In 2013, the Los Angeles Department of Transportation implemented a road diet on Rowena Avenue in the Silver Lake neighborhood in response to the death of a pedestrian. The project reduced the number of travel lanes from four to three and added a new painted bike lane on both sides of the street. The purpose of the project was to make the street safer, and from a safety perspective the evidence is conclusive: Serious crashes dropped in the years following completion of the project.
Despite the improvement to safety, the road diet has generated controversy in the surrounding neighborhood. Local residents complained that the road diet caused an increase in cut-through traffic on nearby residential streets as motorists attempt to avoid the slower-moving Rowena Avenue. The City of Los Angeles and the staff at LADOT engaged DataScience to see if we could use municipal data to answer the following research question: Did the road diet on Rowena Avenue increase cut-through traffic on nearby residential streets?
We previously published our findings in an op-ed in the Los Angeles Times. This post explains, in more detail, our technical methods and collaboration with city officials. We'll also be sharing our methodology at a meetup about the data science of traffic safety on January 31, 2017.
While it might be assumed that removing motor vehicle lanes reduces access for motorists, an analysis of 17 different road diet projects in a diverse group of cities showed a decrease in traffic volume in just two of the projects. Even with fewer lanes, streets still carry just as much traffic as they did prior to the diet — but with a substantial improvement to safety. The City of Los Angeles recently committed to Vision Zero, a worldwide initiative to reduce non-motorized traffic fatalities to zero by 2025. Traffic-calming projects such as road diets are a potential means of achieving this goal.
The Rowena road diet was implemented in response to the death of Ashley Sandau as she walked across the street. The total length of the project is about a half mile, stretching from Hyperion Avenue to the west and Glendale Boulevard to the east.
The cut-through route that has generated the most vocal opposition to the road diet is located along Angus Street to the south of Rowena. LADOT field traffic count surveys have confirmed that commuters are turning right from Griffith Park Bouelvard, and proceeding down Angus to Glendale Boulevard via Lakewood. The route ends at Lakewood and Glendale, completely bypassing the stretch of Rowena where the road diet was implemented.
The Road Diet Improved Safety
As we reported in our op-ed, we observed some substantial improvements in safety in terms of crashes. Our source for this data is SWITRS, a database of California motor vehicle collisions maintained by the California Highway Patrol. In our initial article, we reported crash statistics for selected years based on some analysis we received from LADOT. Due to the data selection, commenters on the article accused us of cherry-picking the data. To address these concerns, we've created a plot containing the full annual counts of crashes that resulted in a fatality or severe injury on Rowena Avenue for all available years. We straight-line estimate the 2016 counts, as the full-year data is not yet available.
Gathering Data and Formulating a Hypothesis
Our study is observational; we obviously cannot design a real-world experiment where we split motorists into two groups and give one group the road diet configuration and the other group the four-lane configuration. We worked with LADOT on our survey of available datasets and identified four applicable datasets:
1. Induction Loops
One of our datasets is generated from inductive-loop traffic detectors installed at various intersections throughout the Rowena corridor.
Image source: Wikimedia Commons
Induction loops are visible on many Los Angeles streets and are typically placed at intersections. They can be identified by the circular groove in the road surface. The sensors provide measurements of vehicle volume, with measurements taken about once per minute. Los Angeles County makes the live data from these sensors publicly available on the RIITS website.
2. Traffic Count Surveys
Field surveys conducted by LADOT personnel have confirmed the presence of cut-through traffic on Angus. While these surveys provide data of current conditions, they were not conducted prior to the construction of the road diet and we are unable to make a before/after comparison.
3. Waze Data
Through the city's partnership with Waze, we considered looking at traffic alerts and routing data in the Rowena corridor. Unfortunately, we don't have access to Waze's historical routing data dating back to the period before the road diet implementation, meaning we are again unable to make a before/after comparison.
4. Live Traffic Conditions via Mapping APIs
We also tested an approach via simulation. We constructed 25 different starting and endpoints that logically flow through the Rowena corridor. We set up an automated script that queried the Google Maps API during peak periods (7 a.m. to 10 a.m. and 4 p.m. to 7 p.m.) for driving directions. We then analyzed the suggested routes, looking for cases where motorists were sent over the cut-through route of Angus.
We never observed the cut-through route as the first-selected choice, but we did see it as the second or third option. Unfortunately, we were not able to test Waze as they do not offer a public web API. We found some anecdotal evidence that Waze is more aggressive than Google Maps at routing motorists onto side streets such as Angus. This approach also only yields information about present conditions; it doesn't give us any insight into what happened in the days before the road diet.
Limitations on Data Prevented Us From Directly Answering Cut-Through Question
Inductive Loop Placement
Unfortunately, inductive loops are typically placed on busy arterial streets rather than on the quiet residential streets relevant to our Rowena analysis. Vehicle volume on Rowena itself is available via the induction loops both before and after implementation of the road diet, but there are no loops installed anywhere along the primary cut-through route on Angus. Rowena itself experienced lane closures in the years immediately preceding the road diet due to construction by the LA Department of Water and Power.
DWP Construction on Rowena
From 2011 through early 2013, a construction project by the LA Department of Water and Power resulted in the closure of two lanes on Rowena. Data from the induction loops at the beginning and end of Rowena are only available from 2012 onward. As a result, we do not have traffic volumes from any time period when Rowena was configured with four lanes. Any intervention analysis we conduct from the induction loop data will be based on volumes affected by the DWP closure.
One confounding factor is the rise of powerful navigation apps such as Waze that grew increasingly popular during the DWP project before the road diet was installed. Waze grew from 20 million users in July 2012 to 50 million users as of June 2013. This period includes the period of DWP construction on Rowena and the construction of the road diet itself. The increase in usage of navigation apps potentially raises awareness of cut-through routes and contributes to an increase in traffic on streets like Angus. Other Los Angeles residents have blamed Waze for increasing cut-through traffic on residential streets.
Summary of Limitations
To summarize, we discovered the following four limitations over the course of our research and data exploration:
- Induction loop traffic detectors are not installed anywhere along the primary cut-through route along Angus
- Field cut-through traffic surveys were not taken prior to the road diet's implementation
- Rowena was affected by lane closures several years prior to the road diet due to DWP construction
- Navigation apps that provide real-time alternative routes became increasingly popular around the time of DWP construction and road diet implementation
Given the limitations of the data, we determined that we are unable to design an effective hypothesis test for answering the question of cut-through traffic. The lack of induction loops on Angus and pre-diet cut-through field surveys, coupled with the adoption of navigation apps, make it impossible to construct an intervention analysis free of confounding factors.
Making Use of Available Data with Intervention Analysis
Despite these limitations, we were able to use the available data to answer a related question: Did the road diet affect traffic volume on Rowena itself?
We demonstrated usage of this framework using the induction loop data we have available on Rowena. To test a hypothesis in the absence of running a controlled experiment, we used a counterfactual technique to measure the impact of the intervention (also referred to as pre-treatment and post-treatment) and compared the post-intervention outcome to our expectations had no intervention occurred.
We have access to traffic counts on Rowena before and after the implementation of the road diet, and can use our framework to estimate the probability that the effect of the interventions was significant, e.g., whether or not the traffic volume decreased as a result of the road diet.
How We Use CausalImpact
We selected an impact analysis approach using Bayesian structural time-series models developed at Google by Kay Broderson. The accompanying R package CausalImpactprovides a convenient implementation of this approach.
We have used CausalImpact in multiple applications at DataScience where were unable to construct a properly controlled experiment, such as the application of marketing campaigns limited to specific geographic areas. Advantages of CausalImpact include easily interpretable outputs, light data requirements, and reasonable defaults that yield good out-of-the-box performance.
Bayesian Structural Time Series
The CausalImpact framework uses Bayesian structural time series to construct the counterfactual scenario. Bayesian structural time series models include standard trend and seasonal components that function identically to ARIMA models.
For our induction loop data, we observe both time-of-day and day-of-week seasonality. An additional regression component is added for the correlated time series (in our case, the sensor readings from nearby Hillhurst). The regression component features a regularizing spike-and-slab prior that adds shrinkage to our coefficient estimates, which is useful when we need to extract the most predictive control time series from a large pool of candidate control series and a small set of observations in the treatment series.
Instead of a comparing single points in pre/post treatment, CausalImpact generates predictions at each time interval in the post-treatment period
Each state of the post-treatment time series is associated with a probability distribution, allowing for better inference around treatment effects
Post-treatment predictions incorporate trend and seasonality effects. The regression component incorporates behavior of control time series, resulting in a reduction in exposure to outliers in the training period.
The CausalImpact approach constructs a synthetic control from three components:
Behavior of the Rowena traffic volume time series before the road diet (treatment series)
Behavior of other traffic volume time series (control series) that was predictive of the Rowena series prior to the road diet
Specification of priors on model parameters
The choice of second component is the most interesting. Our selection criteria for the control is based on pre-treatment behavior, but its behavior after the treatment is what informs the counterfactual scenario. In our case, we are looking for sensors at locations near Rowena that are likely to share similar behavior but were unaffected by the road diet. The control series provide the model with information on what would have happened on Rowena had no treatment taken place. Our model selected locations along Hillhurst Ave in nearby Los Feliz from several candidates.
Analysis of Model Outputs
We constructed a time series of average vehicle volume on Rowena and Hillhurst (measured via inductive loops and stored in the RIITS database) from January 2012 through summer 2016. We set the date of intervention to March 2013, the month the road diet was completed. Due to the presence of noise in the time series data, such as sensor failure and uneven reading intervals, we resampled the data into 15-minute blocks. Within each block, the recorded volume was averaged across all readings.
Due to high variability in the data from minute to minute, averaging smoothed out the time series and helped us determine if a structural change actually occurred. We also limited the time series to peak-demand times between 7 a.m. through 10 p.m. on weekedays. At off-peak times, traffic volumes on Rowena are much lower.
Here's a plot of the time series for the sensor at Rowena/Glendale. We can see that it appears mostly stable even after March 2013.
Below is the same time series, this time shown with the CausalImpact output. The shaded region represents the 95% interval of the posterior distribution of the counterfactual. The actual observed time series, represented by the line, lies completely within the shaded region. Therefore, the evidence suggests that there was not a material change in traffic volume following the introduction of the road diet. Speeds are lower on the stretch according to LADOT (though still equal to or exceeding the speed limit, on average), and travel times may be longer, especially during peak periods.
This lack of change in volume, combined with the safety improvements, were the core findings in our original op-ed. Based on the data, we concluded that the project is working as intended and is consistent with the experience of numerous other cities worldwide that have pursued similar projects.
What We Learned
Collaborating with the city on the Rowena project was a rewarding — and challenging — process. The road diet is several years old at this point, and has many potential confounding factors. As a result, we cannot determine the extent of the project's impact on cut-through traffic, only that is has improved safety while maintaining traffic volume stability.
However, our method of inferring causal impact could be applied to any project where we have time series data and lack a real control. For future projects, we can work to make sure we have the right data available to robust inferences and to increase confidence in our conclusions. And as we continue to support the city's Vision Zero initiative, methods like this one will be valuable for measuring the impact of related projects — and ultimately improving traffic safety in Los Angeles.
Want to learn more? Register now for our upcoming meetup on the data science of traffic safety featuring Tim Fremaux, transportation engineer from LADOT.