Spatial clustering and the temporal mobility of walking school trips in the Greater Toronto Area, Canada

ARTICLE IN PRESS Health & Place 16 (2010) 646–655

Contents lists available at ScienceDirect

Health & Place journal homepage: www.elsevier.com/locate/healthplace

Spatial clustering and the temporal mobility of walking school trips in the Greater Toronto Area, Canada Raktim Mitra a,n, Ron N. Buliung b, Guy E.J. Faulkner c a

Department of Geography and Program in Planning, University of Toronto, 100 St. George Street, Toronto, Ontario, Canada M5S 3G3 Department of Geography, University of Toronto Mississauga, 3359 Mississauga Road N, South Building, Mississauga, Ontario, Canada L5L 1C6 c Faculty of Physical Education and Health, University of Toronto, 55 Harbord Street, Toronto, Ontario, Canada M5S 2W6 b

a r t i c l e in f o

a b s t r a c t

Article history: Received 2 September 2009 Received in revised form 20 January 2010 Accepted 24 January 2010

Interest in utilitarian sources of physical activity, such as walking to school, has emerged in response to the increased prevalence of sedentary behavior in children and youth. Public health practitioners and urban planners need to be able to survey and monitor walking practices in space and time, with a view to developing appropriate interventions. This study explored the prevalence of walking to and from school of 11–13 year olds in the Greater Toronto Area (GTA), Canada. The Getis–Ord (Gni ) local spatial statistic, Markov transition matrices, and logistic regressions were used to examine the spatial clustering of walking trips in the study area, and to document any temporal drift of places in and out of walking clusters. Findings demonstrate that walking tends to cluster within the urban and innersuburban GTA, and in areas with low household income. Temporally persistent cluster membership was less likely within inner-suburban and outer-suburban places. The evidence suggests that interventions to increase active school transportation need to acknowledge spatial and temporal differences in walking behavior. & 2010 Elsevier Ltd. All rights reserved.

Keywords: Active school transportation Physical activity Cluster Spatio-temporal analysis Built environment

1. Introduction Walking and cycling for school travel may make an important contribution to overall daily energy expenditure for children and youth (Faulkner et al., 2009). Engaging in active school transportation (AST: walking and cycling) may also develop into persistent environmentally sustainable travel practices through time (Black et al., 2001), improve psychological well being (Frumkin et al., 2004), and give rise to greater participation in physical activity later in life (Frank et al., 2003). Policy interest in AST, as a utilitarian source of physical activity, has materialized in response to the increasing prevalence of obesity and overweight in children and youth (Frumkin et al., 2004; Tudor-Locke et al., 2001). In response, urban planners have emphasized built environment interventions to increase AST among children (Ontario Professional Planners Institute, 2009; Transportation Alternatives, 2002). The interest of practice in AST has been matched with the development of transdisciplinary research focused on the relationship between the built environment, school transportation, and the level of physical activity among children and youth. Several recent studies have reviewed and summarized this

n

Corresponding author. Tel.: + 1 416 9299803. E-mail address: [email protected] (R. Mitra).

1353-8292/$ - see front matter & 2010 Elsevier Ltd. All rights reserved. doi:10.1016/j.healthplace.2010.01.009

emerging urban planning and physical activity literature (Pont et al., 2009; Sirard and Slater, 2008). Perhaps the largest consensus among the researchers concerns the effect of distance on school travel mode choice; longer school travel distances tend to associate with auto-oriented outcomes (Ewing et al., 2004; McDonald, 2008; Nelson et al., 2008; Timperio et al., 2006). With regard to other built environment characteristics, the empirical results have been less conclusive. Some studies indicate that density (Kerr et al., 2006; McDonald, 2008), land use mix (Larsen et al., 2009; McMillan, 2007), and neighborhood design characteristics (e.g., street facing houses) (McMillan, 2007) statistically associate with the likelihood of taking an active mode to school. In contrast, others have reported no statistically significant association between density and/or land use mix, and school travel mode choice (Ewing et al., 2004; Yarlagadda and Srinivasan, 2008). A few researchers have examined differences in school transport between urban and rural places; not surprisingly, urban children tend to walk more for school purposes than others (Martin et al., 2007; Robertson-Wilson et al., 2007). With regard to transportation connectivity, the presence of sidewalks appears to positively associate with AST (Boarnet et al., 2005; Ewing et al., 2004), while other design characteristics such as intersection density tend to negatively correlate with the likelihood of walking (Mitra et al., 2010; Schlossberg et al., 2006; Timperio et al., 2006). These findings support conclusions

ARTICLE IN PRESS R. Mitra et al. / Health & Place 16 (2010) 646–655

regarding the central role of safety concerns in the school travel debate. Not only do objective measures of the pedestrian experience appear to correlate with AST, but research has also shown that perceptions of risk regarding traffic and personal safety in general may act as key barriers to AST (Kerr et al., 2006; Schlossberg et al., 2006). Parental attitudes toward travel may also mediate school transport decisions. For example, the carcenteredness of a household (Black et al., 2001; McMillan, 2007), and lack of confidence about a child’s capabilities for autonomous mobility (Martin et al., 2007; Wen et al., 2008) likely reduce the odds of walking to/from school. Despite current thinking that the built environment may facilitate healthy and sustainable school travel mode choice decisions, there remains two important conceptual issues, that are not addressed in current AST literature (e.g., Pont et al., 2009; Sirard and Slater, 2008), requiring closer attention. First, while most empirical research has studied travel behavior of an individual child or youth, and explored built environment qualities of neighborhoods, patterns of AST practices at the regional scale, an increasingly relevant planning unit, have received less attention. If the built environment does influence AST, then it is reasonable to assume that AST should vary systematically across an urban region, where characteristics of the built environment also vary. For example, in Western nations, walking school trips are expected to be more common in ‘‘traditional’’ urban areas, developed prior to World War Two (WW2), where development densities, land use mix, street connectivity, and the availability of pedestrian infrastructures combine to provide a built environment that is perceived to be comparably more walkable than what might be found in the typical post-war suburban context. At the scale of an urban region, particularly one like the Greater Toronto Area (GTA), Canada, which has a lengthy development history ranging from the colonial period to present day, a range of urban forms is expected to be present (see Sewell, 2009; White, 2007). This mixture of forms may associate with a diverse range of walkability outcomes, creating a formidable challenge for regional planners who may be attempting to influence walking in the absence of data that adequately describe how outcomes associated with school travel mode choice decisions potentially vary from place to place, in space, and in time. Second, the temporal consistency of mode choice for school travel (i.e., between morning and afternoon period trips), or lack thereof, has received relatively less attention in the literature. Although most studies have focused on trips to school (e.g., Ewing et al., 2004; McDonald, 2008; McMillan, 2007); and while others have explored school travel in general, pooling the responses across the a.m. and p.m. periods ( e.g., Kerr et al., 2006; Timperio et al., 2006; Wen et al., 2008), recent exploratory research has found more parentally escorted car trips to school, than parentally escorted car trips from school to home at the end of the day (Buliung et al., 2009; Schlossberg et al., 2006; van der Ploeg et al., 2008). This third group of studies is particularly relevant because they begin to suggest that the strength and possibly the direction of the relationship between AST and the built environment may change through time, as the spatial, temporal, and institutional constraints (e.g., hours of paid work, suburbanization of employment) facing households intensify and/or subside throughout the day. Also, and regardless of the cause, changes in individual mode choice decisions over time, when aggregated to the neighborhood or some other areal unit, could produce places that generate above average AST rates some of the time, and below average rates at other times. In other words, the relationship between place and AST is likely to be spatially and temporally heterogeneous. This study contributes to the literature on active school travel, and pedestrian behavior, by exploring this spatial and temporal

647

variability in school travel mode share in the GTA. For public health practitioners and urban planners, to be able to identify places where healthy behaviors are either commonly or rarely practiced, may have important implications with regard to developing appropriate interventions. Certainly, the potential spatial and temporal variation of AST outcomes, described previously, could moderate the impact of a policy intervention (e.g., sidewalks development) on the use of active modes. Nesting the school travel mode choice problem within a quantitative spatial analytic framework provides an opportunity to demonstrate the utility of spatial analysis for broadening current thinking about the relationship between active school travel and place. The study addresses three research questions: first, to what extent are high rates of walking geographically clustered? Second, to what extent do places change in time, i.e., between the a.m. (trips to school) and p.m. (trips from school) time periods, with regard to the clustering of walking trips? And third, what spatial characteristics may explain the likelihood of being part of a cluster? The Getis–Ord (Gni ) local spatial statistic, Markov transition matrices, and logistic regressions were used to explore this ‘‘idealized’’ travel behavior (i.e., walking to and/or from school) in space and time. The remainder of the paper is organized into three sections. The following section establishes the scope of the study and outlines the empirical research design. Observations regarding the spatial clustering of AST outcomes, temporal variability of the spatial clusters, and the correlates of cluster membership are described next. The paper concludes with a summary of the major findings, and a brief discussion of their implications for advancing knowledge and policy that are focused on the promotion of AST.

2. Research design 2.1. Data and study area Data from the 2001 Transportation Tomorrow Survey (TTS) were used to explore patterns of walking school trips across space and time. The decision to exclude the other common active mode, cycling, was informed by previous research which found that, for the study area, cycling captured no more than 1% of all school trips (Buliung et al., 2009). The TTS is a repeat cross-sectional travel survey that covers 21 cities and regional municipalities in the province of Ontario, Canada; it has been conducted, once, every five years, since 1986 (Data Management Group, 2008). The TTS data are collected using a computer assisted telephone interview (CATI) method, in the Fall or the Spring; a direct data entry (DDE) instrument is used to record self-reported weekday travel data (e.g., origin and destination of a trip, trip purpose, travel mode). Travel information for all household members aged 11 years and older, for trips made on the day prior to the interview date, is proxy reported by an adult household member. The 2001 TTS collected data from 5% of the households in the study area; this sample produced approximately 817,000 trip records (Data Management Group, 2008). The study area for this research is limited to the City of Toronto, and the four surrounding regional municipalities: Durham, York, Peel, and Halton. These places collectively represent the GTA, which is Canada’s largest urban region with regard to both population and geographic area. Travel data are reported at the level of 1548 traffic analysis zones (TAZs)—small geographical areas that are commonly used for transportation planning and engineering purposes. The TAZs for the 2001 TTS are largely similar in size to census tracts (CT). However, within the suburban GTA where CTs are typically larger, a CT may contain

ARTICLE IN PRESS 648

R. Mitra et al. / Health & Place 16 (2010) 646–655

several TAZs, a feature that facilitates more precise spatial investigation. Previous analysis using the TTS data showed higher rates of walking for children and youth aged 11–13 years than older youth (Buliung et al., 2009). The subpopulation for this study has been fixed to the 11–13 years age group, to explore school travel behavior for children who are more likely to walk. Walking mode share for school trips was estimated for both the a.m. (6:00–11:30, interval includes 99.8% trips to school) and p.m. (14:30–16:00 p.m., interval includes 94.2% trips from school to home) periods.

2.2. Urban area classification A typology of urbanized spaces was created to facilitate the development of an understanding of variation in spatial clustering across the regional landscape. The classification process involved the conceptual and empirical allocation of TAZs to three broad urbanization categories that are largely similar within themselves with respect to the historical period of development, and urban design characteristics: (1) urban, (2) inner-suburban, and (3) outer-suburban (Fig. 1); this classification is widely used and understood within current discourse on urbanism in the study area (Sewell, 2009; White, 2007 for detail). The ‘‘urban’’ area follows the 1996 boundary for the Toronto census-subdivision (CSD), i.e., the boundary of the pre-1998 City of Toronto. This area contains the historic downtown, which is also the largest employment district within the GTA, as well as the neighborhoods constructed during the horse and electric streetcar eras, beginning in the mid-to-late 1800s. Most neighborhoods within this urban boundary were developed prior to WW2 (Sewell, 2009). Many of the neighborhoods (or parts of them) within this urban category have undergone major urban renewal and re-urbanization processes over time, where mid-to-high rise residential or mixed-use developments have replaced older commercial and/or residential buildings. The built environment within this boundary is represented largely by smaller city blocks and high-density development. There are, however, commercial districts in the downtown core where post-WW2 land assemblages resulted in the development of super-blocks that can be found in, for example, the financial districts of most North American cities. Land use is often mixed, and the street network follows mostly a traditional grid-pattern, a harbinger to the development of the transport infrastructures prior to the 20th century, where walking and transit represented the dominant modes of transportation for all activities.

The inner-city is surrounded by the ‘‘inner-suburban’’ neighborhoods of Toronto, developed largely between the mid-1900s and the 1970s (Sewell, 2009). The inner-suburbs represent an interesting mix in terms of built environments, ranging from high density, high-rise, mixed-use developments to low density, suburban-style, and residential neighborhoods. In general, however, the urban form in the inner-suburbs is more mixed-use compared with typical suburban areas. The rest of the GTA, outside the inner-suburbs, is defined as ‘‘outer-suburban’’, and is dominated by post-1970s neighborhood construction, with low to medium density, primarily residential developments. Street patterns are mostly curvilinear (e.g., looped streets) and land uses are often highly segregated. This part of the GTA also hosts some of Canada’s largest emerging suburban cities such as Mississauga, and some of Canada’s early planned suburban neighborhoods.

2.3. Analytical approach School travel practices have been situated within a spatiotemporal analytical framework. Spatial cluster analysis was employed to identify concentrations of AST within the study area. Next, Markov transition matrices were used to document changes in the spatial pattern of AST practices over time. These exploratory analyses were then followed by a logistic regression analysis specified to explain the relationship between a preliminary set of socio-demographic and environmental factors, and the likelihood of zonal membership within the set of AST cluster categories developed from the spatial cluster analysis. Spatial cluster analysis has received considerable attention as an approach to the surveillance and monitoring of disease and illhealth (Carrel et al., 2009; Rogerson and Yamada, 2009); an emerging literature in physical activity research has also employed spatial cluster analysis (Gorely et al., 2007; Huang et al., 2009). Few examples, however, exist of the application of cluster detection in research conducted at the nexus of health and travel behavior. While Huang et al. (2009) used Kulldorff (1997) spatial scan statistic to study the active travel of adults in Southern California; nothing has been published on the spatial clustering of school travel, or AST in particular. To the best of our knowledge, the current study is the first to combine spatial statistics with a Markov transition matrix approach, to describe the spatio-temporal pattern of a health-related outcome (i.e., walking for school travel) within a large metropolitan region. The spatial clustering of walking rates across the GTA was explored in two steps, and separately for the a.m. and p.m. periods. First, a region-wide analysis of the extent to which

Fig. 1. Urban form in the GTA—three different types of built environment with elementary schools at the center. Source: Digital Orthophotos, 1995—Greater Toronto Area, Triathlon Mapping Corporation; obtained through the University of Toronto Libraries, Canada, 2008.

ARTICLE IN PRESS R. Mitra et al. / Health & Place 16 (2010) 646–655

adjacent zones produced similar rates was conducted using global Moran’s I. While the global Moran’s I provides information on whether the spatial pattern in walking school trips does or does not deviate from a random spatial distribution, the index, and global statistics in general, offer no information on the size, location, or statistical contribution of local places to any broader pattern of spatial similarity or difference (Rogerson and Yamada, 2009). The detection of local clusters or walking hot spots within the study area, then, requires a local statistical analysis (Anselin, 1995; Getis and Ord, 1992; Ord and Getis, 1995). Example statistics that support local analysis include Anselin’s (1995) local Moran, the spatial scan statistic (Kulldorff, 1997), and G statistics (Getis and Ord, 1992; Ord and Getis, 1995). The combined global and local scale analyses offer a greater depth of understanding of the relationship between geographical location and a behavioral outcome of interest (Ord and Getis, 1995). A global or general analysis of the spatial pattern of events (e.g., incidence of disease, or preventive behavior) may proceed with minimal a priori knowledge or expectation about the presence or location of clusters (Besag and Newell, 1991). Local analyses include situations where a priori expectations or knowledge regarding the presence and/or location of spatial clusters, may or may not exist. Besag and Newell (1991) distinguish focused tests that are helpful in describing processes that produce clusters of disease around some environmental hazard, from those experiments centered on cluster detection only. In this paper, we subscribe to this second case and adopt the Gni statistic (Getis and Ord 1992; Ord and Getis, 1995) to detect active school travel (i.e., walking) clusters within the study area. We are not as much interested in spatial association, as we are in spatial concentration, and so, for example, the special case of negative spatial auto-correlation, that would be detectable using local Moran’s (i.e., zones with high or low values for the variable in question, surrounded by zones with dissimilar values) has not been explored here. The mathematical expression of the Gni statistic, as employed in our analysis, takes the following form: n P j¼1

wij xj x

n P j¼1

wij

Gi ¼ vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi  2ffi ; 8i ¼ j u n n u P 2 P tn wij  wij j ¼ 1 j ¼ 1 s n1

ð1Þ

where xj is the walking mode share for a TAZ j, x the mean walking mode share for the GTA, s the standard deviation of the xj’s, n the number of observations (zones) in the study area, and wij denotes a binary (i.e., one/zero) spatial weights matrix (Getis and Ord, 1992). We applied a first-order neighborhood contiguity rule for the weight matrix wij, assuming spatial auto-correlation only among the neighboring TAZs. Following this rule, wij equals 1 (one) if a TAZ j is an estimation zone (i.e., TAZ i), or an immediate neighbor to it (i.e., shares a common boundary with zone i), and 0 (zero) otherwise. The Gni statistic was computed using ArcGIS 9.2&, and the results are reported in terms of the standard normal variate, ZGni , for the individual TAZs; the null hypothesis of no spatial association with the neighboring TAZs was rejected at a = 0.05 (i.e., 95% CI). The Gni statistic was estimated separately for the a.m. and p.m. periods, to facilitate further analysis of the extent to which this spatial process changes across these time periods. One limitation of this analysis is that the first-order neighborhood contiguity rule does not account for the geometric properties of the neighbors (e.g., scale and topology). For example, the potential diminishing spatial relation among larger neighbors is mostly ignored (Getis, 2009). In response, some have argued in favor of using a standardized weights matrix, such as one defined

649

based on distance (Craglia et al., 2000; Le Gallo and Ertur, 2003; Ord and Getis, 1995), although the empirical advantage of such an approach remains unclear (Getis, 2009). While a standardized weights matrix may improve the spatial clustering analysis of larger geographical units (e.g., city, county), we argue that the application of such an approach may not be essential when exploring smaller geographical entities, such as the TAZs in this study. In order to examine the mobility of the clustering process between the a.m. and p.m. periods in further detail, a Markov transition matrix approach was adopted (Table 1). A transition matrix can be used to explore how spatial units migrate across classes, defined by a finite class-space, over time (Hammond and Thompson, 2002; Quah, 1996; Rey and Ye, 2010). The data within a transition matrix are the aggregate probabilities associated with the transition of spatial units across multiple pre-defined classes over time. In other words, each cell in the transition matrix represents the probability that a spatial unit will move from one category to another between discrete time intervals (Collins, 1975; Hammond and Thompson, 2002). The approach reveals more about the temporal stability of, in this particular case, walking outcomes over time (i.e., the persistence of walking within zones between the a.m. and p.m. periods), than would an exploration of trends using typical summary measures such as univariate mean, median, or standard deviation. This study explored the mobility of the TAZs with regard to their membership in the spatial clusters of high or low walking mode share. Hence, for the transition matrices, the range of ZGni estimates were discretized into three classes: (1) clusters of high walking rate: zones with significantly positive ZGni at a =0.05, (2) clusters of low walking rate: zones with significantly negative ZGni , and (3) unclustered/random zones: zones where ZGni scores were not statistically significant at a =0.05. To summarize mobility across the cluster classes defined above, the Shorrocks trace index (M) was applied to the transition matrices (Shorrocks, 1978). The index, M, belongs to a broader family of stability indices that provide simple measures for summarizing the mobility of a distribution across time (Hammond and Thompson, 2002; Maasoumi and Zanvakili, 1984; Shorrocks, 1978). This approach was preferred over the test for the Markov property of temporal independence (using a w2 test). While the Markov w2 test would identify the presence or absence of statistically significant mobility among the TAZs (Collins, 1975), the Shorrocks M index enables a relative comparison of stability/ mobility across the urban, inner-suburban and outer-suburban places, providing the opportunity to make qualitative statements about the types of places where walking is typically more persistent across time. The index is calculated as follows: M¼

#Classes-TraceðPt;t þ s Þ #Classes-1

ð2Þ

where #Classes represents the total number of classes within a finite class-space (for the purpose of this research, #Classes=3, see Table 1), and the Trace(Pt,t + s) corresponds with the sum of the diagonal probabilities of the transition matrix (i.e., zones remaining in the same class across the a.m. and p.m. periods). The maximum value for M is sensitive to the number of classes in the matrix. For the transition matrices in Table 1, the value of the index will range between 0 (when the diagonal probabilities are all equal to 1) and 1.5 (when the diagonal probabilities are all equal to 0). The estimated value of M signals the extent to which the production of walking school trips is temporally persistent. The closer the value of M is to its upper limit, the more mobile the TAZs are between the a.m. and p.m. periods with respect to the spatial clustering process. In other words, a value closer to the upper limit (i.e.,  1.5) is indicative of considerable

ARTICLE IN PRESS 650

R. Mitra et al. / Health & Place 16 (2010) 646–655

Table 1 Transition matrices—clustering pattern of high and low walking rates in the GTA, 2001 (a.m. vs. p.m. periods).a,b,c GTA

High walking rate, a.m. (n= 172) Unclustered, a.m. (n= 1313) Low walking rate, a.m. (n= 63) Sharrocks index (a.m.–p.m.), MGTA

High walking rate, p.m. (n= 170)

Unclustered/random, p.m. (n= 1223)

Low walking rate, p.m. (n= 155)

0.75 0.03 0.00 0.21

0.25 0.90 0.06

0.00 0.07 0.94

High walking rate, p.m. (n= 58) 0.86 0.07 0.00 0.12

Unclustered/random, p.m. (n= 96) 0.14 0.89 0.00

Low walking rate, p.m. (n= 5) 0.00 0.04 1.00

High walking rate, p.m. (n= 74) 0.70 0.09 0.00 0.20

Unclustered/random, p.m. (n= 246) 0.30 0.91 0.00

Low walking rate, p.m. (n= 2) 0.00 0.00 1.00

High walking rate, p.m. (n= 38) 0.68 0.01 0.00 0.24

Unclustered/random, p.m. (n= 881) 0.32 0.89 0.06

Low walking rate, p.m. (n= 148) 0.00 0.09 0.94

Urban High walking rate, a.m. (n= 59) Unclustered, a.m. (n= 99) Low walking rate, a.m. (n= 1) Shorrocks index (a.m.–p.m.), Murban

Inner-suburban High walking rate, a.m. (n= 76) Unclustered, a.m. (n= 245) Low walking rate, a.m. (n= 1) Shorrocks index (a.m.–p.m.), Minnersuburban

Outer-suburban High walking rate, a.m. (n= 38) Unclustered, a.m. (n= 967) Low walking rate, a.m. (n= 62) Shorrocks Index (a.m.–p.m.), Moutersuburban a b c

Spatial clusters are reported at the level of traffic analysis zones (TAZ). Classes are defined on the basis of ZGni values, at a 95% CI (high walking rate: Z4 1.96, unclustered: 1.96 Z ZZ  1.96, low walking rates: Zo  1.96). Mmaximum =1.5 (absolute mobility); Mminimum = 0 (absolute stability).

movement of places in and out of the cluster categories (high, low, and random/unclustered) described previously, while a value closer to the lower limit (i.e.,  0) suggests that TAZ membership within cluster classes changes very little between the a.m. and p.m. time periods. In all likelihood, the most desirable and manageable outcome, from a public health and planning perspective, is for the cluster process to exhibit temporal persistence, i.e., having large transition probabilities along the diagonal which are indicative of the TAZs remaining in the same class during the a.m. and p.m. periods. Interventions tailored to persistent or habitual behavior could be more readily constructed. While the Gni statistic describes the relative spatial concentration of walking practices, and the transition matrices (and the Shorrocks M) summarize the stability or persistence of the spatial pattern of walking school trips over time, these exploratory approaches offer less in the way of explanation. It is widely understood that the strength of these techniques lie in their ability to describe processes in space-time, with a view to establishing hypotheses requiring further investigation (Rey and Ye, 2010). In the final part of the analysis, we attempt to explain, albeit in a preliminary manner, TAZ membership in the cluster categories described previously, using a multinomial logistic regression analysis. Separate logistic regression models were estimated for the a.m. period and p.m. periods, precisely because the spatial distribution of high and low walking clusters was shown to vary between these two discrete time periods. The unclustered case (i.e., the random case) was the reference category for both models. While the Gni statistic and transition indices were calculated for all 1548 TAZs, the sample was modified for logistic regression estimation purposes. TAZs having no resident population were excluded, as were cases with missing data; the estimation sample included 1088 cases. The independent variables are described, and estimation results are summarized, in Table 2.

3. Results 3.1. Spatial variation in the walking school trips The walking mode share, of 11–13 year old children, for journeys to and from school is shown in Fig. 2. Global analysis of spatial autocorrelation, using Moran’s I, suggests the presence of spatial autocorrelation in walking rates at the region-wide scale (Ia.m. =0.38, po0.01; Ip.m. =0.38, po0.01). While the global statistic (i.e., Moran’s I) also revealed similar degrees of spatial auto-correlation between the two time periods; specific local differences in the production of walking school trips across zones remain unclear from this global scale analysis. The subsequent local cluster analysis (Gni statistic) identified walking hot-spots in the GTA (Fig. 3). When the walking mode share of a TAZ is high, and that of its surrounding TAZs are also high, the TAZ is part of a hot-spot (i.e., a cluster of zones with higher uptake of walking school trips). Fig. 3 demonstrates that the prevalence of walking as a mode for school transportation varied across the typology of urban spaces. When compared against the outer-suburban GTA, active school commuting was more common in the urban (i.e., Old City of Toronto) and inner-suburban areas. Within the Old City of Toronto, however, walking occurred more commonly in the neighborhoods that are residential in nature. The downtown employment district, located at the central part of the city and extending toward the north, had fewer clusters of high-walking zones. The hot-spot analysis revealed multiple concentrations of spatial clustering in the inner-suburbs of Toronto. Fig. 3 also shows a linear pattern of clustering in the north-south direction both to the east (i.e., Scarborough) and west (i.e., Etobicoke) of the Old City of Toronto. In general, these corridors contain low-to-medium income neighborhoods, with some concentrations of high-rise residential developments. In the outer-suburban GTA, spatial clusters of high walking

ARTICLE IN PRESS 651

mode share were relatively uncommon; which is not surprising given that these suburban locations would typically be expected to have automobile-oriented street designs and lower density development; design qualities that are typically identified as barriers to the use of active modes for school trips (McMillan, 2007; Schlossberg et al., 2006). But interestingly, there were some spatial clusters of high-walking zones in the outer-suburbs, especially to the west of Toronto in the cities of Mississauga and Brampton, which are also known to have the typical suburban design characteristics mentioned above.

b

Coefficients in bold-italics are significant at po 0.01; coefficients in bold are significant at p o0.05, Coefficients in italics are significant at p o 0.10. All variables were measured at the TAZ of residence, i.e., for the a.m. trips, at the TAZ of trip production, and for the p.m. trips, the TAZ of trip destination. c Data source: Population Census 2001, Statistics Canada. Standardized median income derived from dissemination area (DA) level median household income. d Data Source: TTS 2001. Number of schools per sq. km. within the TAZ. Only schools attended by the sample population were included in the analysis. e Data source: TTS 2001.

3.2. Spatial clustering in time

a

0.19(0.03  1.45) 0.00(0.00  2.62e97) 0.92 (0.82 1.04) 0.19(0.06  0.61)  1.654  12.630  0.079  1.646 1.907 1.263  0.005  1.744 1088 1312.424 (df= 2174) 1078.023 (df = 2164) 234.401 (p o 0.001) 0.18 (0.19 adj.)  13.292  12.843  0.116  1.315 7.78(4.37  13.84) 3.02(1.88  4.84) 0.90(0.97  1.26) 0.77(0.32  1.85)

0.00(0.00  9.3e192) 0.00(0.00  1.3e146) 0.89(0.73  1.09) 0.27(0.06 1.19)

6.73(3.94  11.52) 3.54(2.23  5.60) 1.15(1.02  1.30) 0.18(0.08  0.37)

1.00 (1.00-1.00) 0.05(0.01  0.19)  0.000  3.046  0.000 0.141 1.00 (1.00  1.00) 0.00(0.00  0.21)  0.000  7.401 1.00(1.00–1.00) 1.10(0.97  1.26)

Median household income of a TAZ  0.000 0.098 Density of schools within a TAZ d Regional location of the residence (ref.: outer-suburban GTA) Urban GTA 2.051 Inner-suburban GTA 1.105 Employment to population ratio within a TAZ e  0.101 Intercept  0.260 n 1088 Null deviance:  2[L(0)] 1214.086 (df =2174) Residual deviance:  2[L(B)] 920.731 (df= 2164)  2[L(0) -L(B)]: 293.355 (p o 0.001) 0.24 (0.25 adj.) McFadden’s r2

Odds ratio (95% CI) Odds ratio (95% CI)

1.00(1.00  1.00) 1.15(1.02 1.30)

Odds ratio (95% CI) Coef. Coef.

Coef.

Membership of high-walking rate cluster (n =166) Membership of low-walking rate cluster (n= 31) Membership of high-walking rate cluster (n= 170)

Odds ratio (95% CI)

Coef.

c

p.m. Period a.m. Period

Table 2 Logistic regression results—the likelihood of a TAZ’s membership within a high or low walking-rate cluster (reference: unclustered TAZs).a,b.

Membership of low-walking rate cluster (n= 49)

R. Mitra et al. / Health & Place 16 (2010) 646–655

With regard to the spatial-temporal change of AST practices at the local level, Fig. 3 shows that many TAZs that had significantly positive ZGni values in the a.m. period were not statistically significant in the p.m. period, and vice versa; demonstrating temporal heterogeneity. Additionally, the temporal variation in the spatial clustering of walking was different for different types of urban places. While the urban GTA (i.e., Old City of Toronto) appeared to be temporally stable (or perhaps behaviorally consistent) in terms of the locations of walking hot-spots, innersuburban and outer-suburban TAZs demonstrated greater inconsistency, and the locations of spatial clusters drifted across space between the a.m. and p.m. periods. Examination of transition probabilities confirms the qualitative assessment of the mapped ZGni scores. In Table 1, the rows show the distribution of ZGni scores across the three cluster classes in the a.m. period, and the columns show their distribution in the p.m. period. The cell values in the transition matrices represent the transition probabilities between the a.m. and p.m. periods, i.e., the probability that a zone belonging to a particular class in the a.m. period (rows) remains within its a.m. class, or moves to any other class in the p.m. period (column). The diagonal elements are the proportion of zones that remain in the same class; high transition probabilities in the diagonal suggest little temporal mobility in the spatial process. For the entire GTA, the diagonal transition probabilities suggest that there was little mobility in the position of zones within the defined class-space (i.e., changes in the share of zones within the high, random/unclustered, low classes) between the a.m. and p.m. time periods. The Shorrocks M index was calculated using (2) (MGTA = [3–(0.75+ 0.90+ 0.94)]/[3–1]= 0.21). This result changed somewhat when the transition probabilities were estimated separately for the different broad urban-place categories. Table 1 suggests that in the urban GTA, places with or without walking hot-spots essentially retained their a.m. period qualities during the p.m. time period (Murban = 0.12). The zones appeared to exhibit greater temporal change in the innersuburban GTA (Minnersuburban =0.21); and even more so in the outer-suburban areas in the GTA (Moutersuburban =0.24). However, the Shorrocks M index does not yield complete information about the temporal consistency across the classes within a distribution (Hammond and Thompson, 2002). Since the index only relies on the diagonal probabilities, mobility across the non-diagonal elements within a transition matrix remains unaccounted for. An exploration of the non-diagonal elements of the matrix may then provide important insight into the spatial distribution of the clustering pattern. For example, Table 1 reveals that only 70% of the inner-suburban TAZs that represent spatial clusters of high walking mode share for the a.m. period remained in the same class during the p.m. period. In addition, 9% of innersuburban zones that did not demonstrate spatial clustering in the a.m. period became walking hot-spots in the p.m. period. In the outer-suburban GTA, even fewer TAZs (68%)

ARTICLE IN PRESS 652

R. Mitra et al. / Health & Place 16 (2010) 646–655

Fig. 2. Mode share of walking-trips to and from school in the GTA, 2001.

consistently remained walking hot-spots across the a.m. and p.m. periods.

3.3. Correlates of spatial clustering Many of the qualitative interpretations offered in light of the spatio-temporal description of walking school trips were substantiated by the logistic regression analysis (Table 2). Prior to further discussion, it is important to emphasize that the unit of analysis is the TAZ—i.e., we have not modeled the walking behavior of the individual students; any heterogeneity of walking practices within each TAZ was not taken into account. Further-

more, at the TAZ level, the models do not directly measure the probability of the production of high or low walking rates. Rather, the estimation results communicate the independent effects of socio-demographic and environmental correlates on the odds of TAZ membership within a walking hot-spot (high or low), relative to remaining outside of such places (i.e., unclustered). Table 2 reveals that the regional location of a TAZ is associated with the clustering process. The odds of TAZ membership within a high walking cluster, compared with remaining unclustered, was greater within the urban and inner-suburban parts of the GTA, than in the outer-suburban areas. Also, the odds of TAZ membership within a high-walking cluster decreased with an increase in the median household income of the TAZ, although the effect size of this variable

ARTICLE IN PRESS R. Mitra et al. / Health & Place 16 (2010) 646–655

653

Fig. 3. Hot-spots of walking school trips in the GTA, 2001—temporal variation between the a.m. and p.m. periods.

on the clustering processes was minimal. Zonal coverage by elementary schools, measured as the density of schools within the TAZ of residence, did not associate with clustering in the a.m. period; for the p.m. period, school coverage increased the odds of TAZ membership within the high-walking cluster category, compared to the case of being in the unclustered category. In contrast, land use mix, measured as the ratio of employment to population, negatively associated with TAZ membership in the high-walking category in the a.m. period, and did not correlate with spatial clustering for the p.m. period trips. Overall, the models explained a.m. period clustering processes better, compared to the p.m. period (adjusted r2am =0.25, adjusted r2pm =0.19).

4. Discussion and conclusions This study explored walking as a travel mode for school transportation in the GTA. One of the goals of the article was to demonstrate the efficacy of quantitative spatio-temporal analysis as an approach to providing public health professionals and urban planners with a better understanding of the prevalence of a potentially preventive health behavior across an urban region. More specifically, the study examined whether there are patterns of systematic clustering of walking as a mode of school travel across the regional landscape, whether spatial variation in walking persists over time (i.e., for trips to and from school), and whether local/zonal spatial characteristics may associate with the formation of spatial clusters of walking school trips. The results generally confirm a priori expectations regarding the distribution of walking practices across the prescribed urban

typology. In the urban GTA (i.e., the Old City of Toronto), and in the inner-suburbs, active school travel was well-practiced when compared with the outer-suburban GTA. The findings also indicate that although the location of the high and low walking rate clusters appeared to change very little at the scale of the entire study area (i.e., between the a.m. and p.m. periods), the extent to which places became more or less active over time tended to vary with the assignment of zones to our typology of urban, inner-suburban, outer-suburban places (Table 1). The explanatory logistic regression analysis (Table 2) reveals that socio-economic status of the households in a neighborhood may partly explain these local variations; low income neighborhoods were more likely to be parts of walking hot-spots. The finding, in general, is in accordance with the existing literature that reports an association between economically disadvantaged neighborhoods (i.e., low median household income) and higher AST among children (McDonald, 2008a), although, based on our estimated models, the effect size of neighborhood level household income was trivial (odds ratio =1.00). Also, while evidence is inconsistent regarding the association between land use mix and school travel (Ewing et al., 2004; McMillan, 2007; Pont et al., 2009), our results demonstrated that land use mix at the TAZ scale negatively associated with TAZ membership in the high-walking cluster category for the morning period trips. School coverage, measured as the density of schools within the TAZ of residence, was also shown to have little impact on cluster membership during the a.m. period. This relationship reversed in the afternoon period, when clustering processes become more sensitive to the density of schools within the neighborhood, and less sensitive to the land use mix. Importantly, these findings support the

ARTICLE IN PRESS 654

R. Mitra et al. / Health & Place 16 (2010) 646–655

hypothesis that the strength of the relationship between various correlates, particularly built environment correlates, of active travel, may vary across time. In other words, built environment interventions—such as the amelioration of a jobs/housing imbalance (which our land use mix variable captures) could differentially impact the emergence of idealized travel behavior associated with school transportation. Furthermore, and contrary to the general pattern in the spatial distribution of high and low-walking clusters, the local spatial statistical analysis identified several walking hot-spots within the outer-suburban GTA (Fig. 3), which is particularly interesting from a policy perspective. For example, there could be a relationship between suburban hot-spots and the sort of traditional neighborhood design that can still be found within the small towns and cities that remain, in some ways, as historical artifacts (having developed during the 19th century) within the formerly ‘‘rural’’ but now outer-suburban parts of the study area. Although further research is warranted, this interesting dichotomy supports the argument that when formulating policy to affect healthy behavior, interventions should be tailored to the local context. School or neighborhood specific plans/ programs that leverage local enablers while addressing local barriers to AST, could be more successful than ‘‘one-size-fits-all’’ solutions designed and delivered by a regional planning authority. Clearly, the development of a detailed localized empirical foundation can also facilitate the production of appropriate interventions. Although the current literature emphasizes that household socio-demographic characteristics and the built environment are determinants of mode choice behavior for school transportation, employment, and mobility constraints faced by households may also give rise to the observed spatial and temporal variation in active school travel, especially for households living in the inneror outer-suburban areas. Households in these places could be spatially mismatched with respect to employment, and have longer commute times. Recent studies have found associations between parental employment status, work travel, and a child’s likelihood of walking for school transportation (McDonald, 2008b; Yarlagadda and Srinivasan, 2008). Activity and time scheduling in the morning and afternoon, and availability of vehicles, may limit a caregivers’ ability to accompany their children to, or from, school, potentially leading to the choice of a different mode in the afternoon than in the morning period. The density of schools, then, arguably becomes more salient in the afternoon period. For caregivers at work or otherwise occupied, it may be that their child(ren) walking home from school is then perceived as a viable and potentially necessary option, in the presence of an enabling parity between the distribution of schools and residences. Hypotheses with regard to the relationship between household activity-travel behavior and AST remain an important subject for future work. There are several limitations to discuss. First, an ecological approach has been adopted; and the travel behavior of individual households was not examined directly. Statistical analysis had been conducted on the average walking rates for the TAZs, assuming homogeneity of behavior within a TAZ. While the findings from this research may illustrate variation of walking practices across space and over time, and explain some of this spatial and temporal variation, they may not fully explain active travel practices at the household level. Second, the total number of reported school trips within a TAZ was not consistent between the a.m. and p.m. periods. In general, the 2001 TTS data showed fewer school trips in the afternoon period (school to home) than in the morning (home to school), possibly because of underreporting, or because some children tend to travel elsewhere after school hours than to home. This data discrepancy might have

influenced the observed temporal variation in spatial clustering. But, a rank correlation test between a.m. period and p.m. period trip production showed strong association (r =0.98), suggesting that the temporal difference in trip reporting was likely systematic across space, and therefore, should not affect the validity of our analysis. With regard to future research, then, it would be helpful to explicitly model the household AST process separately for different times of a day, while also controlling for household membership within the various classes of the typology of urban areas described in this study. The urban/suburban classification problem (more specifically, the definition of suburban or sprawled development) has remained an issue of considerable intellectual debate (Harris, 1996; Talen, 2003). The classification used in this study, largely influenced by the region’s historical and political geography, can be effective in terms of policy intervention, as it follows municipality (or former municipality) boundaries. However, the approach may not be adequately sensitive to the objective qualities of the built environment characteristics representative of an urban or a suburban area, particularly at the edge, or along the boundary, separating the geographical classes. Careful thought needs to be given to the influence of spatial effects of this sort on any explanatory analysis that is to follow. Ultimately, research findings regarding the role of regional forces in school travel decisions and outcomes, combined with evidence constructed at the local scale will, in due course, give rise to a multi-level/multi-scale understanding of the AST process, which will benefit the development of policy interventions focused on children’s mobility and physical activity. Such interventions will need to be adequately flexible in accommodating the potential spatial and temporal variability in AST outcomes that may exist across a region.

Conflict of interest statement The authors declare that there are no conflicts of interest.

Acknowledgements This research was funded by the Built Environment, Obesity and Health Strategic Initiative of the Heart and Stroke Foundation of Canada and the Canadian Institutes of Health Research (CIHR). References Anselin, L., 1995. Local indicators of spatial association- LISA. Geographical Analysis 27, 93–115. Besag, J., Newell, J., 1991. The detection of clusters in rare diseases. Journal of the Royal Statistical Society, Series A 154, 143–155. Black, C., Collins, A., Snell, M., 2001. Encouraging walking: the case of journey-toschool trips in compact urban areas. Urban Studies 38, 1121–1141. Boarnet, M.G., Anderson, C.L., Day, C., McMillan, T.E., Alfonzo, M., 2005. Evaluation of the california safe roots to school legislation: urban form changes and children’s active transport to school. American Journal of Preventive Medicine 28 (suppl 2), S134–S140. Buliung, R.N., Mitra, R., Faulkner, G.E.J., 2009. Active school transportation in the Greater Toronto Area, Canada: an exploration of trends in space and time (1986–2006). Preventive Medicine 48, 507–512. Carrel, M., Emch, M., Streatfield, P.K., Yunus, M., 2009. Spatio-temporal clustering of cholera: the impact of flood control in Matlab, Bangladesh, 1983–2003. Health and Place 15, 741–752. Collins, L., 1975. An Introduction to Markov Chain Analysis. Headley Brothers, London. Craglia, M., Haining, R., Wiles, P., 2000. A comparative evaluation of approaches to urban crime pattern analysis. Urban Studies 37, 711–729. Data Management Group, 2008. Transportation Tomorrow Survey. University of Toronto. /http://www.dmg.utoronto.ca/transportationtomorrowsurvey/index. htmlS. Accessed 21 December, 2009. Ewing, R., Schroeer, W., Greene, W., 2004. School location and student travel: analysis of factors affecting mode choice. Transportation Research Record 1895, 55–63.

ARTICLE IN PRESS R. Mitra et al. / Health & Place 16 (2010) 646–655

Faulkner, G.E.J., Buliung, R.N., Flora, P.K., Fusco, C., 2009. Active school transport, physical activity levels and body weight of children and youth: a systematic review. Preventive Medicine 48, 3–8. Frank, L.D., Engelke, P.O., Schmid, T.L., 2003. Health and Community Design: The Impact of the Built Environment on Physical Activity. Island Press, Washington, DC. Frumkin, H., Frank, L.D., Jackson, R., 2004. Urban Sprawl and Public Health: Designing, Planning and Building for Healthy Communities. Island Press, Washington, DC. Getis, A., Ord, J.K., 1992. The analysis of spatial association by use of distance statistics. Geographical Analysis 24, 190–206. Getis, A., 2009. Spatial weights matrices. Geographical Analysis 41, 404–410. Gorely, T., Marshall, S.J., Biddle, S.J., Cameron, N., 2007. Patterns of sedentary behavior and physical activity among adolescents in the United Kingdom:Project STIL. Journal of Behavioral Medicine 30, 521–531. Hammond, G.W., Thompson, E., 2002. Mobility and modality trends in US state personal income. Regional Studies 36, 375–387. Harris, R., 1996. Unplanned Suburbs: Toronto’s American Tragedy. John Hopkins, Baltimore. Huang, L., Stinchcomb, D.G., Pickle, L.W., Dill, J., Berrigan, D., 2009. Identifying clusters of active transportation using spatial scan statistics. American Journal of Preventive Medicine 37, 157–166. Kerr, J., Rosenberg, D., Sallis, J.F., Saelens, B.E., Frank, L.D., Conway, T.L., 2006. Active commuting to school: associations with environment and parental concerns. Medicine and Science in Sports and Exercise 38, 787–794. Kulldorff, M., 1997. A spatial scan statistic. Communications in Statistics: Theory and Methods 26, 1481–1496. Larsen, K., Gilliland, J., Hess, P., Tucker, P., Irwin, J., He, M., 2009. The influence of the physical environment and sociodemographic characteristics on children’s mode of travel to and from school. American Journal of Public Health 99, 520–526. Le Gallo, J., Ertur, C., 2003. Exploratory spatial data analysis of the distribution of regional per capita GDP in Europe, 1980–1995. Papers in Regional Science 82, 175–201. Maasoumi, E., Zanvakili, S., 1984. A class of generalized measures of mobility with applications. Economic Letters 22, 97–102. Martin, S.L., Lee, S.M., Lowry, R., 2007. National prevalence and correlates of walking and bicycling to school. American Journal of Preventive Medicine 33, 98–105. McDonald, N.C., 2008. Children’s mode choice for school trip: the role of distance and school location in walking to school. Transportation 35, 23–35. McDonald, N.C., 2008a. Critical factors for active transportation to school among low income and minority students: evidence from the 2001 National Household Travel Survey. American Journal of Preventive Medicine 34, 341–344. McDonald, N.C., 2008b. Household interaction and children’s school travel: the effects of parental work patterns on walking and biking to school. Journal of Transport Geography 16, 324–331. McMillan, T.E., 2007. The relative influence of urban form on a child’s travel mode to school. Transportation Research Part A 41, 69–79. Mitra, R., Buliung, R.N., Roorda, M., 2010. The built environment and active school travel in Toronto, Canada. The 89th Annual Meeting of the Transportation Research Board of the National Academies, Washington, DC.

655

Nelson, N.M., Foley, E., O’Gorman, D.J., Moyna, N.M., Woods, C.B., 2008. Active commuting to school: how far is too far? International Journal of Behavioral Nutrition and Physical Activity 5 (1, doi:101186/1476-. Ontario Professional Planners Institute, 2009. Plan for the Needs of Children and Youth: A Call to Action. OPPI, Ontario, Canada. Ord, J.K., Getis, A., 1995. Local spatial autocorrelation statistics: distributional issues and an application. Geographical Analysis 27, 286–306. Pont, K., Ziviani, J., Wadley, D., Bennett, S., Abbott, R., 2009. Environmental correlates of children’s active transportation: a systematic literature review. Health and Place 15, 849–862. Quah, D.T., 1996. Aggregate and regional disaggregate fluctuations. Empirical Economics 21, 137–159. Rey, S., Ye, X., 2010. Comparative spatial dynamics of regional systems In: Paez, A., Le Gallo, J., Buliung, R.N., Dall’erba, S. (Eds.), Progress in Spatial Analysis: Methods and Applications. Springer, New York, pp. 441–464. Robertson-Wilson, J.E., Leatherdale, S.T., Wong, S.L., 2007. Social-ecological correlates of active commuting to school among high school students. Journal of Adolescent Health 42, 486–495. Rogerson, P., Yamada, I., 2009. Statistical Detection and Surveillance of Geographic Clusters. Chapman & Hall/CRC, New York. Schlossberg, M., Greene, J., Phillips, P.P., 2006. School trips: effects of urban form and distance on travel mode. Journal of the American Planning Association 72, 337–346. Sewell, J., 2009. The Shape of the Suburbs: Understanding Toronto’s Sprawl. University of Toronto Press, Toronto. Sirard, J.R., Slater, M.E., 2008. Walking and bicycling to school: a review. American Journal of Lifestyle Medicine 2, 372–396. Shorrocks, A.F., 1978. The measurement of mobility. Econometrica 46, 1013–1024. Talen, E., 2003. Measuring urbanism: issues in smart growth research. Journal of Urban Design 8, 195–215. Timperio, A., Ball, K., Salmon, J., Robers, R., Giles-Corti, B., Simmons, D., et al., 2006. Personal, family, social, and environmental correlates of active commuting to school. American Journal of Preventive Medicine 30, 45–51. Transportation Alternatives, 2002. The 2002 Summary of Safe Routes to School Programs in the United States. New York. Tudor-Locke, C., Ainsworth, B.E., Popkin, B.M., 2001. Active commuting to school: an overlooked source of childrens’ physical activity? Sports Medicine 31, 309–313 van der Ploeg, H.P., Merom, D., Corpuz, G., Bauman, A.E., 2008. Trends in Australian children traveling to school 1971–2003: burning patrol or carbohydrates? Preventive Medicine 46, 60–62 Wen, L.M., Fry, D., Rissel, C., Dirkis, H., Balafas, A., Merom, D., 2008. Factors associated with children being driven to school: implications for walk to school programs. Health Education Research 23, 325–334. White, R., 2007. The Growth Plan for the Greater Golden Horseshoe in Historical Perspective. NEPTIS Papers on Growth in the Toronto Metropolitan Region. NEPTIS, Toronto. Yarlagadda, A.K., Srinivasan, S., 2008. Modeling children’s school travel mode and parental escort decisions. Transportation 35, 201–218.