Human Factors Assessment of Pedestrian Roadway Crossing Behavior

Full Dataset

Overview of Crossings

Figure 61 through figure 66 depict the breakdown of the total observed crossings across all 20 locations. Because all of the crossings are combined across locations, the proportions of each type of crossing may not be typical of all locations. Because Location 3 was the only location that included crossings at an unmarked intersection, the sum values are not the same in every figure.

Figure 61. Chart. Distribution of all crossings observed across the 20 different locations by the area in which they took place.

Figure 62. Chart. Distribution of crossings observed across all 20 locations in the marked intersections by circumstances of the crossing.

Figure 63. Chart. Distribution of crossings observed across all 20 locations in the unmarked non-intersections by the circumstances of the crossing.

Figure 64. Chart. Distribution of crossings observed across all 20 locations for all rule-breaking crossings by the circumstances of the crossing (including those in the unmarked non-intersections).

Figure 65. Chart. Distribution of crossings observed across all 20 locations for crossings made entirely during the walk light phase in the marked intersections, by the circumstances of the crossing.

Figure 66. Chart. Distribution of crossings observed across all 20 locations for rule-breaking crossings in marked intersections (i.e., crossings that took place at least partially during the don’t walk light phase) by the circumstances of the crossing.

Prediction Model—Crossing Area

Using data from all 20 locations, a model was successfully created in SAS^® to predict whether pedestrians would cross at the marked intersection (and not the unmarked non-intersections). There were 68,056 pedestrian crossings among the 20 locations, 62,838 of which occurred at marked intersections. (Note that crossings at the unmarked intersection at Location 3 were not included in the model.) In other words, approximately 92 percent of the crossings occurred at the marked intersections.

First, a factor analysis with an orthogonal varimax rotation was used to describe the underlying relationships among the 16 environmental variables (see table 46). Based on the greater-than-one rule for the eigenvalues, five factors were kept for the rotation. The factor loadings are included. Combined, these factors accounted for approximately 74 percent of the standardized variance in the data.

Table 46. Environmental factors and their labels used to calculate the crossing location prediction model.

Note: Values in parentheses are the values assigned to categorical variables.

Table 47. Factor loadings for the 16 environmental variables.

… Indicates a loading of 60 or greater in absolute value.

Each of the five factors is assigned a descriptive label based on its correlations with relative variables:

·         Factor 1, Travel Pace and Phasing, is negatively correlated with the length of the walk phase, positively correlated with the length of the do not walk phase, and positively correlated with the travel pace.

·         Factor 2, Traffic Throughput, is positively correlated with the AADT, traffic directionality (one- or two-way street), and the curb-to-curb distance (i.e., the width of the street).

·         Factor 3, Distance to Safety, is positively correlated with the distance to the next marked crosswalk, the presence and type of median, and the presence of cross streets between marked crosswalks.

·         Factor 4, External Objects (Barriers/Vehicles) in Center of Road, is positively correlated with the presence of physical barriers that might prevent a pedestrian from crossing the roadway and the presence of a center turning lane.

·         Factor 5, External Objects (Vehicles) on Sides of Road, is positively correlated with the presence of parking along the roadway, the presence of a right turn only lane, and whether or not the far marked crosswalk is light controlled.

Scores for each of the five factors were generated using the standardized scoring coefficients listed in table 48. These coefficients can only be used with the standardized data and not with the raw, observed values. Each of the five factor scores is graphed against the percentage of intersection crossings for each location in figure 67.

Table 48. Standardized scoring coefficients for the five rotated factors.

Notes: The dots remain in consistent locations from box to box.

solid line = regression line, shaded area = 95-percent confidence limits of the regression line).

Figure 67. Chart. Scatterplot of each location’s score for each of the five factors against the percentage of intersection crossings at that location.

Next, the scores for the five factors were regressed on the probability of an intersection crossing through logistic regression using forward selection. The model iterated three times after the entrance of an intercept term. The scores for Factor 1, (1) = 176.6, p < 0.01, Factor 2, (1) = 41.6, p < 0.01, and Factor 3, (1) = 402.3, p < 0.01, were statistically significant. The resultant model is shown in figure 68 where π represents the probability of crossing in the marked intersection and x is the vector of factor scores for a given location. When using a logit (log odds) model, represents the odds of success when all predictors equal zero, and represents the multiplicative effect of a one-unit increase in the corresponding x on the odds of success, when the other predictors are held fixed.

Figure 68. Equation. The logit of crossing at the intersection.

A good probability cutoff point for this data is 0.9, meaning that any future predictions using the model that are 0.9 or greater should be deemed an intersection crossing. With this cutoff, the model can successfully predict approximately 90 percent of the crossings overall. The resultant sensitivity and specificity are about 96 percent and 10 percent, respectively. Sensitivity measures how many events (intersection crossings) were successfully predicted, and specificity measures how many non-events (non-intersection crossings) were successfully predicted. The predicted probabilities of crossing at the intersection for each location are shown in table 49.

Table 49. Probabilities of correctly predicting a crossing at the marked intersection or the unmarked non-intersection by location and the corresponding upper and lower 95-percent confidence limits.

Prediction Model—Walk Versus All Other Crossings

A second model was developed to predict crossings that occurred entirely during the walk phase in the marked intersection. Crossings were divided into two categories—crossings that occurred within the marked intersection entirely during the walk light phase and all other crossings (e.g., unmarked non-intersection, marked intersection during the don’t walk light phase, etc.). Using this classification system, there were 70,378 pedestrian crossings among the 20 locations, 55,040 of which occurred entirely within the walk phase in the marked intersection. In other words, approximately 78 percent of the crossings were rule-following crossings. Using the factor analysis and standardized scoring coefficients (table 48), each of the five factor scores is graphed against the percentage of rule-following crossings for each location in figure 69.

Notes: The dots remain in consistent locations from box to box.

solid line = regression line, shaded area = 95-percent confidence limits of the regression line).

Figure 69. Chart. Scatterplot of each location’s score for each of the five factors against the percentage of intersection crossings at that location.

Next, the scores for the five factors were regressed on the probability of a rule-following crossing through logistic regression using forward selection (see figure 70). The model iterated five times after the entrance of an intercept term. The scores for Factor 1, (1) = 607.6, p < 0.01, Factor 2, (1) = 1224.5, p < 0.01, Factor 3, (1) = 187.4, Factor 4, (1) = 5783.2, p < 0.01, and Factor 5, (1) = 388.4, p < 0.01, were all statistically significant.

Figure 70. Equation. The logit of crossing at the intersection during the walk phase.

A good probability cutoff point for this data is 0.6, meaning that any future predictions using the model that are 0.6 or greater should be deemed a rule-following crossing. With this cutoff, the model can successfully predict approximately 79 percent of the crossings overall. The resultant sensitivity and specificity are about 94 percent and 23 percent, respectively. The predicted probabilities of crossing during the walk phase at the intersection for each location are shown in table 50.

Table 50. Probabilities of correctly predicting a crossing at the marked intersection entirely during the walk phase by location and the corresponding upper and lower 95‑percent confidence limits.