ate it through proactive planning. Mobility monitoring allows trans-portation professionals to track changes from year to year, preparefor future growth, and implement the most efficient congestion mit-igation measures. Direct mobility monitoring on a regular basis canbe time consuming and expensive; therefore, models can be used toestimate parameters associated with congestion, including volume,travel time, and delay.

It is important to monitor and mitigate arterial congestion, whichis growing as freeways become congested. This is especially truein small to medium-sized communities, in which arterials play vitalmobility roles and financial resources are very limited. Travel timeis a vital statistic for transportation system performance that iseasily communicated to both technical and nontechnical audiences.Common planning applications such as mobility monitoring canbenefit from methods for estimating the travel time index (TTI),which is the ratio of the average peak period travel time to the traveltime at free-flow conditions. In relating the actual (peak) travel timeto an off-peak travel time, the TTI has proven valuable for mobilityanalysis and quantifying transportation system performance (1).

This paper describes a recent research effort with the objectiveof developing and validating models to estimate TTI for suburbanarterial corridors. The models presented here are a function of datainputs that are generally available or easy to obtain for practitionersinterested in estimating TTI.

EXISTING MODELS

Existing published models for arterial studies incorporate manydesirable features, but all have some limitations for transportationplanning purposes. This section evaluates models typically usedby transportation professionals for estimating speed or travel timealong arterials.

Mobility in the arterial environment is best understood whenconveyed in terms of arterial travel time. Previous research forNCHRP states, “Measures related to travel time and speed are the most flexible and useful for a wide range of analyses. . . .Travel time measures are consistent, address transportation andland use systems, and are responsive to concerns of residents,businesses, and travelers” (2, p. 2). The interrupted arterial envi-ronment poses many obstacles and possesses other parameters notseen on uninterrupted facilities, so travel time models developedfor freeways are not very useful for applications related to arterials.

Developing and Applying Models forEstimating Arterial Corridor Travel TimeIndex for Transportation Planning in Smallto Medium-Sized Communities

William L. Eisele, Yunlong Zhang, Eun Sug Park, Yanru Zhang, and Rachael Stensrud

81

This paper describes research with the objective of developing and val-idating a corridor arterial model to estimate the travel time index (TTI)for mobility analysis. The TTI is the ratio of the travel rate (minutes permile) during the peak period to the travel rate (minutes per mile) dur-ing the off-peak period. The models are most useful for sketch planningpurposes in areas where monitoring infrastructure is not in place. Prac-titioners in small and medium-sized communities, particularly those onthe developing fringe of such communities, will benefit from the use of the models for tracking mobility along roadways of interest and pri-oritizing roadway improvements. This paper describes the developmentand validation of two models to assist transportation professionals inestimating TTI in the arterial environment during light and moderatecongestion conditions. A 2.6-mi major arterial corridor in College Sta-tion, Texas, was used to develop the models, and three study corridorsin Virginia were used to validate the models. To address limitations ofexisting models, the models presented in this paper (a) consider drive-way density, (b) are corridor-based, (c) are a function of generally avail-able or easy-to-obtain independent variables, (d) are calibrated andvalidated with extensive field data, and (e) explain a relatively highdegree of variability. The use of TTI also makes the models more trans-ferable. For sketch-planning applications, the models are a function ofrelatively available or easy-to-estimate data including traffic volume,driveway density, signal green time relative to the cycle time, and signalcoordination conditions.

Traffic congestion is a growing problem that many communitiesthroughout the United States, both large and small, experience daily.According to the 2009 Urban Mobility Report, congestion withinthe United States in 2007 was estimated to cause 4.16 billion hoursof travel delay (1). Congestion hinders the economy, and it is imper-ative that it be regularly monitored and that steps be taken to allevi-

W. L. Eisele, and E. S. Park, Texas Transportation Institute, Texas A&M Univer-sity, 3135 TAMU, College Station, TX 77843-3135. Yunlong Zhang and YanruZhang, Zachry Department of Civil Engineering, Texas A&M University, MS 3136,College Station, TX 77843-3136. R. Stensrud, Zachry Construction Corporation,12625 Wetmore Road, Suite 301, San Antonio, TX 78247-3624. Correspondingauthor: W. L. Eisele, bill-eisele@tamu.edu.

Transportation Research Record: Journal of the Transportation Research Board,No. 2244, Transportation Research Board of the National Academies, Washington,D.C., 2011, pp. 81–90.DOI: 10.3141/2244-11

In recent years, various arterial travel time models have beendeveloped, but they often use the equations in the Highway Capac-ity Manual (HCM) and have similar numerous inputs (3). A modelto estimate arterial travel time is desired that requires limited datacollection while remaining useful to transportation engineers andplanners.

The HCM is a common source transportation professionals refer-ence for estimating travel time on an arterial. The 2000 HCM usesstreet classification, segment length, and free-flow speed to deter-mine the running time on an urban street (3). The arterial travel timeis estimated using the running time combined with the three typesof signal delay: uniform delay, incremental delay, and initial queuedelay. During oversaturated conditions, the high variability of traveltime makes deterministic HCM results questionable. In addition, theinformation needed to calculate these delay values can be difficultto locate or expensive to collect in many situations. The HCM modelis an operational model with extensive input requirements. Thevariables that are needed to calculate delay using the HCM modelinclude the progression adjustment factor, volume, road capacity,capacity and green time for each lane group, cycle length, adjust-ment factors for actuated control and filtering or metering, initialqueue, and duration of analysis period and the unmet demand. Addi-tionally, assumptions that are used for the HCM model are not uni-versal for all intersections. Permitted left turns, protected-permittedleft turns, progression, multiple green displays, and protected-per-mitted right turns are just some of the situations where alterationsmust be made for the HCM model to work (4).

In addition, because the input data are link-based, the estimateddelay from the HCM method is link-based in nature, rather than rep-resenting a corridor. Another limitation of the HCM model is that thedelay values increase rapidly as the volume approaches and exceedsthe capacity. Rouphail and colleagues have recognized this problemand written several papers relating to signalized intersections andarterial travel times over the past two decades (5–27 ).

For planning applications, another common place for speed esti-mates is the original Bureau of Public Roads (BPR) Curve that isshown in Equation 1 (2, 28).

where

S0 = free-flow travel speed,S = travel speed at volume v, andc1 = practical capacity ≅ 80% of maximum capacity.

The values of a and b were 0.15 and 4, respectively. Many regionshave modified these coefficients for local conditions. While the sim-plified nature of the BPR function facilitates understanding andapplication, it is also a link-based model and typically only used forlinks of uninterruptible-flow facilities. The model does not inher-ently include variables that consider green time distribution, coordi-nation, or access point density, except to the extent that theseconsiderations can be incorporated into the capacity estimate used bythe practitioner. The arterial model described in this paper to estimatethe TTI directly considers these signal and access characteristics.

Research sponsored by NCHRP developed relationships to esti-mate speed based on signal density and average daily traffic per laneby arterial class (as defined by the 1994 HCM) (2). The models pro-vide a very general guideline for practitioners. The models do notconsider signal coordination or access density considerations. The

SS

avc

b=

+ ⎛⎝⎜

⎞⎠⎟

0

1

1

1( )

82 Transportation Research Record 2244

R2 values for the models were low, varying from 0.07 to 0.35. Theplanning-level model described in this paper considers access den-sity and signal coordination parameters, yet the inputs to the devel-oped model include data that are readily available or easily obtainedby practitioners.

The selection of the most common models here provides a repre-sentative sample of the general limitations of existing models. Thelimitations of existing speed estimation arterial models generallyinclude one or more of the following:

• The model is link based versus arterial (corridor) based. Someexisting models are based on a link or individual intersection insteadof being true arterial models. Generally, link-based models do noteffectively consider signal coordination or other driver interference(e.g., driveway density) that may occur along an arterial corridor.Corridor-based models allow for consideration of driver interference,which is improved with access management treatments.

• The model requires multiple inputs. Some existing modelsrequire multiple inputs that may not be readily available and may becostly to collect.

• The model is not calibrated or validated with field data. Someexisting models have not been adequately calibrated or validatedwith actual field data.

• The model does not explain variability. Some existing mod-els do not have very good estimation ability (e.g., low R2, high rootmean square error) and do not estimate arterial travel time wellunder a variety of traffic volume conditions (e.g., congested anduncongested).

RESEARCH OBJECTIVES

The primary objective of the research upon which this paper is basedis to develop and validate an arterial model for estimating TTI. Inlight of the limitations of current models listed above, the researchdocumented here describes the development of an arterial modelthat is

• Corridor-based with consideration of signal coordination andmotorist interference (e.g., driveway density),

• A function of generally available independent variables forestimating TTI along the corridor,

• Calibrated and validated with extensive field data, and• Able to explain a relatively high degree of variability as mea-

sured by R2 values over relatively low volume, high volume, andextreme (special event) situations.

The resulting models described in this paper provide an estimateof TTI with consideration of driveway density. The models can beeasily implemented by transportation professionals for TTI estima-tion for planning applications. Using an index has several advan-tages, including comparability between long and short arterials andtransferability between different geographic locations or corridors.Comparison of TTI estimates between different corridors can assisttransportation professionals in prioritizing mobility improvements.

STUDY CORRIDORS

Data from four case studies were used to perform this research. Onesite, College Station, Texas, was used to develop the model, andthree additional sites in Virginia were used to validate the transfer-

ability of the model. Table 1 summarizes the key characteristics ofthe sites.

The following sections describe the process of developing themodel using the College Station site, and subsequent sectionsthen describe validation of the resulting models using the threeVirginia sites.

One of the validation study corridors, US-29 (Lee Highway) is inthe Northern Virginia area. An Interstate highway (I-66) crosses thisstudy corridor, and, therefore, it is not a typical site in a small tomedium-sized community. The researchers chose this site becauseit is expected that the congestion level in some small to medium-sized communities will reach the level that is currently experiencedin larger suburban corridors such as this.

DATA COLLECTION AND REDUCTION FOR MODEL DEVELOPMENT SITE

The researchers performed extensive data collection along theCollege Station corridor for model development. Data collectionincluded travel time runs, volume counts, video, and obtainingsignal timings.

Traffic Volume Data

The researchers collected extensive traffic volume data along theCollege Station corridor. With pneumatic tubes, directional traf-fic volumes were collected at 40 locations along University Drive(FM-60). Traffic counts were generally taken between the majorsignalized intersections and cross streets along University Drive, aswell as north and south of the major signalized intersections. Theresearchers collected traffic data from Monday, November 6, 2006,to Monday, November 13, 2006.

The researchers also collected video data along the primary signal-ized intersections along the corridor. Students reduced the data toobtain turning movements and queue length information at the signal-

Eisele, Zhang, Park, Zhang, and Stensrud 83

ized approaches. Data collection by video or digital video recorderwas performed at all but one of the signalized intersections along thecorridor. The researchers performed video data collection from thetop of tall buildings along the corridor or by using video trailers ordigital video recorders connected into the signal systems.

Travel Time Data

The researchers collected travel time data on two normal weekdays:Wednesday, November 8, 2006, and Thursday, November 9, 2006.Travel time data were collected from 7:00 to 10:00 a.m. and from 4:00to 7:00 p.m. on Wednesday and Thursday. Travel time runs were col-lected in both directions, and instrumented test vehicles performedtravel time runs in a circuit in both directions. The researchers usedthe floating car test vehicle driving style and were instructed to floatwith the traffic by attempting to safely pass as many vehicles aspassed them (29). The test vehicles were equipped with Global Posi-tioning System (GPS) instrumentation. The researchers also includedtravel time runs on Saturday, November 11, 2006, from 10:00 a.m. to1:00 p.m. before a Texas A&M University home football game. Thisprovided an opportunity to obtain travel time data during specialevent traffic conditions.

The researchers used extensive data collection and reduction pro-cedures that are documented elsewhere in a report of a study that per-formed GPS data collection (30). The interested reader is encouragedto review this other study for detailed data collection and reductionprocesses, including a guide used for reducing travel time run data.

After quality control was performed and any runs with errorswere cleaned out, there were 283 runs in the eastbound directionand 280 runs in the westbound direction. After all of the runs werereduced, headways were found to be between 2.95 min and 3.36 minfor each direction and each day. To maintain these approximately3-min headways, researchers operated between six and eight testvehicles along the corridor. Figure 1 shows the speed profile forthe afternoon eastbound test vehicles, including the data for bothWednesday and Thursday.

TABLE 1 Study Site Characteristics

Signal DrivewayTotal Density Density

Length No. of No. of (signals/ (driveways/ PrimaryCorridor Location Case Study Use (miles) Median Type(s) Lanes Signals mile) mile) Land Uses

FM-60(University

Drive)

US-60 (Midlothian

Turnpike)

US-29(Seminole

Trail)

US-29(Lee Highway)

NOTE: No. = number, TWLTL = two-way left-turn lane.

Model development

Model validation

Model validation

Model validation

College Station,Texas

Chesterfield,Virginia

Charlottesville,Virginia(AlbemarleCounty)

Centreville,Virginia(PrinceWilliamCounty)

2.6

2.3

2.8

8.0

6

4

6–8

4

4.6

2.6

3.9

1.6

12

6

11

13

27

24

21

9

Retail,university

Retail,undeveloped

Retail,undeveloped

Undeveloped,state forest,retail,residential

Raised median,TWLTL

Grassydepressed,TWLTL,undivided,painted

Raised

Grassydepressedmedian, Jersey barrier, undivided

Additional Data Collection and Reduction

In addition to the extensive traffic volume and travel time datacollection described above, the researchers also

• Used GPS travel time data to determine lengths between inter-sections for developing a microsimulation operations model of thecorridor;

• Collected signal timings and coordination, if any, as well aspermitted-protected left-turn phasing, for both peak and off-peakperiods from each signal cabinet;

• Obtained as-built drawings of the corridor from the TexasDepartment of Transportation to determine turn-bay lengths;

• Documented changes in speed limit throughout the corridor;• Documented driveway locations and counts for density measure

along the corridor;• Computed turning movement volumes and percentages from

video recordings of signal approaches for the microsimulation modelinput; and

• Computed queue lengths for each lane and approach of thesignalized intersections to validate the microsimulation model.Queue lengths were computed by watching the video of each sig-nalized intersection approach, pressing the pause button before thegreen started, and counting the queue.

DEVELOPMENT OF MICROSIMULATION MODEL

To develop a model to estimate TTI for an arterial corridor, addi-tional observations of TTI values were desired to obtain a broaderrange of the dependent variable (travel time through the corridor).Because microsimulation is a stochastic model, multiple replica-tions can be used to capture variability in the response variable oftravel time.

To obtain the additional replications (observations), researcherscalibrated a CORridor SIMulation (CORSIM) microsimulationmodel. The researchers used the traffic counts, turning movementsreduced from video, signal timings, as-built drawings, and postedspeed limits to develop the calibrated model. The model was cali-

84 Transportation Research Record 2244

brated for both directions of travel. The researchers used the queuelength information to validate the microsimulation model. Inter-section approaches were investigated cycle by cycle to investigatequeuing parameters (e.g., average, maximum). The researcherscalibrated parameters such as link speeds and queue dischargecharacteristics to ensure the calibrated model was performingwithin approximately 5% of field data for maximum queues andtravel times.

After development of the CORSIM microsimulation model, theresearch team used the model to produce additional replications(runs) of TTI estimates given the range of operational inputs (e.g.,traffic volumes during different times of the day). The researchersset up the microsimulation model to produce additional replicationsof each 30-min period for which travel time data were available.Because there were 15 h total of data collection, there were thirty30-min periods. Calibrated and validated models were produced foreach of these 30-min periods by direction. Note that this procedurerefers to validation of the CORSIM microsimulation model, notvalidation of the TTI models that were developed and are describedin subsequent sections of this paper and which use the three casestudies from Virginia shown in Table 1. Traffic volumes, turningpercentages, signal timings, and geometric data were input intothe model.

The researchers replicated the microsimulation model 10 timesfor each 30-min period except Wednesday and Thursday afternoonfrom 6:00 to 7:00 p.m. As described later in the paper, this timeperiod was not used in the final analysis. The replications resulted in276 additional replications of the original data set. The researchersultimately had 30 original cases and an additional 276 replicationsof the total corridor travel time and average link volume along thecorridor from the CORSIM output.

DEVELOPMENT OF MODELFOR ESTIMATING TTI

After performing replications of the validated CORSIM model, theresearchers had a populated database with numerous variables fromwhich to develop a model to estimate TTI. Initially, all data were in

Sp

eed

(m

ph

)

Time of Day

60

50

40

30

20

10

015:00 15:30 16:00 16:30 17:00 17:30 18:00 18:30

Thursday

Wednesday

FIGURE 1 Afternoon eastbound test vehicle speeds.

the database at the link (signal-to-signal) level. Because the creationof a corridor-based model was desired, the research team aggregatedthe link-based information to the segment level. The researchers cre-ated six segments (three in each direction) along the corridor becauseof different signal coordination settings along the arterial.

The researchers defined the variables in Table 2 as predictor (inde-pendent) variables for estimating TTI. Table 2 provides the variablenames and descriptions. In general terms, the independent variablesconsider volume, access density, signal density, signal coordination,and the proportion of green time to total cycle time (g/C).

Equation 2 shows the general form of TTI. For this research, anaverage (statistical mean) off-peak free-flow travel rate was used inthe denominator for all the calculations.

The researchers normalized (by length) or weighted many of thevariables shown in Table 2. Normalizing some of the variables bylength (VMT/LN-MILE, DRIVEDENS, SIGDENS) ensures theyare scalable to corridors when the model is applied elsewhere.Weighting driveway density (DRIVEDENSWT) by through VMTensures that the model is scalable. It also incorporates the intuitivefact that driveway density should affect roadway operation as afunction of the number of through vehicles, length of the given link,and the number of lanes. Similarly, weighting g/C by through vol-

TTIaverage travel rate (minutes per mile)

free-f=

llow travel rate (minutes per mile)( )2

Eisele, Zhang, Park, Zhang, and Stensrud 85

ume on the link is meant to capture the intuitive affect that g/C alonga corridor should be averaged to the segment level as a function ofhow many vehicles are influenced by each link g/C value.

Determination of Peak and Off-Peak Traffic Conditions

The researchers investigated speed graphics by segment anddirection, as shown in Figure 1. The researchers also investigatedtravel time and TTI graphics by time of day to investigate the rangesof the TTI values over time. TTI values were computed relative tothe average and 15th percentile travel time; however, the relation-ships using the average values ultimately resulted in the most appro-priate models and are simpler for practitioners to understand andimplement. The researchers selected the time periods for the peakand off-peak analysis and model development with considerationfor the following:

• Signal timing plan. The researchers ensured that the peak andoff-peak periods used for the TTI computation included the sametiming plan (e.g., morning peak period or afternoon peak period)to ensure comparison to an off-peak travel time in the arterialenvironment with similar operational constraints.

• Speed and travel time profiles. The researchers created graphicsas shown in Figure 1 to investigate the speed (and travel time) profilesalong the corridor and the segments to identify peak conditions.

TABLE 2 Candidate Predictor (Independent) Variables for TTI

Variable Name inPotential Predictor Variable Model Description

VMT/lane-mile

Driveway density of segment

Driveway density of the segmentweighted by link through volume

Signal density of segment

Signal coordination

Weighted average green time tocycle time for through direction

Continuous variableVMT of through vehicles per lane-mileThis value reduces to the peak hour volume per lane. It is shown as PHV/LANE in theremainder of this paper.

Discrete variableOne value for each segmentNumber of access points in given direction and segment/length of segment

Continuous variableComputed as

Discrete variableNumber of links per length of segmentOnly one value for a pair of segments (eastbound and westbound)

Discrete variableFavored (There is coordination of signals, traveling during the peak, and in the peak direction.)No coordination (There is no coordination of signals.)Unfavored (There is coordination of signals, traveling during the peak, and in the off-peakdirection.)

Continuous variableAverage traffic signal green time per total cycle time in direction of travelWeighted by through volume of the linkComputed as

through volume

through volumelinks

links

×[ ]∑∑

g C

# access through VMT lengthlink linklink

link×( )∑tthrough VMTlink

link∑

PHV/LANE

DRIVEDENS

DRIVEDENSWT

SIGDENS

COORD

g/C

NOTE: VMT = vehicle miles traveled.

• TTI profiles. The researchers investigated plots of TTI bysegment and time period.

After consideration of these factors, the researchers determinedthat there were two ranges of congestion level in the data: (a) whatmight be termed “light” congestion conditions from 7:30 to 8:30 a.m.(TTI values up to approximately 1.35) and (b) what might be termed“moderate” congestion conditions (TTI values up to approximately2.8) from 5:00 to 6:00 p.m. The researchers used the morning datato develop a model representing light congestion conditions and theafternoon data to develop a model representing moderate congestionconditions. Average speeds in the peak direction during the “light”congestion period (morning) were approximately 25 mph and dur-ing the “moderate” congestion period (afternoon) were approximately18 mph.

Off-peak conditions (denominator in Equation 2) were computedas the average off-peak travel rate from 7:00 to 7:30 a.m. in the lightcongestion conditions model and from 4:00 to 4:30 p.m. in the mod-erate congestion conditions model. The Saturday (special event)data were not used in the model development, as the congestion lev-els were generally not as high throughout the corridor as those dur-ing the morning and afternoon commutes. After determining thesetwo time periods and the two corresponding models to pursue, theresearchers investigated numerous models; the results are providedin the sections that follow.

Equation 3 shows the general form of the fitted models developedin this paper. The βs are the regression coefficients, and the Xs arepotential predictors.

Not all of the variables shown in Table 2 ultimately turned out to besignificant for inclusion in each model after the statistical analysis.The following typical criteria and guidance were used in determiningthe regression model:

1. Correlation between independent variables. Ensuring that therewas no significant correlation among the selected independent vari-ables to ensure there was no collinearity problem in the regressionanalysis.

2. Parameter significance. Ensuring that only predictor (indepen-dent) variables that were statistically significant were kept.

3. Coefficient sign. Ensuring that signs on the coefficients in thefitted model were intuitive.

4. Residual investigation. Ensuring that there were no significantpatterns in the residual plots (plots of the predicted value comparedwith the difference between the predicted value and the actual value).Any patterns present in the residuals indicate that one or more of theunderlying assumptions for the regression was violated.

Moderate Congestion Level Model of TTI

For the model to estimate TTI for the moderate congestion level, theresearchers investigated numerous combinations of the independentvariables shown in Table 2. The researchers ultimately vetted themodels to satisfy the four criteria identified above. The general modelform is shown in Equation 4. This model satisfied the criteria andprovided the best fit (R2 = .94). All independent variables shown inEquation 4 are significant at the α = 0.05 significance level.

TTI = + + + + +β β β β β0 1 1 2 2 3 3 3X X X Xn n. . . ( )

86 Transportation Research Record 2244

For the moderate congestion level conditions, the best-fit modelis a function of PHV/LANE, weighted driveway density, g/C, andcoordination (favored, unfavored, or uncoordinated). In Equation 4,only one term is used to estimate TTI for the signal coordinationvariable (e.g., if coordination is favored, only the “−0.019(COORD[Favored])” term is used, and the other COORD terms are ignored;if coordination is unfavored, a term for coordination is not usedbecause the coefficient is zero). The signs on all terms are intu-itive. For example, increased PHV/LANE or DRIVEDENSWT yieldhigher TTI values (positive sign), and higher values of g/C yield lowerTTI values (negative sign). The other variables shown in Table 2were not statistically significant or did not result in a meaningfulmodel.

The researchers investigated the plot of the residual by predictedvalue for the moderate congestion conditions model, and no signifi-cant patterns were present (i.e., the residuals were generally centeredaround zero); therefore, the assumptions for regression were not vio-lated. The researchers also investigated exponential models, but thelinear models provided comparative TTI predictability, and they areeasier for practitioners to implement.

Light Congestion Level Model of TTI

For the light congestion level model to estimate TTI, the researchersinvestigated numerous combinations of the independent variablesshown in Table 2. The researchers ultimately vetted the models tosatisfy the four criteria identified above. The model shown in Equa-tion 5 satisfied the criteria and provided the best fit (R2 = .68). Allindependent variables shown in Equation 5 are significant at theα = 0.05 significance level.

For the light congestion conditions, the best-fit model is a functionof only PHV/LANE and g/C. It seems intuitive that at the lower con-gestion levels, the COORD term may not affect the model becausetraffic is flowing relatively smoothly. Similarly, access terms (DRIVEDENS or DRIVEDENSWT) do not seem to be significantfor this model. As with the moderate congestion level model,researchers investigated residuals, and no significant patterns werepresent (i.e., the residuals were generally centered around zero); there-fore, the underlying assumptions about errors for regression were notviolated. The researchers also investigated exponential models for thelight congestion level, but the linear models provided comparativeTTI predictability, and they are easier for practitioners to implement.

VALIDATION OF MODEL TRANSFERABILITY

The researchers investigated the transferability and use of the devel-oped models by applying them to three case study locations in Vir-ginia. The characteristics of these three roadways in Virginia areshown in Table 1.

TTI PHV LANE= + ( ) − ( )1 341 0 00041 0 781 5. . . ( )g C

TTI PHV LANE DRIVEDENSWT= + ( ) +2 936 0 00226 0 011. . . (( )− ( ) − [ ]( )+

6 770 0 019

0

. .g C COORD favored

COORD unnfavored

COORD uncoordinated

[ ]( )+ [ ]( )0 364 4. ( ))

Data Collection at Validation Study Corridors

Extensive data were needed to validate the light and moderate con-gestion models. The three Virginia case studies provided access tothe detailed data required. Note that the level of data detail requiredfor validation is more extensive than that required for practitioneruse of the models. Table 3 lists all of the data collected for eachvalidation site.

Model Validation Procedure

The researchers summarized the needed model inputs for each linkand at the corridor level for all three validation study corridors. Themodel independent variables were computed based on the descrip-tions in Table 2 and Equations 4 and 5. TTI was estimated for boththe light and the moderate congestion models after the independentvariables were computed.

Ground truth (field measured) average travel time and free-flowtravel time are needed to estimate TTI with Equation 2. The field traveltime runs and SYNCHRO were used to estimate the average travel ratein the numerator of Equation 2. The researchers used two methods toestimate the free-flow travel rate shown in the denominator of Equa-tion 2. The method depended on the site itself. For US-60 (Chester-field) and US-29 (Albemarle County), the researchers reduced thetraffic volumes in the calibrated SYNCHRO model by 50%,because the SYNCHRO file is created for a peak-hour conditionand the reduction from the peak-hour flow represents an approxima-tion of the free-flow traffic condition. For US-29 in Prince WilliamCounty, the researchers used the running time data as a function ofthe link distance divided by speed. This value was the similar to thetotal travel time at low traffic volumes in the SYNCHRO model,because the corridor was relatively uncongested.

With these estimates of the average travel rate and the free-flowtravel rate, the researchers computed the “ground truth (field mea-sured)” TTI value as the ratio shown in Equation 2. This groundtruth TTI value was compared with the TTI estimate from themodels shown in Equations 4 and 5. Note that when travel timemeasurements for free-flow conditions are available, they can beused directly, and the estimation of free-flow travel times is notnecessary.

Eisele, Zhang, Park, Zhang, and Stensrud 87

The researchers computed an error term as the absolute differencebetween the model estimate from Equation 4 or Equation 5 and thefield-measured TTI value, as shown in Equation 6.

Findings of Model Validation

Table 4 shows the results of applying the light and moderate con-gestion level models along the entire US-60 corridor by available

errormodel TTI estimate

=−ground truth field measured TTI value( )

( )6

TABLE 3 Data Available for Model Validation Sites in Virginia

Data Element US-60 (Chesterfield) US-29 (Albemarle County) US-29 (Prince William County)

ADT, K-factor, and D-factor

Calibrated SYNCHRO models

Field travel time runs (during peak hours)

Number and location of driveways

Predominant land use

Speed limit

Photographs and video logs

NOTE: ADT = average daily traffic, K-factor = percentage of the ADT that occurs in the peak hour, D-factor = percentage of the ADT that occurs inthe peak direction.aThe four SYNCHRO files include models calibrated with field data for a base case (peak and off-peak), which includes a coordinated timing plan inthe field without an adaptive split feature, and an adaptive split feature case (peak and off-peak), which includes a coordinated timing plan in the fieldwith an adaptive split feature. An adaptive split feature means that the split (percentages of green times in a cycle) changes according to the arrivingvolumes of different movements, and the split may change from cycle to cycle.

Yes

4 modelsa (2008)

Yes

Yes

Yes

Yes

Yes

Yes

Morning and afternoonmodels peak (2007)

Yes

Yes

Yes

Yes

Yes

Yes

Morning and afternoon peakmodels, mid-day off-peakmodel (2009)

Yes

Yes

Yes

Yes

Yes

TABLE 4 Model Validation Results for the US-60 Study Corridor

Ground TruthTTI Model (Field Measured)

Conditiona Direction Estimate TTI Value Errorb

Light Congestion Model

Without ASF EB 1.11 1.12 0.01(off-peak) WB 1.02 1.08 0.06

With ASF EB 1.13 1.11 0.02(off-peak) WB 1.05 1.09 0.04

Without ASF EB 1.08 1.10 0.02(peak) WB 0.99 1.11 0.12

With ASF EB 1.13 1.12 0.01(peak) WB 1.05 1.09 0.04

Moderate Congestion Model

Without ASF EB 4.00 1.12 2.88(off-peak) WB 2.83 1.08 1.75

With ASF EB 2.84 1.11 1.73(off-peak) WB 2.85 1.09 1.76

Without ASF EB 2.87 1.10 1.77(peak) WB 2.80 1.11 1.69

With ASF EB 2.91 1.12 1.79(peak) WB 2.85 1.09 1.76

NOTE: EB = eastbound, WB = westbound.aASF = adaptive split feature signal timing. See Table 3 for a description ofASF.bError defined as shown in Equation 6. Differences within 0.01.

SYNCHRO model condition and direction. The light congestionmodel has a smaller error than the moderate congestion model. Theresearchers found similar results for the US-29 Albemarle studycorridor.

For all three validation study sites, the researchers investigatedthe error term when estimating TTI using both the light congestionmodel and the moderate congestion model. Figure 2 shows the rela-tionship between g/C, PHV/LANE, and error when using the lightcongestion model (points shown as solid circles) and the moderatecongestion model (points shown as hollow circles). The points rep-resent link-level data. All directional links from all three validationcorridors are shown. The researchers found the following:

1. Green time allocation (g/C) appears more critical than trafficvolume. A roadway link (or corridor) can have relatively high vol-ume, but if motorists are provided with the green time needed, theycan flow with limited congestion.

2. When g/C < 0.45, the moderate congestion model performs bet-ter (average error of 0.32) than the light congestion model (averageerror of 0.49).

3. When g/C ≥ 0.45, the light congestion model performs better(average error of 0.21) than the moderate congestion model (averageerror of 1.33).

The researchers found that the US-29 (Prince William County)study corridor had the lowest link g/C values. These relatively lowg/C values were at the links adjacent to I-66 along US-29. It appearsthat when there are cross streets that require significant green time(g/C ≥ 0.55 on the cross street), link-level analysis is more appropri-ate with the moderate model adjacent to those cross streets. Themodel can then be used at the corridor level (i.e., combining sim-ilar condition links) with the light model where the green time isincreased on the primary roadway. Because of relatively similaroperational and geometric conditions, the researchers combinedall links to the corridor level for the US-60 and US-29 (AlbemarleCounty) sites.

88 Transportation Research Record 2244

Ultimately, the models provide an adequate sketch-planning tool toassist in identifying directional roadways for more detailed analysis.Such an analysis is most appropriate when transportation profession-als are performing a comparative analysis to prioritize roadway oroperational improvements and want to identify critical areas. TTI is aproven measure for performing such comparative mobility analyses.

These models appear useful for transportation professionals insmall to medium-sized communities for mobility analysis, particu-larly in the fringe areas of cities, where there may be limited, if any,infrastructure in place for mobility monitoring.

CONCLUSIONS AND DISCUSSION

This paper has described the development and validation of twomodels to assist transportation professionals in estimating TTI inthe arterial environment during light and moderate congestion con-ditions. These models are valuable at the sketch-planning level fortransportation professionals in small to medium-sized communitieswhen monitoring congestion and performing comparative mobilityanalysis of corridors to prioritize where infrastructure improve-ments are potentially needed. To address the limitations of existingmodels, the models presented in this paper (a) consider drivewaydensity, (b) are corridor-based, (c) are a function of generally avail-able or easy-to-obtain independent variables, (d) are calibrated andvalidated with extensive field data, and (e) explain a relatively high degree of variability. The use of TTI also makes the modelsmore transferable. The following are the key conclusions from thisresearch:

1. The model for moderate congestion conditions (TTI valuesup to approximately 2.8) is a function of traffic volume, drivewaydensity, g/C, and signal coordination condition. The moderate con-gestion level model has an R2 of .94. The model for light conges-tion conditions (TTI values up to 1.35) is based on traffic volumeand g/C along the corridor. The light congestion conditions model

0500

10001500

0.20.40.60.810

1

2

3

4

5

PHV/LANEg/C

Err

orWhen g/C islow, themoderatecongestionmodel has lesserror.

When g/C ishigh, the lightcongestionmodel has lesserror.

Error (M)Error (L)

FIGURE 2 Illustration of g /C and PHV/LANE and error (link data for all three Virginia corridors).

has an R2 of .68. The independent variables in these models aregenerally available or relatively easy to obtain for transportationprofessionals.

2. The research described here found that the level of drivewaydensity is an important predictive variable for speed-related perfor-mance measures in the arterial environment for relatively higher con-gestion levels. This is intuitive because at relatively higher trafficvolumes, added driveways provide more opportunity for interferencealong the corridor. The significance of driveway density in the modelindicates the importance of considering access management for TTIestimation in the arterial environment.

3. The results show that the effect of driveway density on TTI val-ues can vary throughout the day by traffic volume and driveway(business) activity. Traditional HCM methods set one representativespeed to try to capture driveway density impacts. The results heredemonstrate that setting one speed does not fully account for the realimpact, because the impact of driveway density can vary throughoutthe day. The results demonstrate the need to model different trafficlevels and time periods separately.

4. The researchers tested the transferability of the models inthree case study corridors in Virginia, with promising results. Theresults indicate that g/C allocation is more critical than traffic vol-ume. In other words, an arterial roadway segment can have highvolume, but if adequate green time is provided, congestion can berelatively limited.

5. The researchers found that when g/C < 0.45, the moderatecongestion model performs better (average error of 0.32), and wheng/C ≥ 0.45, the light congestion model performs better (average errorof 0.21). The researchers found that when there are cross streets thatrequire significant green time (g/C ≥ 0.55), link-level analysis is moreappropriate with the moderate model adjacent to those cross streets.The model can then be used at the corridor level (i.e., combining sim-ilar geometric and/or operational links) with the light model whereg/C is increased beyond 0.45 on the primary roadway.

6. One of the validation study corridors was in the Northern Vir-ginia area [US-29 (Lee Highway)]. Because a high-volume Inter-state (I-66) crosses this study corridor, it is not a typical site in a smallto medium-sized community. This site was chosen because it isexpected that the congestion level in some small to medium-sizedcommunities will reach the level that is currently experienced inlarger suburban corridors such as this. The researchers found thatthe moderate congestion model applies to corridors with futurehigher demand in small to medium-sized communities, and itapplies to those corridors in major suburban areas that experience ahigh level of travel and congestion. As indicated in Item 5 of thislist, link-level performance monitoring is more appropriate in thepresence of high-volume cross streets.

7. The models appear to provide an adequate sketch-planningtool to assist in identifying roadways and locations for more detailedanalysis (comparative analysis). The models can be applied by trans-portation professionals in fringe areas and medium-sized communitieswhere monitoring infrastructure is not currently in place. Although theresearchers performed extensive data collection for model develop-ment to capture as many influencing factors as possible, practition-ers do not need an extensive data collection effort to use the models.In fact, all of the data elements are relatively available or easy toestimate for traffic engineers or planners in a city with limitedresources. Further, the models provide a relatively simple methodfor quantifying roadway performance, which is important for justi-fying and documenting infrastructure investments during times oflimited funding.

Eisele, Zhang, Park, Zhang, and Stensrud 89

8. The case studies used for model development and validationare suburban arterials. While the models were developed with road-ways with links covering a range of values for g/C and PHV/LANE,the researchers have yet to test the models along corridors that haveseveral cross streets that require significant green time (g/C ≥ 0.55).Future research is needed to evaluate model performance in theseconditions.

9. Future work could include collection of more travel timedata for peak and off-peak times (free-flow) for model validation.Researchers could collect the travel time data with new technologies(e.g., Bluetooth, a private-sector data source). This would facilitatedata collection and provide a relatively large sample size.

ACKNOWLEDGMENTS

The authors thank the sponsors of the Mobility Measures in UrbanTransportation FHWA pooled-fund study, including 12 state depart-ments of transportation (California, Colorado, Florida, Kentucky,Maryland, Minnesota, New York, Ohio, Oregon, Texas, Virginia,and Washington) two metropolitan planning organizations (Houston–Galveston Area Council, Maricopa Association of Governments),and FHWA. The authors also thank the numerous individuals whoassisted in the data collection and data reduction for this project.

REFERENCES

1. Schrank, D. L., and T. J. Lomax. 2007 Urban Mobility Report. TexasTransportation Institute, Texas A&M University System, College Station,2007.

2. Lomax, T., S. Turner, G. Shunk, H. S. Levinson, R. H. Pratt, P. N.Bay, and G. B. Douglas. NCHRP Report 398: Quantifying Congestion.Volume 1: Final Report. TRB, National Research Council, Washington,D.C., 1997.

3. Highway Capacity Manual. TRB, National Research Council, Washing-ton, D.C., 2000.

4. Strong, D. W., N. M. Rouphail, and K. Courage. New CalculationMethod for Existing and Extended Highway Capacity Manual DelayEstimation Procedures. Presented at 85th Annual Meeting of theTransportation Research Board, Washington, D.C., 2006.

5. Fambro, D. B., and N. M. Rouphail. Generalized Delay Model for Sig-nalized Intersections and Arterial Streets. In Transportation ResearchRecord 1572, TRB, National Research Council, Washington, D.C.,1997, pp. 112–121.

6. Rouphail, N. M., M. Anwar, D. B. Fambro, P. Sloup, and C. E. Perez.Validation of Generalized Delay Model for Vehicle-Actuated TrafficSignals. In Transportation Research Record 1572, TRB, NationalResearch Council, Washington, D.C., 1997, pp. 105–111.

7. Engelbrecht, R. J., D. B. Fambro, N. M. Rouphail, and A. A. Barkawai.Validation of Generalized Delay Model for Oversaturated Conditions.In Transportation Research Record 1572, TRB, National ResearchCouncil, Washington, D.C., 1997, pp. 122–130.

8. Ran, B., N. M. Rouphail, A. Tarko, and D. Boyce. Toward a Class ofLink Travel Time Functions for Dynamic Assignment Models on Sig-nalized Networks. Transportation Research Part B, Vol. 31, No. 4,1997, pp. 227–290.

9. Li, J., N. M. Rouphail, A. Tarko, and L. Velichansky. Overflow DelayModel for Signalized Arterials. In Transportation Research Record 1555,TRB, National Research Council, Washington, D.C., 1996, pp. 1–8.

10. Daniel, J., D. B. Fambro, and N. M. Rouphail. Accounting for Non-random Arrivals in Estimate of Delay at Signalized Intersections. InTransportation Research Record 1555, TRB, National Research Council,Washington, D.C., 1996, pp. 9–16.

11. Fambro, D. B., N. M. Rouphail, P. R. Sloup, J. R. Daniel, J. Li, M.Anwar, and R. J. Engelbrecht. Highway Capacity Manual Revisions ofChapters 9 and 11. FHWA-RD-96-088, FHWA, U.S. Department ofTransportation, 1996.

12. Rouphail, N. M., and Z. K. Khatib. Arterial Signal Optimization Considering Left-Turn Control. In Transportation Research Record1456, TRB, National Research Council, Washington, D.C., 1994, pp. 72–82.

13. Tarko, A., and N. Rouphail. A Distribution-Free Model for EstimatingRandom Queues in Signalized Networks. In Transportation ResearchRecord 1457, TRB, National Research Council, Washington, D.C., 1994,pp. 192–197.

14. Li, J., N. M. Rouphail, and R. Akcelik. Overflow Delay Estimation fora Simple Intersection with Fully Actuated Signal Control. In Trans-portation Research Record 1457, TRB, National Research Council,Washington, D.C., 1994, pp. 73–81.

15. Sisiopiku, V. P., and N. M. Rouphail. Toward the Use of DetectorOutput for Arterial Link Travel Time Estimation: A Literature Review.In Transportation Research Record 1457, TRB, National ResearchCouncil, Washington, D.C., 1994, pp. 158–165.

16. Sisiopiku, V. P., N. M. Rouphail, and A. Santiago. Analysis of Correlation Between Arterial Travel Time and Detector Data from Sim-ulation and Field Studies. In Transportation Research Record 1457,TRB, National Research Council, Washington, D.C., 1994, pp. 166–173.

17. Akcelik, A., and N. M. Rouphail. Overflow Queues and Delays withRandom and Platooned Arrivals. Journal of Advanced Transportation,Vol. 28, No. 3, 1994, pp. 227–251.

18. Tarko, A., N. Rouphail, and R. Akcelik. Overflow Delay at SignalizedIntersection Approach Influenced by an Upstream Signal: An Analyti-cal Investigation. In Transportation Research Record 1398, TRB,National Research Council, Washington, D.C., 1993, pp. 82–89.

19. Akcelik, A., and N. M. Rouphail. Estimation of Delays at Traffic Sig-nals for Variable Demand Conditions. Transportation Research Part B,Vol. 27, No. 2, 1993, pp. 109–131.

20. Rouphail, N. M., and R. Akcelik. Oversaturation Delay Estimates withConsideration of Peaking. In Transportation Research Record 1365,TRB, National Research Council, Washington, D.C., 1992, pp. 71–81.

21. Rouphail, N. M. Cycle by Cycle Analysis of Congested Flows at Sig-nalized Intersections. ITE Journal, Vol. 61, No. 3, 1991, pp. 33–36.

90 Transportation Research Record 2244

22. Rouphail, N. M. A Traffic Model of Platoon and Secondary Flows.Archive of Transport, Polish Academy of Sciences, Vol. 2, No. 2, 1990,pp. 85–97.

23. Rouphail, N. M., and A. E. Radwan. Simultaneous Optimization of Sig-nal Settings and Left-Turn Treatments. In Transportation ResearchRecord 1287, TRB, National Research Council, Washington, D.C.,1990, pp. 1–10.

24. Rouphail, N. M. Progression Adjustment Factors at Signalized Inter-sections. In Transportation Research Record 1225, TRB, NationalResearch Council, Washington, D.C., 1989, pp. 8–17.

25. Mousa, R., and N. M. Rouphail. Effects of Platoons on Permissive LeftTurns: A Pilot Study. Journal of Transportation Engineering, Vol. 115,No. 2, 1989, pp. 208–215.

26. Rouphail, N. M. Delay Models for Mixed Platoon and SecondaryFlows. Journal of Transportation Engineering, Vol. 114, No. 5, 1988,pp. 131–152.

27. Rouphail, N. M. Analytical Warrant for Separate Left-Turn Phasing.In Transportation Research Record 1069, TRB, National ResearchCouncil, Washington, D.C., 1986, pp. 20–24.

28. Bureau of Public Roads. Traffic Assignment Manual. U.S. Departmentof Commerce, Urban Planning Division, 1964.

29. Turner, S. M., W. L. Eisele, R. Benz, and D. Holdener. Travel TimeData Collection Handbook. FHWA-PL-98-035I, Texas TransportationInstitute, Texas A&M University System, College Station, and FHWA,U.S. Department of Transportation, 1998.

30. Eisele, W. L., J. A. Crawford, and R. L. Stensrud. Measures, Methods,and Application of a Mobility Monitoring Process for Small to Medium-Sized Communities. FHWA/TX-08/0-5571-1, Texas Department ofTransportation, Austin, and FHWA, U.S. Department of Transportation,2007.

The Transportation Planning for Small and Medium-Sized Communities Committeepeer-reviewed this paper.