National Institute of Statistical Sciences


P.O. Box 14006
Research Triangle Park, NC 27709-4006
Tel: 919.685.9300
FAX: 919.685.9310
admin@niss.org


Measurement, Modeling and Prediction
for Infrastructural Systems


Project Information


PIs Alan F. Karr, NISS; Eric I. Pas, Duke University; Jerome Sacks, NISS
Source of Funds NSF
Funding Level $1,225,000 in FY 1996
Dates November 1, 1994 - October 31, 1999


Project Activities

Currently, the project has three foci:


Travel Demand

This topic is devoted to modeling and predicting activity--travel patterns.

Vehicle Ownership

The number of cars available to a household is a key factor in travel behavior, including choice of mode and trip chaining. Bayesian methods for estimating the parameters of a multinomial probit model have been applied to car ownership modeling. (The probit model relies on (unobserved) utilities of available alternatives, which depend on individual characteristics as well as attributes of the alternatives. The individual's choice is determined by maximizing utility.) Software has been developed to enable use of Gibbs sampling, modifying methods developed by McCulloch and Rossi to allow incorporation of individual characteristics and to make Markov chain Monte Carlo methods work more rapidly. Similar methods have been applied to model responses to travel demand management policies, and will be used to forecast travel behavior for synthetically generated households.

A multinomial, multiperiod probit model with random effects for the level of car ownership over time has been developed, and has been estimated using panel data collected in the Netherlands. C. Bhat (Civil Engineering, University of Massachusetts), A. Nobile (former postdoc, NISS) and Pas are key participants.

Activity Participation and Travel Behavior

Current efforts focus on activity--travel data collected in Portland, OR, in 1994-95, which contain information on in-home and out-of-home activity participation, as well as travel behavior, for all members of more than 4,000 households for a period of 48 hours. Data sets to be addressed in the future come from Raleigh-Durham, NC, Dallas-Fort Worth, TX, and the San Francisco Bay area.

Sociodemography, Activity Participation and Travel Behavior. Relationships among sociodemographics, time use and travel behavior (number of trips, number of tours, travel time and car mode share) have been studied using a structural equations modeling system (LISREL).

Activity-Travel Patterns for Synthetic Populations. Because transportation surveys typically collect data on only a small number of households, and since census data are inadequate, microsimulation of travel demand and hence network flows, requires generation of both a synthetic population and activity-travel patterns.

One approach models the daily activity-travel pattern of an individual as a Markovian sequence of activities, each described by type, duration and location. Travel is considered an activity, with location being the mode of travel. This work currently is employing time use data collected in Washington, DC, in 1994.

A second effort, aimed to develop activity-travel patterns for all the members of each synthetic household, assumes that a daily activity--travel pattern has a basic skeletal structure (comprising, for example, the number of activities and the number of tours), imposes constraints and simplifies simulation of the fine details of the pattern. Thus, to generate the activity-travel patterns of the members of a synthetic household, generate the skeleton of the daily activity-travel patterns by sampling from observed activity-travel patterns (from a travel behavior survey), and then simulate the details of the activity-travel patterns of household members by models based on observed probability distributions.

In addition, Bhat has developed a joint model of work mode choice and number of stops during the work commute, which provides a basis to evaluate the effect of alternative policy actions to alleviate peak-period congestion. Mode choice is modeled using a multinomial logit model and number of stops is modeled using an ordered response formulation. The model has been applied to data from the Boston metropolitan area. Bhat will extend this modeling framework in the next year to develop and estimate a joint model of activity participation on the work-to-home trip and activity participation after arriving home.

Pas leads this component of the project; other participants are R. Kitamura (Civil Engineering, Kyoto, and also associated with RDC, Inc.), P. Speckman (Statistics, University of Missouri), K. Vaughn (postdoc, NISS), and X. Lu (graduate student, Civil Engineering, Duke).

Travel Demand Management

Results from a travel survey in the Washington, DC, area became available during 1995; the data are stated preference responses to (hypothetical) policies designed to reduce congestion (such as parking taxes and congestion pricing). Several modeling strategies -Bayesian estimation of multinomial probit models, CART, generalized additive models and Bayesian model selection - were applied, but no model fit in a definitively good manner. All models, however, revealed high resistance by commuters to change from private automobile to other modes such as public transit or carpooling. The principal conclusion, though, is the need for very careful design of future surveys.

This component of the research was carried out by M. Clyde (Statistics and Decision Sciences, Duke), M. Lavine (Statistics and Decision Sciences, Duke), Nobile and Pas, with advice from Kitamura and J. Williams (RDC, Inc.).

Synthetic Household Generation

Prediction of travel behavior requires detailed sociodemographic information, tied to location, for all the households in a region. This amounts to reconstructing the joint distribution of some sociodemographic variables at the ("geographically fine") Census block group level, using as ingredients the Census marginal table at block group level (Full tables at the block group level are unavailable because of confidentiality constraints.), together with a 5% sample of the joint table, available for each Public Use Micro Area (PUMA).

One approach to the problem is iterative proportional fitting (IPF), which is rapid computationally, but potentially unrealistic because all reconstructed tables have the same odds ratios as the PUMA sample. An alternative approach has been taken: to simulate from the conditional distribution of the joint table, given the available Census information, by means of a Markov chain Monte Carlo sampler. Results from applying the method to data from the 109 block groups of Napa County, CA, are encouraging. Five sociodemographic variables were considered, with the simulated multiway table having 52320 cells. The results are comparable to those obtained using IPF, yet some clear differences emerge. The main advantage of the method is that a better idea of the variability in the distribution can be obtained.

Lavine and Nobile played key roles in this effort, supported by P. Goel (Statistics, Ohio State) and N. McMillan (former postdoc, NISS).

New Activities in 1996

Location Modeling. As a step in generating activity-travel patterns for synthetic populations, Speckman has developed a continuous version of the gravity model for spatial interaction, in which the zone parameters for origin and destination are replaced by continuous functions giving the attractiveness of arbitrary origin and destination points. A nonparametric estimation technique has been developed and applied to (home, work) location pairs in the Portland data. From the fitted model, the conditional density can be computed for work locations given any specified home location, and used to synthesize destinations, as required for microsimulation. In the next year, the initial model will be extended to include the locations of multiple activities on the trips to and from work.

Model Overfitting. An April, 1996, workshop articulated serious concern that models of travel behavior incorporate too many parameters in efforts to gain better goodness-of-fit. As evidence, in CART analyses undertaken to help identify relevant sociodemographic characteristics for use in sampling actual households observed in the survey, cross-validation performance often deteriorated quite rapidly beyond a very small number of nodes. The issue will be explored, beginning with an examination of the goodness-of-fit and cross-prediction performance for models of trip frequency.


Intelligent Transportation Systems (ITS)

Research on ITS addresses estimation of travel times in large, urban street networks, estimation of travel times on freeways, and network flow models. An investigation of freeway breakdown has been initiated.

Travel Times in Arterial Networks

Prediction of link (and route) travel times is essential in systems that provide dynamic route guidance information to drivers, such as Project ADVANCE. Primary data sources for such Advanced Traveler Information Systems (ATIS) are
Probe vehicles,
which provide real time information on link travel times to a central controller;
Traffic signals,
whose state is crucial to both values and dependence of link travel times; and
Detectors,
which provide count and occupancy data at 5-15 minute aggregations.
Resolving the relative contributions of these data sources and devising means to impute unavailable information have been key efforts.

A central activity was to conduct and analyze data from a field experiment conducted using ADVANCE probe vehicles and data collection mechanisms during the summer of 1995. For a small (11) link network, data collected include

Attention has focused on models that represent travel time as a function of relativized entry time to the link. (Entry time relative to the signal cycle, with zero corresponding to arrival at the signal just as it turns green.) An example:

Java-aware browsers display image here.

Daley has completed initial studies of link-to-link dependence in the same network, and is pursuing the topic.

Bayesian methods are being used to address a variety of conditional distributions. With TT denoting travel time, CT congested time and CD congested distance, these include

A related effort is exploring methods to infer origin-destination data from link flows.

Sen has played central role in evaluation of the (downsized) ADVANCE project, which involves implementing a series of designed experiments to test and validate various algorithms.

Still other issues that have been addressed include:

Principal sites for this activity are NISS headquarters, Duke and the University of Illinois at Chicago (UIC). Karr and A. Sen (Urban Planning and Urban Transportation Center, UIC) have led it; other participants are D. Daley (Statistics, Australian National), N. Rouphail (Civil Engineering, North Carolina State University), S. Stidham (Operations Research, University of North Carolina at Chapel Hill), M. West (Statistics and Decision Sciences, Duke), T. Graves (postdoc, NISS), P. Thakuriah (postdoc, NISS), H. Ashih (graduate student, Statistics and Decision Sciences, Duke), J. Goodstein (graduate student, Operations Research, University of North Carolina at Chapel Hill), C. Tebaldi (graduate student, Statistics and Decision Sciences, Duke) and X. Zhu (graduate student, Urban Planning, UIC).

Travel Time on Freeways

Estimation of travel times on a freeway from single loop detector data is being investigated using a very rich set of data collected by the PATH (Partners in Advanced Highways) Project (Berkeley, CA).\footnote{The data were gathered in order to study the effect of proactively dispatching fleets of tow trucks to ameliorate congestion caused by breakdowns and accidents.} At each of a number of locations spaced approximately one third mile apart, flow, occupancy and velocity were collected at a time resolution of one second from double loop detectors in each lane. Travel times of probe vehicles were also recorded.

The primary aim has been to estimate travel times from single detectors, which are common on US freeways. The single detectors yield information only on flow and occupancy, corrupted by noise and equipment malfunction. The data on velocity from the double loop detectors and from the probe vehicles is used as a benchmark to assess the accuracy of travel time estimates based on single loops.

Methods of travel time estimation have been developed based on correlation measures and on stochastic models for the travel times of individual drivers. The estimation problems are very difficult and ill-posed, but tuning the procedures via initial crude pilot estimates based on a simplified model of constant vehicle length and limiting the support of the estimated travel time distribution has been surprisingly effective. The methods have been successful in estimating daily profiles of travel times over the entire freeway stretch during the morning and evening commute hours, even in the presence of congestion.

This component of the project is housed at the University of California, Berkeley (UCB). Participants are P. Bickel (Statistics, UCB), J. Rice (Statistics, UCB), Y. Ritov (Statistics, Hebrew National), J. Jiang (postdoc, NISS), K. Petty (graduate student, Civil Engineering, UCB) and F. Schoenberg (graduate student, Statistics, UCB), with the cooperation of P. Varaiya (Civil Engineering, UCB, and Director, PATH).

Network Flow Models

Research on evaluating the attributes of models of route choice behavior on urban transportation networks addressed

D. Boyce (Civil Engineering and Urban Transportation Center, UIC) leads this component of the project; other participants are B. Janson (Civil Engineering, University of Colorado at Denver), P. Mirchandani (Industrial Engineering, University of Arizona), D.-H. Lee (graduate student, UIC) and M. Tatenini (graduate student, UIC). Lee and Tatenini completed doctoral dissertations under the auspices of this project.

Freeway Breakdown

In heavy traffic, freeways enter a regime of instability, and can undergo breakdown: without an exogenous cause (such as an accident), flows and speeds drop dramatically. This is depicted below with data from the North Carolina Department of Transportation, which show, for three lanes on I-40, time- and lane-dependent counts and vehicle speed distributions. Each band represents one lane; the "view" is from the shoulder. Heights are (smoothed) one-minute counts, obtained from detectors. Higher speeds are at the top; colors are a histogram of the distributions of vehicle speeds. The drop in speed depicted is from more than 60 to less than 35 miles per hour.

Java-aware browsers display image here.

From the standpoint of congestion management, the central need is to recognize breakdown before it occurs (in order to implement control measures such as changeable speed limits). The intriguing suggestion from that seemingly transient "breakdowns" are precursors of sustained breakdown is being pursued in collaboration with Rouphail, using data from both detectors and the MOBILIZER video system. Controlling the false alarm rate will be an important problem. The issues arise in a safety context as well: treating accidents rather than flow as the response points to need for recognizing incipient unsafe conditions. This question is being studied collaboratively with Rouphail and R. Hughes (Highway Safety Research Center, University of North Carolina at Chapel Hill).


Permeability of Concrete

Three pilot experiments have been completed, whose overall goal is to develop models that predict chloride permeability as a function of
Design ("mix") variables,
principally the volume fraction and grading (size distribution) of aggregate and the water/cement ratio;
Microstructural variables,
including the specimen-level volume fraction and size distribution of aggregate, the separation between aggregates, the paste radius, the area of the aggregate-paste interface and measures of the tortuosity of the interface surface.
Through the former, better concrete can be engineered; through the latter, the science of concrete is advanced.

Three initial experiments have been completed:

  1. Using specimens with two sizes of spherical (alumina) aggregate and pure cement paste, some of which were sliced, as shown below. Using a combination of commercial and NISS-developed software, the entire three-dimensional structure of the specimens was reconstructed, from which microstructural variables were calculated. Statistical models have been fit to the data, and predict, for example, expected but previously unverified "U-shaped" dependence on volume fraction.
  2. Using natural aggregate (but "gapped," that is, of only two sizes) and mortar (containing sand) rather than cement paste. OPTIMAS image analysis software was used to calculate microstructural variables. Modeling of the data is underway.
  3. A pilot experiment to measure fluid permeability of cracked concrete, a key first step in elucidating the feedback relationship between permeability and deterioration. The experimental protocol has been developed and initial results obtained. The first goal is to develop models that predict permeability from characteristics of the crack, such as the width (controlled during the process), length (measured) and number of connected components (measured).
A critical contribution has been to introduce statistical design to the experiment process. In addition, the equipment to measure chloride permeability was developed as part of the project.

Java-aware browsers display image here.

During the summer of 1996, large-scale (120 specimen) experiment was conducted, involving natural aggregate with (standard) size distributions used in practice, together with a designed experiment for chloride permeability of cracked concrete. Data are being analyzed.

This work has been overseen by Karr and S. Shah, Director, Center for Advanced Cement Based Materials (ACBM), Northwestern University. Other participants are B. Ankenman (Industrial Engineering, NU), T. Igusa (Civil Engineering, NU), T. Styer (postdoc, NISS), K. Wang (postdoc, ACBM) and S. Jaiswal (graduate student, Civil Engineering, NU).


Conference Presentations

Conference sessions featuring activities under this project have been held at the Summer Research Conference of the American Statistical Association and the Southern Regional Council on Statistics (June, 1995; Melbourne, FL); the International Congress on Industrial and Applied Mathematics (July, 1995; Hamburg, Germany); the Joint Statistics Meetings (August, 1995; Orlando, FL); the Third Workshop on Bayesian Statistics in Science and Technology (October, 1995; Pittsburgh); the national meeting of INFORMS (October, 1995; New Orleans); the International Federation of Operations Research Societies (October, 1995; St. Louis); the North American Meeting of the Regional Science International (November, 1995; Cincinnati); the Transportation Research Board (January, 1996; Washington, DC); the Spring Regional Meeting of the IMS (March, 1996; Richmond, VA); the American Ceramic Society (April, 1996; Indianapolis); and the national meeting of INFORMS (May, 1996; Washington); the Activity-Based Travel Forecasting Conference (June, 1996; New Orleans); the Workshop on the Relationship Between GIS and Behavioral Travel Modeling (June, 1996; Santa Barbara); the Conference on Theoretical Foundations of Travel Choice Modeling (August 1996; Stockholm); the Joint Statistical Meetings (August, 1996; Chicago); the Fourth World Congress of the Bernoulli Society (August, 1996; Vienna); the 36th Congress of the European Regional Science Association (August, 1996; Zurich); and the national meeting of INFORMS (November, 1996; Atlanta).

More than a dozen presentations are scheduled for the Transportation Research Board (January, 1997; Washington).



NISS Home Page NISS Projects Help