CHAPTER 7: OBJECTIVE 6, ARE THERE DIFFERENCES IN DRIVING PERFORMANCE FOR DRIVERS WHO ARE ENGAGING IN A DISTRACTION TASK VERSUS THOSE DRIVERS WHO ARE ATTENDING TO DRIVING? ARE SOME OF THE SAFETY SURROGATE MEASURES MORE SENSITIVE TO DRIVING PERFORMANCE DIFFERENCES WHEN DRIVING DISTRACTED VERSUS OTHER SAFETY SURROGATE MEASURES?
To determine whether there were any differences in driving performance between inattentive and attentive drivers, the baseline database was evaluated. A discriminant analysis was conducted to determine if any statistically significant differences were present between the baseline epochs that involved drivers engaging in secondary tasks and/or driving while drowsy and those baseline epochs where the driver was attentive. Prior to conducting the discriminant analysis, a stepwise selection procedure was conducted to determine which driving performance measures were accounting for the highest percentage of variance. This provided insight into which driving performance measures (surrogate safety measures) are most sensitive to inattentive driving.
DATA USED IN THIS ANALYSIS
Table 7.1 presents all the driving performance data that were used in the discriminant analysis. Please recall from Chapter 1: Introduction and Method that the vehicle speed could not be 0 mph for the duration of the epoch. The vehicle was in motion for at least a portion of the 6-second segment for all 20,000 epochs.
Table 7.1. Driving Performance Data Used in the Discrimnant Analysis.
Average percent throttle |
Percent that throttle pedal was depressed by driver over the duration of 6-second epoch. |
Maximum percent throttle |
Maximum percent that throttle pedal was depressed by driver over the duration of the 6-second epoch. |
Minimum lateral acceleration |
Minimum absolute value of lateral acceleration over the 6-second epoch. |
Average lateral acceleration |
Average absolute value of lateral acceleration over the 6-second epoch. |
Maximum lateral acceleration |
Maximum absolute value of lateral acceleration over the 6-second epoch. |
Maximum longitudinal acceleration |
Maximum longitudinal positive acceleration across the 6-second epoch. |
Average longitudinal acceleration/deceleration |
Average longitudinal acceleration/deceleration value across 6-second epoch. |
Maximum longitudinal deceleration |
Maximum longitudinal negative deceleration across the 6-second epoch. |
Yaw time differential |
Duration of the maximum peak-to-peak across the 6-second epoch (i.e., jerk). |
Average speed |
Average vehicle speed across the 6-second epoch. |
Maximum speed |
Maximum vehicle speed across the 6-second epoch. |
There were some driving-performance measures that were not included in the analyses. Some of these measures include forward range, range-rate, and TTC. These dependent measures, while useful in identifying crashes, near-crashes, and incidents when used in conjunction with longitudinal deceleration, were too variable to use with the baseline data. There were many epochs with no lead vehicle present as well as difficulties in filtering spurious radar data when using only 6-second segments. Radar data is notoriously noisy and effectively filtering for this task proved to be too time consuming given the resources available. Even with effective filtering, we hypothesize that this data would not have yielded different results than the results that will be presented with the data that were used.
STEPWISE SELECTION PROCEDURE AND CANONICAL DISCRIMINANT ANALYSIS
A stepwise selection procedure was conducted to determine if all of the above variables are necessary to distinguish between a driver who is engaging in a secondary task or is driving while drowsy to a driver who is attentive to the forward roadway. The stepwise selection procedure initially uses a forward selection procedure but after each selection, the procedure checks to ensure that all the variables previously selected remain significant (Johnson, 1998). In this manner, the stepwise selection procedure will select those driving performance variables or surrogate safety measures that can best discriminate between an attentive and an inattentive driver.
Table 7.2 presents those surrogate safety measures that the stepwise selection procedure selected. The standardized canonical coefficient can be used to interpret the relative contribution that each variable is making to the model. The magnitude and the sign of the value are both used in this interpretation; therefore, the average percent throttle is contributing the most to the model whereas yaw time differential is contributing the least.
Table 7.2 The safety surrogate measures that best discriminate between attentive and inattentive drivers.
The stepwise selection procedure also indicated that these four safety surrogate measures together achieved a multivariate measure analogous to an R-squared value of 0.004 indicating that these four variables account for less than 1 percent of the variance associated with inattentive and attentive driving. While differences are present between attentive and inattentive drivers, these surrogate safety measures are not adequately explaining these differences.
DISCRIMINANT ANALYSIS
The discriminant analysis was conducted to determine whether these surrogate safety measures were predictive of inattentive driving. Table 7.3 shows that 51.4 percent of the attentive epochs were correctly classified and 54.5 percent of the inattentive epochs were correctly classified. These results suggest that the predictive linear model using these surrogate safety measures is not able to accurately predict whether the driver is attentive or inattentive as these percentage values are too close to 50 percent accuracy or chance.
Table 7.3. The percent of baseline epochs that the linear discriminant analysis model was successfully able to distinguish.
51.4 |
48.6 |
100 |
45.8 |
54.2 |
100 |
48.5 |
51.5 |
100 |
DISCUSSION
The stepwise selection procedure indicated that the average percent throttle, yaw time differential, average lateral acceleration, and maximum longitudinal deceleration were the safety surrogate measures most sensitive to inattentive driving. While these safety surrogate measures were most sensitive to inattentive driving, they were only able to account for less than 1 percent of the variance. The subsequent discriminant analysis indicated that the predictive abilities of these four safety surrogate measures to distinguish between attentive and inattentive driving was not better than chance or 50 percent accuracy.
Other discriminant analyses using the variance of the above safety surrogate measures were also attempted. These results were similar to the above results in that the surrogate safety measures selected in the stepwise selection procedure accounted for less than 1 percent of the variance. The discriminant analysis also indicated poor predictability that was not significantly different from chance (i.e., 50 percent were correctly identified and 50 percent were incorrectly identified).
There are several hypotheses as to why the surrogate safety measures did not adequately explain the differences in attentive versus inattentive driving. One hypothesis is that the results from these analyses are accurate and that inattentive driving does not in fact differ significantly from attentive driving. Rather it is only in the presence of multiple other contributing factors and extreme circumstances that differences exist in the inattentive driver’s ability to effectively respond versus an attentive driver’s ability to effectively respond to an emergency situation. Testing this hypothesis is possible with the 100-Car Study data but would require specific baseline events to be identified and reduced that match on a variety of environmental and situational variables per individual driver. This reduction and analysis effort is beyond the scope of this project but could be conducted in the future.
A second hypothesis is that there are differences that exist for these safety surrogate measures but these differences are not being captured adequately by using point estimates. A point estimate may not be accurately capturing the differences between inattentive and attentive drivers. A different statistical analysis or what is known as functional data analysis may produce different results. Functional data analysis would use overall rates of change for each baseline epoch rather than a point estimate to summarize the data for that epoch. While this technique could be used, it would require additional data reduction and time spent researching these relatively new data analysis methods. These techniques are generally not attempted unless the point estimate analysis produced some promising results; therefore, this hypothesis should only be tested as a last resort.
A third explanation for these findings is that the 6-second duration for the baseline epochs is too short to accurately assess driving performance. Recall that the baseline epochs were 6 seconds in duration to compare to the time frame used by trained data reductionists to assess whether a particular behavior or action by the driver contributed to the occurrence of the crash, near-crash, or incident. It is unknown whether a point estimate for a longer duration of time would be any better than the analysis already conducted. Also note that lengthening the time duration would require additional data reduction.
After conducting multiple discriminant analyses using a variety of surrogate safety measures, it is clear that the databases that currently exist are not adequate to test the above hypotheses that are listed here. More data reduction that is specifically designed to adequately assess driving performance for individual drivers during specific environmental conditions is required to further assess this research objective.