Trends in Metropolitan Network Circuity

Transcription

1 Trends in Metropolitan Network Circuity David J. Giacomin Luke S. James David M. Levinson Abstract Because people seek to minimize their time and travel distance (or cost) when commuting, the circuity the ratio of network distance traveled to the Euclidean distance between two points plays an intricate role in the metropolitan economy. This paper seeks to measure the circuity of the United States 51 most populated Metropolitan Statistical Areas and identify trends in those circuities over the time period from With many factors playing a role such as suburban development and varying economic trends in metropolitan areas over this timeframe, much is to consider when calculating results. In general, circuity is increasing over time. 1 Introduction This study measures the structure of road networks of 51 of the most populous Metropolitan Statistical Areas 1 in the United States. Circuity is the ratio of network to Euclidean distance, and has previously been estimated about 1.2 times the Euclidean distance for stylized networks [Levinson and El-Geneidy, 2009, Newell, 1980]. In a Metropolitan Statistical Area, more area is covered than just the urban sectors, and networks are not uniform. Interest in the implications of network circuity is nothing new. [Goldman, 1972] realizes its importance to the placement of centralized facilities. Ballou et al. [2002] identifies differences various countries have in road network circuity, examining factors such as network density. Jha [2003] discusses the role circuity plays in selecting highway alignments. With any increase in value placed on efficiency, the value placed University of Minnesota, Department of Civil Engineering, 500 Pillsbury Drive SE, Minneapolis, MN, USA, giaco018@umn.edu University of Minnesota, Department of Civil Engineering, 500 Pillsbury Drive SE, Minneapolis, MN, USA, james359@umn.edu RP Braun-CTS Chair of Transportation Engineering; Director of Network, Economics, and Urban Systems Research Group; University of Minnesota, Department of Civil Engineering, 500 Pillsbury Drive SE, Minneapolis, MN USA, dlevinson@umn.edu, nexus.umn.edu 1 Metropolitan Statistical Area, MSA, and metropolitan area are used synonymously in this paper. As defined by The Office of Management and Budget (OMB), a metropolitan area is that of a large population nucleus, together with adjacent communities having a high degree of social and economic integration with that core. Metropolitan areas comprise one or more entire counties, except in New England, where cities and towns are the basic geographic units. Metropolitan Statistical Area and metropolitan area identify the same geographic regions. 1

2 on network circuity increases as well. Popken [2006] examines its significance in the shipping industry and its opportunity cost with utilization. More recently, Uester and Kewcharoenwong [2011] assess the implications that road network circuity, as well as other factors, have in the truck shipping industry regarding truck company employees and the associated costs. The topic of circuity bears great importance to many economic sectors. For the first time, this paper calculates circuity for 51 metropolitan areas over the course of 20 years, looking at the road networks in 1990, 2000, and Providing a consistent estimate requires a set of definitions and application of those definitions. The next two sections address those topics. To ensure the measures are reasonable, weighted circuity is computed, weighting the network origin-destination pairs by the likelihood of that distance appearing in actual commutes. Survey data from the National Household Travel Survey 2 (NHTS) are employed to weight for home-work trip distances (e.g. more people choose a commute distance of 5-10 kilometers than kilometers). The subsequent sections present the results and compare cities and trends over time. Implications about changing urban form are identified in the Conclusions. 2 Circuity Definitions In this section multiple types of circuity will be discussed. 2.1 Unweighted Circuity On any road network, people must travel a distance that is greater than the Euclidean distance to reach their destination. This can result from many obstacles such as bodies of water, mountains, or some other difficult form of terrain. The added distance traveled to reach one s destination varies from area to area (and with every origin and destination), and the metric of circuity allows us to measure just that. Circuity can be calculated as such: Dn C = (1) DE where C = Average Unweighted Circuity, and D n = Sum of the network distances between all origin-destination pairs in the subsample, and D E = Sum of the Euclidean distances between all origin-destination pairs in the subsample. Table 1 shows the unweighted circuities as calculated in Equation 1 for trips less than or equal to 60 kilometers. Student t-tests were performed between data from 1990 and Respective p-values and confidence intervals are tabulated in Table 1. When comparing Tables 1 and 2 it is almost always the case that the unweighted circuity is lower than the weighted circuity for an MSA in any of the years analyzed. The only exceptions are Honolulu and Salt Lake City for all three analyzed years, the 1990 versions of San Diego and Las Vegas, and the 2000 road network of Las Vegas. It is obvious that people prefer shorter commutes (at least to a point). These observations 2 Actual data was measured in units of miles. For this study, units of kilometers were used. 2

3 agree with previous literature [Levinson and El-Geneidy, 2009] that higher circuities are experienced with shorter commutes. Table 3 corroborates this on an individual MSA scale, for Minneapolis-St. Paul. 2.2 Weighted Circuity Measuring the total network circuity does little to provide insight on how the network is performing. Each metropolitan area has its own profile identifying what percentage of workers travel certain distances to work, and the results of an unweighted circuity calculation will disproportionately measure the longer trips, because longer distances carry more weight in Equation 1. For this reason, a new equation is needed to accurately measure the circuity experienced on a road network. Equation 2 identifies the circuity of a road network for trips less than or equal to 60 kilometers, also accounting for the distribution of home to work trips actually traveled in the MSA: C w = Ii=1 T i C i Ii=1 T i (2) where C w = Weighted circuity of trips less than or equal to threshold t, T i = Number of trips in each interval [U.S Department of Transportation, Federal Highway Administration, 1995, 2001, 2009], C i = Circuity of trips in the interval i (based on network distances) [Office of Management and Budget, U.S. Census Bureau, 2009], interval size s = 5km, threshold t = 60km, total number of intervals I = t/s = 12). It was discussed that weighting all trips in a metropolitan area should be included. This became a cause for concern because some metropolitan areas are smaller than others and do not have trips which reach a distance of 100 kilometers. Therefore not only were those trips likely originating from locations outside the MSA, but any trips simulated in such an MSA were erroneous and should not be included in any calculation. An interval size (s) of 5 km was chosen based on ensuring sufficient sample size in each metropolitan area. A threshold (t) of 60 km included more than 95% of work trips in the tested metropolitan areas. 3 Procedure 3.1 Data Collection and Generation Metropolitan Statistical Areas (MSA) as defined by the Census in 2009 were used as a basis for this study. MSA boundaries change decennially. It was considered to use the definitions from 1990 and 2001 with the Census road network data from 1990 and 2000 respectively, but this was decided against for two reasons. First, the definitions change drastically with some metropolitan areas. For instance, in 1990, the Phoenix MSA definition only included Maricopa County. By 2009, the MSA definition included Pinal County as well. Another issue considered the definition of Consolidated Metropolitan Statistical Areas (CMSA) as opposed to Metropolitan Statistical Areas. This can be contrasted with Census nomenclature used in 1990, where within MSAs there were Primary Metropolitan Statistical Areas (PMSA), different markedly from 3

4 the definitions in A good example of this is Pittsburgh where in 1990 the area was divided into two PMSAs and in 2009 the counties were listed altogether in one MSA. In 2009, the Census did not define CMSAs nor PMSAs, and opted to define Metropolitan Divisions within MSAs. To minimize confusion the 2009 Census MSA definitions were used throughout this study, ensuring a consistent geography for all three points in time. 3.2 Sampling Sampling of origins and destinations occurred randomly (as opposed to based on actual trips, as had been used for Levinson and El-Geneidy [2009]) The reason for not using actual trips was limited data availability. Actual origins and destination blocks (with some privacy protection filters) are available for many (but not all) metropolitan areas for recent years from the Census Longitudinal Employment Household Dynamics survey. These data are not available historically. Points were randomly selected across the network, not always falling on nodes or junctions. The network has more points in built-up areas, so points are more likely to be sampled in developed areas. To weight the data to better reflect actual trips, data collected from the NHTS regarding home to work trip distances accounted for the distribution of points of urban and rural areas. Using GIS software (ArcGIS), 500 points generated over the MSA represented randomly located home origins, and 100 points generated represented work destinations. From there, the shortest distance on the road network of the MSA was measured, after the MSA had been projected onto a plane according to the State Plane System, North American Datum (NAD) Files for the 1990 road networks were created using NAD 1927, and a conversion to 1983 was required. These distances were tabulated and their respective Euclidean distances were collected as well. Ideally, 50,000 origindestination pairs, representing trips were generated, though it was common that a few points would fail, resulting in a total generation of trips slightly under 50,000 for each year for each metropolitan area. 3.3 NHTS Data Not all MSAs had data collected continually from their studies, and when this occurs it is clearly marked detailing how the data from NHTS was treated. The NHTS conducted surveys in 1990, 1995, 2001, and Data from 1995 was better coded to metropolitan areas than 1990, so that was used instead. Those surveys are overlaid with the data generated from the road networks of 1990, 2000, and 2010 respectively. For some metropolitan areas, data was not collected by the NHTS for all three surveys. Metropolitan areas that did not have survey data for the 1995 survey are: Atlanta, Grand Rapids, Honolulu, Jacksonville, Louisville, Providence, and Raleigh. For those metropolitan areas, survey data from 2001 was used. The only MSA that did not have survey data for just the 2009 survey was Honolulu, HI, for which survey data from 2001 was used. Another notable case in which the NHTS data did not match perfectly was that of 4

5 Baltimore and Washington, DC. Though in the 1995 survey they were listed separately, these metropolitan areas were combined for the 2001 and 2009 surveys which represents an added uncontrollable variable in the data for those two metropolitan areas. 3.4 Sample Size Determining an appropriate sample size is essential to minimizing the computational time incurred when analyzing an Origin-Destination (OD) Matrix in a road network. To do so, many simulations were run on the road network of Miami, Florida 3. The MSA of the area includes Broward, Miami-Dade, and Palm Beach Counties. These counties include Miami, Fort Lauderdale, and Pompano Beach. Five iterations were run with origin-destination matrices of 2000, 8000, 50000, and trips respectively, 20 iterations in total. Unweighted circuity values were used as a measure, as this is the raw data. The results of these iterative tests can be seen in Table 4. Obviously from each set of generated origins and destinations the unweighted circuity value will differ slightly from test to test. After a few trials with each matrix one can begin to see how reproducible the results of that matrix are. The standard deviation (σ) of these results will be the measure used to determine what matrix will be used for testing the rest of the metropolitan areas. Miami was chosen because it is a large metropolitan area that has relatively few (three) counties relatively simple to manage. From Table 4, it was determined that an OD Matrix of 500x100 would be used for all metropolitan areas. With a standard deviation (σ) of , this creates a 95% confidence interval of ± in the unweighted circuity value. Each of the 51 metropolitan areas has matrices generated in the maps for 1990, 2000, and That is 150 iterations and with a 95% confidence interval; it is expected that approximately 7 or 8 of those iterations will fall outside the aforementioned confidence interval. Since the standard deviation of unweighted circuity results as calculated in Equation 1 does not decrease beyond 50,000 trips generated, the sample size chosen for all cities was 50,000 trips. 4 Results Shown in Table 2 are the results of weighted circuities of trips up to 60 kilometers in their respective MSAs. The table is comprehensive and includes all 51 MSAs incorporated in this study; showing their results from Equation 2 for 1990, 2000, and As is the case in Table 1, Student t-tests were performed on the data from 1990 and Calculated in the bottom row of Table 2 is the average (weighting all 51 cities equally) for each year analyzed (1990, 2000, 2010). In these averages, a clear increasing trend can be seen. Table 5 has been generated to identify the frequency of the order in which the circuity values ascend. For example, means that the weighted circuity 3 Also referred to as South Florida. 5

6 was lowest in 1990 and highest in 2010, obviously 2000 falling between. As can be seen in Table 5, the first three possibilities listed have 1990 with a lower weighted circuity value than This sums to 43 of the 51 tested MSAs. When looking at how 1990 compares to 2000, again from Table 5 it can be seen that 29 of the 51 metropolitan areas weighted circuity values increase over that 10 year time period. Finally, from 2000 to 2010 thirty-four MSAs exhibit weighted circuity value increases. Though there is not much difference between the number of MSAs that increase in the decade following 1990 to that of 2000 (29 MSAs increase, 34 MSAs increase respectively), there is an obvious trend over the 20 year time period from 1990 to 2010 in which 43 of the 51 MSAs increase (according to a linear fit). The slope of a linear fit has been calculated in the last column of Table 2, and then multiplied by 10 to identify the change over decade. To clarify, the Slope (S) tabulated in Table 2 can be defined as: S = 10 n Y C w Y C w n C w 2 [ C w ] 2 (3) where S = The average change of weighted circuity (of trips less than or equal to 60 kilometers) in an MSA over a 10 year interval, Y = The year, n = The number of points used in the linear fit, and all other variables as aforementioned. 4.1 Circuity decay with distance A linear fit is not ideal for modeling circuity over long periods because for example in the case of a negative trend a linear fit would imply that at some point the weighted circuity value will go below 1.0, a value theoretically impossible. A logarithmic fit is another method that could be easily implemented but in many instances (33 of the 51 MSAs observed) the weighted circuity value obtained for the year 2000 was either higher or lower than both the weighted circuity values of 1990 and Clearly the problem of understanding changes in circuity of an MSA over time is complex. If the values were random over time, roughly half would decrease and half would increase. With all else in the world held constant, they would remain the same over time. Obviously such is not the case. Furthermore, when simply averaging the values of the 51 MSAs (shown in bottom row of Table 2), another clear trend is exhibited as the weighted circuity value increases by over , and increases by from Clearly there is a change in these values over time but what is the driving force? While there are multiple causes, suburbanization [Jackson, 1985] is a likely suspect, and it is possible that this continued trend of progression to lower-density suburban areas is associated with this general increase in circuity over the past few decades. The average population density for all MSAs in 2010 is the density averaged over all 51 metropolitan areas; this gives each metropolitan area equal weight, rather than summing the populations of all 51 MSAs and dividing over the sum of all land area within the MSAs. For more detailed population, and land area regarding all 51 MSAs used in this study see [Bureau, 2010]. 6

7 It was observed that between population density, ρ Y, and weighted circuity, C w,y, the correlation coefficient in Year Y = 2010 is r = Though this is by no means a strong correlation, it is statistically significant. It indicates that MSAs with higher population densities tend to have lower weighted circuity values. Correlations were explored for a few other variables thought to potentially be related to weighted circuity in metropolitan areas. Table 6 tabulates those correlations. As is seen in Table 6, there are many areas that can be explored in future research as factors that could potentially influence weighted circuity values. 4.2 Circuity Profile In Table 3, various statistics about the trips generated within the MSA of Minneapolis, MN are shown. C i shows an unweighted circuity value according to Equation 1 over a certain network distance interval (i). C µ shows perhaps a more interesting value, taking the circuity of each trip and then averaging those values, giving each trip equal weight regardless of distance. Equation 4 provides further detail on the calculation. Ct C µ = (4) T g where C t = the circuity of each trip, and C µ = Average of the trip circuities, T g = Number of trips generated in simulation. σ µ is the standard deviation of the trip circuities, C t. Obviously, circuity cannot go below a value of 1.0. For example in 1990 when i=0-5 [km], σ µ =0.609 and C µ =1.443, one standard deviation (σ µ ) below the average would be below a value of 1. This indicates that the underlying distribution at this interval is not normal. However at higher intervals say again in 1990 when i=55-60 [km], σ µ =0.146 and C µ = Though the Average of the trip circuities has decreased so has the standard deviation (σ µ ). Using a Two-Sample Kolmogorov-Smirnov Test 4 it is easily to quickly see that all of the various average circuities about each 5 kilometer interval are not from the same distribution. Another way to characterize the trips generated in a metropolitan area is as a function of trip time, and exploring the circuities of certain thresholds as opposed to intervals. For example measuring unweighted circuity between 0-10 min, 0-20 min, 0-30 min as opposed to 0-10 min, min, min. Such a profile can be seen in Table 7, where C th = the unweighted circuity (as discussed in Equation 1) of all trips within a certain time threshold. As expected, the circuity decreases as the travel time increases. 4.3 Fitting the Circuity Profile Levinson [2012] discusses that circuity can be modeled exponentially. In searching for the best exponential fit, when looking at unweighted circuity and travel time, linear fits were explored for all 51 MSAs in the study on linear plots, double-log plots, and both semi-log plots. The highest average r-squared value (over the 51 MSAs) was the 4 For more information, see: 7

8 statistic used in determining what plot gave the best fit. The double-log plot gave the best fit for all three years analyzed in the study (1990, 2000, and 2010). The results of those calculations are tabulated in Table 8. The equation that corresponds to the double-log plot is as follows: C = e b t ζ (5) where C = unweighted circuity at some time (t), ζ = the slope calculated using an Ordinary Least Squares Regression (OLS Regression) of the data on a double-log plot, b = the ordinate intercept of the OLS Regression of the data on a double-log plot. Table 9 shows the values calculated from OLS Regressions to be placed in Equation 5 as well as their corresponding r 2 values. A few metropolitan areas have very low r 2 values (examples include Las Vegas and San Diego). For these MSAs it can be assumed that the model provided in Equation 5 is not a good fit. 5 Conclusions The implications of network circuity are vast. Though a general decrease in the distance traveled is a benefit, all else equal, such is rarely the case as people will accept a greater distance traveled for a savings in travel time. This resonates with the benefits of highways, interstates, and other high-speed arterials; accepting a greater fuel cost for a time savings. Though causality is somewhat speculative, a negative correlation has been observed between metropolitan population density and weighted circuity. Furthermore, over the period from , 43 of the United States 51 most populated MSAs (84%) experience positive trends in weighted circuity. So in the most populated metropolitan areas in the United States, random points have not only become farther apart in distance, they are becoming more circuitous, suggesting that the more recently constructed parts of street networks are laid out less efficiently than older parts of the network. It is also not surprising that new areas are less well connected than older areas, as that is part of the general process of network growth and infill [Levinson and Huang, 2011]. But the trend is that they stay less well connected, or that the amount of new network is becoming disproportionately significant. This is consistent with many critiques of US suburban development over the past decades. The underlying motivations and causes to explain these increases across many of the most populated MSAs in the United States remain unanswered and signal areas of future research. Though Equation 5 can reliably model many of the metropolitan areas observed in this study (as evidenced with high r 2 values), clearly the model fails for some MSAs. More research is necessary to better understand the mechanisms which drive the relationship between circuity and travel time/travel distance. Exploring the connections between intra-metropolitan location and network structure and daily travel is an important line of future research. Determining causality, and whether influencing the circuity of a metropolitan area could influence metropolitan economic productivity and agglomeration economies bears great potential for future research. Other network structure properties for future research include but are not 8

9 limited to the entropy, connectivity, treeness, and accessibility of a road network as discussed in [Levinson, 2012], and their relationships with road network circuity. Acknowledgments The authors would like to thank Paul Anderson for his work in data preparation and Andrew Owen for reviewing an earlier draft. References Ronald H. Ballou, Handoko Rahardja, and Noriaki Sakai. Selected country circuity factors for road travel distance estimation, volume 36. November U.S. Census Bureau. American factfinder, URL gov/faces/tableservices/jsf/pages/productview.xhtml?pid=dec_10_sf1_ GCTPH1.US23PR&prodType=table. A. J. Goldman. Approximate localization theorems for optimal facility placement. Transportation Science, pages , Kenneth T. Jackson. Crabgrass Frontier: The Suburbanization of the United States. Oxford University, MK Jha. Criteria-based decision support system for selecting highway alignments. Journal of Transportation Engineering-Asce, 129(1):33 41, Jan-Feb doi: DOI /(ASCE) X(2003)129:1(33). D. Levinson and A. Huang. A positive theory of network connectivity. Environment and Planning Part B, 39(2): , David M. Levinson. Network structure and city size. Plos One, 7(1):e29721, January doi: DOI /journal.pone David M. Levinson and Ahmed El-Geneidy. The mimimum circuity frontier and the journey to work. Regional Science and Urban Economics, 39: , G. F. Newell. Traffic flow on transportation networks. MIT Press, Office of Management and Budget, U.S. Census Bureau. Metropolitan and Micropolitan Statistical Areas and Components, December URL population/metro/files/lists/2009/list1.txt. Douglas A. Popken. Controlling order circuity in pickup and delivery problems. Transportation Research Part E-Logistics and Transportation Review, 42(5): , September doi: DOI /j.tre Halit Uester and Panitan Kewcharoenwong. Strategic design and analysis of a relay network in truckload transportation. Transportation Science, 45(4): , November doi: DOI /trsc

10 U.S Department of Transportation, Federal Highway Administration. National household travel survey, URL U.S Department of Transportation, Federal Highway Administration. National household travel survey, URL U.S Department of Transportation, Federal Highway Administration. National household travel survey, URL 5 t-statistic based on a normal two-tailed distribution comparing 2010 and 1990 data. For example, a t-statistic greater than 1.96 corresponds to a 95% confidence interval stating that the numbers come from different distributions. 6 Data was not collected in the NHTS 1995 Survey. NHTS 2001 data was used. 7 Data in NHTS 2001 and 2009 was combined between Washington, DC and Baltimore, MD. Surveys were performed separately in Data was not collected in the NHTS 2009 Survey. NHTS 2001 data was used. 10

11 MSA t-stat 5 p-value C.I. Atlanta % Austin % Baltimore % Boston % Buffalo % Charlotte % Chicago % Cincinnati % Cleveland % Columbus % Dallas-Fort Worth % Denver % Detroit % Grand Rapids % Hampton Roads % Hartford % Honolulu % Houston % Indianapolis % Inland Empire % Jacksonville % Kansas City % Las Vegas % Los Angeles % Louisville % Memphis % Milwaukee % Minneapolis-St.Paul % Nashville % New Orleans % New York % Oklahoma City % Orlando % Philadelphia % Phoenix % Pittsburgh % Portland % Providence % Raleigh % Rochester, NY % Sacramento % Salt Lake City % San Antonio % San Diego % San Francisco % San Jose % Seattle % South Florida % St. Louis % Tampa Bay % Washington % Average Table 1: Unweighted Circuities based on Equation 1 ( ) 11

12 MSA t-stat 5 p-value C.I. Atlanta % Austin % Baltimore % Boston % Buffalo % Charlotte % Chicago % Cincinnati % Cleveland % Columbus % Dallas-Fort Worth % Denver % Detroit % Grand Rapids % Hampton Roads % Hartford % Honolulu % Houston % Indianapolis % Inland Empire % Jacksonville % Kansas City % Las Vegas % Los Angeles % Louisville % Memphis % Milwaukee % Minneapolis-St.Paul % Nashville % New Orleans % New York % Oklahoma City % Orlando % Philadelphia % Phoenix % Pittsburgh % Portland % Providence % Raleigh % Rochester, NY % Sacramento % Salt Lake City % San Antonio % San Diego % San Francisco % San Jose % Seattle % South Florida % St. Louis % Tampa Bay % Washington % Average Table 2: Weighted Circuities ( ) 12

13 Interval (i) [km] C i C µ σ µ C i C µ σ µ C i C µ σ µ Table 3: Circuity Profile of Minneapolis, MN MSA based on Trip Distance Matrix Trips C σ 100x20 2, x40 8, x100 50, x , Table 4: Sampling iterations performed on Miami, Florida MSA Year (Ascending) Count Total 51 Table 5: Frequencies of Circuity Trends Variable C w Land Area Population Density (ρ) ζ Density (ρ) Population Land Area Table 6: Correlations (r) of selected variables with Circuity (C w ) 13

14 Time [min] C th Table 7: Circuity Profile of Minneapolis, MN MSA based on Travel Time r 2 Plot Linear-Linear Log-Linear Linear-Log Log-Log Table 8: r 2 Data for Linear Fits on Various Plots. Values are the average of all 51 MSAs (evenly weighted). Circuity and Time data were displayed on various plots (listed) and the best linear fit was sought. Best linear fit was on a double-log plot. 14

15 MSA ζ b r 2 ζ b r 2 ζ b r 2 Atlanta Austin Baltimore Boston Buffalo Charlotte Chicago Cincinnati Cleveland Columbus Dallas-Fort Worth Denver Detroit Grand Rapids Hampton Roads Hartford Honolulu Houston Indianapolis Inland Empire Jacksonville Kansas City Las Vegas Los Angeles Louisville Memphis Milwaukee Minneapolis-St.Paul Nashville New Orleans New York Oklahoma City Orlando Philadelphia Phoenix Pittsburgh Portland Providence Raleigh Rochester, NY Sacramento Salt Lake City San Antonio San Diego San Francisco San Jose Seattle South Florida St. Louis Tampa Bay Washington Average σ ζ Table 9: Summary Table of ζ, b, and respective r 2 values 15

16 Distribution of ζ Values ζ (Zeta) Year Figure 1: A summary of all calculated ζ (zeta) values. Blue box indicates 50% confidence interval. Red Line indicates the median. Whiskers represent 95% confidence interval, with outliers marked as red crosses. 16

17 C w [60 km] Travel Time Figure 2: An average weighted circuity function averaged over all 51 MSAs studied in 1990, 2000, and When looking at Figure 2, the most important observation is that these are very similar lines, 2000 and 2010 are virtually identical. However circuity for short distances has increased from 1990 to 2000 (and thus to 2010) as indicated by the constant term. This is not surprising since more short distance trips (like all trips) are suburban in nature and suburban trips, especially short suburban trips on tree-like subdivision streets, are likely to be more circuitous than urban trips. 17