University of Michigan-Flint campus in downtown Flint, Michigan, U.S. (Figure 1). ... Flint is largely a commuter campus with only one student housing complex ...
This is an accepted manuscript of an article published by Taylor & Francis in International Journal of Sustainable Transportation on Accepted 25 December 2017, available at http://www.tandfonline.com/eprint/XDFBEh6eqYSUzXZh3MAP/full
Toward a spatial understanding of active transportation potential among a university population Greg Rybarczyk, Ph.D. Abstract Universities and surrounding communities stand to benefit when active travel mode choices are elevated. Despite this, there is little research on travel mode choice at commuter universities and, in particular, the non-linear spatial relationships among active travel potential and various contextual and compositional factors. The purpose of this study was to examine and visualize linkages among personal, household, density, diversity, and design factors, and active travel (bicycling, walking, and mass-transit modes) among a commuter-university population residing throughout southeastern Michigan, USA. This was accomplished by employing exploratory spatial data analysis (ESDA), ordinary least squares (OLS) regression, and a geographically weighted regression (GWR) model. The GWR model outperformed the traditional OLS model in terms of goodness of fit (R2 = .534 and R2 = 461, respectively). A novel cartographic mapping technique was employed to depict where statistically significant parameter estimates negatively or positively influenced active travel. The main finding was that personal, household, density, diversity, and design estimates varied in both magnitude and spatiality throughout the university’s study area. Interestingly, distance was not a universal barrier to active travel potential. These variations emphasize the importance of promoting active transportation through localized interventions as well as coordinating efforts among universities and surrounding communities.
1. Introduction Determining how to increase use of active transportation among students, faculty, and staff has become a prominent item on policy and planning agendas at many universities. Shifting campus travel mode choice from reliance on cars to walking, bicycling, and use of public transit has numerous potential benefits. It can enhance a university’s image in the community; increase recruitment and retention of faculty, staff, and students; improve personal health; change personal attitudes toward sustainable transportation; and facilitate a travel mode shift in
surrounding communities (Balsas, 2003; Delmelle, 2012; Lundberg & Weber, 2014; Shannon et al., 2006; S. B. Sisson & Tudor-Locke, 2008; Toor & Havlick, 2004). 1.1 Background It is commonly accepted that active transportation includes walking, bicycling, masstransit travel modes (Ewing & Cervero, 2010). Past research has suggested these modes are influenced by a slew of multilevel factors (Bopp, 2011; Sallis et al., 2006), including personal attitudes (Lundberg & Weber, 2014), distance (Rybarczyk & Gallagher, 2014), monetary cost (Shannon et al., 2006), urban density (Saelens, Sallis, & Frank, 2003), roadway quality (Landis, Vattikuti, & Brannick, 1997), land-use mix (L. Frank, Engelke, & Schmid, 2003), socioeconomic conditions (Cervero & Duncan, 2003; Larsen, 2010), and the built environment (P. Zhao, 2013), including the availability of sidewalks (Rodríguez & Joo, 2004) and dedicated bicycling lanes (Cervero et al., 2009). Despite this evidence, our understanding of active travel determinants is limited because little of this prior research has explicitly addressed (1) the empirical and spatial variability of active travel determinants, or (2) specific factors influencing active travel at a local scale. The current study addressed these shortcomings using a disaggregated analysis coupled with impactful data visualization that provides valuable insight into where interventions will best encourage active travel. Finally, this study focused on a university environment rarely addressed in the research: the commuter university. Use of active transportation is relatively high at most major universities (Lundberg & Weber, 2014); however, it is very low at commuter universities, which remain unquestionably auto-centric. Commuter universities are characterized by minimal on-campus housing and a large number of non-traditional students, many of which have full or part-time jobs, are parents, and commute long distances to campus (Roe Clark, 2006). Furthermore, they typically contend with
a unique set of challenges, including scant options for active travel, lack of university-level support systems, and time constraints stemming from varied work schedules and multiple household/life roles (Jacoby, 1995; Roe Clark, 2006). Thus, encouraging active travel in this population is a challenge. The current study set out to approach this challenge by examining global and local influences on active transportation. In previous studies, ordinary least squares (OLS) regression models have been used to determine how attitudes, personal, household, and environmental variables affect travel mode choices on a global scale. For instance, Kerr et al, (2010) used multivariate OLS to determine car commuting predictors among students from three universities, and Shannon et al., (2006) used multivariate regression modeling of survey responses from staff and students at a large urban university to determine influences on travel mode choice. The major drawback of these global modeling approaches is that geographic variation is not recognized. Global model results are static and do not reveal the potentially complex interactions occurring at the local scale (Cardozo, 2012). Moreover, global models do not reveal the non-linear parametric or spatial relationships among the dependent and independent variables, thereby compromising parameter estimation. Geographically weighted regression (GWR) models improve upon OLS models by accounting for spatial heterogeneity. The GWR model has become a popular means of examining non-linear relationships and spatial variations among dependent and independent variables (S. Fotheringham, C. Brundson, M. Charlton, 2002b). The model has largely been used in the areas of development (Yu, 2006), geology (C. Zhang, Tang, Xu, & Kiely, 2011), ecology (Shaker, 2014), and geomorphology (Atkinson, 2003), with very little use in the area of transportation. Nonetheless, a few transportation studies are worth noting. GWR was used to
forecast traffic accidents (Zheng, Robinson, Khattak, & Wang, 2011), to estimate average annual daily traffic (F. Zhao, N. Park, 2004), and to measure accessibility (Mountain, Tsui, & Raper, 2007). More recently, Li et al., (2013) used a geographically weighted poisson regression model to determine spatially significant accident predictors at the county level. The results underscored the advantages of using a spatial model (rather than a generalized linear model) for crash analysis. Similarly, Cardozo et al., (2012) used a GWR model to forecast transit ridership at the station level and found the outcomes were more realistic than those produced using an OLS model. The mounting evidence that spatiality should be accounted for in transportation research validates the use of a GWR model in the current research. Furthermore, the present study sought to build on prior work by using a GWR model to highlight the heterogeneous parametric and spatial relationships among significant explanatory factors and active transportation potential among a commuter university population. 1.2 Study Area The study area used in this research consisted of a region surrounding and including the University of Michigan-Flint campus in downtown Flint, Michigan, U.S. (Figure 1). The city is located in mid-Michigan and contains a population of 102,434 as of 2010 (Bureau, 2010). UMFlint is largely a commuter campus with only one student housing complex on campus. The student body is predominantly non-traditional; the percent of students age 25 and older was 39% in 2011 (Michigan-Flint, 2012). The predominant travel mode choice among the campus, city, and regional population is the automobile.
Figure 1. State of Michigan and distribution of UM-Flint faculty, staff, and students
1.3 Objectives The specific aims of this study were to 1) identify important active travel correlates (personal, household, density, diversity, and design factors) using a combination of methods, including ESDA, GWR), geographic information systems (GIS), and a nuanced cartographic technique, 2) ascertain the utility of a spatially explicit model (i.e., GWR) for identifying and visualizing statistically significant local active travel predictors throughout the university’s region, and 3) suggest appropriate interventions that commuter universities and their local communities can use to encourage active travel mode choices. 2. Data The variables included in this research were selected in accordance with the ecological modeling framework. Grounded in public health, researchers have theorized that conditions at multiple levels influence physical activity (Sallis et al., 2006; Stokols, 1992). The datasets utilized in this research are founded on this past research, and others which have posited that
individual, household, and environmental factors influence active transportation usage (Bopp, 2011; Cervero et al., 2009; Whalen, Paez, & Carrasco, 2013). 2.1 University transportation survey For this paper, we drew on data acquired from an attitudinal survey disseminated to the UM-Flint population during 2010 and 2011. The purpose of the survey was to identify barriers to and facilitators of active travel among UM-Flint students, faculty, and staff. The participants included all faculty and staff; and a random sample of graduate and undergraduate students (15%). After quality assurances (i.e., data cleaning) were met, we amassed a total of 520 usable observations for this study. The survey response rate was 22 percent for students, 24 percent for faculty, and 28 percent for staff. We collected the following general information: university classification (i.e., student, faculty, or staff), residential address, commute distance, primary mode of transport to campus, and as previously mentioned, attitudes towards several hypothetical walking and bicycling facilitators and barriers. The primary variable of interest in this study was the percent active travel usage (i.e., proportion of faculty, staff, and students choosing active transportation) and was therefore the dependent variable in all exploratory and inferential analysis. In accordance with Shannon et al., (2006), active transportation was defined as use of mass transit, bicycling, and walking. The choice set was conflated for two reasons: 1) to provide reasonable policy and planning recommendations that will universally promote active transportation across a broad spectrum of travelers-as suggested by Sallis et al., (2006) and Handy (2005); 2) remain in line with past works which have treated mass-transit, bicycling, and walking as one mode in exploratory and inferential analysis (Celis-Morales et al., 2017; Klöckner & Friedrichsmeier, 2011). In addition to obtaining the proportion of active transportation use within the university sample, each participant’s residential address was
geocoded using Environmental Systems Research Institute (ESRI) ArcGIS software, version 10.0. The distance between each participants home and the university was then calculated using Euclidian distance. We considered distance a linear explanatory variable in this research, as it has been suggested that active commuting among a university population decreases as distance increases (Moniruzzaman & Farber, 2017; Rybarczyk & Gallagher, 2014). All survey data was entered into Microsoft Excel version 2010 and SPSS (IBM, Inc.) version 20 for empirical data analysis. 2.2 Government databases While personal and household variables were controlled for and obtained from the aforementioned university transportation survey, spatial and non-spatial variables used in this research fell under the design, diversity, and density (i.e., 3-Ds) framework established by Cervero et al., (1997). The majority of these variables were obtained from the U.S. Environmental Protection Agency (EPA) Smart Location database (SLD) version 2.0. The database contains 90 different travel-related correlates drawn from the 2010 decennial Census and American Community Survey (Ramsey, 2014). All factors were aggregated to the 2010 U.S. Census block group (CBG) level for the entire nation. The temporality of this data matches the timeframe of the survey distribution (2010-2011) and the university’s physical environment during the time of this study, ensuring that the survey participants’ responses were reflective of current conditions. The associations between the “3-D” variables and active travel have been well established. For instance, past research has shown that public transit, bicycling, and walking increases proportionally as land-use diversity and intensity increases (Cervero, 1996; Schwanen & Mokhtarian, 2005). Personal and household level socioeconomic factors also affect active travel, especially within a university population. Klockner et al., (2011) posited that university
students choose travel mode partly based on situational factors such as the cost of travel, and Maxwell (2001) found that students will drive if they have the economic means. The current study also incorporated bicycle and pedestrian crash density as a barrier (i.e., dangerous transportation environment) to active travel. Crash incident data from 2011 was obtained from the State of Michigan GIS data clearinghouse and summarized to each CBG. Due to the low sample size and non-normal frequency distribution, this variable was transformed using the square root. All possible explanatory factors were summarized for each 2010 CBG and normalized by area to reduce errors associated with the Modifiable Areal Unit Problem (MAUP). Each survey respondent’s trip origin (residence) was then joined to the CBG using a point-in-polygon operation in ArcGIS as this procedure is commonly used as a means to link neighborhood features to individuals (Cromley, 2012). The candidate variables, including the dependent variable, are listed in Table 1.
Table 1 Descriptive statistics for initial model variables (n = 520) Variable Description
Mean
Std. dev.
Dependent variable Active commuting
Percent faculty, staff, student who currently bicycle, walk, or take mass transit 11.61 to campus
29.09
Independent variables Personal Malea
Gender
.38
.487
Facultyb
University classification
.20
.399
Studentb
University classification
.49
.500
Distance to Campus
Euclidean distance to UM-Flint
13.81
14.99
Socioeconomic
Working age pop
Percent of population that is working aged, 2010
0.78
0.06
One-car households
Percent of one-car households in CBG
0.07
0.10
Two-car households
Percent of two-plus-car households in CBG 0.34 Percent households that do not own an 0.07 automobile # of workers earning $3333/month or more 0.17 (home location), 2010
0.17
Zero-car households High-wage workers
.102 0.21
Environment Density Res. density Population density Employment density Activity density
Gross residential density (housing units/acre) on unprotected land Gross population density (people/acre) on unprotected land Gross employment density (jobs/acre) on unprotected land Gross activity density (employment + HUs) on unprotected land
1.45
1.54
2.95
2.81
2.62
5.50
4.07
6.01
Diversity Land-use mix. Calculation based on population and total employment. Output is Regional job diversity 0.21 deviation of CBG ratio of jobs/pop from regional average of jobs/pop Employment mix I Jobs per household 3.73
0.26 46.84
Employment mix II
Employment and household entropy
0.45
0.23
Employment mix III
Household Workers per Job, by CBG
9.36
22.64
2.05
3.36
11.12
7.92
1.54
2.20
1.85
1.36
7.73
6.13
39.73
39.63
Design Road intersection density Road density Auto-orientated links Multi-modal links Pedestrian links Weighted street intersection
Auto-orientated street intersection density per square mile Total road network density Network density of facility miles of autooriented links per square mile Network density of facility miles of multimodal links per square mile Network density of facility miles of pedestrian-oriented links per square mile Street intersection density (weighted, autooriented intersections eliminated)
Intersection density of multi-modal intersections having three legs per square mile Intersection density of multi-modal Multi-modal intersections having four or more legs per intersections (4 legs) square mile Intersection density of pedestrian-oriented Pedestrian intersections having three legs per square intersections (3 legs) mile Intersection density of pedestrian-oriented Pedestrian intersections having four or more legs per intersections (3 legs) square mile Number of bicycle and pedestrian crash Bike/pedestrian crash incidents with automobiles per CBG area density (acres) a = reference – female, b = reference – UM-Flint staff Multi-modal intersections (3 legs)
8.92
11.18
4.27
7.42
27.19
25.83
11.38
17.79
0.01
.028
3. Methods 3.1 Exploratory data analysis The exploratory segment of this research consisted of carefully selecting explanatory variables, developing a global model, and testing for spatial autocorrelation using ESDA. Variables were selected by developing a correlation matrix to examine positive or negative relationships to the dependent variable. In addition, variables with a variance inflation factor (VIF) less than 10 were retained. Using SPSS software (IBM Inc.) version 20, the VIF procedure was used to test for multicolinearity. Values greater than 10 indicate multicolinearity, which reduces the explanatory power of the model (Cardozo, 2012; Mason, 1989). The final predictors were linked to previous literature, linearly related to the dependent variable, and uncorrelated among themselves. SPSS software was also used to develop the OLS model, which expressed the global relationships among dependent and independent variables. The OLS model essentially served as a baseline for comparison to the spatial model (i.e., GWR model). Once the OLS
model was developed, the homoscedastic assumptions were tested as these are normally violated when spatial data is used (Shoff, 2012). This involved carefully testing the OLS residuals for spatial heterogeneity, which is a common process when evaluating regression models for errors (O Sullivan, 2010). An ESDA was used to test for autocorrelation of the dependent variable, independent variables, and OLS model residuals. We utilized the global Moran’s I in an ArcGIS software environment to carry this out. The index is a common means to test the validity of global models, such as OLS (A. S. Fotheringham, M.E. Charlton, and C. Brunsdon, 2002a). If OLS residuals are clustered we can infer that spatial randomness is nonexistent, indicating the model is missspecified (Legendre & Legendre, 2012) and thus hampering interpretation of any inferential relationships (Dormann, 2007). The index produces one statistic for each variable; the autocorrelation index values range from -1 to + 1. Positive values indicate positive spatial autocorrelation and negative values highlight an inverse spatial relationship (Burt, 2009). 3.2 Spatial modeling The presence of spatial autocorrelation is nearly ubiquitous in the environment. This is in contrast to spatial homogeneity, where the mean and variance of the phenomenon are constant across space (Chi & Zhu, 2008). The GWR model accounts for spatial varying relationships; in other words, the relationships between dependent and independent variables are assumed to be spatially heterogeneous (Brunsdon, Fotheringham, & Charlton, 1996). The model is an extension of the OLS model in that it formalizes the spatially distinct influence that the explanatory factor(s) has on the dependent variable at each incident’s location (C. Zhang et al., 2011). The weighting strategy in GWR is based on distance decay which allows local influences to have a
stronger impact than those further away when producing estimations. The model produces an estimate for each independent variable X at location i; these estimates can then be visualized in GIS to determine the spatial variability of the localized parameter estimates (A. S. Fotheringham, M.E. Charlton, and C. Brunsdon, 2002a). The GWR model takes the form: 𝑦𝑦𝑖𝑖 = 𝛽𝛽0 (𝑢𝑢𝑖𝑖 , 𝑣𝑣𝑖𝑖 ) + � 𝛽𝛽𝑘𝑘 (𝑢𝑢𝑖𝑖 , 𝑣𝑣𝑖𝑖 )𝑥𝑥𝑖𝑖𝑖𝑖 + 𝜀𝜀𝑖𝑖 𝑘𝑘
Where (𝑢𝑢𝑖𝑖 , 𝑣𝑣𝑖𝑖 ) denotes the coordinates of the dependent variable y where (𝑢𝑢𝑖𝑖 , 𝑣𝑣𝑖𝑖 ) denotes the
coordinates of i, and 𝛽𝛽0 and 𝛽𝛽𝑘𝑘 represent the local estimate intercept and influence of factor k at
location i, respectively, and 𝜀𝜀 is the random error term to account for varying values across space (S. Fotheringham, C. Brundson, M. Charlton, 2002b; Matthews & Yang, 2012). As a result, a
continuous set of points are produced which represent the parameter values. The key characteristic of the GWR equation is locations closer to i possess a stronger influence on the estimation of 𝛽𝛽𝑘𝑘 (𝑢𝑢𝑖𝑖 , 𝑣𝑣𝑖𝑖 ) than locations further away. The spatial weighting scheme in GWR utilizes one of two functional methods: bi-square and Gaussian. Both methods utilize a spatial weighting matrix. The matrix essentially relaxes the hypothesized spatial dependency among the dataset (A. S. Fotheringham, C. Brunsdon, M. Charlton, 2000; O'Sullivan & Unwin, 2010). The weights are applied within a specified neighborhood that is either based on distance or a quantity of neighbors. In this research, a Gaussian distance-decay kernel weighting function was utilized as it is most commonly used with GWR (L. Zhang, Gove, & Heath, 2005). The Gaussian function takes the form:
𝑤𝑤𝑖𝑖𝑖𝑖 = 𝑒𝑒𝑒𝑒𝑒𝑒 �−
𝑑𝑑𝑖𝑖𝑖𝑖 � 𝜃𝜃 2
Where 𝑑𝑑𝑖𝑖𝑗𝑗 is the distance between observation i and j and 𝜃𝜃 is the bandwidth. The bandwidth parameter essentially regulates the scale of analysis. In this research, a golden search adaptive bandwidth kernel that set out to minimize the AICc (Corrected Akaike Information Criteria) was chosen because the distribution of the university population under investigation was inhomogeneous throughout the study area (Figure 1). This technique also prevents model over-fitting and provides an optimal bias-variance trade-off (A. S. Fotheringham, R. Crespo, J. Yao, 2015). The golden search adaptive bandwidth method assures that the scale of analysis changes to optimize the AICc throughout the study area. The study here included a minimum of 5% and a maximum of 20% of the neighbors around each point. The resultant AICc and R2 diagnostics were used together to measure the performance and goodness-of-fit for the OLS and GWR models. Both metrics are acceptable means to compare global and local models (S. Fotheringham, C. Brundson, M. Charlton, 2002b; O'Sullivan & Unwin, 2010). The OLS and GWR models were fitted using Spatial Analysis Macroecology (SAM) GIS software version 4.0 (Rangel, 2010). 3.3 Visualization of GWR outputs The visualization strategy consisted of highlighting the significant spatially varying relationships among the outcome and explanatory variables. The GWR model coefficients were mapped in ArcGIS using a novel cartographic strategy employing statistical significance and GWR parameter estimates. The model outputs used to examine the spatial distribution and influence of the explanatory variables on the outcome variable included the t-statistics and
unstandardized parameter estimates. Rather than display the coefficients in the traditional manner (i.e., a choropleth map with a sequential no-hue color scheme and an equal-step classification strategy) the current research combined the two coefficients to produce a bivariate choropleth map in accordance with Matthews et al., (2012), Mennis (2006), and Olson (1975). This strategy offers the advantages of reducing the stress of interpreting two maps simultaneously and focusing the map reader on the true magnitude of influence on the dependent variable. The first step of this cartographic process was the creation of an interpolated raster surface for each standardized parameter estimate and t-value. The interpolation technique used the common inverse distance weighting (IDW) method in ArcGIS. After the two surfaces were created for each explanatory variable, a masking technique was used to screen out insignificant estimates. The statistically significant t-values (95% CI, ± 1.96) were used to define the mask by applying 100% transparency to values greater than or less than 1.96, and effectively masking values within ± 1.96 by applying 100% opaque white. The results were then overlaid onto the continuous surface estimates. The final map displayed the statistically significant parameter estimates (±) for each explanatory variable from the GWR model. The magnitude and directionality of influence was represented by adjusting the color saturation. 4. Results and discussion First this study explored whether non-spatial factors (i.e., personal and household) and spatial factors (density, diversity, and design) significantly affected active travel potential among university students, staff, and faculty. Several multivariate OLS models were fitted using combinations of variables that were well suited based on correlation analysis, low
multicolinearity, and logical relationships to active travel. Final model selection relied on the selection of independent variables that were easily reproducible and offered the greatest model explanatory power. These explanatory variables were assessed using ESDA and then incorporated in the GWR model for statistically, and visually, examining their local relationship to the likelihood of increasing active transportation mode choices. 4.1 Results of exploratory analysis OLS model results are summarized in Table 2, which highlights the estimated magnitude and direction of influence for each selected active travel variable. The VIF among all variables was less than the reasonable cut-off value (i.e., < 10); therefore, multicolinearity was not an issue. The direction of influence for most variables is intuitive. For example, an increase in either distance to campus or bicycle/pedestrian crash density was associated with a decrease in the likelihood of active travel. The estimated magnitude of the association for bicycle/pedestrian crash density was not statistically significant, possibly due to differing behavioral responses to perceived versus real safety from active travelers, such as bicyclists (Heinen, 2010); whereas, the estimate for distance to campus was (i.e., p-value < 0.05). For every 1 percent increase in distance from campus, a person was estimated to be 0.24 percent less likely to use active travel (i.e., walk, bike, or use mass transit). This outcome is partially supported by previous research from Shannon et al., (2006), who identified travel time as a significant barrier to bicycling in a university environment. The remaining eight OLS model variables were shown to have positive influences on active travel. Three personal attributes male, student, and facultywere positively associated with active travel. These positive relationships are in part supported by Winters et al., (2007),
where they found that male students are more likely to bicycle than female students and Balas (2003) who posited that faculty are likely to bicycle as much as students. Among the personal attributes, the current study found student status had the strongest positive association with active travel. Two household factors representing demographics and socioeconomic status (SES)— percentage of zero-car households and percentage of population of working age—were also positively associated with active travel. The estimated magnitude of association for percentage of population of working age was not statistically significant. As expected, percentage of zero-car households possessed the greatest explanatory power. Each 1 percent increase in zero-car households resulted in an estimated 48 percent increase in the likelihood of using active travel. This outcome partly reflects the findings of Lundberg et al., (2014), who discovered a significant association between percentage zero-car households and increased pedestrian commuting. The last three variables to display positive influences on active travel potential were residential density, regional employment diversity, and road intersection density. Among all study variables, regional employment diversity was the second strongest predictor of active travel (percentage of zero-car households was the strongest predictor). Each 1 percent increase in regional job diversity resulted in an estimated 20 percent increase in active travel. Assuming job diversity relates to the degree of mixed-land uses, this estimate is supported by previous research such as Lamiquiz et al., (2015) and Saelens et al., (2003). The current study’s finding that residential density had a positive influence on active travel is supported by prior research linking population density and alternative mode-choices (L. D. Frank & Pivo, 1995). Finally, the positive influence of road intersection density on active travel is encouraging, as Marshall et al., (2014) recently reported that more compact street networks (i.e., elevated number of intersections) with smaller major roads are correlated with better health.
The associations generated by the OLS model appear to be in line with previous research. However, as in past works, the OLS model coefficients are generalized and there may be model violations due to spatial heterogeneity of the factors in the study area. Global models cannot identify spatial variations in the associations among certain sets of variables; therefore, ESDA and models such as GWR can overcome this limitation by identifying and accounting for spatial heterogeneity. The results of this finer analysis are outlined in the sections that follow.
Table 2 Summary of OLS model coefficients Variable estimate std-error
t-statistic
p-value
VIF
Constant
-4.371
1.927
-2.268
.024*
--
Personal Male
.881
.213
4.137
.000*
1.047
Faculty
.797
.298
2.678
.008*
1.373
Student
1.309
.235
5.577
.000*
1.343
Distance to campus
-.033
.007
-4.516
.000*
1.141
% zero-car households
4.862
1.437
3.383
.001*
2.076
% working age pop
4.582
2.577
1.778
.076
1.995
.256
.075
3.433
.001*
1.188
-5.942
4.333
-1.371
.171
1.530
2.979
.003*
2.382
6.528
.000*
1.891
Socioeconomic
Environment Residential density Bike/pedestrian crash density
Regional job diversity 1.759 .590 Road intersection .315 .048 density * statistically significant at the .05 level 4.2 ESDA outcomes
Before building the GWR model, the response and explanatory variables were tested for spatial autocorrelation (i.e., clustering) using ESDA. The Global Moran’s I results were expected as most covariates exhibited spatial clustering (Table 3). All but one independent variable (faculty)
displayed higher than normal estimated Moran’s I values. Statistically significant variables with the largest positive spatial dependency among them included: active travel, male, student, percentage zero-car households, residential density, road intersection density, regional employment diversity, crash density, and distance to campus. Although percentage of workingage population was clustered, the degree of spatial dependency was not statistically significant. Table 3 shows that clustering was significant among most variables, which means that the level of influence on the dependent variable is spatially invariant and a simple linear model (i.e. OLS) should not be used.
Table 3 Global Moran’s I for response and explanatory variables Variable Moran’s I Pattern
p-value
Dependent variable Active Travel (% bus, bike, walk) 1.355
Clustered
0.00*
Independent variables Distance to campus
0.839
Clustered
0.00*
Male
0.498
Clustered
0.00*
Faculty Student
-0.038 0.124
Random Clustered
.513 .022*
% working age pop
0.099
Clustered
.068
% zero-car households
1.301
Clustered
0.00*
Residential density Road intersection density
1.300 1.268
Clustered Clustered
0.00* 0.00*
Regional job diversity
2.120
Clustered
0.00*
Bike/pedestrian crash density
1.137
Clustered
0.00*
4.3 OLS and GWR comparison The second objective in this research was to determine the utility of incorporating a GWR model for inferential and visual examinations of active travel correlates. In doing so, we used the same explanatory variables in both OLS and GWR models. As shown in Table 4, the OLS model explained 46 percent of the total variance (adjusted R2) in active travel potential among university members. Stated differently, nearly half of the variability in active travel potential can be explained by the final ten explanatory variables (i.e., the independent variables listed in Table 3). The F-statistic and p-value for the OLS model show statistical significance and supported the inclusion of the selected variables; however, violation of the homoscedastic assumption is likely a problem. The GWR model explained more than 50 percent (R2 = .534) of the total variance in active travel potential. Compared to the OLS model (R2 = .461), this represents a 7 percent increase in robustness (Table 4). There was an increase in model fit as the AICc for the GWR model exhibited a 1.65 percent decrease from that of the OLS model. This decrease also indicated that the two models are significantly different (i.e., difference > 3), with the GWR model being the most parsimonious. Likewise, the residual sum of squares was less in the GWR model (Table 4) further proving its superiority. The test for autocorrelation among both model residuals indicated that the Moran’s I value was closer to expected for the GWR model compared to the OLS model (Table 5). Therefore, the GWR residuals indicated that the null hypothesis (i.e., complete spatial randomness) could be accepted as the p-value was not statistically significant.
Table 4 Estimated diagnostics of OLS and GWR models (n = 520) Variables OLS GWR 2 R .473 .582 2 Adjusted R .462 .534 AIC 4670.95 4606.95 AICc
4671.26
4614.10
CV Residual sum of squares
472.30 231,516.88
436.86 183,216.68
Table 5. Global Moran’s I statistics for OLS and GWR residuals Model
Moran’s I
variance
z-score
p -value
OLS GWR
.239 -0.049
0.003 0.003
4.37 -.862
.000 .388
4.4 GWR outputs The last objective in this research was to suggest effective policy and planning interventions by linking pertinent active travel predictors to place vis-à-vis GWR analysis. Table 6 shows the variability among the parameters used in the GWR model. The values are useful for interpreting the empirical differences between the reference model (OLS) and the spatially local model, and in understanding how different predictors affect active travel potential. That said, the GWR coefficients are generally in line with those of the OLS model (Table 2); however, significant empirical variation is now noticeable. As an example, the locally modelled distance to campus parameter estimate is both negative and positive, whereas it is only slightly negative in the OLS model. Further localized visualizations are therefore warranted.
Table 6 GWR parameter summary results Variable Min Max -265.33 28.75 Constant
Mean -75.40
Std. Dev. 64.92
Personal Male Faculty Student Distance to campus Socioeconomic % working age pop % zero-car households Environment Residential density Bike/pedestrian crash density Regional job diversity Road intersection density
-0.49 4.96 2.42 -7.70
28.93 51.56 68.12 19.20
7.48 7.73 12.40 -.019
2.39 1.67 2.89 0.31
-61.33 -2.26
209.72 80.82
45.71 49.23
17.55 6.77
0.84
1 2.48
2.30
0.58
-119.49
36.66
53.91
1 2.24
-14.18 -0.39
41.68 4.25
19.93 2.92
3.99 0.19
The GWR method allows one to plot and visualize the spatially heterogeneous coefficients of the explanatory variables, which is useful for implementing targeted interventions. The spatiality and degree of influence on active travel for the statistically significant (95% CI, ± 1.96) parameter estimates are illustrated at two scales representing the full university travel-shed (approximately 40 miles) and a focused area surrounding the university (i.e., 5 miles). The mass-transit routes and car-pool lots have also been added to provide transportation context. Figure 2 displays the spatially varying relationships between active travel and the explanatory variables resulting from the GWR models. The following personal explanatory variables: distance to campus (a, b), male gender (c, d), faculty status (e, f), and student status (g, h) are highlighted in figure 2. Notably, throughout a large portion of the study region, distance had either an insignificant, negative, or positive effect on the likelihood of active travel (Figure 2a). In areas southeast of campus distance appeared to be a significant barrier to active travel. The distance cost associated with use of active travel modes (e.g., mass transit) in these areas
may be minimized by ensuring that the current car-pool parking lots are safe, introducing safe bicycle facilities in these areas, or programming initiatives to encourage active travel. In areas furthest southwest and northwest, distance had a positive effect on active travel potential. This illogical result could stem from a survey interpretation error. In other words, the participants may be using active travel for recreation or commuting in their communities versus actually commuting to campus. Figures 2c and 2d shows that the likelihood of university males (students, staff, or faculty) to use active travel modes was highest for those living in areas surrounding and north of campus and lowest for those living in areas southeast and southwest of campus. In the area farthest south of campus, this male influence appeared negative. A possible explanation for this outcome may be that females (reference group) in this particular area use active travel modes more so than males. This explanation is feasible given previous studies that found females are just as likely as males to bicycle in a university setting (Sisson et al., (2008). An opposite tendency was observed for faculty members in the same region (south of campus) (figures 2e and 2f). Faculty in this area appeared very likely to use active transportation to reach campus when compared to faculty living in areas near the university (figure 2f) and northeast of campus. This overall trend may be attributed to the areas immediately south of campus being unsafe for active travel, or fewer faculty reside there. Interventions to promote active transportation in this local area should include efforts to increase the safety for walking and bicycling, and examining ways to increase mass transit usage. Overall, the likelihood of active travel use among faculty members is positive (except in areas immediately south of campus). Past research generally supports this finding; for instance, Winters et al., (2010) found that persons with higher education are more likely to bicycle. This may also explain why university students, like faculty members, generally favored the prospect of using active travel to reach campus. Student
estimates for active travel potential are largely positive and encompass a vast area north of and immediately surrounding campus (figures 2g and 2h). In areas far south of campus, however, students appear less willing to use active transportation to campus (Figure 2g). The difference between active travel potential between this group and faculty in the same area may be reflective of the unique challenges that commuter students face. These students may be constrained for time due to employment or home responsibilities which could hamper active travel potential. Considering the accessibility of mass-transit in this area, the university and surrounding communities could coordinate efforts to promote the use of the regional connector busses. An intervention that may increase active travel closer to campus should include “flex-pass” program as suggested by Whalen et al., (2013), This would allow students to pay for a daily parking pass rather than a seasonal pass where the parking costs are not realized. The program also has the advantage of inducing an alternative mode choice, such as bicycling. By and large, it appears that like faculty, students favor the prospect of using alternative transportation to reach campus throughout the region and especially in close proximity to campus (figure 2h). Figures 2i and 2h reveals a significant relationship between active travel and percentage worker-age persons at the trip origin (i.e., residence). As the density of working-age persons increased, the likelihood of alternative travel increased significantly north of campus (figure 2i). For each new working-age person added to areas north of campus, the likelihood of alternative travel use increases by approximately 200 percent (i.e., threefold). Given this outcome, we can infer that as density of workers increases active modes of travel will likely be utilized. The result is supported by previous research; for example, Rajamani et al., (2003) posited that individuals with increasing incomes will likely utilize carpooling, mass-transit, walking, or bicycling more than driving, albeit only to a middle-income income level. From a university perspective,
concentrated efforts to encourage active travel use should be focused in all other areas (as insignificant estimates prevail throughout the region). The second household SES indicator, percentage of zero-car households, is notable due to this factor’s heterogeneous spatiality and level of influence. Figures 2k and 2l depict that, except for insignificant estimates located close to campus (figure 2l), the study region was generally split between percent zero-car households having either a positive or a negative influence on active travel potential near the trip origin (figure 2k). The increased proportion of zero-car households had a positive effect on active travel south of campus. The area includes Detroit and its suburbs and consists of an intercounty transit network, albeit of poor quality, and a significant number of car-less households (Grabar, 2016). Thus, it is likely that the concentration of lower income families and access to mass-transit increases the probability of active mode-share usage by UM-Flint personnel in this region. An opposite condition was observed northwest of campus. One explanation is that the reduced quantity of car-less households in this zone (i.e., elevated incomes) and lack of a connected multi-county transit system hampers the propensity of using active transportation. This interpretation is partially supported by Sardianou et. al., (2015) who found a (although insignificant) relationship between car ownership and eco-cycling. To convince UM-Flint car owners north of campus to adopt alternative transportation, active travel modes must be perceived as both convenient and safe. Concerted efforts to promote active travel should include one or more of the following: lobby policy-makers and transportation agencies to extend transit routes north of the university, installing covered mass transit shelters in key areas to protect riders from the weather, developing programming materials related to the social efficacy of active travel, offering mass-transit subsidies (university or community sponsored), and advertising bicycle and pedestrian routes in this area and around campus.
According to Figures 2m and 2n, population density had a positive influence on active travel use throughout most of the region (although areas directly south and west of campus displayed no significant relationships). For each new person added to areas south of campus active travel potential increased approximately 12 percent. This outcome is supported by previous studies which found that population density had a significant positive influence on active travel modes such as walking (Saelens et al., 2003). In sparsely populated neighborhoods near campus (figure 2n), it is important to implement active travel interventions which will universally induce active travel modes. For instance, to encourage more UM-Flint personnel residing in this area to use active travel, concerted efforts between the university and the Metropolitan Transit Authority (MTA) focused on educational programming on how to utilize bus-bicycle racks could elevate this mode-share. The environmental design factor, road intersection density, was statistically significant, and either had a positive or a slightly negative influence on active travel throughout the region (figures 2o-2p). The density of intersections had a negative influence on active travel north and beyond the localized university travel-shed. In this area we can hypothesize that the environment may not facilitate active travel, perhaps due to traffic intensity or lack of active travel facilities. Active travel could be promoted in these areas by redesigning the travel environment to make it friendlier to bicyclists and pedestrians. The implementation of a “complete streets” policy, modifying street intersections to include traffic-control devices designed for bicycling and walking, or the university may work with regional mass-transit agencies to promote bus-bicycle rack usage on major transit routes. In contrast, the intersection density had a positive influence on active travel in areas west and east of campus (figure 2o). This positive correlation is likely attributed to the increased number of intersections, which by default increases connectivity. The
finding has been partly supported elsewhere; for instance, Winters et al., (2010) discovered that increased bicycling rates were associated with increased road intersection densities and Badland et al., (2008) found that street connectivity encouraged transport-related physical activity. The influence of the diversity metric, employment diversity (i.e., mixed land-use), on active travel varied substantially-both in directionality and magnitude –throughout the region (figures 2q and 2r). In areas immediately south of campus (figure 2r), increased regional employment diversity correlated with increased active travel potential. We can infer from this that an increase in employment diversity is correlated to increased potential for active travel. The outcome was expected and is supported by previous works related to pedestrian mobility (Lamiquiz, 2015). Areas farther south of campus exhibited an inverse relationship (figure 2q). For every 1 percent increase in employment diversity, there was an approximately 14 percent decrease in active travel. This counterintuitive result may indicate that increases in employment diversity will only encourage additional car commuting, despite access to regional mass-transit. We can conclude that the mix of employment (i.e., land-use) isn’t encouraging active mode shares. Therefore, to counter this effect communities and the university need to work on progressive policies and planning initiatives such as: increasing positive perceptions of regional and local mass transit, introducing facilities/programs that support walking or biking, or the implementation of an effective campus parking management plan. As an example, expanding bus routes into areas south of campus would facilitate overall mass transit, while universitysubsidized bus fares would lower the travel cost for university faculty, staff, and students. Figure 2s shows a statistically significant area of concern southwest of campus, where bicycle/pedestrian crash density negatively affected active travel. The magnitude of the negative correlation in this area (indicated by the deepest shade of green) indicates the likely presence of a
major transportation-environment safety issue, and may hinder active travel near campus. The perceived and real safety concerns in this area may become formalized as a negative stigma which could hinder the likelihood of using active travel on or near campus. Not surprisingly, this phenomenon is not new. Shannon et al., (2006) found that university staff were concerned about dangerous traffic scenarios more so than students from the same university. The result highlights why a combined effort between the university and local community would be a value-added approach to counter the deleterious psychological effects from exposure to unsafe transportation environments. Efforts may include a university programming effort to educate the campus population on bicycling safety or working with local communities on crash reduction strategies set on universally increasing travel safety, especially for active commuters.
Figure 2. Spatial distribution of statistically significant (95% CI, § 1.96) GWR estimates at a regional and local scale
5. Conclusion Universities are well positioned to increase use of active transportation on campus and throughout their local communities. Commuter universities, in particular, are positioned to increase active travel throughout their surrounding area. Accomplishing this requires a clear understanding of where and to what degree various factors affect active travel potential is affected by design, density, diversity, personal, and household factors. Prior research has explored the magnitude and direction of influence of these types of factors on active travel in various settings. Yet minimal research has either addressed the commuter university setting or the spatial variation in active travel correlates. The current research addressed this gap by examining the empirical and spatially varying influences of these factors on active travel within a commuter university context. This research contributes to a growing body of literature focused on disentangling connections between active travel determinants and university travel behavior. It also adds to recent research on effective spatial modeling strategies by establishing the utility of a GWR model for developing spatially targeted policy and planning recommendations. Using a combination of OLS, ESDA, and GWR methods, this research provided evidence that a) personal, household, density, diversity, and design measurands are spatially clustered, b) the relationships between the selected explanatory factors and active travel potential are empirically and spatially non-linear, and c) placed-based policy and planning initiatives to increase active travel need to be coordinated among commuter universities and community stakeholders throughout their region. Through the use of a global and spatially explicit model, we have provided a deeper understanding of the complexities surrounding active travel potential in a university environment. The OLS results in this article support similar past research that found personal,
household, diversity, density, and design factors universally affect active travel potential. The GWR model proved more robust than the OLS model, as evidenced by a 1.65 percent decrease in AICc and 7 percent increase in R2 (i.e., indicating better goodness-of-fit). The GWR mappings of statistically significant coefficients clearly revealed the nuanced spatial relationships between active travel and explanatory variables within the university travel-shed. Moreover, the GWR model provided finer, localized output from effectively creating a regression equation for each explanatory variable. Percentage of zero-car households displayed the greatest degree of spatial variation on active travel potential. The GWR model revealed that this SES factor possessed both positive and negative influences on active travel potential throughout the study area; however, it had no effect in neighborhoods near campus. The outcome suggests that efforts to increase active travel in areas with elevated incomes should differ than in areas where the proportion of zero-car households is elevated. Similarly, the personal factor distance also affected active travel both negatively and positively depending spatiality. This contradicts the common assertion that distance is linearly related to mode-choice (see Delmelle and Delmelle, 2012). Both of these outcomes suggest that active travel mode potential are dynamically reliant on intangible (personal and household factors) and tangible factors (design, diversity, density) that are ultimately based on geography. The influence of regional employment diversity was also important. The parameter estimate was statistically significant throughout most of the study area, including neighborhoods near campus; however, its spatial dependency and direction of influence were heterogeneous, suggesting that any intervention set on increasing employment diversity (i.e., land-use mixing) to encourage active travel for the campus population should be fitted to local conditions promoted within the university.
This study was intended to increase our understanding of how and where various predictors of active travel are statistically significant and, based on that understanding, to identify effective means of encouraging this active travel in selected areas. As such, it has necessary limitations. First, the GWR model is exploratory in nature; therefore, coefficients and mapped parameter estimates should not be interpreted as predictions. Nonetheless, the results of this study could be used to drive forecasting models. Second, additional mediating variables would likely increase the robustness of the models in this study. Therefore, follow-up studies should include the following additional contextual and environmental variables: topography (Heinen, 2010), weather (Dill & Voros, 2007), crime (Ferrell, Mathur, & Mendoza, 2008), and travel pathway visibility (Rybarczyk, 2014). Follow-up studies may also explore psychosocial factors (derived from a survey) to lend insight into how active travel is perceived among a university population. For instance, it has been previously noted that these variables significantly affect active commuting in a university setting (Molina-Garcia, Castillo, & Sallis, 2010). Despite its limitations, the current study succeeded in coupling the explanatory power of a GWR model with a conventional OLS model to improve our understanding of how various personal, and environmental conditions related to household, density, diversity, design factors, influence active travel within a commuter university context. Specifically, the study revealed that these influences varied in both their magnitude and spatiality throughout the study area. It is hoped that the results of this research will spur commuter universities to collaborative with surrounding communities interested in developing universal interventions to promote active transportation locally and regionally.
Acknowledgements The author would like to acknowledge the invaluable feedback provided by Ms. Mary Hall during the composition of this manuscript and for support from Theresa Landis, Director of UMFlint Recreation Services and Fawn Skarstan, Director of UM-Flint Institutional Analysis. 6. References Atkinson, P. M., S.E. German, D.A., Sear, M.J. Clark. (2003). Exploring the relations between riverbank erosion and geomorphological controls using geographically weighted logistic regression. Geographical Analysis, 35(1), 58-82. Badland, H. M., G.M.Schofield, N.Garrett. (2008). Travel behavior and objectively measured urban design variables: Associations for adults traveling to work. Health and Place, 14(1), 85-95. Balsas, C. J. L. (2003). Sustainable transportation planning on college campuses. Transport Policy, 10(1), 35-49. Bopp, M., A. Kaczynksi, P. Wittman. (2011). Active commuting patterns at a large midwestern college campus. Journal of American College Health, 59(7), 605-611. Brunsdon, C., Fotheringham, A. S., & Charlton, M. E. (1996). Geographically weighted regression: a method for exploring spatial nonstationarity. Geographical Analysis, 28(4), 281-298. Bureau, U. S. C. (2010). 2010 Census: U.S. Census Bureau. Burt, J., G. Barber, D. Rigby. (2009). Elementary statistics for geographers (3rd ed.). New York: The Guilford Press. Cardozo, O. D., J.C. Garcia-Palomares, J. Gutierrez. (2012). Application of geographically weighted regression to the direct forecasting of transit ridership at the station-level. Applied Geography, 34, 548-558. Celis-Morales, C. A., Lyall, D. M., Welsh, P., Anderson, J., Steell, L., Guo, Y., . . . Sattar, N. (2017). Association between active commuting and incident cardiovascular disease, cancer, and mortality: prospective cohort study. bmj, 357, j1456. Cervero, R. (1996). Mixed land-uses and commuting: Evidence from the American Housing Survey. Transportation Research Part A, 30(5), 361-377. Cervero, R., & Duncan, M. (2003). Walking, Bicycling, and Urban Landscapes: Evidence From the San Francisco Bay Area. American Journal of Public Health, 93(9), 1478-1483. Cervero, R., & Kockelman, K. (1997). Travel demand and the 3Ds: density, diversity, and design. Transportation Research Part D, 2(3), 199-219. Cervero, R., Sarmiento, O. L., Jacoby, E., Gomez, L. F., Neiman, A., Social, F., & a Colombia, B. (2009). Influences of Built Environments on Walking and Cycling: Lessons from Bogotá. International Journal of Sustainable Transportation, 3(4), 203-226. Chi, G., & Zhu, J. (2008). Spatial regression models for demographic analysis. Population Research and Policy Review, 27(1), 17-42. Cromley, E. K., S.L., McLafferty. (2012). GIS and Public Health (2nd ed.). New York: The Guilford Press. Delmelle, E. M., E.C. Delmelle. (2012). Exploring spatio-temporal commuting patterns in a university environment. Transport Policy, 21, 1-9. Dill, J., & Voros, K. (2007). Factors affecting bicycling demand: initial survey findings from the Portland, Oregon, region. Transportation Research Record: Journal of the Transportation Research Board, 2031(-1), 9-17. Dormann, C. F., J.M McPherson, M.B. Araujo, R. Bivand,J. Bolliger,G. Carl, G Davies, A. Hirzel, J. Walter, D. Kissling, I. Kuhn, R. Ohlemuller, P. Peres-Neto, B. Reineking, B. Shroder, F. M. Schurr, R. Wilson. (2007). Methods to account for spatial autocorrelation in the analysis of species distributional data: a review. Ecography, 30(5), 609-628.
Ewing, R., & Cervero, R. (2010). Travel and the built environment: a meta-analysis. Journal of the American Planning Association, 76(3), 265-294. Ferrell, C. E., Mathur, S., & Mendoza, E. (2008). Neighborhood Crime and Travel Behavior: an Investigation of the Influence of Neighborhood Crime Rates on Mode Choice (FHWA A 200633). Retrieved from San Jose State University: http://works.bepress.com/shishirmathur/12/ Fotheringham, A. S., C. Brunsdon, M. Charlton. (2000). Quantitative geography: Perspectives on spatial data analysis (1st ed.). London: Sage. Fotheringham, A. S., M.E. Charlton, and C. Brunsdon. (2002a). Geographically weighted regression: The analysis of spatially varying relationships. London: Wiley. Fotheringham, A. S., R. Crespo, J. Yao. (2015). Geographical and Temporal Weighted Regression. Geographical Analysis, 1-22. Fotheringham, S., C. Brundson, M. Charlton. (2002b). Geographically weighted regression: the analysis of spatially varying relationships. New York: Wiley. Frank, L., Engelke, P. O., & Schmid, T. L. (2003). Health and community design: The impact of the built environment on physical activity. Washington, DC: Island Press. Frank, L. D., & Pivo, G. (1995). Impacts of mixed use and density on utilization of three modes of travel: single-occupant vehicle, transit, and walking. Transportation Research Record, 1466, 44-52. Grabar, H. (2016). Can America's worst transit system be saved? Slate. Handy, S. (2005). Does the built environment influence physical activity?: Examining the evidence. Washington D.C. : Transportation Research Board. Heinen, E., B.V. Wee, K. Maat. (2010). Commuting by bicycle: An overview of the literature. Transport Reviews, 30(1), 59-96. Jacoby, B. (1995). Adapting the institution to meet the needs of commuter students. In D. M. Johnson, Bell, D.A. (Ed.), Metropolitan Universities: An emerging model in American higher education (1st ed., pp. 51-62). North Texas: University of North Texas. Kerr, A., Lennon, A., & Watson, B. (2010). The call of the road: factors predicting students' car travelling intentions and behaviour. Transportation, 37(1), 1-13. Klockner, C. A., & Friedrichsmeier, T. (2011). A multi-level approach to travel mode choice-how person characteristics and situation specific aspects determine car use in a student sample. Transportation Research Part F: Traffic Psychology and Behaviour, 14(4), 261-277. Klöckner, C. A., & Friedrichsmeier, T. (2011). A multi-level approach to travel mode choice–How person characteristics and situation specific aspects determine car use in a student sample. Transportation Research Part F: Traffic Psychology and Behaviour, 14(4), 261-277. Lamiquiz, P. J., J. Lopez-Dominguez. (2015). Effects of built environment on walking at the neighbourhood scale. A new role for street networks by modelling their configurational accessibility. Transportation Research Part A, 74, 148-163. Landis, B. W., Vattikuti, V. R., & Brannick, M. T. (1997). Real-Time Human Perceptions: Toward a Bicycle Level of Service. Transportation Research Record, 1578(1997), 119-126. Larsen, J., A. El-Geneidy, and F. Yasmin. (2010). Beyond the quarter mile: Re-examining travel distances by active transportation. Canadian Journal of Urban Research: Canadian Plannign and Policy (supplement), 19(1), 70-88. Legendre, P., & Legendre, L. F. (2012). Numerical ecology (Vol. 24): Elsevier. Li, Z., W. Wang, P. Liu, J.M. Bigham, D.R. Raglund. (2013). Using geographically weighted poisson regression for county-level crash modeling in California. Safety Science, 58, 89-97. Lundberg, B., & Weber, J. (2014). Non-motorized transport and university populations: an analysis of connectivity and network perceptions. Journal of Transport Geography, 39, 165-178. Marshall, W. E., D.P. Piatkowski, N.W., Garrick. (2014). Community design, street networks, and public health. Journal of Transport and Health, 1(2014), 326-340. Mason, R. L., R.F. Gunst, J.L. Hess. (1989). Statistical design and analysis of experiments: Applications to engineering and science. New York: Wiley.
Matthews, S. A., & Yang, T.-C. (2012). Mapping the results of local statistics: Using geographically weighted regression. Demographic Research, 26, 151-166. Maxwell, S. (Ed.) (2001). Negotiating car use in everyday life. Oxford: Berg. Mennis, J. (2006). Mapping the results of geographically weighted regression. The Cartographic Journal, 43(2), 171-179. University of Michigan-Flint 2011-2012 Overview, (2012). Molina-Garcia, J., Castillo, I., & Sallis, J. F. (2010). Psychosocial and environmental correlates of active commuting for university students. Preventive medicine, 51(2), 136-138. Moniruzzaman, M., & Farber, S. (2017). What drives sustainable student travel? Mode choice determinants in the Greater Toronto Area. International Journal of Sustainable Transportation(just-accepted). Mountain, D. M., Tsui, J. L. Y., & Raper, J. F. (2007). Modelling accessibility via transportation networks based upon previous experience: a Geographically Weighted Regression approach. Geocomputation 2007. O'Sullivan, D., & Unwin, D. J. (2010). Geographic Information Analysis (2nd ed.). Hoboken: John Wiley & Sons. O Sullivan, D., D.J. Unwin. (2010). Geographic Information Analysis (2nd ed.). Hoboken: John Wiley & Sons Inc. Olson, J. (1975). Spectrally encoded two-variable maps. Annals of the Association of American Geographers, 71, 259-276. Rajamani, J., Bhat, C. R., Handy, S., Knaap, G., & Song, Y. (2003). Assessing impact of urban form measures on nonwork trip mode choice after controlling for demographic and level-of-service effects. Transportation Research Record: Journal of the Transportation Research Board, 1831(1), 158-165. Ramsey, K., Bell, A. (2014). Smart Location Database, Version 2.0 User Guide. Rangel, T. F. L. V. B., Diniz-Filho, J.A.F and Bini, L.M. (2010). SAM: a comprehensive application for Spatial Analysis in Macroecology. Ecography, 33, 46-50. Rodríguez, D. A., & Joo, J. (2004). The relationship between non-motorized mode choice and the local physical environment. Transportation Research Part D: Transport and Environment, 9(2), 151173. Roe Clark, M. (2006). Succeeding in the city: challenges and best practicies on urban commuter campuses. About Campus, 1-8. Rybarczyk, G. (2014). Simulating bicycle wayfinding mechanisms in an urban environment. Urban, Planning and Transport Research, 2(1), 89-104. Rybarczyk, G., & Gallagher, L. (2014). Measuring the potential for bicycling and walking at a metropolitan commuter university. Journal of Transport Geography, 39, 1-10. Saelens, B., Sallis, J. F., & Frank, L. D. (2003). Environmental Correlates of Walking and Cycling: Findings From the Transportation, Urban Design, and Planning Literatures. Annals of Behavioral Medicine, 25(2), 80-91. Sallis, J. F., Cervero, R. B., Ascher, W., Henderson, K. A., Kraft, M. K., & Kerr, J. (2006). An ecological approach to creating active living communities. Annu. Rev. Public Health, 27, 297-322. Sardianou, E., E. Nioza. (2015). Who are the eco-bicyclists? Transportation Research Part D, 34, 161167. Schwanen, T., & Mokhtarian, P. L. (2005). What affects commute mode choice: neighborhood physical structure or preferences toward neighborhoods? Journal of Transport Geography, 13(1), 83-99. Shaker, R., Ehlinger, T. (2014). Exploring Non-linear Relationships between Landscape and Aquatic Ecological Condition in Southern Wisconsin: A GWR and ANN Approach. International Journal of Applied Geospatial Research, 5(4). doi:10.4018/ijagr.2014100101 Shannon, T., Giles-Corti, B., Pikora, T., Bulsara, M., Shilton, T., & Bull, F. (2006). Active commuting in a university setting: Assessing commuting habits and potential for modal change. Transport Policy, 13(3), 240-253.
Shoff, C., T. Yang, S. Mathews. (2012). what has geography got to do with it? Using GWR to explor place-specific associations with prenatal care utilization. GeoJournal, 77, 331-341. doi:10.1007/s10708-010-9405-3 Sisson, S. B., & Tudor-Locke, C. (2008). Comparison of cyclists' and motorists' utilitarian physical activity at an urban university. Preventive medicine, 46(1), 77-79. Sisson, S. B., Tudor-Locke, C. (2008). Comparison of cyclists' and motorists' utilitarian physical activity at an urban university. Preventative Medicine, 46, 77-79. Stokols, D. (1992). Establishing and maintaining healthy environments: toward a social ecology of health promotion. Am. Psychol, 47, 6-22. Toor, W., & Havlick, S. W. (2004). Transportation & sustainable campus communities: Issues, examples, solutions: Island Pr. Whalen, K. E., Paez, A., & Carrasco, J. A. (2013). Mode choice of university students commuting to schooland the role of active travel. Journal of Transport Geography, 31, 132-142. Winters, M., Friesen, M. C., Koehoorn, M., & Teschke, K. (2007). Utilitarian Bicycling A Multilevel Analysis of Climate and Personal Influences. American journal of preventive medicine, 32(1), 5258. Yu, D. (2006). Spatially varying development mechanisms in greater Beijing area: a geographically weighted regression investigation. Annals of regional science, 40(1), 173-190. Zhang, C., Tang, Y., Xu, X., & Kiely, G. (2011). Towards spatial geochemical modelling: Use of geographically weighted regression for mapping soil organic carbon contents in Ireland. Applied Geochemistry, 26(7), 1239-1248. Zhang, L., Gove, J. H., & Heath, L. S. (2005). Spatial residual analysis of six modeling techniques. Ecological Modelling, 186(2), 154-177. Zhao, F., N. Park. (2004). Using geographically weighted regression models to estimate annual average daily traffic. Transportation Research Record: Journal of the Transportation Research Board(1879), 99-107. Zhao, P. (2013). The impact of the built environment on individual workers’ commuting behavior in Beijing. International Journal of Sustainable Transportation, 7(5), 389-415. Zheng, L., Robinson, R. M., Khattak, A., & Wang, X. (2011). All Accidents are Not Equal: Using Geographically Weighted Regressions Models to Assess and Forecast Accident Impacts. Paper presented at the 3rd International Conference on Road Safety and Simulation.
