Below is the uncorrected machine-read text of this chapter, intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text of each book. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.
100 Other Approaches to the Use of Disaggregated Socioeconomic Data in Air Passenger Demand Analysis The previous chapter described the results from incorporating a particular example of regional disaggregated socioeconomic data into the econometric modeling framework that is commonly used (with aggregate socioeconomic data only) in airport analyses of passenger enplanements over time. This chapter presents analyses of new forms or sources for disaggre- gated socio economic data about air passenger as well as other approaches to incorporating dis- aggregated socioeconomic data into air passenger demand analysis. The new types of data result from the capture of passenger data through their credit card purchase behaviors. These data are of increasing interest to marketers and others because they provide increasingly individualized insights in consumer purchasing patterns and decisions. The examination of other approaches involves two ways of characterizing air passenger demand analysis and the potential roles for disaggregated socioeconomic data in each of them. The first approach makes use of the differences in propensities to travel by air that can be developed from the analysis of passenger and consumer survey data. These differences in propensity are suggestive of additional ways to specify econometric models of air passenger behavior that include disaggregated information on population cohorts that exhibit greater propensities to travel by air. New Forms of Disaggregated Socioeconomic Data for Passenger Demand Analysis Thus far, this report has highlighted how traditional sources of disaggregate socioeconomic data have been used in airport demand modeling applications. Traditional data sources typi- cally include surveys of air travelers (e.g., airport surveys, travel surveys, consumer expen- diture surveys, etc.). As discussed in Kressner and Garrow (2012), there has been increasing interest in using non-traditional data sources for travel demand modeling applications. The interest is motivated in part by the explosion of large, third-party data sources that offer the potential of lower data collection costs per respondent. These big datasets, which range from mobile phone signal traces and GPS data to credit card spending patterns, collectively provide detailed spatial and temporal data about individualsâ behaviors and mobility patterns, often in real time. This chapter provides an overview of new sources of socioeconomic data, including credit card transaction and cell phone data that may be useful for airport demand modeling applications. Given a broad overview of the potential applications of nontraditional data sources, the section concludes with a description of an in-depth study conducted as part of the current project to assess the potential of using credit card reporting data for air travel demand modeling applications C H A P T E R 5
Other Approaches to the Use of Disaggregated Socioeconomic Data in Air Passenger Demand Analysis 101 Credit Card Transaction Data In many sectors of the economy, it is becoming increasingly recognized that the credit card transaction records maintained by credit card companies contain detailed information about spending patterns by individual card holders that can be potentially highly valuable for mar- keting and other purposes. Although the credit card companies of course know the identity of each card holder, privacy considerations prevent the release of data in a form that would allow the individual card holder to be identified. Even so, the credit card companies recognize the potential value of suitably de-identified information derived from the transaction data and are actively exploring ways to market this information. At the same time, potential users of this information are exploring ways to utilize the transaction data and link these data to broader socioeconomic characteristics, given the constraints that the specific individual making each transaction is not known. Multiple firms have explored ways to link credit card transaction information to disaggre- gate, individual-level consumer characteristics, with varying levels of success. For example, both American Airlines and the Airlines Reporting Corporation (ARC) explored whether they could merge socioeconomic data to individual ticket transactions. The key challenge is that the zip code associated with the airline ticket purchase is that of the merchant, not the individual traveler. Thus, all customers who purchase an airline ticket through Expedia, for example, have an Atlanta zip code associated with the ticket purchase. Other information, such as the passengerâs name and origin airport, can be provided to companies that maintain databases of customer addresses and demographics [see Binder, et al.(2014); Kressner and Garrow (2012); and MacFarlane, Garrow, and Mokhtarian (2015) for example applications in travel demand modeling]. The companies can use this information to link to their customer database which contains socioeconomic information. However, the individualâs name and home metropolitan area may not provide enough information to uniquely identify a home zip code or census sub-region for the passenger, especially for common last names, such as Smith or Li (Carvalho, 2015; Howard, 2015). ARC estimates that his approach results in a unique match for approximately 10% of its tickets. For some applications, most notably determining catchment areas, this match rate is sufficiently large to provide useful information for airports (Howard 2015). Sometimes, additional information about the customers is available that enhances the abil- ity to identify detailed travel histories for individual travelers. For example, a major credit card approached ARC because they were concerned that, due to the recent wave of mergers in the airline industry, many of their elite customers were no longer allowed to use their elite credit card to access airlinesâ airport lounges. The major credit card company wanted to identify which airports its elite customers were using, so that it could invest in an airport lounge for those cus- tomers. The major credit card company sent a list of credit card numbers and customer names to ARC, which ARC then matched to its ticketing database. The ability to use customer name and credit card numbers allowed ARC to identify travel histories for these elite customers (e.g., how often they travelled and which airports they most often used). The credit card company merged these travel histories with its internal customer database to link its elite customersâ travel histo- ries and socio-demographic information. Armed with this information, the credit card company built a number of airport lounges for its customers in the United States (Howard 2015). These are some examples of how credit card information has been used for concession plan- ning/facility location in the airline industry. Depending on the level of transactional detail that is available, these data could also potentially be used to identify how often individuals travel by air, which airport(s) and airline(s) they use, and whether they pay for parking at or near the airport. As part of the current project, we obtained credit card transaction data from a financial planning firm to explore this research question in more depth.
102 Using Disaggregated Socioeconomic Data in Air Passenger Demand Studies Cell Phone Data Over the past decade, there has been increased interest in using mobile phone signal traces and GPS data from cell phones to anonymously track individual customers through time and space. According to Bill King of AirSage (King 2014; King 2016) several airports and plan- ning organizations have used cell phone data to identify catchment areas and/or to identify the number of passengers that have traveled in a specific travel corridor. The latter is particularly helpful for understanding the overall market potential in high speed rail corridors served both by both auto and air, as in the case for a study conducted for the AtlantaâCharlotte corridor (King, 2015). For privacy reasons, the identification number assigned to a particular cell phone is scrambled every 30 days, which effectively limits any data analysis to one based on cross- sectional (versus longitudinal) data. There is also a growing number of airport ground access studies that have made use of cell phone data to identify travel patterns to and from airports. With some processing, it is possible to distinguish between trips by air passengers and those by airport employees and between trips by residents of the region and visitors to the region. Past studies using cell phone data have included an analysis of the geographic distribution of ground trip ends for travel to and from Ontario International Airport in California on a regular weekday, a regular weekend day, and the day before Thanksgiving, and an analysis of travel between Detroit Metropolitan Airport and hotels in downtown Detroit (King 2014). These are some examples of how cell phone data have been used to identify ground trans- portation movements to an airport or passenger volumes by mode in a specific corridor. Due to the need to protect individualsâ privacy, limited demographic and socioeconomic informa- tion about the travelers is available; however, some information can be obtained by identifying a (relatively large) geographic area in which the individual likely lives, such as the individualâs home zip code. These data could also be helpful for identifying resident trips from home to airport, resident trips from work to airport, and non-resident trips from the airport. Other Data Sources There are numerous other potential sources of disaggregated data on traveler behavior. These sources include airline ticket booking and reservation data, other location-tracking devices that take advantage of cell phone signals, social media data, vehicle license plate data that is routinely collected in airport parking facilities, web search data from visitors to web sites, and transaction data collected in the course of airport concession sales or other airport operations. ARC has also worked with airports to understand the (likely) movements of connecting passengers in its terminals. This is useful for concession planning. By modeling which gates con- necting passengers arrive at and depart from, airports can forecast high foot-traffic areas. ARC built a model of connecting passenger movements by linking its ticketing data, which provided the sequence of flights the customer purchased, to gate information. ARC also explored whether it could assign a value to connecting passengers by associating aggregate socio-demographic Cen- sus information (such as the average household income for the passengersâ home zip code). How- ever, the low match rate with passenger names and origin airports made it difficult for ARC to identify large differences in customer demographics across connecting airports (Howard 2015). Other airports have explored whether they can use location-based devices, such as beacons, to assist in airport operations planning and concession planning. Some airports, such as London Gatwick, have installed low-cost beacons throughout their concourses (McCartney 2015). These beacons allow the airports to passively collect precise information about the location of indi- viduals who have cell phones with Bluetooth-enabled technologies, that is, beacons provide complete journey information for a sample of passengers [i.e., when did the passenger enter the
Other Approaches to the Use of Disaggregated Socioeconomic Data in Air Passenger Demand Analysis 103 airport, when did the passenger enter (and leave) security, which security line did the passen- ger use, which stores did the passenger pass by or visit, which restrooms did the passenger use, how long was the passenger in the gate area, when did the passenger board the aircraft, etc.]. This precise location information is anonymous, meaning no information other than the loca- tion about the passenger is known. (Cosmas and Wollersheim, 2015). However, individuals who choose to identify themselves or provide details about their trip can receive âcustomized GPS-like directions to their gate,â as in the London Gatwick application (McCartney 2015). Nonetheless, this information can be used to help airports better understand their operations (such as where and when queues form) and plan the layout of their concessions and passenger facilities. The precise location of this technology also allows airports to push surveys to customers, e.g., to survey passengers boarding a particular flight when they are in the boarding area (Cosmas and Wollersheim 2015). Many airports, such as Akron-Canton, use social media (including Twitter, Facebook, LinkedIn, etc.) to connect with their passengers. Through parsing unstructured texts, these data feeds can help airports identify general sentiments of passengers towards an airport, e.g., do passengers like the concession offerings? Social medial also allows airports to more directly interact with the local community and better understand who are their most vocal advocates. This can be particularly helpful during the airport expansion planning process (Cosmas and Wollersheim 2015). License plate information has been used to identify the catchment area for an airport. For example, Frankfurt Airport uses the license plates of individuals parking at the airport to model its catchment area. This is possible because German license plates use one to three letters to link the vehicle to the county in which the vehicle is registered. The license plate data can then be used to predict catchment areas, which can be validated by results from airport surveys of its customers. This approach can also be used in airports in other countries that have a way to associate license plates to geographic areas, such as the United States on a state level (Cosmas and Wollersheim 2015). Similar to the use of beacons in airports, customers who elect to pro- vide more information can receive more personalized services. For example, passengers who provide their flight itinerary information at the time they park in DÃ¼sseldorf airport can have a robot park their car and return it to a designated pickup area after they land (McCartney 2015). Web search data can also be used to determine which individuals are searching for multiple airports. Many airlines are using these data to better understand airport choice in multi-airport regions (Hotle and Garrow 2014). Many airports have duty free stores that scan boarding passes of passengers making purchases. These data are often used to model customer purchase behavior as a function of the customerâs destination, flight number, nationality, and gender (Cosmas and Wollersheim, 2015). Some examples of how disaggregated data on traveler behavior has been used in the air- line industry include revenue-generating applications (e.g., through concession planning and/or targeted marketing) and cost-reduction applications (e.g., through improving opera- tional efficiencies and better allocation of airport staffing levels). Passively collecting data pro- vides the advantage of collecting âmoreâ and potentially âmore completeâ information about travelers, but it is not perfect. The need to protect individualâs privacy often prevents the ability to directly link disaggregated socioeconomic information to individual-level transaction data. Spotlight: A Study of How Credit Transaction Data Can Be Used to Model Air Passenger Behavior As part of the current project, we conducted an in-depth study to see how credit transac- tion data could be used to model air passenger behavior. Our analysis (detailed in Appendix E)
104 Using Disaggregated Socioeconomic Data in Air Passenger Demand Studies highlights several challenges we faced in re-purposing existing (and de-identified) data for a new application. As part of the analysis, we needed to: â¢ Develop algorithms to determine the home zip code for the credit card holder; â¢ Use text parsing to identify which transactions involved the purchase of an airline ticket; and â¢ Use the sequence of transactions to identify when an individual left their home zip code and (likely) traveled by air to a different destination. We conducted a case study of the Los Angeles area that included all zip codes located within 100 miles of five airports in the Los Angeles area: Burbank (BUR), John Wayne (SNA), Long Beach (LGB), Los Angeles (LAX), and Ontario (ONT). Based on a sample of 50 households in this study area, we determined how often the trip characteristics shown in Table 41 could be identified. As seen in the table, information about the airline, price, group size, and booking data was easily identified from the data, but information on the travel dates, where the indi- vidual actually went, and what airports the individual used were much more difficult to infer from the transactional data. We also discovered that the distribution of air travel party size for the air trips in the financial transaction database is quite different from that found in recent air passenger surveys. As shown in Table 42, the financial transaction database contains a much larger percentage of air trips with two more travelers than the air passenger surveys. Based on the analysis of the financial transaction database, we conclude that financial trans- action data show promise for being used in the future, but currently lack some critical infor- mation (most notably the ability to consistently identify the airports used). Further research is needed to identify how complete and representative these financial transaction databases are. It can be expected that, if the market penetration and consumer acceptance of using online financial tools grows, so too will the value of the collected financial transaction data for plan- ning applications. Trip Characteristic Percent Identified Airline 100 Price 97 Group size 99 Booking date 95 Destination 38 Outbound departure date 26 Outbound departure date range 39 Inbound return date 30 Inbound departure date range 40 Outbound departure airport (or inbound arrival airport) 6 Outbound arrival airport (or inbound departure airport) 3 Table 41. How often trip characteristics were identified in sample of financial data. Group Size Financial Transaction Data SFO 2014/15 Air Passenger Survey1 MWCOG 2013 Air Passenger Survey2 1 4.2% 53.7% 51.5% 2 74.1% 32.4% 35.6% 3+ 21.7% 13.9% 12.9% Average 2.46 pax 1.71 pax 1.71 pax 1 Survey responses for U.S. residents making personal trips (domestic and international) 2 Survey responses for MWCOG regional residents making personal trips (domestic and international) Table 42. Comparison of group size distributions.
Other Approaches to the Use of Disaggregated Socioeconomic Data in Air Passenger Demand Analysis 105 Approaches to Air Passenger Demand Model Specification The case study analysis in the previous chapter has demonstrated one approach to incorpo- rating disaggregated socioeconomic variables into air passenger demand models, albeit in a fairly simple way. However, the research team has identified four different ways disaggregated socio- economic data could be incorporated into air passenger demand models: â¢ Use of variables that reflect the shape of the distribution of the explanatory variable in a single relationship in addition to variables that reflect the aggregate or average value of the explana- tory variable (the approach used in the models described in the previous chapter); â¢ Use of separate variables for different ranges of the explanatory variable in a single relationship; â¢ Use of separate relationships for different ranges of each explanatory variable; and â¢ Use of a simulation approach that generates a measure of trip propensity for individuals with specific values of the explanatory variables. Alternative Approaches Use of Variables Reflecting the Shape of the Distribution of Explanatory Variables This approach might include, for example, variables for the 20th and 80th percentiles of the household income distribution in addition to the average household income. Since air travel propensity varies with income, the resulting demand model could be expected to indicate how air travel demand is sensitive to changes in the distribution of household incomes as well as the average or aggregate income. For models where the explanatory variables are multiplied together (such as the common log-linear model), the variables for the distribution percentiles should be expressed as ratios of the average income rather than in monetary terms to avoid having two income values multiplied together. Although in principle this approach could be applied to any socioeconomic variable, it makes most sense to apply it to the income variable, since it is known that income distributions have been changing over time, that air travel demand varies with household income, and data on household income distributions for each year are readily available from the U.S. Census Bureau. The U.S. Census Bureau report Income and Poverty in the United States: 2015 (Proctor, Semega, and Kollar 2016) includes a table showing the household income distribution for each year from 1967 to 2015. The data show that households with an income of $200,000 or more in constant 2015 dollars increased from 2.1% of all households in 1985 to 5.1% of all households in 2005 and 6.1% of all households in 2015. Conversely, households with an income of $25,000 per year or less declined from 24.6% of all households in 1985 to 22.0% of all households in 2005 and 22.1% of all households in 2015. Perhaps of greater relevance for air travel demand, the percent of households with incomes between $50,000 and $100,000 declined from 33.3% of all households in 1985 to 30.4% of all households in 2005 and 28.8% of all households in 2015. Although the annual changes in these percentages are not large and the trend has been fairly stable over the last 30 years, some of the increase in air travel over that time period is clearly due to the changing income distribution (a higher percentage of households in the higher income brackets that have a higher air travel propensity) rather than the change in the average household income. This showed greater year-to-year variability, increasing from $61,049 in 1985 to $76,878 in 2005 (a 25.8% increase) and to $79,263 in 2015 (3.2% increase from 2005), but declining in some years, particularly during recessions. In addition to income, there is the question of whether to include distributional variables for other demographic or socioeconomic variables in air travel demand models. Analysis of air
106 Using Disaggregated Socioeconomic Data in Air Passenger Demand Studies passenger and household travel surveys has shown that air travel propensity varies by age, gen- der, race/ethnicity, and educational attainment, as well as income, although some of these effects may be correlated. For example, the analysis of demographic trends presented in Chapter 3 shows that the median income of households with a head of household age 65 or older has grown faster than all other households. The extent to which changes in the distributions of these characteristics may help explain changes in the demand for air travel can be addressed in air pas- senger demand model development by examining the correlation between model residuals and variables measuring the changes in the distributions of these characteristics. One potential challenge to the application of this approach that needs to be explored in model development work is the extent of correlation between the distributional and aggregate vari- ables, which would make it difficult or impossible to obtain statistically significant coefficients for both types of variables in the same model. Both types of variables are slow-moving and gener- ally change in a consistent direction, so are quite likely to be highly correlated. Use of Separate Variables for Different Ranges of Explanatory Variables This approach provides greater flexibility in reflecting the effect of the underlying distribu- tion of particular socioeconomic variables, such as income or air traveler age, by defining sepa- rate variables for different income or age ranges. However, this imposes two limitations on the analysis. The first is a more complex functional form, since the variables for each value range for a given factor cannot simply be multiplied together. Rather, terms for the demand generated by each subset of the total population formed by the various value ranges (each of which may well have a multiplicative form) must be added together to give the total demand. For example, a simple model using total household income (H) and average airfare (P) might take the following form: ) ) )( ( (= Ã Ã + Ã Ã + + Ã ÃPax a1 a2 . . . anb1 c1 b2 c2 bn cnH1 P H2 P Hn P where H1, H2, . . . , Hn represent the total household income for different income ranges. This formulation allows for a different income and airfare elasticity for each income range, although the model structure could be simplified by restricting each income range to have the same airfare elasticity (i.e., c1 = c2 = . . . = cn = c) or even the same income elasticity (i.e., b1 = b2 = . . . = bn = b) although it would be surprising if in fact air travel demand for households in different income ranges had the same sensitivity to changes in income. Nonetheless these are research issues that could be explored. The second limitation is the need for a much larger number of data points for estimating the larger number of model coefficients implied by the approach. This suggests that this approach would be more appropriate to use with a large panel dataset, such as the demand in multiple O&D markets or for a single model that is estimated across a number of airports or regions. The resulting model segmentation is also likely to result in model functional forms (such as the one shown above) the coefficients of which cannot easily be estimated using standard linear regression estimation techniques. Therefore, research is needed to identify appropriate statisti- cal techniques to estimate such model functional forms, such as the use of maximum likelihood estimation or nonlinear regression. Use of Separate Relationships for Different Ranges of Each Explanatory Variable This approach avoids some of the technical challenges of the second approach by estimating separate models for different subsets of the population. Although this is likely to require a less
Other Approaches to the Use of Disaggregated Socioeconomic Data in Air Passenger Demand Analysis 107 complex functional form for each model, it requires an estimate of the demand generated by each subset of the population. Since there is no way to obtain this from the reported air passen- ger traffic, it requires air passenger survey data to segment the reported traffic into the air trips made by each subset of the population. However, segmenting the models in this way not only simplifies the model estimation process (since the model for each segment can have the same functional form as a model developed using aggregate data), but allows a much finer definition of the population subsets than would be practical with the second approach. The challenge with applying this approach is that most airports have limited air passenger survey data (or none at all) that can be used to estimate the proportion of the total air passenger traffic in each subset of the population. Even those airports that have undertaken several air pas- senger surveys over time typically do not have survey data for every year. In these cases it would be necessary to interpolate the proportions between the years for which survey data is available. To assess the likely validity of this approach, it would be helpful to undertake some analysis of how much these proportions vary from year to year (at least for those years for which survey data is available). For airports where air passenger survey data is only available for one year, the proportions for other years can be estimated by applying the trends from airports for which multiple surveys are available. This is obviously less satisfactory than having survey data for multiple years, but may still be better than ignoring the differences between subsets of the population entirely. One potential approach, although one that would require a significant amount of data analy- sis, would be to examine trends over time in the spending on air travel and number of air trips by households in the different subsets of the population, using data from the CES, which is available on an annual basis. Although the sample size of the CES is not large enough to obtain reliable data for a given geographical area served by a specific airport or regional airport system, it is large enough to allow some segmentation between major metropolitan areas and smaller communities in general, as well as differences between different regions of the United States (defined broadly). Use of Simulation to Generate Estimates of Air Travel from Trip Propensity Data This approach takes a much more disaggregate approach to forecasting air travel demand based on air travel propensity relationships identified through analysis of air passenger survey data rather than a conventional econometric approach that attempts to fit a model to observed data using regression or other techniques. Although this approach requires a much greater amount of data than traditional econometric techniques that are based on reported air pas- senger traffic and fairly aggregate measures of potential causal variables, it avoids the inherent constraints imposed by the functional form of any given econometric model and provides much greater flexibility to vary the assumptions used for future values of the causal variables. However, while past air travel propensity relationships can be determined from survey data, these relationships do not explicitly consider the effect of changes in pricing. Any practical application of this approach must also include a way to incorporate changes in airfares and other travel costs as well as changes in air travel propensity that result from trends in the level and distribution of socioeconomic variables. The effect on air travel propensity of changes in airfares and other air travel costs can be assessed by applying estimates of the price elasticity of air travel demand. This has been exten- sively studied for airfares and estimates exist for different types of air trip, as discussed in the literature review documented in Chapter 2. Generally, it has been found that the airfare price elasticity for business travel is somewhat less than â1 in an absolute sense (i.e., air travel declines
108 Using Disaggregated Socioeconomic Data in Air Passenger Demand Studies somewhat less than proportionately to an increase in cost) while airfare price elasticity for non- business travel is somewhat greater than â1 in an absolute sense, reflecting that nonbusiness travel is generally more discretionary than business travel. If airfares decline in real terms, house- holds choose to make more air trips at the expense of other consumption, and vice versa if real airfares rise. Although the exact airfare price elasticity for households with a given set of characteristics will not be known to any degree of precision, using an approximate elasticity value will correct for much of the effect of airfare changes and any errors this introduces can be corrected by calibrating the results to actual air passenger traffic levels. Where air passenger survey data is available for surveys undertaken at the same airport over a period of time (as is the case for several of the airports or regions studied in the current proj- ect), trends in air travel propensities for survey respondents with a given set of socioeconomic characteristics can be determined and applied to years for which survey data is not available. If these changes in trip propensity can be shown to be reasonably consistent across different air- ports, even if the actual trip propensity values differ, then they could be applied to airports for which survey data is only available for one year, or even to airports for which no air passenger survey data is available, by adjusting the resulting estimates of total air passenger demand to conform to the actual air passenger traffic levels. Given estimates of air travel propensity for households with any given set of household char- acteristics for a given year, taking into account the effect of changes in airfares and travel costs, the total number of air trips at an airport or for a region in a given year can be estimated by generating a synthetic sample of households with appropriate characteristics from the regional distributions of household characteristics and then simulating the number of air trips that each of these households would be expected to take in the year. The resulting projected air trips can then be calibrated to the actual passenger traffic for past years and the resulting calibration fac- tors used to forecast future air travel, based on scenarios for future trends in socioeconomic characteristics and future trends in airfares and other travel costs. Implementation Considerations Business and Personal Travel The analysis of air passenger and household travel surveys has shown, not surprisingly, that the distributions of household characteristics of those making business and personal trips are significantly different. More important from the perspective of air passenger demand model- ing, the factors influencing the demand for business and personal air travel are also likely to be different. In reality, it is not households that generate business trips but businesses (although of course those making business trips are members of households). Therefore the level of busi- ness trips at an airport is likely to depend on the composition and size of the local economy, as well as other factors unrelated or only indirectly related to the distributions of household characteristics in the region served by the airport. This suggests that one dimension of disaggregation in air passenger demand modeling would be to distinguish between business and personal trips. This could be fairly easily accomplished by the last three approaches discussed earlier, but would be more difficult to address using the first approach discussed, although one such approach has been demonstrated in the detailed analysis of the Baltimore-Washington region described in Chapter 4. The underlying concept in each of the latter three approaches is to add terms to the demand function to cover business trips or (in the case of the third approach) simulate personal and business trips separately. Of course, this requires knowing the split between business and personal travel, but this can be obtained from
Other Approaches to the Use of Disaggregated Socioeconomic Data in Air Passenger Demand Analysis 109 air passenger or household travel survey data in exactly the same way as estimating the propor- tion of trips by households with different characteristics. Although the simulation of business travel could be based on the number of households with given characteristics in the region and the business trip propensity by household, it would be more logical to base the simulation of business trips on employment by sector. This would allow forecasts of future air passenger demand to reflect projected or assumed changes in the growth of employment by sector. Similarly, in the second and third approach, the demand function terms for business travel could use variables reflecting the composition of the local economy and employment levels rather than household characteristics. Developing estimates of business air trip propensity relative to sectoral employment would require analysis of air passenger or household travel survey data, or other data sources, to iden- tify the distribution of business trips by economic sector. Unfortunately, few air passenger sur- veys ask about the type of firm or organization that business travelers are employed in and given the large number of economic sectors and the likely variation in business travel propensity per employee, obtaining this information in sufficient detail from air passenger or travel sur- veys would be very unwieldy. Furthermore, including sufficient terms in a single air passenger demand model to address many different sectors would result in an excessively complex func- tional form for which it would be extremely difficult, if not impossible, to obtain statistically significant coefficient estimates. A more practical approach would be to use total employment and an estimate of average business air trip propensity obtained from a separate analysis of business air trip propensity in each sector and the sectoral composition of the local economy. This is still sensitive to changes over time in the sectoral composition and can reflect anticipated future changes. Since estimat- ing business air trip propensity by sector from air passenger or travel survey data is problematic, it may be more productive to analyze differences in business spending on air travel by sector. There is a considerable amount of data on business travel from a wide range of sources (see https://www.creditdonkey.com/business-travel-statistics.html for examples) that could provide estimates of trends in business travel expenditures and average costs per trip (which would allow expenditures to be converted to trips). The 2007 input-output model of the U.S. economy avail- able on the website of the U.S. Bureau of Economic Analysis provides a detailed breakdown of spending on air transportation by economic sector. Although these data are only for one point in time, the relative business air trip propensities across different economic sectors are likely to be fairly stable, since they are largely determined by the types of activities undertaken by employees in firms or other organizations in each sector. Therefore these data can be combined with trend data from other sources to obtain estimates of business air travel propensity by sector for other years. Air Trips by Residents and Visitors Analysis of air passenger survey data has also shown that there are differences in the composi- tion of air passenger trips at a given airport between those made by residents of the region served by the airport and visitors to the region. Although the approaches described earlier make sense for air trips generated by residents of a region, it is less clear that they apply to air trips by visitors. At a minimum, the distributions of household characteristics for visitors are likely to be differ- ent from those for residents, and in any case the air trips to a region made by visitors are not the only air trips that those people made. There may be a degree of symmetry for some types of trip between trips made by residents and those made by visitors. For example, trips by residents to visit family and friends elsewhere may be balanced by trips by visitors to visit family and friends in the region. However, for other types of trip, such as those for vacations, attending college, or medical treatment, there is no reason to expect that the levels of air trips by visitors are likely to be similar or proportional to those by residents.
110 Using Disaggregated Socioeconomic Data in Air Passenger Demand Studies Data from air passenger surveys performed at the same airport over time can provide some indication of whether the composition of the different types of trip appears to be fairly stable over time. The U.S. DOT 10% O&D data can be analyzed to show the extent of the variability in the split of air trips at a given airport between residents and visitors over time as well as any over- all trends. Since these data are available on an annual basis, they can be used to divide the total air passenger traffic at an airport into three components: outbound trips by residents, inbound trips by visitors, and connecting passengers. Since the latter are largely a consequence of airline network and hubbing strategies rather than the demographic and socioeconomic composition of the region where the hub airport is located, forecasting connecting traffic requires a different approach from that for O&D traffic that is beyond the scope of the current project. Although all four approaches to incorporating disaggregated socioeconomic data in air pas- senger demand models can be applied as well to modeling visitor trips as resident trips, the variables used may well be different, particularly for personal trips by visitors for purposes other than visiting friends and family. In using the 10% O&D data to separate total traffic into the three directional components, care should be taken to make adjustments in the analysis for one-way trips. While these undoubtedly do include some genuine one-way trips, most are an artifact of travelers purchas- ing two separate tickets, typically on different airlines to obtain a lower overall round-trip fare or more convenient flight schedules. Even for those that are genuine one-way air trips, it is not clear whether the traveler is a resident of or visitor to the region containing the first airport in the itinerary. It is therefore reasonable to assume that the proportion of one-way trips in each direction that are outbound trips by residents of the region containing the first airport in the itinerary reflects the directional split given by the round trip itineraries in the data. Where appropriate questions have been included in air passenger surveys to identify respondents who were making a one-way air trip, these data can be used to estimate the true proportion of actual one-way trips in each direction.