Airline Network Effects and Consumer Welfare

As we describe in more detail below, the concept that air carriers compete on dimensions of quality and that effects on quality should be balanced against effects on price has been recognized in the economics literature at least since Douglas and Miller (1974). However, these concepts have not been consistently applied to policy related to air travel.

The concepts of quality-adjusted prices and consumer welfare are closely related. Indeed if there were only one product in the market, the quality adjusted price (Equation 5) would be identical to the consumer surplus (Equation A1). See generally Willig (1978).

To be clear, this paper does not offer an analysis of the sources of possible hub market power, which may be accounted for by the specifics of the current regulatory environment, including elements that encourage entry.

This paper focuses on presenting methods for evaluating prospective airline mergers. Evaluating actual mergers on a retrospective basis is beyond the scope of this paper. Mehta and Miller (2012) and Luo (2013) have both evaluated the impact on fares of the Delta-Northwest merger and concluded that it led to no substantial increases in nominal fares.

The most recent estimates indicate that even the nominal hub premia have been declining in recent years.

Numerous authors have investigated the fare effects of airline hubs, including Borenstein (1989, 2005, 2013), Berry (1990), Brueckner et al. (1992), Evans and Kessides (1993), Lee and Luengo Prado (2005), Lederman (2007, 2008), Borenstein and Rose (2013), Berry and Jia (2010), and Ciliberto and Williams (2010).

Indeed, Borenstein (1989) notes: “Though the link between airport dominance and high fares seems clear, a welfare analysis of increased airport concentration must also include the benefits that may accrue from hub operations. [note omitted] … Greater flight frequency, easier connections, and more nonstop flights may also be associated with these route systems. In this regard, the estimated impact of these quality factors on price, presented in the previous section, should not be interpreted as hedonic prices. [note omitted] These possible benefits of mergers or other increases in airport shares should be weighed against the higher prices that seem likely to result.”

Our approach is closest to Berry (1990).

Most papers estimate either a nested logit model with the outside good in one nest and all inside goods in another nest or a simple logit model with no outside good. For a general description of the restrictive substitution patterns implicit in the logit functional form, see Berry (1994).

Peters (2006) points to only one other working paper that allows for imperfect substitution across airports.

For simplicity, we suppress the f subscript since we only observe itinerary-level data and not flight-level data. For a similar model, see Ackerberg and Rysman (2005).

The model allows a consumer’s preferences for flights within the same itinerary to be correlated. This specification allows each consumer to realize a draw from the logit error distribution for each flight in addition to each itinerary. The intuition is that each consumer has idiosyncratic preferences over each flight, perhaps because actual and preferred departure times vary. Note that this means that consumers only receive a different number of draws from the error distribution for flights in the event of a merger if the merger results in addition (or subtraction) of flights. If a merger were to simply “smush” the networks of the merging carriers, then consumers would receive an identical number of draws from the error distribution for flights.

The GEV model proposed by Peters (2006) allows substitution patterns to depend on two product characteristics: i) whether the product was non-stop or connecting; and ii) the origin and destination airports. However, the modeling of flexible substitution patterns across multiple product dimensions substantially increases the computational burden. In particular, with multiple overlapping nests, the market shares cannot be inverted analytically so the computation requires a contraction mapping algorithm. See Berry (1994).

Formally, using the notation of Peters’ (2006), we assume that ρ_D=ρ₀. As a result, a=1 and the market share equation in the appendix collapses to the equation shown below.

In our model, consumer heterogeneity enters only through the error term. Armantier and Richard (2008) and Berry and Jia (2010) are two notable papers that estimate more flexible functional forms that allow for greater consumer heterogeneity. Those papers suggest that price sensitivity is negatively correlated with the strength of preferences for non-price (quality) attributes. As a result, welfare estimates based on averages across all passengers such as those presented here are likely to understate the magnitude of impact from improvements in quality.

Formally, we assume that the idiosyncratic error term, ε_ijt, is distributed Type I Extreme Value. For further details, see Cardell (1997).

Own-price elasticity with respect to price can be derived from the share equation as:

εj=αρApj(1−(1−ρAρ0)sj|gj−ρAρ0(1−ρ0)sj|gj×sgj|G−ρAsj).

This is a standard assumption in the discrete choice literature; see Berry (1994).

We use linear instrumental variable methods to minimize the difference between observed and predicted shares. We compute observed shares as the ratio of the number of passengers choosing a given product in a quarter to the geometric mean of the endpoint MSA-level populations.

Our approach is similar to that used by Berry and Jia (2010).

We assign regional operating carriers to their mainline parent on a route-by-route basis using OAG data.

For example, for distances <350 miles, circuity must be <4.15, while for distances >2000 miles, circuity must be <1.45.

We define the following cities with multiple airports: CH2 (Chicago O’Hare (ORD) and Chicago Midway (MDW)), CL2 (Cleveland Hopkins International (CLE) and Akron Canton (CAK)), DA2 (Dallas/Ft. Worth (DFW) and Dallas Love Field (DAL)), HO2 (George Bush Intercontinental (IAH) and Hobby (HOU)), LA3 (Los Angeles International (LAX), Burbank (BUR), and Long Beach (LGB)), MI2 (Miami International (MIA) and Ft. Lauderdale (FLL)), NY3 (LaGuardia (LGA), Newark (EWR), and John F. Kennedy (JFK)), SF2 (San Francisco (SFO) and Oakland (OAK)), DC3 (Reagan National (DCA), Washington Dulles (IAD), and Baltimore-Washington (BWI)), and TA2 (Tampa International (TPA) and St. Petersburg/Clearwater International (PIE)).

Note that because we define a product by the ticketing carrier, we treat code-shared flights as separate products. For example, suppose NW and DL codeshare on a particular route that is operated by DL. This would show up in the data as two sets of observations, one for a flight operated by DL and ticketed by NW and a second for a flight operated by NW and ticketed by NW. In general, the physical characteristics of the two flights will be identical, but ticketing characteristics (for example, price and a fixed effect for ticketing carrier) will be different.

These screens are standard in the literature. See, for example, Borenstein (1989); Peters (2006).

We also eliminate routes that do not have both endpoints in the continental US.

Peters (2006) applies a similar screen. The reason for doing so is that the logit model assumes that choice probabilities (shares) are integrated over multiple individuals, each with her own i.i.d. logit error term. Without sufficient product-level observations, it is not possible to differentiate between the product-level unobservable quality and the individual idiosyncratic error term. For example, suppose we observe a product with just one passenger in the quarter. Further, suppose that the reported fare for this product is twice the average fare on the route, while other observable characteristics are equal to the average values on the route. With just a single individual observation, the model cannot determine whether the individual chose to fly on the flight – despite its high price – because the product itself has a high level of unobserved quality or because the individual has an idiosyncratic preference for the product and is therefore willing to pay a high fare. With multiple observations, the logit model integrates over the idiosyncratic error term in order to identify the unobserved quality.

We estimate the model using the outbound segment of the round-trip itinerary.

Our results are substantively similar when we examine all routes.

Our “discrete choice” model of demand similar to that used by many economists and operations research specialists who have studied airline demand. See, for example, Morrison et al. (1989), Berry (1990), Coldren et al. (2003), Peters (2006), Armantier and Richard (2008), and Berry and Jia (2010). The route fixed effects in the model generate a route-specific constant term. This route-specific constant term captures the aggregate level of flying (relative to the outside good) on a route-by-route basis. We assume that the quality of the outside good stays constant over the sample period.

See note 11 for a description of our econometric treatment of frequency.

For the development of this methodology, see Willig (1978).

In particular, we count the number of products offered by competing carriers on a route, the percent of products offered by competing carriers on a route that are non-stop, passenger-weighted mean circuity for products offered by competing carriers on a route, the mean of the number of itineraries making use of segments for competing carriers on a route, the mean of the endpoint populations for those itineraries for competing carriers on the route, the mean of non-stop network quality for products offered by competing carriers on a route, and the mean of connecting network quality for products offered by competing carriers on a route.

By focusing on itineraries that fly to or through hubs, we eliminate the demand effects associated with flying from a hub.

A variety of test statistics reject the null hypothesis that the instruments are weak. Specifically, the Angrist-Pischke (AP) first-stage χ² and F statistics reject the null hypothesis that each endogenous regressor is under- or weakly identified.

Table 4 reports the results from the first stage regressions of the right-hand-side endogenous variables on exogenous instruments.

Berry and Jia (2010) estimated an average own-price elasticity of –2.1 using data from 2006 and a different functional form. They estimate an average own-price elasticity of –2.9 when aggregating across certain airports within cities, a methodological choice similar to ours.

While more recent papers such as Peters (2006), Armantier and Richard (2008), and Berry and Jia (2010) do not focus on valuing the characteristics of air travel, such values can be derived from the parameters of the logit models that they estimate. For example, Armantier and Richard (2008) find evidence on preferences for travel time that are consistent with the predictions of Morrison and Winston (1995).

To see this, note that “inconvenience” enters the utility function in logs. The value of a change in inconvenience from 6 to 3 h is equivalent to $19.48 × (ln(6) – ln(3)) = $13.51, where $19.48 is the value coefficient reported in Table 3. We have also experimented with allowing “inconvenience” to enter linearly and find similar results.

The average “inconvenience” in our sample is approximately 7.3 h. Recall that our measure of convenience incorporates both time in transit as well as the extent to which the schedule matches desired departure times. For example, suppose a passenger desires to depart at noon on a flight that takes 3 h to travel from the origin to the destination. If a carrier that previously only offered a flight that departs at 3 PM adds a second flight that departs at noon, this would be equivalent to reducing the inconvenience of the schedule from 6 to 3 h. We separately control for the duration of the individual flight itinerary with distance and connection time variables.

See GAO (2001). See also (1999), GAO (1989 and 1990), and DOT (2001).

For example, DOT (2001) rejects quality as an explanation for higher fares based on its conclusion that hub carriers typically lower fares (and maintains quality) in response to entry at hubs by low cost carriers. The fact that airlines respond to competition is not surprising and does not undermine the conclusion that higher quality could explain part or all of the observed hub premia, especially today after significant entry by low cost carriers has already occurred.

Borenstein (2013) finds that the nomimal hub premium has declined since the mid-1990s. He attributes these declines to a number of factors, including LCC competition, decreases in costs (in part from increases in load factors), and declining market power of legacy carriers.

Evans and Kessides (1993) also suggest that service quality may explain at least part of the observed hub premia.

We compute hub premia, controlling for distance, following the methodology in Borenstein (2005). This method was presented to the TRB in 1999. The academic literature has often used regression analysis to more fully control for factors affecting price. (See, e.g., Evans and Kessides 1993) We follow the simpler approach for two reasons. First, our goal is to inform the policy debate. So it is helpful to use a similar starting point. Second, price regressions in the airline industry can raise subtle econometric questions, such as the endogeneity of market share measures typically used in the academic literature, which can be difficult to resolve.

Borenstein analyzed the 50 busiest airports. To remain consistent with the rest of our analysis, which focuses on the continental US, we exclude Honolulu (HNL) and Kahului (OGG). We explain the methodology in detail in the Appendix.

We note that the fare premium calculation does not account for the mix of passengers. To the extent that hubs tend to be located in commercial centers, one explanation for the higher observed average fares at some hub airports may be the presence of relatively more business travelers. See Lee and Luengo-Prado (2005).

An alternative way to estimate the average nominal fare premium is to run a regression of average nominal fare (in logs) on a dummy for hub airport. We include a polynomial in non-stop distance, year-quarter fixed effects to control for the possibility that passenger distributions change across routes over time, and we weight by passengers. This approach indicates that the nominal fare hub premia is positive and significantly different from zero.

See Willig (1978) for a discussion of the conditions under which this is appropriate.

This scaling is embedded in the constant term and route fixed effects, which reflect the size of the inside goods relative to the size of the market and are captured in δ.

In order to compare nominal and quality-adjusted fare premia, we measure all premia relative to nominal fares.

A regression of quality-adjusted fares (in logs) on hub status, again controlling for distance and time and weighting by passengers, indicates that the average quality-adjusted hub premium is negative and significant.

For the purposes of results reported here, we do not consider the route-specific fixed effects, which capture the average benefits of flying versus not flying, the airport-specific fixed effects, which capture the benefits of flying to specific airports within multi-airport cities, or unobservable route-product-specific elements of quality. We do this (1) to focus most directly on the elements of quality that can be clearly interpreted and (2) in recognition of the fact (as explained in the Appendix) that we do not estimate effects on all routes and therefore do not have estimates of the value of unobserved quality for all observations. Our substantive conclusions are unchanged if we account for unobserved quality on those observations where estimates are available.

The “other” category captures aspects of air travel that are less directly related to hubs. This category includes factors like carrier fixed effects, which capture unobserved attributes specific to certain carriers but invariant across routes.

A regression of partially quality-adjusted fares (in logs) on hub status, again controlling for distance and time and weighting by passengers, also indicates that the average partially quality-adjusted hub premium is negative and significant.

We do not attempt to evaluate how the relationship between hub premia and quality-adjusted fares would affect structural changes, such as the banning of frequent-flyer programs or making airport facilities available to new entrants were made. For a discussion of the Aviation Investment and Reform Act for the 21st Century – AIR-21, see Snider and Williams (2001).

Atlanta also serves as a hub for AirTran, which was recently acquired by Southwest.

DoJ concluded that “[o]ur best estimates of the likely increases in consumer welfare significantly exceeded the feared harm to consumers in the overlap routes served by the two carriers.” See Heyer et al. (2009).

Morrison et al. (1989) is the first paper that we are aware of to propose this approach in the context of evaluating airline mergers.

We model changes in quality characteristics using scheduling data from the third-quarter of 2008. We simulate the effect of the merger by comparing stand-alone schedules to schedules in which all Northwest flights are recoded as Delta flights. This analysis therefore does not incorporate any estimates of changes in schedule resulting from the merger. As discussed above, we do not attribute any increase in consumer welfare to the combination of frequencies within the merged carrier over and above any benefits arising from other quality attributes such as scheduling convenience. To the extent that the merger would have caused the combined carrier restructure the post-merger network, these effects could also be accounted for in the analysis. Indeed, carriers often construct detailed plans for post-merger networks as part of the merger planning process. In some cases, those plans indicate reductions in capacity on some routes (e.g., where the pre-merger carriers fly similar schedules), and increased capacity on other routes (e.g., where the combined carrier produces sufficient flow to sustain increased service). In general, it is not possible to say which direction these effects will go without access to the precise schedule planning.

See Brueckner et al. (2013), at Table 3. We base our estimates on their “market” model, which generates the highest estimate of two-to-one effects and is therefore most conservative. Our results are robust to using alternative specifications. The Brueckner et al. model controls for whether an LCC is present on a route, but does not allow the competitive effects of reducing the number of legacy carriers to vary with LCC presence. In the case of the Delta-Northwest merger, the combined carrier faced competition from Airtran, particularly on routes out of Atlanta.

Our use of the Small and Rosen welfare formula is conceptually equivalent (and numerically approximately equivalent) to computing the changes in consumers′ surplus due to changes in quality-adjusted prices, given the specifications of the demand functions we employ. As discussed in the Appendix, for these demand functions the Small and Rosen welfare formula permits this calculation in one integrated step. In contrast, quality-adjusted prices are a more natural way to address the literature on hub premia, which analyzes price differences across airports rather than the welfare effects of a transaction.

This approach is conservative in that it increases the nominal prices of the non-merging firms without letting them respond to the improved quality of the merging carrier via enhanced quality and/or lower nominal fares. A recent paper about the price effects of the 1987 merger between USAir and Piedmont found that competing carriers did lower their nominal fares in response to the merger. The paper goes on to suggest that this pattern could occur because of “better service by the merged carrier” or “’S-curve’ effects” (analogous to our network breadth benefits). See Kwoka and Shumilkina (2010).

We weight by pre-merger passengers. Negative values indicate that the schedule becomes more convenient.

We weight by pre-merger passengers.

This measure of consumer welfare captures welfare changes realized by passengers of non-merging carriers as well as passengers who choose one of the merging carriers. The welfare of those customers who choose non-merging carriers is impacted for two reasons. First, the nominal fares may change due to changes in the competitive structure on the route. Second, those customers benefit from the option of choosing a merged carrier with its merger-induced quality improvements. For some passengers, the improvement in quality on the merging carrier may be so great that they will be induced to switch.

For the purposes of these calculations, we assume that airlines do not face capacity constraints. Airlines typically employ sophisticated “spill and recapture” algorithms that can be used to assess the extent to which capacity constraints bind. We also do not model any changes in network configuration. To the extent that airlines cannot seat all passengers desiring a seat post-merger, the welfare gains from quality improvements may be reduced. On the other hand, airlines may be able to respond to capacity constraints by flying larger aircraft on a route or adding frequencies.

See US Department of Justice 2008. “Statement of the Department of Justice’s Antitrust Division on its Decision to Close its Investigation of the Merger of Delta Air Lines Inc. and Northwest Airlines Corporation.” http://www.justice.gov/atr/public/press_releases/2008/238849.htm(last updated October 29, 2008).

For example, applying the valuations from Table 3, the change in convenience implies average consumer benefits of approximately $1–$6 per flight (in each direction). Similarly, the increase in nonstop destinations translates into $2–$28 of consumer benefits on average per flight (in each direction). The variation in these benefits is driven by differences across routes in the changes in scheduling convenience and network quality as a result of the merger. Consumers get additional benefits from connecting flights as well as the conversion of codeshare itineraries to online itineraries. In contrast, the assumed increases in nominal prices based on the estimations in Brueckner et al. (2013) come to $3–$11 depending on the type of route.

To see how this translates into welfare, consider the first row based on ATLDTW. The convenience benefits are worth $3.45 per passenger while the improved network is worth $25.13. Based on pre-merger passenger counts (that is, not allowing passengers to switch to flights that become more desirable), this accounts for $6.6 million of the $7.7 million in total benefits. The remaining benefits are accounted for by changes in other attributes (for example, codeshare status) and benefits to switchers (as described in the Appendix). In contrast, the offsetting fare effects are predicted to be just 1.4% of total revenue or $0.5 million.

As shown in Equation (A1) in the Appendix, the changes in welfare depend on the changes in product characteristics and a non-linear function of shares that determines how much those changes matter. So, for example, changes in characteristics of products with low share get relatively less weight in the welfare calculation than changes in characteristics for more highly valued characteristics. As we discuss in the Appendix, we estimate our demand parameters using data from 2009–2010. Consequently, our demand model does not estimate the ξ’s for the period surrounding the DL-NW merger. To address this issue, we solve for the implicit ξ’s such that the predicted shares match the observed shares in the third quarter of 2008. Using these ξ’s allows us to weight the welfare function according to pre-merger observed shares.

Benefits can arise on routes where there is no non-stop overlap for a few reasons. First, some interline itineraries that were previously excluded from the model could become online itineraries, creating new options for passengers. Second, some codeshare itineraries could become online itineraries. Third, on routes on which there are connecting overlaps, convenience could improve due to better flight times. Fourth, even on routes where there is no overlap at all, to the extent that there are airport-level overlaps, benefits from improved network breadth may be realized.

Merger simulations are often used for the evaluation of price effects from mergers. For more details, see, for example, Budzinski and Ruhmer (2010). However, research has shown that merger simulations can be quite inaccurate in the airline industry. See, for example, Peters (2006).

Willig (2011) shows how quality-adjusted prices can be incorporated into UPP analysis.

Delta Air Lines, Inc., Form 10-K For the fiscal year ended December 31, 2009 at 29.

However, we conservatively assume that convenience and network breadth are measured only on the basis of airports.