Transferability of Household Travel Data Across Neighborhood Types and Geographic Areas Using NHTS

Transcription

1 Transferability of Household Travel Data Across Neighborhood Types and Geographic Areas Using NHTS Jie Lin a* Liang Long b a Departent of Civil and Materials Engineering, & Institute of Environental Science and Policy, University of Illinois at Chicago, 842 W. Taylor St. (MC246), Chicago, IL b Departent of Civil and Materials Engineering, University of Illinois at Chicago, 842 W. Taylor St. (MC246), Chicago, IL * Corresponding author. Phone: Fax: Eail addresses: janelin@uic.edu (J. Lin), llong6@uic.edu (L. Long) 1

2 Abstract Travel data transferability is defined as the feasibility of transferring household travel attributes across geographic areas and the trip generation odeling built on the transferred data. Transferability has particular iportance to areas where there is no or little disaggregated inforation available for the area. The fundaental research question is can the data be transferred across areas or segents? In this paper, we propose a two-level rando coefficient odeling structure to study transferability across geographic areas. We show that transferability is a typical rando effects odeling effort. We then apply the odel fraewor to the sapled households in the 2001 national household travel survey (NHTS) across fourteen MSA/CMSAs with population greater than 3 illion. The odel results show that in general the variability of household auto wor trip rates across geographic areas can be ignored, with a few exceptions. The odel results also confir the influence of neighborhood specific features (e.g., housing density, worer density, intersection density, average auto wor trip travel tie) on travel behavior of the households living within. Keywords: data transferability, rando coefficient, hierarchical odel, fixed effects, rando effects, neighborhood type 2

3 1.1 Bacground Metropolitan areas with populations of over 50,000 are required to conduct transportation planning. In order to be eligible for federal funding, confority regulations require etropolitan planning organizations (MPOs) to have their odels calibrated on a continuing basis using new data. On the other hand, surveys are expensive instruents and the data required to support the planning process can becoe outdated. The 2001 National Household Travel Survey (NHTS) was designed to provide statistically valid estiates of trips rates and travel statistics at the Census Division level. Although the NHTS data is rich in inforation, it is not recoended for drawing conclusions on travel in areas saller than a Census Division. Extrapolating NHTS data within saller areas could ris developing and subsequently using unreliable estiates in planning process. As a result, any questions have been eerging with respect to the transferability of household travel surveys data. A nuber of recent studies addressing the data transferability issues have focused on ethods of updating or siulating local data. Those studies include Wilot and Stopher (2001), Greaves and Stopher (2000), and the Oa Ridge National Laboratory (Reuscher et al., 2002) using the 1995 Nationwide Person Travel Survey (NPTS) data. The ost coon ethods are Bayesian updating and Monte Carlo siulation. The general approach of studying data transferability involves dividing the sapled households into relatively hoogenous groups (or clusters) with respect to a set of household attributes and coparing the travel patterns or one step further to update or siulate the travel data across groups (Wilot and Stopher, 2001; Greaves and Stopher, 2000; Reuscher et al., 2002). Good clustering results provide a good basis for understanding the data transferability. The past studies have concentrated on internal paraeters related to individual households such as household incoe, size, nuber of ids, nuber of vehicles, etc. With increasing attention 3

4 paid to the iportance of external factors such as residential location and neighborhood lifestyle to household travel, the grouping of the households needs to tae into consideration of the neighborhood types where the households reside. Extending travel data transferability is the transferability of travel deand odels (or odel paraeters). The idea of odel transferability is to apply previously estiated odel paraeters to a new context (Karasaa, 2001). Model transferability has been priarily focused on trip generation and ode choice. The debates with respect to the feasibility of transferring travel deand odels have never stopped. The discussion focuses not only on feasibility of transferring odel in space and tie, but also on the odel specification and level of aggregation (Cotrus et al., 2005). A nuber of studies showed that spatial transferability of odels could reach reasonable satisfaction (Karasaa, 2001), while others did not arrive at the sae conclusion (Sith and Cleveland, 1976). Nonetheless literature tends to agree that disaggregated odels be ore transferable than aggregated odels. For exaple, Wilot (1995) found that disaggregate odels are preferable because the odel paraeters are not dependent on the zonal structure. Koppelan and Wilot (1986) investigated the effect of oission of variables on choice odel transferability and concluded that the inclusion of relevant variable leads to an absolute iproveent in transfer effectiveness. They concluded that iniu odel specification exists in order to achieve reasonable levels of odel transferability. 1.2 Forulation of Transferability Proble In this paper we define travel data transferability as the feasibility of transferring household travel attributes across geographic areas and the trip generation odeling built on the transferred data, which has the following for, 4

5 y Ks = β + β x + ε s,, i s,, 0 s,, s,,i, s,,i (1) =1 where, y s,,i = ith dependent variable for household (i=1,, N s, ) in hoogeneous group s (s=1,, S) and geographic area (=1,, M) β s,, = th odel coefficient of intercept (=0) or household attribute (=1, 2,, K s ) x s,,i, = th continuous 1 independent variable (household attribute) for household i in hoogeneous group s and geographic area ε s,,i = rando error ter for household i in hoogeneous group s and geographic area In the proble of household travel data transferability, independent variables, x s,,i, s, are generally unavailable (or very liited) for the study area. What is often available is the aggregated inforation of households and their travel at the zonal level (e.g., bloc group, census tract, or travel analysis zone). Ultiately, we hope to create a transferred dataset for the study area so that odel (1) can be constructed, i.e., to be able to estiate coefficients β s,, s. Suppose we have already clustered households into hoogeneous groups across geographic areas including the area of interest (yet no local inforation is available). The aggregated inforation at the group level is available for all of the areas let us call that bacground attributes associated with the group. Hence, for the area of interest the individual household attributes are not available but the bacground attributes of the group to which those households belong are nown. Then we as the question can we associate the dependent variable y s,,i with the bacground attributes? or can we estiate coefficients β s,, s based on 1 Continuous household attributes in this paper refer to non-duy or non-categorical variables including count data such as nuber of vehicles and nuber of ids. 5

6 the area s bacground attributes given that individual household attributes are generally not available?. For exaple, if household auto ownership is the priary deterinant of nuber of household auto trips for households in the areas belonging to a hoogeneous group, then it is reasonable to thin that nowing the household auto ownership level of a new area, which belongs to the sae group, will allow us to predict, to soe degree, the level of household auto trips for that area. If the predictions are accurate we say the odel (or ore accurately odel coefficients β s,, s) is transferable. This concept is, in fact, not new. In aret research, for exaple, preferences of consuers at new locations or for new segents, where detailed inforation is at first not available or very liited, are predicted through preference weights (i.e., β s,, s). The relationship between bacground attributes and preference weights is used to ae predictions in new geographic regions and/or segents and then copared with the locally estiated preference weight to deterine transferability (Uesh, 1987). Expressed in atheatical ters, if odel (1) coefficient β s,, can be estiated with the bacground attributes of group s and geographic area, i.e., Q s,,= γs,, + q =1 β 0 γ w + υ (2) s,,q s,,,q s,, where, w s,,,q = qth bacground attribute associated with coefficient β s,, (=1, 2,, K s ) of group s and geographic area γ s,,q = qth weight for intercept (q=0) or bacground attribute (q=1,2,, Q) υ s,, = rando error ter associated with coefficient β s,, of group s and geographic area 6

7 then the dependent variable y s,,i can be associated with the bacground attributes in the following way, y s,, i = γ s, 0,0 + Q q = 1 γ s, 0,q w s,,0, q + Ks γ + Q s,, 0 = 1 q = 1 υ γ s,, 0 s,,q + w Ks = 1 s,,,q υ s,, x x s,, i, s,,i, + + ε s,,i (3) Rewriting equations (1) (3) in atrix for, we have, Y s, = X s, β s, + ε s, (4) β s, = W s, γ s + υ s, (5) and, Y s, = X s, W s, γ s + (X s, υ s, + ε s, ) (6) where, Y s, is the (N s, x 1) dependent variable colun vector = (y s,,1, y s,,2,, y s,,i, y s,,ns, ) T ; X s, is the N s, x (K s +1) covariate atrix (with unities in the first colun); β s, is the (K s +1) x 1 coefficient vector = (β s,,0, β s,,1,., β s,,,, β s,,ks ) T ; W s, is the (K s +1) x (Q+1) bacground covariate atrix (with unities in the first colun); γ s is the (Q+1) x 1 weight vector = (γ s,,0, γ s,,1,, γ s,,q,, γ s,,q ) T ; and ε s, (N s, x 1) and υ s, ((K s +1) x 1 ) are rando error vectors. Equations (1)-(3) (or (4)-(5)) define the so-called rando coefficient odel because of the rando error ters υ s,, s. Rando coefficient odel is one ind of hierarchical odels or ultilevel odels because equation (2) represents the upper level regression odel defined by (s, ) and equation (1) represents the lower level regression odel defined by individual i. Specifically to this study, for any group s, we define two-level odeling structure, i.e., M geographic areas (level 2) and N s, households within each area (level 1), as shown in Figure 1. [Insert Figure 1 here] 7

8 Hierarchical odeling is widely used in edicine and epideiological research (e.g., Sullivan et al., 1999; Greenland, 2000a & b; Burgess Jr. et al., 2000), econoics (e.g., Nunes Aaral et al., 1997; Goodan and Thibodeau, 1998), and educational, social, and behavioral sciences (e.g., Hox and Kreft, 1994; Kreft, 1995; Singer, 1998). Hierarchical odeling represents a suite of odeling techniques, for exaple Bayesian (and epirical-bayes), penalized lielihood, ixed-odel, rando coefficient regression, and variance-coponent analysis of nested experiental design (Bry and Raudenbush, 1992; Searle et al., 1992; Goldan, 1995; Gelan et al., 1995; Kreft and de Leeuw, 1998; Steel et al., 1997). In the odel definition (1)-(3) or (4)-(6), the following properties of ε s, and υ s, are assued. 2 εs, ~ N( 0, R), where R = σ I (7) τ 0,0 L τ 0, L τ 0, K s υ = s, ~ N( 0, G), where G M L τ, L M τ Ks,0 L τ Ks, L τ Ks, Ks (8) cov( ε s,, υ s, ) = 0 (9) Vectors ε s, and υ s, are both norally distributed with ean equal to zero and covariance atrix R and G, respectively. Eleents within vector ε s, are independent with the equal variance (σ 2 ). Covariance of two eleents (υ s,,, υ s,,l ) is τ,l. Rando error vectors ε s, and υ s, are independent. In this study we consider the rando coefficient odels separately for each group s as illustrated in Figure 1, and thus the subscript s can be dropped for convenience in the rest of the discussion. If we include duy variables then the odel has the following final for, 8

9 y Ks, i β,0 + β, x, i, + = 1 j = α z + ε (10), j, i, j, i Q β, = γ,0 + γ, q w,, q + υ, q= 1 (11) y, i = γ 0,0 + q γ 0, q w j,0, q α +, j z γ,0 + γ, q w q + υ + υ, i, j,0,,, q x x, i, + + ε, i,, i (12) or in atrix for, Y = X β + Z α + ε (13) β = W γ + υ (14) Y = X W γ + Z α + (X υ + ε ) (15) where, z,i,j = jth duy variable (j = 1,, J) for household i in geographic area (and hoogeneous group s) α,j = coefficient for duy variable z,i,j. Z = duy variable atrix (N x J) α = coefficient vector (J x 1) Equation (15) (or (12)) contains two types of effects, fixed effects defined by (X W γ + Z α ) and rando effects defined by X υ. Fixed effects represent the deterinistic, average relationships between the dependent variable and covariates (i.e., average effects). Rando effects represent the deviations of fixed effects (i.e., deviation fro the average effects). Recalling the definition of transferability in this study, we will be able to infer the transferability by conducting hypothesis tests of the odel s fixed and rando effects. 9

10 1.3 Hypotheses There are three basic hypothesis tests of a hierarchical odel related to the fixed effects, rando effects, and covariance coponents. a. Hypothesis tests for fixed effects (γ,q and α,j ) The null hypothesis for fixed effects is of the for H 0 : γ,q =0 and α,j =0, for,,q. That is, the fixed effects of the covariates on the dependent variable are zero if the null hypothesis is accepted; there are statistically significant fixed effects if the null hypothesis is rejected. The following statistic follows the t-distribution and thus t test is used. t ˆ γ, q = (or var( ˆ γ ), q ˆ α, j var( ˆ α, j ) (16) ) where ˆ, αˆ γ q and, j are estiates of γ,q and α,j respectively. Alternatively, fixed effects of the odel coefficients can be tested using F test, whose statistic values equal the square of t-statistics. b. Hypothesis tests for rando effects (υ, ) The null hypothesis for rando effects is of the for H 0 : υ, = 0, for,. If the null hypothesis is accepted, there are no deviations fro the fixed effects across all areas. We then consider the household travel data (and the deand odels) are transferable. Siilar to the fixed effects, t-test is applied to the following statistic, t ˆ υ, = (17) var( ˆ υ ), where υ ˆ is an estiate of υ,,. 10

11 c. Hypothesis tests for covariance coponents (R and G) The null hypothesis H 0 : τ,l = 0 tests the goodness-of-fit of the hierarchical odel. The axiu lielihood estiates (ML) (when saple size is relatively sall) and restricted axiu lielihood (REML) estiates (when saple size is relatively large) of τ,l s are found by axiizing the following log lielihood function (Sullivan et al., 1999; see also Littell et al., 1996; Searle et al., 1992). + = N N r r N l T ML 2π log 1 2 log 2 log 2 1 ) ( 1 Σ Σ G R, (18) + = ) ( 2π log 1 2 ) ( log 2 ) ( log 2 1 log 2 1 ) ( p N p N r r p N W X X W l T T T REML 1 1 Σ Σ Σ R,G (19) where, var( ) Y = Σ (19) ( ) T T T T T Y X W W X X W W X Y r 1 1 = Σ Σ (20) p = ran (X W ) (21) and, (22) = M N N 1 It has been shown that the following test statistic follows Wald Z test asyptotically (Littell et al, 1996), ) var(ˆ ˆ,, l l τ τ Z = (23) Alternatively, the following statistic follows the Chi-square distribution, ( ) = = M 2 2, 2 1 M N σ γ β χ 1 0,0 0 ˆ ˆ ˆ (24) 11

12 1.4 Data Both Census Transportation Planning Pacage (CTPP) 2000 and the 2001 National Household Travel Survey (NHTS) data are used in this study. CTPP is a set of special tabulations of household/individual and coute trip attributes derived fro the long for of decennial. CTPP has three parts of surveyed inforation for place of residence (Part I), place of wor (Part II), and journey to wor (Part III). The data is suarized at different geographic levels including state, county, census tract, bloc group, etropolitan statistical area (MSA), consolidated etropolitan statistical area (CMSA), and traffic analysis zone (TAZ). This study extracts the census tract level attributes fro the CTPP Part I dataset, which contains one hundred and twenty different tabulations of attributes of households located in 65,315 census tracts throughout fifty states and Washington D.C. Table 1 suarizes the final sixty-four tract attributes assebled. The sixty-four variables are categorized as sociodeographic, land use, and journey-to-wor features. The socio-deographic attributes are age, sex, occupation, race/ethnics, household incoe, household size, industry, vehicle counts, and nuber of households, etc. Journey-to-wor features are travel tie to wor and eans of transportation to wor. The land use attributes such as road density and intersection density were derived fro the census Tiger/line shape file. [Insert Table 1 here] The 2001 NHTS is used to obtain individual household attributes and travel characteristics. NHTS consists of five datasets of households, persons, travel day trips, vehicles, and long-distance trips, respectively. Long distance trips were not considered in this study. There are a total of 69,817 household observations nationally, of which over 40,000 households were 12

13 fro the nine NHTS add-on areas 2 and the rest were fro the national saple. While the total nuber of observations is large, ajority of the census tracts have few (often single-digit) household observations. NHTS has various weights developed to account for selection probability and non-response of households, persons, trips and vehicles (Federal Highway Adinistration, 2004). Details and highlights of the 2001 NHTS socioeconoics and household travel can be found in Pucher and Renne (2003) and Bureau of Transportation Statistics report (2003). 1.5 Research Design To test transferability using the hierarchical odel, we design the following study. First, we define the hoogeneous groups of households in ters of neighborhood types using the tract attributes extracted fro CTPP shown in Table 1. After the neighborhood types are defined, the saple households fro the NHTS dataset are associated with one of the neighborhood types accordingly. For each neighborhood type, we construct the hierarchical odels across MSA/CMSAs and tested the hypotheses. We then infer the transferability of household travel data based on the odel results and hypothesis tests. In this study, we only consider those MSA/CMSAs with population greater than 3 illion, which resulted in fourteen MSA/CMSAs (Table 2). The household travel easure odeled is individual household nuber of autoobile wor trips, including to wor, worrelated, to school, and school related trips. [Insert Table 2 here] 2 The nine add-ons are: Baltiore, MD, Des Moines, IA, Edonson, Carter, Pulasi, and Scott Counties, KY, Lancaster, PA, Oahu, HI, State of Hawaii except Oahu, State of New Yor, State of Texas, and State of Wisconsin 13

14 1.5.1 Defining neighborhood types using CTPP data We have applied a statistical clustering ethod coupled with geographic inforation syste (GIS) spatial analysis to classifying census tracts into classes that define the types of neighborhood. This step was accoplished by using the sixty-four variables in Table 1. A neighborhood in this study is thus identical to a census tract 3. The detailed clustering ethod has been discussed in Lin and Long (2006) and will not be repeated here. As a result, ten neighborhood types were deterined to best categorize the census tracts throughout 50 states and Washington D.C. The clustering results will be presented later in the results section Hierarchical odeling For each of the ten neighborhood types, we construct the hierarchical odels defined by equations (10) (12) or (13) (15) using the household observations across the fourteen MSA/CMSA areas listed in Table 2. In this study, the dependent variable y,i of interest is the square root nuber of autoobile trips of household i in MSA/CMSA. The transforation was the result of the error ter exaination. Bacground variables w,,q s are neighborhood (tract) attributes fro Table 1. Table 3 is the list of the assebled continuous household variables x,i, s and duy variables z,i,j s. Household variables x,i, s are non duy household characteristics fro the NHTS household file. Those attributes are household size, nuber of ids, nuber of vehicles owned by the household, nuber of worers of the household, etc. Duy variables z,i,j s are race (White or non-white), with college degree or without, with low incoe or not, and hoe owner or not. 3 We have also applied exactly the sae ethod using the sae data and variables at the bloc group level (the sallest zonal classification in CTPP) and obtained the very siilar results as that at the census tract level. This result suggests the robustness of the clustering ethod used. The reason we did not use the bloc group clusters is because only eleven states in CTPP 2000 contain bloc group classification. 14

15 [Insert Table 3 here] 1.6 Results Neighborhood type Table 4 is the final clusters (i.e., neighborhood types) of the census tracts and households. Table 5 suarizes the priary features of the ten neighborhood types. Each neighborhood type is naed to reflect the priary features of the inclusive census tracts. Each neighborhood type represents certain lifestyle defined by socio-econoics, deographics, and geographic/land use characteristics. For exaple, Urban elite represents an urban neighborhood with priarily young non-hispanic White woring in professional, anagerial or technical fields and earning $45,000 and ore. Copared to the suburban households (e.g., suburban id-age wealthy), any of the urban elites (17.7%) use transit to wor and have sall household size (an average of 1.91 persons per household) suggesting any of the are single or arried with no children. Places lie Boston, North Chicago (e.g., Lincoln Par), and New Yor City (e.g., Lower Manhattan) (see Figure 2) are the typical exaples of urban elite neighborhoods. [Insert Tables 4 & 5] [Insert Figure 2] Figure 3 is the national display of the ten neighborhood types. Urban and suburban census tracts heavily concentrate in the Northeastern coastal, West coast, and the Great Laes. Natural scenic attractions include the Grand Canyon, Death Valley, Yellowstone, Alasa, and Hawaii. Non-Blac Hispanics are seen in south and west of Texas, Central Valley, California, and South Florida. Costal Florida is the popular choice for the retirees. The clustering result also reveals racial residential separation in Aerica. Urban and suburban census tracts are predoinantly White. Central cities are concentrated with poor non-hispanic Blacs and other 15

16 inorities (Figure 2). Non-Blac Hispanic neighborhoods can be seen in cities, suburbs and rural. [Insert Figure 3 here] Each of the ten neighborhood types uniquely defines a lifestyle. An urban elite and a suburban id-age wealthy neighborhood ay possess siilar socio-econoic characteristics, such as occupation coposition and incoe levels, they differ, however, by life cycle (e.g., one adult with no children, two adult with children of different ages, etc.) and residential location, which in turn affect their odes of transportation. For exaple, if an urban elite household later oves to a suburban wealthy neighborhood for ore space and better education for their children after starting a faily, its lifestyle changes. So does its travel behavior. The travel characteristics of the ten neighborhood types have been discussed in length in Lin and Long (2006). Recent studies have even suggested the endogeneity of residential location choice to the relations between travel and land use (e.g., Bagley and Mohtarian, 2002; Boarnet and Crane, 2001). That is, attitudes towards lifestyle and travel choice are found to be the deterinants of residential location, travel distances and ode choice Hierarchical odeling and hypothesis test results Three neighborhood types, rural, natural scenic, and Hispanic doinant, were excluded fro the study because of the very sall saple sizes fro the 14 MSA/CMSAs considered. For the reaining seven neighborhood types, hierarchical odels were constructed separately. Here we present only three of the seven odels for suburban retired, urban elite and suburban wealthy in Table 6 as exaples. Note that these odels are all weighted to the entire population. 16

17 Hierarchical odel fixed effects [Insert Table 6 here] a. Suburban retired The suburban retired odel contains two neighborhood covariates, housing density and percent population with age 75 years and older, one continuous (nuber of ids) and two duy household variables (low incoe and owned house). The signs of the coefficients are intuitively correct. A negative housing density coefficient suggests a higher housing density of the neighborhood produces fewer household auto wor trips. Percent population with age 75 years and older has a negative, albeit insignificant, coefficient, which eans nuber of auto wor trips decreases as tact level percent population with 75 years and older increases. The coefficients of the two duy variables low incoe (-) and owned house (+) are also statistically significant, suggesting low incoe households ae less auto wor trips and hoe owning households tend to ae ore auto wor trips. The interaction ters, housing density*nuber of ids and percent population with age 75 years and old*nuber of ids, represent how the neighborhood and household covariates are correlated. Housing density of the area and nuber of ids per household in the area are virtually non-correlated (p value of ), which aes intuitive sense there is no evidence that the authors are aware of between how any of ids a household has and the housing density of the neighborhood. On the other hand, percent population with age 75 years and old in the neighborhood is significantly and positively correlated with the nuber of ids per household, the reason of which is arguably self-explanatory. b. Urban elite 17

18 In this odel, there are three neighborhood covariates (worer density, intersection density, and average auto wor trip travel tie), two household level covariates (household size and household nuber of vehicles), and six subsequent interaction ters. Aong the three neighborhood covariates, worer density shows positive effect on nuber of auto wor trips (significant at the 10% level). Interestingly intersection density shows significant, negative effect, suggesting that high intersection density in this urban neighborhood type be associated with few auto wor trips. Average auto wor trip travel tie has negative (albeit insignificant) effect on auto wor trips, eaning that high average auto wor trip travel tie ay result in fewer nuber of auto wor trips (for exaple switching to other odes of transportation to wor). Larger household size is shown to be associated with higher nuber of auto wor trips. Household auto ownership is an insignificant covariate to household auto wor trips for urban elite households, which is an interesting finding. It sees to reflect the less auto-dependent travel behavior of those urban elite households and their urban lifestyle preferences. Aong the six interaction ters, household size is positively correlated with neighborhood worer density but negatively correlated with average auto wor trip travel tie, which tends to suggest that saller urban elite households (e.g., urban single professionals) live away (in ters of travel tie) fro wor copared to larger urban elite households (e.g., urban arried professionals). Household nuber of vehicles is negatively correlated with neighborhood worer density and positively correlated with neighborhood intersection density and average auto wor trip travel tie. A possible explanation is that, for urban elite neighborhoods, high urban intersection density ay be associated with urban residential environent as opposed to non-residential environent, and thus is positively correlated with 18

19 household nuber of vehicles. The positive correlation between household nuber of vehicles and average travel tie for auto wor trip ay suggest autoobile becoes the preferred ode of transportation when coute tie is long. c. Suburban id-age wealthy The only significant neighborhood characteristic is average auto wor trip travel tie, which is positively correlated with household nuber of auto wor trips. This ay see counter intuitive at first. However, this relationship ay siply reflect the lac of (or no easy access to) other coute options for these suburban residents. There ay be other confounding factors not accounted in the odel, such as job-housing tradeoffs. Not surprisingly, the two household covariates, nuber of ids and nuber of household vehicles, are significant, positive covariates. The low incoe duy variable is significantly but negatively correlated with the household auto wor trips, saying that low incoe households ade fewer auto wor trips. Household nuber of ids is positively (albeit insignificantly) correlated with neighborhood average auto wor trip travel tie. This ay suggest the location trade-offs aong job, housing and school. Higher nuber of household owned vehicles ay indicate better obility of the household and result in shorter average auto wor trip travel tie Rando effects Tables 7 9 are the rando effect tests for the three odels. A positive (negative) estiate eans an upward (downward) deviation fro the fixed effect of the sae covariate. If the estiate is statistically significant, the associated rando effect is significantly different fro zero, which eans the total effect of the covariate deviates statistically fro the average effect (i.e., fixed 19

20 effect) seen in Table 6. For exaple, in Table 7, the rando effect of the intercept and nuber of ids for the households in Atlanta, GA, MSA are both positive ( and respectively), eaning that there are additional and units of effects by the intercept and nuber of ids on the dependent variable (i.e., household nuber of auto wor trips). However, neither of the is statistically significant. In other words, the households in the suburban retired neighborhoods of the Atlanta MSA are of no difference fro the average households in the sae neighborhood of the 14 MSA/CMSAs, in ters of household nuber of auto wor trips. For suburban retired neighborhoods, only Detroit CMSA, Phoenix CMSA, and San Francisco CMSA are showing significant difference fro the average (Table 7). In particular, nuber of ids of a household in Phoenix CMSA has units lower than the average of effects (-) while that of San Francisco CMSA is higher than the average (+), suggesting a bigger (+) or saller (-) effect of the household nuber of ids on auto wor trips than that of the average households in the sae neighborhood across the 14 MSA/CMSAs. [Insert Table 7 here] For households in urban elite neighborhoods, New Yor CMSA stands out as expected (see Table 8). Its household size has larger effect than average and its household nuber of vehicles has less effect than average on household nuber of auto wor trips. That is to say, for the young professional New Yorers, household vehicles are less associated with auto wor trips. [Insert Table 8 here] For suburban id-age wealthy neighborhoods, Phoenix, San Francisco, and Seattle CMSAs are different fro the rest (Table 9). That is, nuber of ids atters ore to households 20

21 in San Francisco and Seattle CMSAs and less to households in Phoenix CMSA. The odel does not reveal the causality as to why that is. [Insert Table 9 here] In ost cases, all three odels suggest that households living in the sae neighborhood type tend to ae siilar nuber of auto wor trips within the siilar size MSA/CMSAs, whether they live in the San Francisco Bay Area, greater Boston, or Chicago etropolis, or whether they live in Detroit, MI or Miai, FL. This sees to suggest household auto wor trips (and consequently trip generation odel coefficients) are transferable across those regions. 1.7 Conclusions Household travel data transferability has iportant practical values to MPOs in regional transportation planning. Transferability issue eerges when there is no or little disaggregated inforation available. The past data transferability studies have focused on ethods to update (or transfer) the local household travel data using data available fro other regions or sources (e.g., NPTS or NHTS) (Wilot and Stopher, 2001; Greaves and Stopher, 2000; Reuscher and Schoyer, 2002). However, a basic research question of transferability has been overlooed in those studies. That is, is it feasible to transfer travel data across geographic areas? In this paper, we have showed that transferability can be forulated into a two-level rando coefficient odel and thus transferability can be statistically tested. By allowing rando effects in the regression odel coefficients, we are able to test odel variability across geographic areas or segents and infer transferability of travel data and the (trip generation) odels built on the data. Furtherore, the rando coefficient odel structure allows us to incorporate environental covariates related to neighborhood characteristics, which have been shown to have significant influence on household travel. In other words, we argue that 21

22 transferability is affected not only by the internal household characteristics but also the surrounding environent. The case study of transferability across the fourteen MSA/CMSAs showed that in general the odel variability across geographic areas can be ignored. But there are exceptions. For exaple, households in urban elite neighborhoods within the New Yor CMSA shows highly distinctive travel behavior (in ters of nuber of auto wor trips) fro the rest of the households fro the other thirteen MSA/CMSAs in the sae neighborhood type. The odel results also see to suggest that suburban households (both in suburban retired and wealthy neighborhoods exained) in Phoenix and San Francisco CMSAs show different patterns fro the rest. The odel results also confir the influence of neighborhood specific features on travel behavior of the households living within For exaple, housing density and percent population with age 75 years and older to suburban retired neighborhoods, worer density and intersection density to urban elite neighborhoods, and average auto wor trip travel tie to suburban id-age wealthy neighborhoods. On the other hand, it is iportant to recognize the liitations of this study. We assebled a set of household and neighborhood variables available to us for the odeling. However, we do not intend to conclude that the variables, especially the neighborhood ones, have fully characterized the households or neighborhoods. Other features, e.g., proxiity to highway/transit and ore detailed land use categorization, will be of great interest. The hierarchical odel is useful in revealing the interactions of variables in a hierarchical structure. However, it is iportant to eep in ind when interpreting the odel results that the odel does not explain the causal relationships between the dependent variable and covariates. Finally, only household auto wor trip rates were exained. It is reasonable to assue that household wor trips are less 22

23 variable after controlling for household characteristics. Hence, the conclusion that household auto wor trip rate is transferable is ore or less expected. Different conclusions ay be drawn for other easures such as ode share and nuber of shopping trips. These will be the future research. 1.8 Acnowledgeent This study is part of the transferability study funded by the Federal Highway Adinistration (FHWA). 1.9 References Bagley, M. N., and P. L. Mohtarian (2002) The Ipact of Residential Neighborhood Type on Travel Behavior: A Structural Equations Modeling Approach. The Annals of Regional Science 36, pp Boarnet, M.G. and R. Crane (2001) Travel by Design: the Influence of Urban For on Travel. Oxford University Press, Oxford, UK Bry, A.S., S.W. Raudenbush (1992) Hierarchical Linear Models. Sage, Newbury Par, CA Bureau of Transportation Statistics (2003) NHTS 2001 Highlights Report. BTS U.S. Departent of Transportation, Washington D.C. Burgess Jr., J.F., C.L. Christiansen, S.E. Michala and C.N. Morris (2000) Medical profiling: iproving standards and ris adjustents using hierarchical odels, Journal of Health Econoics, Volue 19, Issue 3,

24 Cotrus, A.V., J.N. Prasher, Y. Shiftan (2005) Spatial and Teporal Transferability of Trip Generation Deand Models in Israel. Journal of Transportation and Statistics, Vol. 8, No. 1, ISSN Federal Highway Adinistration (2004) 2001 National Household Travel Survey: User s Guide. U.S. Departent of Transportation, Washington D.C., June Gelan, A., J.B. Carlin, H.S. Stern, D.B. Rubin (1995) Bayesian Data Analysis. Chapan and Hall, New Yor Goldan, H. (1995) Multilevel Statistical Models, 2 nd edition. Edward Arnold, London, UK Goodan, A.C. and T.G. Thibodeau, Housing Maret Segentation, Journal of Housing Econoics, Volue 7, Issue 2, Greaves, S.P. and P.R. Stopher (2000) Creating a Siulated Household Travel/Activity Survey- Rationale and Feasibility Analysis. Transportation Research Record 1706, Transportation Research Board, Washington D.C., pp Greenland, S. (2000a) When Should Epideiologic Regressions Use Rando Coefficients? Bioetrics 56, Greenland, S. (2000b) Principles of ultilevel odeling. International Journal of Epideiology 29: Hox, J.J., Kreft, I.G.G. (Eds.) (1994) Multilevel analysis ethods (special issue), Sociological Methods and Research, 22 (3) 24

25 Karasaa, N. (2001) The Spatial Transferability of the Helsini Metropolitan Area Mode Choice Models. Paper presented at the 5 th worshop of the Nordic Research Networ on Modeling Transport, Land-use, and the Environent, Septeber 28-30, Nynashan, Sweden Koppelan, F.S., and Wilot, C.G. (1986) The Effect of Oission of Variables on Choice Model Transferability. Transportation Research, Part B: Methodological, Vol. 20 B, No. 3, pp Kreft, I.G.G. (Ed.) (1995) Hierarchical linear odels: probles and prospects (special issue). Journal of Educational and Behavioral Statistics, 20(2) Kreft, I.G.G., J. de Leeuw (1998) Introducing Multilevel Modelling. Sage, London, UK Lin, J. and L. Long (2006) What Neighborhood Are You In? Epirical Findings of Relationships between Residential Location, Lifestyle, and Travel, Copendiu CD ROM of the 85 th Annual Meeting of Transportation Research Board, Washington D.C., January 22-26, 2006 Littell, R.C., G.A. Millien, W.W. Stroup and R.D. Wolfinger (1996) SAS Syste for Mixed Models, SAS Institute Inc., Cary, NC Nunes Aaral, L.A., S.V. Buldyrev, S. Havlin, P. Maass, M. A. Salinger, H.E. Stanley and M.H.R. Stanley (1997) Scaling behavior in econoics: The proble of quantifying copany growth, Physica A: Statistical and Theoretical Physics, Volue 244, Issues 1-4, Pages 1-24 Pucher, J. and J.L. Renne (2003) Socioeconoics of Urban Travel: Evidence fro the 2001 NHTS. Transportation Quarterly, Vol. 57, No. 3, pp

26 Reuscher, T.R., R.L. Schoyer, Jr., and P.S. Hu (2002) Transferability of Nationwide Personal Transportation Survey Data to Regional and Local Scales. Transportation Research Record 1817, Transportation Research Board, Washington D.C., pp Searle, S.R., G. Casella and C.E. McCulloch (1992) Variance Coponents, Wiley, New Yor Singer, J.D. (1998) Using SAS PROC MIXED to Fit Multilevel Models, Hierarchical Models, and Individual Growth Models. Journal of Educational and Behavioral Statistics, Vol. 24, No. 4, pp Sith, R.L. and D.E. Cleveland (1976) Tie Stability Analysis of Trip Generation and Predistribution Modal Choice Models. Transportation Research Record 569, Transportation Research Board, Washington D.C., pp Steel, R.G.D., J.H. Torrie, D.A. Dicey (1997) Principles and Procedures of Statistics A Bioetrical Approach, 3 rd Edition. WCB/McGraw-Hill, ISBN Sullivan, L.M., K.A. Dues, E. Losina (1999) Tutorial in Biostatistics An Introduction to Hierarchical Linear Modelling. Statistics in Medicine 18, Uesh, U.N. (1987) Transferability of Preference Models across segents and Geographic Areas. Journal of Mareting, Vol. 51, No. 1, Wilot, C.G. (1995) Evidence on transferability of trip-generation odels. Journal of Transportation Engineering, Vol. 121, pp

27 Wilot, C.G., and P.R. Stopher (2001) Transferability of Transportation Planning Data. Transportation Research Record 1768, Transportation Research Board, Washington D.C., pp

28 Table 1 Sixty-four census-tract level variables used for clustering Variable Definition Sociodeographics of a census tract Average household size Average nuber of household ebers Eployent rate Ratio of nuber of worers per household to household size Vehicle count Average nuber of vehicles per household Race/ethnics (8) Percents of households in 8 racial/ethnic groups: Incoe (15) Hoe tenure status (2) Age (6) Resident industrial type (14) Occupation (5) 1 =White Hispanic 2 =White non-hispanic 3 =Blac Hispanic 4 =Blac non-hispanic Percents of households in 15 incoe levels: 01 = <$5, =$5,000-$9, =$10,000-14, =$15,000-19, =$20,000-24, =$25,000-29, =$30,000-34, =$35,000-39,999 Percent of hoes owned or rent Percents of persons in 6 age groups: 1 = under 16 years 2 =16-24 years 3 =25-44 years 5 =Asian Hispanic 6 =Asian non-hispanic 7 =Other Hispanic 8 =Other non-hispanic 09 =$40,000-$44, =$45,000-49, =$50,000-54, =$55,000-59, =$60,000-74, =$75,000-99, = >$100,000 4 =45-64 years 5 =65-74 years 6 =75 years and older Percents of worers in 14 industries: 1 =agriculture, forestry, fishing and hunting, and ining 2 =construction; 3 =anufacturing; 4 =wholesale trade; 5 =retail trade; 6 =transportation and warehousing, and utilities; 7 =inforation; 8 =Finance, insurance, real estate and rental and leasing; 9 =Professional, scientific, anageent, adinistrative, and waste anageent services; 10 =educational, health and social services; 11 =Arts, entertainent, recreation, accoodation and food services; 12 =Other services (except public adinistration); 13 =Public adinistration; 14 =Ared forces Percents of worers in 5 occupation categories: 01 =Sales or service; 02 =Clerical or adinistrative support; 03 =Manufacturing, construction, aintenance, or faring; 04 =Professional, anagerial or technical; 91 =Other Land use features of a census tract Population density Per sq ile population Job density Per sq ile nuber of jobs Housing density Per sq ile housing units Road density Per sq ile road length (in iles) Intersection density Per sq ile nuber of intersections Journey-to-wor features of a census tract Auto usage Fraction of worers using autoobile Transit usage Fraction of worers using transit Auto users Nuber of worers using autos to wor Transit users Nuber of worers using transit to wor Travel tie to wor (2) By transit or autoobile (in inutes)

29 Table 2 MSA/CMSA s in NHTS with population greater than 3 illion NHTS MSA/CMSA code MSA/CMSA nae 0520 Atlanta, GA 1122 Boston-Worcester-Lawrence, MA-NH-ME-CT 1602 Chicago-Gary-Kenosha, IL-IN-WI 1922 Dallas-Fort Worth, TX 2162 Detroit-Ann Arbor-Flint, MI 3362 Houston-Galveston-Brazoria, TX 4472 Los Angeles-Riverside-Orange County, CA 4992 Miai-Fort Lauderdale, FL 5602 New Yor- Northern New Jersey-Long Island, NY-NJ-CT-PA 6162 Philadelphia-Wilington-Atlantic City, PA-NJ-DE-MD 6200 Phoenix-Mesa, AZ 7362 San Francisco-Oaland-San Jose, CA 7602 Seattle-Tacoa-Breerton, WA 8872 Washington-Baltiore, DC-MD-VA-WV Table 3 Suary of household attributes considered Nae of the variable HHSIZE HHVEHCNT DRVRCNT NUMKID NUMADLT LOWINC Description Household size Household nuber of vehicles Household nuber of drivers Household nuber of ids Household nuber of adults Household is a low incoe household (LOWINC=1, i.e., household incoe less than $45,000 a year), otherwise 0 WHITE Household head is White (WHITE=1), otherwise 0 COLLEGE Household head has college degree (COLLEGE=1), otherwise 0 N_HOWN Household owns a house (N_HOWN=1), otherwise 0 Table 4 Saple sizes of ten neighborhood types # Census % of # census % of # NHTS % of Neighborhood type Tracts Total households Total Households Total Urban elite 2, ,279, , Urban/2 nd non-hispanic Blac doinant 5, ,974, , City low incoe, priarily inority 2, ,394, , Suburban id-incoe woring class 10, ,056, , Suburban id-age wealthy 8, ,666, , Suburban young 8, ,279, , Suburban retired 7, ,746, , Rural 14, ,779, , Natural Scenic 1, ,134, Non-Blac Hispanic doinant 4, ,226, , Valid Cases 64, ,537, , Excluded Cases Total 65, ,537, ,

30 Table 5 Neighborhood characteristics fro ten clusters Urban elite Urban/2 nd city poor, non-hisp Blac doinant City low incoe, priarily inority Suburban id-incoe woring class Pop density 13,547 7,394 55,644 Housing density 7,236 3,163 22,947 Job density 7,562 2,336 21,749 Road density Intersection density Transit usage 17.7% 14.2% 50.7% Auto usage 68.2% Resident Edu/Progn/arts industrial type(14) 1 (49.7%) Occupation(4) 2 Professional (51.1%) Suburban wealthy Suburban young Suburban retired Rural 1,477 3,018 4,952 3, ,112 2,079 1, ,508 2,252 1, Natural Scenic Non-Blac Hispanic doinant 4,334 9,870 1,197 3,024 1,110 3, % 4.1% 3.2% 3.0% 0.8% 7.7% 7.2% 77.6% 36.4% 93.9% 89.4% 90.2% 90.6% 92% 68.1% 85.5% Edusoc/Manusoc/ad/fi- Edu- Manu/Retail Edu/Pro/Fi- Retail/Ac- Health/Re- Agr/Manu Rec/Ared/ Manu/food/ /Const nance co/cons tail/rec / Retail Edu-soc Const enter nance (36.5%) (45.7%) t (30.7%) (43.0%) (38%) (52.2%) (36.4%) (47.7%) (44.9%) Sales/ service (24.5%) Professional (30.5%) Professional (51.5%) Sales/ service (24.7%) Professional (38.8%) Sales/ service (26.6%) Manuconstfar (26.1%) Manuconstfar (34.7%) Manuconst-far (37.3%) Age(6) (43.0%) 24 (40.9%) (47.6%) (55.1%) (26.5%) (46.4%) 65+ (21.1%) (24.1%) (37.8%) 44 (76.3%) HH Incoe(15) 2 $45,000 (50.9%) $24,999 (55.6%) $34,999 (53.1%) $40,000-99,999 $75,000 (53.8%) $10,000-44,999 $45,000 (51.9%) $5,000-39,999 $29,999 (55.8%) $29,999 (51.6%) (52.9%) (53.2%) (55.4%) HH Race/ethnics(8) White non- Hisp (71.7%) Blac non-hisp (69.6%) Non White (62.8%) White non- Hisp (88.5%) White non- Hisp (84.5%) White non-hisp (71.5%) White non- Hisp (87.6%) White non-hisp (88.6%) White non- Hisp (61.1%) (White- & Othr-) Hisp (57.6%) HH size HH vehicles Hoe owned 36% 44.3% 31.7% 82.5% 85.4% 54.6% 69.7% 77.5% 34.3% 49.4% 1. Top three categories above national averages are considered 2. Highest percent Age/occupational/Incoe categories above national averages at the census tract level

31 Table 6 Fixed effects (dependent variable: square root of household vehicle wor trips) y ˆ, i = ˆ γ 0,0 + ˆ γ 0, q w,0, q + ˆ γ,0 + ˆ γ, q w,, q x, i, + ˆ α, j z, i, j q q j Effect Estiate Standard Error t value Pr > t a. Suburban retired (obs = 759) -2 Res Log Lielihood = Intercept < Housing density (w 1 ) Percent population with age 75 years and older (w 2 ) Nuber of ids (x 1 ) Housing density * nuber of ids (w 1 *x 1 ) 1.018E Percent population with age 75 years and older*nuber of ids (w 2 *x 1 ) Low incoe (z 1 ) Owned house (z 2 ) b. Urban elite (obs = 552) -2 Res Log Lielihood = Intercept < Worer density (w 1 ) E Intersection density (w 2 ) Average auto wor trip travel tie (w 3 ) Household size (x 1 ) Household nuber of vehicles (x 2 ) Worer density*household size (w 1 *x 1 ) 6.297E E Intersection density*household size (w 2 *x 1 ) Average auto wor trip travel tie*household size (w 3 *x 1 ) Worer density*household nuber of E vehicles (w 1 *x 2 ) Intersection density*household nuber of vehicles (w 2 *x 2 ) Average auto wor trip travel tie*household nuber of vehicles (w 3 *x 2 ) < c. Suburban id-age wealthy (obs = 2669) -2 Res Log Lielihood = Intercept < Average auto wor trip travel tie (w 1 ) Nuber of ids (x 1 ) Nuber of household vehicles (x 2 ) < Average auto wor trip travel tie*nuber of ids (w 1 * x 1 ) Average auto wor trip travel tie*nuber of household vehicles (w 1 * x 2 ) Low incoe (z 1 )

32 Table 7 Rando effects of hierarchical odels for suburban retired neighborhoods υ +, 0 υ, x, i, Effect MSA/CMSA Estiate Standard Error t value Pr > t Intercept Atlanta, GA Nuber of ids Intercept Boston-Worcester-Lawrence, Nuber of ids MA-NH-ME-CT Intercept Chicago-Gary-Kenosha, IL-IN Nuber of ids WI Intercept Dallas-Fort Worth, TX Nuber of ids Intercept Detroit-Ann Arbor-Flint, MI Nuber of ids Intercept Houston-Galveston-Brazoria, Nuber of ids TX Intercept Los Angeles-Riverside-Orange Nuber of ids County, CA Intercept Miai-For Lauderdale, FL Nuber of ids Intercept New Yor-Northern New Nuber of ids Jersey-Long Island, NY-NJ- CT-PA Intercept Philadelphia-Wilington Nuber of ids Atlantic City, PA-NJ-DE-MD Intercept Phoenix-Mesa, AZ Nuber of ids Intercept San Francisco-Oaland-San Nuber of ids Jose, CA Intercept Seattle-Tacoa-Breerton, Nuber of ids WA Intercept Washington-Baltiore, DC Nuber of ids MD-VA-WV

33 Table 8 Rando effects of hierarchical odels for urban elite neighborhoods υ +, 0 υ, x, i, Effect MSA/CMSA Estiate Standard t value Pr > t Error Intercept a Atlanta, GA 0 Household size Household nuber of vehicles Household size Boston-Worcester Household nuber of vehicles Lawrence, MA-NH- ME-CT Household size Chicago-Gary Household nuber of vehicles Kenosha, IL-IN-WI Household size Dallas-Fort Worth, TX Household nuber of vehicles Household size Detroit-Ann Arbor Household nuber of vehicles Flint, MI Household size Houston-Galveston Household nuber of vehicles Brazoria, TX Household size Miai-For Lauderdale, Household nuber of vehicles FL Household size Los Angeles-Riverside Household nuber of vehicles Orange County, CA Household size New Yor-Northern Household nuber of vehicles New Jersey-Long Island, NY-NJ-CT-PA Household size Philadelphia Household nuber of vehicles Wilington-Atlantic City, PA-NJ-DE-MD Household size Phoenix-Mesa, AZ Household nuber of vehicles Household size San Francisco-Oaland Household nuber of vehicles San Jose, CA Household size Seattle-Tacoa Household nuber of vehicles Breerton, WA Household size Washington-Baltiore, Household nuber of vehicles DC-MD-VA-WV a. Intercept is fixed (i.e., no rando effect) in this odel to obtain restricted axiu lielihood estiates and excluded in the rest of the table.

34 Table 9 Rando effects of hierarchical odels for suburban wealthy neighborhoods υ +, 0 υ, x, i, Effect MSA/CMSA Estiate Standard t value Pr > t Error Intercept Atlanta, GA Nuber of ids Household nuber of 0 vehicles b Intercept Boston-Worcester-Lawrence, Nuber of ids MA-NH-ME-CT Intercept Chicago-Gary-Kenosha, IL Nuber of ids IN-WI Intercept Dallas-Fort Worth, TX Nuber of ids Intercept Detroit-Ann Arbor-Flint, MI Nuber of ids Intercept Houston-Galveston-Brazoria, Nuber of ids TX Intercept Los Angeles-Riverside Nuber of ids Orange County, CA Intercept Miai-For Lauderdale, FL Nuber of ids Intercept New Yor-Northern New Nuber of ids Jersey-Long Island, NY-NJ- CT-PA Intercept Philadelphia-Wilington Nuber of ids Atlantic City, PA-NJ-DE- MD Intercept Phoenix-Mesa, AZ Nuber of ids Intercept San Francisco-Oaland-San Nuber of ids Jose, CA Intercept Seattle-Tacoa-Breerton, Nuber of ids WA Intercept Washington-Baltiore, DC Nuber of ids MD-VA-WV b. Household nuber of vehicles is fixed (i.e., no rando effect) in this odel to obtain restricted axiu lielihood estiates and excluded in the rest of the table.

35 For any group s, Level 2 Area 1 Area 2 Area M Level 1 Household 1 2 N N N M Figure 1 Hierarchical structure of the odel

36 Figure 2 Clustering results of (clocwise fro upper left) (a) Boston, MA, (b) Chicago, IL, (c) New Yor City, NY, and (d) Los Angeles, CA

37 Figure 3 Visualization of the ten cluster