0.217
1.4 2.7 1.6
R-squared Adj. R-squared Observations
0.118 0.085 57
0.011 1932.0 Binary logit
Table 8 Distance to airport from city centre. Dependent variable: log(DISTCITY) Coefficient
t-Statistic
C LOG(PAXTOTO)
0.094 0.166
0.1 3.3
R-squared Adj. R-squared Observations
0.164 0.149 57
and motion studies. Instead we suggest that airlines’ minimum check-in times should be a proxy, since these will be set in the knowledge of how long it will likely take a passenger to get from the check-in counter to the departure gate. In our sample of 57 origin airports, 30 specify a 30 min minimum or recommended check-in time (i.e. check-in at least 30 min before scheduled departure time); 13 asked for 40 min, and the remaining 14 specified either 45 or 60 min. The average size of these three groups, in terms of annual passengers handled, was 10.8 million, 22.2 million and 39.7 million, respectively. That is, while we have not done any significance testing on these numbers, they do support the reasonable hypothesis that there is a tendency for larger airports to require passengers to spend more time in the terminal.17 4. Summary and implications We have found solid econometric evidence of diseconomies of scale with respect to air travel time, generated by larger airports. For example, the journey of a traveller taking off from and landing at airports handling 40 million passengers/year would tend to be scheduled by the airline at about 6 min longer total duration than a journey between two 20 million passengers/year airports, other things equal. Less formally established are our findings that a traveller’s trip to the airport from downtown would likely be about two minutes longer in the case of the larger airport, and this passenger would have quite likely been required to check-in for their flight at least five minutes earlier. These figures are statistically significant, but there is the question of whether there are they also economically significant. That is, are the time effects large enough to materially affect traveller wellbeing? We will not attempt a formal cost-benefit analysis, but will give some idea of the scale of the scale effects relative to two benchmarks: the operating revenues of airports, and the calculated welfare savings from implementing optimal congestion pricing at
17 These are minimum recommended check in times. Many airports (and airlines) suggest allowing more time for checking in to international flights. Note that, in our sample, the largest airports (more than 40 million passengers/year) have, on average, the same proportions of domestic flights (55%) as the full sample.
US airports, from Daniel and Harback (2009). These authors calculate welfare savings for the 27 largest US airports (using 2003 data) at about $4,500,000/day, or around $60 million/year per airport, on average. We have some data for these airports for 2004, and that year the average large airport handled 35 million passengers, and had total18 operating revenue of $289 million. Continuing with our comparison between a forty million passenger airport and two twenty million passenger airports hypothetically carrying the same travellers, our figures imply total additional annual passenger travel time in the former case of ten minutes per passenger,19 or more than six million hours e worth, say, $180 million, at an hourly rate of $30. On top of this should be added additional costs to the airlines of longer taxi times, etc, which Morrison and Winston’s (2008) estimates of costs of flight delays (late departures) put at about 80% of the on-board passenger time costs. So, relative either to the economic size of airports measured by total revenues, or to the estimated costs of congestion-induced delays, our scale effects are quite large and undoubtedly “significant” in their size. However, we should note unresolved difficulties in valuing traveller time in these contexts. Many may be likely to agree on the irksomeness of ground travel to and from the airport20, and of trudging through the terminal. However it is harder to assign a value to on-board time. For some travellers, each minute trapped inside an aircraft may be another minute of claustrophobic terror; for others it may be a relaxing break, well away from outside cares and responsibilities; and, for some, flying may remain an interesting, fun experience. We note, too, limitations on the generality of our econometric analysis, based as it is on a sample of just 57 of the world’s origin airports (and 375 flight destinations), and around 2000 actual flights. However, given the statistical and economic significance of our quantitative results, we do consider it reasonable to recommend that any cost benefit analysis of a major urban airport construction or extension project with spatial implications be required to include consideration of likely time diseconomies of scale along with the other relevant factors.
References Chang, Y.-C., Yu, M.-M., Chen, P.-C., 2013. Evaluating the performance of Chinese airports. Journal of Air Transport Management 31, 19e21. Daniel, J., Harback, K., 2009. Pricing the major US hub airports. Journal of Urban Economics 66, 33e56.
18 Operating revenue is generated about 60% from aeronautical services (e.g., landing fees) and 40% from terminal income (retail concessions, parking), on average. 19 10 ¼ 6/2 þ 2 þ 5. 20 Ishii et al. find evidence from trips taken by San Francisco Bay Area residents that time costs associated with airports are significant determinants of airport choice; in particular, ‘distaste for access time’ (2009, p225).
T. Hazledine, R. Bunker / Journal of Air Transport Management 32 (2013) 17e23 Gillen, D., Lall, A., 1997. Developing measures of airport productivity and performance: an application of data envelopment analysis. Transportation Research Part E 33, 249e260. Ishii, J., Jun, S., Van Dender, K., 2009. Air travel choices in multi-airport markets. Journal of Urban Economics 65, 216e227. Martin, J., Roman, C., Voltes-Dorta, A., 2011. Scale economies and marginal costs in Spanish airports. Transportation Research Part E 47 (2), 238e248. Mayer, C., Sinai, T., 2003a. Network effects, congestion externalities, and air traffic delays: or why not all delays are evil. American Economic Review 93 (4), 1194e1215.
23
Mayer, C., Sinai, T., 2003b. Why Do Airlines Systematically Schedule Their Flights to Arrive Late?. Working Paper: The Wharton School, University of Pennsylvania. Morrison, S., Winston, C., 2008. The effect of FAA expenditures on air travel delays. Journal of Urban Economics 63, 669e678. Rupp, N., 2009. Do carriers internalize congestion costs? Empirical evidence on the internalization question. Journal of Urban Economics 65 (1), 24e37. Santos, G., Robin, M., 2010. Determinants of delays at European airports. Transportation Research Part B 44, 392e403.
Contents lists available at SciVerse ScienceDirect
Journal of Air Transport Management journal homepage: www.elsevier.com/locate/jairtraman
Airport size and travel time Tim Hazledine*, Rory Bunker University of Auckland, Auckland, New Zealand
a b s t r a c t JEL classification: L93 R41 Keywords: Airport size Travel time Scale diseconomies
Econometric modelling of the scheduled duration of 2010 flights between 57 origin and 375 destination airports in the year 2009 supports hypotheses (a) that airlines will incorporate realistic predictions of aircraft time on the ground into their published schedules, and (b) that this time will depend positively on airport size, as well as other factors. That is, larger airports generate time diseconomies of scale. A corollary is that actual lateness of flights is not related to airport size. The value of additional time is significant compared with airports’ operating revenues and costs of slot congestion at large airports. Ó 2013 Published by Elsevier Ltd.
1. Introduction Viewed as a part of the urban landscape, perhaps the most striking feature of airports is just how big they are. London’s Heathrow airport, for example, covers twelve square kilometres e more than the city’s largest open area, Richmond Park. Whereas wide open spaces in a park are part of their point, for a commercial transport operation such as an airport, space, in itself, is a nuisance e ground to be covered by people or aircraft at a cost in time, fuel and other resources which of itself contributes nothing to the travellers’ goal of getting from where they do not want to be to where they do as quickly and conveniently as possible. It is a largely unavoidable nuisance, of course, since aircraft need flat space to maneuver in, and even airport terminal buildings cannot sensibly be extended in three dimensions to more than about three floor levels. Nevertheless, given (i) that in general the average distance between two randomly selected points on a twodimensional space increases with the area of the space; (ii) that there has been no empirical investigation of the implications of this for airports, and (iii) airport size is often a private or public investment decision variable (as, for example, when possible expansion is mooted), then it seems reasonable to explore the matter further. We do this by testing the following hypothesis. Airlines do not like running late (or early), because their customers do not like it, and because it creates further operational problems down the line. Accordingly, airlines will tend to incorporate realistic forecasts of the actual time taken by a particular flight in their published schedule
* Corresponding author. Tel.: þ64 93767141. E-mail address: [email protected] (T. Hazledine). 0969-6997/$ e see front matter Ó 2013 Published by Elsevier Ltd. http://dx.doi.org/10.1016/j.jairtraman.2013.06.003
for this. This forecast will be the sum of three elements: expected time on the ground at the origin airport; expected flight time, and expected time on the ground at the destination airport. 1Then, time on the ground at both ends of the flight will be positively related to the size of the airports, because of longer taxiing distances and possibly other factors such as the need to build in a precautionary margin of time because of a higher probability of “lateness” on the part of passengers and/or ground handling operations. We test this hypothesis by estimating econometric models of scheduled flight duration on a cross section of data from a sample of flights and airports around the world, incorporating a measure of airport size as well as other controls (such as, obviously, flight distance) as regressors, and breaking down flight duration into ground and air time. These models are the main outputs of the paper. We also test a corollary of the hypothesis: airport size should not explain late departures, because if it systematically did so, then airlines would be failing to use this information to adjust their schedules to compensate. More informally, we look for evidence that airport size is related to passenger time spent in the terminal; and that size tends to be related to duration and distance of the passenger’s trip to the airport. The normative implications of our results of course depend on the proposition that “time is money”, which places our work in the context of the quite large literature on airport productivity or efficiency. We do not attempt a full survey of this literature in the paper, but will note its salient features. Most studies focus on what could be called “hard” measures of productivity, relating measures
1 By time on the ground we mean time from when the aircraft’s brakes are released for push-off from the gate to time of lift-off from the runway, and equivalently for the arrival.
18
T. Hazledine, R. Bunker / Journal of Air Transport Management 32 (2013) 17e23
of physical output such as passengers handled or aircraft movements to inputs such as terminal area, number of employees, and number of runways. These studies are of two general types. One type seeks to identify the technological frontier, assumed to represent efficient operations, and then calculates the relative (in) efficiency of actual airports as their distance from the frontier (e.g. Gillen and Lall, 1997). The second type of study focuses on the characteristics of airport technologies by means of econometric estimation of production or cost functions e an approach which naturally lends itself to the identification of scale (dis)economies in the production of airport outputs. A recent example which includes a comprehensive literature survey is Martin et al. (2011). These authors find that returns to scale are increasing throughout the size range of their sample of Spanish airports, a result which they say contradicts the ‘standard view’ in the literature that returns to scale are exhausted at fairly low airport sizes.2 A smaller literature explores the “soft” efficiency dimension of use of time: in particular, delays to flights due to airport congestion or other factors. Here it is usually the behaviour or characteristics of airlines, not airports, that is focused on e for example, a tendency for airlines to create congestion by cramming their flights into particular time slots. Santos and Robin (2010), Rupp (2009) and Mayer and Sinai (2003a) link flight delays to use of hub-and-spoke route structures, and to the interesting issue of whether airlines with significant market shares at an airport have an incentive to partially internalize the congestion they cause. The latter has implications for the effectiveness of congestion pricing of airport slots, explored in a number of papers, including Daniel and Harback (2009). We note here that delay e meaning flights departing and/ or arriving later than scheduled e is not necessarily linked to congestion, since the latter can be predicted and allowed for in scheduling. Finally, in an unpublished working paper, Mayer and Sinai (2003b) do empirically model scheduled flight time (i.e., the same variable we will model), and find, with US data, that an airline which has a “hub” at a particular airport is likely to schedule more time for its flights originating at that airport, though the effects are not large (less than a minute, on average). Section 2 describes the data and sources, Section 3 presents the econometric and other empirical results, and Section 4 discusses normative implications of our findings that airport size matters for all three components of a trip: journeys to/from airports; moving through airport terminals, and time in the aircraft on the ground. 2. Database A key source of data was the website www.flightstats.com, on which is recorded information on just about every commercial flight taking off and landing at just about every commercial airport in the world every day of the year. The information includes: operating airline; flight number; equipment type; scheduled and actual departure time; scheduled and (in most but not all cases) actual arrival time at destination. Only direct (non-stop) flights are recorded, and there are no data on numbers of passengers, so the basic unit is the aircraft movement, not the passenger’s journey, which of course may involve stop-overs and changes of plane. The flightstats data are e at least in their public record e ephemeral, being on display for just five dates: two days before; flight date and two days after (at which times any divergences between scheduled and actual arrival times can be observed). We and our research assistants observed the flight data manually,
2 Chang et al. (2013) use a combination of frontier and econometric analysis of the productivity of 41 Chinese airports, and find that airports serving larger cities (populations greater than 2 million) tend to be relatively more efficient.
printing out pages from the website and then recording these on a spreadsheet. This is a quite time-consuming process which eventually yielded data on 2010 actual flights, over various dates in 2009 and 2011. The flight data were supplemented with data culled from search engines, websites and other sources on airport3 and aircraft characteristics. Although we did not undertake any formal randomization of the observation process, we did work with the research assistants to get a good spread of flights: across days of the week; time of day; length of flight; type of aircraft, and size of origin and destination airports. There is some regional bias in the choice of the 57 origin or departure airports: 51% of these are in the United States; 23% in Europe; 16% in Canada, and 10% in Australia or New Zealand. This means that nearly all flights to small airports are also within these regions, which in turn means that English-language information from websites and Wikipedia would likely be available for the small airports, few of which have basic data such as annual numbers of passengers recorded on standard databases. There are 366 different destination airports in the sample, including all of the origin airports. A key concept to quantify is the size or scale of an airport. We are interested in economic size, and will measure this in terms of physical output e the quantities of goods and services handled. A quite standard measure of output in the airport productivity literature is the number of “Workload Units” (WLUs), where one WLU is either one passenger4 handled (either departing or arriving) or 100 kg of freight loaded on or off aircraft. For smaller airports, information on freight handled is often not available from websites or annual reports, and so we will simply use total numbers of passengers as our measure of the functional size of each airport. For the 27 largest US airports, freight handled, in hundreds of kilogrammes, is on average 16% of the number of passengers handled. We do check the correlation between total passengers and two other size-related variables: total commercial aircraft movements,5 and size of the airport terminal in square metres. Ordinary Least Squares results using the EViews 7.2 regression package are shown in Tables 1 and 2. The Table 1 regression has the log of total annual passengers handled by an airport (PAX) explained by the log of total commercial aircraft movements and a dummy variable equal to one if the airport is in North America. The correlation between passengers and movements is very strong and the coefficient (greater than one) implies that, overall, larger airports tend to have larger aircraft (more seats) using them,6 which is unsurprising. The large negative coefficient on the dummy variable tells us, at least in our sample of airports, average aircraft size is smaller in North America, which in turn may reflect a greater tendency there to use air travel, with relatively small aircraft, for shorter trips that in Europe might be carried out by rail. In any case, the regression suggests that passenger numbers are a more generally comparable indicator of the output of an airport than in aircraft movements. The Table 2 model explores the link between output and one of the major airport inputs: terminal size in square metres. Larger flows of passengers do seem to require a larger terminal, though
3 For larger airports, data on passenger numbers are available from the Airports Council International (ACI) World Airport Traffic Report. 4 These are passengers on scheduled or commercial charter flights. This excludes travellers on private aircraft and corporate jets. A “passenger” is someone either taking off or landing at an airport. 5 That is, excluding general aviation movements (e.g. corporate aircraft). A “movement” is an aircraft either taking off or landing at an airport (i.e., a round trip counts as two movements). 6 Dividing number of passengers by number of movements gives the average number of seats utilized per aircraft.
T. Hazledine, R. Bunker / Journal of Air Transport Management 32 (2013) 17e23 Table 1 Number of passengers and aircraft movements. Dependent variable: log(PAX) Coefficient
t-Statistic
Constant Log(MOVEMENT) NA
1.565 1.253 0.586
5.3 49.2 9.7
R-squared (R2) Adj. e R-squared (R2) Observations
0.924 0.924 214
Table 2 Terminal size and number of passengers. Dependent variable: log(TERMINAL) Coefficient
t-Statistic
Constant log(PAX)
2.461 0.863
2.1 12.3
R-squared (R2) Adj. R-squared (R2) Observations
0.605 0.601 101
with a coefficient somewhat less than one, consistent with there being economies of scale in the processes of handling passengers.7 So, measures of passengers using an airport are quite closely related to two other plausible metrics for airport size. A valuable advantage of the passenger numbers variable is its availability e only 214 of the 366 airports in our sample had accessible data on aircraft movements, and for only 101 could we find a measure of terminal building size. The variables used in the main part of the paper are defined as follows: DURATION: Scheduled duration in hours of the flight as reported to flightstats.com, where duration is measured from the scheduled time of pushback from the origin airport gate to the time at which the aircraft brakes are scheduled to be applied at the destination. The flightstats data are in local time, so that time-zone conversions have to be made, for most flights. DIST: Great circle distance between origin and destination airports, in kilometres. PAXO; PAXD: Total number of commercial passengers handled annually (i.e., arrivals plus departures) by origin and destination airport, usually measured for 2009. So, these measures exclude “general aviation” traffic (private planes, charters). PROP: Dummy variable ¼ 1 if aircraft is a turboprop (i.e., not jetpowered). SPEED: Manufacturer’s reported economical cruising speed of the aircraft type; km/h. SEATS: total number of seats (all classes) on the aircraft as usually configured by the airline. LCC: Dummy ¼ 1 if operating airline is a “Low-cost Carrier”. HUBO; HUBD: Dummy ¼ 1 if the origin or destination airport is designated as a “hub” for the operating airline (US only) EAST: Dummy ¼ 1 if the net direction of the flight is easterly. xxORIGIN: Dummy ¼ 1 if the flight originates in region xx. We also culled data from airport websites for all the originating airports, giving estimates of typical travel times and distance
7 Another major input is land. We do not have readily available data on the land holdings of airports, which in any case will be affected by the incidence of “land banking” e holding currently unutilized land to meet expected future expansion needs.
19
between airport and city centre (TIME CITY, DIST CITY). The descriptive characteristics of the data are shown in Table 3. Table 3 tells that the average flight in our database was just over 3 h in scheduled duration, and 2000 km in distance. The average cruise speed of the operating aircraft was 792 km/h, and this aircraft had 133 seats and was in 88% of cases a jet. The operating airline was a low-cost carrier for one flight in five. Table 3 also shows descriptive statistics for two variables measuring the ground journey time from origin airport to the central area of the nearest city (TIMECITY) by the generally quickest mode of transport, and the distance of this trip (DISTCITY). These data were culled from airport and other websites. 3. Results We will first show results for our main variable of interest, scheduled flight duration, then for the other dimensions of the airport size issue. Econometric models are estimated using the EViews 7.2 package. 3.1. Scheduled flight duration First we model total scheduled duration, then break this down to ground and air time. For total flight duration we use a log model to capture non-linearities, shown in Table 4. The least squares model for the log of scheduled flight duration appears to be highly successful. Log(DIST) is hugely significant, with a coefficient well below one, which is likely to be because of on-theground flight time, and, perhaps, time to and from cruising altitude. The size of this coefficient, however, results in the model seriously under-predicting duration of long-haul flights (greater than seven hours). Despite this, the overall goodness of fit is very high for a cross-sectional regression model, with an R2 of 0.97. Both the origin and destination airport size measures (PAXO, PAXD) are significant with positive coefficients, in support of the main hypothesis of this paper. We will return to this. Then we have a number of successful control variables. Bigger aircraft (SEATS) apparently require slightly more generous flight scheduling. Faster aircraft (SPEED) get there more quickly, naturally, and Low-cost carriers (LCC dummy) spend (it seems) less time on the ground. Even controlling for aircraft speed, turboprop aircrafts’ flights (PROP dummy) take longer, which is a surprise. The prevailing winds around the Earth are such that Easterly flights are faster (EAST dummy). Perhaps surprisingly, a (non-stop) flight on an airline which has a major Hub (in the US) takes less time, other things equal.8 Of the regional dummies, only that for New Zealandoriginating flights (NZORIGIN) was significant, with a negative coefficient, perhaps reflecting the more relaxed security procedures for domestic flights within NZ.9 The regression model shown in Table 4 is encouraging in that it shows we have found a set of variables significantly related to scheduled flight duration, including the airport size measures that are our main point of interest.10 However the single-equation specification is awkward, given that determinants of the two
8 Of course one-stop journeys using the hub-and-spoke system are likely to take longer than direct flights. 9 Turboprop flights within NZ have no security checkpoints, and even for jet flights check-in on Air New Zealand does not close until 20 min before scheduled departure time. 10 The regression model of Table 4 was arrived at after a limited amount of experimentation with different sets of regressors, including the other regional dummies, and unsuccessful attempts to identify a time-of-day effect on flight duration, with “peak” period departures and/or arrivals expected to be scheduled to take longer.
20
T. Hazledine, R. Bunker / Journal of Air Transport Management 32 (2013) 17e23
Table 3 Descriptive statistics for airports and flights.
Mean Min Max
DURATION h
DIST km
PAXO number
PAXD number
PROP dummy
SPEED km/h
SEATS number
LCC dummy
TIME CITY min
DIST CITY km
3.1 0.4 15.9
2011 73 13,552
32,765,473 194,864 84,846,639
19,984,785 3354 84,846,639
0.12 0 1
792 440 907
133 8 517
0.20 0 1
27 6 60
19 4 45
Table 4 Modelling total scheduled flight duration. Dependent variable: log(DURATION) Coefficient
t-Statistic
Constant log(DIST) log(PAXO) log(PAXD) SEATS SPEED LCC PROP EAST HUBO HUBD NZORIGIN
4.487 0.686 0.037 0.018 0.000 0.001 0.058 0.103 0.054 0.016 0.044 0.066
51.4 172.3 12.0 9.9 8.4 7.1 9.0 4.4 10.4 1.9 4.3 5.3
R-squared (R2) Adj. R-squared (R2) Observations
0.972 0.972 2009
additive components of flight duration e ground time and flying time e are likely to be quite different. For example, the log specification generates a constant elasticity response of duration to airport size, implying a smaller absolute effect on shorter flights, which may not be true. Accordingly, we now break down total scheduled duration into those two components, and model them separately. We do not have any direct information on the size of the two components in each airline’s flight time scheduling computations, but if we could estimate a simple linear model of duration, using only determinants of in-the-air time (flying time) as regressors, then we could reasonably interpret the intercept of this model as an estimate of average ground time. So, we posit that the major determinant of flying time will be distance to be travelled divided by aircraft speed, and estimate the models shown in Table 5. The first model shows distance divided by aircraft type cruise speed is a hugely successful predictor of duration, with a coefficient very close to one. Slightly “tweaking” this variable based on Table 4 results, allowing for a five percent direction speed bonus for Easterly flights further improves the fit. In both variants the intercept is around 0.63e0.65, which we interpret as an average total ground time of nearly two thirds of an hour, or a bit less than 20 min in each airport, which seems reasonable. Accordingly, we take the predicted duration from the second Table 5 model, subtract 0.64 from this, and then subtract the result from actual scheduled duration to get an estimate of
(scheduled) time on the ground, multiplied by 60 to turn it into minutes, for easier interpretation. This variable is then dependent in a linear model bringing in all the putative ground timedeterminants as regressors. In this model we include quadratic terms for the two airport size variables, to allow for possible nonlinear effects. Table 6 shows three estimated versions of this model: first with the full sample, then with flights longer than six hours removed, and then with fixed effects for the origin airport. Not surprisingly, since we are now trying to explain variation in a calculated variable that is the difference between two variables, overall explanatory power is much lower than in the previous models, and we do not find so many strongly significant regressors. Eliminating long-haul flights, for which predicted ground time has quite a large variance, including some negative numbers, improves the model. The truncated-sample model tells us that LCCs spend about three minutes less on the ground11 than legacy carriers; prop planes are nearly 7 min quicker (perhaps due to quicker taxis to/ from smaller specialised terminals); there is no origin-hub effect, but flights headed for an airline’s hub as destination save about 4 min on the ground; New Zealand and US origin flights are about two and four minutes faster and slower, respectively. As regards airport size impacts, the quadratic effects on groundtime of origin and destination airport size are concave, and positive up to total annual millions of passenger numbers of around 57 and 58, for origin and destination airports respectively.12 In terms of the size of these effects, a doubling of both origin and destination airport sizes from the sample mean of around 20 million passengers/year to 40 million would increase predicted on-the-ground time for each flight by about 6 min, with one third of this at the origin airport and two thirds at the destination. We will discuss in Section 4 the normative significant of these numbers. The third model in Table 6 has fixed effects for the origin airport. Estimation of this model requires dropping any variables which are constant for each airport, so both the regional origin dummies and total origin airport passenger numbers are excluded, with their effects now subsumed in the fixed effects. The jump in the R2 over the second model tells us that (not surprisingly) there are some airport-related explanatory factors which we have not measured specifically in this study, although the coefficients of the included variables are reasonably stable, with the exception of the LCC dummy, of which the coefficient falls away in size and significance, suggesting that the airports used by low-cost carriers tend to have some characteristics that differ from other airports.13 In what follows we will work with the second Table 6 model, which gives us estimates of both airport size effects.
Table 5 Flying time models. Dependent variable DURATION Coefficient
t-Statistic
Coefficient
t-Statistic
Constant DIST/SPEED
0.647 1.018
73.6 411.6
0.634 1.045
81.4 464.1
R-squared Adj. R-squared Observations
0.988 0.988 2010
0.991 0.991 2010
11 Note that this is additional to the well-known “turn-around” time advantage of LCCs, whereby time between flights is reduced by less time spent restocking galleys or handling baggage or passengers in transit, as well as less need to bunch arrival and departure times of flights to fit hub connecting flights. 12 The positive range covers most of the sample. Only three of the 56 origin airports (LAX, ATL, LHR) processed more than 57 million passengers in 2009. 13 In the US and Europe, but not in Australia, Canada or New Zealand, low-cost carriers often use smaller, less congested “secondary” airports, these often being sited quite far from the city that is the ostensible origin or destination of the flight.
T. Hazledine, R. Bunker / Journal of Air Transport Management 32 (2013) 17e23
21
Table 6 Ground-time models. Dependent variable: GROUNDTIME*60 Flights <6 h
Full sample
Origin fixed effects
Coefficient
t-Statistic
Coefficient
t-Statistic
Coefficient
t-Statistic
C LCC PROP HUBO HUBD NZORIGIN USORIGIN PAXO PAXO^2 PAXD PAXD^2
31.106 2.661 6.868 0.487 1.064 5.208 6.449 0.090 0.001 0.316 0.003
26.4 3.1 6.1 0.4 0.8 3.3 7.5 1.7 1.6 6.3 4.4
29.408 2.964 6.962 0.145 4.064 1.999 4.200 0.199 0.002 0.426 0.004
32.5 4.6 8.2 0.2 3.7 1.6 6.2 4.8 4.0 10.7 6.6
34.731 0.767 6.487 0.146 2.601
72.4 1.2 8.5 0.2 2.7
0.380 0.003
10.5 6.0
R-squared Adj. R-squared Observations
0.137 0.132 2009
3.2. Is lateness ‘scheduled’? We have found that airlines tend to build into their schedules extra time on the ground at larger airports, but whether this is enough time needs to be considered. If not, then we could expect to find some relationship between late departures and airport size. For official statistical and reporting purposes, “late” is defined as a flight pushing off from the gate 15 or more minutes after the scheduled departure time.14 Of the 1934 flights in our sample for which actual departure times were noted on flightstats.com, just over 14% were officially late. A Logit model predicting the probability of late departure (DLATE ¼ 1) found no significance in the overall airport size variable (PAXO). If this is replaced by dummy variables for airport sizes in ranges of 10 million passengers/year, then only the dummy for the largest airports (greater than 50 million passengers) showed any significance. This model is shown in Table 7. That is, there is apparently some tendency for airlines to be optimistic about delays at the very large airports (i.e., to not fully allow for the likelihood of delays in their scheduling), but not otherwise. Our calculated scheduled ground-time variable has no effect on actual lateness, consistent with airlines being fully aware of the determinants of time on the ground at each airport (excepting perhaps the very largest). The probability of late departure is higher in the late afternoon/early evening peak period, presumably because of the cumulative effect of delays earlier in the day.15 3.3. Passenger ground time Finally, we move from what happens after the cabin door is closed on an aircraft to what happens before that event, from the passengers’ perspective. Specifically, we look at time getting to (or from) the airport, and time spent in the airport terminal. Our empirical strategy here in both cases is to use readily available proxies rather than attempt to construct comprehensive data. So, to estimate passengers’ ground travel time accurately would require either surveys of actual travellers, or elaborate
14 In the US, the Bureau of Transportation Statistics (BTS) defines “departure” as “the instance when the pilot releases the aircraft parking brake after passengers have loaded and aircraft doors have been closed”. Arrival time is defined analogously (www.transtats.bts.gov/glossary.asp). 15 PMPEAK takes the value 1 for flights scheduled to depart after 3pm and before 8pm.
0.239 0.235 1787
0.450 0.430 1787
manipulations of (census) data on the distribution of population in the airport’s catchment and distances (and mode of travel) to the airport from each location. This is a major data gathering exercise even for a single airport.16 However, there is a plausible single number proxy for airport accessibility, which is simply the distance of the airport from the Central Business District of the major city it serves. This figure can invariably be quickly culled from airport or other websites, and this fact alone suggests that airports know that distance from downtown is a useful piece of information for travellers. Our hypothesis is that the economics of land prices will tend to push larger airports to locate further away from the CBD, which we test for our group of 57 origin airports, with results in Table 8. There is indeed a moderately robust relationship, with an elasticity of distance with respect to airport size of 0.165. If we use city population instead of airport size as the regressor, we get an almost identical result, which of course reflects the tendency of larger cities to build larger airports. Actual time spent getting to the airport will depend, as well as on distance, on the transport infrastructure e whether there is a large capacity freeway, and/or whether there is direct train service between the CBD and the airport. We can get rather approximate estimates from airport websites of the likely minimum journey time from city centres, and we model this in Table 9. Here, size of airport is still significant, and there is some evidence that trains are faster than buses or taxis. However, these two variables are not independent: if we do a simple Probit model of the probability of a direct airport-CBD rail link with airport size as the regressor, there is indeed a significant positive relationship. This is presumably because the high fixed costs of rail service tend to need a large airport with many passengers to justify them. So, the Table 9 regression model implies that a doubling of airport size goes with a 12% longer travel time to the airport e perhaps two or three minutes onto the average travel time, which is 23 minutes e unless the larger airport now has train service, in which case the net effect is a small decrease in the predicted ground travel time. Secondly, we explore passenger time spent in the airport. Personal experience has shown that the larger terminals of larger airports result in longer walks to and from gates, but this has to be documented systematically. As with travel to/from airports, generating accurate data would require a substantial amount of work for each case, perhaps entailing customer surveys and/or time
16 Ishii et al. (2009) make good use of a survey of passengers travelling in the San Francisco/Oakland Bay area.
22
T. Hazledine, R. Bunker / Journal of Air Transport Management 32 (2013) 17e23
Table 7 Modelling late departures.
Table 9 Time to airport from city centre. Dependent variable: DLATE
Constant PAXTD50 HUBO GROUNDTIME NZORIGIN PMPEAK McFadden R-squared Observations Procedure:
Dependent variable: log(TIMECITY)
Coefficient
z-Statistic
1.853 0.481 0.308 0.021 0.520 0.375
10.0 2.7 1.5 0.1 1.6 2.4
Coefficient
t-Statistic
Constant LOG(PAXTOTO) TRAIN
1.083 0.126 0.217
1.4 2.7 1.6
R-squared Adj. R-squared Observations
0.118 0.085 57
0.011 1932.0 Binary logit
Table 8 Distance to airport from city centre. Dependent variable: log(DISTCITY) Coefficient
t-Statistic
C LOG(PAXTOTO)
0.094 0.166
0.1 3.3
R-squared Adj. R-squared Observations
0.164 0.149 57
and motion studies. Instead we suggest that airlines’ minimum check-in times should be a proxy, since these will be set in the knowledge of how long it will likely take a passenger to get from the check-in counter to the departure gate. In our sample of 57 origin airports, 30 specify a 30 min minimum or recommended check-in time (i.e. check-in at least 30 min before scheduled departure time); 13 asked for 40 min, and the remaining 14 specified either 45 or 60 min. The average size of these three groups, in terms of annual passengers handled, was 10.8 million, 22.2 million and 39.7 million, respectively. That is, while we have not done any significance testing on these numbers, they do support the reasonable hypothesis that there is a tendency for larger airports to require passengers to spend more time in the terminal.17 4. Summary and implications We have found solid econometric evidence of diseconomies of scale with respect to air travel time, generated by larger airports. For example, the journey of a traveller taking off from and landing at airports handling 40 million passengers/year would tend to be scheduled by the airline at about 6 min longer total duration than a journey between two 20 million passengers/year airports, other things equal. Less formally established are our findings that a traveller’s trip to the airport from downtown would likely be about two minutes longer in the case of the larger airport, and this passenger would have quite likely been required to check-in for their flight at least five minutes earlier. These figures are statistically significant, but there is the question of whether there are they also economically significant. That is, are the time effects large enough to materially affect traveller wellbeing? We will not attempt a formal cost-benefit analysis, but will give some idea of the scale of the scale effects relative to two benchmarks: the operating revenues of airports, and the calculated welfare savings from implementing optimal congestion pricing at
17 These are minimum recommended check in times. Many airports (and airlines) suggest allowing more time for checking in to international flights. Note that, in our sample, the largest airports (more than 40 million passengers/year) have, on average, the same proportions of domestic flights (55%) as the full sample.
US airports, from Daniel and Harback (2009). These authors calculate welfare savings for the 27 largest US airports (using 2003 data) at about $4,500,000/day, or around $60 million/year per airport, on average. We have some data for these airports for 2004, and that year the average large airport handled 35 million passengers, and had total18 operating revenue of $289 million. Continuing with our comparison between a forty million passenger airport and two twenty million passenger airports hypothetically carrying the same travellers, our figures imply total additional annual passenger travel time in the former case of ten minutes per passenger,19 or more than six million hours e worth, say, $180 million, at an hourly rate of $30. On top of this should be added additional costs to the airlines of longer taxi times, etc, which Morrison and Winston’s (2008) estimates of costs of flight delays (late departures) put at about 80% of the on-board passenger time costs. So, relative either to the economic size of airports measured by total revenues, or to the estimated costs of congestion-induced delays, our scale effects are quite large and undoubtedly “significant” in their size. However, we should note unresolved difficulties in valuing traveller time in these contexts. Many may be likely to agree on the irksomeness of ground travel to and from the airport20, and of trudging through the terminal. However it is harder to assign a value to on-board time. For some travellers, each minute trapped inside an aircraft may be another minute of claustrophobic terror; for others it may be a relaxing break, well away from outside cares and responsibilities; and, for some, flying may remain an interesting, fun experience. We note, too, limitations on the generality of our econometric analysis, based as it is on a sample of just 57 of the world’s origin airports (and 375 flight destinations), and around 2000 actual flights. However, given the statistical and economic significance of our quantitative results, we do consider it reasonable to recommend that any cost benefit analysis of a major urban airport construction or extension project with spatial implications be required to include consideration of likely time diseconomies of scale along with the other relevant factors.
References Chang, Y.-C., Yu, M.-M., Chen, P.-C., 2013. Evaluating the performance of Chinese airports. Journal of Air Transport Management 31, 19e21. Daniel, J., Harback, K., 2009. Pricing the major US hub airports. Journal of Urban Economics 66, 33e56.
18 Operating revenue is generated about 60% from aeronautical services (e.g., landing fees) and 40% from terminal income (retail concessions, parking), on average. 19 10 ¼ 6/2 þ 2 þ 5. 20 Ishii et al. find evidence from trips taken by San Francisco Bay Area residents that time costs associated with airports are significant determinants of airport choice; in particular, ‘distaste for access time’ (2009, p225).
T. Hazledine, R. Bunker / Journal of Air Transport Management 32 (2013) 17e23 Gillen, D., Lall, A., 1997. Developing measures of airport productivity and performance: an application of data envelopment analysis. Transportation Research Part E 33, 249e260. Ishii, J., Jun, S., Van Dender, K., 2009. Air travel choices in multi-airport markets. Journal of Urban Economics 65, 216e227. Martin, J., Roman, C., Voltes-Dorta, A., 2011. Scale economies and marginal costs in Spanish airports. Transportation Research Part E 47 (2), 238e248. Mayer, C., Sinai, T., 2003a. Network effects, congestion externalities, and air traffic delays: or why not all delays are evil. American Economic Review 93 (4), 1194e1215.
23
Mayer, C., Sinai, T., 2003b. Why Do Airlines Systematically Schedule Their Flights to Arrive Late?. Working Paper: The Wharton School, University of Pennsylvania. Morrison, S., Winston, C., 2008. The effect of FAA expenditures on air travel delays. Journal of Urban Economics 63, 669e678. Rupp, N., 2009. Do carriers internalize congestion costs? Empirical evidence on the internalization question. Journal of Urban Economics 65 (1), 24e37. Santos, G., Robin, M., 2010. Determinants of delays at European airports. Transportation Research Part B 44, 392e403.