An Analysis of Transportation Demand in : How Much Will Atlanta’s Proposed MARTA Expansions Increase Ridership?

By SAM TUCKER*

Transportation in Atlanta is considered poor by national standards. I look at the MARTA light rail expansion proposals from the Atlanta City Government. I use survey data from the 2011 Atlanta Regional Commission’s Household Travel Survey to estimate travel demand. Since explicit coordinates are not given for each survey participant, I use Traffic Analysis Zones as proxies to estimate distances. A discrete choice multinomal logit model is then estimated by maximum likelihood estimation. Change in predicted ridership is then calculated by adding the proposed train stations to the choice set. Clifton and Campbellton are predicted to be the most successful lines, and then I examine the demographics of those who benefit the most from the MARTA expansion.

Public transportation is an important duty of city governments, and it is becoming more important than ever with the environmental and traffic concerns of the 21st century. The need for improvements to public transportation could not be overstated in Atlanta. In 2017, Atlanta ranked in the top ten out of 1,360 cities worldwide for most time drivers spent in congestion, according to INRIX’s Global Traffic Scorecard. In addition, Atlanta ranked in the bottom ten of eligible American metropolitan areas in access to transit, according to a Brookings Institution report in 2011. Also, the Federal Transit Administration has found that MARTA’s passenger trips decreased by 2.6 percent in 2017 (Green, 2018). Attempting to remedy the dire state of public transportation in Atlanta, in 2016 the city approved a half-cent sales tax that will raise $2.5 billion over the next 40 years (Freemark, 2017). However, the list of all transportation improvements for eligible funding covers $10.3 billion in capital projects operations and maintenance. Many projects were considered initially. However, according to the Atlanta Journal-Constitution, as of May 2018 the list of light rail projects being considered has been narrowed to six: Campbellton Line (C), Clifton Corridor (D), Beltline-Loop Northeast (A), Beltline-Loop Southwest (B), Crosstown Downtown West Extension (F), and the Crosstown Downtown East Expansion (E).1 Some other projects are being considered, such as bus rapid transit and arterial rapid transit, but I will be focusing my analysis on the light rail expansion projects. The goal is to understand what drives demand for public transportation, and how much each of the expansions will increase MARTA ridership. To answer this question, I will use the multinomial logit framework laid out by McFadden, et al. (1977) and expanded upon by Train (2009) and apply it to the survey data from Atlanta Regional Commission (2011). The survey

* This paper was submitted as a graduation requirement to the Department of Economics, Terry College of Business, University of Georgia, Athens, GA (email: [email protected]). Thanks to Professor Josh Kinsler for advice and guidance. 1 A map of the proposed new lines can be accessed at https://www.ajc.com/news/local-govt--politics/local-update- marta-plan-boosts-atlanta--cuts-funding-for-emory-line/ILPofMcJKbAw0e3KNBBIUI/ AUGUST 2019 THE UGA JOURNAL OF ECONOMICS 2 data contains demographic information, information on home and work location, and transportation mode of choice for Atlanta commuters in 2011. This model will estimate parameters so that I can determine which factors are relevant for demand of transportation. Afterwards, I plot the coordinates for the proposed stations and add them to the choice set, to see which stations give the highest predicted increase in ridership. I find that the Clifton Corridor (A), and Campbellton Line (C) are predicted to be the most successful lines.

I. Literature Review

Billings (2011) looks at light rail transit (LRT) in Charlotte, North Carolina. He notes that economists seem to be puzzled by the popularity of LRT expansion in U.S. cities, due to their limited time savings over bus transit and their high costs. However, urban planners notice that LRT usually commands increased ridership over bus. Billings (2011) claims that increased utility of living near an LRT station should be capitalized in housing values. Previous literature has identified four ways through which rail transit affects housing values: 1) the direct distance between a residential property and a transit station; 2) the positive effect of transit stations on commercial development; 3) increased parking congestion near transit stations; and 4) greater opportunity for crime with improved access to neighborhoods with LRT stations. The author also notes that areas with LRT stations are usually targeted by cities for future investment to improve land use. Using a differences-in-differences estimator, he finds positive capitalization in neighborhoods with LRT, specifically, 4.0 percent for single-family properties, 11.3 percent for condominiums. However, no capitalization is found for commercial properties. This evidence suggests that LRT may be a better investment in economic development for certain neighborhoods rather than specifically as a transportation option. In conclusion, Billings (2011) estimates a neighborhood benefit of $97.2 million due to LRT expansion in Charlotte. However, costs for the new South Line were $450 million. He notes that the full benefits may not be captured in his estimates, because pollution and traffic congestion reduction as well as other unobserved factors are not captured in his estimates. Baum-Snow et al. (2005) notes that while 16 American cities have spent a combined $25 billion on rail transit expansion between 1970 and 2000, the number of Americans taking public transportation has decreased by half during that time period. However, ridership increased in areas within two kilometers of a new line and beyond ten kilometers of the city center. The issue is that since cities are becoming decentralized, a smaller fraction of the population now has access to the existing stations. Atlanta is one of the cities in the analysis, and while only five percent of people were taking public transit in 2000, the authors estimate that 11 percent would take public transit in Atlanta if the population were at its 1970 spatial distribution. Also, thirteen percent of commuters lived within two kilometers of a transit line in 2000, a two percent increase from 1990. Their model of transit use insists that those who live close to the center of the city will tend to walk or take a bus a short distance to the rail line and those who live further out will drive to the line. This is useful information, because I will define costs of transit options in my model assuming a cost of driving to the nearest station. The authors present several arguments that could justify large public investments in public transit. Since rail transit exhibits increasing returns, it may be optimal to subsidize transit up to the point where the average social cost of taking public transit is less than the average social cost of driving. Also, negative externalities caused by driving, such as congestion and pollution, could be mitigated by more people taking public transit. They estimate that a little over 19,000 people switched to public transit between VOL. 1 NO. 2 Tucker: Transportation Demand in Atlanta 3

1980-2000, with a daily savings of 31,100 hours and $624,000. In conclusion, they find that there are significant increases in ridership in areas greater than ten kilometers from the city center. Most of the effects close to the city center are found in bus riders switching to rail. While this does not increase share of the population taking public transit, it does have value in reducing commute times. II. Data To estimate demand I use data from the Atlanta Regional Commission’s 2011 Household Activity Travel Survey. Here is how the ARC describes the survey: The purpose of the survey was to update ARC’s travel demand model and get a better understanding of travel behavior in a 20 County Region. Households were randomly selected to participate in the survey. Those agreeing to participate were assigned a one-day travel period and asked to track of all trips during that period. Information was collected from 10,278 households. When weighted to reflect the population of the region, the households represent 25,810 persons, 93,713 trips and 21,270 vehicles. (ARC, 2011) The sample size for my analysis is 5,077 due to incomplete information in the survey. I do not believe that incompletes will bias work trip estimates, because most of the incompletes are either children under eighteen or senior citizens who are likely retired. What is important for my analysis is the location of each person’s house, their place of work, and their transportation mode of choice. Table 1 shows the demographics of my sample divided by transportation modes of choice. Table 1: Demographics by Mode of Transportation to Work

Variable Train Bus Car % of Sample 1.65 1.2 97.15 % with Income over $75,000 46.4 37.7 61.1 % with Bachelors Degree or higher 61.9 45.9 54.4 Average Age 42.54 43.67 45.55 Average Household Size 2.35 2.72 2.95 % Owning Home 73. 68.8 88.9 % with Employer Provided Parking 37 57 93 % with Employer Subsidized Transit 51 37 11 Average Distance to Work (miles) 12.87 18.04 11.29 Average Distance to nearest Train Station from Home (miles) 4.62 11.54 14.28 Average Distance to nearest Train Station from Work (miles) 1.16 1.84 10.37 Average Distance to nearest Bus Station from Home (miles) 1.95 7.41 9.33 Average Distance to nearest Bus Station from Work (miles 0.48 0.59 6.41 Average Estimated Daily Cost of Travel $4.14 $4.52 $3.81

I limited modes of choice to bus, train, or car. Some people walk or bike to work, but their shares are small enough that they can be excluded. In this sample we see that roughly 1.6 percent of people take the train to work and 1.2 percent take the bus to work. In all, about 2.8 percent of people take public transportation. This number is close to an estimate by the Atlanta Regional Commission in 2013 of about 3.1 percent. AUGUST 2019 THE UGA JOURNAL OF ECONOMICS 4

The main obstacle is that coordinate information is not given for each person’s home and work, information that was likely left out to preserve the privacy of the survey participants. Fortunately, the Traffic Analysis Zones (TAZ) for each person’s house and place of work were included. These TAZs are the units of geography most commonly used in conventional transportation planning models. In general, each individual zone must contain somewhere between 600-1200 people, should be compact in nature, and must not cross county lines. Specifically, for the ARC TAZ delineation strategy, these physical barriers to travel were examined as TAZ boundaries: lakes, rivers, streams, railroad tracks, and utility corridors. I acquired coordinates for current bus and train stations from MARTA’s website. Coordinates for stations for the potential new lines are placed directly if specific information is available. Otherwise, I placed stations based on mean distance between current MARTA stations. III. Model Specification The behavioral approach to travel demand estimation assumes that the behaviors of a given sample are representative and will emulate the underlying patterns of human conduct. In order to estimate the travel demand model, I follow McFadden, et al. (1977) by using a multinomial logit framework. According McFadden, et al., The MNL model can be derived from the theory of individual choice behavior by assuming that individual utility deviations from mean utility in a homogeneous market segment are statistically independent for different alternatives and have a probability distribution called the extreme value distribution. Train (2009) expands on this work, defining utility for decision-maker n from alternative j as

1)푈푛푗 = 푉푛푗 + 휖푛푗

푉푛푗 is defined as representative utility, which is a function of some characteristics of the alternatives and some characteristics of the decision maker. By assumption, the error term 휖푛푗 is independent and identically distributed. This means that unobserved factors are uncorrelated over alternatives and have equal variance. It is also important to note that utility is constructed as a measure of well-being and is ordinal. In other words, the absolute level of utility is irrelevant. Three conditions must be met for a discrete choice model to be valid: 1) the alternatives must be mutually exclusive; 2) the choice set must be exhaustive; and 3) the number of alternatives must be finite. Also, the decision maker is designed to be utility maximizing. Since equation (1) contains unobserved factors, the researcher does not observe utility. 푉푛푗 is defined for three alternatives in my paper: train, bus, and car. Mean utility can vary through the following individual specific characteristics: gender, age, household income, education level, and dummies for employer provided parking, and employer subsidized transit. Household Income is listed in discrete categories starting from under $10,000 to greater than $150,000. Finally, education level ranges from high-school dropouts to graduate degree holders. For alternative-specific variables, I use distances to the closest train and bus station from each person’s home and work and travel costs. 푉푛푗 will contain parameter estimates for both bus and train. Representative utility of car is set to 0, since only differences in utility matter. It is important to note that a multinomial logit framework is only valid if the ratio of probabilities of two alternatives are independent of any other alternatives. This is called VOL. 1 NO. 2 Tucker: Transportation Demand in Atlanta 5

Independence from Irrelevant Alternatives, or the IIA property. This property is not appropriate in some circumstances, for example when a blue bus is added to the current choice set of red bus or car. Under IIA, any positive probability of taking the blue bus must be subtracted equally from red bus and car. This clearly makes no sense, as you would expect that the entrance of a different color bus would have no effect on probability of taking the car. McFadden et al. (1977) introduces a method to test this property: removing an alternative from the choice set to see if the coefficients remain the same. If they differ, then IIA could be violated. I use this method to test the IIA assumption later. In contrast to linear models, the coefficients cannot be considered as the marginal effects of the explanatory variables on the dependent variable. However, the coefficients can be interpreted through various transformations. For alternative-specific variables the sign of the coefficient is directly interpretable. On the other hand, the sign is not necessarily interpretable for individual specific coefficients. The sign of the marginal effect can only be established with certainty for alternatives with the lowest and greatest coefficients. The sign of the marginal effect is given by (훽푗 − ∑ 푗 푃푖푙훽푙), which is positive if the coefficient for the jth alternative is greater than the weighted average of the coefficients for every alternative. While interpreting the sign is useful, coefficients are marginal utilities; therefore, their marginal effect cannot be interpreted since utility is ordinal. Another property of logit discrete choice models is that the ratio of coefficients have meaning. They can be used to understand marginal rates of substitution. Croissant et al. (2012) gives an example where if 훽푙 is the coefficient for transport time in hours and 훽2is the coefficient for transport cost then the ratio can be interpreted as the amount that a person would be willing to pay to reduce travel time by an hour. I use this method to estimate a willingness to pay for a reduction in distance to the nearest bus or train station, since time is not available. Finally, this equation gives the logit choice probability for alternative i. I use this information to see how probabilities change when new train stations are added to the model

휖푉푛푖 2) 푃푛푖 = 푉 ∑ 푗휖 푛푗

I modeled straight-line distances between a person’s home and work as the distance between the coordinates of the center of the two zones. I plotted the nearest train stations to home and work and the nearest bus stations to home and work, yielding the minimum distance from the center of the zone to a bus or train station. Cost for car is defined as two times the dollar per mile times the straight-line distance between home and work in miles. Dollar per mile is calculated using the average price of a gallon over the sample period in the Lower Atlantic Region, which was $3.46, and the average miles per gallon of cars on the road in 2011, which was 21.4 according to the U.S. Energy Information Administration and U.S. Bureau of Transportation Statistics. Cost of train and bus trips are defined to be $3.17 (average cost per day for 90-day MARTA holders) plus two times the cost of driving from home to the nearest train or bus stations. Unfortunately, there is no way to know how those who commuted by bus or train arrived at the station. However, if they are close enough to bike or walk then the cost associated with driving will be minuscule, and it will more accurately reflect the true cost for those that live far from a bus or train station.

AUGUST 2019 THE UGA JOURNAL OF ECONOMICS 6

IV. Model Output and Interpretation Table 2 shows the coefficient estimates from maximum likelihood estimation.

Table 2: Maximum Likelihood Estimation Output

Explanatory Variable α Coefficient t-Statistic Bus intercept (2) 0.169 0.215 Train intercept (3) 1.670 2.266* Combined distance of nearest bus station to home and work (2) -0.039 -2.231* Combined distance of nearest train station to home and work (3) -0.165 -6.963*** Daily cost of travel ($0-$40,000) (2-3) -0.107 -1.309 Daily cost of travel ($40,000-$60,000) (2-3) -0.310 -3.327*** Daily cost of travel ($60,000-$75,000) (2-3) -0.531 -4.336*** Daily cost of travel ($75,000-$100,000) (2-3) -0.330 -3.907*** Daily cost of travel ($100,000-$150,000) (2-3) -0.362 -3.398 *** Daily cost of travel (> $150,000) (2-3) -0.245 -1.674 Income ($40,000-$60,000) (2) -0.611 -1.483 Income ($40,000-$60,000) (3) -0.016 -0.037 Income ($60,000-$75,000) (2) -0.734 -1.392 Income ($60,000-$75,000) (3) 0.819 1.763 Income ($75,000-$100,000) (2) -1.176 -2.755** Income ($75,000-$100,000) (3) -0.775 -1.666 Income ($100,000-$150,000) (2) -1.773 -3.430*** Income ($100,000-$150,000) (3) -0.803 -1.720 Income (>$150,000) (2) -1.921 -2.834** Income (>$150,000) (3) -0.711 -1.348 Gender (2) -0.108 -0.388 Gender (3) -0.122 -0.481 Age (2) -0.016 -1.371 Age (3) -0.026 -2.424* Employer Provided Parking (2) -1.983 -6.743*** Employer Provided Parking (3) -2.359 -9.050*** Employer Subsidized Transit (2) 1.360 4.542*** Employer Subsidized Transit (3) 1.572 6.060*** Education (2) -0.249 -2.522* Education (3) -0.152 -1.619

α Variable enters modes in parentheses and is zero in other modes. Modes: 1.Auto alone, 2.Bus with auto access, 3.Train with auto access. *** Significant at 0.001, ** Significant at 0.01, *Significant at 0.05. Daily cost of travel is cost of travel for bus or train minus car costs.

After using the formula 훽푗 − ∑ 푗 푃푖푙훽푙 from Croissant et al. (2012), the signs of all individual specific coefficients in the table can be interpreted as shown. We can see that distance to the closest train station negatively impacts a person’s utility of taking the train much more than a distance to the nearest bus station. This reinforces the hypothesis that MARTA rail needs to be more accessible in order to increase ridership. As shown in Table 3, the 4 highest income VOL. 1 NO. 2 Tucker: Transportation Demand in Atlanta 7

brackets make up 70.15 percent of the total sample. From here, it looks like what we would expect. Households in the $60,000-$75,000 range are more concerned with cost of travel than those in the $75,000-$100,000 range and those in the $100,000-$150,000 range. As you would expect, those in the > $150,000 range are the least concerned about cost of travel out of the top four household income brackets. We also see that high-income people are significantly less likely to take the bus, elderly people are significantly less likely to take the train, and highly educated people are significantly less likely to take the bus. As expected, an individual who works for an employer that provides parking is more likely to drive, and an individual who works for an employer that provides subsidized transit is more likely to take the bus or train. In Table 3, the fourth and fifth columns show each person’s willingness to pay for a one-mile reduction in distance to a station from home or work for train and bus respectively.

Table 3: Statistics by Income Bracket

Household Income # in sample % of sample WTP Train WTP Bus $0-$40,000 718 14.14 $1.54 $0.37 $40,000-$60,000 797 15.70 $0.53 $0.13 $60,000-$75,000 485 9.55 $0.31 $0.07 $75,000-$100,000 1247 24.56 $0.50 $0.12 $100,000-$150,000 1172 23.08 $0.46 $0.11 > $150,000 658 12.96 $0.68 $0.16

V. Comparing Proposed MARTA Lines I examine the six light-rail lines proposed for MARTA expansion. In order to understand how many stations should be in each proposed line, I looked at current MARTA lines and calculated the average difference between stations. The average distance in miles between stations on the Gold, Red, Green, and Blue lines is 1.176, 1.181, 0.668, 0.960. The average of the lines is approximately one station per mile. I attempted to stick to this rule of thumb as often as possible when proposing coordinates for new stations. I then calculate new probabilities for train ridership using new variables for minimum distance and daily cost, given the existence of the additional stations. Percent change in train ridership is then calculated and shown in Table 4.

Table 4: Proposed MARTA Lines

Proposed Line Estimated Estimated Estimated Estimated Length cost # New % (miles) (millions) Stations increase in ridership Campbellton 5 $263.7 5 1.23 Clifton 4 $503.6 4 1.76 BeltLine-Loop NE 3 $174 3 0.46 BeltLine-Loop SW 4 $196.2 4 0.17 Crosstown Downtown W Extension 3 $171.6 4 0.11 Crosstown Downtown E Extension 2 $189.8 3 0.49 All Proposed Stations 21 $1,498 23 3.17 AUGUST 2019 THE UGA JOURNAL OF ECONOMICS 8

Now that I have calculated what the percent change in ridership looks like, I can examine the people who benefit the most from the MARTA expansion. I looked at change in probability of train ridership for everyone in the sample. Table 5(a) shows some statistices for the 140 people whose change in probability of train ridership is more than one standard deviation above the mean. Interestingly, 19.2 percent of the 140 work at or Emory Hospital. Since this is where the Clifton line ends, it could help explain why Clifton is expected to be the most successful line. This is in line with the analysis from Baum-Snow et al. (2005), because Emory is very close to ten kilometers away from downtown Atlanta. Table 5(b) shows that these people would be expected to save 22 cents a day and shows how much closer the distance to the nearest train station is from home and work for these people. One final thing to examine is the IIA hypothesis. Earlier, I mentioned that the IIA property must hold for this model to be valid. I am testing this heuristically by removing bus from the choice set and seeing if the coefficient estimates are wildly different. By comparing Table 6 below to Table 2 we can see that the coefficient estimates are fairly similar, therefore we can assume that IIA is a reasonable assumption for this model. VI. Conclusion It is possible that all these lines will be built, since the total estimated cost is under the $2.5 billion currently raised for transportation expansions. Furthermore, if a line needs to be prioritized, it should be the Clifton line. Even though it is the most expensive, it is estimated to increase MARTA ridership by three times more than the next most effective line. However, this model produces very rough estimates for a few reasons. One, I cannot account for general equilibrium effects. For example, I assumed home and work are fixed for everyone. With new stations, it is likely that people will factor that into future decisions for where to live and work. Secondly, distance is used as a proxy for time. People care more about the time it takes to get to their destination than the distance. Thirdly, selection is definitely at play here. People who currently live closer to train stations are different from those who live far away. Since random assignment is not possible, this problem cannot be possibly addressed. It is still important to note, because it affects the analysis. Even the measure of distance is noisy, since I am assuming people can take the straight-line distance to the nearest station or to work. Obviously, this is not the way people travel. All of that being said, the effects may not seem that large, but metro Atlanta contains almost six million people. A four percent increase in those taking MARTA trains represents thousands of people off the road, which will help alleviate Atlanta’s traffic problems.

REFERENCES

ARC (2011). Household Travel Survey. https://atlantaregional.org/ transportation- mobility/modeling/household-travel-survey/. Baum-Snow, N., M. E. Kahn, and R. Voith (2005). Effects of urban rail transit expansions: Evidence from sixteen cities, 1970-2000 [with comment]. Brookings-Wharton papers on urban affairs, 147–206. Billings, S. B. (2011). Estimating the value of a new transit option. Regional Science and Urban Economics 41 (6), 525–536. VOL. 1 NO. 2 Tucker: Transportation Demand in Atlanta 9

Croissant, Y. et al. (2012). Estimation of multinomial logit models in r: The mlogit packages. R package version 0.2-2. URL: http://cran. r-project. org/web/packages/mlogit/vignettes/mlogit. pdf . Freemark, Y. (2017). Atlanta’s Raising $2.5 Billion to Invest in Transit. Will It Be Money Well- Spent? https://usa.streetsblog.org/2017/06/02/ atlantas-raising-2-5-billion-to-invest-in- transit-will-it-be-money-well-spent/. Green, J. (2018). Feds say MARTA ridership is decreasing (gasp), but why? https:// atlanta.curbed.com/2018/5/30/17410516/marta-atlanta-ridership-decreasing. McFadden, D., A. Talvitie, S. Cosslett, I. Hasan, M. Johnson, F. Reid, and K. Train (1977). Demand model estimation and validation. Urban Travel Demand Forecasting Project, Phase 1. Train, K. E. (2009). Discrete choice methods with simulation. Cambridge university press. Wickert, D. (2018). MARTA’s expansion plans for Atlanta: A detailed look. https://www. myajc.com/blog/commuting/marta-expansion-plans-for-atlanta-detailed-look/ PyuO7EtPFU0N4TZEYqNP6H/.