Effects of zoning structure and network detail on traffic demand modeling

Transcription

1 Environment and Planning B: Planning and Design 2002, volume 29, pages 37 ^ 52 DOI: /b2742 Effects of zoning structure and network detail on traffic demand modeling Kang-tsung Chang Department of Geography, University of Idaho, Moscow, ID , USA; chang@uidaho.edu Zaher Khatib Department of Civil Engineering, University of Idaho, Moscow, ID , USA; zkhatib@uidaho.edu Yanmei Ou} Engineer Department, City of McKinney, McKinney, TX 75070, USA; Received 23 February 2001 Abstract. By using traffic analysis zones (TAZs), centroids, and a road network as the inputs, traffic demand modeling aggregates trips from the locations of individual tripmakers to TAZ centroids and estimates trips generated between TAZs on the network. Like spatial analysis, traffic demand modeling is subject to the modifiable areal unit problem (MAUP). In this paper we report the findings of a simulation study, which uses eleven zoning structures and two levels of network detail to assess the effects of TAZs, centroids, and network detail on statewide traffic demand modeling in Idaho. First, trips generated between smaller TAZs have shorter trip lengths, higher proportions of interzonal trips, more accurate estimated volume-to-ground-count ratios, and lower percentage root mean square errors (E RMS ) between estimated volumes and ground counts. Second, the effect of centroid locations is mixed and generally slight with small TAZs. Third, the level of network detail impacts E RMS values in two ways: larger TAZs produce lower E RMS values than smaller TAZs on the less detailed network, and the detailed network outperforms the less detailed network, regardless of the size of TAZs. 1 Introduction Transportation planners have used traffic demand models over the past three decades to forecast travel demand for long-term planning activities. The structure of a traffic demand model consists of a transportation network, traffic analysis zones (TAZs), and centroids. TAZs represent areas from which and to which trips are allocated in a traffic demand model. Ideally, a TAZ should be homogeneous in population, socioeconomic, and land-use characteristics. A centroid represents the `center of activity' of a TAZ and the origin and destination for all trips to and from the TAZ. Trips generated between TAZs are channeled to the transportation network through connectors joining centroids and the network. Since its initial development in the 1960s, traffic demand modeling has been closely related to urban transportation planning. But in the United States, statewide transportation planning has become important because of the introduction of the Intermodal Surface Transportation Efficiency Act (ISTEA) of 1991 (now the Transportation Equity Act for the 21st Century, or TEA-21) and the Clean Air Act of Each state is required to carry out a statewide transportation plan. Besides meeting the mandated planning requirements, a statewide model is useful for forecasting rural and intercity travel, supplying data such as through trips to urban area models, and providing project-level forecasts in rural areas (FHWA, 1999). Although there are important differences between statewide models and urban area models, many states have adopted an urban-modeling framework to develop their } Current address: Tele Atlas North America, Inc., Menlo Park, CA 94025, USA; yanmei.ou@na.teleatlas.com

2 38 K-t Chang, Z Khatib, Y Ou statewide models (Horowitz and Farmer, 1999). One main difference lies in the zoning structure. An urban area model uses parcels, street blocks, or neighborhoods as TAZs, and shopping centers, business establishments, or geometric centers as centroids. In contrast, the zoning structure for a statewide model usually follows the geography of census such as block groups, census tracts, or counties to take advantage of demographic and socioeconomic data available from the US Bureau of the Census (O'Neill, 1991). Use of census units in a statewide model has raised new questions in traffic demand modeling. First, which level of census units or how many TAZs should be used in a statewide model? Among statewide models that have been developed, the number of internal TAZs ranges from 5 in Wyoming to well over 2000 in Michigan and New Jersey (FHWA, 1999). Second, how should centroids be defined in a statewide model? In an urban area model, a centroid corresponds to a shopping center, a business establishment, or the geometric center of a TAZ. It becomes difficult to define a centroid in a statewide model when the TAZ is as large as a census tract or a county. Besides having heterogeneous land-use patterns, a large TAZ is likely to have more than one city or activity center. Third, how should the level of transportation network detail be matched with the size of TAZs? Is a network of major roads detailed enough for TAZs at the county or census-tract level? Or, is a network with both major and minor roads a better choice regardless of the geographic scale of TAZs? In this study we attempt to answer the above questions by using a statewide model of Idaho as a case study. The paper is organized into three main sections. In the first section we review the modifiable areal unit problem (MAUP) in spatial analysis and traffic demand modeling and discuss the commonality of spatial data aggregation in both topics. In the second section, we describe the research design of the study including transportation networks, zoning structures, and data processing. In the third section we present results of the effects of TAZs and centroids on various measures of traffic forecast, statistical analysis of the effects, and the interaction between network detail and TAZs. 2 Literature review 2.1 MAUP The MAUP addresses the `modifiable' nature of area data used in spatial analysis and the influence it has on the analysis and modeling results. A classic example of MAUP effects is the relationship between estimated correlation coefficients and the geographic level of census data used in the computation (Robinson, 1950). MAUP effects have two components: the scale effect relates to the level of spatial data aggregation, and the zoning effect relates to the definition or partitioning of units for which data are collected (Openshaw and Taylor, 1979; Wong and Amrhein, 1996). Effects of the MAUP have been reported in bivariate regression (Clark and Avery, 1976), multiple regression (Fotheringham and Wong, 1991), spatial interaction (Batty and Sikdar, 1982a; 1982b; 1982c; 1982d; Openshaw, 1977), and location ^ allocation (Goodchild, 1979). At least one study has shown that MAUP effects can vary with the statistics calculated: no apparent effects on means and variances but dramatic effects on regression coefficients and correlation statistics (Amrhein, 1995). Like spatial analysis, traffic demand modeling requires spatial data aggregation becauseitisnotfeasibleöeconomically, computationally, and for privacy issuesöto use data at the level of individual tripmakers. The choice of block groups, census tracts, or counties as TAZs therefore becomes a scale problem. Revision of preliminary TAZ boundaries at a selected scale to match roads, rivers, or other physical features, or use of cluster analysis and geographic information systems (GIS) in

3 Effects of zoning structure and network detail on traffic demand modeling 39 configuring homogeneous and contiguous TAZs from census units, is a zoning problem (Ding, 1998; O'Neill, 1991; You et al, 1997a; 1997b). Although one expects that MAUP effects can influence results of traffic demand modeling, they have so far received little recognition in transportation planning and GIS for transportation (GIS-T) (Miller, 1999). Three studies relating TAZs to traffic demand forecasts deserve mentioning here. By questioning whether travel demand forecasts could be improved through the subdivision of TAZs, Crevo (1991) compared traffic assignments using an original zone system and a modified zone system and reported that the result did not support the hypothesis that a greater number of zones would improve travel demand forecasts. By using a GIS-based computer simulation system, Ding (1994; 1998) showed how traffic characteristics, such as trip length and proportion of intrazonal trips, could change with different numbers of TAZs on a fixed transportation network. Traffic demand modeling requires centroids and a transportation network in addition to TAZs. Therefore, spatial data aggregation in traffic demand modeling involves not only use of area units (TAZs) but also use of points (centroids) and lines (roads). As defined earlier, a centroid is a single point location, to which all trips to and from a TAZ are assigned. The use of centroids, which represents an area to point aggregation, actually has its counterpart in spatial analysis. A location ^ allocation analysis typically aggregates population to be served by a public facility, such as a school or a public library, at demand points (Francis et al, 1999; Goodchild, 1979). The location of demand points can obviously influence the analysis result, especially over an implicit network. Network detail represents yet another form of spatial data aggregation: a less detailed road network has a higher level of traffic data aggregation than a more detailed network. Previous studies by Crevo (1991) and Ding (1994; 1998) concentrated on the effect of TAZs on traffic demand forecasts. In this study we expanded the scope of the research question by including not only the simple effects of TAZs and centroids but also the interactions between TAZs and centroids and between TAZs and network detail. At the same time, this study broadened the MAUP by dealing with different forms of spatial data aggregation and by studying their interactions in the context of traffic demand modeling. A general recommendation in transportation planning is to choose the detail of network and the size of TAZs that are appropriate to the analysis required or the objectives of the statewide model (FHWA, 1999; Wilson and Wang, 1995). But the recommendation, which assumes that the level of spatial data aggregation should be compatible between input data, has not been tested empirically (Ding, 1998). 2.2 Traffic demand modeling Traffic demand models typically follow a four-step process of trip generation, trip distribution, modal choice, and network or trip assignment (Edwards, 1999). The steps are chained in a sequence, and the outputs of each step become inputs of the following step. Because TAZs and centroids are defined and used at the beginning of the modeling process, they affect the subsequent outputs including trip assignments on the network Trip generation Trip generation estimates the number of person-trips to and from a TAZ for a typical day. Trip generation models cover trip production and trip attraction. Trip production refers to trip ends (in other words, trip origin or destination) at the traveler's home, whereas trip attraction refers to trip ends at a nonhome location, such as a workplace, shopping center, or school. Estimates of trip production and attraction are usually grouped by trip purpose because the travel behavior of tripmakers depends on the

4 40 K-t Chang, Z Khatib, Y Ou trip purpose. This study dealt with three trip purposes: home-based work, home-based other, and non-home-based. The most common trip generation models are regression models and cross-classification models. A regression model for trip production uses the total trips produced by a zone as the dependent variable and household characteristics as the independent variables. A regression model for trip attraction uses the total trips attracted by a zone as the dependent variable and nonresidential attributes, such as retail employment, service employment, and other employment, as the independent variables. Mainly used for trip production estimates, a cross-classification model first classifies household types into a set of categories by income, car ownership, and household size that are highly correlated with tripmaking. Trip production rates are then estimated for each type of household statistically. Data compiled from a variety of travel surveys are needed to develop and calibrate a trip generation model (Crevo et al, 1995; Virkud and Keyes, 1995). Because statewide travel survey data for Idaho were not available, we used trip production rates and attraction equations from Martin and McGuckin (1993). By holding the rates and equations constant in traffic demand modeling, we were able to achieve the primary purpose of the study, in other words, to isolate the effects of TAZs, centroids, and network detail Trip distribution Trip distribution estimates trip interchanges between all pairs of trip-producing and trip-attracting zones. Trip distribution assumes that all trip-attracting zones compete with each other to attract trips produced in each trip-producing zone and that zones having higher levels of `attractiveness' attract more trips. In addition, trip distribution considers other intervening factors, such as distance and travel time, which affect travel behavior. The most widely used trip distribution model is the gravity model (Papacostas and Prevedouros, 1993). The gravity model predicts that the relative number of trips made between two TAZs is directly proportional to the number of trip productions and attractions in each TAZ, and inversely proportional to a function of the spatial separation or travel time between the two TAZs Modal choice Modal choice estimates the number of trips by each available transportation mode between a production ^ attraction pair. An urban traffic demand model may have automobile, public transit, bicycle, and walking as mode choices. A statewide traffic demand model usually ignores this step because of the low public transit ridership for intercity travels. In this study we did not consider modal choice and dealt only with the vehicle trips Trip assignment Trip assignment models load the vehicle trips onto the transportation network by using a range of path-building algorithms, which may be used alone or in combination (FHWA, 1999; Papacostas and Prevedouros, 1993). The all-or-nothing algorithm assigns all trips between each pair of zones to the shortest path based on distance or travel time between them. All other paths are assigned nothing. After all trip interchanges are assigned, the traffic flow on a particular road link is computed by summing all interzonal flows that include the link on their minimum paths. Capacity restraint, or equilibrium loading, balances trip assignment with the link capacity and speed by using an iterative procedure. The first iteration is an all-ornothing assignment. On successive iterations, link travel-time is adjusted link by

5 Effects of zoning structure and network detail on traffic demand modeling 41 link according to capacity restraints. After adjustment, the algorithm recomputes new minimum paths and uses another all-or-nothing method to assign trips on the new paths. Because the algorithm cannot guarantee the convergence of traffic volumes from successive iterations to a stable value, the final assignment is usually an average of the assigned volumes, or based on a weighting scheme that assigns greater weights to late iterations (FHWA, 1990). After experimenting with the all-or-nothing and capacity restraint algorithms, we used capacity restraint in trip assignment because it yielded better results. The number of iterations was set to twenty five, and the link capacity function came from the standard Bureau of Public Roads formula (Papacostas and Prevedouros, 1993). 3 Research design 3.1 Data Network We built two statewide networks for our study. The first network included all interstates, principal arterials, minor arterials, and major collectors in both urban and rural areas, and minor collectors in rural areas (figure 1). This detailed network stitched all interstate, US, and state highways from TIGER files with the rural and city functional classification road maps from the Idaho Transportation Department. Urban areas included three metropolitan planning areas (Boise ^ Nampa ^ Caldwell, Pocatello, and Interstates US highways Other roads N km Figure 1. Major and minor roads in Idaho.

6 42 K-t Chang, Z Khatib, Y Ou Idaho Falls) and thirteen other cities. The rest of the state was classified as rural. Table 1 shows the number of links and the mileage of link groups by functional class and by area type. The second networköa less detailed networköincluded only interstate, US, and state highways (figure 2). Link attributes for the network included traffic related data. Compiled from the Idaho Transportation Department database, traffic-related data for major roads Table 1. Number of links and mileage by functional classification. Link group Number of links Mileage functional class area type (% total) (% total) Interstate rural 390 (3.18) 532 (3.97) Principal arterial rural (11.02) (12.52) Minor arterial rural (9.11) (9.61) Interstate urban 190 (1.55) 76 (0.57) Principal arterial urban (8.43) 245 (1.83) Minor arterial urban (13.27) 507 (3.97) Collector rural (48.64) (66.61) Collector urban 588 (4.80) 147 (1.10) Total (100) (100) Interstates US highways State highways N km Figure 2. Major roads in Idaho.

7 Effects of zoning structure and network detail on traffic demand modeling 43 included speed limit, capacity, functional classification, annual average daily traffic (AADT), and number of lanes. The only information available for minor roads was functional classification, which was used in this study to estimate speed limit and capacity from published tables Zoning structures A zoning structure consisted of TAZs and their centroid locations. We prepared 11 zoning structures by varying TAZs and centroid locations (table 2). TAZs included three levels of census geography: county (C), census tract (T), and census block group (B). Idaho has 44 counties, 269 census tracts, and 1122 block groups (figure 3, see over). Centroids included four types of location: geometric center (G), city location (C), population-weighted center (PW), and household-density-weighted center (HW). Irrespective of its location, a centroid was connected to the closest point on the transportation network by a straight line or dummy link without crossing the network or other connectors. Table 2. Abbreviations and descriptions of eleven zoning structures. Abbreviation Description C ± G County-level TAZs and geometric centroids C ± C County-level TAZs and city centroids C ± PW County-level TAZs and population-weighted centroids C ± HW County-level TAZs and household-density-weighted centroids T ± G Census-tract-level TAZs and geometric centroids T ± C Census-tract-level TAZs and city centroids T ± PW Census-tract-level TAZs and population-weighted centroids T ± HW Census-tract-level TAZs and household-density-weighted centroids B ± G Census-block-group-level TAZs and geometric centroids B ± C Census-block-group-level TAZs and city centroids B ± PW Census-block-group-level TAZs and population-weighted centroids TAZ, traffic analysis zones. The geometric center option used the physical center of a TAZ as centroid, whereas the other three types were derived centroid locations. For the city location and population-weighted center options, we first overlaid a digital map of 650 places (cities hereafter) in Idaho with the TAZ map. The city location option used a city's location as centroid if it was the only city within a TAZ or the largest city as centroid if a TAZ contained two or more cities. The population-weighted center option applied to a TAZ with two or more cities by using the following equations: X ˆ X n X n x i P i P i, Y ˆ X n X n y i P i P i, where X and Y are the x and y coordinates of the population-weighted centroid, x i and y i are the x and y coordinates of city i, n is the number of cities in a TAZ, P i is the population of city i. If a TAZ did not contain any city, then its geometric center was used as the centroid. For example, a city might include ten census tracts within its boundary but only one city location was included in the city map. In that case, nine census tracts

8 44 K-t Chang, Z Khatib, Y Ou Census block groups Census tracts Counties N km Figure 3. Counties, census tracts, and block groups in Idaho. would use geometric centers as centroids. Therefore, the zoning structures listed as having city locations or population-weighted centers as centroids also included geometric centers in urban areas. The household-density-weighted option required subareas with household data within each TAZ and was available only at the census tract and county levels. For a centroid at the census-tract level, we computed the household density for each block group within a census tract and then derived a household-density-weighted center for the census tract by using the following equations: X ˆ X n X n x i H i H i, Y ˆ X n X n y i H i H i, where X and Y are the x and y coordinates of the household-density-weighted centroid, x i and y i are the x and y coordinates of the geometric center of block group i, n is the number of block groups in a census tract, H i is the household density of block group i. The same procedure was followed to compute household-density-weighted centers at the county level with census tracts replacing block groups in the above equations.

9 Effects of zoning structure and network detail on traffic demand modeling Data processing We used ARC/INFO, a GIS package from ESRI in Redlands, CA, to prepare data files including the networks, centroids, and centroid connectors. ARC/INFO commands were used directly to locate geometric centers and centroid connectors. Centroid locations other than geometric centers were derived by using macro programs (AML programs) in ARC/INFO. We used TRANPLAN, a transportation-planning analysis package from the Urban Analysis Group in Dauville, CA, for traffic demand modeling. To isolate the effects of TAZs, centroids, and transportation network, we used the same values for trip generation rates and other parameters in TRANPLAN with each zoning structure. The output from each run showed the trip assignment result on the transportation network for a specific zoning structure. All eleven zoning structures were used to assess the effects of TAZs and centroids, while six zoning structures (C ^ G, C ^ PW, T ^ G, T ^ PW, B ^ G, and B ^ PW) were used to evaluate the interaction between network detail and TAZs. 4 Analysis Our data analysis included three parts. The first part examined the general performance of traffic demand modeling by comparing travel characteristics such as trip length and assigned interzonal trips. The trip length was the average travel time for the trip purpose of home-based work, home-based other, or non-home-based. Interzonal trips were trips that began and ended in different TAZs, as opposed to intrazonal trips which began and ended in the same TAZ. Only interzonal trips were assigned to a network in a traffic demand model. The literature suggests that smaller TAZs tend to result in higher levels of interzonal trips (Ding, 1994; 1998; Ord and Cliff, 1976), and TAZs must be small enough to ensure that intrazonal trips do not exceed 15% of total trips (Crevo, 1991). The second part compared the estimated traffic volume (V )withtheaadtor ground count (A) by link. The Idaho Transportation Department had ground counts for all interstate links, 98% of principal and minor arterials in rural areas, 55% of principal arterials in urban areas, and 8% of minor arterials in urban areas. Without ground counts, rural and urban collectors were not included in this part of data analysis. We used two methods for the link comparison. For the first method we computed the ratio V=A for each link, and grouped the V=A ratios into three classes: <0.8; 0.8 ^ 1.2; and >1.2. The 0.8 ^ 1.2 class with estimated volumes within 20% of their ground counts was considered satisfactory. For the second we computed the percentage root mean square error (E RMS ) between V and A (Barton-Aschman Associates and Cambridge Systematics, 1997): E RMS " # 2, ˆ 100 Xn V i A i 2 X n n 1 A i n, where n is the number of links, V i is the estimated traffic volume of link i, A i is the AADT count of link i. In the third part, we took samples from different zoning structures and conducted the analysis of variance (ANOVA) to test whether the means of jv Aj=A were equal among the samples. We then used Fisher's least significant difference (LSD) test (Ott, 1993) to identify the zoning structures that had significant differences. We also used two-way ANOVA to test the effects of the geographic level of TAZs, the location of centroids, and their interaction.

10 46 K-t Chang, Z Khatib, Y Ou 5 Results Two notes should be made about the interpretation of the results. First, rural links were more important than urban links in Idaho: the total mileage of rural interstates, principal arterials, and minor arterials was more than four times longer than their urban counterparts in Idaho (table 1). Most rural links in the above three functional classes had ground counts for comparison with estimated traffic volumes as well. Second, V=A ratios and E RMS percentages were computed in this study for the single purpose of comparison between zoning structures: a better zoning structure would have V=A ratios closer to 1.0 and smaller values of E RMS among the link groups. Results from this study should not be compared with percentage V=A difference targets defined by the Federal Highway Administration or other studies based on calibrated models (Barton-Aschman Associates and Cambridge Systematics, 1997; FWHA, 1990). 5.1 Trip length Table 3 shows the average trip lengths in minutes by trip purpose. Smaller TAZs have shorter trip lengths: a difference of up to 17 minutes between C ^ G and T ^ G, and a difference of up to 4 minutes between T ^ G and B ^ G. The exception is T ^ HW, which has the shortest trip length for all trip purposes. Replacement of geometric centers by derived centroid locations reduces the trip length by 10 to 13 minutes at the county level, 3 to 7 minutes at the census tract level, and about a minute at the block-group level. Table 3. Average trip lengths, in minutes, by trip purpose and zoning structure. Home-based work Home-based other Non-home-based C ± G C ± C C ± PW C ± HW T ± G T ± C T ± PW T ± HW B ± G B ± C B ± PW Interzonal and intrazonal trips Table 4 shows the percentages of assigned interzonal and unassigned intrazonal trips that make up the total trips. Smaller TAZs have higher percentages of interzonal trips. The percentage of interzonal trips increases from 18% for C ^ G, to 75% for T ^ G, to 89% for B ^ G. Only TAZs at the block-group level have proportions of intrazonal trips lower than 15%. The percentage of interzonal trips also increases with use of derived centroid locations: 15% ^ 19% increases at the county level, 6% ^ 10% increases at the census-tract level, and 3% ^ 4% increases at the block-group level. 5.3 Comparison of estimated volumes and ground counts V=A ratios Table 5 shows the proportions of links in three V=A ratio classes by zoning structure and by functional classification link group. Smaller TAZs have higher proportions of V=A ratios in the 0.8 ^ 1.2 or satisfactory class for nearly every rural and urban link group. One exception is the urban minor arterial group, for which TAZs at the censustract level perform slightly better than TAZs at the block-group level. Use of derived

11 Effects of zoning structure and network detail on traffic demand modeling 47 Table 4. Proportions of interzonal trips and intrazonal trips by zoning structure. Percentage interzonal trips Percentage intrazonal trips C±G C±C C ± PW C ± HW T±G T±C T ± PW T ± HW B±G B±C 92 8 B ± PW 93 7 Table 5. Percentage of links by V=A ratio class, link group, and zoning structure. V=A C±G C±C C±PW C±HW T±G T±C T±PW T±HW B±G B±C B±PW Interstate (rural) < ± > Principal arterial (rural) < ± > Minor arterial (rural) < ± > Interstate (urban) < ± > Principal arterial (urban) < ± > Minor arterial (urban) < ± > centroid locations increases the proportion of V=A ratios in the satisfactory class for all link groups at the county level. Derived centroid locations have mixed effects at the census-tract and block-group levels: positive for rural interstate and urban minor arterial, but generally negative for other link groups.

12 48 K-t Chang, Z Khatib, Y Ou E RMS percentage values Table 6 shows E RMS percentage values by zoning structure and by link group. Smaller TAZs have lower E RMS percentages, as shown by steady decreases of E RMS from C ^ G, to T ^ G, to B ^ G on every link group except rural interstate. The effect of replacing geometric locations with derived centroid locations is overall negative at the county level but slight and mixed at the census-tract and block-group levels. No single zoning structure has the best performance for all link groups. B ^ C and B ^ PW have the lowest E RMS (39% ^ 66%) for every link group except urban principal arterial and urban minor arterial. B ^ G has the lowest E RMS for urban principal arterial, whereas T ^ G has the lowest E RMS for urban minor arterial. Table 7 shows E RMS percentages by AADT value range and by zoning structure. Vertically, E RMS steadily decreases from low to high AADT values; in other words, from less than 1000 to over , for almost every zoning structure. Smaller TAZs have lower E RMS percentages for all link groups, as shown by constant decreases of E RMS from C ^ G, to T ^ G, to B ^ G. The overall effect of using derived centroid locations is negative. No single zoning structure in table 7 outperforms the others Table 6. Percentage E RMS values by link group and zoning structure. C±G C±C C±PW C±HW T±G T±C T±PW T±HW B±G B±C B±PW Interstate (rural) Principal arterial (rural) Minor arterial (rural) Interstate (urban) Principle arterial (urban) Minor arterial (urban) Overall Table 7. Percentage E RMS and zoning structure. values by link group of annual average daily traffic count value range C±G C±C C±PW C±HW T±G T±C T±PW T±HW B±G B±C B±PW Ground counts (1000) < ± ± ± ± ± > Overall

13 Effects of zoning structure and network detail on traffic demand modeling 49 decisively, although B ^ G has the lowest overall E RMS and the lowest E RMS values for every link group except the 5000 ^ and ^ link groups. B ^ C and B ^ PW tie for the second best overall percentage E RMS value. 5.4 Statistical analysis The result from one-way ANOVA shows that the means of E RMS values are not equal among the eleven zoning structures (the computed F-ratio of is greater than the critical F-ratio of 1.83 at the 5% significant level). Multiple comparisons using Fisher's LSD show that the means at the block group level are significantly different from the means at the census-tract or county level, and the means at the census-tract level are significantly different from the means at the county level. The result from two-way ANOVA shows that the effect of the geographic level of TAZs is significant at the 5% level (the computed F-ratio of is greater than the critical F-ratio of 3.00), but the effect of centroid location is nonsignificant. The interaction between the two variables is significant at the 5% level (the computed F-ratio of 2.62 is greater than the critical F-ratio of 2.37). 5.5 Network detail Trip length Table 8 shows average trip lengths by trip purpose, zoning structure, and network. The detailed network (D) included major and minor roads, whereas the less detailed network (LD) included only major roads. The effect of network detail on the trip length is slight. The trip lengths for home-based work and home-based other are all slightly higher on a less detailed network. But the trip length for non-home-based is higher for some and lower for others on the less detailed network. Table 8. Average trip lengths, in minutes, by trip purpose, zoning structure, and network detail. Home-based work Home-based other Non-home-based D LD D LD D LD C ± G C ± PW T ± G T ± PW B ± G B ± PW D, detailed network; LD, less detailed network. Table 9. Proportions of interzonal trips and intrazonal trips by zoning structure and network detail. Percentage interzonal trips Percentage intrazonal trips D LD D LD C±G C ± PW T±G T ± PW B±G B ± PW D, detailed network; LD, less detailed network.

14 50 K-t Chang, Z Khatib, Y Ou Proportion of interzonal and intrazonal trips Table 9 shows the proportions of interzonal and intrazonal trips by trip purpose, zoning structure, and network. Similar to the trip length, the effect of network detail is slight E RMS percentage values Table 10 shows E RMS values by link group, zoning structure, and network. The less detailed network has higher E RMS values than the detailed network for all link groups and zoning structures. When percentage E RMS values are compared among different zoning structures on the less detailed network, C ^ G has the smallest overall E RMS, followed by C ^ PW. Table 10. Percentage E RMS values by link group, zoning structure, and network detail. C±G C±PW T±G T±PW B±G B±PW D LD D LD D LD D LD D LD D LD Interstate (rural) Principal arterial (rural) Minor arterial (rural) Interstate (urban) Principal arterial (urban) Minor arterial (urban) Overall D, detailed network; LD, less detailed network. 6 Conclusion This study has found the following three patterns about the effects of zoning structure and network detail on traffic demand modeling. First, smaller TAZs generated shorter trip lengths, higher proportions of assigned interzonal trips, better V=A ratios, and lower overall percentage E RMS values between V and A. This finding confirms the general observation in the transportation planning literature that smaller TAZs will achieve better results (for example, Sosslau, 1973). Second, the effect of using derived centroids instead of geometric centroids increased with the size of TAZs. Derived centroids generated shorter trip lengths and higher proportions of interzonal trips but mixed results on V=A ratios and E RMS values. At the county level, use of derived centroids improved V=A ratios in the 0.8 ^ 1.2 class but negatively impacted E RMS values apparently owing to the increase in extreme estimated volumes. Therefore, the extra effort in computing derived centroids may not be justified without further study. Third, the level of network detail had a negligible effect on trip length or proportion of interzonal trips, but impacted the value of E RMS in two ways. On one hand, larger TAZs resulted in lower percentage E RMS values on the less detailed network. On the other hand, the detailed network resulted in lower E RMS values than the less detailed network for all link groups, regardless of the geographic level of TAZs. This third finding reinforces the benefit of using smaller TAZs in traffic demand modeling but directly challenges the general observation of matching the level of network detail with the size of TAZs in transportation planning.

15 Effects of zoning structure and network detail on traffic demand modeling 51 This simulation study has shown effects of spatial data aggregation on statewide traffic demand modeling. Transportation planners can no longer ignore the impact that their decisions on TAZs, centroids, and network detail will have on estimated traffic volumes. Manipulation of the zoning structure and transportation network used to be difficult and time consuming. GIS can now assist transportation planners in processing and exploring the input data to traffic demand modeling. Acknowledgements. We would like to thank two anonymous reviewers for their helpful comments on an earlier draft. References Amrhein C G, 1995, ``Searching for the elusive aggregation effect: evidence from statistical simulations'' Environment and Planning A ^ 119 Barton-Aschman Associates Inc, Cambridge Systematics Inc, 1997 Model Validation and Reasonableness Checking Manual prepared for Federal Highway Administration, US Department of Transportation, 400 Seventh Street, SW, Washington, DC Batty M, Sikdar P K, 1982a, ``Spatial aggregation in gravity models. 1. An information theoretic framework'' Environment and Planning A ^ 405 Batty M, Sikdar P K, 1982b, ``Spatial aggregation in gravity models. 2. One-dimensional population density models'' Environment and Planning A ^ 553 Batty M, Sikdar P K, 1982c, ``Spatial aggregation in gravity models. 3. Two-dimensional trip distribution and location models'' Environment and Planning A ^ 658 Batty M, Sikdar P K, 1982d, ``Spatial aggregation in gravity models. 4. Generalisations and largescale applications'' Environment and Planning A ^ 822 Clark W A V, Avery K L, 1976, ``The effect of data aggregation in statistical analysis'' Geographical Analysis ^ 437 Crevo C C, 1991, ``Impacts of zonal reconfigurations on travel demand forecasts'', in Transportation Research Record number 1305, pp 72 ^ 80, Transportation Research Board, National Research Council, Washington, DC Crevo C C, Niedowski R S, Scott D J, 1995, ``Design and conduct of a statewide household travel survey in Vermont'', in Transportation Research Record number 1477, pp 26 ^ 30, Transportation Research Board, National Research Council, Washington, DC Ding C, 1994, ``Impact analysis of spatial data aggregation on transportation forecasted demand: a GIS approach'', in URISA Proceedings URISA, 1460 Renaissance Drive, Park Ridge, IL 60068, pp 362 ^ 375 Ding C, 1998, ``The GIS-based human-interactive TAZ design algorithm: examining the impacts of data aggregation on transportation-planning analysis'' Environment and Planning B: Planning and Design ^ 616 Edwards J D (Ed.),1999 Transportation Planning Handbook 2nd edition, Institute of Transportation Engineers, th Street, NW, Washington, DC FHWA, 1990, ``Calibration and adjustment of system planning models'', US Department of Transportation, 400 Seventh Street, SW, Washington, DC 20590, DOCS/377CAS.html FHWA, 1999, ``Guidebook on statewide travel forecasting'', US Department of Transportation, 400 Seventh Street, SW, Washington, DC Fotheringham A S, Wong D W S, 1991, ``The modifiable areal unit problem in multivariate statistical analysis'' Environment and Planning A ^ 1044 Francis R L, Lowe T J, Rushton G, Rayco M B, 1999, `À synthesis of aggregation methods for multifacility location problem: strategies for containing error'' Geographical Analysis ^ 87 Goodchild M F, 1979, ``The aggregation problem in location ^ allocation'' Geographica Analysis ^ 255 Horowitz A J, Farmer D D, 1999, ``Statewide travel forecasting practice: a critical review'', in Transportation Research Record number 1685, pp 13 ^ 20, Transportation Research Board Martin W A, McGuckin N A, 1993, ``Travel estimation techniques for urban planning'', prepared for the National Cooperative Highway Research Program, Transportation Research Board, The National Academies, 2101 Constitution Avenue, NW, Washington, DC Miller H J, 1999, ``Potential contributions of spatial analysis to geographic information systems for transportation'' Geographical Analysis ^ 399 O'Neill W A, 1991, ``Developing optimal transportation analysis zones using GIS'' ITE Journal 61 (December) 33 ^ 36

16 52 K-t Chang, Z Khatib, Y Ou Openshaw S, 1977, ``Optimal zoning systems for spatial interaction models'' Environment and Planning A ^ 184 Openshaw S, Taylor P J, 1979, `À million or so correlation coefficients: three experiments on the modifiable areal unit problem'', in Statistical Applications in the Spatial Sciences Ed. N Wrigley (Pion, London) pp 127 ^ 144 Ord J K, Cliff A D, 1976, ``The analysis of commuting patterns'' Environment and Planning A ^ 946 Ott L R, 1993 An Introduction to Statistical Methods and Data Analysis 4th edition (Duxbury Press, Boston, MA) Papacostas C S, Prevedouros P D, 1993 Transportation Engineering and Planning 2nd edition (Prentice-Hall, Englewood Cliffs, NJ) Robinson W S, 1950, ``Ecological correlation and the behavior of individuals'' American Sociological Review ^ 357 Sosslau A B, 1973, ``Traffic assignment'', FHWA, US Department of Transportation, 400 Seventh Street, SW, Washington, DC Virkud U, Keyes C S, 1995, ``Design and implementation of a statewide roadside origin destination survey in Vermont'' Transportation Research Record ^ 25, Transportation Research Board, National Research Council, Washington, DC Wilson E M, Wang J, 1995 Statewide Transportation PlanningöAn Interactive Modeling Process Wyoming Department of Transportation, 5300 Bishop Boulevard, Cheyenne, WY Wong D, Amrhein C, 1996,``Research on the MAUP: old wine in a new bottle or real breakthrough'' Geographical Systems 3 73 ^ 76 You J, Nedovic-Budic Z, Kim T J, 1997a, `À GIS-based traffic analysis zone design: technique'' Transportation Planning and Technology ^ 68 You J, Nedovic-Budic Z, Kim T J, 1997b,`À GIS-based traffic analysis zone design: implementation and evaluation'' Transportation Planning and Technology ^ 91 ß 2002 a Pion publication printed in Great Britain