Below is the uncorrected machine-read text of this chapter, intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text of each book. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.
Incorporating Safety into Long-Range Transportation-Planning Appendix D: Developing a Planning Level Safety Forecasting Model (PLANSAFE) APPENDIX D DEVELOPING A PLANNING LEVEL FORECASTING MODEL (PLANSAFE) Appendix C described the application of a PLANSAFE model for forecasting crashes at the planning level. The focus in Appendix C was on forecasting crashes (total, fatal, pedestrian, etc.) in future periods or for build scenarios for use in planning applications. Primary uses include the setting of safety performance targets and for feedback on development and/or growth scenarios. This Appendix, in contrast, provides the details necessary to develop (as opposed to apply) a planning level forecasting model. This appendix is intended to serve as a resource for an agency that has both the desire and ability to develop their own set of regression models for forecasting safety at the planning level. The motivation for such an undertaking would be the desire to increase the confidence in the relationships captured in the models using local or regional data instead of data from other regions (Pima County, Maricopa County, and Michigan State). This section is organized as follows. First, the limitations of planning level safety forecasting models are described. The data requirements for such a model are then discussed, followed by software requirements and required expertise. Development of the datasets is followed by a discussion of the development of the statistical models. Detailed development of the planning level safety predictions models is then provided. Finally the methodology for GIS processing required to develop the datasets are discussed. LIMITATIONS OF PLANNING LEVEL SAFETY FORECASTING MODEL A safety model at the planning level is fundamentally different than corridor and site level safety models with which most safety professionals are familiar. The differences need illumination so that model misuses are avoided. Following are the limitations of these models. ⢠the model can only be used at a TAZ area level: it can not be used for corridor or project-level-related assessments and analysis, ⢠the model is not suitable for bolstering arguments for or against particular safety, land use, or transportation investments. In other words these models are predictive in nature and intend to inform the analyst as to when certain outcomes will occur; however, it they are not explanative models that describe why certain outcomes occur. ⢠a geo-coded road network and linked accident and other transportation data (refer to the section discussing data requirements) are required to develop the model, ⢠the creation of the data sets necessary to develop the model requires the transformation of census block group data to TAZ area which requires GIS expertise, ⢠the modelling requires the careful identification of independent variables and the selection of these variables requires considerable statistical modelling expertise, and ⢠special expertise is required to prepare the dataset and to develop the model (refer to Exhibit 87). The model uses the linear regression model with logarithmic transformation of the dependent variable. This distribution is sensitive to any correlation between variables in the model and the selection of independent variables is therefore essential for the successful development of this model. The professional can use a 147
Incorporating Safety into Long-Range Transportation-Planning Appendix D: Developing a Planning Level Safety Forecasting Model (PLANSAFE) correlation matrix to assist with in the selection of independent variables during the model development process. DATA REQUIREMENTS OF PLANNING LEVEL SAFETY FORECASTING MODEL Both the development and use of the prediction model requires data by traffic analysis zone (TAZ). TAZs are the smallest analysis unit. Larger units can be analyzed by aggregating TAZs. For example, a change to a commute corridor that impacts numerous TAZs can be modeled by considering the impacts of the project on all affected TAZs. The models require data sets referring to geographical areas such as census block groups and transportation facility datasets in geospatial information systems. Geographical information systems (GIS) are used extensively to develop the data sets in support of these models. GIS layers in the development of the prediction model include: ⢠The TAZ areas that makes up the area for the prediction model, as defined by the transportation agencies of the area, ⢠Tracts and/or block groups as defined by the U.S. Census (the use of block groups is recommended) with the associated demographics, socio-economics and other data, ⢠The entire road network of the area: i.e., including facilities managed by the state, counties, regional agencies, and local agencies, ⢠The federal functional classification of the entire road network of the area, ⢠The vehicle miles traveled on the road network on the area (can be calculated by generating known section lengths and multiplying it with known section traffic volumes), ⢠Bike facilities and routes, ⢠Transit facilities, ⢠Unique accident record identification numbers for accidents for a minimum of one year and ideally three years, and ⢠Locations of institutions such as schools and police stations. The details for the development of these datasets are described later in this section. SOFTWARE REQUIREMENTS The analyst develops the model by using GIS software and statistical analysis software, such as LIMDEP, SPLU.S., GENSTAT, SPSS, SAS, aML, etc. The researchers at the University of Arizona used ArcGIS, and LIMDEP for the development of models described in this section and in Appendix C. REQUIRED EXPERTISE The estimation of planning level safety forecasting models requires the following expertise. Development of datasets. GIS software-related expertise is required for the preparation of data needed in the development of the model. The individual will have to perform various types of GIS processing to assign data to the TAZ areas and have a fair knowledge of vector and raster modeling and spatial analysis in the GIS environment. 148
Incorporating Safety into Long-Range Transportation-Planning Appendix D: Developing a Planning Level Safety Forecasting Model (PLANSAFE) Development of the models. The development of the statistical models, using the dataset created for the model, requires experience in statistical modelling and transportation safety. Knowledge about basic hypothesis testing, regression and the ability to evaluate a model using goodness-of-fit are basic requirements. As the development requires the use of statistical software such as LIMDEP or STATA, the individual also has to be able to use the software and interpret the results provided by the software. The individual should also be knowledgeable in the field of transportation safety as the evaluation of the variables in the generated models requires an understanding of the relationships between the variables and accident- related variables. DETAILED DEVELOPMENT OF A SAFETY PREDICTION MODEL AT THE TAZ LEVEL Exhibit 99 depicts the process that the analyst follows to develop the planning level safety prediction model. The process consists of three basic steps: ⢠data collection, ⢠development of a dataset containing variables used in modelling, and ⢠development/estimation of the statistical models used for forecasting. All of these activities support development of the planning level safety prediction model. Before one begins this process, it is important to recognize that the ultimate model drives all the activities preceding it. So, a review of the safety model and what factors are thought to affect safety at the aggregate level is worthwhile at this point. Safety, as defined by total crashes, severe crashes, injury crashes, pedestrian crashes, and bicycle crashes are influenced by numerous factors. These factors must be viewed in the framework of aggregated data and crashes cannot be examined in isolation. Exhibit 89 lists potential variables that may capture the underlying effects listed in the first column. For example, weather is known to affect crashes, with wet, ice, and snow affecting crashes considerably. At the TAZ level, the proportion of wet pavement days may help to capture the variability in crashes observed within a TAZ. Similarly, high risk driving populations are involved in crashes more frequently than average drivers. Identifying the proportion of high risk drivers residing within a TAZ may help to capture some of this effectâpredominately those crashes that occur close to home (which is a significant proportion). The list of variables listed in the table is meant to provide a basis from which TAZ data collection is conducted. The list is not exhaustive, but captures most of the major factors involved with crashes at the TAZ level. 149
Incorporating Safety into Long-Range Transportation-Planning IDENTIFY INDEPENDENT VARIABLES FOR USE IN THE MODEL DEVELOP THE MODEL Prepare a dataset by TAZ area using GIS technologies such as dynamic segmentation and spatial joining. a) Prepare a correlation matrix for the dataset using software such as Limdep or Stata b) Identify variables that do not correlate with other variables c) Prepare dataset with independent variables that can potentially be used for the model Develop a set of independent variables using road network characteristics, socio-economic and demographics, and crash history Select the model with the best goodness-of-fit a) Collect road network related data: usually includes local, county and state road network & use dynamic segmentation to assign mileage and other attributes to the particular TAZ b) Collect census related data by block group and assign with GIS technologies to the TAZ polygons. Prepare a dataset with potential socio-economic and demographic variables by TAZ. c) Collect crash data for at least one year and develop a dataset with potential crash related variables by TAZ zone. DATA COLLECTION Develop an initial model using a set of independent variables generated in previous step using Lindep or Stata and linear regression with the transformation of the dependent variable. Test the model by: a) Determining the significance of each of the variables in the model b) Determine whether the relationship provided by the model can be logically explained Repeat process and estimate a number of candidate models using variations of variables and by adding, maintaining or dropping variables based on tests required in previous step Exhibit 99: Process followed to develop PLANSAFE by TAZ for planning level safety prediction Appendix D: Developing a Planning Level Safety Forecasting Model (PLANSAFE) 150
Incorporating Safety into Long-Range Transportation-Planning Appendix D: Developing a Planning Level Safety Forecasting Model (PLANSAFE) Major Contributing Factor Potential Aggregate (TAZ level) Variables that may capture effect of Major Factor (assumes time scale is year) Weather Proportion of wet pavement days per year Proportion of icy pavement days per year Proportion of snow days per year Proportion of fog/reduced visibility days per year Proportion of sunny days per year High risk driving populations Population/number of licensed drivers Proportion of population between 16 and 24 Proportion of population over 60 Number of DUI arrests Employed/unemployed workers High risk non motorized populations Number of crosswalks Number of schools (elementary, middle, high, college) Percentage/mileage of sidewalks (of street mileage) Percentage/mileage of bicycle facilities Speed, design standards of facilities, and access control Total street mileage Proportion of local road mileage Proportion of collector road mileage Proportion of arterial road mileage Proportion of rural highway mileage (urban/rural) Proportion of interstate (urban and rural) Conflicts Number/proportion of signalized intersections Number/proportion of stop-controlled intersections Intersection density Total area DATA COLLECTION AND PREPARATION During the data collection and preparation process, the analyst develops datasets that tabulate the particular variable(s) per TAZ area. The major factors and their associated variables (or similar ones) listed in Exhibit 100 serve as motivation for obtaining certain information in the data collection phase. The data collection effort for the TAZ based (planning level) safety prediction model requires cooperation among the different transportation agencies in the region. Data are collected at the different levels of agencies and sharing of data between these agencies can present difficulties, it is therefore recommended that the support of the state DOT, county and metropolitan/regional level be sought at the start of the data collection process. Typically, data will be gathered from the State DOT, the included counties, and, in some cases, metropolitan/regional/local agencies. In some areas, there may also be other agencies to consider and data sources will vary from area to area. Typical data per TAZ area considered for inclusion into the model are: ⢠road network mileage by federal functional classification, ⢠accident data: a variety of variables can be generated varying from degree of injuries sustained in the accidents, number of injuries and fatalities, or accident types, ⢠census data: population, age distribution within a TAZ (e.g., number of individuals age 17 and younger), employment, housing units: vacant and occupied, persons with disabilities, etc., and ⢠traffic volume data: vehicle miles traveled. Exhibit 100: Major contributing factors in crashes at the TAZ level and potential variables 151
Incorporating Safety into Long-Range Transportation-Planning Appendix D: Developing a Planning Level Safety Forecasting Model (PLANSAFE) This section describes the data preparation process, the development of a dataset for modeling, and the creation of a crash prediction model. Data Preparation As listed in Exhibit 100 the model development process uses information related to the census, the road network, and historical accident data. The development of the model requires a data matrix by TAZ area number. During the data preparation process GIS technology is utilized to develop the datasets. Specific issues that arise with respect to GIS are described in the next subsection of this appendix. The ArcGIS environment is used but similar processing can be performed in other GIS environments as the description is intended to provide the sequence for processing operations in command line or graphical user interface environments; or for scripted batch processing. Refer to the section titled Using GIs in the Development of the Planning Level Safety Forecasting Model for a discussion of the GIS processing procedures. This section describes the four different data categories that can be considered for a PLANSAFE model. Road Network Data During the development of the model, the following road network information per TAZ, among others, the analyst can consider the following as potential variables: ⢠total mileage per functional class of all the roads, i.e., all state, county, regional, and local streets, ⢠total number of intersections, ⢠positions of bus stop and transit facilities, ⢠mileage of bike facilities, ⢠portions of signalized and stop controlled intersections, and ⢠population and vehicle-miles-traveled. Vehicle miles traveled by TAZ area is recognized as an important element of the development of accident prediction models and the researchers recommended that the data collection efforts for the Highway Performance Monitoring System (FHWA) can be used for this purpose unless the agency has VMT data available for all the road sections. It is also possible, however, that population serves as a sufficient exposure metric, as it is probably more accurate than VMT in its measurement. Having both may be the best approach for model testing and refinement. VMT may be approximated by multiplying average annual daily traffic (AADT) for a particular road section by the length of the road section. This requires that the analyst ensures that road segments that make up the road network be provided with a unique segment identifier that can be linked to a unique road segment identifier within the HPMS data set. In some cases it may be necessary to obtain the HPMS data on a county level and also on a state level to ensure that such unique route identifiers exist. Careful attention needs to be paid during the assignment of mileage to the different TAZ areas to ensure that arcs representing the road network do not get lost due to complex GIS-related calculations. It would therefore be valuable to calculate the total mileage per functional class for the entire area and then for the different TAZ areas and compare the total mileage per class with the sum of the mileage per class per TAZ values to ensure that all the sections are included in the dataset. 152
Incorporating Safety into Long-Range Transportation-Planning Census Data The U.S. Census data for SF1 and SF3 is used to identify potential variables related to socio-economic, demographic, and employment data. Based on the case studies presented as part of this section, it is recommended that the census data be transformed from a block-group level to the TAZ level. Census data is not reported by a sub-area where the data can be personally identifiable, i.e., variables with low frequencies in an area may be presented as zero values in the data from the census. This causes false zeros in block data. The tract areas, on the other hand, is large compared to the TAZ area and is therefore expected to generalize the data too much when it is transformed to the TAZ area. Census data can either be downloaded from the U.S. Census website through the American Fact Finder web page at http://factfinder.census.gov/ or by creating datasets by using the U.S. Census 2000 Data Engine CDâs that are available per state per SF1 and SF3. NCHRP 8-48 is currently reviewing the use of the new American Community Survey data for transportation-planning and can potentially be a source of data for the development of the prediction model. In some cases transit and other transportation studies generate data that can be used in the development of the model. These data are generally available per census tract and in these cases can be transformed into TAZ level data. The next section presents step-by-step instructions to transform census block- group data or data per tract or other sub area to TAZ areas in ArcGIS (refer to the section titled Using GIS in the Development of the Planning Level Prediction Model). In the GIS environment, the block group data are assumed uniform and the assignment to the TAZ is done using proportion per area of overlap. Institutions The number of relevant institutions per TAZ, such as police stations, schools, colleges, and universities are considered as potential variables for the model. The final section of this appendix provides step-by-step instructions to calculate the frequencies of each of these institutions per TAZ area. Accident History Accident data is geo coded in a number of different ways and the GIS environment is used to generate the outcome variables that are considered during the model development process. The analyst uses a shape file containing the point events, i.e., accidents, by unique accident report number, together with a shape file containing the TAZ boundaries, to generate of a data set that contains the unique accident report number and the TAZ area it is located in (refer to the step-by-step instructions to calculate the frequencies of each of these institutions per TAZ. The data set can then be used to generate a table of frequencies of accidents per TAZ by summarizing the data points per TAZ. Accident-related variables to be investigated as possible accident outcome predictions: accident severity, injuries sustained in the accident, pedestrian involved crashes, fatal crashes, and other accident-related variables. Appendix D: Developing a Planning Level Safety Forecasting Model (PLANSAFE) 153
Incorporating Safety into Long-Range Transportation-Planning Appendix D: Developing a Planning Level Safety Forecasting Model (PLANSAFE) Development of a dataset containing modelling variables The next step in the process of developing a planning level safety prediction model is the development of a data set containing independent variables. It is recommended that a correlation matrix be prepared to assist the statistical specialist in this effort. The correlation matrix is helpful for identifying which variables are capturing essentially the same or similar underlying phenomenon. The use of variables described in previous sections will motivate the development of this variable list. This step requires the use of database management software such as MS Excel, Access, or other database management system. Finally, prior to modelling, all variables should be examined individually to determine whether the variables make sense. Reasonable checks for reasonableness include computing means, medians, modes, maximums, and minimum values of all variables in the database. Often times coding and transcription errors can be detected during this process so as to avoid negative influences on the modelling results. Development of Crash Prediction Model The researchers of NCHRP 8-44 developed a safety prediction model by using the following approach and assumptions: ⢠Accident count distribution. Accident counts are assumed to be well approximated by the negative binomial distribution when observed per unit area or per unit time (e.g., crashes at intersections for one year each). A linear regression model with logarithmic transformation of the count data will produce a satisfactory model when data are aggregated at the TAZ level (i.e., lots of intersections, road segments, etc.) and TAZs are of varying sizes. Mean crash frequencies are thought to vary across TAZs due to unobserved characteristics of the TAZs. ⢠Simultaneity of accident occurrences. Simultaneous model estimation techniques may be used to model the simultaneity of the accident occurrences (see Washington, Karlaftis, and Mannering, 2004, âStatistical and Econometric Methods for Transportation Data Analysisâ, Chapman Hall, for details on simultaneous model estimation techniques). This need arises due to the likely correlation of error terms across crash prediction models. If modeled separately (and not simultaneously) the coefficients will be inefficient. ⢠Variables maintained due to statistical significance and agreement with expectation. Variables are maintained in the models by determining the significance level (95% is accepted as a minimum) and by assessing whether the relationships between the particular variable and accident outcome, including direction of the effect, agrees with theoretical expectations of accident outcomes. ⢠Error terms correlated across models. The error terms in the models are thought to consist of omitted variables and measurement errors. Omitted variables are assumed to affect all accident injury outcomes (e.g., fatal, serious, slight, total injuries) and the original error term in the model is not correlated to the observable variables. ⢠Contemporaneous correlation. During model estimation additional information from contemporaneous correlation is used. The simultaneous equations are solved by using system estimation methods such as the three-stage least squares. ⢠Simultaneous negative binomial equations. An iterative estimation process is followed using a likelihood maximization algorithm until convergence is achieved and parameters are estimated ⢠Measurement of Goodness of Fit. The goodness of fit for the simultaneous model system is assessed using the RP2P statistic, and individual t-statistics for variables. 154
Incorporating Safety into Long-Range Transportation-Planning Appendix D: Developing a Planning Level Safety Forecasting Model (PLANSAFE) Modelling trial and error is used to derive meaningful and useful models. Knowledge of transportation safety is used to derive a model that is consistent and in agreement with current knowledge of motor vehicle crashes and safety. U.S.ING GIS IN THE DEVELOPMENT OF THE PLANNING LEVEL SAFETY FORECASTING MODEL The Planning Level Safety Prediction Model requires the analyst to perform various calculations within the GIS environment. The purpose of this section is to describe a general methodology for the processing of data within the GIS environment. ⢠Creating census data sets per TAZ, i.e., distribution of demographic data in block groups to TAZ areas by assuming uniformity of values in block groups, ⢠Creating accident data sets per TAZ, i.e., assignment of total road mileage to each TAZ ⢠Creating road mileage and VMT summary sets for the road network by TAZ, i.e., association of accident events (points) with the TAZ. The ArcGIS environment is used but similar processing can be performed in other GIS environments as the description is intended to provide the sequence for processing operations in command line or graphical user interface environments; or for scripted batch processing. Conceptual Framework This section places the described methodologies within a conceptual framework for conceptualizing the data processing. The association of the attributes of TAZ by their spatial relationship with the attributes of other spatial themes, such as traffic accidents and census block groups is a fundamental function of GIS. Overlay functions handle the association of the attributes of one feature class with those in another feature class. Once the attributes are feature classes are associated the values of an attribute of one feature class can be summarized by the values of another. For example, the summarization of demographic data by TAZ to produce proportional population counts for each TAZ. Since the transportation data (daily trip counts, etc.) are associated with the zones of the TAZ, it is the proportional demographic data, for example, that will be associated with the TAZ numbers. The proportional population counts can then, be summarized by TAZ number for further statistical processing. One of the important assumptions of this method is the uniform distribution of persons and person characteristics within a census block group. Methods This section discusses the methodologies that could be used to perform the GIS processing needed in the process of creating census, road mileage, and accident data per TAZ. Distribution of demographic data in block groups to TAZ areas Census data sets can be obtained from the U.S. Census or the agency responsible for the area. To enable the analyst to summarize census data per TAZ, the following are needed: ⢠A shape file with the geographic boundaries of the census block groups for the corresponding census data collection year â this file should match the datum, projection coordinate system and units of any other shape files. The boundaries 155
Incorporating Safety into Long-Range Transportation-Planning Appendix D: Developing a Planning Level Safety Forecasting Model (PLANSAFE) are then associated with a database file, either in Microsoft Access or in dbf format, that contains the census data. ⢠A shape file with the geographic boundaries of the TAZs for the area. Exhibit 101 describes the data that are required to perform the GIS processing. Type Name Description Feature class polygon block_groups U.S. Census Block Groups Feature class polygon TAZ Traffic Analysis Zones The GIS processing steps are as follows: 1. Obtain required digital data sets with metadata 2. Verify spatial and attribute domains 3. Normalize spatial data sets to common projection and datum 4. Vertically integrate data sets 5. Calculate density for census block groups 6. Overlay TAZ and census block feature classes 7. Calculate population for unioned polygon feature class 8. Summarize counts by TAZ for output unioned feature class polygon Assignment of total road mileage to each TAZ Some of the variables considered during the development of a planning level safety prediction model and subsequently required during the application of the model, includes the length of roads within a particular TAZ with a particular functional classification or characteristic. To generate such a data set, the analyst needs the following: ⢠A shape file containing the TAZ boundaries ⢠A shape file containing the road network and associated characteristic values for the road sections that makes up the road network. Exhibit 102 describes the data that are required to perform the GIS processing. Type Name Description Feature class polygon TAZ Traffic Analysis Zones Feature Class line Street_network Line theme of road network The GIS processing steps are as follows: 1. Obtain required digital data sets with metadata 2. Verify spatial and attribute domains 3. Normalize spatial data sets to common projection and datum 4. Vertically integrate data sets 5. Overlay street network and TAZ boundaries 6. Summarize counts by output intersected feature class line 7. Associate summary street length values with TAZ polygons Exhibit 101: Data required to distribute demographic data in block groups to TAZ areas Exhibit 102: Data required to assign road mileage to TAZ areas 156
Incorporating Safety into Long-Range Transportation-Planning Appendix D: Developing a Planning Level Safety Forecasting Model (PLANSAFE) Association of accident events (points) with the TAZ In the planning level safety prediction model, the analyst uses the frequency of accidents or severity of accidents or any other related events per TAZ. The analyst therefore has to develop a data set that summarizes the particular data points within each TAZ. Exhibit 103 describes the data that are required to perform the GIS processing. Type Name Description Feature class polygon TAZ Traffic Analysis Zones Accident Location Data Accidents Database Table The GIS processing steps are as follows: 1. Obtain required digital data sets with metadata. 2. Verify spatial and attribute domains. 3. Classify and scrub accident data for subprocessing procedures. ⢠Build route systems ⢠Calibrate route systems ⢠Create event theme for linear reference accidents OR ⢠Verify reference theme for address matching ⢠Create address locator service ⢠Geocode addresses 4. Derive point feature class for georeferenced accident locations. 5. Overlay point feature class accidents on TAZ polygons. 6. Summarize accidents by TAZ number. References Dixon, Michael P., Brent Orton, and Karl Chang. GIS Input Processing Methodologies for Transportation-planning Models. HTUhttp://www.featureanalyst.com/UserConf/papers/Orton/Orton%20GIS%20Pa per.pdfUTH (March 8, 2005). OâNeill, Wende A. and Daniel Baldwin Hess. 1999. Using GIS to Evaluate a New Source of Transportation Census Data: The American Community Survey. Available at HTUhttp://www.fcsm.gov/99papers/oneill.htmlUTH. (March 9, 2005). Exhibit 103: Data required to assign accidents to TAZ areas 157
