Associating Ridesourcing with Road Safety Outcomes: Insights from Austin Texas

Improving road safety and setting targets for reducing traffic-related crashes and deaths are highlighted as part of the United Nation's sustainable development goals and vision zero efforts around the globe. The advent of transportation network companies, such as ridesourcing, expands mobility options in cities and may impact road safety outcomes. In this study, we analyze the effects of ridesourcing use on road crashes, injuries, fatalities, and driving while intoxicated (DWI) offenses in Travis County Texas. Our approach leverages real-time ridesourcing volume to explain variation in road safety outcomes. Spatial panel data models with fixed effects are deployed to examine whether the use of ridesourcing is significantly associated with road crashes and other safety metrics. Our results suggest that for a 10% increase in ridesourcing trips, we expect a 0.12% decrease in road crashes (p<0.05), a 0.25% decrease in road injuries (p<0.001), and a 0.36% decrease in DWI offenses (p<0.0001) in Travis County. Ridesourcing use is not associated with road fatalities at a 0.05 significance level. This study augments existing work because it moves beyond binary indicators of ridesourcing presence or absence and analyzes patterns within an urbanized area rather than metropolitan-level variation. Contributions include developing a data-rich approach for assessing the impacts of ridesourcing use on our transportation system's safety, which may serve as a template for future analyses of other US cities. Our findings provide feedback to policymakers by clarifying associations between ridesourcing use and traffic safety, while helping identify sets of actions to achieve safer and more efficient shared mobility systems.


INTRODUCTION
Ridesourcing is a term used to describe the operation of transportation network companies such as Uber, Lyft, DiDi Chuxing, and RideAustin, where the rider can hail a vehicle from the convenience of their smartphone using a web or smartphone application. Ridesourcing services today are offered in more than a hundred cities in the US and hundreds cities worldwide. As an example, ridesourcing trips have grown exponentially with Uber providing ten billion trips globally in 2018 (Uber Technologies Inc., 2019).
Transportation network companies' service can bridge mobility supply gaps, offering a convenient and competitive mode alternative with pooled-service capabilities (Lavieri et al., 2018), enhancing urban transportation options (Hall et al., 2018), potentially reducing private vehicle ownership (Ward et al., 2019), and even providing emergency services by replacing expensive ambulances (Moskatel and Slusky, 2019). However, empirical studies uncover negative effects of ridesourcing travel, including associations with congestion (Erhardt et al., 2019), competition with public transportation modes (Hall et al., 2018), increased net energy use (Wenzel et al., 2019) and environmental externalities (Fry et al., 2019).
Ridesourcing operations could lead to road safety benefits by reducing road crashes, driving while intoxicated (DWI), and the welfare losses associated with those (Greenwood and Wattal, 2017).
Road safety improvements are a focus of several initiatives around the world (e.g., the United Nations sustainable development goals) and in the United States through vision zero efforts (United Nations Department of Economic and Social Affairs, 2015). Ridesourcing use in a city could be associated with a reduction of DWI offenses in locations where other transport modes are not provided or are less attractive than driving one's own vehicle under the influence of alcohol. On the other hand, ridesourcing use could be positively associated with crashes, based on existing empirical evaluations of the impact of ridesourcing on road safety at the metropolitan level. Some authors found that ridesourcing is associated with an overall increase in vehicle miles traveled, due to drivers cruising between offering rides (Erhardt et al., 2019;Henao and Marshall, 2018;Wenzel et al., 2019), which not only results in congested city centers but is also related to increased road crash rates (Barrios et al., 2019).
In this study we investigate how the use of ridesourcing influences the rate of vehicular crashes, injuries, fatalities, and DWI offenses and we capture safety-related externalities of ridesourcing use.
To our knowledge, this is the first empirical study that uncovers such associations, while accounting for the intensity of ridesourcing use by leveraging real-world data from this transportation service.
Previous evaluations of ridesourcing impacts on road safety outcomes focused on metropolitan-level effects and dichotomous indicators of ridesourcing presence or absence, e.g., Barrios et al., 2019. Instead, we utilize more granular data on ridesourcing trips per census tract for Travis County, Texas.
Spatial panel data models are employed to demonstrate associations between ridesourcing trips and road safety outcomes.
In the following section we review existing literature analyzing empirical associations between ridesourcing travel and road safety. The Methods section documents the methodology followed and the Data section describes the data used in this study. The Results & Discussion section presents the main modeling specification parameter outcomes and results of robustness checks. The final section concludes and proposes future research questions.

LITERATURE REVIEW
Existing studies focus on uncovering associations between ridesourcing entry and road safety outcomes, DWI offenses, and alcohol consumption. As shown in Table 1, almost all empirical studies examine the impact of ridesourcing services entry on the relative change in road crash rates, e.g., Brazil and Kirk, 2016. Out of seven economics and epidemiology studies in this field, only one attempts to capture ridesourcing intensity (Barrios et al., 2019). Specifically, Barrios et al. use Google search count data as a ridesourcing exposure proxy, due to the absence of real-world information on this type of trips. Across studies, safety outcomes examined vary between crash, injury and fatality rates, as well as DWI offense rates. A subset of studies focuses on modeling alcohol-related fatalities and injuries (Greenwood and Wattal, 2017;Martin-Buck, 2016;Morrison et al., 2018). Week (2010)(2011)(2012)(2013)(2014) Note: DiD stands for difference-in-differences, OLS for ordinary least squares, QMLE for Poisson quasimaximum likelihood estimator, and ARIMA for autoregressive integrated moving average. network companies entry is not significantly associated with any of the categories of traffic fatalities examined in their work for the US (Brazil and Kirk, 2016); a similar conclusion is reached by Huang et al. (Huang et al., 2019) in their South African-focused study. Morrison et al. hypothesize that the resumption of ridesourcing operation in specific US cities is associated with a decrease in alcoholinvolved road crashes (Morrison et al., 2018); their hypothesis is partially supported for the cities of Portland Oregon and San Antonio Texas. However, the studies above do not account for granular ridesourcing travel, much needed to better capture such dampening effects. Other studies also report findings that are in general agreement with the aforementioned outcomes: Dills and Mulholland indicate that Uber's launch is associated with a reduction in fatal traffic crashes and DWI offenses after certain number of months of operation, uncovering lags (Dills and Mulholland, 2018 increase with ridesourcing use and that these trends persist over time (Barrios et al., 2019). They also demonstrate that the road crash rate increase is significant in cities with greater population levels, higher income quartiles, greater vehicle ownership share and public transportation use, as well as higher carpooling use. This is actually the only work that captures such exposure, using a proxy for adoption intensity, by adopting Google trends search counts. Existing literature fails to assess how the intensity of ridesourcing services in geographic areas smaller than a metropolitan area is related to road safety outcomes.
Ridehailing operation resembles quite a lot that of a taxi service. Evidence from safety research of the operation of taxis examine the impact of fatigue (Dalziel and Job, 1997), driver behavior and working conditions (Wang et al., 2019), as well as taxi driver offenses (La et al., 2013) on taxi crashes.
Associations of these factors with taxi crashes and fatalities are examined through Poisson and logistic regression models without taking into account spatial and temporal characteristics. After systematically searching for evidence, we have not observed studies that examine associations of taxi use and road safety outcomes the way that ridesourcing use relationships have been studied with safety outcomes.
Our analysis aims to bridge literature gaps by shedding light on the relationship between transportation network companies' operation and road safety outcomes. In our analysis, these associations are captured by using real-world ridesourcing trip data. It is crucial to uncover these effects, since ridesourcing could be considered as a road safety solution by city planners, engineers, and policymakers and can be utilized to curb traffic crashes and injuries under a vision zero's plan umbrella. Prioritizing successful interventions is key to achieving substantial road safety improvements (Rakauskas et al., 2009).

METHODS
To measure the association between ridesourcing travel and road safety, we use spatial panel data models. Such models have been previously applied in transportation safety research (Abdel-Aty and Wang and Kockelman, 2007;Xie et al., 2014). The natural logarithm of crashes, injuries, fatalities, and DWI offenses, as well as ridesourcing exposure addresses the variables' right skewness via normalization (Washington et al., 2003). We leverage data on safety outcomes, ridesourcing and traffic volumes, and socio-demographics from January 2012 to the beginning of April 2017. To conduct the proposed analysis, we aggregate variables by census tract and month-year units. The data used here (described in greater detail in the next section) are longitudinal, containing repeated observations of the same census tract units over time. Cross-sectional data and models suffer from an inability to capture intertemporal dependence of events, which panel data models used here address.
The proposed models with spatial and time fixed effects enable eliminating bias from unobserved factors that are changing over time but are constant over each spatial unit and controls for unobserved factors that change over space but are constant over time (Washington et al., 2003).
Previous studies attempt to answer a similar research question with empirical data through difference-in-differences estimation, which compares how the trajectory of road safety outcomes differs before and after the launch of ridesourcing. Due to absence of a control group in our study, since RideAustin ridesourcing operation was launched in Travis County in all census tracts simultaneously (June 2016), we are not able to deploy a DiD model. Instead, we employ spatial panel fixed effects lag and error models as well as Spatial AutoRegressive with additional AutoRegressive error structure (SARAR) models that allow for the disturbances to be generated by a spatial autoregressive process (Anselin and Florax, 1995). The index of each spatial unit is ∈ {1, … , I} and each time unit is ∈ {1, … , T}. A fixed effects spatial lag specification is presented in equation [1], according to Anselin and Hudak, 1992: -. = 1 -3 3. -43 where -. is each road safety outcome (e.g., the number of total crashes that occurred in the spatial unit during month , and similar for injuries, fatalities, and DWI offenses), -3 is a spatial weighted matrix in which neighborhood relationships are defined between the spatial units of analysis (constant over time , with diagonal elements equal to zero), ∑ -3 3. -43 is the spatially lagged dependent variable (denoting that the value of at time is explained not only by the values of exogenous independent variables but also those neighboring the spatial unit ), is the scalar spatially autoregressive coefficient of the spatially lagged dependent variable, -is the spatial unit fixed effect, . is the time unit fixed effect, the vector of parameters to be estimated, -. a vector of explanatory variables, and the error terms -.~( 0, A ). The census tract fixed effects control for all time invariant census tract specific factors that are potentially correlated with safety outcomes, such as area.
Similar assumptions hold for the time fixed effects that control for census tract invariant factors that vary by month like travel patterns. For all models, -3 weights are defined based on binary contiguity, where -3 = 1 when the intersection of the boundaries of and spatial units is not empty, otherwise The spatial error with fixed effects specification is presented in equations [2] and [3], according to the specification by Baltagi et al., 2003: [3] where the disturbance term -. follows the first order spatial autoregressive process of the from presented in [3], and is spatial autoregressive coefficient where | | < 1. The rest of the terms have been specified in the previous paragraph. Comparing the spatial lag and spatial error models, the former suggests a diffusion, where road safety crashes in one spatial unit predict an increased likelihood of road crashes in neighboring places; the latter suggests that we have omitted spatially correlated covariates that would affect inference.
Last, the SARAR model is defined as presented in equations [4] and [5] (Anselin and Florax, 1995;Debarsy and Ertur, 2010): -43 [5] where both and are spatially autoregressive coefficients. SARAR accounts for both neighboring effects and omitted spatially correlated covariates. Maximization of the likelihood function results in the estimation of the unknown coefficients , in accordance with Anselin and Hudak 1992 notation for the spatial lag and error models and Millo and Piras 2012 for the SARAR model.
We conduct specification tests to confirm that fixed effects are most appropriate over random effects. The Hausman test (Hausman, 1978) is applied that denotes that a fixed effects model is at least as consistent as the random-effects specification, and thus, fixed effects is chosen. Locally robust Variables of interest include median household income, population density, employment density, and percent of zero vehicle ownership (U.S. Census Bureau, 2018).

Road safety outcomes rates
The analysis period is from January 2012 to April 2017. We showcase in Figure 1 the time series of monthly crash, injury, fatality, and DWI offense rates (per thousand people) averaged by census tract in Travis County. Note that we use the KABCO scale for injury severity; the injuries used to compute the corresponding rate include incapacitating (A), non-incapacitating (B), and possible injuries indicated by behavior but not visible wounds (C) (Federal Highway Administration, 2018). We observe similarities between the crash and injury rates seasonal trends, even though for the former the fitted linear trend over the total period of time is increasing and for the latter decreasing.
In Figure 1,   The spatial distributions of the average monthly crash rate and RideAustin passengers' pick up and drop off rate in Travis County are portrayed in Figure 3. We present these before and after the entry of ridesourcing in the region (without accounting for the excluded period denoted in Figure 1). Note that in June 2016, RideAustin started offering rides mainly downtown and expanded to serving all tracts in the Travis County over the months.

FIGURE 3 Road crash rate and ridesourcing rate before and after the introduction of the ridesourcing service in Travis County Texas census tracts.
With highway networks spanning the suburban Austin region, we observe an increase of average crashes per 1000 people in the aftermath of RideAustin launch there. Figure 3 also suggests that after RideAustin's entry, the rate of total crashes increased in the suburbs of Travis County but not in the downtown regions that the ridesourcing operation covered at a greater rate. This could be associated with slower population and firm growth in the suburbs and could reflect changes in travel patterns that resulted from spatial differences in the region's economic development.

Other controls
Various measures are introduced as control variables in our model. Their selection is based on prior evidence (as show in Table 1  In short, the hypotheses that support the inclusion of the aforementioned control variables are as follows. Population density, employment percentage, and traffic (in the form of OD trip index provided through StreetLight Data's platform) are expected to capture the market's size. The median household income is controlling the market's wealth as in (Greenwood and Wattal, 2017), and the percentage of zero vehicle ownership is expected to capture likely users who do not own a personal automobile and may rely on other modes of transport, as in (Hall et al., 2018).    The spatial fixed effects panel data model estimates, with all controls included, are presented in Table 4. Lagrange Multiplier tests results are noted there. Signs and magnitudes of the coefficients are consistent across the three models, supporting robustness of our findings.

RESULTS & DISCUSSION
Ridesourcing use is found significantly associated with three road safety outcomes: crashes, injuries, and DWI offenses. For a 10% increase in ridesourcing use, we expect a 0.12% decrease in road crashes (p<0.05), a 0.25% decrease in road injuries (p<0.001), and a 0.36% decrease in DWI offenses (p<0.0001). Ridesourcing use is not found associated with fatalities at the 0.1 significance level. These results are well aligned with Dills and Mulholland findings as well as Morrison et al. that associate the entry of ridesourcing with a decrease in DWI offenses and alcohol-involved crashes, respectively (Dills and Mulholland, 2018;Morrison et al., 2018). However, Morrison et al. find no significant association between ridesourcing entry and road injuries. Our findings are also aligned with two other studies which found no significant association between ridesourcing entry and fatalities (Brazil and Kirk, 2016), while we also account for the ridesourcing induced demand effects that the aforementioned studies do not capture.
The magnitude of percentage decrease of road safety externalities associated with ridesourcing use can be considered small compared to the effectiveness of other strategic road safety interventions such as seatbelt laws enforcement (up to 9%) (Carpenter and Stehr, 2008), speed limit reduction and traffic calming measures (10-15%) (Elvik, 2001). Therefore, promoting the use of ridesourcing as vision zero policy might not be as effective as infrastructure improvements, speed limit changes, and other educational or policing vision zero initiatives. However, a ridesourcing initiative to reduce DWI offenses and increase road safety could benefit groups most likely to ride those: younger drivers (Dills and Mulholland, 2018). Road crashes and fatalities are significantly associated with population density, as expected. Crashes increase with increased population density, but road fatalities decrease. Population density as fatalities' predictor might serve as proxy for speed's effect, given that roads in lower density environments tend to be characterized by higher speed limits. The percentage of vehicle owners is positively associated with road crashes and injuries in Travis County. This result is aligned with other literature findings (Barrios et al., 2019); lower levels of vehicle ownership hint greater public transit and/or active transportation usage. We note though that ridesourcing use could endogenously influence vehicle ownership positively, by incentivizing more to drive for transportation network companies. In our analysis, ridesourcing use is positively (but weakly) correlated ( = 0.12) with the percentage of zero vehicle ownership.
The number of DWI offenses would decrease when a greater number of alternative transportation options are available, as expected. This finding also suggests that ridesourcing in Travis county may not only serve as a substitute for taxis and other modes of transportation, but also for drunk driving. However, this outcome could also be a result of enforcement, educational and other efforts, which we are unable to capture here.

Robustness Checks
The concurrent launch of RideAustin and the exit of Uber and Lyft from the Austin region (Solomon, 2017)  Even though the significance of the relationship between ridesourcing use and road safety outcomes is weaker in this case, we observe that the sign of the parameter remains negative. In accordance with the base RideAustin operating period definition, we observe that ridesourcing use is associated with a reduction in DWI offenses and traffic injuries but does not seem to be related to road crashes and fatalities.
The assumption that once Uber and Lyft services depart from Austin, the crashes and DWI offenses decline is based on Barrios et al. (2019) findings. In such a case, we would expect that some of the observed decline in road crashes to be simply due to reductions in the number of cars on the road. Therefore, the treatment driving the results might be the departure of Uber and Lyft, not the arrival of RideAustin. We acknowledge that the current definition of the after RideAustin launch period might result in misinterpreting the modeling outcome.

CONCLUSIONS
We use RideAustin OD trips to examine the effect of ridesourcing exposure on road safety outcomes, such as road crashes, injuries, fatalities, and DWI offenses. Spatial fixed effects panel data models are employed to establish that RideAustin use is significantly associated with a decrease in total road crashes, injuries, and DWI offenses in Travis County Texas. On the contrary, our findings do not demonstrate significant relationships between ridesourcing use and road fatalities. Given the significant costs associated with road safety outcomes and DWI offenses (Connelly and Supangan, 2006; National Highway Traffic Safety Administration, 2015), ridesourcing services can be a low-cost option that could assist cities and counties with meeting goals for road injuries and drunk-driving reduction. At the same time, the magnitude of the road safety externalities reduction associated with ridesourcing trips is smaller compared to the effectiveness that has been documented after the application of other interventions including seatbelt laws, reduced speeds, and traffic calming design.
This outlines the need for determining population segments that ridesourcing-related solutions or policies could be more impactful for improving road safety (e.g., younger ridesourcing demographics focus).
Our analysis augments existing work in this field by accounting for spatial distributions of ridesourcing use, road safety outcomes, and other socio-economic characteristics in the given region.
Instead of testing associations of the launch of ridesourcing with road injuries and the rest of safety outcomes, we account for spatio-temporal characteristics and capture actual ridesourcing use via realtime trip data analytics in Travis County. The spatial panel data modeling outcomes show that spatial dependence is of significance, thus, granular longitudinal travel, road safety, and socio-demographic data can provide transportation and traffic safety agencies that opportunity to uncover associations and plan for appropriate safety interventions.
Additional research efforts should be put towards addressing this study's limitations, including: 1) testing the effect of alternate travel demand exposure methods, such as vehicle miles traveled instead of the OD trips index, 2) exploring whether effects might not manifest immediately from ridesourcing use, accounting for lags (Greenwood and Wattal, 2017), and 3) examine results robustness by performing similar analysis in additional regions in the US and around the world, since the generalizability of the results can be questioned. Future research should also uncover for which populations and subpopulations road safety outcomes can be improved through ridesourcing use by: (a) exploring the relationship of ridesourcing and road safety outcomes for different household income and employment percentage panels and (b) identify key drivers of where potential public health benefits of ridesourcing utilization can be the greatest.

APPENDIX A1. Correlations
We present the correlation matrix results of the four safety outcome variables and the ridesourcing use and control variables to showcase associations between those in Figure A1.  121.85 *** Note: Symbol *** corresponds to p<0.0001, ** to p< 0.001, * to p< 0.01, and . to p<0.05.