A HEDONIC PRICE INDEX FOR AIRLINE TRAVEL

A HEDONIC PRICE INDEX FOR AIRLINE TRAVEL

by David H. Good

Indiana University

Robin C. Sickles*Rice University, Texas

and

Jesse C. Weiher

Office of Federal Housing Enterprise Oversight, Washington DC

Our research takes an exhaustive approach to measurement issues in price index construction for theBLS airfare index. We pursue a number of the objectives for dealing with the biases that the 1997 CPICommission recommended and detail a protocol for data collection and analysis that can be replicatedand can be enhanced by availability of additional data sources. We find an upward bias in the BLSairfare index over the period considered. However, because of issues of practicality and implement-ability of the methods we utilize in our analysis, the goals of the Commission recommendations remainillusive and problematic in being more broadly applied to other components of the CPI.

1. Introduction

Ideally, the Consumer Price Index (CPI) measures the price of a fixed marketbasket over time. The CPI has a wide variety of uses as both a measure of theoverall inflation in the economy and as a cost of living index (COLI). The CPICommission outlined several features of the CPI which tend to make it less usefulas a COLI. The issues raised by the CPI Commission are discussed at length inBoskin et al. (1997, 1998) and Boskin and Jorgenson (1997). First, because themarket basket is fixed, the CPI does not allow for consumers to respond to pricechanges by substituting away from commodities with higher price increasestoward commodities with lower price increases (substitution bias). Second,because price changes are considered only within outlets, the index does not allowfor consumers to substitute away from higher priced outlets toward lower pricedoutlets for identical items (outlet substitution bias). Third, since the market basketis typically sparse in details about quality, the CPI does not always consider the

Note: This research was funded by a grant from the Bureau of Labor Statistics. Preliminary draftswere prepared for the BLS conference “Issues in Measuring Price Changes and Consumption,” June2000 and the Brookings Institution’s “Workshop on Transportation Output and Productivity.” Wewould like to thank without implicating Dennis Fixler, John Greenlees, Robert Gordon, Clint Oster,Jack Triplett and Cliff Winston for providing suggestions and criticisms that substantially improved thepaper. We thank two anonymous referees for their comments on the paper. The usual caveat applies.

*Correspondence to: Robin C. Sickles, Rice University, 6100 South Main Street, Houston, TX77005-1892, USA ([email protected]).

Review of Income and WealthSeries 54, Number 3, September 2008

© 2008 The AuthorsJournal compilation © 2008 International Association for Research in Income and Wealth Publishedby Blackwell Publishing, 9600 Garsington Road, Oxford OX4 2DQ, UK and 350 Main St, Malden,MA, 02148, USA.

438

mailto:[email protected]

possibility that price increases may be caused by unmeasured quality improve-ments in the items that constitute the fixed market basket that increase consumerwelfare (quality change bias).

The CPI Commission concluded that changes in the CPI overestimated thechange in the Cost of Living by about 1.1 percent per year. They estimated theportions of bias that could be attributed to substitution bias and outlet substi-tution at 0.4 and 0.1 percent per year. They also estimated the portion that couldbe attributed to failing to control for the quality of existing commodities and toincorporate new commodities at 0.6 percent per year. The commission concludedthat impacts of this bias were far reaching since many government expenditureincreases, most notably Social Security, are tied into the CPI. Consequently, thebias has led the federal government to overcompensate individuals for the cost ofliving, leading some to conclude this overcompensation has accelerated inflationas more dollars in benefits chase a fixed level of goods. The commission gener-ated a number of recommendations. Among them were: (i) use hedonic statisti-cal methods to adjust for quality change; (ii) reweight the consumptionbasket more frequently; (iii) increase the pace of sampling so that new goodscan be accounted for more rapidly; (iv) study the individual componentsof the CPI to determine which components provide the most information aboutfuture movements in the index and which components have movements whichare mostly irrelevant to movements in the total; (v) move toward the notion ofa new “basket” every year; and (vi) use new sources of data such as scannerdata.

This paper describes the implications of implementing some of these recom-mendations in the context of one specific component of the CPI, domestic airlinefares. To our knowledge, ours is the first paper to integrate so many of theserecommendations simultaneously: re-weighting every period, hedonic modelsand the use of scanner-like data. We view the consumer’s market basket ofairline travel commodities as composed of a number of alternative city-pairroutes. The airline fares portion of the CPI is amenable to the CPI Commission’srecommendations because data sets of ticket sales already exist. In otherindustries, improved data collection may prove difficult because national leveldata collection requires cooperation that these firms are unlikely to willinglyprovide.

Based on the U.S. Department of Transportation’s (USDOT) origin anddestination data (O&D) we estimate a quality-adjusted price index for airline faresusing a hedonic price regression which includes quality characteristics as well astime dummies. From these estimates we are able to construct quality adjustedchanges in price over time. Basing our adjustments on the O&D database is alsouseful in dealing with market basket composition issues. Because the weightsconsists of sales data, they can be adjusted as the quantities sold adjust. When anew product is introduced (e.g. an airline adding a new route, the introduction ofa business class fare, etc.) it shows up automatically in the data. Interregionalcomparisons can be made because the data collection techniques are the sameacross regions. Still there are some limitations: airline data are collected quarterlyrather than monthly (making construction of a monthly airfare index difficult),smaller airlines tend to underreport and some information about fare class is

Review of Income and Wealth, Series 54, Number 3, September 2008

© 2008 The AuthorsJournal compilation © International Association for Research in Income and Wealth 2008

439

censored. Even so, the aviation O&D data have a much broader geographicalcoverage and longer time series than any other single source publicly availablesales data series in any other industry of which we are aware.

The paper is organized as follows. In the next section we discuss developmentsin the literature on price indices and on specific problems with the construction ofan air travel price index. In Section 3, we discuss the methods employed by theBureau of Labor Statistics (BLS) to calculate the Consumer Price Index forAirfares. In Section 4, we present the construction of measurements for a numberof fare characteristics and present their trends during the period of our study,1979I–1992IV. In Section 5, we construct a new method that explicitly recognizesthe role of quality characteristics on the consumer’s valuation of airline service.This reduced form approach utilizes hedonic price index methods to weight downthe actual reported output and thus weight up the effective price of airline service.Our final section presents results and provides concluding remarks.

2. Sources of Bias in the CPI: Suggested Corrections and theirImplications for the Airline Price Index

As we pointed out in the introduction, the CPI Commission concluded thatchanges in the Consumer Price Index overestimated the change in the Cost ofLiving by about 1.1 percent. The bias was attributed to three factors: the substi-tution bias, outlet substitution bias, and quality change bias (or new goods bias).Boskin et al. (1998) discuss the Commission findings and describe the sources ofthese biases in detail. They also discuss another type of bias that few have deemedimportant. This is a “when” bias (referred to in this paper as “intertemporalsubstitution bias”). The “when” bias occurs because price data tend to be collectedduring the week, whereas an increasing share of purchases are made on weekendsand holidays. Many outlets even emphasize weekend sales.

The Commission’s conclusions generated a large quantity of furtherresearch—both supporting and refuting the Commission’s findings. Abrahamet al. (1998) provide the BLS response to the CPI Commission’s report anddescribe initiatives at the BLS that are intended to improve the accuracy of the CPI(some of which were enacted before the Commission’s report). Pollak (1998)claimed that changing the CPI so that it could more accurately reflect a cost-of-living index was not possible without addressing very fundamental issues whichstill appear to be unresolved. Whose cost-of-living do we want to model? How canwe deal with an economy in which the law of one price does not hold andconsumers must search for the best price? What exactly is a good? Baker (1998)summarizes criticism of the CPI Commission’s report. Among several issues, heasserts that several estimates of bias depend exclusively upon introspection andseveral other estimates are the result of either misinterpreting previous research ormaking dubious extrapolations from that research.

Possible bias in government inflation statistics is not just a concern in theUnited States and when the findings of the Boskin Commission were made public,the international community responded. In 1997 Statistics Canada published ajoint study examining CPI bias in four other OECD countries (Australia, Canada,



440

France, U.K.) (Ducharme, 1997).1 Referring to the international impacts of thereport it states, “. . . in a number of countries it raised the awareness of the need toreview formulae, improve underlying household surveys (make them more fre-quent and more comprehensive) and adjust for quality variation all goods andservices, especially complex goods and services typically offered in the marketplace as bundles.” It goes on to state, “The five national statistical agencies agreethat quality change is the most difficult and important problem in estimating theCPI . . . [A]ll feel that the only way to progress is by developing a research agendaand, with the help of other experts, to systematically analyze the effects of qualitychange on estimated prices.”

The BLS has implemented several changes in the construction of the CPI inorder to improve the accuracy of the index. Some of the improvements came as aresult of the CPI Commission’s report and some were already being contemplated.To address lower level substitution bias the BLS moved toward using a geometricmean instead of an arithmetic mean. To address upper level substitution bias theBLS began issuing alternative superlative indexes for comparison. Other changeswere also made to ameliorate outlet substitution bias. To address quality changebias the BLS commissioned several studies to investigate how hedonic price func-tions might be used to control for changes in quality of items in the market basket.We discuss below how such hedonic approaches have been applied to other indus-tries and why the airline travel index in particular could benefit from controllingfor quality characteristics hedonically.

2.1. A Brief Review of Hedonic Literature

The automobile industry was one of the first industries to utilize hedonicmethods (Griliches, 1961). The analysis of marginal prices of housing attributesthrough computation of hedonic indices began in the late 1960s (e.g. Ridker andHenning, 1967) and 1970s (Rosen, 1974). Hedonic indices were applied to com-puters in the mid 1980s (e.g. Triplett, 1984) and more recently information tech-nology products in general (Triplett, 2004). Other markets whose prices wererevalued using hedonic techniques include the medical field (Primont andKokoski, 1990; Trajtenberg, 1990) and university education (Schwartz and Scafidi,2000). To our knowledge, our development of a price index for airline fares via ahedonic price approach is new to the literature. For a clear and concise review ofhedonic regression theory we suggest the reader review Quigley (1982). Armknechtand Ginsburg (1992) offer hedonic regressions as a way to address the issue ofquality that arises when new products enter the market. Hedonic models not onlyhelp in estimating quality differences resulting from a change in characteristics.They also enhance the analyst’s information about the quality composition ofservices offered. Specifically, the models can identify those quality characteristicsthat provide the largest impacts on price.

Armknecht and Ginsburg (1992) suggest two important benefits related tothese insights. First, analysts can create a formal statistical test for whetherchanges in a quality characteristic render the old and new services incomparable.

1The study also included an updated discussion of efforts in the United States to address the issuesraised by the Boskin Report.



441

Second, the factors that provide the most impact on price can be used to redesigncollection documents used by the CPI. By ordering the quality characteristicsaccording to importance, for example, field staff can pick the most appropriatesubstitute by the most important characteristics according to the appropriateorder. Pakes (2003) furthers this line of thinking by directly comparing matchedmodel indexes to hedonic indexes in the market for PCs.

Pakes (2008) cautions the researcher that imperfect competition prevents thehedonic coefficients from being accurately interpreted as the consumer’s marginalwillingness-to-pay. He makes this point with a hypothetical example involvingmedications that illustrates the problem of interpreting the derivatives of a hedonicprice function as either willingness to pay derivatives or cost derivatives, sinceempirical estimates are essentially reduced form solutions formed from a complexequilibrium process. Pakes makes it clear, however, that losing the interpretationof hedonic coefficients derivatives of a hedonic price function as either willingnessto pay derivatives or cost derivatives does not invalidate the use of hedonics tocontrol for quality in a price index. Pakes’ concern about the interpretation ofhedonic coefficients involves situations where the commodity is defined toobroadly. An example is one in which the consumers buying drug A are not thesame consumers buying drug B. If one were to run a separate regression for eachdrug, and if the toxicity of drug B were to reduce over time, we would expect to seea negative parameter estimate for toxicity.

Using this line of reasoning the large number of customers shopping amongthe various characteristics of an airline ticket would validate the interpretability ofhedonic coefficients. Parameter estimates in our analysis below thus retain theirinterpretation as measures of the marginal willingness to pay because airline travelis not so broad a service as to apply to completely different consumer groups.Airlines do have different marketing strategies for different groups of consumers(business versus leisure travelers, for example). However, these groups of consum-ers are not exclusive sets of consumers. An airline passenger who is traveling onbusiness now will be traveling for leisure later. Someone who buys a first classticket today may well buy a discount coach ticket tomorrow. It seems reasonableto assume, for example, that a first class ticket was bought when the premium forfirst class was sufficiently small and that a coach ticket was bought when thepremium was too large. In Pakes’ drug example, it is hard to conceive of aconsumer buying the more toxic drug if the price were lower. However, we cannotthink of any quality attribute associated with an airline ticket that consumerscould not be induced to change through a sufficient change in price.

2.2. CPI Bias and Quality Issues in the U.S. Airline Industry

In this section we discuss how CPI bias might appear in the index for airlinefares as well as discuss research into quality issues in the U.S. airline industry. Itshould be noted that the BLS airline fares index is not the only price index forairline fares. The Bureau of Economic Analysis (BEA) produces the consumerexpenditure deflator for airline fares as part of the National Accounts Estimationand the Bureau of Transportation Statistics is now publishing the Air Travel PriceIndex (ATPI) based on a Fisher index. Lent and Dorfman (2005) conduct a



442

detailed comparison of the ATPI to the CPI for airfares and the BEA’s expendi-ture deflator. The ATPI is not available prior to 1995 while the quality-adjustedindex constructed here ends in 1992. Consequently there can be no direct com-parison between the two indices. It should be noted that the Fisher index does notnecessarily represent a lower bound for a COL index and, as such, may be subjectto less substitution bias. Most quality adjustments made by the CPI use an impu-tation procedure that estimates the price change of a new product (or old productwith new quality feature) by the average price changes of similar products cur-rently in the market. This assumes that the underlying price change of the servicenot currently available would have been the same as the price changes of servicescurrently in the market. However, if this is not the case, then the imputationprocedure would miss important price changes. Armknecht and Ginsburg (1992)illustrate this problem in the context of airline fares:

. . . As an example of this situation consider airfares. The airlines at one pointin time introduced a new set of discount airfares to replace supersaver fares.Originally, supersaver fares required a 30-day prepayment to obtain thereduced fare. The new discount fares introduced a lower price structure thanthe supersavers and required the 30-day prepayment. However, along with thelower fares came a 50 percent cancellation penalty, a quality differencebetween the fares that may/should make them noncomparable. If no otherairline price changes occurred during this month, the exclusion of the discountquotes would result in an imputed price change of zero because all otherairline fares (coach and first-class) remained the same. The index, therefore,misses any possible price change occurring with the introduction of the lowerdiscount fares under current procedures (this problem would exist even ifother fares changed but at a different rate).

Gordon and Griliches (1997) point out that some quality improvements arequite problematic when attempting to incorporate them into a price index becausethey represent an abatement of “bads” that also were not considered in the index.For example, increases in the price of a flight due to increased safety measures maybe a price change due to quality improvement. However, unless we also deal withthe cost that was imposed due to increased risk of plane malfunction or terrorism,we will not be accurately reflecting the COL associated with it. It may be easilyargued that patrons were better off before terrorism was ever a threat than aftersecurity measures were put in place. By not specifically accounting for the costs toincreased risks, we will be overestimating the effect of security on welfare andunderestimating the cost of living. The security issue is not explicitly accounted forin our analysis, but this is an interesting point. One possible modeling approach touse in this regard that we have not pursued is the directional distance functionparadigm introduced by Färe et al. (1994) and recently utilized in the context ofthe impact of environmental “bads” on growth accounting by Jeon and Sickles(2004).

Although they did not construct a price index, Morrison and Winston (1995)calculated marginal valuations of quality characteristics via an airline choicemodel. The marginal value of an additional mile for travelers who accumulated3,501 to 15,000 miles was estimated to be 13 cents while for travelers who accu-



443

mulated 15,001 to 80,000 miles it was 21.5 cents. They estimated the cost ofrestrictions on a discount flight via a binary, probit model of choice betweenunrestricted vs. restricted fairs. Although there appeared to be no significant costsfor the leisure travel the business traveler incurred some significant costs. Saturdaynight restrictions added $219 dollars of disutility. Advanced reservation require-ments imposed a cost of $3.68 a day. Overall, business travelers (via a compensatedvariation method) were willing to pay $87 dollars per trip to avoid restrictions, asubstantial portion of which was due to the Saturday night stay ($67). They alsoestimate a fare equation for each carrier to investigate whether multimarketcontact has an affect on prices. They found that multimarket contact did have aneffect on fares in the industry but that the effects were highly cyclical. When theeconomy is growing rapidly fares were higher where there was multimarket contactthan where there was none. However, there are recessionary periods where theopposite effect occurs. Multimarket contact might be considered a benefit toconsumers because it allows for the possibility of greater competition (e.g. Perloffet al., 2003). Even in the presence of collusion on prices, airlines might stillcompete along quality dimensions. An example is Continental competing withUnited in the Dallas–Fort Worth/Washington DC market, in which there is anec-dotal evidence of price collusion, by allowing for more carry-on luggage space.United tried to implement a carry-on size limit. Continental sued United in orderto stop the carry-on limits.2

A number of studies have found that the consumer has benefited from air-line deregulation. Caves et al. (1987) showed that deregulation increased thepassenger-mile productivity growth rate between 1.3 and 1.6 percent per year.Morrison and Winston (1995) highlighted a number of important benefits accruingto consumers: (1) fares were about 22 percent lower with deregulation; (2) a benefitof about $10.3 billion per year resulted from the elimination of route restrictionsand an increased use of hub-and-spoke networks; (3) the proportion of passengerswith direct nonstop flights increased, even though passengers often had to take anindirect route associated with the hub-and-spoke network structure; and (4) pas-sengers who had to change planes rarely had to change airlines due to increases innetwork size.

It should be noted, however, that the benefits of such innovations as thehub-and-spoke systems may be overstated due to longer travel time, longer flightsegments, and longer waiting times for the flights themselves because of schedulingand because of congestion. Ground time has increased by five minutes regardlessof distance traveled. It would seem, however, that the cost of increase travel time($2.8 billion per year in 1995) was more than offset by the benefit of hub-and-spokesystems (Morrison and Winston, 1995).

2Continental Airlines, Inc. v. United Airlines, Inc., 126 F. Supp.2d 962 (E.D. Va. 2001). Conti-nental Airlines challenged a rule at Dulles Airport imposing limits on the size of carry-on baggage. TheAssociation of Air Carriers serving Dulles had imposed the rule at the insistence of United Airlines. Thecourt struck down the rule, finding that it was not justified by safety and efficiency concerns, and thatUnited attempted to influence the standard-setting process in an anticompetitive manner in order torelieve it from competitive pressure from air carriers which permitted larger carry-ons. However, thatdecision has been vacated, citing “unique structure of Dulles,” and remanded for further proceedings.See Continental Airlines, Inc. v. United Airlines, Inc., 277 F.3d 499 (4th Cir. 2002). The parties havesubsequently agreed to dismiss the case.



444

3. Simulating the Current BLS Approach for Constructing anAirfares CPI

The airline fares index is the largest component (61 percent) of the publictransportation index. It accounts for 0.814 percent of the total CPI. The BLSselects approximately 850 observations for its airline fares index. All regularlyscheduled commercial airline trips departing from airports in the 87 cities in theCPI are eligible for use in the index. When the city does not have a qualifyingairport, the qualifying airport from the nearest city is designated as the point ofdeparture. Reflecting the average pattern of trips from the USDOT’s origin/destination survey, 2 percent are first class, 4 percent are full coach fare, and 94percent are discounted coach fares. The origin/destination survey does not identifytypes of discount fares. Thus the BLS assumes that half of the discounted coachfares are the lowest available fare regardless of restrictions. The BLS assumes thatthe other half includes specific restrictions. The fare at the time of sampling isfound in the SABRE reservation system. This is the same system that historicallyhas been used by much of the travel industry, though its use has declined substan-tially over the last two decades. When an airline discontinues a discount fare thatis being priced, the BLS substitutes the closest available alternative. However,applicable restrictions may change, thus changing the “quality” of the flight. BLSestimates the value of this change. If the value is high, the two fares will not becompared. Usually, however, the change is not large and the fares are comparedand reflected in the CPI.

Below we more formally describe the process used by the BLS during ourstudy period (1979 through 1992) and the new procedure that they used begin-ning in 1999. Since we are interested in evaluating the current procedure of BLSbut must evaluate it using older data we construct price indices using bothmethods. The first method is used for comparison to historical indices. Thesecond is used to identify the historical index had current procedures beenused. Below we describe the historical procedure and then how it changed in1999.

Step 1 (primary sampling): Conduct random price sampling for all items inthe market basket. The market basket is a set of origin–destination–fare classtriples in each Price Sampling Unit (PSU). The PSU is a county within one of themetropolitan areas which is covered by the Consumer Price Index for all urbanconsumers (CPI-U).

Step 2 (aggregation process pre-1999): Compute the Air Fare CPI

AFCPIAFCPI

s wp

pt

tj ij

ijt

ijtij− −

=⎛⎝⎜

⎞⎠⎟∑∑

1 1

where the subscript i represents the item in the basket, j represents the PSU and trepresents the time period. Weights wij are obtained from the Consumer Expendi-ture Survey and weights sj represent the fraction of sales which occur in PSUj

compared to the entire U.S.Step 2� (aggregation process post-1999): Compute the Air Fare CPI



445

AFCPIAFCPI

s wp

pt

tj ij

ijt

ijtij− −

=⎛⎝⎜

⎞⎠⎟∏∑

1 1

.

The difference between step 2 and step 2′ is the use of arithmetic means priorto 1999 and geometric means after 1999 at the stage where price ratios are aggre-gated to the level of the PSU.

Implementation of these procedures is complicated by inadequate informa-tion on the appropriate market basket and the proper weights. Unfortunately, theonly market basket and weights available from BLS are those currently used andthey were constructed in 1998. Market baskets and weights for earlier years areunavailable because the BLS did not archive this data. Information on deeplydiscounted fares with corresponding restrictions thus was not available in many ofthe years that our study covers.3

4. Data Sources Used to Compute a New Quality Adjusted AirfarePrice Index

The CPI is best computed when the products under consideration are simplehomogeneous commodities. This does not describe the air transport industry,particularly over our study period, because air travel takes place between manydifferent origins and destinations over alternative paths and with differing levelsof service. In this section we discuss several of the important dimensions of airtravel and how they affect the price of travel through changes in quality. Theseissues fall into seven broad quality categories: network configuration, flight con-venience, passenger amenity, airport and flight delay, ticket restrictions, yieldmanagement, and frequent flier programs (zero coupon tickets). These categorieswill at times overlap. In many cases our choice of study period, from 1979through 1992 is both interesting and restrictive. It is interesting because manyfare innovations and service quality changes occurred during this time periodfollowing the deregulation of the industry in 1978. It is restrictive because manyof the data sets which would be useful to identify service quality changes do notspan our study period.

4.1. Network Configuration

Prior to deregulation, airline networks were constructed on a piecemealmanner. Airlines were granted routes out of “public convenience and necessity.”This often led to rather inconvenient routing for passengers, particularly to smallcommunities. Deregulation opened access to all routes and by 1982 airlines couldserve whatever routes they wished. Airlines responded by developing hub-and-

3One possible way to deal with the lack of weights would be to use the market basket from 1998while acknowledging that the basket and weights changed over time. An alternative approach would beto use fare mixes each quarter from the Origin and Destination Survey. The first approach uses anincorrect market basket but keeps a constant fare-class mix. The second approach uses fare-class mixesthat are closer to what would have been in older market baskets but loses some of the fixed-basketqualities of the CPI. We settled on eliminating fare class information all together in constructingweights. This produced a series that was quite similar to the CPI and appeared to be a better approxi-mation to the use of an obviously incorrect fare mix.



446

spoke networks, adding some destinations and dropping others. In our study weconsider three hedonic characteristics of the airlines. network configuration: theindirectness of routing, change of planes, and the number of airlines used tocomplete the trip.

Airlines find different network configurations useful because they allow forhigher passenger densities on individual routes, potentially allowing the airlines totake advantage of economies of equipment size (larger aircraft tend to have lowercosts per passenger mile) and higher load factors (filling otherwise empty seats onan aircraft costs the airline very little). In a simple network involving five cities (oneserving as the hub), the cities can be connected with at most one change of planeservice with as few as four round trip flights. Connecting the cities together with anetwork providing nonstop service would require 10 round trip flights. Adding anadditional city would require one more round trip flight with a hub-and-spokenetwork compared to five more round trip flights using direct flights only.

Because passenger time is valuable, indirect routing is a cost to the passenger.The cost increases when the indirect routing involves changes of planes through acongested hub. On the other hand, indirectly routed passengers often will accruemore frequent flyer miles than a directly routed passenger.

When an airline does not serve both the origin and destination or serves themonly indirectly, passengers must take part of their trip on one airline and theremainder on at least one other airline (interlining). Passengers perceive a lowerquality of service when interlined. Baggage will more likely be mishandled ormisdirected, the distance between gates at the connecting airport is usually greater,and there is the increased likelihood of a missed connection.

We measure the network characteristics at the individual ticket level. Usingthe DB1A, we identify the origin and the ultimate destination as indicated by a tripbreak. Approximately 30 percent of trips involve one way tickets (no trip breaks),65 percent are one way (one trip break at the destination), and the remaining 5percent involve open jaw or multibreak tickets. In order to gain an understandingof the bulk of trips we limit our attention to either one way or round trip ticketsand develop separate models for each. The changing pattern of these tickets isdescribed in Figure 1. For each trip type, we identify the complexity of the tripfrom the DB1A at the ticket level. The travel itinerary allows us to measure thenumber of segments (changes of planes with the same carrier) required in theitinerary, the number of airlines that were used to provide service, and the numberof changes required to a different airline (interlines). For each of these variables thehigher the variable the lower the quality of service. Figure 2 displays the patternsof one way trips over the sample period while the patterns of round trip itinerariesare shown in Figure 3. For one way tickets 28 percent of the tickets used more thanone airline on a trip in 1979I. By 1992IV only 4 percent of the one way itinerariesrequired that passengers interline. There is little change in number of segmentsover the study period. The pattern for round trip tickets is similar. In 1979I nearlyhalf of the round trip itineraries involved more than one airline. Some of theseinvolved more than one interline as an itinerary started with one carrier, switchedto a second, then went back to the first carrier on the return. There is a slightincrease in the number of segments, indicating that while passengers were notdecreasing the number of segments over the study period they were switching to



447

79

0.0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

10

80 81 82 83 84 85 86

Date

87 88 89 90 91 92 93

Frac

tion

One

Way

Figure 1. Fraction of One Way Tickets Sold

79

0.00

0.25

0.50

0.75

100

125

150

175

2.00

80 81 82 83 84 85 86

Date

87 88 89 90 91 92 93

Ave

rage

Num

ber

SegmentsAirlinesInterlines

Figure 2. Patterns of One Way Itineraries



448

preferred itineraries on the same airline. This is what would be expected underhub-and-spoke systems.

In summary, the itinerary data suggest that there has been some improvementin the quality of service over the study period. If ignored in the computation ofprice indices these service quality improvements would artificially inflate the rateof growth in prices.

4.2. Flight Convenience and Availability

Passengers typically have clear preferences regarding the time of travel andlocation of departure. They may choose the expected time of departure or the timeof arrival at the final or intermediate destination. The willingness to accept otherflight times or locations varies a great deal with trip purpose. Ideally, we wouldmeasure when and where the passenger wanted to leave and when and where thepassenger left. This is not possible. Instead we proxy this aspect of flight conve-nience and availability with a hedonic characteristic of service quality measured asthe number of departures at the airport of origin. We use the total scheduleddepartures from DOT’s Airport Activity Statistics Form 41 Schedule T3 tomeasure departures. The more departures that occur (higher flight frequency), themore likely that the time of departure will match the most desired time for thepassenger, indicating a higher level of service.

The number of flights at different sized airports has changed since deregula-tion when airlines were allowed to offer service at some airports and eliminateservice to other airports, typically those serving small communities. The pattern of

79

0.00

0.25

0.50

0.75

100

125

150

175

2.00

2.25

2.50

2.75

3.00

80 81 82 83 84 85 86

Date

87 88 89 90 91 92 93

Ave

rage

Num

ber

SegmentsAirlinesInterlines

Figure 3. Patterns of Round Trip Itineraries



449

change in departures for different sizes of airports is displayed in Figure 4. Thegroup of large airports is represented by the 20 airports with the highest totaldepartures between 1979 and 1992. The medium airports are represented by the 20airports ranked between 100 and 119 in terms of total departures. Small airportsare represented by those ranked between 300 and 319, and very small airports arerepresented by the 20 airports ranked between 400 and 419. Departures increasedby 34 percent over the study period for large airports, and by 20 percent formedium and small airports, while departures decreased by 80 percent for verysmall airports. There was a clear shift from very small to larger cities. From theairline’s perspective the shift was economical since offering service to very smallairports that service small municipalities was an expensive proposition when fewenplanements could be expected. However, from the passenger’s perspective, offer-ing service to small communities was desirable since it reduced travel time to analternative city’s airport. For our study, we view departures from small commu-nities as an indication of high service quality and a corresponding willingness ofpassengers to pay higher prices than they would be willing to pay at the closestmedium or large airport.

Another aspect of higher flight frequency, however, is higher congestion andan increased likelihood of flight delay. This suggests that at some point, higherflight frequency has negative consequences on passenger convenience. Since moretickets are likely to be generated at airports that have high flight frequencies, theimpact of the congestion and flight delay consequences of flight frequency is likelyto be more apparent in our sample than aspects that represent convenience.

79 80 81 82 83 84 85 86

Date

87 88 89 90 91 92 93

0

5

10

15

20

25

30

35

40

45

50

LargeMedium

SmallVery Small

Medium

, Sm

all and Very S

mall A

irports

0

50

100

150

200

250

300

350

400

450

500

Lar

ge A

irpo

rts

Figure 4. Number of Departures by Airport Size



450

4.3. Passenger Amenity

We use hedonic characteristics to measure two dimensions of passengeramenity: food, and class of service. First class service brings with it an array ofamenities, among them more space, priority boarding, more attentive service, anda better meal. Class of service is information directly provided for each itinerary inDB1A and thus we know when a particular segment of the itinerary is first class orbusiness class. We construct a dummy variable for this class of service (FirstCl).For carriers that offer only one class of service, typically new entrants, all ticketsare reported as first class, even though the characteristics of service more closelyresemble coach. Consequently, we have constructed the dummy variable (FCNew)to indicate first class service on carriers which only designate one class of servicefor their itineraries during the quarter. It is difficult to differentiate this service witha coach ticket designation.

When no-frills airlines began offering service, one of the amenities they mostvisibly compromised was airline food. It seems clear that the more an airlinespends on food the higher will be the quality of service received by consumers andthe more they would be willing to pay for service. We obtain information on acarrier’s food expenditures from Form 41 Schedule P6 and standardize it by thenumber of enplanements during that quarter. We then construct an itineraryspecific measure by adding the food cost per enplanement for each airline provid-ing service for each segment in the itinerary to measure this service quality char-acteristic (Food). We also experimented with using the airline’s expenditures perpassenger mile for proxying food service. This yielded very similar results. In eithercase, our use of food expense on the itinerary proxies several amenities that arealso correlates of food expense in our model. An airline’s willingness to spendmoney on food often indicates a willingness to provide many more amenities.

Figure 5 indicates there has been a gradual increase in food expenditures perpassenger. Even when one controls for inflation in food prices, we find real expen-ditures per passenger increased by about 25 percent during the sample period. Iffood expenditures were simply passed through, it would indicate that about 3percent of the price of travel, on average, could be attributed to food expense. Thisfigure masks the large variability in food expenses across airlines. Some carriers,for example Pan Am, spent nearly 20 dollars per passenger on average prior totheir merger with National. Low cost carriers like Southwest Airlines spend almostnothing on passenger food.

4.4. Ticket Restrictions

A major feature of airline fare structures is ticket restrictions. These may shiftthe risk of nontravel from airlines to consumers (non-refundability), provide theairlines with improved predictability about demand (advanced booking), orenhance the airline’s ability to use price discrimination information separatingprice sensitive consumers from business travelers with more inelastic demands(Saturday night stay-overs). A major liability of our use of DOT’s DB1A as theprimary source of ticket information is that it includes very limited information onticket restrictions. What can be consistently identified from the fare class codes isthat some kind of restriction was used on a particular ticket leading to the ticket



451

being discounted. However, we are unable to identify what the restriction is.Figure 6 shows the fraction of all tickets which were discounted coach tickets forone way and round trip travel during the sample period. The data show a steadyincrease in both one way and round trip tickets that are discounted coach. Forround trip tickets, only 30 percent of tickets were discounted coach in 1979I,increasing to nearly 90 percent by 1992IV.

Yield management practices on the part of airlines add further complicationsto the discounting process. While two airlines may offer service with an identicalfare, their fare structure may be very different because they offer different numbersof seats at that fare. Consequently, it would be inappropriate to model the situa-tion as though the fare described the price at the margin of an individual’s will-ingness to accept the particular restriction. Because of the unavailability ofdifferent fares, consumers typically will meet the requirements for higher sets offare restrictions than those actually required by their ticket. For example, a con-sumer may meet the 30 day book in advance, no refund, Saturday night stayrestrictions, but because this fare is sold out, they may accept a 14 day book inadvance, no refund, with a Saturday night stay fare.

4.5. Frequent Flier Programs and Zero Coupon Tickets

Frequent flyer miles were introduced in the mid 1980s by then CEO RobertCrandall of American Airlines. The purpose behind this program was to increasecustomer loyalty by offering them free travel at a later date. The program hasproliferated to other industries. Grocery stores now offer discounts for frequent

79

0.0

0.2

0.4

0.6

0.8

10

12

14

16

18

2.0

80 81 82 83 84 85 86

Date

87 88 89 90 91 92 93

Food

Pri

ce I

ndex

RealNominal

Figure 5. Airline Food Expenditures



452

shoppers. The CPI currently ignores zero coupon tickets. The pattern of thesetickets from 1979 to 1992 can be seen in Figure 7 for one way and round trip travel.For both round trip and one way travel there is a spike in 1987 associated with theintroduction of these tickets to about 10 percent for both one way and round triptravel. By 1988, these decline to about 5 percent of round trip travel and about 4percent of one way trips.

The use of zero coupon tickets by consumers is determined by a mix ofexpectations about their value and the rules that a carrier has for their redemption.During periods of increasing prices we see an increased use of free tickets. Animplication of consumers choosing when and where to use zero coupon tickets isthat the price that would have been paid for a zero coupon itinerary tends to be morethan the price for the typical itinerary. In other words, zero coupon tickets tend togo to more exotic places (trip to Hawaii rather than to Florida) than other trips.

During the period of our study, zero coupon tickets were generated solely byair travel, indicating that they should be used to adjust the effective price of airtravel. This view is seriously blurred when, for example, credit card and longdistance phone companies began offering frequent flyer miles as part of the induce-ment to use their services. This raises the issue about where they should be includedin the CPI: as adjustments to air travel or the services that generated the frequentflier miles. While conceptually it makes sense to adjust the price of the service thatgenerated the benefit, this is certainly beyond the scope of this paper. It wouldrequire that we partition benefits earned by travel from benefits earned by othermeans and to constantly revise the value of those benefits based on expectationsabout future air travel prices, the usefulness of the miles based on an airline’s

79

0.000.050.100.150.200.250.300.350.400.450.500.550.600.650.700.750.800.850.900.95100

80 81 82 83 84 85 86

Date

87 88 89 90 91 92 93

Frac

tion

of

Tic

kets

DiscountedFirst Class

Figure 6. Fraction of Discounted Coach and First Class Tickets



453

network or partnership arrangements, and anticipations about the airline’sredemption policy. Instead we adopt an alternative approach that allocates all ofthe benefits of frequent flier programs to reductions in the price of air travel. Weimpute the value of zero coupon tickets by our hedonic equation and then con-struct an expenditure adjustment factor:

λ = value of tickets purchasedvalue of tickets used

The result of this adjustment factor is to lower the effective price index in quarterswhere a large number of zero coupon tickets were used rather than the quarterwhere the ticket is earned.4

4We recognize that our approach is not perfect with respect to zero coupon tickets for at least threereasons. First, there are zero coupon tickets which are systematically excluded from our analysis.Because we use DB1A as our source of information, we do not have consistent information aboutforeign travel (DB1A includes only those trips which incorporate a trip segment within the U.S.).Consequently, because of data limitations, we suspect that the most expensive zero coupon tickets,those associated with international travel, are systematically left out of the analysis. As a result, ouranalysis systematically overestimates the price of domestic travel where many of those flights areearned. Of lesser importance, passenger upgrades offered as part of a frequent flier program may alsosystematically be left out of our analysis if they are not explicitly recorded (passengers get a higherquality of service than we have recorded on their itinerary), again leading to an overstatement of price.Third, those domestic zero coupon tickets that we do include in our analysis almost always come withrestrictions including blackout dates, which tend to make them less valuable than the average tripbetween those city pairs, leading to an overstatement of price. To the extent that these actors areimportant, data limitations suggest that we will not fully capture the effects of frequent flier programsin our adjusted price index.

79

0.00

0.02

0.04

0.06

0.08

0.10

0.12

0.14

0.16

0.18

0.20

80 81 82 83 84 85 86

Date

87 88 89 90 91 92 93

Frac

tion

of

Tic

kets

One WayRound Trip

Figure 7. Fraction of One Way and Round Trip Zero Coupon Tickets



454

4.6. Route Specific Factors

Several route characteristics clearly affect the price of travel. Not the least ofthese is the distance between the origin and destination. Other variables whichmany have attempted to incorporate into modeling the demand side of longdistance travel include weather related variables such as mean temperature differ-ence in an attempt to capture vacation travel in the winter months. Others haveincluded variables which attempt to capture the demand for business travel such asthe number of white collar jobs in an area. In our model we assume that thesefactors are either very slow to change or that they are strongly correlated withother factors in our model (for example, white collar jobs are likely correlated withper capita income). We capture these slowly moving factors with fixed routespecific effects which describe the origin–destination pair. In our model thisamounts to approximately 115,000 route effects for the one way models and134,000 for the round trip models.

4.7. Other Factors

We also considered the use of other factors. These included safety, circuity oftravel, complaint data, and controls for local demand characteristics such asgrowth of local GDP and employment rates at origin and destination. Safety wasexcluded primarily because there is very little data on safety. Aside from theextremely rare event of an actual crash, there is no database that tracks mechanicalfailures, expenditures on security, etc. Even if changes in security could be tracked,the likelihood is that the security procedures were mandated by unmeasurableincreases in security risks. Circuity of travel, defined as the ratio of actual milesflown to the minimum flight length of the route, was included in some models, butdid not perform as well as the number of flight segments. Complaint data was notused because the data are driven by expectations. Persons who expect poor servicerarely find it worthwhile to complain when they get it. Local demand factors wereexcluded for several reasons. With over 115,000 route pairs, collecting local eco-nomic data is daunting. Additionally, there is no theoretical reason for regionaleconomic factors to enter into consumer demand equations such as the hedonicindex. Combining these facts with significant uncertainty as to usefulness led us toexclude local demand factors from our model.

5. Results

Our intended question is how alternative methodologies to those the BLScurrently utilizes affect the Air Fare Index. This requires that we simulate theircurrent methodology over the study period, without the advantage of using theirdata. We then compare these results to price indices which allow for increasedsubstitutability in market baskets by using sales information and adjust the priceindex by holding attributes of service quality fixed. These different indices aresummarized in Table 1.

The replicated airfare indices for arithmetic and geometric means and theairfare index calculated by the BLS are presented in Figure 8. The replicatedindices match fairly well with the actual index reported by the BLS. What differ-



455

ences there are between the replicated arithmetic mean series and the replicatedgeometric mean series are in the direction expected. The replicated arithmeticmean series, with weights that are not entirely fixed, has a higher rate of inflation(average annual inflation is 10.19 percent). The replicated geometric mean series,also with weights that are not entirely fixed, has a lower rate of inflation (averageannual inflation is 8.84 percent). This supports the general notion that geometricmeans lead to lower estimates of inflation.

TABLE 1

A Comparison of Different Indices

BLSAirfareIndex

DB1AArithmetic

Mean

DB1AGeometric

Mean

BenchMark

HedonicModel

HedonicIndexw/o

Restrictions

HedonicIndexw/o

Restrictions

Data Source SABRE DB1A DB1A DB1A DB1A DB1AFixed Route Yes Yes Yes No No NoFixed Class Yes No No No No NoFix Wgt. Area Yes No No No No NoFix Wgt. U.S. Yes Yes Yes No No NoGeo Mean No No Yes NA NA NAArith Mean Yes Yes No NA NA NAQuality Adj. No No No Yes Yes YesZero Coup Adj. No No No Yes Yes YesInf. w/o Zero Adj – – – 4.82% 5.11% 6.71%Avg. Inf. 9.80% 10.19% 8.84% 4.42% 4.69% 6.27%

79

60

80

100

120

140

160

180

200

220

240

80 81 82 83 84 85 86

Date

87 88 89 90 91 92 93

Pri

ce I

ndex

BLS Air Fare IndexSimulated Arithmetic IndexSimulated Geometric Index

Figure 8. BLS and Simulated Air Fare Indexes



456

As we noted in the last section we consider one way and round trip tickets tohave key differences in terms of both the passenger and the airline. Consequently,we also develop separate demand models for one way and round trips and discussthese results before combining them in an overall air travel demand model. We firstestimate a base model that does not include any aspects of service quality:

ln P Dijt j t t ijt= + +α β ε

for ticket i on route j at time t. bt is the coefficient on a dummy time variable wherethe reference category is time period 1979I. Omitted service quality characteristicsmay be correlated with fixed time and/or route effects and thus the ordinary leastsquares estimates of the ′α j s and the ′βts may be biased and inconsistent. Thepredicted price ratio from 1979I to time t is:

P

Pjt

j I

j t

jt

,

exp

expexp .

1979 0=

+( )+( ) = ( )

α βα

β

The estimates for the one way and round trip models are provided in Table 2.Summary statistics including the number of observations, N, and the number ofroute city pairs, Nj, are also listed. The relatively high R2 values given suchdisaggregated data reveal the substantial explanatory power associated with thesefixed route characteristics.

The results of this basic model are useful. While they do not adjust for thecharacteristics of service they do provide a comparison between our method whichprovides for changing weights among ticket types as opposed to the constantweights used in the BLS Consumer Price Index. This is particularly importantbecause the BLS index has fixed proportions of heavily discounted tickets whileour analysis shows that these proportions change substantially over time.

Our hedonic model includes the effects of restrictions and service quality

ln P D Xijt i t t t kijt ijtk

= + + +∑α β γ ε

where the Xkijt are the k measures of service quality and ticket discounting for ticketi on route j. In this model we control for the expenditures on food, the number ofdepartures, the number of segments (more segments indicating a more circuitoustrip and required changes of planes), interlining (the requirement that passengersalso change airlines when they change planes), the class of service (dummy indi-cating first class) and whether or not the ticket had fare restrictions. Food, inter-lining and departures are incorporated in logged form. Parameter estimates for thehedonic model without incorporating restrictions model are presented in Table 3while those for the hedonic model with restrictions incorporated as additionalexplanatory variables are presented in Table 4.

The signs and magnitudes for the coefficients on food and first class serviceare as expected. Also, restrictions and discounting lower the price of the ticket. Thecoefficients on other aspects of service quality remain substantively similar acrossmodels. Restrictions mostly affect the price of round trip tickets (where Saturday



457

TA

BL

E2

Ba

seM

od

elR

egr

essi

on

Res

ul

ts

(ex

clu

des

qu

alit

ym

easu

res

)

Coe

ff

One

Way

Mod

elR

ound

Tri

pM

odel

Coe

ff

One

Way

Mod

elR

ound

Tri

pM

odel

Est

imat

eSt

dE

rror

Est

imat

eSt

dE

rror

Est

imat

eSt

dE

rror

Est

imat

eSt

dE

rror

1979

-10.

0000

000.

0000

000.

0000

000.

0000

0019

87-1

0.53

1658

0.00

0669

0.33

6607

0.00

0587

1979

-20.

1316

470.

0007

170.

0843

860.

0006

2719

87-2

0.55

7406

0.00

0660

0.32

7748

0.00

0576

1979

-30.

1833

850.

0007

030.

1202

220.

0006

3419

87-3

0.64

2275

0.00

0653

0.39

7713

0.00

0582

1979

-40.

2712

590.

0006

880.

2386

980.

0006

3019

87-4

0.70

2656

0.00

0667

0.45

5133

0.00

0584

1980

-10.

3411

440.

0006

990.

3260

220.

0006

3819

88-1

0.75

0258

0.00

1217

0.65

5742

0.00

1002

1980

-20.

3684

450.

0006

770.

3301

740.

0006

3619

88-2

0.68

4097

0.00

0659

0.50

2266

0.00

0571

1980

-30.

3471

910.

0006

770.

3032

500.

0006

4819

88-3

0.77

6370

0.00

1141

0.68

6379

0.00

1008

1980

-40.

3230

890.

0007

120.

3185

640.

0006

3919

88-4

0.68

6545

0.00

0669

0.51

6745

0.00

0574

1981

-10.

4069

690.

0007

030.

4735

440.

0006

6619

89-1

0.87

5029

0.00

1078

0.90

7902

0.00

0929

1981

-20.

3910

120.

0006

790.

4214

430.

0006

4719

89-2

0.66

6478

0.00

0658

0.56

7267

0.00

0573

1981

-30.

3662

080.

0006

730.

4230

670.

0006

6919

89-3

0.56

6361

0.00

0648

0.49

4588

0.00

0570

1981

-40.

3417

810.

0006

830.

4187

450.

0006

4419

89-4

0.59

6645

0.00

0661

0.51

0470

0.00

0562

1982

-10.

3415

140.

0006

820.

3874

780.

0006

5119

90-1

0.64

7923

0.00

0664

0.60

0511

0.00

0567

1982

-20.

3616

500.

0006

670.

4093

080.

0006

3619

90-2

0.67

6124

0.00

0655

0.59

8749

0.00

0561

1982

-30.

3893

780.

0006

590.

4542

540.

0006

4019

90-3

0.65

1849

0.00

0654

0.54

0321

0.00

0563

1982

-40.

3768

220.

0006

820.

4091

820.

0006

3219

90-4

0.70

0685

0.00

0672

0.58

1347

0.00

0569

1983

-10.

3752

010.

0006

790.

3458

010.

0006

2919

91-1

0.61

9265

0.00

0674

0.65

4331

0.00

0577

1983

-20.

4382

980.

0006

570.

4405

450.

0006

2419

91-2

0.61

1315

0.00

0662

0.57

0308

0.00

0563

1983

-30.

4507

130.

0006

490.

5095

890.

0006

4319

91-3

0.66

0826

0.00

0655

0.58

5231

0.00

0570

1983

-40.

4480

370.

0006

550.

5302

970.

0006

2819

91-4

0.67

0174

0.00

0669

0.63

6296

0.00

0570

1984

-10.

4805

020.

0006

590.

5573

380.

0006

3019

92-1

0.68

7399

0.00

0668

0.75

7830

0.00

0582

1984

-20.

4913

750.

0006

450.

5309

160.

0006

0819

92-2

0.64

6160

0.00

0658

0.54

3558

0.00

0565

1984

-30.

4737

860.

0006

400.

5107

740.

0006

1819

92-3

0.64

1092

0.00

0652

0.42

7392

0.00

0559

1984

-40.

4985

200.

0006

460.

5011

070.

0006

0819

92-4

0.67

1827

0.00

0661

0.63

4672

0.00

0572

1985

-10.

5007

580.

0006

510.

4872

700.

0006

0819

85-2

0.50

8989

0.00

0638

0.41

3476

0.00

0595

R2

0.68

112

0.44

185

1985

-30.

5055

680.

0006

360.

4389

100.

0006

04F

113,

511

120,

919

1985

-40.

4232

350.

0006

350.

4256

520.

0006

01N

64,0

24,1

4195

,279

,537

1986

-10.

4293

160.

0006

370.

3650

510.

0005

93N

j79

,248

79,1

3919

86-2

0.38

6636

0.00

0641

0.32

6974

0.00

0594

1986

-30.

4543

210.

0006

280.

3324

520.

0006

0519

86-4

0.49

3267

0.00

0653

0.38

6131

0.00

0597



458

TA

BL

E3

Hed

on

icR

egr

essi

on

Res

ults

wit

ho

ut

Res

tr

ict

ion

s

Coe

ff

One

Way

Mod

elR

ound

Tri

pM

odel

Coe

ff

One

Way

Mod

elR

ound

Tri

pM

odel

Est

imat

eSt

dE

rror

Est

imat

eSt

dE

rror

Est

imat

eSt

dE

rror

Est

imat

eSt

dE

rror

lnF

ood

0.05

7350

0.00

0062

0.07

0076

0.00

0122

1986

-10.

4459

760.

0006

320.

4012

040.

0005

79ln

Dep

-0.0

7745

70.

0002

10-0

.021

254

0.00

0186

1986

-20.

4135

810.

0006

360.

3664

090.

0005

80ln

Inte

r0.

3196

610.

0005

470.

3147

210.

0001

9419

86-3

0.47

7949

0.00

0624

0.37

3821

0.00

0591

Fir

stC

l0.

0098

400.

0002

490.

6220

130.

0003

4619

86-4

0.51

9626

0.00

0648

0.43

0882

0.00

0582

2Se

gs0.

0047

740.

0002

2119

87-1

0.55

3903

0.00

0664

0.38

3329

0.00

0573

3Se

gs0.

0006

170.

0002

1719

87-2

0.58

2817

0.00

0656

0.38

3257

0.00

0564

4Se

gs-0

.109

097

0.00

0208

1987

-30.

6620

000.

0006

490.

4475

290.

0005

7119

79-1

0.00

0000

0.00

0000

0.00

0000

0.00

0000

1987

-40.

7244

730.

0006

620.

5035

070.

0005

7319

79-2

0.11

9343

0.00

0708

0.07

2007

0.00

0607

1988

-10.

7406

590.

0012

050.

6815

580.

0009

7519

79-3

0.17

7663

0.00

0695

0.11

1818

0.00

0614

1988

-20.

7211

650.

0006

540.

5493

430.

0005

6119

79-4

0.26

7474

0.00

0680

0.23

0940

0.00

0610

1988

-30.

7681

250.

0011

300.

7075

440.

0009

8219

80-1

0.33

3729

0.00

0691

0.31

7276

0.00

0619

1988

-40.

7297

680.

0006

650.

5625

440.

0005

6619

80-2

0.35

5464

0.00

0670

0.32

1492

0.00

0617

1989

-10.

8623

610.

0010

680.

9085

090.

0009

0619

80-3

0.33

2617

0.00

0670

0.29

8420

0.00

0629

1989

-20.

7016

870.

0006

530.

6041

310.

0005

6519

80-4

0.28

8215

0.00

0705

0.31

1909

0.00

0621

1989

-30.

6060

470.

0006

440.

5368

350.

0005

6219

81-1

0.37

5977

0.00

0695

0.45

7073

0.00

0647

1989

-40.

6343

990.

0006

560.

5457

390.

0005

5519

81-2

0.36

5019

0.00

0672

0.42

3952

0.00

0628

1990

-10.

6782

610.

0006

620.

6295

400.

0005

6219

81-3

0.33

0924

0.00

0667

0.42

0699

0.00

0651

1990

-20.

7153

180.

0006

550.

6314

470.

0005

5819

81-4

0.30

5416

0.00

0677

0.41

3868

0.00

0627

1990

-30.

6851

300.

0006

540.

5700

960.

0005

6219

82-1

0.30

9437

0.00

0675

0.38

2742

0.00

0633

1990

-40.

7383

590.

0006

720.

6118

600.

0005

6619

82-2

0.33

6192

0.00

0661

0.42

2916

0.00

0618

1991

-10.

6511

710.

0006

730.

6822

290.

0005

7419

82-3

0.36

1963

0.00

0653

0.46

2897

0.00

0623

1991

-20.

6433

030.

0006

630.

6034

640.

0005

6219

82-4

0.35

1123

0.00

0675

0.42

6553

0.00

0615

1991

-30.

6963

900.

0006

570.

6154

490.

0005

7119

83-1

0.34

8629

0.00

0673

0.36

3538

0.00

0612

1991

-40.

7143

630.

0006

700.

6661

610.

0005

6919

83-2

0.41

1501

0.00

0651

0.46

2283

0.00

0607

1992

-10.

7327

430.

0006

700.

7789

810.

0005

8219

83-3

0.42

6179

0.00

0643

0.53

1412

0.00

0625

1992

-20.

6922

130.

0006

600.

5733

280.

0005

6519

83-4

0.41

7013

0.00

0649

0.54

6810

0.00

0611

1992

-30.

6957

540.

0006

560.

4629

760.

0005

5919

84-1

0.45

3261

0.00

0654

0.56

9750

0.00

0613

1992

-40.

7229

490.

0006

640.

6672

300.

0005

7119

84-2

0.46

6200

0.00

0640

0.54

9884

0.00

0593

1984

-30.

4486

230.

0006

350.

5315

230.

0006

02R

20.

6825

40.

4833

419

84-4

0.48

0478

0.00

0640

0.52

5912

0.00

0592

F13

3,34

722

2,15

019

85-1

0.48

2905

0.00

0645

0.51

1242

0.00

0592

N64

,024

,141

95,2

79,5

4219

85-2

0.49

4214

0.00

0633

0.44

6156

0.00

0579

Nj

79,2

4879

,139

1985

-30.

5234

230.

0006

310.

4715

290.

0005

8919

85-4

0.44

2321

0.00

0629

0.46

0142

0.00

0586



459

TA

BL

E4

Hed

on

icR

egr

essi

on

Res

ul

ts

wit

hR

est

ric

tio

ns

Coe

ff

One

Way

Mod

elR

ound

Tri

pM

odel

Coe

ff

One

Way

Mod

elR

ound

Tri

pM

odel

Est

imat

eSt

dE

rror

Est

imat

eSt

dE

rror

Est

imat

eSt

dE

rror

Est

imat

eSt

dE

rror

lnF

ood

0.03

5667

0.00

0064

0.07

3075

0.00

0119

1986

-10.

5501

540.

0006

230.

5626

860.

0005

69ln

Dep

-0.0

6982

10.

0002

05-0

.025

848

0.00

0181

1986

-20.

5236

650.

0006

280.

5364

470.

0005

71ln

Inte

r0.

3110

140.

0005

360.

2824

820.

0001

9019

86-3

0.57

8357

0.00

0616

0.54

0619

0.00

0581

Fir

stC

l0.

4508

800.

0005

100.

5110

570.

0004

6719

86-4

0.62

1167

0.00

0639

0.60

2813

0.00

0573

FC

New

-0.7

4657

10.

0006

010.

0799

690.

0009

2019

87-1

0.65

6667

0.00

0654

0.54

5567

0.00

0564

FR

estr

t-0

.163

372

0.00

0917

-0.3

3605

00.

0008

6319

87-2

0.68

5043

0.00

0647

0.55

0922

0.00

0555

YR

estr

t-0

.143

883

0.00

0135

-0.2

9473

10.

0001

3419

87-3

0.76

9153

0.00

0640

0.61

9987

0.00

0562

2Se

gs0.

0047

740.

0002

2119

87-4

0.82

1586

0.00

0652

0.67

5525

0.00

0564

3Se

gs-0

.002

569

0.00

0211

1988

-10.

7636

610.

0011

790.

6714

670.

0009

504

Segs

-0.1

0866

60.

0002

0219

88-2

0.81

0117

0.00

0644

0.72

2428

0.00

0552

1979

-10.

0000

000.

0000

000.

0000

000.

0000

0019

88-3

0.79

8768

0.00

1106

0.69

7766

0.00

0957

1979

-20.

1303

430.

0006

930.

0851

490.

0005

9219

88-4

0.80

9356

0.00

0654

0.73

6161

0.00

0557

1979

-30.

1944

260.

0006

800.

1274

500.

0005

9819

89-1

0.87

6850

0.00

1046

0.87

8056

0.00

0883

1979

-40.

2767

230.

0006

650.

2400

010.

0005

9519

89-2

0.78

3943

0.00

0653

0.76

6493

0.00

0555

1980

-10.

3440

440.

0006

760.

3107

140.

0006

003

1989

-30.

6848

380.

0006

430.

7093

090.

0005

5319

80-2

0.38

3116

0.00

0656

0.32

4051

0.00

0601

1989

-40.

7114

960.

0006

450.

7206

620.

0005

4719

80-3

0.36

0788

0.00

0656

0.31

1709

0.00

0613

1990

-10.

8629

750.

0006

570.

8007

340.

0005

5419

80-4

0.32

1839

0.00

0690

0.35

1941

0.00

0605

1990

-20.

9016

810.

0006

510.

8078

700.

0005

5119

81-1

0.40

6595

0.00

0681

0.49

3380

0.00

0631

1990

-30.

8772

810.

0006

510.

7495

200.

0005

5419

81-2

0.40

3843

0.00

0659

0.47

4508

0.00

0612

1990

-40.

9307

240.

0006

680.

7887

130.

0005

5919

81-3

0.37

7461

0.00

0654

0.47

5235

0.00

0635

1991

-10.

8502

310.

0006

700.

8634

070.

0005

6619

81-4

0.35

1862

0.00

0664

0.47

1488

0.00

0611

1991

-20.

8443

760.

0006

590.

7981

290.

0005

5519

82-1

0.35

9685

0.00

0662

0.44

1142

0.00

0618

1991

-30.

9019

370.

0006

540.

8136

270.

0005

6419

82-2

0.39

0038

0.00

0648

0.49

2552

0.00

0603

1991

-40.

9225

020.

0006

680.

8582

370.

0005

6319

82-3

0.41

6864

0.00

0640

0.54

0316

0.00

0608

1992

-10.

9430

080.

0006

670.

9629

370.

0005

7519

82-4

0.40

6352

0.00

0662

0.52

3693

0.00

0601

1992

-20.

9068

640.

0006

590.

7752

200.

0005

6019

83-1

0.40

4044

0.00

0660

0.46

3098

0.00

0598

1992

-30.

9104

790.

0006

540.

6673

210.

0005

5419

83-2

0.47

2558

0.00

0638

0.56

3095

0.00

0593

1992

-40.

9416

170.

0006

630.

8695

380.

0005

6519

83-3

0.48

8926

0.00

0631

0.62

3636

0.00

0611

1983

-40.

4836

800.

0006

370.

6573

170.

0005

98R

20.

7028

50.

4969

919

84-1

0.51

6296

0.00

0641

0.66

7603

0.00

0599

F17

7,38

030

1,48

419

84-2

0.56

8221

0.00

0630

0.67

7280

0.00

0581

N64

,024

,141

95,2

79,5

4219

84-3

0.55

7910

0.00

0625

0.67

9142

0.00

0591

Nj

79,2

4879

,139

1984

-40.

5883

240.

0006

310.

6803

850.

0005

8119

85-1

0.58

5641

0.00

0635

0.65

8956

0.00

0581

1985

-20.

5970

590.

0006

230.

5960

440.

0005

6819

85-3

0.62

0830

0.00

0622

0.62

1574

0.00

0578

1985

-40.

5518

320.

0006

220.

6114

160.

0005

76



460

night stayovers can be enforced) and have relatively little effect on one way trips:lowering the price of one way trips by approximately 15 percent and the price ofround trip tickets by approximately 30 percent. These reductions hold regardlessof class of the ticket (first class or coach).

To generate prices indices for the base and hedonic models, it is necessary toaggregate the one way and round trip price indices. This is accomplished by takingtheir expenditure weighted geometric mean value:

PP

t

t

t

t

k

t

t

kt t

− − −

= ⎛⎝⎜

⎞⎠⎟

⎛⎝⎜

⎞⎠⎟

1

1

11

2

12

1 2

ββ

ββ

where βt1 and βt

2 are coefficients based on one way and the round trip tickets,respectively; and kt

1 and kt2 are the expenditure shares of one way and round trip

tickets, respectively.5 We have constructed these indices for the base model and thehedonic model. The results are presented in Figure 9.

Finally, we turn to adjusting the indices based on the base model and ourhedonic model for zero coupon tickets. Since we have imputed values associatedwith all zero coupon tickets, we can use this to obtain the effective price as adeflator. The indices generated by our base and hedonic regression models are

5The formula we use is a geometric mean, but only of two components, one way and round trip,and not of each individual route as with Lent and Dorfman. Also, our weights are different for eachtime period. This index number is referred to as the Young index. All variants of the Young index passthe standard requirements, except that when there are exceptional relative price or quantity changes,monotonicity breaks down. This is not the case for our data (Balk, 2008).

79

60

80

100

120

140

160

180

200

220

240

260

80 81 82 83 84 85 86

Date

87 88 89 90 91 92 93

Pri

ce I

ndex

Basic Price IndexHedonic IndexHedonic with Restrictions

Figure 9. Hedonic and Base Model Regression Indexes



461

presented in Figure 10. As an additional reference, we include an index for theunadjusted revenue yield (revenue per passenger mile flown, normalized to be 100in 1979I). Notice that the yield index (growing 3.65 percent per year over our studyperiod), which incorporates no quality adjustment, grows at a much lower rateafter airline deregulation in the U.S. gave carriers authority to set their own pricesand determine their own routes in 1980. Airlines initially used this authority torationalize their route structures, eventually leading to a use of hub-and-spoketype networks and an indirect routing of some passengers, particularly betweenless common origins and destinations. Together with an increase in trip length(revenue per mile for long trips is lower than revenue per mile for short trips), thissuggests that the yield index substantially underadjusts for changes in price.

Comparing these indices with the hedonic index leads to an interesting con-clusion. The hedonic index has an average annual inflation of 6.27 percent. This ismuch lower than any of the other indices (with the exception of the revenue yieldindex). Specifically, the difference between the replicated indices and the hedonicindex is much larger than the difference between the replicated indices and theactual CPI. This seems to imply that quality adjustment issues are more importantto correcting upward biases in the airfares index than both fixed weight/basket andaggregation issues. However, this result is directly contradicted by our bench mark

79

80

100

120

140

160

180

200

220

240

260

280

300

320

340

360

380

80 81 82 83 84 85 86

Date

87 88 89 90 91 92 93

Pri

ce I

ndex

BLS Airfare IndexHedonic IndexAdjusted Hedonic IndexYield Index

Figure 10. Zero-Coupon Adjusted Hedonic Index Compared to other Indexes



462

regression. The bench mark regression estimates a simple fixed effects model withtime dummies but no quality variables. The average annual inflation resultingfrom this model is 4.42 percent, indicating that adjusting for quality increasesinflation by a small amount. The appropriate conclusion can only be that qualityfell during the period 1979 to 1992. Also, the difference between the replicatedindices and the bench mark hedonic index is much larger than the differencebetween the bench mark hedonic index and the hedonic index with quality vari-ables. This suggests that having a market basket determined by sales in each period(and appropriately aggregating) lowers inflation much more than adjusting forquality raises inflation. Thus substitution bias is more important than quality biasin our sample period. One reason may be the very small amount of routes includedin the BLS market basket (approximately 670). The DB1A data set can identifyover 134,000 different routes in the sample period. It is unlikely that all substitu-tion biases are dealt with when the consumer can only substitute among 0.5percent of the total routes. Quality adjustments are still important in order toavoid all bias (in this case downward bias). It seems likely that this fact would holdfor many of the subindices that make up the CPI (especially indices trackingservice oriented industries and high-tech appliances). Thus, complete correction ofbias in the CPI (upward or downward) may not be attainable without methods tomore accurately control for quality.6

6. Conclusions

In our research we have attempted to take an exhaustive approach to mea-surement issues in price index construction for the BLS airfare index. While ourfindings are constrained by data availability and other resource constraints thatmay not be as constraining for government agencies, we have nonetheless beenable to pursue a number of the objectives for dealing with the biases that the 1997CPI Commission recommended be addressed. We have also provided a detailedprotocol for data collection and analysis that can be replicated and can beenhanced by availability of additional data sources. We find an upward bias in theBLS airfare index over the period considered. However, we have also found that

6Lent and Dorfman (2005) have used the same origination–destination information as we do,though their study interval is between 1995 and 2002. Their approach follows the standard BLSapproach based on matching, though they require that like items match on both fare class and the exactitinerary. Compared to the BLS index, their approach leads to the inclusion of a much broader basketof trips. However, a shortcoming in their approach is that nearly half of the itineraries do not have exactmatches. They carry out an imputation to adjust for the non-matching items based on trip distance.Though the contrast may appear subtle, our approach relies on the DOT supplied trip-break indicatorsto identify the destination in round trip travel. Travel between the origin and the destination is, after all,what is ultimately demanded by the passenger. Were we to implement matching over our study period,there would be an even more substantial number of non-matches since deregulation and mergers led tomany reconfigurations of the airlines’ networks. There were approximately 15 mergers and servicecessations among major carriers over our study period, while the industry was much more stablebetween 1995 and 2002 with only two major service cessations. Indeed, we chose our study periodbecause of its volatility. We explored the use of an alternative circuity measure (itinerary distance/origination–destination distance) as an additional variable in our hedonic model, but found it to behighly collinear with our measure of indirect routing (number of ticket coupons), and consequently, weleft it out of our final model. Of course, our hedonic model allows us to incorporate other attributes forwhich matching is difficult to implement, and that Lent and Dorfman do not consider, such aspassenger interlining.



463

the data and computational resources required to implement the Commissionrecommendations for the airline industry, one which for historical reasons has asubstantial amount of detailed data, are quite onerous and the corrections via thehedonic approach are not without their own problems. Because of issues of prac-ticality and implementability, the goals of the Commission recommendationsremain illusive and problematic.

References

Abraham, Katharine, John S. Greenlees, and Brent R. Moulton, “Working to Improve the ConsumerPrice Index,” The Journal of Economic Perspectives, 12, 27–36, 1998.

Armknecht, Paul A. and Daniel H. Ginsburg, “Improvements in Measuring Price Changes in Con-sumer Services: Past, Present, and Future,” in Zvi Griliches (ed.), Output Measurement in theService Sectors, University of Chicago Press, Chicago, IL, 1992.

Baker, Dean, Getting Prices Right: The Debate Over the Consumer Price Index, M. E. Sharpe, Armonk,NY, 1998.

Balk, Bert M., Price and Quantity Index Numbers; Models for Measuring Aggregate Change andDifference, Cambridge University Press, New York, 2008.

Boskin, Michael J. and Dale W. Jorgenson, “Implications of Overstating Inflation for IndexingGovernment Programs and Understanding Economic Progress,” American Economic ReviewPapers and Proceedings, 87, 89–93, 1997.

Boskin, Michael J., Ellen R. Dulberger, Robert J. Gordon, Zvi Griliches, and Dale W. Jorgenson, “TheCPI Commission: Findings and Recommendations,” American Economic Review Papers andProceedings, 87, 78–83, 1997.

Boskin, Michael J., Ellen R. Dulberger, Robert J. Gordon, Zvi Griliches, and Dale W. Jorgenson,“Consumer Prices, the Consumer Price Index, and the Cost of Living,” The Journal of EconomicPerspectives, 12, 3–26, 1998.

Caves, D., L. R. Christensen, M. W. Tretheway, and R. J. Windle, “An Assessment of the Efficiencyof U.S. Airline Deregulation via an International Comparison,” in Elizabeth E. Bailey (ed.), PublicRegulation: New Perspectives on Institution and Policies, MIT Press, Cambridge, 1987.

Ducharme, Louis-Marc (ed.), Bias in the CPI: Experiences from Five OECD Countries, StatisticsCanada Analytical Series: Prices Division, 10, 1997.

Färe, R., S. Grosskopf, M. Norris, and Z. Zhang, “Productivity Growth, Technical Progressand Efficiency Change in Industrialized Countries,” American Economic Review, 84, 66–83,1994.

Gordon, Robert J. and Zvi Griliches, “Quality Change and New Products,” American EconomicReview Papers and Proceedings, 87, 84–8, 1997.

Griliches, Zvi, “Hedonic Price Indexes for Automobiles: An Econometric Analysis of Quality Change,”in Price Statistics of the Federal Government, U.S. Government Printing Office, Washington DC,1961.

Jeon, B. M. and Robin C. Sickles, “The Role of Environmental Factors in Growth Accounting,”Journal of Applied Econometrics, 19, 567–91, 2004.

Lent, Janice and Alan Dorfman, “A Transaction Price Index for Air Travel,” Monthly Labor Review,128, 16–31, June 2005.

Morrison, S. A. and C. Winston, The Evolution of the Airline Industry, Brookings Institution, Wash-ington DC, 1995.

Pakes, Ariel, “A Reconsideration of Hedonic Price Indices with an Application to PC’s,” The AmericanEconomic Review, 93, 1578–96, 2003.

———, “Hedonics and the Consumer Price Index,” forthcoming in Les Annales d’Economie et deStatistique, 2008.

Perloff, Jeffrey, Robin C. Sickles, and Jesse C. Weiher, “An Analysis of Market Power in the U.S.Airline Industry,” in Daniel J. Slottje (ed.), Measuring Market Power, North-Holland, Amster-dam, 309–23, 2003.

Pollak, Robert A., “The Consumer Price Index: A Research Agenda and Three Proposals,” TheJournal of Economic Perspectives, 12, 69–78, 1998.

Primont, Diane F. and Mary F. Kokoski, “Comparing Prices Across Cities: A Hedonic Approach,”BLS Working Papers #204, 1990.

Quigley, John M., “Nonlinear Budget Constraints and Consumer Demand: An Application to PublicPrograms for Residential Housing,” Journal of Urban Economics, 12, 177–201, 1982.



464

Ridker, Ronald G. and John A. Henning, “The Determinants of Residential Property Values withSpecial Reference to Air Pollution,” The Review of Economics and Statistics, 49, 246–57, 1967.

Rosen, S., “Hedonic Prices and Implicit Markets: Product Differentiation in Pure Competition,”Journal of Political Economy, 82, 34–55, 1974.

Schwartz, Amy Ellen and Benjamin P. Scafidi, “Quality Adjusted Price Indices for Four Year Col-leges,” Mimeo, New York University and Georgia State University, 2000.

Trajtenberg, Manuel, Economic Analysis of Product Innovation: The Case of CT Scanners, HarvardUniversity Press, Cambridge, 1990.

Triplett, J. E., “Measuring Technological Change with Characteristics-Space Technique,” BLSWorking Papers #141, 1984.

———, Handbook on Hedonic Indexes and Quality Adjustments in Price Indexes: Special Application toInformation Technology Products, Brookings Institution, Washington DC, 2004.



465

A HEDONIC PRICE INDEX FOR AIRLINE TRAVEL

Documents