An Update on Public Transportation's Impacts on Greenhouse Gas Emissions (2021)

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.

47   Transit Multiplier Methodology This appendix provides the full methodology for the model used to develop the transit multiplier for this project. This method and its application in the public transportation GHG analysis are summarized in the body of the report and Appendix A. Appendix B is provided for those who want to explore the details of the transit multiplier structural equation model. The transit multiplier model was developed for this project by Sadegh Sabouri, Ph.D. Candidate in Metropolitan Planning, Policy & Design, College of Architecture and Planning, University of Utah, and Reid Ewing, Ph.D., Professor, City & Metropolitan Planning, Director of the Metro- politan Research Center, University of Utah. Data and Method 1 – Regional Household Travel Survey and Built Environment Data This study uses geocoded household travel data to explain household VMT in terms of sociodemographic, built environment, and travel variables. The main criterion for inclusion of regions in this study was data availability. Regions had to offer regional household travel surveys with XY coordinates, so the authors could geocode the precise locations of residences and capture the built environment for households more accurately. It is not easy to assemble databases that meet this criterion because confidentiality concerns mean that metropolitan planning organizations are often unwilling to share XY travel data. At present, there are consistent datasets for 31 regions. The resulting pooled dataset consists of almost 900,000 trips by 81,573 households (see Table B-1 and Figure B-1). The regions are as diverse as Boston and Portland at one end of the urban form continuum and Houston and Indianapolis at the other. To the researchers’ knowledge, this is the largest sample of household travel records ever assembled for such a study outside the National Household Travel Surveys (NHTSs) of 2009 and 2017. And relative to the NHTS, the database for this project provides much larger samples for individual regions and permits the calculation of a wide array of built environmental variables based on the precise location of households. The NHTS provides geocodes (identifies households) only at the census-tract level. In this project, for each of the 31 regions, the built environment variables (known as 5Ds: density, land use diversity, street design, distance to transit, and destination accessibility) have been computed at the traffic analysis zone (TAZ) level based on these data: • Population and employment at the block, block group, and TAZ level; from these, activity density can be computed. • Travel times for auto and transit travel from TAZ to TAZ (so-called travel time skims); from these and TAZ employment data, regional employment accessibility measures for auto and transit can be computed. A P P E N D I X B

48 An Update on Public Transportation’s Impacts on Greenhouse Gas Emissions • Parcel-level land-use data with detailed land-use classifications; from these, detailed measures of land use mix can be computed. • A geographic information system (GIS) layer for street networks and intersections; from these, intersection density and percentage of four-way intersections can be computed. • A GIS layer for transit stops; from these data, transit stop densities can be computed. 2 – Variables, Conceptual Framework, and Research Design Variables The outcome variables are household VMT, activity density, and the number of transit trips per household. There are mediating variables between the percentage of jobs reachable within 30 minutes by transit and household VMT. There are confounding influences, such as household income and household size. The selection of explanatory variables to predict the outcome variables is based on common sense as well as theory. In addition, these are mostly the variables that were used in TCRP Report 176: Quantifying Transit’s Impact on GHG Emissions and Energy Use—The Land Use Component. (Gallivan et al. 2015) Different variables are tested for significance as predictors of transit trips, activity density, and VMT, while the final model only includes the independent variables that had an expected Regions Surveyed Households Surveyed Trips Mean of Household Vehicles Albany, NY 1,309 12,618 2.02 Atlanta, GA 8,729 93,681 2.11 Boston, MA 7,181 86,915 1.64 Burlington, NC 539 5,111 2.24 Charleston, SC 223 2,098 2.04 Dallas, TX 2,824 27,066 2.05 Denver, CO 5,021 55,056 1.94 Detroit, MI 819 14,690 1.49 Eugene, OR 1,592 16,563 1.82 Greensboro, NC 1,768 17,561 2.09 Hampton Roads–Norfolk, VA 1,797 16,495 2.16 Houston, TX 5,233 59,552 2.27 Indianapolis, IN 3,447 37,473 1.89 Kansas City, MO 2,816 31,779 1.84 Madison, WI 128 1,316 2.12 Miami-Dade, FL 1,283 11,580 1.76 Minneapolis–St. Paul, MN 7,439 79,236 1.81 Orlando, FL 775 7,315 2.00 Phoenix, AZ 3,541 37,811 1.92 Portland, OR 4,135 47,551 1.86 Provo-Orem, UT 1,314 19,255 2.08 Richmond, VA 560 5,123 2.13 Rochester, NY 3,393 23,145 1.81 Salem, OR 1,551 16,231 1.82 Salt Lake City, UT 3,034 44,565 2.04 San Antonio, TX 1,547 14,952 1.90 Seattle, WA 4,883 47,877 1.49 Springfield, MA 778 8,456 1.70 Syracuse, NY 592 5,752 1.94 Tampa, FL 2017 17,538 1.79 Winston-Salem, NC 1,305 12,168 2.15 Total 81,573 883,695 1.92 Table B-1. Combined household travel survey dataset from 31 regions of the United States.

Transit Multiplier Methodology 49   sign and a statistically significant relationship to the outcome variables. Table B-2 presents the definition and descriptive statistics for all the endogenous (outcome) and exogenous variables investigated in the model. Hypothesis The hypothesis is that households that live in areas that have greater accessibility to jobs by transit will generate more transit trips, which will lead to a decrease in household VMT. This will be the direct effect of transit accessibility on VMT. By transit accessibility is meant job accessibility by transit within 30 minutes (the variable labeled accessibility_by_transit30 in Table B-2). For transit trips, also developed were a dummy variable of zero (household with no transit trip) and one (household with at least one transit trip) to control for the excessive number of zero values—in this case, 89.2% of households with no transit use at all. In every city in the United States, infrastructure dedicated to private-vehicle travel dwarfs public transpor- tation infrastructure. It is not surprising then that transit represents a very small proportion of total travel in the United States. Hence, use of transit is the exception rather than the rule overall in the United States, even as it is a major mode of travel in many communities. The indirect effect will be through activity density because the hypothesis is that areas with better transit accessibility will have higher activity density and generate more non-auto trips. In addition, there is a direct path from transit accessibility to VMT to capture all the remaining reductions in VMT that result from having better accessibility by transit. Note that this path is considered as another indirect effect. The complex causal chain described previously is best modeled with structural equation modeling (SEM). SEM is a statistical technique for evaluating complex hypotheses involving Figure B-1. Location of 31 regions in the United States (Boston not pictured).

50 An Update on Public Transportation’s Impacts on Greenhouse Gas Emissions multiple, interacting variables (Grace 2006). The estimation of structural equation models involves solving a set of equations. There is an equation for each “response” or “endogenous” variable in the system. They are affected by other variables and may also affect other variables. Variables that are solely predictors of other variables are termed “influences” or “exogenous” variables. They may be correlated with one another but are determined outside the system. There are several related and distinctive features of SEM that make it appropriate for this analysis (Ewing et al. 2015). On the other hand, the data and model structure are hierarchical, with households and TAZs “nested” within regions. The best statistical approach for nested data is multilevel modeling (MLM), also called hierarchical modeling. MLM accounts for spatial dependence among observations. Ordinary least squares and other regression methods produce biased standard Table B-2. Variables used to estimate household VMT. Variable Description N Mean Std. Dev. Endogenous Variables lnvmta vehicle-miles traveled 81,573 31.56 31.67 transit_trips transit trips per household 81,573 0.28 1.09 lnactivity_density activity density within TAZ (pop + emp per square mile in 1,000s) 81,573 1.67 0.85 Exogenous Variables: Socioeconomic Characteristics vehicle number of vehicles owned by household 81,573 1.93 1.04 household_size household size 81,573 2.48 1.34 employed number of employed persons in household 81,573 1.24 0.88 income household income in 1,000s 81,573 77.24 49.42 Exogenous Variables: Built Environment Characteristics within TAZ jobpopb job–population balance 81,573 0.57 0.27 entropyc land use mix 81,554 0.43 0.25 intersection_density intersection density 81560 104.2 81.3 percentage_4way percentage of four-way intersections 81,540 25.38 20.05 accessibility_by_auto10 percentage of regional employment within 10 minutes by auto 81,573 7.83 11.87 accessibility_by_auto20 percentage of regional employment within 20 minutes by auto 81,573 30.46 26.87 accessibility_by_auto30 percentage of regional employment within 30 minutes by auto 81,573 52.52 30.46 accessibility_by_transit30 percentage of regional employment within 30 minutes by transit 81,573 20.83 24.28 Exogenous Variables: Regional Characteristics regional_pop regional population in 1,000s 31 2094.14 1794.37 regional_emp regional employment in 1,000s 31 1055.57 917.82 pop_density regional population density 31 2186.41 747.25 fuel average metropolitan fuel price 28 2.76 0.12 freeway freeway lane miles per 1,000 population 28 0.72 0.28 other other lane miles per 1,000 population 28 2.20 0.40 route_density transit route density per square mile 28 3.02 1.97 transit_freq transit service frequency (annual revenue miles/route miles) 28 8722.01 3233.50 tpm annual transit passenger miles per capita 28 130.05 97.05 compactness the compactness index 26 99.90 26.58 a “ln” at the beginning of a variable indicates the “natural logarithm” version of the variable. b job–population balance = 1 − [ABS(employment − 0.2 * population)/(employment + 0.2 * population)]; ABS = absolute value of expression in parentheses. The value 0.2, representing a balance of employment and population, was found through trial and error to maximize the explanatory power of the variable (Ewing et al. 2015). c land use entropy = − [residential share * ln(residential share) + commercial share * ln(commercial share) + public share * ln(public share)]/ln(3), where ln is the natural logarithm (Ewing et al. 2015).

Transit Multiplier Methodology 51   errors and inefficient regression coefficients. MLM overcomes these limitations, accounting for the dependence among observations and producing more accurate coefficient and standard error estimates (Raudenbush and Bryk 2002). In this study, a multilevel SEM model was used to explain outcome variables and compute the transit land-use multiplier. Note that by using multilevel SEM and considering regional char- acteristics, separate multipliers for each individual region can be estimated, which is the whole purpose of this project. Also note that since the dataset is massive and the model is extremely complex due to the multilevel component, it will take a huge amount of time (nearly impossible) from statistical software such as R (i.e., lavaan package) or Stata (i.e., GSEM command) to estimate the model that is sought. In addition, historically, SEMs have been estimated using a maximum likelihood approach to select parameter values that best reproduce the entirety of the observed variance–covariance matrix. The goodness of fit of the SEM can then be evaluated using a chi-square test compar- ing the estimated to the observed covariance matrix (Grace 2006). This approach, however, assumes that all observations are independent, and all variables follow a (multivariate) normal distribution (Grace 2006). It also restricts the minimum number of observations necessary to fit the SEM since there need to be sufficient degrees of freedom to estimate the whole variance– covariance matrix (the “t rule,” Grace 2006). These restrictions led to the parallel development of directed acyclic, or piecewise, SEMs based on applications from graph theory. In piecewise SEM, the path diagram is translated to a set of linear (structured) equations, which are then evaluated individually. The switch from global estimation, where equations are solved simultaneously, to local estimation, where each equation is solved separately, allows for the fitting of a wide range of distributions and sampling designs (Shipley 2000, 2009). Because of that, in this study, a multilevel piecewise SEM was used, which allows estimation of a model in a very short period of time, using the piecewiseSEM package in R software for a total of 81,573 households with no missing data. Since the household’s choice to have a transit trip or not (variable called any transit) is a dummy outcome variable, hierarchical binary logistic regression might seem to be the best approach. However, there is a growing body of literature that explains the problem of unobserved heterogeneity in logistic regression models (Mood 2010). The literature then suggests the use of linear prediction models, especially when the sample size is huge. Because of that, and also to make the model less complex, the authors used a linear probability model for this dummy variable as well. Result and Discussion Figure B-2 shows the best-fit model from the multilevel piecewise SEM analysis. The diagram has been simplified to avoid confusion in understanding the paths from exogenous explana- tory variables to each of the endogenous variables. In this regard, all of the socioeconomic and built environment variables that were used are called “Socioeconomic and D Variables,” except job accessibility by transit, so that the direct and indirect effects of transit on VMT can be illustrated. Unfortunately, many of the regional variables and relationships are neither theoretically nor statistically significant, possibly due to the small sample size at the regional level. There- fore, they can be dropped. One possible solution is to apply a factor analysis or principal component analysis; these are techniques for reducing the dimensionality of such a dataset, increasing interpretability but at the same time minimizing information loss. However, since the ultimate goal of this project is to provide a simple formula so that other transit agencies

52 An Update on Public Transportation’s Impacts on Greenhouse Gas Emissions across the United States can compute their own multipliers, it was decided not to use these data reduction techniques. In addition, all efforts have been made to have a parsimonious model with great explana- tory predictive power. The causal path diagram can undoubtedly become more complex by adding the number of vehicles owned by a household as another mediating variable. However, this will lower the interpretability of the results and make the computation of multipliers extremely difficult. As explained in the conceptual framework, job accessibility by transit (within 30 minutes) and socioeconomic (e.g., household size) and other built environment variables (e.g., accessibility by auto within 30 minutes) have direct paths to activity density. On the other hand, these exogenous variables have direct paths to transit trips (as well as any transit) made by households. Finally, activity density, transit trips, and exogenous variables are directly affecting VMT. The direct and indirect effects of transit accessibility on VMT, which was explained in the previous section, are shown in red and blue, respectively, in Figure B-2. Annual transit passenger miles per capita (i.e., tpm), which is the only regional level variable, has a direct path to each of the outcome variables. In addition, those two paths from the regional variable to the blue and red paths represent the interactions between tpm and activity density, any transit, and transit trips. The value of these interactions will allow separate transit land-use multipliers for each region. Note that while there are consistent datasets for 31 regions, tpm has been measured in 28 regions. Hence, three regions were dropped due to data availability. Table B-3 presents the path coefficient estimates (regression coefficients) and associated statistics for direct effects of explanatory variables on outcome. Note that all variables are mean-centered to have a more meaningful interpretation of the coefficients and interactions Figure B-2. Final causal path diagram (simplified).

Transit Multiplier Methodology 53   in the multilevel analysis. In terms of the goodness-of-fit measures, piecewise SEMs do not have an equivalent to the Tucker–Lewis index or comparative fit index. The main diagnostic is Fisher’s C, which is a test to see if there are missing paths. The model has a low p-value, which suggests that one or more of the missing paths is important. However, it is suspected that this is due to the huge sample size and, hence, huge degrees of freedom, which make Fisher’s C significant. Almost all of the explanatory variables in Table B-3 are significant at the 0.05 probability level and also have the expected signs. As expected, accessibility by transit, household size, and number of workers in a household are positively associated with any transit and transit trips. The same trend can be seen for built environment variables [i.e., land use entropy, percentage of four-way intersection (only for any transit), and annual transit passenger miles per capita vis-à-vis any transit and transit trips]. Outcome Predictor Estimate Std. Error Crit. Value P-Value transit_trips accessibility_by_transit30 0.0010 0.0001 7.1795 <0.001 transit_trips household_size 0.0259 0.0021 12.1036 <0.001 transit_trips employed 0.0141 0.0034 −4.1927 <0.001 transit_trips income −0.0008 0.0001 −14.8235 <0.001 transit_trips entropy 0.0160 0.0106 1.5102 0.1310 transit_trips anytransit 2.8708 0.0085 337.1042 <0.001 transit_trips tpm 0.0001 0.0002 0.5939 0.5580 anytransit accessibility_by_transit30 0.0017 0.0001 27.5868 <0.001 anytransit household_size 0.0086 0.0009 9.5751 <0.001 anytransit employed 0.0182 0.0014 12.8674 <0.001 anytransit income −0.0004 0.0000 −18.3775 <0.001 anytransit entropy 0.0226 0.0044 5.0751 <0.001 anytransit percentage_4way 0.0013 0.0001 21.316 <0.001 anytransit tpm 0.0006 0.0001 5.7948 <0.001 activity_density accessibility_by_transit30 0.0931 0.0035 26.2775 <0.001 activity_density entropy 3.0119 0.2227 13.5227 <0.001 activity_density percentage_4way 0.1743 0.0030 57.3886 <0.001 activity_density accessibility_by_auto30 0.0759 0.0034 22.2922 <0.001 activity_density tpm.m 0.0351 0.0073 4.8376 0.0001 vmt household_size 4.0699 0.0884 46.0288 <0.001 vmt employed 4.9736 0.1386 35.8733 <0.001 vmt income 0.0565 0.0023 24.589 <0.001 vmt accessibility_by_auto30 −0.2001 0.0069 −29.2041 <0.001 vmt percentage_4way −0.0803 0.0061 −13.0769 <0.001 vmt entropy −2.7304 0.4378 −6.2372 <0.001 vmt accessibility_by_transit30 −0.0907 0.0071 −12.6878 <0.001 vmt tpm −0.0306 0.0386 −0.7932 0.4468 vmt transit_trips −2.8822 0.1137 −25.3414 <0.001 vmt activity_density −0.0978 0.0142 −6.8716 <0.001 vmt accessibility_by_transit30: tpm −0.0001 0.0001 −1.9948 0.0461 vmt transit_trips: tpm 0.0020 0.0007 2.7341 0.0063 vmt activity_density: tpm 0.0004 0.0001 4.7044 <0.001 Goodness-of-fit Measures Akaike Information Criterion 1431.086 Bayesian Information Criterion 1857.176 Fisher’s C 1339.086 (p-value < 0.01) Table B-3. Path coefficient estimates (regression coefficients).

54 An Update on Public Transportation’s Impacts on Greenhouse Gas Emissions On the other hand, household income is negatively and significantly correlated with transit trips. This is mainly due to the fact that households with higher income prefer to use their own vehicles rather than using other modes such as transit. In terms of the third endogenous variable, all of the built environment variables, including accessibility by auto and transit within 30 minutes as well as the regional variable, are highly significant and positively associated with activity density, which makes sense both intuitively and theoretically. Household VMT is positively correlated with household size, workers, and income. These relationships suggest that households with more members, workers, and higher incomes are likely to drive more and probably own more vehicles, which is consistent with most previous studies. In terms of built environment variables, accessibility by transit and auto, activity density, land use entropy, and percentage of four-way intersections are all negatively asso- ciated with VMT. That is, households that live in dense, mixed, and well-connected areas and with greater accessibility to jobs are more likely to use other modes of travel rather than their personal vehicle. Unsurprisingly, the any transit and transit trips variables are significantly negatively correlated with VMT. The last three rows of Table B-3 (before the goodness-of-fit measures) show the interaction between two of the outcome variables (i.e., activity density and transit trips) as well as accessibility by transit and the transit passenger miles per capita. This inter- action is positive for activity density, which means that the negative coefficient of this variable alone (−0.0978) will decrease slightly (−0.0004) should the regional variable be included (i.e., transit passenger miles per capita). In other words, an increase in the regional variable will result in a smaller effect of activity density on VMT. For transit trips, the interaction is positive too, which means that higher transit passenger miles for a region will result in a slightly smaller reduction of VMT due to the impact of transit trips. The interaction term for the transit accessibility variable, however, is negative, suggesting a larger effect of this variable on VMT reduction in regions with higher transit passenger miles per capita. With the results presented in Table B-3, the transit land-use multiplier for each region can now be computed. If we did not have any regional variable, the direct and indirect effects could be readily computed as follows: Direct effect = (Btransit accessibility → transit trips × Btransit trips → VMT) + (Btransit accessibility → anytransit × Banytransit → transit trips × Btransit trips → VMT) = (0.001 × −2.882) + (0.0017 × 2.871 × −2.882) Indirect effect = (Btransit accessibility → activity density × Bactivity density → VMT) + Btransit accessibility → VMT = (0.093 × −0.098) + (−0.091) Since we have transit passenger miles per capita and its interaction with activity density, transit accessibility, and transit trips variables, the multiplier should be computed in this way: Direct effect: [Btransit accessibility → transit trips × (Btransit trips → VMT + (Btransit trips: tpm × (transit passenger miles per capita – 130.05)))] + [Btransit accessibility → anytransit × Banytransit → transit trips × (Btransit trips → VMT + (Btransit trips: tpm × (transit passenger miles per capita – 130.05)))]

Transit Multiplier Methodology 55   By plugging in the coefficient values, this direct effect can be simplified to: [0.001 × (−2.882 + (0.002 × (transit passenger miles per capita – 130.05)))] + [0.0017 × 2.871 × (−2.882 + (0.002 × (transit passenger miles per capita – 130.05)))] Indirect effect: [Btransit accessibility → activity density × (Bactivity density → VMT + (Bactivity density: tpm × (transit passenger miles per capita – 130.05)))] + [Btransit accessibility → VMT + (Btransit accessibility: tpm × (transit passenger miles per capita – 130.05))] Which can be simplified to: [0.093 × (−0.098 + (0.0004 × (transit passenger miles per capita – 130.05)))] + [−0.091 + (−0.0001 × (transit passenger miles per capita – 130.05))] Where 130.05 is the average of the annual transit passenger miles (per capita) in the sample of 28 regions. So, each new region should use its annual transit passenger miles per capita and subtract that by 130.05 and use the previous equations to compute its multipliers. Table B-4 shows direct, indirect, and total effects as well as the transit land-use multiplier for each of the 28 regions. These results suggest that transit-rich regions tend to have higher transit multipliers than transit-poor regions. The direct effect of transit accessibility on VMT is about the same in all regions. A transit trip is a transit trip, and it partially substitutes for an auto trip. It is the indirect effect that varies from region to region, which causes variance in the transit multiplier. Transit-poor regions have a smaller indirect effect of transit on VMT than do transit-rich regions, which makes sense. The indirect effect is the main variable in the transit multiplier equation, which causes the transit multiplier to be larger in transit-rich regions. In transit-rich regions, the effect of transit is primarily through land use changes, favoring density and its effect on travel variables other than transit passenger miles. In these regions, the main effect of transit is to boost walk trips and shorten auto trips, which are both indirect effects of transit service quality. In Seattle, for instance, with a multiplier of 9.31, each increment of additional transit acces- sibility produces a direct effect of 0.0137 reduction in VMT per capita, while each increment of additional transit accessibility produces an indirect effect of 0.1171. On the other hand, in a region like Detroit, with a multiplier of 6.51, each increment of additional transit acces- sibility produces almost a comparable direct effect of 0.0175, but a much smaller indirect effect of 0.097. This implies that there are increasing returns to scale to transit supply. The range of the multiplier is from 6.1 to 9.5 for the regions in the sample. Note that this is not the transit land-use multiplier as usually defined, but a multiplier of total VMT reduction relative to VMT reduction directly due to transit passenger miles. The main effect of transit is not due to model shifts from auto use to transit use, but rather is due to changes in the built environments that are well served by transit. If all regions are considered simultaneously and households in all regions are treated as a single database, the average transit land-use multiplier is 7.43, which isn’t far from the midpoint of the multilevel model. Transit agencies can either compute the annual transit passenger mile per capita and from that a specific multiplier, or they

56 An Update on Public Transportation’s Impacts on Greenhouse Gas Emissions can use the average or range of the multiplier for the entire sample to estimate the total effect of transit on VMT from the direct effect. To recapitulate, transit reduces automobile travel in two different ways: directly when a traveler shifts a trip from automobile to rail or bus, and indirectly when it creates more acces- sible land use and reduces automobile ownership in an area. These indirect impacts can be large, depending on the region that households reside in. The multipliers computed from these direct and indirect impacts are in an acceptable range and consistent with some of the previous research. For instance, Holtzclaw (2000) found that VMT reductions per transit passenger mile for San Francisco (a transit-rich region) and Walnut Creek were 8 in his 1991 study and 9 in his 1994 study, which are comparable to the findings of this study. Based on the formula that the authors have provided, other transit-rich regions such as New York or Los Angeles should have a multiplier with a range between 9 and 10. Overall, the models and multipliers developed in this study have external validity missing from earlier studies, including that of TCRP Report 176. This is due to the fact that household data were used (as the most disaggregate level of analysis) for 28 diverse regions across the United States. For instance, the authors of TCRP Report 176 used hundreds of UZAs as their unit of analysis (Gallivan et al. 2015). While there are some doubts about the global and single-level Direct Effect Indirect Effect Total Effect Multiplier Boston, MA −0.0137 −0.1171 −0.1308 9.5404 Seattle, WA −0.0139 −0.1159 −0.1298 9.3102 Portland, OR −0.0154 −0.1083 −0.1236 8.0470 Denver, CO −0.0158 −0.1057 −0.1215 7.6700 Atlanta, GA −0.0160 −0.1048 −0.1209 7.5515 Minneapolis–St. Paul, MN-WI −0.0163 −0.1033 −0.1196 7.3413 Eugene, OR −0.0163 −0.1031 −0.1195 7.3187 Miami, FL −0.0165 −0.1020 −0.1186 7.1716 Salt Lake City, UT −0.0166 −0.1019 −0.1185 7.1589 Madison, WI −0.0170 −0.0996 −0.1166 6.8657 San Antonio, TX −0.0171 −0.0992 −0.1162 6.8112 Houston, TX −0.0171 −0.0989 −0.1160 6.7781 Phoenix, AZ −0.0173 −0.0980 −0.1153 6.6663 Syracuse, NY −0.0174 −0.0975 −0.1149 6.6077 Greensboro, NC −0.0174 −0.0973 −0.1147 6.5883 Dallas, TX −0.0174 −0.0971 −0.1146 6.5674 Orlando, FL −0.0175 −0.0971 −0.1146 6.5656 Salem, OR −0.0175 −0.0970 −0.1145 6.5521 Rochester, NY −0.0175 −0.0970 −0.1144 6.5463 Albany, NY −0.0175 −0.0969 −0.1144 6.5376 Detroit, MI −0.0175 −0.0966 −0.1142 6.5062 Tampa, FL −0.0178 −0.0953 −0.1131 6.3551 Richmond, VA −0.0178 −0.0953 −0.1131 6.3536 Springfield, MA −0.0178 −0.0950 −0.1129 6.3240 Kansas City, MO −0.0180 −0.0942 −0.1122 6.2351 Winston-Salem, NC −0.0181 −0.0936 −0.1117 6.1683 Charleston, SC −0.0181 −0.0936 −0.1117 6.1633 Indianapolis, IN −0.0182 −0.0932 −0.1114 6.1285 Note that these values are slightly different than the values used in the GHG calculations, which were transit-agency based. In all cases, the transit multiplier is calculated as [direct effect + indirect effect] / direct effect. Table B-4. Transit multipliers for each of the 28 regions (sorted by the multiplier).

Transit Multiplier Methodology 57   model estimated in that report (which should not let the impact of land use on VMT vary across urbanized areas, yet the authors were somehow able to do so), the main drawback of it is related to aggregation bias, which might lead to the “ecological fallacy”—the conclusion that what is true for the group (i.e., UZAs) must be true for the subgroup (i.e., households). The authors of TCRP Report 176 then sought to validate their results by using household travel survey data. However, they only used data for nine regions and again developed a global model without any interaction term between regional and household variables. The output, hence, will not generate different land-use effects for different regions. Instead, the authors used the average value for all of the variables (by each region) to compute the land use effects, which is a less accurate method.