Impacts of the COVID-19 Pandemic on Pedestrian Push-Button Utilization and Pedestrian Volume Model Accuracy in Utah

Abstract

This work investigated the impacts of COVID-19 on pedestrian behavior, answering two research questions using pedestrian push-button data from Utah traffic signals: How did push-button utilization change during the early pandemic, owing to concerns over disease spread through high-touch surfaces? How did the accuracy of pedestrian volume estimation models (developed pre-COVID based on push-button traffic signal data) change during the early pandemic? To answer these questions, we first recorded videos, counted pedestrians, and collected push-button data from traffic signal controllers at 11 intersections in Utah in 2019 and 2020. We then compared changes in push-button presses per pedestrian (to measure utilization), as well as model prediction errors (to measure accuracy), between the two years. Our first hypothesis of decreased push-button utilization was partially supported. The changes in utilization at most (seven) signals were not statistically significant; yet, the aggregate results (using 10 of 11 signals) saw a decrease from 2.1 to 1.5 presses per person. Our second hypothesis of no degradation of model accuracy was supported. There was no statistically significant change in accuracy when aggregating across nine signals, and the models were actually more accurate in 2020 for the other two signals. Overall, we concluded that COVID-19 did not significantly deter people from using push-buttons at most signals in Utah, and that the pedestrian volume estimation methods developed in 2019 probably do not need to be recalibrated to work for COVID conditions. This information may be useful for public health actions, signal operations, and pedestrian planning.

The outbreak of the coronavirus disease COVID-19 first started in Wuhan, China, in December 2019. In March 2020, the World Health Organization announced COVID-19 as a global pandemic after it spread rapidly all over the world, including in the United States. To slow the spread of the virus, different countries took various public health actions. Many states and communities in the United States implemented stay-at-home orders or recommendations, schools and restaurants were closed or constrained, working from home became the norm in some fields, and many public events and large gatherings were canceled. Mandates or recommendations also often included social distancing (6 ft or 2 m), face coverings (masks), and frequent hand washing and surface cleaning. The COVID-19 pandemic led to major changes in travel patterns around the world and across multiple modes ( 1 – 5 ). Travel restrictions also appear to have resulted in significant changes in walking activity in Utah ( 6 ). There is a need and desire to be able to accurately monitor traffic patterns, including pedestrian activity, to inform agencies in their traffic management and other operational and planning decisions. It is also of scientific interest to know how pedestrian behavior changed in response to COVID-19 concerns. In this paper, we focus on one specific area of COVID-19 influences on pedestrian behavior: pedestrians’ utilization of push-buttons at traffic signals (i.e., button-press behavior), and the corresponding impacts on the accuracy of pedestrian volume estimation from push-button traffic signal data.

Background and Research Questions

Early in the pandemic, fears over virus spread and contracting COVID-19 by interacting with high-touch public surfaces—including pedestrian push-buttons at traffic signals—led some transportation agencies to eliminate the need to press the push-button to get a walk indication ( 7 ), a signal timing technique called pedestrian recall. For example, Salt Lake City and the Utah Department of Transportation (UDOT) placed several dozen signals in downtown Salt Lake City on pedestrian recall from April through June 2020 ( 8 ). These actions were in response to fears which may manifest in different pedestrian behaviors when interacting with traffic signals that have not been switched to pedestrian recall. Specifically, people may have been less willing to press the pedestrian push-button in times and locations of community spread of COVID-19. However, these actions were taken without knowing whether pedestrian push-button utilization or button-press behavior had actually changed. In fact, later in the pandemic, studies showed that infected surfaces (especially those exposed to sunlight) were not a leading cause of COVID-19 spread.

• Research Question 1: How did the utilization of pedestrian push-buttons at traffic signals change during the early months of the COVID-19 pandemic?

• Hypothesis 1: Pedestrians were slightly less likely to press pedestrian push-buttons, owing to concerns about COVID-19.

Recent research in Utah has investigated the use of pedestrian push-button data from traffic signals for pedestrian traffic monitoring and pedestrian volume estimation ( 9 ). UDOT has been a leader in developing the automated traffic signal performance measures (ATSPM) system ( 10 , 11 ), which allows access to high-resolution traffic signal controller event logs ( 12 ), including information about pedestrian push-button presses ( 13 ). Work by Singleton and Runa in 2019 recorded more than 22,000 crossing-hours of video and collected observed counts of over 170,000 pedestrians at 90 signals throughout Utah ( 14 ). Comparisons of pedestrian counts to pedestrian signal data (including pedestrian actuations and time-filtered pedestrian push-button presses) used simple quadratic or piecewise linear regression models, applied to different situations (e.g., pedestrian recall or not, short versus long cycle lengths). Overall, the model-estimated pedestrian crossing volumes had a low error (±3.0 pedestrians per hour) and were strongly correlated (0.84) with observed volumes ( 14 ).

The application of these models allows for the estimation of pedestrian volumes (directly from traffic signal data) at around 1,500 signals throughout Utah, providing information that is useful for pedestrian planning and safety analysis efforts ( 15 , 16 ). However, these models rely on empirically derived relationships from 2019 about pedestrian behavior at signals: specifically, the utilization of pedestrian push-buttons. Any change in pedestrian push-button press behavior resulting from COVID-19 might yield less accurate volume estimates and require a recalibration of these pedestrian volume estimation methods.

• Research Question 2: How did the accuracy of pedestrian volume estimation models based on traffic signal data (developed pre-COVID) change during the early months of the COVID-19 pandemic?

• Hypothesis 2: Pedestrian push-button utilization (button-press behavior) did not change enough to degrade the accuracy of the pedestrian volume estimation models.

Data and Methods

To answer our research questions, we had to first collect pedestrian data from recorded videos, then assemble pedestrian push-button data from traffic signals, and finally analyze push-button utilization and the accuracy of the pedestrian volume estimation models.

Pedestrian Data Collection

Observed pedestrian data were obtained from recorded videos at different signals in Utah. In 2019, we collected pedestrian volume data at 90 signals ( 9 ). For 2020, we collected data at 11 signals—see Figure 1 and Table 1—where we had the most 2019 data, to increase the likelihood that any differences in pedestrian behavior were not because of random chance. These locations also captured a range of estimated traffic volumes—annual average daily pedestrian crossing volumes and entering motor vehicle traffic volumes—as well as different regions and urban contexts. (None of these locations were placed under continuous pedestrian recall by Salt Lake City or UDOT.) For each location in each year, we used UDOT traffic cameras to record more than 200 crossing-hours of video. We then watched the videos and tabulated the number of pedestrians (walking, running, on a skateboard, or in a wheelchair) using each crossing in each hour.

Figure 1.

Map of locations with data collected in 2019 and 2020.

Table 1.

Details About Data Collection in 2019 and 2020

Signal	Location	Crossing-hours		Months		Crossing AADP a		Entering AADT b
		2019	2020	2019	2020	2019	2020	2019	2020
1021	1300 S & 300 W, Salt Lake City	489	310	03	06, 07	1,871	1,578	37,100	33,200
1094	S Temple & 700 E, Salt Lake City	294	212	04	06, 07	305	196	41,800	37,300
1229	2100 S & 1300 E, Salt Lake City	274	253	03	06, 07	1,489	1,274	62,800	56,100
4130	6200 S & Margray Dr, Taylorsville	471	215	02, 07, 08, 11	05	57	86	24,800	22,200
5108	Antelope Dr & Hill Field Rd, Layton	416	326	03, 10	06	470	318	44,500	39,800
5306	400 N & Main St, Logan	498	251	01, 02, 09, 11	04, 05	291	225	50,900	44,300
5349	2600 S & US-89, Bountiful	508	288	07, 11	05	380	277	47,300	42,300
6307	800 N & Palisades Dr, Orem	281	353	07	06, 07	83	137	37,200	32,400
7184	900 S & 700 E, Salt Lake City	558	694	01, 02, 08, 11	04, 05, 06	822	790	54,700	44,200
8119	St. George Blvd & 400 E, St. George	730	368	01, 02, 08, 11	04, 05	102	84	27,100	26,900
8302	Center St & Main St, Moab	789	358	02, 03, 06, 07, 10	05	6,146	5,163	18,500	17,600

Note: AADP = annual average daily pedestrians; AADT = annual average daily (entering motor vehicle) traffic.

^aEstimated AADP values were calculated by applying the modeling methods developed by Singleton and Runa ( 14 ) to a full year of pedestrian push-button data. Pedestrian crossing volumes across the 11 signals decreased by an average 16% from 2019 to 2020.

^bEstimated AADT values were obtained from products of Utah Department of Transportation’s traffic monitoring program. Motor vehicle traffic volumes across the 11 signals decreased by an average 11% from 2019 to 2020.

Pedestrian Push-Button Data Assembly

Time-stamped pedestrian push-button presses are recorded in high-resolution traffic signal controller log data ( 12 ). We used UDOT’s ATSPM system ( 10 ) to obtain push-button data for the time periods corresponding to the videos at each signal. Based on earlier work by Singleton et al. ( 9 ), we then calculated—for each hour and pedestrian phase (crossing)—several different measures of pedestrian traffic signal activity.

• Pedestrian push-button presses: The most direct measure of pedestrian push-button utilization or button-press behavior is Event Code 90 (“pedestrian detector on”). This occurs whenever a pedestrian push-button is activated (pressed), which could happen multiple times per cycle.

• “Unique” pedestrian push-button presses: Because one person may press the push-button multiple times in quick succession, we used time filters to remove successive push-button presses within a certain time interval. Testing indicated that a 15-s filter was the best fit to the observed volume data ( 14 ).

• Pedestrian actuations: Other research ( 17 ) has used actuations rather than push-button presses to correlate pedestrian volumes. An actuation occurs the first time a push-button is pressed before being served, so usually just once per cycle. This measure was the best predictor of pedestrian volumes for crossings when on pedestrian recall ( 14 ).

Analysis of Changes in Pedestrian Push-Button Utilization

To determine whether pedestrian push-button utilization changed during the COVID-19 pandemic, we compared the ratio of pedestrian push-button presses with pedestrian crossing volumes, which we defined as the push-button use rate or utilization (presses per person). Our use of rates to measure pedestrian push-button utilization behavior was admittedly simplistic, but it was appropriate for our hourly data collection method and it provided a first-stage look at COVID-related changes. Also, our second analysis (model prediction accuracy) better addressed the (nonlinear) relationship between push-button presses and pedestrian volumes.

To statistically analyze changes in utilization between the two years, we estimated a fixed-effects multilevel linear regression model (hours, i, at level one; signals, j, at level two) with no intercept (Y_ij = β_j × X_ij), predicting hourly pedestrian volumes (Y_ij) as a function of pedestrian push-button presses (X_ij) across all crossings/phases, where the slopes (β_j) are fixed parameters that vary across signals, j. Note that the slope (β) is the average number of pedestrians per push-button press, whereas the inverse slope (1/β) is the average number of push-button presses per pedestrian, our utilization rate. We allowed β_j to be different for each signal in each year (Y_ij = β_j,2019 × X_ij,2019 + β_j,2020 × X_ij,2020); also, by dummy coding for 2020 (Y_ij = β_j × X_ij + β_j,Δ2020 ×X_ij,2020), null hypothesis tests of β_j,Δ2020 provided statistical significance of the change in slope at each signal from 2019 to 2020. Specifically, a decrease in the utilization rate (an increase in the slope) would suggest that people may have been avoiding push-buttons out of fears of contracting COVID-19.

Analysis of Changes in the Accuracy of Pedestrian Crossing Volume Estimates

To assess any changes in the accuracy of the pedestrian volume estimation models, we compared the model prediction errors between the two years. First, we applied the models developed by Singleton and Runa to estimate hourly pedestrian crossing volumes from traffic signal and pedestrian push-button data for both 2019 and 2020, and calculated the prediction errors (observed minus estimated) ( 14 ). As previously mentioned, Singleton and Runa developed five piecewise linear or quadratic linear regression models for different situations: pedestrian hybrid beacon signals, crossings with pedestrian recall at high or low volume signals, and crossings with pedestrian recall at signals with short or long cycle lengths ( 14 ). To aid with application, the models used just one independent variable: whichever pedestrian signal activity measure best fit the data (unique push-button presses or pedestrian actuations).

Then, for each signal, we performed a Welch’s (unequal variances independent samples) t-test on the model prediction errors for 2019 versus 2020. Specifically, a significant difference (especially an increase) in prediction error would suggest that the pedestrian volume estimation models may need to be adjusted to remain accurate during the COVID-19 pandemic.

Results and Discussion

Analysis of Changes in Pedestrian Push-Button Utilization

Table 2 summarizes the findings of the first analysis of changes in pedestrian push-button utilization. For most signals (7 out of 11 signals), the change in the utilization rate (push-button presses per person) from 2019 to 2020 was not statistically significant (change in slope: p > 0.10). Two other signals (1094 and 7184) had significant decreases in push-button utilization, whereas the utilization rate increased significantly at the final two signals (5108 and 8302). Aggregating across all 11 signals, the utilization rate increased from 1.08 in 2019 to 1.40 in 2020, which would suggest that people were pressing push-buttons more often at signals during the early months of the COVID-19 pandemic. However, aggregate results appeared to have been greatly influenced by the noticeably different results at signal 8302; when removing that signal, the new aggregate results (for 10 signals) indicated a statistically significant reduction in push-button utilization from 2.11 to 1.47 presses per pedestrian.

Table 2.

Pedestrians Push-Button Utilization, 2019 Versus 2020, by Signal and Overall

Signal	Push-button presses per pedestrian (utilization rate)			Pedestrians per push-button press (β from model)
	2019	2020	Δ	2019	2020	Δ	Δ SE	Δp
1021	1.18	1.10	NS	0.85	0.91	0.06	0.04	0.104
1094	3.62	2.08	−	0.28	0.48	0.21	0.10	0.043
1229	1.26	1.35	NS	0.80	0.74	−0.06	0.08	0.511
4130	3.13	2.83	NS	0.32	0.35	0.03	0.31	0.912
5108	2.44	3.65	+	0.41	0.27	−0.14	0.06	0.032
5306	3.34	2.50	NS	0.30	0.40	0.10	0.09	0.259
5349	3.19	2.56	NS	0.31	0.39	0.08	0.07	0.286
6307	5.84	4.10	NS	0.17	0.24	0.07	0.10	0.479
7184	2.16	1.38	−	0.46	0.72	0.26	0.03	<0.001
8119	6.05	3.24	NS	0.17	0.31	0.14	0.16	0.381
8302	0.50	0.63	+	2.01	1.58	−0.43	0.07	<0.001
All 11 signals	1.08	1.40	+	0.93	0.71	−0.21	0.02	<0.001
10 signals (not 8302)	2.11	1.47	−	0.47	0.68	0.20	0.01	<0.001
9 signals (not 7184, 8302)	2.10	1.64	−	0.48	0.61	0.13	0.01	<0.001

Note: + = significant positive change; – = significant negative change; NS = change not significant; SE = standard error.

Figure 2—showing plots of the relationships between pedestrians and push-button utilization at the four signals with the highest pedestrian activity in our study—illustrates these varied findings. Signal 1021, located in a transit-accessible area of Salt Lake City with numerous big-box stores (i.e., physically large retail establishments), experienced a small (but not statistically significant) decrease in push-button utilization (from 1.18 to 1.10 push-button presses per pedestrian). Signal 1229, located in a neighborhood commercial district in Salt Lake City, saw a small (but not statistically significant) increase in the utilization rate (from 1.26 to 1.35). In both cases, 2020 observations generally fell in the same range as 2019 observations.

Figure 2.

Pedestrian push-button use, 2019 versus 2020, for signals 1021, 1229, 7184, and 8302. Each data point is one crossing observed for 1 h on a given day, either in 2019 (empty red downward-pointing triangles) or in 2020 (filled blue upward-pointing triangles).

Signal 7184, located in a residential neighborhood of Salt Lake City near a large park, saw a significant decrease in pedestrian push-button utilization that is also apparent from the increased slope in the figure. Most pedestrians crossing at this intersection were observed going to/from the park, so it may be that people who were walking for recreation (rather than for transportation purposes) during the early pandemic were more cautious and concerned about COVID-19 spread from touching pedestrian push-buttons.

Signal 8302, located in downtown Moab in eastern Utah, was one of two signals to see a significant increase in utilization rate (push-button presses per person) from 2019 to 2020. Owing to its proximity to popular Arches and Canyonlands National Parks, Moab is a tourist-oriented small city that attracts many visitors annually, making signal 8302 one of the highest pedestrian volume intersections in Utah ( 14 ). The COVID-19 pandemic hit Moab hard after the National Park Service closed the parks to all visitors on March 28, 2020. Thus, the most noticeable difference for this signal in Figure 2 is that 2020 saw dramatically fewer pedestrians (during the months studied) than in 2019. We suspect that the “increase” in pedestrian push-button utilization found in our analysis (for this signal and overall) is more the result of lighter crowds and smaller pedestrian group sizes (perhaps because of social distancing) than any major change in pedestrians’ willingness to press push-buttons resulting from COVID-19 concerns. In fact, results by Singleton et al. suggest that the relationship between pedestrians and push-button presses is nonlinear: the slope increases (more pedestrians per push-button press) as push-button activity (per hour) increases ( 9 ). This also highlights a limitation of our linear analysis method: if overall activity decreases (as it did at signal 8302), then the overall slope will also decrease. Thus, we also performed a second analysis (described in the following subsection) that accounted for nonlinear relationships between pedestrian volumes and push-button utilization.

Analysis of Changes in the Accuracy of Pedestrian Crossing Volume Estimates

Table 3 shows the results of the analysis of changes in accuracy of the pedestrian volume estimation models, including the mean and standard deviation of the model prediction errors (observed minus estimated) in 2019 and 2020, and the results from the Welch’s t-tests on those errors. Most signals (9 of 11) showed no statistically significant difference (p > 0.05) in the average error (pedestrians per hour) between the 2 years. Furthermore, the small changes in the mean errors did not show a consistent trend for all signals: some moved closer to zero (four) or farther from zero (seven), and some became more negative (five) or more positive (six). Aggregating across all 11 signals, the average error actually became less negative from 2019 to 2020 (−1.38 to −0.68); however, this improvement in accuracy was completely driven by significant differences at two signals (7184, 8302), as discussed later. The change (or lack thereof) in error does not seem to have been caused by more extreme but counteracting (i.e., larger positive and negative) errors, because the overall standard deviation of the prediction errors was smaller in 2020 than in 2019.

Table 3.

Pedestrian Volume Model Prediction Errors, 2019 Versus 2020, by Signal

Signal	2019		2020		Welch’s t-test
	Mean	SD	Mean	SD	t	df	p
1021	−5.02	11.58	−5.23	12.12	0.24	635	0.809
1094	−0.08	5.59	0.36	2.60	−1.17	440	0.241
1229	−4.00	10.00	−3.04	5.33	−1.38	423	0.168
4130	−0.19	1.26	−0.40	1.51	1.81	356	0.071
5108	−0.54	5.61	−0.86	2.06	1.10	549	0.272
5306	−0.16	2.61	−0.32	2.46	0.79	529	0.430
5349	−1.05	2.85	−0.80	1.66	−1.55	793	0.122
6307	−0.50	2.47	−0.85	3.51	1.49	623	0.137
7184	−1.46	5.78	0.20	7.10	−4.55	1,250	<0.001
8119	0.06	1.71	0.25	1.60	−1.78	781	0.076
8302	−2.42	26.51	1.69	9.89	−3.81	1,114	<0.001
All 11 signals	−1.38	11.59	−0.68	6.31	−3.68	8,553	<0.001
10 signals (not 8302)	−1.20	5.91	−0.94	5.72	−1.94	7,170	0.052
9 signals (not 7184, 8302)	−1.16	5.93	−1.24	5.25	0.60	5,951	0.550

Note: SD = standard deviation.

Results suggest that, in general, the models were still producing similarly accurate (if not more accurate) estimates of pedestrian crossing volumes during the COVID-19 pandemic. We suspect this may be owing to the modest changes in pedestrian push-button utilization behavior identified in the first analysis: no statistically significant change for 7 of 11 signals. Also, it could be that the models’ methods of time-filtering the push-button data (the 15-s filter for “unique” presses and the use of actuations in some situations) were robust to COVID-induced changes in push-button utilization.

As noted, we did find statistically significant differences in the prediction errors for two signals: 7184 and 8302. Nevertheless, we should also note that the absolute value of the mean errors for these two signals in 2020 was smaller than the absolute value of the mean errors in 2019, indicating that the model was actually more accurate (on average) during the COVID-19 pandemic. We have a few potential explanations for why these signals in particular saw changes and why the accuracy of the models might have increased.

As previously mentioned, signal 8302 in Moab saw greatly reduced pedestrian activity during the early months of the COVID-19 pandemic. Smaller crowds and pedestrian group sizes (and lower activity overall) provides fewer opportunities for large prediction errors, and the models tend to be more accurate (smaller magnitude errors) for lower-activity signals ( 14 ).

In contrast, this explanation cannot account for the improved accuracy at signal 7184, since this location—near a popular large park in Salt Lake City—saw increased pedestrian activity early in the pandemic, especially on days with pleasant weather. One potential explanation unique to this location is that there was an open-streets installation on 900 S ( 18 ) during the study period that converted the outer travel lanes to space for active transportation, including a pop-up bike lane in the west-bound direction (an east-bound bike lane already existed). In 2019, we noticed that many people bicycling through this intersection used the crosswalks and push-buttons; thus, these cyclists added push-button presses but were not counted as pedestrians. Therefore, in 2020 perhaps there were fewer people bicycling on the sidewalk and “contaminating” the push-button counts, yielding more accurate model estimates of pedestrian volumes.

Conclusion

The first objective (Research Question 1) of this research was to examine how the utilization of pedestrian push-buttons at traffic signals changed during the early months of the COVID-19 pandemic. We expected push-button utilization to have decreased slightly owing to concerns about using high-touch surfaces. At 7 of the 11 signals, the change in utilization rate (push-button presses per pedestrian) was not statistically significant. Push-button utilization decreased at two signals (1094 and 7184) but increased at two others (5108, 8302). Aggregated across all 11 signals, push-button utilization (presses per person) actually increased slightly in 2020, yet this was mostly driven by unique changes at one signal: 8302. Given this signal’s location in a tourist town with high pedestrian volumes, we suspect this is the result of reductions in pedestrian group sizes. If there are fewer crowds or people travel in smaller groups (because of social distancing), then we would expect the observed increase in push-button presses per person (decrease in pedestrians per push-button press). Excluding signal 8302, the new aggregate results (for 10 signals) showed a statistically significant decrease in push-button utilization, from 2.1 to 1.5 presses per person. Thus, Hypothesis 1 was partially supported.

Our second objective (Research Question 2) focused on the accuracy of pedestrian volume estimation models, based on traffic signal (push-button) data and developed in 2019, during the early months of the COVID-19 pandemic. We expected that button-press behavior had not changed enough to degrade the accuracy of the models (especially considering the 15-s filtering of extraneous push-button presses). Indeed, 9 of the 11 signals saw no statistically significant change in accuracy between 2019 and 2020, whereas two signals (7184 and 8302) actually had more accurate pedestrian volume estimates in 2020 than in 2019. Aggregated across all 11 signals, the average model prediction error decreased from −1.4 to −0.7; this may have been the result of smaller crowds and pedestrian group sizes at signal 8302. Excluding signals 7184 and 8302, the new aggregate results showed effectively no change in average error (−1.16 to −1.24). Thus, Hypothesis 2 was supported. This is not surprising, given the results of the first analysis, but it is still “good news” that the pedestrian volume estimation models seem to be similarly (if not more) accurate and do not need to be recalibrated to work during COVID conditions.

Overall, this research has provided insights into the impacts of the COVID-19 pandemic on walking and pedestrian behavior, specifically in relation to push-button utilization at traffic signals. Despite this narrow focus, the research addressed an important public health and signal operations question, indicating that, overall, people in Utah were not significantly deterred from using pedestrian push-buttons from fears of contracting/spreading COVID-19. More recent understanding of COVID transmission sources (i.e., more from air than from surfaces) suggests that even modest changes in button-press behavior may not persist postpandemic. Also, by investigating the accuracy of pedestrian volume estimation models based on traffic signal data, this research also addressed an important question for planning. Our results suggest that pedestrian volume estimates obtained during the COVID-19 pandemic (using models calibrated on prepandemic data) were no less accurate and may even be more accurate. Greater model accuracy could be the result of reduced pedestrian activity overall (by an estimated 16%; see Table 1) as well as smaller pedestrian group sizes (because of social distancing), both of which reduce large prediction errors. This indicates that the models can continue to be used. Pedestrian volume estimates from traffic signal data have been used for various planning ( 15 ) and safety analysis ( 16 ) purposes.

Still, this study was not without limitations that could be addressed through future work. There may be other factors for which we are not controlling that might explain differences in push-button utilization. Reduced motor vehicle traffic volumes (by an estimated 11%; see Table 1) may have encouraged/allowed some pedestrians to cross against traffic or in midblock locations without pressing the push-button. Our data collection covered different months in different years, and button-press behavior may vary over the year; however, studying a seasonal effect is challenging, because utilization or accuracy differences may be related to volume differences, since we know that pedestrian volumes are affected by weather ( 19 ) and the models are more accurate for lower-volume locations ( 14 ). Although we studied 11 signals of different types and in different locations, certain locations could have seen larger or different changes in pedestrian push-button use behavior. Locations with different compositions of users or travel purposes (e.g., walking for recreation versus transportation) might have seen different results. Even though the models remained similarly accurate during the pandemic, relationships may change or not be applicable outside of Utah. Thus, there is a continued need to validate the models with new data on a periodic basis (every couple of years). Also, our method of data collection and analysis limited us to hourly and more aggregate observations. Studying other pedestrian behaviors at traffic signals that may be of interest during the COVID-19 pandemic—group sizes, social distancing, walking speed, signal violations, and so forth—would require more fine-grained data collection from videos. We encourage such research to continue to advance our limited understanding of how the COVID-19 pandemic affected pedestrian travel.