National Academies Press: OpenBook

Using Electronic Devices While Driving: Legislation and Enforcement Implications (2021)

Chapter: Appendix B - Proposed Sample of Jurisdictions and Methodology

« Previous: Appendix A - Protocol for Ranking the Strength of Distracted Driving Laws
Page 45
Suggested Citation:"Appendix B - Proposed Sample of Jurisdictions and Methodology." National Academies of Sciences, Engineering, and Medicine. 2021. Using Electronic Devices While Driving: Legislation and Enforcement Implications. Washington, DC: The National Academies Press. doi: 10.17226/26082.
×
Page 45
Page 46
Suggested Citation:"Appendix B - Proposed Sample of Jurisdictions and Methodology." National Academies of Sciences, Engineering, and Medicine. 2021. Using Electronic Devices While Driving: Legislation and Enforcement Implications. Washington, DC: The National Academies Press. doi: 10.17226/26082.
×
Page 46
Page 47
Suggested Citation:"Appendix B - Proposed Sample of Jurisdictions and Methodology." National Academies of Sciences, Engineering, and Medicine. 2021. Using Electronic Devices While Driving: Legislation and Enforcement Implications. Washington, DC: The National Academies Press. doi: 10.17226/26082.
×
Page 47
Page 48
Suggested Citation:"Appendix B - Proposed Sample of Jurisdictions and Methodology." National Academies of Sciences, Engineering, and Medicine. 2021. Using Electronic Devices While Driving: Legislation and Enforcement Implications. Washington, DC: The National Academies Press. doi: 10.17226/26082.
×
Page 48
Page 49
Suggested Citation:"Appendix B - Proposed Sample of Jurisdictions and Methodology." National Academies of Sciences, Engineering, and Medicine. 2021. Using Electronic Devices While Driving: Legislation and Enforcement Implications. Washington, DC: The National Academies Press. doi: 10.17226/26082.
×
Page 49
Page 50
Suggested Citation:"Appendix B - Proposed Sample of Jurisdictions and Methodology." National Academies of Sciences, Engineering, and Medicine. 2021. Using Electronic Devices While Driving: Legislation and Enforcement Implications. Washington, DC: The National Academies Press. doi: 10.17226/26082.
×
Page 50
Page 51
Suggested Citation:"Appendix B - Proposed Sample of Jurisdictions and Methodology." National Academies of Sciences, Engineering, and Medicine. 2021. Using Electronic Devices While Driving: Legislation and Enforcement Implications. Washington, DC: The National Academies Press. doi: 10.17226/26082.
×
Page 51
Page 52
Suggested Citation:"Appendix B - Proposed Sample of Jurisdictions and Methodology." National Academies of Sciences, Engineering, and Medicine. 2021. Using Electronic Devices While Driving: Legislation and Enforcement Implications. Washington, DC: The National Academies Press. doi: 10.17226/26082.
×
Page 52

Below is the uncorrected machine-read text of this chapter, intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text of each book. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.

B-1 This appendix discusses the methodology used to select a sample of jurisdictions for in-depth review of distracted driving laws, educational and outreach efforts, enforcement strategies, and legislation challenges and successes. B.1 Proposed Sample of Jurisdictions As indicated in the project work plan, following an initial scan of existing distracted driving legislation, a cluster analysis of states and provinces with similarly structured legislation was conducted to group the jurisdictions for further study. The cluster analysis was based on the pro- posed Protocol for Ranking the Strength of the Law (now Appendix A, but delivered March 11, 2019, and reviewed by the panel) as well as additional characteristics, as described in detail in Section B.2. Table B-1 presents the selected sample of 20 states and provinces (and backups) representing the different groupings. These jurisdictions were included in the in-depth analysis of distracted driving law implementation. A P P E N D I X B Proposed Sample of Jurisdictions and Methodology Stratum (cluster) General Description of Cluster Sampled Jurisdiction(s) Alternate(s) 1 Canadian provinces Alberta, Manitoba, Ontario, Quebec New Brunswick 2 U.S. states, relatively stronger laws Connecticut, Maryland, Maine, Oregon, West Virginia, Vermont Georgia, New Jersey 3 U.S. states, moderate-to-small populations, relatively weaker laws Idaho, Kentucky, Louisiana, New Mexico, South Carolina, South Dakota, Tennessee Missouri, North Dakota 4 U.S. states, large populations, relatively weaker laws Pennsylvania, Virginia California 5 U.S. states, weakest laws Nebraska Arizona Note: Maryland and Ontario were selected with certainty. Alaska, District of Columbia, Hawaii, Montana, and Prince Edward Island were excluded from selection. Table B-1. List of proposed jurisdictions by sampling stratum (cluster). Figure B-1 displays a map of the sampled jurisdictions (shaded in blue) for the in-depth review. B.2 Methodology B.2.1 Motivation As previously noted, the objective is to obtain a sample of 20 out of 61 jurisdictions (50 U.S. states plus the District of Columbia, and 10 Canadian provinces) to contact for a more detailed review. This sample has a relatively diverse set of laws related to distracted driving, and a range

B-2 Using Electronic Devices While Driving: Legislation and Enforcement Implications of population characteristics, so that the information gathered covered a range of laws, chal- lenges, and populations. The sample represents jurisdictions with stronger distracted driving laws somewhat more heavily, since the objective of this project is to develop model legislation as well as best practices for education and enforcement. Additionally, the details of these laws often vary, and having a diverse sample will allow any findings to be generalizable to other jurisdic- tions rather than limited to unique aspects of a particular jurisdiction. A simple random sample would not guarantee that the sample covers a range of law and popu- lation characteristics and would not result in a sample with more “strong law” jurisdictions. At the other extreme, a hand-picked sample runs the risk of being biased since it would be selected at least in part by subjective, rather than objective, criteria. An ideal sample is between these two extremes: data-driven, but informed by expert opinion. To accomplish this, the researchers used a clustering algorithm to group all 61 jurisdictions into five clusters based on characteristics of their distracted driving laws and populations. (Before settling on this approach, several clustering methods were tested; these are discussed in Section B.2.4.) The researchers then drew a fixed sample of jurisdictions from each cluster, along with at least one alternate jurisdiction per cluster. This form of stratification, while somewhat rough, guaranteed that jurisdictions with a variety of law and population characteristics were represented in the final sample. B.2.2 Data Sources The analysis file was composed of data from several different sources, as discussed in the following. B.2.2.1 Jurisdiction Laws Access Database This is the database prepared by the Westat study team, which coded different aspects of the law for each jurisdiction. Using the database, jurisdiction laws were scored based on the protocol outlined in Appendix A: Protocol for Ranking the Strength of Distracted Driving Laws. Each variable level was assigned 0 to 3 points, depending on the variable. Any jurisdiction without a Figure B-1. Map of sampled jurisdictions.

Proposed Sample of Jurisdictions and Methodology B-3 law, or without a particular element in the law, was assigned a score of 0 for the variable. In general, more points indicate a stricter law. More details on each of these variables can be found in Appendix A. The variables, levels, and assigned points are: • Violation Type (None = 0, secondary = 1, primary = 2); • Behaviors Covered (none = 0, texting = 1, manipulating/dialing = 2, handheld = 3, reading/ speaking = +1 extra point); • Drivers Covered (none = 0, specific populations only = 1, all drivers = 2, all drivers + more stringent for specific populations = +1 extra point); • When Law Is Enforceable (none = 0, only when vehicle is in motion = 1, at all times = 2); • Penalty Range (warning/0 pts = 0, 1–2 pts = 1, 3–4 pts = 2, 5+ pts = 3); • Incremental Penalty (yes = 1, no = 0); • Fine Range (no fine = 0, $1–$75 = 1, $76–$150 = 2, $151+ = 3); and • Incremental Fine (yes = 1, no = 0). For the remaining jurisdiction-level data sources, the researchers used data from 2017 since it was the most recent year available for all data sources. For data on provinces, territories were excluded (Yukon, Northwest Territories, Nunavut) as outside the scope of this project, although they are included in most Canadian source data. B.2.2.2 Fatal Injuries, 2017 For states, data were obtained from FHWA (https://www.fhwa.dot.gov/policyinformation/ statistics/2017/). For provinces, data were not directly available. A table was obtained from Transport Canada (https://www.tc.gc.ca/eng/motorvehiclesafety/canadian-motor-vehicle-traffic-collision-statistics- 2017.html). The table gives fatalities per 100,000 population rather than fatalities directly, so the authors obtained 2017 population estimates by province from Statistics Canada (https:// www150.statcan.gc.ca/t1/tbl1/en/tv.action?pid=1710000501) and multiplied these fatality rates by 100,000 population to obtain fatality counts directly. Due to rounding, these estimates of fatalities may differ slightly from the actual counts, but they should be comparable to the U.S. state fatality counts for analysis purposes. B.2.2.3 Licensed Drivers, 2017 For states, data were obtained from FHWA (https://www.fhwa.dot.gov/policyinformation/ statistics/2017/). For provinces, these data were not directly available. As a proxy, the authors used a table of fatalities by province from Transport Canada (https://www.tc.gc.ca/eng/motorvehiclesafety/ canadian-motor-vehicle-traffic-collision-statistics-2017.html). This table presents fatality rates per 100,000 population and per 100,000 licensed drivers. As described previously, the authors first used population to obtain an estimate of province-level fatalities, then used that estimate of fatalities to estimate the number of licensed drivers, since the fatality rate per 100,000 licensed drivers should simply be fatalities divided by licensed drivers/100,000. Again, due to rounding, this calculation may not result in the exact number of licensed drivers but should be a close enough proxy for this analysis. B.2.2.4 Motor Vehicle Registrations, 2017 For states, these data were obtained from FHWA (https://www.fhwa.dot.gov/policyinformation/ statistics/2017/). For provinces, these data were obtained from Statistics Canada (https://www150.statcan. gc.ca/t1/tbl1/en/tv.action?pid=2310006701).

B-4 Using Electronic Devices While Driving: Legislation and Enforcement Implications In both data sources, publicly owned vehicles were excluded because privately owned vehicles are of primary interest for this analysis and because for both states and provinces the data sources note that the publicly owned vehicle counts may be of poor quality. B.2.2.5 Vehicle Miles Traveled, 2017 For states, these data were obtained from FHWA (https://www.fhwa.dot.gov/policyinformation/ statistics/2017/). For provinces, the latest direct province-level estimates of vehicle kilometers traveled (VKT) were from 2009, when the Canadian Vehicle Survey was ended. Rather than use outdated infor- mation, the authors again used the table of fatalities by province from Transport Canada (https:// www.tc.gc.ca/eng/motorvehiclesafety/canadian-motor-vehicle-traffic-collision-statistics-2017. html), since fatalities are also provided by billion VKT. As described previously, the researchers first used population to obtain an estimate of province-level fatalities, then used that estimate of fatalities to estimate VKT, since the fatality rate per billion VKT should simply be fatalities divided by VKT/1,000,000,000. Again, due to rounding, this calculation may not result in exact VKT, but should be a close enough proxy for this analysis. The researchers then converted VKT into VMT (vehicle miles traveled) (1 km = 0.6213712 mi) for comparability across states and provinces. B.2.2.6 Population, 2017 For states, state-level population estimates were obtained from the 2017 American Com- munity Survey of the U.S. Census Bureau, both overall and for 15- to 24-year-olds alone. This was because studies such as NOPUS suggest that younger drivers may use electronic devices at a significantly higher rate than older drivers. (See, for example, https://crashstats.nhtsa.dot.gov/ Api/Public/ViewPublication/812665; this publication also suggests that electronic device use rate may be higher among males than females. However, since for nearly all states males make up between 48% and 50% of the population, it is unlikely that variation in that small range would be useful for this analysis.) For provinces, similar population data were obtained from Statistics Canada (https://www150. statcan.gc.ca/t1/tbl1/en/tv.action?pid=1710000501). B.2.3 Data Preparation The laws database was coded from text variables into numeric variables for analysis. When necessary, blank values were coded as 0. For example, since Montana has no electronic device use laws, it was assigned a score of 0 for all law components. After merging all data sources into a single file, the researchers ran quality-control checks to verify that all merging was done correctly and to check for any outlying or unexpected data points. During these checks, it was noted that the correlations between population and fatalities, licensed drivers, vehicle registrations, VMT, and number of 15- to 24-year-olds were very high (ranging between 0.90 and 0.99). Since this degree of multicollinearity is likely to cause problems with modeling, the authors converted fatalities, licensed drivers, vehicle registrations, VMT, and number of 15- to 24-year-olds to rates per 1,000 population. Before clustering, all variables were standardized to have a mean of 0 and a standard error of 1. This effectively weights each variable in the analysis equally. If unstandardized variables were used, variables on a larger scale (such as population) would dominate the models, and the categorical law-related variables would have virtually no impact. As a final check, the authors looked at the correlation between normalized fatality rate and normalized law strength, calculated by totaling all law strength points. The correlation is

Proposed Sample of Jurisdictions and Methodology B-5 -0.57 (p < 0.0001), which is moderate-to-strong, negative, and significantly different from zero. This means that jurisdictions with stronger laws tend to have lower fatality rates, and vice versa. Even though fatality rates include all types of traffic-related fatalities, not just distracted driving–related fatalities, at a very high level it is encouraging to see that the law strength scoring has the relationship one would expect with fatality rates. B.2.4 Clustering Analysis Three different clustering methods were tested. This was done as a sensitivity analysis; each clustering method has strengths and weaknesses, and the researchers wanted to ensure that the different methods independently produced roughly consistent results. B.2.4.1 Hierarchical Clustering Hierarchical clustering works from the bottom up, combining elements that are the most similar to each other. Each jurisdiction starts out as its own cluster. Then, the algorithm calcu- lates a dissimilarity score for each pair of jurisdictions. [The researchers used the hclust function in R, with Ward’s minimum variance method (“ward.D2”). This method tends to result in the most compact clusters.] The jurisdictions with the lowest dissimilarity score (most similar) are grouped together, and a new set of dissimilarity scores were calculated using the new cluster. This process continued iteratively until all jurisdictions were grouped together into a single cluster. The output of this process is a tree-like object called a dendrogram. The dendrogram is a visual representation of the order in which observations were grouped together and is helpful to visualize the major groupings. Observations that were grouped together further up the tree are less similar from each other. The height of the “arms” joining clusters together also indicates how similar the clusters are—a small height means that the two clusters are very similar, while a large height value would indicate a larger difference. Figure B-2 shows the dendrogram for the dataset, where each number represents a jurisdic- tion. Based on this figure, it looks like five major groupings of jurisdictions would be reason- able (identified by the red boxes). Although these clusters could be divided further, for the purposes of this study, it was generally preferable to have fewer, larger clusters than to overly restrict the sampling by defining too many clusters. Figure B-2. Dendrogram of hierarchical clustering output with 5 clusters defined.

B-6 Using Electronic Devices While Driving: Legislation and Enforcement Implications B.2.4.2 K-means The K-means clustering algorithm is somewhat more advanced because instead of working iteratively by joining pairs, it is able to find optimal clusters across the whole dataset at once. It works by first randomly assigning each observation to a cluster. Then, it calculates the centroid (center mass) of each cluster and reassigns each observation to the cluster with the closest centroid. It iterates through these two steps until no observations change cluster assignments. The final clusters are optimal in the sense that they have the lowest within-cluster variation. [The researchers used the K-means function in R, with the default (Hartigan-Wong) algorithm and 20 random starts. Using multiple random starts protects against the algorithm stopping at a local optimum value or failing to converge.] The downside to K-means clustering is that the analyst must specify the number of clusters, and this is often unknown. However, since the hierarchical clustering algorithm was run first, the researchers were able to specify that five final clusters were wanted. For sensitivity testing purposes, the researchers also ran the K-means algorithm with 10 clusters. B.2.4.3 Model-Based Cluster Analysis (Mclust) The Mclust algorithm uses Gaussian (normal) mixture models to determine the optimal number of clusters, optimizing over the Bayesian information criterion (BIC), a measure of overall model fit that adds a penalty for models with more parameters (more clusters). This means that the algorithm does not suggest splitting into more clusters unless the split improves model fit more than the size of the penalty term. (The researchers used the Mclust package in R, with default settings except for adding a normal prior.) The final number of clusters and how the observations are clustered comes from the model with the largest BIC; that is, the best overall fit according to the BIC. The researchers chose to add a normal prior to the model to smooth out some of the noise (“regularization”). After regu- larization, this algorithm recommended using five clusters as well. As a further sensitivity analysis, the researchers also tested running each algorithm with and without the extra points for driver type and behaviors covered. The results without the extra points were not markedly different from those presented here, so they were not pursued further. B.2.5 Clustering Results Since all three methods suggested five clusters, and the final clusters (see Table B-2) were very similar between algorithms, only the clusters from the final K-means model with five clusters are presented. Because hierarchical clustering looks for optimal pairwise combinations only (combining two jurisdictions or two clusters), it can often fail to identify an overall optimal solu- tion; the researchers used this method primarily to search for a reasonable number of clusters to input into the other methods. Model-based clustering is another good alternative, but it relies on the assumption that data are a mixture of multivariate normal distributions. This is not true in these data: even after normalization, the categorical variables used for law coding will always be bunched into 2- to 4-point masses rather than being approximately continuous. K-means, however, does not place any assumptions on the distributions of the variables, and it is designed to find an overall optimal solution. It is the method best suited to this particular problem, so it makes most sense to rely on the K-means results. The heat map depicted in Table B-3 shows how the authors were able to generalize and describe each cluster. Red means that a cluster is, on average, low in terms of the row variable, and green indicates clusters with a higher average value for the row variable. For example, looking at the total population row, Cluster 1 (primarily Canadian provinces) has the lowest average popula- tion, and Cluster 4 (U.S. states with large populations) has the highest average population.

Proposed Sample of Jurisdictions and Methodology B-7 Note that these cluster descriptions are intentionally general, and each jurisdiction in a cluster may not perfectly agree with the high-level description. However, since the goal is simply to group jurisdictions that are roughly similar in order to end up with variety in the final sample, these clusters and descriptions work well. As a final check of the clustering, the researchers plotted (see Figure B-3) each jurisdiction against total population (x-axis) and against its total law strength (y-axis), where the law strength is simply the sum of all law-related variables. Both variables were normalized before plotting, so that 0 represents the average of all jurisdictions. This means that a jurisdiction with a law strength greater than 0 has stronger than average laws, based on the scoring procedure. Each jurisdiction name is color-coded based on the cluster, so that jurisdictions in the same cluster are the same color. This is a crude method of checking whether the clusters make intuitive sense; although the clusters were formed using the individual law variables rather than overall law strength, and population was not the only non-law factor used, one would still expect jurisdictions in the same cluster to be relatively close to each other on this plot, which is the case. The pink jurisdic- tions represent the stricter-law jurisdictions from Cluster 2, and as expected they all appear in Stratum (Cluster) General Description of Cluster Jurisdictions 1 Canadian provinces and Colorado, stronger laws, lower population Alberta, British Columbia, Colorado, Manitoba, New Brunswick, Newfoundland and Labrador, Nova Scotia, Prince Edward Island**, Quebec, Saskatchewan 2 U.S. states and Ontario, relatively stronger laws Connecticut, Delaware, District of Columbia**, Hawaii**, Georgia, Maine, Maryland*, Massachusetts, Nevada, New Hampshire, New Jersey, Ontario*, Oregon, Rhode Island, Vermont, Washington, West Virginia 3 U.S. states, moderate-to-small populations, relatively weaker laws Alabama, Alaska**, Arkansas, Idaho, Indiana, Iowa, Kansas, Kentucky, Louisiana, Mississippi, Missouri, New Mexico, North Carolina, North Dakota, Ohio, Oklahoma, South Carolina, South Dakota, Tennessee, Utah, Wisconsin, Wyoming 4 U.S. states, large populations, relatively weaker laws California, Illinois, Michigan, Minnesota, Pennsylvania, New York, Texas, Virginia 5 U.S. states, weakest laws, high vehicle usage/fatality rates Arizona, Florida, Montana**, Nebraska Notes: *Maryland and Ontario were selected with certainty for the final sample. **Alaska, District of Columbia, Hawaii, Montana, and Prince Edward Island were included in the clustering process but excluded from selection. Table B-2. Final clusters of jurisdictions. Variable Cluster Number 1 2 3 4 5 Law-Related Variables Violation type Behaviors covered Drivers covered Law enforceable Penalty range Incremental penalty Fine range Incremental fine Non-Law-Related Variables Total population Fatality rate % population ages 15–24 Licensed driver rate Vehicle registration rate VMT per population Table B-3. Heat map of average clustering variable value, by cluster.

B-8 Using Electronic Devices While Driving: Legislation and Enforcement Implications the upper portion of the plot. The Canadian provinces are shown in tan (also with stricter laws, since they all have law scores > 0), while the weaker-law U.S. states are split between the blue (Cluster 3, small-to-moderate populations) and coral (Cluster 4, larger populations). The green jurisdictions in Cluster 5 are spread across the plot more widely; this is because the clustering is based on individual variables rather than the totals on this plot. These jurisdictions are similar in law components, while their overall total scores are different; they also all have higher fatality rates, licensed driver rates, and vehicle registration rates than the average jurisdiction. Overall, these clusters seem to be effective at grouping together jurisdictions that are similar in terms of population characteristics or law strength. B.2.6 Sampling Before drawing the sample, the researchers decided to discard Montana because there was no distracted driving law in place. In addition, Alaska, Hawaii, Prince Edward Island, and the District of Columbia were discarded because the goal of the study was to provide generaliz- able findings, and each of these jurisdictions had unique features that would make it difficult to generalize any findings to other jurisdictions. The researchers also selected Ontario and Maryland with certainty because these jurisdictions had the strongest laws (highest law scores) in Canada and the U.S., respectively. After these exclusions, a sample of 18 jurisdictions was selected from a pool of 55, with rep- resentation from each cluster. The researchers selected three jurisdictions from the Cluster 1 because of the similarities in the different characteristics across these jurisdictions (primarily Canadian provinces). Five out of 13 eligible jurisdictions in the stronger laws, U.S. states cluster were selected, because the researchers wanted to slightly oversample jurisdictions with strong laws. This left 10 jurisdictions to sample out of the remaining three clusters (32 total jurisdic- tions). The researchers sampled from each cluster proportional to its size in order to obtain the required sample, plus at least one alternate for each cluster. [Sampling was performed using the sample() and sample_frac() functions in R.] The final sample is presented in Table B-1. Figure B-3. Scatterplot of population by law strength, with cluster groupings.

Next: Appendix C - Discussion Guides with Key Stakeholders »
Using Electronic Devices While Driving: Legislation and Enforcement Implications Get This Book
×
MyNAP members save 10% online.
Login or Register to save!
Download Free PDF

Distracted driving is a complex and ever-increasing risk to public safety on roadways. Drivers’ use of electronic devices significantly diverts human attention resources away from the driving task. The enforcement community faces significant challenges as electronic device use has expanded beyond simply texting or talking. Legislation regulating electronic device use while driving is inconsistent in content and implementation.

The TRB Behavioral Traffic Safety Cooperative Research Program's BTSCRP Research Report 1: Using Electronic Devices While Driving: Legislation and Enforcement Implications presents the results of an examination of the current state and provincial legislation on electronic device use while driving; evaluates the benefits and impediments associated with enacting, enforcing, and adjudicating electronic device use; and proposes model legislation and educational materials that can be used by relevant stakeholders to enact a law and educate key individuals on the importance of the law.

Supplemental the report is a presentation for law enforcement.

  1. ×

    Welcome to OpenBook!

    You're looking at OpenBook, NAP.edu's online reading room since 1999. Based on feedback from you, our users, we've made some improvements that make it easier than ever to read thousands of publications on our website.

    Do you want to take a quick tour of the OpenBook's features?

    No Thanks Take a Tour »
  2. ×

    Show this book's table of contents, where you can jump to any chapter by name.

    « Back Next »
  3. ×

    ...or use these buttons to go back to the previous chapter or skip to the next one.

    « Back Next »
  4. ×

    Jump up to the previous page or down to the next one. Also, you can type in a page number and press Enter to go directly to that page in the book.

    « Back Next »
  5. ×

    To search the entire text of this book, type in your search term here and press Enter.

    « Back Next »
  6. ×

    Share a link to this book page on your preferred social network or via email.

    « Back Next »
  7. ×

    View our suggested citation for this chapter.

    « Back Next »
  8. ×

    Ready to take your reading offline? Click here to buy this book in print or download it as a free PDF, if available.

    « Back Next »
Stay Connected!