Urban Analytics Final Project

I used GDOT’s Crash Data Dashboard to download a dataset including all accidents reported in the Atlanta Metro between 2021 & 2023. In the code below, I cleaned the data for analysis.

In the following code, I spatially joined both sf dataframes to result in a final dataframe with a binary column ‘involved_truck’ that specifies whether a truck was involved or not in all of the accidents reported in the full dataset.

Next, I use a search pattern to filter the dataset to only include data points that occur on one of the metro Atlanta highways I am studying. The method I used was very effective at filtering out accidents on other roads, but it is not perfect due to the inconsistency in the naming of roads in GDOT’s data. However, the results of ther filtering are sufficient for their purpose: a map-based visualization of accidents.

I used GDOT’s Traffic Analysis & Data Application to download a dataset that includes average traffic flow (daily number of cars) and the percentage of that flow that is trucks (freight vehicles) at specified sensors alongside different highways in metro Atlanta.

Each point was used to estimate flow & truck percentage values for short segments of highway defined later in this report. To note: each point covers only one direction of the highway, and values for traffic in each direction of various highways was unavailable from GDOT. Therefore, in this report, I am assuming that flow rates are the same in each direction of traffic. In the code below, I specify the specific traffic sensors I will use in my analysis from GDOT’s traffic flow dataset and clean the data from GDOT and specify the sensors useful for the highways I am studying.

A small number of the sensors I’m studying had missing values for the percentage of the flow that was trucks. For these points, I used imputatiuon to estimate the percentage at each point based on values from the same highway but further from the city center and the difference between points near and far from the city on other highways.

I imported a shapefile of all highways in Georgia for mapping purposes. I then created a linestring spatial feature objects for each segment of highway I will be analyzing.

My next data source is TomTom, which measures traffic backups and speeds based on cell phone GPS data. From their website, I specified 44 segments of highway (in both directions of traffic) from across the metro area in order to study the impact of traffic speeds.

First, I read each of the 44 shapefiles from a folder and assigned id’s.

Next, I imported and cleaned a data file that included variable data for each segment.

Note: The data I gathered is from the entirety of August 2024, as my free trial only allowed for the download of one month’s data.

Note: From the TomTom data, the most useful variable I will include in my models is the speed differential value. In the dataset I downloaded, I gathered travel times (speeds) at four different time periods for each day: morning rush hour (7am-10am), midday (12pm-2pm), evening rush hour (3:30pm-7pm), and overnight (12am-2am). The differential value for each row in the dataset measures the difference in average speed between the given time of day and the base value, defined as the overnight speed. For example, a morning rush value of 1.5 would mean that the overnight speed divided by the morning speed = 1.5. A higher differential value means bigger difference in overnight vs daytime travel times, or in other words, more traffic/delays. This data will allow me to use a measure of traffic as an independent variable in modeling.

Mapping Traffic on Atlanta Highways:

In these maps, the truck flow percentage is indicated by segment line width.

By mapping the speed differentials, I can get a better view of where traffic backups occur. I mapped differentials in the evening rush hour for traffic exiting Atlanta and in the morning rush hour for traffic entering Atlanta. I see that the morning differentials are much higher on the highways on the northside of Atlanta than the southside, possibly indicating a larger number of commuters using those roadways. I also see that differentials are far higher on the portions of highway outside of the perimeter, and lower inside the perimeter where truck traffic is limited. In the afternoon, the differentials are more evenly distributed across the region, although you see again that differentials are higher outside of the perimeter than inside. Also, the highest and lowest differentials happen to be on I-85 north of I-285 and on I-85 south of I-285, which is interesting. These two maps indicate that there could certainly be a relationship between the truck percentage on a highway and the buildup of rush hour traffic. Additionally, the fact that differentials are higher going towards the city rather than away could be a result of heavy congeestion at each of the I-285 interchanges, which is also heavily impacted by the number of trucks on I-285. Lastly, I note that speed differentials are actually highest on many portions of I-285, indicating that it is the highway most impacted by traffic delays in the metro.

Mapping Accident Frequency:

In these maps, the truck flow percentage is indicated by segment line width.

These two maps have a few interesting characteristics. In the first, you can see that each of the highways segments with a low truck percentage (<6%) also has a lower than average frequency of accidents. Of segments with a higher truck percentage, some segments have few accidents and other have many; no clear conclusion can be made spatially. The second map shows that highway segments with the highest averade daily traffic flow rates (I-75 in Cobb, I-85 in Gwinnett, I-285 North of I-20) tend to have some of the highest frequencies of accidents as well. From both maps, I can gather that the portion of I-285 between GA400 & I-85 has a strikingly high rate of accidents, which may be explained by the fact that the interchanges at each endpoint of this segment are known as some of the worst bottlenecks in the country (source). Additionally, GA400 has a very low accident rate, which may be explained by the fact that it has much fewer freight vehicles given that it is a state highway, not an interstate. The portions of I-85 & I-75 south of I-285 have much lower accident frequencies than the metro average.

Mapping Injury Frequency:

In this map, the truck flow percentage is indicated by segment line width.

From this map, it is very evident that the frequency of injuries is much higher on I-285 than on other highways, and the southern portion of I-285 has a particularly high frequency of injuries. Additionally, GA400 (both within and outside of the permiter), which has the lowest truck percentage in the metro, has the lowest rate of injuries out of any metro Atlanta highway. Injury frequency is high on most highway segments with a particularly high truck percentage, with I-75 south of the I-675 intersection being an exception; this could possibly be explained by the fact that this segment begins far outside of I-285 and extends further from the city center than any other segment. Another interesting characteristic is that while accidents are more frequent on the eastern half of I-285 than the western half, the opposite is true for injuries.

Mapping Fatality Frequency:

In this map, the truck flow percentage is indicated by segment line width.

The results of the third map are very interesting. Again, you see a very low rate of fatalities on GA400. I also see a low rate of fatalities on the western portion of I-285 despite it having a high rate of injuries. South of the I-675 interchange, I-75 does not have a particularly low rate of fatalities despite having a low rate of accidents & injuries; this highway segment has the highest truck percentage in the metro. I also see that, similarly to accident rates, the southern section of I-285 (which has a very high truck percentage) also has a very high frequency of fatalities. Inside of the perimeter, where truck traffic is limited, fatality rates tend to be lower, with the exception of the portion of I-85 between GA400 & I-285.

Number of Vehicles per Accident:

## 
##  Wilcoxon rank sum test with continuity correction
## 
## data:  num_vehicles by involved_truck
## W = 1733508958, p-value < 2.2e-16
## alternative hypothesis: true location shift is not equal to 0

There is an average of 2.02 vehicles involved in accidents without trucks, and 2.09 vehicles involved in accidents with trucks. Although this difference in means is small, I have a very large number of data points, and therefore the Wilcox test results in a p-value much lower than 0.05, which allows us to conclude with strong statistical significance that the average number of vehicles involved in each crash involving trucks is higher than the average number of vehicles involved in each crash not involving trucks.

Number of Serious Injuries:

## 
##  Wilcoxon rank sum test with continuity correction
## 
## data:  serious_injuries by involved_truck
## W = 1839658436, p-value = 3.822e-09
## alternative hypothesis: true location shift is not equal to 0

There is an average of 0.011 serious injuries involved in accidents with trucks, and 0.017 vehicles involved in accidents without trucks. This would indicate that serious injuries are 50% more likely in accidents not involving trucks. I have a very large number of data points, and therefore the Wilcox test results in a p-value much lower than 0.05, which allows us to conclude with strong statistical significance that the average number of injuries involved in each crash involving trucks is lower than the average number of injuries involved in each crash not involving trucks.

Number of Fatalities:

## 
##  Wilcoxon rank sum test with continuity correction
## 
## data:  fatalities by involved_truck
## W = 1.832e+09, p-value = 0.9066
## alternative hypothesis: true location shift is not equal to 0

The average number of fatalities involved in accidents with or without trucks is almost identical. The Wilcox test shows no statistically significant difference between the groups. The results of this may be skewed by the fact that despite having a very large dataset, the number presence of fatal accidents is uncommon.

Accident Frequency:

The first regression model identifies which independent variables impact the frequency of accidents. The results show a statistically significant positive correlation between all three independent factors and the frequency of accidents. This means that highways with more severe daytime delays, a higher number of vehicles on the road in general, or a higher truck percentage are more likely to see car accidents. The R-Squared value of 0.27 is not all that strong, which suggests that there could be other factors not included in the model that may also have an effect. It is important to note, though, that the R-Squared value may be lower due to the fact that I am studying only 44 highway segments. The following scatterplot also displays the relationship between the truck percentage and the frequency of accidents:

Injury Frequency:

The second regression model identifies which independent variables impact the frequency of serious injuries. The results show a statistically significant positive correlation between the truck percentage and the frequency of serious injuries. This means that serious injuries occur more frequently as the percentage of trucks on a highway increases. The R-Squared value of 0.09 is weak, which suggests that there could be other factors not included in the model that may also have an effect. It is important to note, though, that the R-Squared value may be lower due to the fact that I am studying only 44 highway segments. The following scatterplot displays the relationship between the truck percentage and the frequency of serious injuries:

Fatality Frequency:

The third regression model identifies which independent variables impact the frequency of fatalities. The results show a statistically significant positive correlation between the truck percentage and the frequency of fatalities. This means that serious injuries occur more frequently as the percentage of trucks on a highway increases. The R-Squared value of 0.06 is weak, which suggests that there could be other factors not included in the model that may also have an effect. It is important to note, though, that the R-Squared value may be lower due to the fact that I am studying only 44 highway segments. The following scatterplot displays the relationship between the truck percentage and the frequency of fatalities:

Number of Vehicles:

The fourth regression model identifies which independent variables impact the number of cars involved in accidents. The results show a statistically significant positive correlation between all three independent factors and the number of cars involved in accidents. This means that highways with more severe daytime delays, a higher number of vehicles on the road in general, or a higher truck percentage are more likely to see car accidents involving a large number of vehicles. The R-Squared value of 0.28 is not very strong, which suggests that there could be other factors not included in our model that may also have an effect. It is important to note, though, that the R-Squared value may be lower due to the fact that I am studying only 44 highway segments.

Predicting Annual Accidents:

## [1] 0.2075943

## [1] -7863.456

The model output finds that in the event that trucks were to be diverted off of I-285, there would be a 21% reduction in the frequency of accidents, which would amount to roughly 7,863 fewer accidents annually. The following bar chart shows an estimate of the current number of accidents occurring annually and the predicted value from the linear regression model for each segment of I-285.

Predicting Annual Injuries:

## [1] 0.3467532

## [1] -194.8694

The model output finds that in the event that trucks were to be diverted off of I-285, there would be a 35% reduction in the number of serious injuries, which would amount to roughly 195 fewer serious injuries annually. The following bar chart shows an estimate of the current number of serious injuries occurring annually and the predicted value from the linear regression model for each segment of I-285.

Predicting Annual Fatalities:

## [1] 0.1848804

## [1] -33.36767

The model output finds that in the event that trucks were to be diverted off of I-285, there would be an 18% reduction in the number of fatalities, which would amount to roughly 33 lives saved every year. The following bar chart shows an estimate of the current number of fatalities occurring annually and the predicted value from the linear regression model for each segment of I-285.

Note: In this portion, I selected all segments of I-285, covering the entire perimeter. However, when drawing the segments, I did not include the many interchanges along I-285, as this would certainly alter the data. Because those short (but certainly dangerous) sections of highway are not present, the reductions in accidents, injuries, and fatalities that I calculated are actually underestimates.