UC Berkeley

Agencies and practitioners are often testing new and innovative strategies for improving driver compliance with traffic regulations. However, in evaluating these strategies, researchers often rely on simple before-and-after methods that suffer from several flaws that can result in misleading results and an inaccurate assessment of a strategy's effectiveness. Specifically, such studies frequently omit control groups to account for other factors that may influence driver behavior aside from the experimental change. Furthermore, these studies often focus on only one compliance measure, and their results are often poorly suited for making comparisons to other compliance strategies that have been evaluated through other before-and-after studies due to the unique details of the experimental sites chosen in each case. Finally, analyses based on the traditional before-and-after approach do not properly account for the period of instability following the experimental change, nor do they make any attempt to characterize it--rather, these studies typically rely on assumptions about how long the instability will last beforehand (and subsequently ignore this period), or fail to account for it at all.

In this dissertation, we examine these flaws and propose a framework that avoids or corrects for them. Among the key features of our proposed framework are a model to describe the driver response to an experimental change (e.g., the increase in compliance following the implementation of a compliance strategy), the inclusion of a baseline prediction model that incorporates control group compliance rates along with other relevant covariates to project what the behavior of the experimental group would have been in the absence of the experimental change, and a measure of effectiveness based on the estimated long-term performance of the compliance strategy after accounting for the period of instability immediately following the experimental change. The framework incorporates the previously documented Novelty Effect, which refers to a short-term boost in compliance due to the novelty of the change, and combines it with a Driver Awareness or driver learning effect, which describes the tendency of the behavioral response (e.g., the boost in compliance) to occur gradually after the experimental change--rather than instantaneously after it--as a result of users taking some time to become aware of the change and to respond appropriately to it. The result is a characteristic driver response curve that is initially increasing after the experimental change, rather than decreasing as is conventionally assumed. When we take a detailed look at compliance data following the implementation of two compliance strategies, we find that the data support this pattern of initially increasing compliance predicted by our framework.

To illustrate the use of this framework, we consider the case of drivers failing to yield on a series of freeway entrance ramps along Interstate 10 in Los Angeles. The problem occurs at several locations along the freeway, with each site having the same general ramp design but a different set of users. Consequently, the entrance ramps are well suited for the application of our framework, as different compliance strategies can be implemented at various locations and comparisons made regarding the resultant improvements in compliance in each case. Furthermore, a subset of the locations can be kept unchanged, to function as control groups for our analyses.

The first compliance strategy to be analyzed was increasing the size of the signage used. Specifically, the size of the Yield sign was increased to from 36 inches across to 48 inches, which increased its viewable area by 78%. After identifying the best prediction model for the behavior of drivers at this experimental location using the pre-change (or "before") data, the baseline compliance predictions were made for the post-change weeks. By comparing the observed levels of compliance to these predicted rates, we could then evaluate the improvement in compliance week by week in the "after" period due to the larger sign. Through curve-fitting techniques, we identified the driver response curve that best characterized the observed degree of improvement by week after the change, and from this we concluded that the larger Yield sign resulted in a 22% reduction in non-compliance at the on-ramp, with a 95% confidence interval of 10% to 26%. This measure of effectiveness excludes an estimated 11% of drivers who are deliberately noncompliant at this location and would not be expected to respond to anything apart from increased enforcement. The Driver Awareness effect was found to be significant in the model for the larger Yield sign at a 5% level, although the Novelty Effect was not.

The second experimental change we considered was the widening of the shoulder on one of the entrance ramps, which was found to have a slightly negative impact on driver compliance as it gave some drivers additional room to negotiate inappropriate merges into the stream of cross-traffic. Using our framework, we found that a wider shoulder resulted in a 16% decrease in compliance. In this case, both the Novelty Effect and Driver Awareness effect were significant on a 5% level.

Finally, we examined the effect of a Yield Line on driver compliance and found it to be 27% effective, with a 95% confidence interval of between 17% and 34%. This measure of effectiveness excludes an estimated 14% of drivers who were deliberately non-compliant at this location. Both the Driver Awareness effect and the Novelty Effect were significant on a 5% level in the Yield Line driver response model.

From our results for the larger Yield sign and the Yield Line, we conclude that the Yield Line was 23% more effective than the larger signage at reducing non-compliance at the yield-controlled entrance ramps in the long term, although these results were obtained for a specific set of environmental conditions: midday Sunday traffic in an uncongested suburban environment, excluding rainy days.

In our analysis of the Larger Sign data and the Yield Line data, we found that the durations of the Novelty Effect and the Driver Awareness effect were best described by Normal distributions, although the Driver Awareness effect was also reasonably modeled by a Uniform distribution. Thus, as part of a sensitivity analysis, we explore the effect of using a Uniform distribution in place of a Normal distribution to describe the duration of the Driver Awareness effect.

We finish our analysis by comparing of our measures of effectiveness to the estimates we would have obtained if we had used a more conventional before-and-after approach instead of our framework, to emphasize the improvements achieved by the latter. We also explore alternate versions of the baseline prediction model to confirm that the form we have chosen is optimal with respect to accurate predictions of compliance rates at the experimental locations.

Our results show that the traditionally-held notion of a driver response (following the implementation of a compliance strategy) that starts off strong and tapers off over time is inaccurate, and that the true driver response includes an initial period of increasing compliance instead. This has important consequences for analyses of compliance strategies, as it indicates that data immediately following a strategy's implementation is likely to under-represent, rather than over-represent, its long term effectiveness.

As more studies are carried out following the framework described here, agencies and practitioners will have a growing library of data to reference regarding the durations of the Novelty Effect and Driver Awareness effect for various kinds of compliance strategies. Such information could enable researchers to establish guidelines regarding appropriate "waiting periods" that must be observed before the driver response may be expected to stabilize. With information regarding the time to stability for a particular experimental change, agencies would be able to perform more simplified analyses of long-term effectiveness without needing to fit driver response curves to the data--in these cases, they would simply exclude the unstable portion of the "after" period instead. However, even in these situations, baseline prediction models would still be needed to ensure that other external factors affecting driver behavior are properly controlled for.

Although we have applied our proposed framework to evaluate compliance strategies on a set of yield-controlled ramp merges, it may also be applied to a wide range of driver compliance situations, such as yielding at pedestrian crossings, speed limit obedience, and proper anticipation of roadway features such as speed humps or intersections. Additional research will be needed to reveal the suitability of our framework (and our characteristic driver response curves) for describing the driver response to strategies implemented in these situations and others. Furthermore, additional research may also provide insight into the extent to which our findings regarding Yield Lines and larger signage can be generalized to other contexts, such as inclement weather or night-time driving.