1 Background

1.1 Introduction

In the early days of automobiles when, for the first time in history, large numbers of people had opportunities to travel well beyond their local areas, finding directions was a problem. Prior to that, the range of most peoples’ travels was limited to a relatively short distance from their home, and people quickly became familiar with the small network that they regularly used. Signage was not needed. However, as new drivers roamed into unfamiliar areas, the lack of signage made getting lost a common occurrence.

Technology soon provided solutions to this problem (French 1986). For example, it was possible to buy and install an in-vehicle cylindrical or disc-shaped device that advanced at a rate that was synchronized with wheel rotation. The cylinders or discs had way-finding information printed on them. When initialized with the correct trip starting location, information about the direction options at each major decision point would be displayed prior to arriving there. Some enhanced versions also included static travel information such as road conditions, and locations of unimproved railroad crossings and speed traps.

Over time, of course, major investments in signage and road maps made such devices less useful, and research in traveler information systems was limited to relatively specialized applications such as for military vehicles. It is only relatively recently, with traffic congestion and the externalities of automobile use becoming more of a concern, that “advanced” traveler information systems (ATIS), have again become of interest. Technological progress in vehicle location, traffic monitoring, and data processing and communications have made possible applications that were probably not imaginable in the early days of the field.

Travel-related messages may be derived from static or dynamic information about the network. Static messages provide fixed information about the network and the destinations that it serves, and may be of use in tasks such as way-finding or preliminary trip planning; however, they do not recognize actual traffic conditions and so cannot respond to them. Dynamic messages reflect either prevailing or predicted conditions on the network, and require capabilities for collecting and possibly processing network data in real time. Such messages may describe the network conditions, or make recommendations based on the conditions, or both. A variety of presentation media (graphical, spoken or text) and levels of quantitative or qualitative detail in the messages are possible.

With these recent enhancements of traveler information system technological capabilities has also come an increased interest in understanding how travelers react to information provided in this way. There are many reasons for this interest:

· public and private organizations developing travel information products need to know what product features are most valued by travelers and why; this knowledge enables better products to be designed, and appropriate pricing strategies to be elaborated;

· public agencies investing in travel information infrastructure also need to know how travelers perceive and value the benefits that they will derive from the provided information, as guidance in making economically worthwhile investment decisions;

· many of these same agencies are examining the contribution that traveler information systems may make towards improving the overall operation and performance of their transportation systems, either by themselves or in combination with advanced traffic management systems (ATMS). Such network-level ATIS impacts can best be determined by aggregating the individual responses of many travelers to the information that they are provided by ATIS, but in doing this the interactions of the travelers on the network may become important and then must also be taken into account;

· finally, ATIS technologies currently under development will eventually be able to provide information based on predictions of future travel conditions; however, such information must incorporate a forecast of how travelers will react to it. For example, on the basis of short-term traffic forecasts, an ATIS may inform drivers that a certain route is expected to become congested in the next hour. If drivers react to this information by choosing a different route, their response may invalidate the forecast, leaving the original route free flowing but creating even worse congestion on an alternate route. Generating guidance based on forecast traffic conditions requires being able to forecast how drivers will respond to the guidance that they receive, determining the aggregate network-level impacts of the responses, and incorporating those responses and impacts into the guidance itself.

1.2 Scope and purpose of this document

In view of these reasons for an interest in traveler response to information, the Federal Highway Administration commissioned a review of published information on the subject; this report is one of the products of the study. It is a review of the literature published as of mid-2001 on the topic of traveler response to real-time information at the individual and network levels. (Static travel information is only considered in passing because of its rather limited scope for improving individual decisions or affecting network conditions.) The report’s intent is to summarize what is currently known about traveler response to information, in a form that provides a useful high-level understanding of the main issues.

This is not a comprehensive review – it could not possibly be, given the volume of material that has been (and continues to be) published in relevant areas. Several criteria were applied in deciding what to review:

· recent (past few years) publications with relevant research or applications results;

· publications providing summaries of long-term research or operational programs;

· selected early (pre-1990) publications, chosen for their historical interest or because their results are still relevant;

· selected publications from the mid-1990s, again chosen for their relevance or historical interest.

It will be seen that, despite the number of publications in the field, understanding of traveler response to ATIS is still in its initial stages. No one is yet able to accurately predict, for a VMS displaying a particular message at a particular location in a particular network, what the effect on individual travelers or on overall network conditions will be. Only limited data is available on individual responses to information, from operational deployments or from surveys investigating user reactions to hypothetical systems. Available data tends to be concentrated in specific areas such as commuter driving behavior; much less is known about information effects on non-commute trips, transit riders and commercial vehicle operators, for example. Efforts to develop models of traveler response based on these data are, for the most part, cutting-edge academic research far removed from the capabilities and needs of mainstream practitioners. Network-level forecasting models capable of predicting ATIS system impacts are also still mostly ad hoc in nature, frequently involving the cobbling together of two different model systems.

This state of affairs is not entirely surprising. Automobiles and modern transit systems were in use for roughly half a century before systematic and comprehensive travel data collection efforts were undertaken, and useful individual- and network-level transportation planning models began to be developed and routinely applied. While the pace of research and development is much faster now, a decade of experiments with ATIS is not foundation enough to support the development of a full understanding of its effects.

For these reasons, this review does not devote excessive effort to documenting the complete sets of results from available user surveys, or the full details of current model systems. For the same reasons, too, it discusses survey and analysis methods as well as with results, because robust and powerful methods will be needed to obtain further useful results in the future. At this point in the development of the field, the creation of appropriate tools and methods is just as necessary and important as their application.

This document may perhaps best be regarded as a source of raw materials that can be used in many different ways. Material can be extracted from it to prepare more specialized documents, focused on particular topics or audiences. It provides extensive references to and discussions of the published literature, enabling the original detailed results on particular subjects to be easily located. Although it mostly highlights what has been done to date, this focus also illuminates some of the gaps in current knowledge, and suggests actions that need to be taken in the future to advance the state of knowledge. In one particular area – the modeling of network-level ATIS impacts – the report makes suggestions regarding specific directions for future development approaches.

A companion document provides a number of specific recommendations for Department of Transportation actions to further knowledge in the field of traveler response to information, based in large part on the gaps identified here.

1.3 Structure of the document

This document is in two parts:

· a high-level summary of the state of the art in a number of areas related to traveler response to information. It attempts to summarize what is known in the area, and also to point out major gaps in current knowledge; and

· a series of reviews (annotated extended abstracts) of relevant documents. These documents provided the knowledge and data that were used in preparing the high-level summary.

The summary discussion covers:

· traveler behavior without information (Section 2);

· traveler behavior with information (Section 3);

· network impacts of ATIS (Section 4); and

· modeling ATIS network impacts (Section 5).

In the document reviews, a single review sometimes covers several documents because of their logical or organizational connections; frequently these are cases where a series of articles describes a line of research pursued over time. To facilitate locating particular document reviews, a listing is provided following the table of contents; it references each document with the number of the page where it is reviewed.

2 Traveler behavior without information

Before beginning a review of the literature on the effect of information on traveler decision-making, it is worthwhile to briefly summarize current approaches and understanding of such decision-making in the absence of information. This is useful for a number of reasons.

Traveler behavior exhibits many features that do not depend in a significant way on whether information from external sources is available or not. Many of these features have been identified and elucidated through studies of behavior without external information. Furthermore, it is likely that many aspects of traveler behavior in the presence of information are variations on similar behavior without information. For example, if travelers are sensitive to travel time in selecting their travel path, it is likely that many aspects of their behavior when they have reliable information on travel times will be similar to their behavior when they had to estimate these times. However, the availability of more precise and reliable time estimates may lead to modified or new behaviors that were not present when only low-quality information could be had.

Understanding of the factors that travelers consider when making trip-related decisions, and of the relative importance of these, can suggest which types of information an ATIS should provide.

Many of the methods that have been developed over the decades to analyze traveler behavior in the absence of information remain applicable to the analysis of behavior with information, so it is worthwhile to briefly review these in the simpler no-information context.

Finally, in some ATIS technologies, travelers will make portions of their trips without information and other portions with information. Consider an ATIS that transmits traffic information over a short range only: a VMS or low power radio transmitter, for example. A driver might leave home having made her travel decisions without input from the ATIS, and only receive reliable real-time information in the middle of her trip. The trip thus consists of two portions: an initial segment without information, and a final segment with information. Accurate predictions of driver behavior and of the network impacts of ATIS would require reliable models of decision-making in both contexts.

In short, traveler behavior with information cannot be understood without knowing something about traveler behavior without information.

2.1 Route choice

Transportation professionals since the beginning have had to consider the question of traveler route choice behavior, since it directly affects network-level traffic flow patterns and costs. For simplicity and convenience, any analyses have assumed that travelers choose, from among a set of alternative routes under consideration, the one that offers the lowest travel time or travel cost. From introspection and observation, however, it is not difficult to conclude that this is usually only an approximation of a more complex decision-making process.

There have been many efforts over the years to obtain a more detailed understanding of how travelers decide which routes to consider and then select one to follow. Many of these have been directed towards understanding the decision mechanism that underlies travelers’ route choice behavior and establishing an appropriate modeling theory and modeling form. A number of selected research articles were reviewed to highlight some of these modeling efforts and the methods they employ.

One of the basic approaches to understand drivers’ route choice behavior is descriptive data analysis. Data collected in the field and from driver surveys are used to infer drivers’ route choice criteria and their relative importance drivers’ decision-making processes. Descriptive statistics of the data form the basis of this approach. (Huchingson, McNees et al. 1977) and (Ratcliffe 1972) used this kind of approach to find the driving habits of the drivers – routes taken, reasons for selecting these routes, and the most important factors influencing the selection. (Heathington, Worrall et al. 1971) conducted a similar study. They found that drivers were more likely to divert to avoid delays or to save travel time on the trip to work than on the trip home; they further found that drivers were more likely to divert in order to avoid delay rather than to save travel time.

Another distinct approach in the existing literature is to use different statistical techniques like principal component factor analysis, canonical correlations, multiple regressions and grouping techniques. (Wachs 1967) used principal component factor analysis to determine whether different reasons that individuals gave to explain their route choices indicated the same or different underlying values. Respondent’s attitudes were examined to determine whether they were influenced by the performance characteristics of the routes. Statistical explanation of the attitudes, in terms of driver and route characteristics, was approached by three methods: canonical correlation, multiple regression and grouping techniques. The results of these analyses are presented and conclusions are drawn regarding the dependence of attitudes toward route choice upon persons and route characteristics. (Heathington, Worrall et al. 1971) also conducted a factor analysis to determine whether relationships existed between diversion frequency and other selected respondent characteristics. However, they did not find any meaningful relationship. (Pedersen 1998) used principal component factor analysis to identify the factors that influence person’s route choice. Four orthogonal factors involved in selecting automobile routes were obtained: safety, interest, purpose and hindrances. A profile analysis was also performed to find if these factors were differentially rated by men and women.

Route choice can also be modeled as a continuous variable in a variety of ways. (Duffell and Kalombaris 1998) identified the main route serving various trip origins and destinations, then used regression analysis to estimate the percentage of drivers using a route other than the main route under consideration.

Disaggregate (i.e., individual-level) choice analysis methods based on random utility models have been widely applied to model drivers’ decision making processes. In the context of disaggregate route choice modeling, the routes available to a traveler make up the choice alternatives, and the model predicts the probability that each of the routes in the set will be chosen. In this class of models, simple multinomial logit models are the simplest and perhaps most commonly used. However, the IIA (independence from irrelevant alternatives) property of the simple logit model restricts its applicability to general route choice analysis. This property results from the logit model assumption that path utilities include a random error term, and that the error terms of different paths are statistically independent of each other. Particularly in urban road networks, where alternative paths may overlap over significant portions of their length, the IIA property can be violated because of correlations in unobserved path attributes.

A number of modifications to the basic multinomial logit specification have been proposed to address this problem. For example, a size variable or a commonality factor may be included in the utility function to account for overlap between paths in the choice set. Another approach is the scaled paired combinatorial logit model, which scales the path utilities by a pair-wise similarity parameter. These models retain much of the simplicity and computational convenience of the basic logit model form, but overcome the unrealistic consequences the IIA property by coping with the correlation between paths.

The nested logit model, a generalization of the simple logit model, has also been used for route choice modeling. The advantage the nested logit model is that, by construction, it avoids the IIA property of the standard logit model. Estimation of nested logit models is only slightly more complex than that of simple logit models; software is readily available for this purpose.

Application of discrete choice modeling methods to route choice behavior is made complicated by the very large number of practically feasible routes between most origin and destinations, and the complex overlapping of these routes. The paper by (Ben-Akiva, Bergman et al. 1984) treats these difficulties by developing a two-stage model structure: choice set generation followed by selection from the choice set.

In the first stage, a labeling approach is used to reduce the huge number of potential routes to a much smaller number of routes, each of which reflects a criterion that might be relevant to route choice. These criteria (minimize time, minimize distance, maximize scenery along routes, etc.) are called labels. For each label, a criterion (or a generalized impedance) function is defined so that a network minimum path algorithm can be used to build trees that are minimal with respect to the criterion. Paths in these trees emphasize the corresponding label characteristics. For example, when considering the scenery label, time spent on roads with poor scenery would be weighted much more heavily (i.e., have greater impedance) than time on scenic roads. In specifying and selecting these labels, the objective is to generate a reasonable set of paths that include the actual paths chosen by the drivers. The selection of labels is made to maximize the coverage by the label set of the actually chosen paths, and the optimal values of the parameters of the impedance functions are the values that maximize this coverage. A deterministic choice set generation model is estimated for this purpose.

In the second stage, a model of choice from the set of labels is applied to predict the chosen route. A discrete choice model in the form of nested logit model is used for this stage. Path attributes specified in the utility function include generic variables like time and distance that describe the physical path, as well as dummy variables. The resulting model formulation was too complicated to be estimated using available software. Estimations were made with a series of successively less severe restrictions imposed on the general model.

In the study of individual route choice behavior, it is important to capture the heterogeneity in drivers’ tastes (preferences). In general, taste variations across individuals results in differences regarding their responses to alternative attributes and their preference to various choices. Similarly, when studying the behavior of an individual over time (because of repeated surveys, for example, or when modeling a learning process), it is important to recognize potential correlations between the individual’s choices. A logit model with fixed coefficients is not capable of fully accounting either for the variations in taste between individuals or the correlation between repeated choices by the same individual over time. Accurate modeling of route choice behavior requires a model that can capture differences in intrinsic preferences and subjective evaluation of alternative attributes due to both observed and unobserved heterogeneity.

The mixed multinomial logit (MML) model provides the flexibility to cope with these issues. In the MML model, an additional error term is added to the utility specification. Depending on the model, the additional error term may have a normal, uniform, log-normal or other distribution, with parameters to be estimated. The additional term captures heteroscedasticity among individuals and allows correlation over alternatives and time. However, this generality comes at a cost: choice probabilities cannot be computed analytically as they can, for example, in a logit model. Simulation techniques must be used to approximate the choice probabilities needed for model estimation and application. Recent advances in simulation-based estimation procedures make this more computationally feasible than it formerly was.

(Han, Algers et al. 2001) used an MML formulation to model route choice. Different error term distributions and model specifications were tested. The models with log-normal error distributions could not be estimated due to computational difficulties, leaving three alternative distributions – fixed, normal, and uniform. The logit model tested with fixed coefficient values differs from the standard logit model by incorporating the correlation between repeated choices by an individual. Dramatic improvement in the statistical performance of the models was found by allowing the coefficients of observed variables to vary randomly across individuals. The change in the estimated parameters caused by using the MML model was also significant. Parameter coefficients are generally larger in the MML relative to the simple logit model.

2.2 Departure time choice

Peak period congestion is one of the most persistent problems facing the transportation system. Transportation planners and transit operators have become increasingly aware of the need to spread the concentration of peak period travel. Various strategies proposed to combat the peak period problem are based on encouraging commuters to alter the time at which they travel to work. One way of assessing the potential impact of these strategies is to develop an understanding of the factors that affect commuters’ departure time decisions. A significant amount of research has been done on modeling commuters’ departure time choice in the absence of information.

A number of research papers on this topic have been reviewed. Again, given the amount of published research and the limited time frame available for the literature review, this cannot be considered a comprehensive survey of available material; rather, it highlights a number of interesting and representative research efforts and their conclusions.

Many research efforts apply disaggregate random utility models, of which the simple multinomial logit model is perhaps the most widely used. In the context of departure time choice modeling, discrete departure time intervals are used as the choice alternatives.

Departure time was modeled in combination with mode choice by (Hendrickson and Plank 1984): mode and departure time choices were treated as a simultaneous interactive decision. They developed a logit model that included up to twenty-eight alternatives, representing combinations of four modes (drive alone auto, shared ride, transit with walk access and transit with auto access) and seven different departure time intervals of 10 minutes each. The modal utility specification included: free flow in-vehicle travel time, the portion of total travel time due to congestion; monetary cost divided by income; walking time on the home end of a transit trip; wait time; minutes of late arrival at work and a quadratic function of that; minutes of early arrival at work and a quadratic function of that.

Much departure time research has focused on auto commuters; transit users have been neglected from consideration. One exception is a discrete choice modeling study by (Abkowitz 1981) of departure time choice. Among the objectives of this research were to extend the study of commuter departure time to include transit commuters, to include consideration of a wide range of socio-demographic characteristics, to account properly for the travel time uncertainty in departure time choices, and to improve the definition of arrival measures. Departure time choice was modeled conditional on mode choice. Departure time was represented as a discrete choice, using a logit model formulation. Each alternative represented a five-minute departure time interval, and the data input for each alternative represented an average of departure attributes for the interval. It was assumed that transit service frequency was sufficiently high during the peak period that all transit users were given a full set of choices.

Although the multinomial logit model structure is appealing to researchers because of its simple formulation, its IIA property is not always appropriate. In the context of departure time modeling, the IIA property implies that that the comparison of two departure time intervals does not need to consider whether they are adjacent or non-adjacent. In reality, two adjacent intervals are likely to be perceived similarly due to unobserved attributes common to both.

The ordered generalized extreme value (OGEV) structure generalizes the MNL structure by allowing an increased degree of sensitivity between adjacent departure time alternatives compared to between non-adjacent departure time alternatives and avoids the IIA restriction. (Steed and Bhat 2000) attempted to model departure time choice using an OGEV structure. However, the dissimilarity parameter in the OGEV model was greater than 1, implying inconsistency with utility-maximization theory. Hence, only the MNL structure was used for the analysis.

The argument in support of the treatment of departure time as a discrete choice is that travelers can only distinguish among a few prevailing traffic conditions over a specified departure period. However, discretizing departure time imposes an arbitrary structure of time intervals on the decision model. (Abu-Eisheh and Mannering 1987) develop and estimate a model that treats departure time as continuous variable and thereby avoids any a priori restrictions due to time discretization. Departure time is modeled as a function of the work start time, travel time, work access time and delay cushion (defined as the time difference between work start time and arrival time). Work start and work arrival times are assumed to be exogenous to the route and departure time choices. Travel time on a route is modeled as a function of route specific characteristics, commuter socio-economic characteristics and vehicle characteristics. However, since travel time on a route and the route choice are interrelated, there is a selectivity bias. The expected value method is used to correct this problem, where every route specific variable included in the travel time equation is replaced by its expected value. Delay cushion on a route is also modeled as a function of route specific characteristics, commuter socio-economic characteristics and commuter preferences for early or late arrival. The delay cushion model is also corrected for possible selectivity bias. The travel time and the delay cushion models are estimated by ordinary least squares.

Another approach to departure time modeling uses Poisson regression. The motivation for this is the belief that commuters never completely settle on a fixed departure time and route because they continually experiment with travel options and because of random effects such as weather. Within this context, a Poisson distribution is found to be a reasonable description of the number of departure time changes. Such a methodological approach is commonly referred to as Poisson regression. (Mannering 1989) and (Jou and Mahmassani 1994) used this approach to model the number of departure time changes by commuters within a month and a week respectively.

A novel approach to model driver departure time decisions is to investigate the cognitive aspects of the decision. This approach treats the departure time choice as a problem of decision-making problem under uncertainty. It criticizes the expected utility theory approach that is frequently applied to departure time modeling because expected utility theory is felt to ignore the cognitive processes underlying observed travel behavior. Depiction of travel behavior under uncertainty requires cognitive models, rather than probability theory, to capture the mental representation of uncertainty. Another finding of this kind of approach is that the decision frame, i.e. the subjective interpretation of the decision problem, critically affects decision-making. It has also been pointed out that the uncertainty of outcome is perceived as an interval of possible resultant values. Based on these findings from cognitive science, (Fujii and Kitamura 2001) propose a model of commuter departure time choice based on a cognitive task and a mental representation of uncertain travel time. By using departure time choice data, the study shows the presence of decisional phenomena, which are poorly explained by expected utility theory, but are explained well by the proposed model.

Most of the research on departure time modeling considers peak period work trips exclusively. In contrast, (Steed and Bhat 2000) modeled departure time choices for home-based recreational and shopping trips. This research examines the effect of socio-demographic characteristics, employment-related attributes, and trip characteristics on individuals’ departure time choices. The departure time alternatives are represented by several temporally contiguous discrete time periods such as early morning, a.m. peak, a.m. off-peak, p.m. off-peak, p.m. peak, evening. The choice among these alternatives is modeled using a discrete choice model. Two alternative discrete choice structures were explored. The first is the multinomial logit (MNL) structure and the second is an ordered generalized extreme value (OGEV) structure.

2.3 Mode choice

The literature on mode choice modeling is vast, and no attempt was made to review or summarize it. The following paragraphs simply note some modeling approaches commonly applied.

As travel modes are by their nature discrete alternatives, discrete choice models suggest themselves as a natural modeling approach. In this approach, all the modes available to a traveler constitute the choice set. Simple logit models are often applied to compute the probability of choosing each mode. The utility to a traveler for a particular mode can be a function of travel time (in-vehicle and out-of-vehicle) on that mode, out-of-pocket costs on that mode, perceived costs on the mode, socio-economic and demographic characteristics of traveler, workplace dummy and lots of other dummy and continuous variables. Many of these variables can be specified either generically or as specific to one alternative.

As has been mentioned above, the standard logit model has the independence from irrelevant alternatives (IIA) property. This means that for a specific individual the ratio of the choice probabilities of any two alternatives is entirely unaffected by the systematic utility of any other alternative. This can be unrealistic in mode choice modeling, because some modes in the choice set may have similar unobserved attributes and so have correlated utilities. An individual choosing between auto, commuter rail and express bus, for example, is likely to have somewhat similar (positive and/or negative) feelings about bus and rail, so treating them as completely independent vis-à-vis the auto could lead to unrealistic choice predictions.

The simplest generalization of the logit model that avoids this problem is the nested logit model; properly specified, it does not suffer from the IIA property. In this modeling approach, alternative modes that are likely to have unobserved common attributes should be put in a single nest and the resulting model should be used. The model incorporates a higher-level choice between nests, and a lower-level choice among the alternatives in a nest. In the previous example, it would be reasonable to group the commuter rail and express bus in a single “commuter transit” nest. The high-level choice would be between auto and commuter transit, with a lower-level choice between bus and rail in the transit nest.

3 Traveler behavior with information

This section considers the question of traveler behavior in the presence of real-time travel information.

This general question actually involves a number of closely inter-related sub-questions:

· which kinds of travelers would use real-time travel information if it were available? What kinds of trips would they want to use it for?

· how would they respond to the information once they received it? How would directly affect decisions about a trip being contemplated or made? How would it affect the context in which trips are made?

· what specific types of information would these travelers want to access?

· how much would they be willing to pay to receive the information?

· what would be their assessment of the benefits they received from accessing the real-time travel information and responding to it?

· how would this assessment of their experience affect the answers to all these questions the next time they have the opportunity to use it?

Because of their deep interdependence, all these questions should ideally, perhaps, be addressed and answered simultaneously. However, it is necessary to begin somewhere. Therefore, this section starts with a review of some of the literature that analyzes and characterizes the potential users of ATIS. From this, it turns to examine the various kinds of user response to travel information that have been studied. It then looks at users' preferences and willingness to pay for different types of information. There follows a discussion of the dynamic effects that can occur when day-to-day learning behavior is considered. Finally, a number of specific topics in traveler response data collection, analysis and modeling are discussed.

3.1 Who are the potential users of real-time travel information?

Understanding who are the potential users of advanced travel information services is essential both for designing and marketing those services and for predicting the users' responses to them. It is intuitively clear that ATIS can serve a variety of different kinds of users, and that these different kinds of users may react to ATIS messages in substantially different ways. The better these differences are understood, the better user needs can be met and user response can be predicted.

Studies of travel behavior are increasingly drawing on ideas and methods of market research. These methods typically attempt to identify subgroups ("segments") of the total market having the property that individuals within a subgroup share many similarities with respect to variables of interest in a study (e.g., travel behavior, socio-economic characteristics), and individuals in different subgroups differ significantly along these dimensions. Each homogeneous market segment can be more efficiently studied than can the mixed population as a whole.

A straightforward way of implementing these ideas is to identify segments on the basis of the exhibited behavior of interest (e.g., ATIS users), and to correlate membership in the segment with other measurable characteristics (e.g., socio-economic characteristics). Although useful, this approach has the disadvantage of being able to identify only relatively simple correlations, and perhaps also of reflecting the analyst's a priori beliefs and preventing a more exhaustive exploitation of the data.

More sophisticated market research methods such as cluster analysis can statistically identify population subgroups whose members share high degrees of similarity across many dimensions. While outputs of statistical procedures always need to be interpreted with insight and caution, clustering methods are often capable of identifying previously unknown significant population segments that might not have otherwise been recognized in the data.

Factor analysis is another method of identifying structure in a data set consisting of multiple observations, each one involving multiple variables of interest. Factor analysis identifies sets of linear combinations of the variables that distinguish as much as possible among the observations. Given a particular linear combination of variables (a factor), an observation's score with respect to the factor is the numerical value of the linear combination evaluated using the particular values of the observation's variables. Factor analysis identifies factors such that (i) the distribution of scores with respect to each one has maximum variance (i.e., the factors have maximum discriminatory power), and (ii) different factors are orthogonal to (i.e., uncorrelated with) each other. When a factor's linear combination includes some variables with very high coefficients and others with very low coefficients, its interpretation may be relatively easy. Factors involving more general linear combinations with arbitrary coefficients on the variable may be more difficult to interpret. In such cases, identified factors may subsequently be "rotated" to facilitate their interpretation in terms of specific variables or sets of variables, and this rotation may introduce correlations between them.

The combination of factor and cluster analysis is a particularly powerful means of identifying market segments, and has come to be a standard method in market research. Factor analysis is first applied to a data set of survey results to identify a set of factors that efficiently and parsimoniously distinguishes the observations. Each observation's scores with respect to the different factors are computed, and then cluster analysis is applied to identify subgroups of observations having similar factor scores. It remains for the analyst to impose a meaningful interpretation of the subgroups so obtained.

(Proussaloglou, Haskell et al. 2001) describe an application of combined factor and cluster analysis to identify transit user market segments in the San Diego metropolitan area. They then develop (fairly conventional) transit mode choice models for each distinct market segment.

Turning to analyses of the potential market for ATIS services, most surveys of potential ATIS users have carried out simple correlations or other descriptive analyses of stated use propensity with socio-economic or characteristics. Work pursued over a number of years by a group at the University of Washington (Barfield, Haselkorn et al. 1989; Haselkorn, Spyridakis et al. 1989; Haselkorn, Barfield et al. 1990; Wenger, Spyridakis et al. 1990; Spyridakis, Barfield et al. 1991; Conquest, Spyridakis et al. 1993) is among the first examples of the application of cluster analysis techniques to investigate the characteristics of potential ATIS users. Based on an mail-in driver survey and follow up personal interviews, the researchers were interested in the respondents' use of traffic information (commercial radio and TV traffic reports, HAR, VMS) and response to it, and in the influences that affect these responses. Cluster analysis of the survey results was intended to identify subgroups that differ significantly in their use of traffic information. The four groups identified by the cluster analysis were (in decreasing order of frequency in the sample): departure time and route changers; non-changers; route changers; and pre-trip changers. (Although mode change behavior in response to travel information was also investigated, the number of respondents who reacted to travel information by changing mode was not significant.) Descriptive statistical analysis was then used to further characterize each of the identified market segments in terms of its use of and attitudes towards different information sources; its priorities with respect to different information features; its tripmaking and activity constraints; and its demographics.

(Mehndiratta, Kemp et al. 1999b) (see also (Mehndiratta, Kemp et al. 2000; Mehndiratta, Peirce et al. 2000)) illustrate the application of combined factor and cluster analysis techniques, as described above, to delineate distinct segments of ATIS users. A detailed collection of data on travel behavior including use of travel information was conducted as part of the ongoing Puget Sound Regional Council's travel diary panel survey. The survey included conventional demographic and socio-economic information as well as responses to attitudinal questions. From this data, individuals with a high propensity to use travel information were identified. An initial attempt to correlate membership in this group with socio-economic characteristics, based on stereotypes of expected users types (e.g., road warriors, commuting mothers) proved only partially successful. Accordingly, a factor analysis of the entire survey population’s attitudinal question responses was performed, and a cluster analysis using the factor scores was carried out to identify distinct segments. Although the segments were defined uniquely in terms of their attitudes, subsequent analysis showed that the segments also differed with respect to their travel behavior, demographic profile, and propensity to use ATIS. The incidence in each segment of individuals likely to use ATIS was then determined.

Eight distinct market segments were identified through the combined factor/cluster analysis. The segments with higher-than-average incidence of ATIS users were termed:

· control seekers: people who travel a lot, are comfortable with technology, like to plan ahead but are not set in their ways;

· web heads: people who are interested in cutting-edge technology and traffic information, although they are less interested in portable electronics.

· rigid routines: people who usually follow the same routine but listen to traffic information and will make small adjustments to their trips;

· value-added service buyers: people uncomfortable with maps and computers who appreciate things that facilitate their daily lives;

· wired with children: people with high incomes, long commutes and children, for whom convenience is important.

Subsequent application of this approach to a wider sample of people who had used ATIS during the various MMDI programs revealed an additional potentially important segment:

· mellow techies: people with little interest in traffic conditions or trip planning, and little concern about being late, but who have high levels of internet and computer use.

It is clear that application of techniques such as these can provide considerable insight into the structure of the market for ATIS services, and allow much more focused investigation of the characteristics, system preferences and behavioral responses of potential ATIS users.

3.2 Traveler response to real-time information

(Polydoropoulou and Ben-Akiva 1999) have described a number of successive stages that travelers typically go through before they become regular ATIS users. These are:

· awareness, where the traveler begins to have basic information about the availability and attributes of a travel information system;

· consideration set formation, where the traveler generally begins to think of ATIS as a possible option to consider before making trips;

· choice set formation, where ATIS is definitely included as an option to assess in response to a specific identified travel need;

· trial use, where the traveler decides to try ATIS to gain more familiarity with its characteristics and potential benefits and costs;

· repeat use, where ATIS is assimilated into a traveler’s continued or habitual travel behavior, although further experience may cause the continued use to be reconsidered.

At the point where repeat usage becomes established, it becomes possible to speak of a systematic traveler response to real-time information. These responses are divided here into two general categories: those involving the tripmaking context, and those involving tripmaking itself. The sections below discuss these responses, drawing on the literature review to indicate the extent of current qualitative and quantitative knowledge about the responses.

Responses to ATIS involving the tripmaking context include behavior that affects the way that trips are scheduled or integrated into daily activities. These include adjustments to residential and/or employment location decisions; adjustments to daily activity schedules; changes in habitual tripmaking behavior; effects on non-travel activities; and trip-related stress or anxiety relief.

Responses to ATIS involving tripmaking itself cover a wide range of trip-related decisions: the decision to travel or not; the choice of destination or destinations (trip chaining); choice of departure time, mode and route; the re-routing decision in response to an incident; driving behavior; and the choice of parking location.

These various possible responses are discussed individually below, despite the fact that in many cases the responses are inter-related. The discussion also examines a number of specific examples of traveler response that merit separate consideration; these include ATIS impacts on shopping trips, transit information systems, variable message signs, and driver compliance with prescriptive information.

It will be seen that, in most cases, very little quantitative information is available. The available information tends to be highly specific to particular situations; very few quantitative conclusions of a generally applicable nature can yet be drawn regarding user responses to ATIS. This is not entirely surprising: significantly research into and deployment of ATIS has only been taking place for the past decade or so. Highways and transit systems were in use for many decades before generally reliable data and models on traveler response to them began to be developed. The pace of research and investigation is faster now, and the methods of data collection and analysis more efficient and sophisticated. Still, the current state of knowledge provides at best general qualitative conclusions regarding traveler response to ATIS. More deployments, more experience with deployed systems, and more research and analysis will be required to move ahead.

3.2.1 Trip context responses to ATIS

3.2.1.1 Reduce stress and anxiety

Many surveys have found that tripmakers appreciate having travel information available even if they do not or cannot modify their tripmaking behavior in any way because of it. Some analysts see this reaction as similar to peoples’ appreciation of weather forecasts. Respondents typically claim that the information reduces the level of anxiety or stress associated with not knowing what travel conditions are going to be. (Khattak, Schofer et al. 1995) and (Khattak, Yim et al. 1999), for example, discuss survey results where users mention this reaction.

(Lee 2000) has attempted to make the notion of travel stress relief more precise by arguing that the value of time spent in travel includes at least two distinct components: the opportunity cost of the activities foregone by traveling, and the disutility of the travel experience itself. This disutility is likely to be higher when a lack of information about travel conditions ahead causes one to be anxious or under stress; conversely, receiving travel information may make one more “serene” during a trip. The value of time spent traveling is likely to be higher in the former case than in the latter, and the benefit of the stress-relieving impacts of ATIS can be estimated as a function of the difference in value of time and the total time spent traveling.

3.2.1.2 Affect non-travel activities at the trip endpoints

Travel information may enable tripmakers to beneficially adjust the activities that they undertake at the departure or arrival ends of a trip. A person stuck in traffic may be able to call ahead with an accurate arrival time estimate and, before arriving, re-arrange her schedule at the destination to minimize the impacts of the delay on other activities. A person who wants to complete a task at one location but also needs to arrive at another location on time may be able to make use of accurate travel time information to determine if there is sufficient time to complete the task before departing. In the absence of such information, the person may abandon the task even if there was enough time to complete it; or complete it, and arrive late at the next location.

A Mitretek study ((Shah, Toppen et al. 2001; Wunderlich, Hardy et al. 2001); see also (Shah, Wunderlich et al. 2001)) provides evidence from simulated yoked driver experiments involving the Washington DC and Minneapolis/St. Paul metropolitan areas that pre-trip ATIS can significantly reduce the early and late schedule delays, and reduce the number of late arrivals. These studies compared the travel time and arrival time reliability of pairs of simulated drivers with identical origin, destination and desired arrival time at the destination. One driver was assumed to have access to pre-trip ATIS information on link travel times, and the other not. (The link travel time information was empirical data, compiled by polling an on-line traffic information service for conditions at five-minute intervals over a large number of days.) Drivers without access to information were assumed to base their path and departure time choices on average link conditions experienced over time, while those with access were assumed to utilize the “real-time” (but non-predictive) link times to make these decisions. In each case, the consequences of the decisions, in terms of travel and arrival time, were determined by reference to the compiled data on actual link times. (Compiled values were slightly perturbed to account for the variability in the time estimates.)

The study found that pre-trip information had only a minor effect on the average travel times experienced by its users. However, ATIS users reduced their number of late arrivals by 62%, and the total late schedule delay by 72%. (These benefits varied significantly by time of day.) The conclusion is that pre-trip ATIS is likely to impact travel time reliability much more than travel time itself. The study also suggests that, in the travel contexts considered, pre-trip ATIS is more likely to produce departure time changes than path choice changes.

3.2.1.3 Adjust daily activity schedule

People schedule the activities that they need to accomplish in a day based in part on the time taken by each activity and the time required to travel between activities in different locations. Because of uncertainty about travel times, people tend to incorporate “slack” in their scheduling decisions to reduce the risk of schedule disruptions due to worse-than-expected travel conditions.

Reliable information on travel times and traffic conditions will allow people to eliminate some of this slack. The time freed up in this way could be used in a wide variety of ways. At one extreme, it could be used to sleep or relax more; at the other, it could lead to a significantly different organization of the day’s activities including new activities and shifts in the order of activities. In terms of tripmaking, the additional time could lead to new trips, to trips made at different times, or to trip chaining.

Although these kinds of behavioral adjustment are entirely plausible, there is as yet very little evidence that they have occurred among users of currently-deployed ATIS.

3.2.1.4 Adjust habitual tripmaking behavior

There is considerable evidence that tripmakers rely to a large extent on habit when making their travel decisions. Over time, they establish a set of default behaviors that influence their tripmaking behavior on particular trips. These default or habitual behaviors do not necessarily dominate the decision-making process; rather, their effect is to increase the likelihood that, in any particular decision context, the default choice will be made. (Aarts, Verplanken et al. 1997) provide an analysis of bicycle use by students that supports this view.

(Uchida, Iida et al. 1994) surveyed commuters in a three-route corridor in Osaka, Japan following the installation of a VMS network that provided predicted travel time information. They identified two types of response to the information:

· tactical response, meaning the immediate decision to divert or not based on reported travel times for the three routes; and

· strategic response, the change over time in drivers’ selection of their habitual route.

The VMS was found to significantly affect both types of response. However, decision inertia was also found to be important in both. In the case of the tactical response, drivers showed a reluctance to switch away from their habitual route, other things being equal. In the case of the strategic response, drivers were reluctant to change their habitual route, even when the VMS repeatedly showed it to be an inferior alternative.

(van Berkum and van der Mede 1998) present a sophisticated modeling and analysis framework that accounts for the effects of ATIS in immediate travel decision-making and longer-term habit formation and change. The article presents empirical results that support their framework and highlight the importance of habit in tripmaking behavior. Similar results are presented, in another problem situation, in (van Berkum and van der Mede 1999).

One potentially important factor not considered in these studies is the possible effect of ATIS-produced changes in the daily activity pattern on the formation of travel habits. If, as was discussed in the preceding section, accurate travel condition information from an ATIS leads to a reorganization of a persons’ daily activity pattern, it is probable that habitual travel behavior will also change as a result.

3.2.1.5 Adjust residence and/or employment location

The variety of changes brought about by ATIS in the tripmaking context could lead people to reconsider their decisions regarding residential and/or employment location. As one example, if more predictable travel times became available from an ATIS, households could move farther away from job locations while still maintaining the same average commute time. Again, rearrangements in daily activity schedules brought about by ATIS could allow more time for outdoor activities, and incite households to take advantage of this by moving. Through these kinds of effect, ATIS could ultimately have an impact on urban form and structure. (Boyce 1988), in an early paper, evoked this possibility. (Hamerslag and van Berkum 1991) presented a simple network model that exhibits such location decision effects. However, it is likely that ATIS deployment on a much larger scale than today’s will be required before such effects become noticeable or significant.

3.2.2 Tripmaking responses to ATIS

3.2.2.1 Decision to travel or not

Relatively little information is available regarding the effects of ATIS on the decision to travel or not; however, it is not inconceivable that information about sufficiently bad travel conditions could induce tripmakers to cancel their intended trips, particularly discretionary trips.

(Khattak, Yim et al. 1999) cite evidence for this effect from CATI and mail questionnaire surveys carried out as part of the San Francisco-area TravInfo project. The surveys covered automobile and transit travelers and commute and non-commute (e.g., shopping or personal) home-based trips. The surveys asked respondents about the effects of pre-trip travel information (available from television, radio or telephone sources) on their tripmaking decisions. Analysis of the survey results revealed a number of general aspects of traveler response to the available information sources, some of which are discussed in sections below.

One of the findings was that non-commuters would occasionally decide to cancel their (presumably discretionary) trips because of unfavorable travel conditions reported by the various information sources, and particularly by radio. It is widely agreed that the demand for non-commuting trips is relatively elastic with respect to travel times and costs – in other words, an increase in travel times or costs leads to a reduction in tripmaking. In view of this, it is not surprising that information about bad travel conditions would lead, at the individual level, to non-commute trips being canceled. However, this is the only empirical evidence that was encountered in the literature review for such an effect.

3.2.2.2 Choice of destination or destinations

Similarly, relatively little information is available in the literature regarding the effects of ATIS on destination choice, or on the decision to visit several destinations and accomplish several purposes in one trip through trip chaining. Trips offering a choice of destination alternatives are likely to be for shopping or personal purposes, rather than for commuting. The opportunities to group multiple purposes and destinations into a trip chain are more varied and difficult to characterize and analyze.

The effects of ATIS on shopping trip destination choice was investigated in a set of internet-based stated preference surveys by (Kraan, Mahmassani et al. 2000) and (Mahmassani, Huynh et al. 2001). In the survey, respondents were asked to make a (simulated) shopping trip from a central location in Austin, Texas to a major suburban mall. Different pre-trip and en route messages about travel conditions were provided in the course of the decision-making process. Following notification of a change in traffic conditions while en route, the respondent was given the options of continuing on the same route; continuing to the same mall but via a different route; or switching to a different shopping mall entirely. Appropriate information was provided in each case. A sequential decision framework was developed to capture the conditional nature of the choices. It was found that the decision to switch route or destination was not influenced by age, gender, education and income. Respondents who were less familiar with the Austin area were more likely to switch destination, but not route. Those who visit the same mall on a frequent basis were less likely to switch destination and route. In general, switching response was greatest when information on traffic delays (as opposed to other kinds of traffic data) was presented.

Again, these are the only references located during the literature search on the topic of destination choice and trip chaining impacts of ATIS. Indeed, these questions are not well covered in the broader transportation literature; data on trip chaining, in particular, is difficult to collect and analyze.

3.2.2.3 Departure time choice

Departure time and route choice are often considered together in discussions of travel behavior. Many surveys of pre-trip user behavior collect data on both types of decision. They are considered separately in this discussion of ATIS because route choice can potentially be influenced by both pre-trip and en route information, whereas departure time choice is by its nature a pre-trip decision only.

There are a number of indicative data elements regarding the influence of ATIS on departure time choice but, again, the available data is not complete enough to draw broadly general conclusions or to develop widely applicable models.

An early study of commuting behavior (Mahmassani and Chang 1985; Mahmassani and Chang 1986) gave some indication of the slack that commuters feel they need to build into their departure times. Around 40% of survey respondents stated that they schedule their commute trip to arrive at work at least 15 minutes before the official start time; furthermore, the early schedule delay was found to increase with increasing distance from work. This suggests that travel time variability influences the departure time decision, and that commuters leave their homes early in order to reduce the risk of late arrival from longer-than-expected travel times.

(Barfield, Haselkorn et al. 1989) (Haselkorn, Barfield et al. 1990) (Mannering, Kim et al. 1994) discuss results of surveys of Seattle-area commuters who receive travel information from radio, television and telephone services. Of the commuters surveyed, 40% indicated that they had some flexibility in scheduling and selecting the route for their morning commute trip; 23% indicated no flexibility. However, 64% responded that they rarely changed their departure time because of pre-trip information.

(Khattak, Schofer et al. 1991) and (Khattak, Yim et al. 1999) report that the perceived accuracy of pre-trip reports is important in determining whether commuters take account of it in their decision-making. The importance of perceived pre-trip accuracy was also reported by (Polydoropoulou and Ben-Akiva 1999) based on analyses of San Francisco commuter surveys.

(Srinivasan and Mahmassani 2001) investigated using travel choice simulators the mechanisms by which drivers arrive at a departure time decision based on ATIS messages. They hypothesized that a driver undertakes a sequence of decisions to arrive at an adjustment to her habitual departure time. First, the driver decides whether or not to adjust the habitual departure time. Conditional on the decision to adjust, departure time alternatives are evaluated sequentially in about five minute increments. The directionality of adjustment (i.e., towards earlier or later departure) is governed largely by the direction of schedule delay experienced on the preceding day, with an earlier switch following prior lateness and vice versa. The results illustrate that the observed departure time adjustment behavior is influenced by dynamic transportation system attributes encountered such as trip time variability in the network, trip-makers’ short and longer term experiences, and the nature, type and quality of real-time information supplied by the ATIS.

3.2.2.4 Mode choice

Relatively little detailed information is available about the mode choice impacts of ATIS, although there is some evidence for this effect.

As reported in (Yim and Miller 2000), less than 1% of the early callers to San Francisco’s Travinfo service asked to be rerouted to the transit menu after learning about bad traffic conditions from the traffic menu. However, as experience with the system increased over the duration of the Travinfo field test deployment, it was found that up to 5% of the callers asked to be rerouted to the transit menu, a significant increase. Of those who accessed transit information, 90% of them chose transit for their travel mode. (Of course, a large fraction of the callers probably consisted of habitual transit users; it cannot be concluded that the information that they received caused them to choose transit.)

(Polydoropoulou and Ben-Akiva 1999) (see also Khattak, Polydoropoulou et al. 1996) discuss an analysis of San Francisco data that showed that prescriptive recommendations to take public transport have a detectible effect on mode choice, particularly in situations of unexpected delay.

3.2.2.5 Route choice

Many surveys and travel choice simulator studies have demonstrated the ability of ATIS to influence route choice. (Khattak, Yim et al. 1999), for example, presented survey results in which over 50% of respondents reported that they had made travel route or departure time changes in response to pre-trip information received by radio, television or telephone. (Owens 1980) describes an early travel choice simulator study that demonstrated drivers’ willingness to divert in response to highway advisory radio (HAR) messages about incidents. Some researchers have estimated sophisticated econometric models of route choice or route switching probabilities in response to ATIS, for example (Uchida, Iida et al. 1994) and (Polydoropoulou and Ben-Akiva 1999).

However, as stated above, from these various surveys and modeling efforts it is difficult to extract generally applicable quantitative conclusions regarding traveler response to information. The state of knowledge does not yet allow the development of a general model capable of predicting that, on a given network, X% of drivers will divert to route Y if they receive message Z while driving. Unfortunately, sufficient experience with and data about these systems is still lacking. Accordingly, this section will focus on qualitative conclusions that have been obtained from the various analysis efforts that were alluded to above.

Based on analysis of driver route choice responses to both VMS and radio information, (Emmerink, Nijkamp et al. 1996) have suggested that some people have a natural propensity to use traffic information of any kind and from any source. (See the discussion in Section 3.1 above.) Nonetheless, there is considerable evidence that the nature of the guidance information, and the conditions experienced prior to its dissemination, can strongly affect driver route choice response to it.

(Khattak, Schofer et al. 1995) and others have found that drivers tend to prefer messages that are descriptive (information about traffic conditions) rather than prescriptive (route recommendations). They found in particular that drivers are most receptive to near term predictions of traffic conditions on congested routes with rapidly changing conditions.

However, drivers’ perception of the accuracy and reliability of the messages is a key determinant of their response. (Kantowitz, Hanowski et al. 1997a; Kantowitz, Hanowski et al. 1997b) have found that there exists an accuracy “threshold”, beneath which drivers will simply ignore ATIS messages. Factors that increase drivers’ confidence in the accuracy of the messages tend to increase the likelihood that the drivers will react to them. In the context of route choice, such factors include a driver’s own observation of congestion prior (and particularly just prior) to receiving the message, and favorable experiences with the ATIS in prior uses.

Prescriptive messages do generally have an effect on route choice, as shown in many travel choice simulator studies and surveys of driver behavior. Combining a prescriptive recommendation to change routes with descriptive information justifying the recommendation has been found in travel choice simulator experiments to result in the highest route switching compliance rates. More generally, (Polydoropoulou and Ben-Akiva 1999) found that, in en route switching situations, the switching rate increased with the elaborateness (level of detail, care in justification) of the guidance messages.

(Owens 1980) found that drivers who received prescriptive information about incident diversion routes were generally more successful in avoiding the incident than those who received descriptive messages only. The success of the latter drivers depended strongly on their knowledge of the network around the incident. However, he found that the travel costs incurred by the two sets of drivers in diverting were not notably different.

(Llaneras and Lerner 2000) also investigated the ability of drivers to translate guidance messages into effective route choices. They considered “simple” and “enhanced” in-vehicle ATIS capabilities; the latter provided basic descriptive and qualitative information on incidents and congestion, while the latter provided the simple information as well as details about incidents, alternate routes, and real-time congestion conditions as well. Overall, drivers were able to use both types of system to divert around incidents. However, he also found that drivers sometimes made incorrect route choices with both types of system. The prevalence of these errors was significantly higher with the basic system; furthermore, the mistakes made with that system were generally more costly (in terms of excess delays) than those made with the enhanced system.

A number of generally idiosyncratic factors also condition a driver’s route choice response to ATIS messages. A freeway bias has been observed in several studies ((Hato, Taniguchi et al. 1995; Kitamura, Jovanis et al. 1999)). Because of this bias, drivers receiving messages that suggest diverting from a non-freeway to a freeway facility are considerably more likely to switch than drivers who receive the opposite message, other things being equal.

The influence of habit or inertia on route choice response has also been noted in a number of studies ((Uchida, Iida et al. 1994) (Hato, Taniguchi et al. 1995) (Srinivasan and Mahmassani 2000b)). Drivers tend over time to establish a preferred route for particular trips. Guidance messages that suggest switching from the preferred route to another are less likely to be accepted than messages that suggest the opposite. Of course, habit does not always over-rule information received from guidance messages. A sufficiently strong message, corroborated by the driver’s observations and confidence in the ATIS, will be considered. Over time, the effect of ATIS may be not only to affect particular route choice and switching decisions, but in fact to change the habitual route choices themselves.

3.2.2.6 Incident diversion response

A special case of the route choice decision occurs when a driver becomes aware, during the trip, of an incident affecting traffic conditions on the path currently being followed. Incident-related and other non-recurrent congestion is a major contributor to total congestion delays on highway networks; for example, it has been estimated that roughly half of all delays on freeways in the U.S. are due to non-recurrent causes. Driver response to an incident situation determines in part the severity of its consequences. It is expected that ATIS can be of considerable help in incident situations by providing drivers with timely information about the location and characteristics of the incident and by suggesting routing alternatives in what are, by their very nature, unexpected and unfamiliar circumstances.

The two key aspects of driver incident response are: whether the driver diverts at all; and, if the driver diverted and avoided the incident, whether she returns to the original route or continues on the alternate route. In the former case, the route switch represents a temporary detour around the cause of delay; in the later, the route switch entails choosing a completely new route to follow to the destination. The choice considerations at work in these two situations may well be different.

A number of studies have examined drivers’ route diversion behavior in the presence of non-recurring congestion, applying a variety of methodologies for this purpose. This is actually one of the better-studied aspects of traveler response to ATIS, perhaps because of the natural interest in applying ATIS to alleviate incident conditions.

(Khattak, Koppelman et al. 1993) investigated factors that influence auto commuters’ en-route diversion propensity. Data on propensity to divert and related factors were collected through a stated preference (SP) questionnaire survey. The effects on drivers’ willingness to divert of incidents and recurring congestion, real-time traffic information, driver and roadway characteristics and situational factors were investigated using conjoint measurement.

Disaggregate discrete choice models are a natural approach for investigating drivers’ diversion and return choices. Multinomial logit models (MNL) and nested logit models (that remove the undesirable IIA property of MNL) are logical model forms. (Khattak, Schofer et al. 1993) examined diversion and return choices using these two forms. The model structure represents these choices as interrelated to take account of the likelihood that drivers’ diversion choices will depend, in part, on their expectation that they will or will not return to the original route. That is, the driver chooses among three alternatives: no diversion (ND), diversion and no return (DNR), and diversion and return (DR). The authors used a joint multinomial logit model of the choice among these three alternatives and a nested logit model in which the return choice is nested within the diversion decision. Both these models were estimated with equivalent systematic utility function specifications; they yielded very similar coefficient values (i.e. identical behavioral interpretation). Commuters’ diversion and return behavior varied with their personal characteristics and with the characteristics of the trip they were making at the time when the choice arose. Individuals making longer trips, facing longer delays and facing less expected congestion on alternate routes were more likely to divert. Commuters who made longer trips were significantly more likely to return after diversion.

(Abdel-Aty 1998) also considered alternative logit model forms to model the three diversion options (ND, DNR and DR), in an investigation of the preferred modeling structure for the incident-related routing decision. In addition to the joint multinomial logit, two nesting structures were tested. In one of them, the DR and DNR choices were modeled under a “diversion” nest, reasoning that these choices are conditional on diverting because diversion has occurred. The other specification places the ND and the DR choices under a “maintain route” nest, reasoning that the choices are conditional on staying on the same route because the majority of the route will be the same. It was concluded that the nested logit model with the ND and DR choices grouped in a nest provided the best structural fit for the observed distribution of the routing decision in case of an incident. The superiority in this application of a nested logit structure over the simple MNL form was also established.

Use of ordered categorical response data is very common in these kinds of modeling, where the bulk of the data is obtained through stated preference questionnaires. Use of multinomial logit or probit models or linear regression may lead to biases in estimation using this kind of data. (Khattak, Koppelman et al. 1993) estimated multivariate models of diversion propensity to explore the effects of several variables simultaneously. The multivariate model used was an ordered probit model with diversion propensity as a function of delay characteristics, reported trip and route attributes and socio-economic characteristics of the respondent drivers. The ordered probit model was selected for estimation because of its ability to analyze ordered categorical response data.

Another method to investigate drivers’ route diversion behavior is to analyze reported and stated data about route diversion obtained through surveys. (Khattak, Kanafani et al. 1994) analyzed a survey of commuting behavior in the San Francisco Bay Area in 1993. The questionnaire was designed to use reported diversion behavior (a measure of the true behavior) as the basis of a sequence of stated preference questions about the propensity to divert with a future in-vehicle ATIS device. This methodology increases the validity of the stated preference technique by relating the response to ATIS technology to a specific behavior that was actually practiced by the respondent. The objective of the stated preference question was to determine how incremental amounts of information provided by an ATIS device would influence the propensity to divert. It appeared that respondents overstated their propensity to divert when compared with reported behavior. Around 22% of the respondents stated that they would divert even though they reported not having diverted. On the other hand, only 5% of the people stated that they would not divert even though they actually diverted when they faced the unexpected delay. To explore the correlation between reported behavior and stated preference, a linear regression model relating the answers to each question was developed.

3.2.2.7 Driving behavior

Traveler information can be used not only to improve trip-related decision-making, but also to influence driving behavior during the trip.

For example, messages might warn drivers before they arrive at hazardous road conditions (road work, accidents, bad weather) so that they drive more cautiously. (Ng and Mannering 2000) report on vehicle simulator experiments to determine the effectiveness of such advisory messages. They developed a very realistic simulation of an actual mountain road in Washington State, and included the ability to generate fog and place snowplows in the simulation. They considered the effect on driving speed of VMS messages, in-vehicle messages and both; messages warned about the presence ahead of fog, road curves, and snowplows. They found that over short distances, the messages did cause drivers to reduce their speeds; however, over longer distances there was no noticeable speed effect. This suggests that after slowing down in response to the message, drivers drove faster in order to compensate for the delay.

(Smulders 1990) describes a subtle application of this idea. He found that merely suggesting appropriate freeway speeds to drivers by VMS – but with no obligation on the part of the drivers to comply – had a small but noticeable effect on average travel speeds but significantly reduced the variability in these speeds across drivers on the facility. This reduction in speed variability considerably delayed the onset of the breakdown of traffic conditions at maximum flow levels, and actually increased the capacity of the freeways where the method was applied. Speed advisory VMS are now deployed on a number of freeways in the Netherlands.

3.2.2.8 Parking search and choice

Parking guidance and information (PGI) systems inform drivers about the availability of parking at various locations or recommend facilities for use.

In general, a PGI system consists of four components:

· a counting mechanism at parking facilities to track vehicle entries and exits and thus determine facility occupancy and available spaces at a given time;

· a control center that processes data on facility occupancies and generates messages about parking availability or recommendations. Messages may also include information about other attributes of parking facilities (prices, location, etc.);

· a communications network that transmits occupancy data from facilities to the control center and disseminates messages from the control center to users;

· information access technologies by which users obtain the messages generated by the control center.

The information access technology generally consists of a system of variable message signs, arranged so that traffic traveling towards the city center receives progressively more detailed data with each VMS encountered. The messages may be based either on current occupancies or on the occupancies predicted to hold at the time a vehicle passing a VMS actually arrives at the parking facility. Occupancy data may be quantitative (actual spaces available) or qualitative ("ample space", "nearly full", "full").

A number of such systems are in use in cities around the world. In some instances, both the parking facilities and the PGI system are operated by the municipal government, but this is not a requirement. In England, for example, there are arrangements where privately-operated parking facilities provide data to a PGI system run by the local government. The hardware needed to implement the system components is commercially available.

(Allen 1993) provides a useful summary of the benefits of PGI systems. These include:

· benefits to drivers by being able to proceed directly to a parking facility with available spaces, without having to spend time searching and waiting;

· benefits to traffic and environmental conditions from the elimination of parking search traffic which, according to some estimates, can be 30% or more of all traffic on roads in city centers;

· benefits from more efficient utilization of available facilities: higher occupancy levels and increased parking revenues;

· benefits from information availability about facility usage, making possible better management of the parking system (e.g., fraud monitoring, pricing policy analyses).

Many of the issues that arise in modeling general driver response to traffic-related information also occur in modeling response to parking-related information. It appears from a review of the literature, however, that parking choice and PGI systems have been less intensively investigated to date than route choice and ATIS.

Among the articles reviewed, (Teng, Falcocchio et al. 2001) surveyed parking facility users in New York City to determine the types of information they considered most useful in a PGI system, and to investigate relationships between user or trip characteristics and the ranking of information types. For a parking information web site, the information of greatest interest included fee structure, hours of operation, location, the predicted probability of having a space available at the time of arrival, and traffic conditions in the vicinity of the facility. For roadside displays, the information of greatest interest included hours of operation, number of available spaces, location and fee structure. These preferences were observed to vary by gender, trip purpose, and familiarity with parking options and conditions, among other factors. Internet-based information and in-vehicle devices were preferred to a kiosk for obtaining pre-trip information, while VMS were preferred to in-vehicle devices for obtaining en route information.

(Allen 1993) conducted stated preference surveys in an outer borough of London to investigate the effects on parking facility choice of VMS message, parking price and walk time to the destination. The survey concentrated on weekday shopping trips and distinguished three different user groups. The considered messages provided qualitative information on the occupancy of different nearby parking facilities. Two message dictionaries were considered in the SP experiments, differing most notably in that one explicitly identified facilities as "nearly full" while the other displayed a blank message for such facilities. The authors present multinomial logit model estimation results. Within the range of prices and walk times considered, the displayed message had a determining effect on parking facility choice, while price and time had secondary effects. It is not clear, however, if these conclusions would hold over a larger range of prices and times. It should also be noted that this work did not consider the practical problem of how to indicate parking facility locations to unfamiliar drivers using a VMS with very constrained message space.

The incorporation of parking choice and PGI systems in general-purpose traffic models is considered by (Chatterjee and Hounsell 1999), with specific reference to the dynamic traffic model RGCONTRAM. The authors show how parking-related movements and the associated times and costs can be represented as special links in a traffic network model. They discuss the application of a travel choice simulator to investigate joint route and parking facility choice with and without PGI messages. They describe simulator experiments that varied parking prices, expected risk of waiting to park, waiting times and PGI messages, but do not present specific model specifications and estimation results. However, regardless of the particular form of a parking choice model, it is clear from the author's discussion how an information-based traffic model (i.e., one that allows en route path diversions based on messages) could represent and integrate parking information and choice as well.

Since parking search traffic is a poorly-understood but potentially significant component of city center traffic, it would seem that further research on driver choice of parking facility and driver response to PGI messages would be fully justified. Research results could be incorporated in information-based traffic models without requiring extensive modifications. With relatively little lead time after the research results became available, the resulting model systems could be applied to the practical analysis of parking search traffic and its impacts, and ultimately to the design of PGI systems.

3.2.3 Specific systems and examples

This section discusses a few specific examples of driver response to information. The examples were chosen either because of their intrinsic interest, or because a considerable amount of information is available about them, thus allowing a more detailed discussion than was usually possible in the preceding section.

3.2.3.1 Variable message signs

Variable message signs (VMS) have been widely installed for freeway traffic management in most metropolitan areas. VMS are electronic message boards located in the close proximity to roadways. They represent a cost-effective mechanism to display short real-time messages to drivers approaching them. Of course, their effectiveness in real-time traffic operations is highly dependent on user response to the displayed information. A compounding factor is that, unlike an in-vehicle navigation system that can provide personalized routing information, VMS are constrained to display generic information to all nearby drivers. It follows that seemingly minor details of the displayed message may have a considerable impact on system performance.

This provides motivation to study the relationship between VMS messages and user response. A few studies have investigated this relationship. (Peeta, Ramos et al. 2000) examined the effect of different message contents on driver response under VMS. The issue was addressed through an on-site stated preference (SP) user survey. Logit models were developed for drivers’ diversion decisions. The analysis suggested that content and level of detail of relevant information are factors that significantly affect drivers’ willingness to divert. Other significant factors included socioeconomic characteristics, network spatial knowledge, and confidence in the displayed information. Results also indicated differences in the response attitudes of semi-trailer truck drivers compared to other travelers. These results provide substantive insights for the design and operation of VMS-based information systems.

A somewhat similar study was performed by (Wardman, Bonsall et al. 1997), also using a stated preference approach to undertake a detailed assessment of the effect on drivers' route choice of information provided by a variable message sign (VMS). Although drivers' response to VMS information will vary according to the availability of alternative routes and the extent to which the routes are close substitutes, the research findings showed that route choice can be strongly influenced by the provision of information about traffic conditions ahead. This has important implications for the use of VMS systems as part of comprehensive traffic management and control systems. The principal findings were that the impact of VMS information depends on: the content of the message, such as the cause of delay and its extent; local circumstances, such as relative journey times in normal conditions; and drivers' characteristics, such as their age, sex and previous network knowledge. The impact of qualitative indicators, visible queues and delays were examined. It was found that not only is delay time more highly valued than normal travel time (which is to be expected) but also that drivers become more sensitive to delay time as delay times increased across the range presented.

Most disaggregate-level studies of drivers’ response to VMS use stated preference (SP) survey data rather than actual traffic data. It is generally not possible to infer from traffic measurements the effects of a VMS on individual driver behavior, since the drivers’ intentions prior to receiving the messages are not usually known. One study encountered during the literature survey used aggregate traffic data. (Yim and Ygnace 1996) used loop detector data from the Système d'Information Routière Intelligible aux Usagers (SIRIUS) information network in Paris to investigate the effects of VMS on link flows. Time-series traffic data were analyzed to measure changes in mean flow rates at a selected link. It was found that variable message signs influence drivers to choose less congested routes when the drivers are provided with real-time traffic information, and that a driver's decision to divert is closely associated with the information pertaining to the level of congestion. In the Paris region, drivers received prevailing queue length information from the VMS. According to the data analysis, a reported queue length of 3 km seems to be a threshold at which a significant number of drivers choose to divert to an alternative route.

3.2.3.2 Compliance with prescriptive guidance

The question of user compliance with ATIS messages arises when those messages consist of prescriptive recommendations about pre-trip departure time or route choices, or about en route path switching decisions. It is of considerable interest to understand the factors that influence whether or not a driver will follow the recommendation, both as a means towards better design of prescriptive ATIS messages (to ensure higher compliance), as well as to model more accurately the effects of such messages at the individual or the system level.

Given basic traffic data on travel times or other measures of network conditions, either descriptive or prescriptive messages could equally well be generated from them. However, it does not follow that drivers' responses to these two different types of messages would be identical. The format and content of the two types of messages would necessarily be different, and could well elicit different reactions from drivers.

Descriptive guidance is in some sense more "neutral", in that it simply conveys information about network conditions, which drivers will interpret as they wish and are able. In contrast to this, prescriptive guidance is a specific recommendation to do a particular thing; drivers may question whether the recommendation is based on sufficiently reliable data, on decision-making criteria consistent with their own, or indeed on a knowledge of the network equal to their own.

On the other hand, prescriptive guidance may potentially provide a traffic control center with a more direct influence over drivers' tripmaking decisions and so on network-level traffic conditions. A considerable amount of underlying traffic data may be efficiently synthesized in the form of a simple recommendation to drivers. Particularly under incident situations, a center may feel it appropriate to intervene aggressively in drivers' choice processes in order to minimize avoidable traffic impacts and to restore normal conditions as rapidly as possible.

It should be mentioned that the distinction between descriptive and prescriptive guidance messages is not an absolute one. Indeed, as will be seen below, there is considerable evidence that the most effective ATIS messages combine descriptive and prescriptive aspects: information that describes a traffic situation together with recommendations that suggest an appropriate reaction. The information explains or justifies the recommendation in some sense, and drivers are more likely to comply.

The most common source of prescriptive guidance currently in operation is variable message signs. These may be used to suggest routes to drivers based on broad destination locations ("take route XYZ for points north"). The limited space available for message display is a major constraint, and the messages must be carefully designed to be clear and understandable. (Bonsall and Palmer 1999) discuss various aspects of VMS message design, and (Summala and Hietamaki 1984) present an earlier study of factors influencing traffic sign effectiveness.

In-vehicle units have the possibility of making much more detailed and personalized recommendations, but are not yet in common use. An early prototype system of this type was Siemens' Ali-Scout system, which was used in West Berlin's LISB deployment (Bonsall and Joint 1991b) and also in Michigan's FAST-TRAK program. It is based around an in-vehicle device that provides a simple keypad for entering data, and outputs both visual (simple text and direction arrows) and audible (synthesized voice) messages to the user. System beacons are installed at key locations on the network; these both transmit data to the in-vehicle units, and receive from vehicles information on their recent travel times. At the beginning of a trip, when a user inputs his or her intended destination, Ali-Scout first indicates the general direction to follow based simply on compass direction. However, when the equipped vehicle passes a beacon, it receives real-time travel time information from which it can determine a minimum time path. The in-vehicle unit then provides detailed driving directions (direction to take at each intersection) until the vehicle arrives in the vicinity of the destination. At that point, the in-vehicle unit reverts to a compass-direction mode, since the density of beacons is not high enough for the system to be able to provide detailed local area directions.

It is difficult to obtain information on prescriptive guidance compliance rates from aggregate traffic measurements such as link volume counts. Determining whether a driver complied or not with a recommendation requires knowing what the driver's original intention was, and also depends on knowing whether a particular message is relevant to the driver's situation. Such information is not generally available at the aggregate level, although license plate survey methods and driver questionnaires have occasionally been successfully used for this purpose (Dudek, Weaver et al. 1978), (Richards, Stockton et al. 1978).

For this reason, most research on driver compliance behavior has been based on experiments with individual drivers using travel choice simulators. Travel choice simulators place experimental subjects in a decision-making situation and record their response. Travel choice simulators focus on decision-making related to travel behavior such as route and departure time choice. They are less elaborate than the (much more expensive) vehicle simulators that attempt to faithfully replicate all aspects of the driving experience. Rather, they provide only the key elements of a choice situation under study, with enough detail to establish the context and to motivate users to respond in a realistic fashion. A travel choice simulation experiment can be viewed as a kind of stated preference survey in which the hypothetical choice scenarios are presented in a rather realistic manner.

For their research in to VMS compliance, Bonsall and co-workers developed first the IGOR travel choice simulator (Bonsall and Parry 1990; Bonsall and Parry 1991) and then the more sophisticated VLADIMIR travel choice simulator (Bonsall, Firmin et al. 1997). Both of these were PC-based programs that allowed subjects to "drive" through a network from a given origin to a given destination, following a route of their choosing. During the "trip", the program displays information on local traffic conditions and, at decision points, may also provide ATIS messages. The user chooses how to proceed at each such decision point, and the program records each such decision along with data about the conditioning factors such as traffic conditions, messages displayed, and others. The experimenter can vary these factors from one run to another in order to investigate their effects on drivers' decisions. In VLADIMIR the display took the form of actual photographs of locations along the routes being driven, along with a simple sketch map of the nearby network, text describing traffic conditions and any ATIS messages, and basic information regarding the progress of the simulation (elapsed time, etc.) After careful comparisons of driver choices in the simulator with actual decisions by the same drivers in comparable situations on the network, (Bonsall, Firmin et al. 1997) concluded that the simulator was able to replicate driver behavior with a high degree of fidelity.

Similarly, Mahmassani and co-workers (Chen and Mahmassani 1993; Srinivasan and Mahmassani 2000b) used a travel choice simulator interfaced to the Dynasmart mesoscopic traffic model. The model represents 20 minutes of peak period traffic in a freeway corridor carrying roughly 11,000 simulated vehicles on three parallel facilities with several opportunities to switch from one to another. Experimental subjects (possibly several at a time) make departure time and route choice and switching decisions. These decisions are taken into account by the traffic model, which computes the traffic conditions that result from them (as well as those of the many simulated vehicles). The ATIS messages provided to the subjects are derived from the computed decisions, and so are consistent with those decisions (rather than being exogenously specified.) Strictly speaking, the messages are descriptive rather than prescriptive: they indicate the travel time to the (unique) destination on each of the three main alternative routes. However, in the simple context studied, the minimum time route is clearly the recommended one; the other considerations mentioned above that might affect compliance behavior do not come into play.

Based primarily on the results of travel choice simulator experiments, a number of general conclusions about driver compliance with prescriptive guidance have been obtained. Examples of such general conclusions include:

· drivers will reject prescriptive messages that they do not find credible. Factors affecting message credibility include the extent to which it is corroborated by local evidence, visible to the driver, about the alternatives and their conditions; and the quality of advice previously (and particularly very recently) received from the system;

· compliance is strongly affected by the driver's familiarity with the network. For a given prescriptive message, the compliance by drivers familiar with the network is generally about 10% less than that by unfamiliar drivers;

· compliance is highest for messages that combine information and recommendations; next highest for those that provide information only; and lowest for those that make recommendations only;

· one minute of delay mentioned in a VMS message has the same effect, in terms of affecting path choice decisions, as 1.75 minutes of actual delay in driving time;

· compliance is higher for recommendations about an immediate action than for vaguer advice about actions in the future. A recommendation that refers to a nearby problem location is more likely to be followed than one that does not;

· drivers have a certain reluctance to switch to a new route from one that they are already following. The reluctance is greatest if the recommended route seems to follow an alignment significantly different from that of the current route;

· socio-economic characteristics of the driver are also important influences on compliance. Among these characteristics are gender, age, level of driving experience, and (as already mentioned) degree of familiarity with the network.

A number of attempts have been made to model compliance behavior. (Srinivasan and Mahmassani 2000b) consider route choice behavior as influenced by both compliance and inertia mechanisms. The inertia mechanism reflects a driver's reluctance to modify a decision already made, while the compliance mechanism reflects a driver's tendency to follow (or to reject) routing advice. They specify and estimate multinomial probit route choice models that include these two mechanisms as latent variables, and conclude that the effects are significant. Simpler route switching models often include a dummy variable that penalizes routes if they are different from the one currently followed.

(Bonsall and Palmer 1999) discuss more particularly the modeling of driver choice of exit link at an intersection when guidance is provided. They estimate a number of simple multinomial logit models that incorporate variables such as travel time, message specific indicators (e.g., mention of accident or of road works), alignment of the exit link relative to current path, and so on. These models are intended for use in traffic simulation systems to predict the probability that individual drivers will proceed via the different possible exit links.

3.2.3.3 Shopping trips

Most studies on ATIS have focused primarily on commuting trips, which go to a fixed destination, tend to be repetitive in nature and involve tripmakers who are familiar with the transportation network. But it also of considerable interest to examine traveler response to ATIS during non-commute trips, where travelers have some flexibility in terms of destination choice and may not be as familiar with the transportation system.

(Mahmassani, Huynh et al. 2001) and (Kraan, Mahmassani et al. 2000) examined behavioral responses of non-commuters under real-time information during shopping trips. Utilizing results from an interactive stated-preference internet-based survey, the authors developed discrete choice models to investigate factors that influence en-route switching to alternate destinations and alternate routes during such trips. The fundamental difficulty in modeling this phenomenon derives from the manner in which information is provided to assist trip-making. The information provided and resulting user choices are interdependent. That is, the choice set presented to a tripmaker at a particular decision point is predicated on his/her previous decisions. Conversely, a tripmaker's decision in turn alters his/her subsequent information and choice sets.

The authors specified a model structure that overcomes this difficulty. It explicitly captures the conditional nature of the decision process. The model that they developed provides insight on en-route diversions during the shopping trip together with the factors affecting these decisions, especially with regard to the role of real-time information.

3.2.3.4 Transit information systems

Transit information systems provide transit users with static information on service such as routes, schedules, transfers and fares. They may also offer real-time information such as the anticipated arrival time of the next bus or train, and individualized information such as the route to follow or the expected travel time of a particular trip the user intends to make. Ultimately, transit information systems may offer their users a full range of trip planning, ticketing and real-time information services, integrated across the range of public transport modes; the Transport Direct system, currently under development in the U.K., ((Lyons, Harman et al. 2001)) is an ambitious step in this direction. (Casey, Labell et al. 2000 Section 3) provide a useful summary of the North American state of the art in transit information systems as of the year 2000. It is fair to say that currently deployed systems still have very rudimentary capabilities compared to their ultimate potential.

Real-time transit information can reduce the anxiety that users feel due to uncertainty regarding the duration of their wait. More generally, it may improve the quality of service perceived by transit users and likely increase transit’s retention of its current patrons. Furthermore, providing information may possibly change non-users’ attitudes toward public transit, and entice more travelers to use public transit.

A relatively limited number of studies have been undertaken to investigate the usefulness of these systems in attracting new transit passengers and improving the level of service of existing passengers. (Abdel-Aty 2001) used ordered probit models to study the effect of Advanced Traveler Information Systems (ATIS) on transit ridership. A computer-aided telephone interview was conducted in two metropolitan areas in northern California. The survey included an innovative stated preference design to collect data that address the potential of advanced transit information systems. The study's main objectives are to investigate whether advanced transit information would increase the acceptance of transit, and to determine the types and levels of information that are desired by commuters. The survey included a customized procedure that presents realistic choice sets, including the respondent's preferred information items and realistic travel times. The results indicated a promising potential of advanced transit information in increasing the acceptance of transit as a commute mode. It also showed that the frequency of service, number of transfers, seat availability, walking time to the transit stop and fare information are among the significant information types that commuters desire. Commute time by transit, income, education, and whether the commuter is currently carpooling, were factors that contributed to the likelihood of using transit following information provision.

Although such transit information systems are assumed to be of benefit, methods for evaluating these benefits under various conditions are limited. (Mishalani, McCord et al. 2000) developed a methodology that focuses on the potential benefits of bus arrival information systems to passengers waiting at bus stops under various supply and demand characteristics. Transit bus operations and passenger arrivals are modeled as a stochastic system where the operator uses real-time bus location data to provide to waiting passengers bus arrival time information that maximizes passengers' utilities. Simulation results reveal how the value of such information systems depends on the type of real-time data available to the operator, on bus operations characteristics, and on demand patterns. Results indicated that while the first two influence the value of information to passengers, demand patterns do not have a significant impact.

3.2.3.5 ATIS for maintenance and protection of traffic around construction zones

It is natural to think of applying ATIS to help manage traffic in and around construction zones. Such zones can create significant traffic disruptions. Because of their temporary and changing nature, most travelers will not be able to learn by experience what “typical” conditions are or how to avoid the most impacted areas. It is logical to suppose that providing real-time information to drivers in such circumstances would produce real benefits both to individual drivers and to network traffic conditions overall.

Surprisingly, there are very few examples of the use of real-time traffic information systems for construction zone traffic management in the U.S., and very little literature on the subject.

(Kratofil 2001) provides a brief but useful review of relevant literature. Based on his literature review, he then proposes a framework for quantifying the benefits of ATIS in construction zone traffic management, applying for this purpose a standard breakdown of ITS impacts into a number of such as mobility, safety, etc., and distinguishing between impacts to drivers, to the implementing agency, and to the community at large. He compares a “with” and “without” ATIS situation for a specific interstate highway reconstruction project using this framework. In most cases, quantification of the impacts of ATIS relies on values (for example, accident rate reduction impacts) that were derived for situations other than construction zone traffic management. He concludes his paper with a recommendation for collection of traffic data before and during the operation of the ATIS for construction zone traffic management, in order to begin accumulating quantitative results that could be useful for future design and evaluation efforts involving such systems. He also recommends the execution of surveys to better understand people’s usage and valuation of information from ATIS.

There is clearly considerable scope for ATIS MPT applications. Very little definite knowledge is available regarding either the design and operation of such systems, or traveler response to them.

3.3 What kinds of information do users want? How much will they pay for it?

In 1991, (Green, Sarafin et al. 1991) discussed the results of a study by a panel of experts of features that should be in driver information systems by the year 2000. To determine this, features were evaluated on the basis of three objectives that had been set by USDOT: (1) their effect on accidents; (2) their impact on traffic conditions; and (3) their fulfillment of driver needs. The analysis considered a very broad range of possible functions including communications, entertainment, office capabilities, way-finding, vehicle status monitoring, display of traffic signs and signals inside the vehicle, road hazard alerts, and traffic information. For each such function, the experts considered a variety of possible features that might implement the function. (In the entertainment function, for example, the possible features considered were cassette/CD player, radio and television.) Each feature was then ranked according to its contribution towards the stated objectives.

The five highest-ranked features were crash site hazard notification, in-car display of external traffic control signals, information on traffic congestion, indication of the presence of multiple compounding hazards in a driving situation, and information about road construction activities. All of these features are components of what we would now call an Advanced Traveler Information System, although some are still more advanced than is anything that has been prototyped to date. Features considered in the study which were given some of the lowest priorities, such as cellular telephone communications capabilities and radar detectors, are by now commonplace.

Considerable work since that time has attempted to identify user preferences for travel information system features. In this context, the term "features" refers to the different kinds and qualities of messages that might be provided by a traveler information system. By kinds of messages is meant the nature of the data provided in the messages – information on travel times or delays, location of incidents, specific route recommendations, etc. By quality of messages is meant their usefulness as it might be judged by a user – how up to date they are (their currency), their accuracy, precision, network coverage, the degree to which the message relates to the traveler's individual situation, and so on.

User preferences are obtained from various kinds of traveler surveys. In some cases, survey respondents are simply asked to express an opinion about various possible features: to state whether a feature is desirable or not, or to indicate the strength of their desire for the feature on an ordinal scale. Other survey methods involve placing the respondent in (hypothetical) situations where they must state their preference between alternative features, and so illuminate his or her tradeoffs between the features. Survey results may be analyzed by computing simple descriptive statistics or by estimating some form of econometric model. A number of these were discussed in the preceding sections.

(Llaneras and Lerner 2000), in a recent study of this type, compared user response to and preference for “basic” and “enhanced” ATIS services in the context of en route decision making; he used travel choice simulation experiments for this purpose. In these experiments, basic ATIS services consisted of descriptive information on incidents and congestion levels, and qualitative estimates of travel delays; enhanced services included all the basic services, but added information on alternative routes, various details about incidents, and a map display showing real-time traffic conditions. By analyzing the effectiveness with which users were able to translate the information received into travel improvements, the authors concluded that the following types of information were most valuable: data on incident location, type and delay; data on queue lengths; and recommendations about alternative routes, with directions to them. The real-time map display of traffic conditions was the information most frequently referred to by drivers in the experiments; however, human factors questions remain unsolved regarding the best way to present such information with minimal interference to driving.

When the survey choice situation involves both information features and money, it becomes possible to estimate an implicit willingness to pay for the feature, defined (in a utility based model) as the negative ratio of the marginal utilities of the feature and of money.[1] It must be emphasized that, to date, very few travelers have ever paid any money to receive travel information.[2] Answers about money in stated preference surveys are frequently biased because respondents know that they will not actually have to pay anything, regardless of what they say. Therefore, conclusions about willingness to pay for travel information are fraught with uncertainty, and the numbers obtained from such surveys should be interpreted in relative rather than absolute terms.

(Wolinetz, Khattak et al. 2001) list six broad factors that they hypothesize may affect travelers' willingness to pay for information:

· uncertainty: if there is little variability in traffic conditions from trip to trip, these is little need for real time traffic information. Non-recurrent congestion increases travel time uncertainty. Recurrent congestion, even through it is relatively more predictable, also adds uncertainty. Both of these effects may increase with trip length;

· information awareness: travelers who are aware of available ATIS services are more likely to express a willingness to pay for future services;

· access to information: individuals who are willing and able to access real-time information through communication or computing devices may be more likely to pay for ATIS services;

· information use: individuals who already receive travel information via phone, radio or other conventional sources may be more willing to pay for ATIS services;

· situational and contextual factors: such as trip purpose, departure or arrival time flexibility, trip chaining requirements, and many others;

· socio-economic factors: background variables such as age, gender, income and education may be important influences on the willingness to pay for ATIS.

Research at the University of Michigan Transportation Research Institute (UMTRI) (Wallace and Streff 1993) studied the stated rankings of different types of travel information by drivers on different kinds of trips (commute trips, trips in a familiar area and trips in an unfamiliar area). This research compiled descriptive statistics on respondents' rankings of the relevance of different types of information on the route choice decision. The researchers were particularly interested in the influence of the different information types in the en route decision to switch from one route to a different one. For commute trips and those in familiar areas, information on travel delays and travel time reliability on the original and alternate routes were ranked the most highly. For trips in unfamiliar areas, the availability of travel directions for the alternate route was a highly ranked consideration.

(Mehndiratta, Kemp et al. 1999a) (see also (Kemp and Lappin 1999)) surveyed drivers who had had significant experience with prototype in-vehicle navigation devices in three recent field operational tests. Drivers' preferences with regard to information update frequencies, network coverage and information personalization were investigated in a series of attitudinal and tradeoff questions. The survey results were analyzed in a number of ways, including by estimating logit-form models of preference probabilities as a function of information quality and price. In general, the authors found that the most basic improvements in information quality over currently-available sources (general radio traffic reports, for example) were highly valued, but that further information quality improvements exhibited a pattern of decreasing incremental utility.

Geographic coverage and update frequency were both important attributes; logit model coefficients for the minimal level of provision of both of these had approximately similar coefficients. With respect to geographic coverage, door-to-door coverage was perceived as having little or no incremental benefit compared to coverage of freeways and arterials. Similarly, information updates several times an hour were clearly preferred to static information, but the added value of nearly continuous updates was small to negligible. Personalized information provision was not highly valued.

Few respondents were indifferent to the type of guidance – prescriptive or descriptive – provided by the system; they strongly preferred either one or the other. A majority of all respondents preferred to receive descriptive information (delays), although about 20% preferred prescriptive route guidance. Where sample sizes were large enough to allow such investigation of gender-related effects, it was found that women were more likely than men to prefer prescriptive guidance.

Most respondents indicated some willingness to pay for real-time traffic information; few indicated that they would not pay anything. The estimated willingness to pay ranged from $8-$10/month in Seattle, from $28-$36/month in Chicago and from $8-$20/month in Boston, depending somewhat on the particular information types and qualities considered. These values are higher than what is generally expected from other, perhaps more informal, analyses of user willingness to pay.

(Wolinetz, Khattak et al. 2001) is another recent investigation into user preferences and willingness to pay for different types of travel information. The survey covered both automobile and transit users in the San Francisco Bay Area, and asked respondents to rank possible information features of a hypothetical traveler information system; it also included pricing questions. Survey analysis was based on the computation of descriptive statistics. The most desirable information content options were constant updates, alternate route information, in-car computer information, expected delay data and route time comparisons. Many respondents indicated a willingness to pay at least some positive amount for high-quality real-time traffic information. The majority of these people prefer to pay on a per-request basis (as opposed to a flat monthly subscription fee.) Most expressed a willingness to pay up to $1 per request.

(Tsai 1991) reports on the results of focus groups held with commercial vehicle operators (truckers and bus drivers) regarding their preferences for information about the highway environment: traffic and weather. Desirable features included in the traffic data were: information on traffic congestion, accidents, lane closures, bridge closures, construction updates, alternate routes, low bridges, road weight restrictions and legal truck routes. Truckers identified specific areas (generally around the largest metropolitan areas) where such information would be particularly useful. Weather information needs included: notice of adverse or severe weather conditions, fog conditions, and identification of areas experiencing black ice. However, the expressed willingness to pay for such information was quite low.

(Ng and Barfield 1997) report on surveys of ATIS feature requirements of both private and commercial vehicle operators. Alternate route information was highly valued by all these users. Respondents indicated that the main reasons for choosing an alternate route were accidents, traffic volume levels and road construction activities. Around half the private and commercial drivers cited the gain in time by rerouting as the reason for switching routes. Accuracy and currency were found to be the most important attributes of the information provided by an ATIS or CVO application. Because drivers' observations of traffic conditions play an important role in motivating a route switch, the authors suggest providing live displays of real-time traffic conditions as a component of a traffic information system. They also suggest providing information (either en route or post trip) that confirms and validates the decisions actually made by a driver, in order to build confidence in the use of ATIS.

Survey and analysis issues that arise in investigations of user preferences for possible travel information system features are addressed directly or indirectly in a number of references in the literature. (Ng, Barfield et al. 1997) provide a high-level overview of survey design and analysis methods that might be applicable to such investigations, and furnishes extensive details about survey design issues and their resolution in several case studies. (Mehndiratta, Kemp et al. 1999a) discuss a number of stated preference survey design and analysis issues, including the possible presence of response bias (respondents give positive answers thinking it will please the surveyor) and non-commitment bias (respondents overstate their willingness to pay because no money is actually committed by answering). The authors also investigated the econometric problem of correlated error terms in the response by a single person to multiple related questions. They addressed the problem by specifying and estimating random parameter logit models, but found that this computationally-intensive technique did not result in estimates significantly different from those obtained using simple logit models.

3.3.1 ATIS message reliability

Reliability is a feature of particular prominence in analyses of ATIS message attributes. Generation of high accuracy ATIS messages is a challenging technical task, for a number of reasons. Measurements of traffic conditions on a network will generally be made using a limited number of data collection devices (traffic detectors, probe vehicles, cameras, etc.) The measurements will inevitably be imperfect (imprecise and inaccurate) for a variety of technical reasons. Information of particular interest, such as assessments of the severity and clearance time of incidents, may not even be available until after special personnel (police, traffic crews) physically reach the incident site. Data communications and processing limitations mean that traffic measurements cannot be instantaneously converted into meaningful traffic messages. From imperfect measurements of a limited number of variables processed at time intervals that are large compared to characteristic times of traffic dynamics, it will be difficult, to say the least, to obtain and maintain a detailed and up-to-date picture of prevailing traffic conditions.

Furthermore, data on prevailing conditions may not be an accurate basis for determining the conditions that a vehicle will actually encounter on a path. (Ben-Akiva, de Palma et al. 1996) show analytically that use of prevailing conditions for ATIS messages can lead to a worsening rather than an improvement in traffic conditions, and explore the sensitivity of ATIS messages to inaccuracies and imperfections in traffic conditions. (Chen and Mahmassani 1991) investigated, using a mesoscopic traffic simulator, the reliability of route guidance recommendations based on prevailing times. They compared minimum paths and path times based on prevailing times with the actual minimum paths and path times using true (i.e. time-varying) link times and concluded that real-time ATIS messages based on "information on currently prevailing link trip times, with no attempt to predict future travel time or traffic conditions, may not be very reliable, especially at high levels of market penetration." However, guidance based on predicted traffic conditions requires forecasts and models, which may not be particularly accurate, and involves large amounts of computation, which will add to the time delays of the provided information. Thus, predictive guidance, even if it has the theoretical possibility of better matching a driver's actual travel experience, may be constrained in its accuracy by practical and computational factors.

A number of studies of user preferences for ATIS features have included consideration of message accuracy, as has been seen. (Madanat, Yang et al. 1995) included drivers' perceptions of information reliability as a latent (unobservable) variable in a route switch model and found it to have both direct and indirect effects on the probability of switching in response to information; the indirect effect came through its influence on drivers' general attitudes towards route diversion (another latent variable in the model). (Hato, Taniguchi et al. 1995) developed stated choice models of route switching behavior in which the accuracy of reported travel times was explicitly varied in different choice situations, and found that the information accuracy level was a significant variable in determining switching probability.

(Kantowitz, Hanowski et al. 1997a; Kantowitz, Hanowski et al. 1997b) explicitly consider the question of how much inaccuracy ATIS users will tolerate. They pose the issue in terms of the relative strengths of drivers' self-confidence in their knowledge of traffic conditions, and their trust in the ATIS messages. The authors conducted experiments using a travel choice simulator in which information on link conditions (light or heavy traffic) was intentionally degraded. They considered situations in which either 73% or 41% of the links had correct information. (These numbers come from prior work by the authors on reliability issues in human factors. Of course, in some cases, the incorrect information is harmful –when driver chooses a heavily congested link because it is reported to have light traffic, for example –while in others the error may be relatively benign.) They found that when 73% of the link reports were accurate, drivers still took account of the messages; while when only 41% were accurate, drivers ignored them. Drivers did not use accurate information as effectively in the familiar setting as in the unfamiliar setting. Also, inaccurate traffic information was more harmful in a familiar setting. Thus, it would appear that drivers are tolerant of a certain amount of error in ATIS messages, although drivers familiar with an area will expect a higher degree of accuracy from the information system.

3.4 User benefits from ATIS

The economic benefits that an ATIS user derives from ATIS services are very closely tied to the user’s response to ATIS and to his or her willingness to pay for ATIS information: they are all aspects of the same internal evaluation and decision-making process. The discussions in Sections 3.2 and 3.3 have covered many aspects of ATIS user benefit evaluation.

It has been seen that the spectrum of possible user responses to ATIS information is vast, ranging from relatively simple behavioral responses like route switching to complex responses such as re-arranging ones schedule of daily activities. This range exceeds the gamut of responses conventionally considered in transportation benefit evaluation exercises, and indicates that considerable care must be taken in thinking about and quantifying their benefits.

In conventional evaluation approaches, user benefits are usually computed as a change in consumer’s surplus, defined as the total difference between what each user is willing to pay (in money or in time) for something and the amount actually paid. Willingness to pay is deduced from the travel demand curve, expressing the amount of travel that would be made at different cost or time levels. The evaluation thus assumes that user benefits are tied to travel cost or time reductions.

This assumption is unlikely to lead to a complete and comprehensive approach to evaluating ATIS-produced user benefits. For example, peoples’ re-arrangement of their daily activity schedules may lead to more rather than less time being spent in travel, as they are able to carry out more activities because of more precise planning. If one were to ask such people if they were better off because of ATIS, they would reply affirmatively, even though they spend more time traveling: the benefits they derive from the additional things they do more than offsets the opportunity cost and disutility of the time spent traveling. If this were not true, they would not have re-arranged their schedule.

Special cases of ATIS-produced benefits can be distinguished, and may lead to simplified evaluation procedures when it is known what are the preponderant impacts of ATIS in a particular situation. In general, of course, it will not be possible to know a priori what the main impacts of an ATIS on user behavior are likely to be.

For example, if the only effect of an ATIS is to cause someone to switch routes, it might be reasonable to evaluate the ATIS user benefits via the resulting savings in travel time or cost. (If the user has confidence in the ATIS, an additional benefit may derive from the reassurance of having made an informed route switch, as opposed to the stress that could accompany an uninformed decision.) However, as noted above, there are indications that ATIS-produced reductions in travel times are likely to be small, and that the most common effects of pre-trip ATIS will be in terms of departure time rather than path choice changes.

If the only effect of the ATIS is to provide more precise estimates of the travel time between activities at two locations, and so allow the user to spend more time at either trip endpoint, then it might be reasonable to evaluate the ATIS user benefits via the benefits of pursuing those endpoint activities. In this way, an estimate of the benefits of improved travel time reliability could be obtained.

(Small, Noland et al. 1999) carried out and analyzed stated preference surveys investigating the value of travel time savings in congested conditions, and the value of travel time reliability, to travelers and freight carriers. They found that travelers definitely impute a monetary value to travel time reliability; however, this value can be entirely explained in terms of the early or late schedule delay costs at the destination (i.e., the cost to a traveler of arriving earlier or later than her intended arrival time). After the schedule delay costs were accounted for, no residual valuation of travel time reliability could be detected from the survey results. Similar results were found for freight carriers, although the conclusions were less statistically robust: travel time reliability had a value to freight carriers, but this value was entirely attributable to the costs of late arrival compared to a scheduled time.

Brand (1998) has proposed a more general user benefit evaluation method that returns to the original economics approach based on willingness to pay. However, instead of attempting to estimate willingness to pay from a conventional time- or cost-based demand curve, he suggests estimating it directly, using stated preference surveys of current or potential users of ATIS services. Such surveys can pose questions in which respondents trade off service attributes against cost and, properly conducted and analyzed, can provide reliable information on users’ willingness to pay for different service attributes or for entire systems. A number of willingness to pay results from stated preference surveys were discussed in the preceding section.

By obtaining willingness to pay in this direct fashion, many of the complications of a model-based approach are avoided. There is no need, for example, to estimate how ATIS users might re-arrange their daily activity schedules and tripmaking behavior, and then to evaluate the travel and non-travel benefits and costs of the re-arrangement: the effects of such possible changes are already incorporated in the users’ responses to the stated preference surveys. This user benefits estimation method has the potential of being both simpler and more accurate than adaptations of conventional transportation evaluation methods to the very different properties of ATIS, although the usual caveats regarding stated preference surveys continue to apply.

3.5 Day-to-day effects and learning

ATIS is a new and evolving set of technologies, and new ATIS users will need to learn about its features, capabilities and performance. While learning about and using ATIS, individuals will inevitably have a variety of experiences with it, both positive and negative. Over time, these experiences will in some way shape peoples’ attitudes towards and use of ATIS. At a larger scale, the mechanisms by which people learn about ATIS and filter their experiences with it will strongly affect the public’s overall acceptance or rejection of ATIS technologies.

Perhaps for these reasons, a number of researchers have investigated the day-to-day learning processes associated with ATIS. Indeed, it appears that this subject has already been more intensively investigated than learning processes associated with conventional traffic equilibrium.

(Iida, Akiyama et al. 1992) provide an example of a study of learning processes in a conventional equilibrium context. They analyzed the dynamics of route choice behavior in simulator-based experiments that asked the participants to respond to repeated hypothetical route choices. In the analysis, travelers depart from a single origin to a single destination connected by two parallel alternative routes. Day-to-day variations in traffic conditions are represented by route travel time changes. Travel time prediction errors (the difference between predicted and actual travel time) as well as actual travel times are treated as "experiences" accumulating through the experiments. It was found that assumptions about learning behavior strongly affected the day-to-day variability of traffic flow; however, none of the assumptions considered led to flow equilibrium. The authors conclude from this that existing traffic assignment models may not be adequate representations of actual traffic phenomena.

In the context of ATIS, (Iida, Uno et al. 1999) performed a study to identify changes in drivers’ route choice mechanisms following the introduction of ATIS. They also investigated the influence of the accuracy guidance information on the route choice mechanism. The study used a travel choice simulator with which subjects repeatedly traveled between the same origin and the destination in the morning. During the experiment, the subjects learned about, and accumulated knowledge of, the network and information system. It was found that introduction of ATIS did change the decision mechanism that drivers applied, and that the quality of the provided information affected the nature and permanence of the change.

A similar study performed by (Vaughn, Abdel-Aty et al. 1993b) analyzed the accuracy of information provided in modeling drivers’ sequential route choices. This study also used discrete choice modeling framework to model sequential route choices. Experimental sequential route choice data under the influence of ATIS was collected using a PC-based travel choice simulator. The experiment collected information on drivers' pre-trip route choice behavior at three levels of information accuracy: 60 percent, 75 percent and 90 percent. An analysis of variance was performed on the data to investigate the interrelationships among the different variables in an attempt to identify factors that significantly influence route choice behavior and learning. An attempt was made to model sequential route choice behavior using a binary logit model formulation; the results were mixed. It was assumed that drivers update their knowledge of the system based on their previous experiences; therefore an information updating function was specified and incorporated into the model. The results indicate that drivers can rapidly identify the accuracy level of information being provided and that they adjust their behavior accordingly. There is also evidence that indicates that an accuracy threshold level exists, below which drivers will not follow advice and above which drivers readily follow advice.

(van Berkum and van der Mede 1999) proposed a very general dynamic model of ATIS-guided route choice that includes behaviors based on perceived utility maximization, habitual choice and compliance with prescriptive guidance. Irrespective of the choice rule operating, individuals learn from their experiences. After each trip, the experienced travel time is used to update the mean expected travel time and the travel time variance for the chosen route. Descriptive and prescriptive guidance information influence route choices in different ways. Descriptive information may be incorporated into the perceived utility of alternatives for the subsequent choice. Prescriptive guidance can overrule the perceived utility maximization and habitual choice behaviors. The degree to which guidance affects the decision depends on the credibility of the information, and the credibility is influenced in turn by previous experiences with the information system.

In modeling dynamics, it is necessary to observe the behavior of a decision-maker over time. Investigating route switching in a dynamic context enables the calibration and testing of richer model specifications by incorporating repeated measurements, heterogeneity, within-day and day-to-day influences of variables, and state dependence effects. The multinomial probit framework (MNP), though well suited to tackle these challenges in dynamic models with a few periods, is prohibitively expensive for panels of longer duration.

To address the needs of modeling dynamic route switching over a large number of decision periods, (Srinivasan and Mahmassani 2000a) proposed a dynamic kernel logit model that retains the flexibility of multinomial probit while exploiting to some extent the computational tractability of the logit model. They applied the model to analyze the influence of systematic effects on route-switching behavior under ATIS. The effect of trip maker characteristics, trip characteristics and traffic conditions, experiences in traffic, and attributes of ATIS information are examined in this context. They also investigated heterogeneity effects in route switching behavior. Finally, time-dependent effects in route switching behavior are examined in two ways. First, at the systematic level, the influence of past experiences on current behavior is assessed. Second, dynamic effects were investigated via the structure of the utility disturbance terms. At the unobserved level, time dependence effects are examined by specifying suitable variance components. The variance-covariance structures are tested for the presence of temporal correlation (both within day and day-to-day), in addition to serial correlation (due to repeated measurements).

Many analyses of driver-network transportation systems assume that the systems are in equilibrium. Equilibrium analyses presuppose that the driver is rational and homogeneous, and has perfect information. (Nakayama, Kitamura et al. 2001) suppose, on the contrary, that people have cognitive limitations. A driver is assumed in this study to adopt simple rules when choosing a route. The authors develop a simulation system in which drivers’ learning is simulated through a genetic algorithm that, over time, that generates and modifies a set of route choice decision rules. The results of simulation analyses can be summarized as follows: Drivers do not become homogeneous and rational as equilibrium analyses presuppose; rather, there are less rational drivers even after a long process of learning, and heterogeneous drivers make up the system. Drivers' attitude toward and perceptions of each route do not become homogeneous either, but become bipolar. The results point to the need for a critical appraisal of the foundation of the equilibrium analysis of network flow.

(Ozbay, Datta et al. 2001) proposed the use stochastic learning automata (SLA) to analyze drivers’ day-to-day route choice behavior. This model addresses the learning behavior of travelers based on experienced travel time and day-to-day learning. In order to calibrate the penalties of the model, an Internet based Route Choice Simulator (IRCS) was developed. The IRCS is a traffic simulation model that represents within day and day-to-day fluctuations in traffic and was developed using Java programming. The calibrated SLA model was then applied to a simple transportation network to test if global user equilibrium, instantaneous equilibrium, and driver learning have occurred over a period of time. It was observed that the developed stochastic learning model accurately depicts the day-to-day learning behavior of travelers. Finally, it is shown that the sample network converges to equilibrium, both in terms of global user and instantaneous equilibrium.

While many travel behavior studies that deal with day-to-day learning have focused on modeling route choice behavior under information, fewer have examined day-to-day processes in departure time choice behavior with ATIS. The motivation in modeling departure time choice dynamics stems from the following considerations. The departure time decisions of commuters on a given day significantly influence the within-day distribution of traffic, congestion and queuing patterns on the network in the peak period. Accurate models of departure time adjustments can translate into a robust time-dependent OD prediction capability that is an essential component for dynamic traffic modeling and assignment techniques. In addition, since departure time variations influence the network flow evolution from day-to-day, models of departure time choice dynamics are important for characterizing and analyzing dynamic network states and the associated costs. Dynamic models of departure time choice play an important role in demand forecasting, as an integral component of activity-based demand modeling framework.

(Mahmassani and Chang 1986) performed an exploratory analysis that included 1) the explicit treatment of the day-to-day dynamics of departure time decisions, 2) the specifications of mechanisms by which individual users adjust their decisions on a daily basis, given prior experience, 3) the boundedly-rational heuristics that are assumed to govern individual tripmakers’ behavior, and their use in a modeling framework that recognizes the interaction between user behavior and system performance, and 4) the use of a special-purpose traffic simulation model to study the dynamics of user behavior. An extension of this work was conducted by (Mahmassani and Stephan 1988) in two directions: 1) the inclusion of the route choice dimension in addition to that of departure time and 2) the consideration of two user groups with different information availability levels interacting in the same simulated commuting system. The effect of information availability on the behavior and performance of given user group was of particular interest. In this regard, the results of this experiment are broadly consistent with a priori expectations; that is, users with more information clearly outperform those with limited information when both are competing in the same system. The interdependence between route choice and departure time decisions is another important aspect of user behavior addressed in this paper. The exploratory aggregate analysis considered here points to the precedence of departure time shifts over route shifting in dealing with experienced unpredicted congestion in the system.

The above mentioned works in day-to-day departure time choice modeling do not propose specific models of the departure time adjustment process. (Srinivasan and Mahmassani 2001) addressed this by investigating alternative mechanisms commuters' day-to-day departure time adjustment behavior. The mechanisms they considered include: utility maximization from unordered alternatives; ordinal response mechanism (where thresholds are corresponding to choice alternatives are ordered); sequential greedy search process; and a two-stage nested adjustment process. Econometric models are proposed corresponding to these mechanisms and implemented using departure time adjustment data obtained from interactive simulator-based experiments. The results indicate that the observed departure time choice dynamics is consistent with a sequential greedy search process. Under this mechanism, users continue to search for acceptable adjustment alternatives in a sequential and ordered fashion, until a satisfactory departure time choice is obtained. The results also indicate that network conditions, users' past experiences in the short and longer-term, and the nature and type of real-time information supplied by ATIS significantly influence the adjustment behavior of commuters. The models and results have significant applications in demand forecasting, network state prediction, and the evaluation of transportation control measures.

All of the above studies considered single-purpose trips from origin to destination. In fact, many trips involve multiple purposes and intermediate stops; this is called trip chaining. Trip chaining can significantly impact travelers’ route and departure time switching behavior. Trips with intermediate stops are more likely to involve switching than trips without stops. (Mahmassani, Hatcher et al. 1991) addressed the daily variation of trip-chaining behavior of commuters, and related it to various attributes of the commuter, the workplace, and the commute. The paper addresses the day-to-day variation of three key aspects of the home-to-work commute: 1) the time of departure from home; 2) the frequency, purpose, and duration of intervening stops between home and work; and 3) the path actually followed through the network. It is based on two-week detailed diaries of actual commuting trips completed by a sample of auto commuters in Austin, Texas. About 25 percent of all reported commutes contained at least one non-work intermediate stop, underscoring the importance of trip chaining in commuting behavior. These multipurpose trips are shown to influence significantly the departure time and route-switching behavior of commuters.

Although considerable attention has been given to incorporate day-to-day learning in route and departure time choice modeling, the same cannot be said about modeling mode choice. The only work that was found during the course of the literature review is by (Aarts, Verplanken et al. 1997). This study focuses on travel mode choice behavior in order to test theoretical propositions as to habitual decision making. It investigates the effects of habit on information processing during judgments of travel mode use. The study used multiple regression analysis to test the hypothesis that habit is negatively related to the elaborateness of information processing preceding judgments of travel mode use. The study focused on the judgment of bicycle use for short distance trips. It is expected that individuals who have developed a strong bicycle choice habit apply less elaborate information processing strategies compared to those who have not developed such a habit.

3.6 Human factors issues

A driver’s ability to navigate through a complex environment is largely dependent on the type and extent of cognitive structures representing that environment, the goals of the driver, and the ability of the driver to stay oriented. These three areas, founded in psychology and environmental cognition, are functionally related. First, a destination and travel plan must be formed. Second, knowledge of the local or global network must be known or acquired. Finally, a reference system must exist to relate the driver to the environment. The cognitive map has been hypothesized as the basis for mentally storing or representing information about the physical world. The internal format of remembering this information could have profound effects on the ease with which one can assimilate information presented by an Advanced Traveler Information System (ATIS). If the information is mentally stored in a prepositional format, then specific verbal directions may be desirable. However, if the information if the information is in a format analogous to the real world, a different representation, the map for example, may be desired. In addition, the spatial and verbal skills of drivers may vary significantly among individuals; thereby influencing their ability to use different navigational display formats. Human factor issues of concern include the format and coding of navigation system information, the attentional demand and safety issues of displays and controls, and agreement on general guidelines for the development and manufacture of ATIS.

A number of research studies have been reviewed that deal with human factors involvement in the design and use of ATIS. Some of them deal with the application of human factors guidelines and design decision aids for ATIS and ATIS displays. The questions that the designers must answer when developing displays for ATIS, which will affect or have an impact on both the safety and usability of the system are: (i) What information should be included in the ATIS that is being developed? (ii) What functions of the ATIS should the driver be allowed to use? (iii) To which sensory modality (e.g., auditory, visual, tactile) should information items be allocated? (iv) What format (e.g., text, map, tone, voice) should be used to present the information?

(Mollenhauer, Hulse et al. 1997) explored the decisions that designers must make when developing ATIS displays. They described a design support process that has been developed to help formulate answers that reflect current human factors research and accepted design principles. Examples of decision tools that make up this process are provided along with a description of how these tools can be used together to aid in the design process. To analyze the information format options, “trade study” analysis is used to aid in design decisions. These analyses serve as systematic aids for complex decision making. In addition, specific results are also presented and discussed.

(Landau, Hanley et al. 1997) reviewed the following topics for guideline availability and applicability to an ATIS:

· input methodology: The design of the input mechanisms for an in-vehicle system must consider the accuracy and speed required for transactions;

· display and information characteristics: The research covers guidelines related to both legibility and readability of a display;

· auditory display characteristics: Auditory displays include both nonverbal and verbal aural displays. Nonverbal displays use auditory alerting signals to signify events. Verbal displays use voice signals or messages to signify events and to provide more complex information. Auditory displays can supplement visual systems;

· human-computer interaction: The interaction between a driver and an ATIS system will be modeled to a great degree on human-computer systems because the nature and complexity of the transactions are so similar to current computer interfaces. Therefore, the applicability of human-computer interface guidelines is reviewed; and

· navigation information format: Navigation information is typically portrayed by maps that provide direction and distance relationships in a plan view presentation. Another type of navigational format is turn-by-turn sequential list.

The successful implementation of ATIS depends on user acceptance of its products and services. Information on user acceptance could be applied to the design of ITS products and services, as well as to the development of ATIS implementation strategy. User acceptance is particularly important to the successful implementation of ATIS because the accuracy of traffic information it conveys is dependent on the number of ATIS equipped vehicles. Receiving inaccurate information from an ATIS device may break the trust the driver has in the system and lead to user rejection. Consumer rejection of ATIS, in turn, may lead to decreased system reliability and accuracy. ATIS is unique in that the degree of consumer use affects system effectiveness. Thus, to optimize ATIS accuracy, initial acceptance of ATIS should be maximized.

The results of the study done by (Wochinger and Boehm-Davis 1997)


U.S. Department of Transportation		John A. Volpe National Transportation Systems Center
Research and Special Programs Administration