Category Archives: travel behavior

Charts of the Day |

Over at I have three charts of the day on Journey-to-Work Mode Share in Minnesota, the Greater Minneapolis-St. Paul region, and the City of Minneapolis from the Census American Community Survey 2012 data. Nothing terribly surprising to me, but it is good to document these things. It is worth noting that walk share > bike share and walk + bike ~ transit, though expenditures do not reflect that. (And for non-work trips walk is even more important).

Minneapolis Journey-to-Work Mode Share from 2012 ACS
Minneapolis Journey-to-Work Mode Share from 2012 ACS

Accessibility and non-work destination choice: A microscopic analysis of GPS travel data


Congratulations to Dr. Arthur Huang for successfully completing and defending his dissertation: Accessibility and non-work destination choice: A microscopic analysis of GPS travel data

The advancements of GPS and GIS technologies provide new opportunities for investigating vehicle trip generation and destination choice at the microscopic level. This research models how land use and road network structure influence non-work, non-home vehicle trip generation and non-work destination choice in the context of trip chains, using the in-vehicle GPS travel data in the Minneapolis-St. Paul Metropolitan Area. This research includes three key parts: modeling non-work vehicle trip generation, modeling non-work, single-destination choice, and modeling non-work, two-destination choice. This research contributes to methodologies in modeling single-destination choice and multiple-destination choice and tests several hypotheses which were not investigated before.

In modeling non-work vehicle trip generation, this research identifies correlation of trips made by the same individual in the trip generation models. To control for this effect, five mixed-effects models are systematically applied: mixed-effects linear model, mixed-effects log-linear model, mixed-effects negative binomial model, and mixed-effects ordered logistic model. The mixed-effects ordered logistic model produces the highest goodness of fit for our data and therefore is recommended.

In modeling non-work, single-destination choice, this research proposes a new method to build choice sets which combines survival analysis and random sampling. A systematic comparison of the goodness of fit of models with various choice set sizes is also performed to determine an appropriate choice set size. In modeling non-work, multiple-destination choice, this research proposes and compare three new approaches to build choice sets for two-destination choice in the context of trip chains. The outcomes of these approaches are empirically compared and we recommend the major/minor-destination approach for modeling two-destination choice. The modeling procedure can be expanded to trip chains with more than two destinations.

Our empirical findings reveal that:

  1. Although accessibility around home is not found to have statistically significant effects on non-work vehicle trips, the diversity of services within 10 to 15 minutes and 15 and 20 minutes from home can help reduce the number of non-work vehicle trips.
  2. Accessibility and diversity of services at destinations influence destination choice but they do not exert the same level of impact. The major destination in a trip chain tends to influence the decision more than the minor destination.
  3. The more dissimilar the two destinations in a trip chain are, the more attractive the trip chain is.
  4. Route-specific network measures such as turn index, speed discontinuity, axis of travel, and trip chains’ travel time saving ratio display statistically significant effects on destination choice.
    Our findings have implications on transportation planning for creating flourishing retail clusters and reducing the amount of vehicle travel.

After working at Valparaiso University last year, he is currently teaching at the University of Minnesota Duluth.

Exploring universal patterns in human home/work commuting from mobile phone data

Technology Review writes about a new working paper from Kevin S. Kung, Stanislav Sobolevsky, and Carlo Ratti: [1311.2911] Exploring universal patterns in human home/work commuting from mobile phone data: “Exploring universal patterns in human home/work commuting from mobile phone data

(Submitted on 12 Nov 2013)
Home-work commuting is known to be one of the major components of human mobility and therefor always attracted much research attention. One of the well-known assumptions being the focus of many works in this area is the universal uniformity of commute times. However, quantifications of commute patterns have often been baffled by the intrinsic differences in the data collection methods, which make the observations from different countries incomparable. In the present work we use mobile phone data offering a common methodology for investigating into the mobility pattern in different parts of the world including entire countries as different as Portugal and Ivory Coast as well as cities (Boston) also comparing results with those obtained from vehicle GPS traces in Milan. We showed that despite substantial spatial and infrastructural differences, the commute time distributions and average values are indeed largely independent of commute distance or country.

So cell phone data corroborates a mean commute time of about 1 hour each day. Yacov Zahavi and others have been talking about this for decades. (I have done a few papers on this topic myself.) It is always good to see more empirical evidence, and such a large data set. The method uses some inferences to determine when someone left home (last phone call at home (most frequented) cell phone tower) and arrived at work (first phone call at work (2nd most frequented) cell phone tower).

There are some details I am not clear about.

I am not always on the phone, and often don’t call just before departing or just after arriving … I would have picked the minimum time of all days when they had calls at home and work (since on the extreme day they are off and on the phone continuously, but on others not), but they don’t say they did that. Also towers serve big areas, and entering a tower zone does not mean arriving at work or home. Still, good to see data used in interesting ways, it is just important to be careful about interpretation.

I think the claim of universality needs to be tempered, since mean commute time (Figure 5(a) varies from under 60 minutes average in Boston to almost 80 minutes in Ivory Coast, a non-trivial difference.

I also think “human” in the title is unnecessary, until we find other species that have home/work commutes.


7 Ways to Reduce Transportation Waste

Some 20 years ago a book came out “Stuck in Traffic” by the brilliant Anthony Downs. One of his key points was the “Iron Law of Congestion”, sometimes called “Triple Convergence”, and now called “Induced Demand” which basically said if you expand a road, the extra capacity gets used up by people switching routes, modes, and time of travel. We might also add other effects of road expansion include changing destinations for non-work trips (making longer trips), making trips that would otherwise be foregone, and even changing jobs and houses, as well new development. While Downs did not discover this idea, (e.g. Lewis Mumford had said something similar) he popularized it.

Since Downs wrote the book in 1992, remarkably little has changed in how we travel. Not nothing of course, (travel rose for a few more years and has leveled off overall, and dropped in more recent years on a per-capita basis) but a lot less than you would expect given the changes in information technologies over the same period. No-one is satisfied with this status quo. Everyone is crying out for something different. We believe we can do better than daily congestion, excess pollution, devastating crashes, and all the other ailments associated with our existing transportation systems. In recent posts I identified peak travel, and made a speculative scenario of how traffic might disappear “on its own”. But of course, that won’t happen everywhere, and there are lots of things we can do to manage better.

Roads are perhaps the slowest changing technology. Once laid, they are difficult to move. Parts of the Appian Way, from two millennia ago, are still in use. Famously London could not change its street grid after the 1666 London Fire destroyed most of its buildings, despite an able plan from Sir Christopher Wren.

We can think of transportation as a layered system. There is the earth, on top of which are rights-of-way, within the rights of way are pavements (themselves layered). On the pavements are markings denoting lanes and directions. Above these are traffic control devices like signs and signals. Only then do we get to services, people driving their vehicles, trucks carrying freight, passenger buses, taxis, bicycles, pedestrians, etc. all riding on the layer of roads.

To the dismay of many transportation planners and engineers (many of whom got into the field to build things), the physical layer of the surface transportation network in the United States is largely complete. The projects that are left are projects that were too expensive to build the first (or second, or third) time, (much like the Second Avenue Subway in New York). These Zombie projects do not die, while construction is essentially irreversible, non-construction is easily reversed. In the end though, these are tinkering at the edges. Given the small amount of new construction, most travel 20, 30, or 50 years from now will be on roads that already exist.

Until we go airborne for short trips, we are highly constrained. As the world gets more developed, building new roads is progressively more expensive. The world adapts to the infrastructure that is provided, and builds as close to the right-of-way as possible, making expansion that much more difficult.

While we have scarce right-of-way (scarce in that it is limited, and finite, and at times fully utilized given the applied technologies for its use). We lack time. We have limited energy. This waste has both supply and demand aspects. Yet collectively we don’t do relatively easy things that would reduce the waste of these scarce resources.

  1. Most roads are under-utilized most of the time. We have plenty of capacity outside the peak.
  2. Most of the pavement is unused even at peak times, there are large gaps between vehicles both in terms of the headway between vehicles [1] and the lateral spacing between vehicles (we drive 6′ wide cars in 12′ lanes, often on highways with wide shoulders).
  3. Most seats in most cars are unoccupied most of the time.
  4. Most cars are carrying around far more weight than required to safely move the passenger. While bigger cars tend to be safer for the occupants, they are less safe for non-occupants. This is an inefficient arms race.
  5. Most roads are so wide we use them for storage of vehicles most of the day.
  6. There is a tremendous amount of excess delay at traffic lights, especially at off-peak periods, wasting time (and space).
  7. Most trips during peak periods are not work trips and have temporal flexibility, yet these trips travel in the peak because they are underpriced.
  8. Most trips produce negative externalities (pollution, congestion, noise, risk of crash) in excess of the price paid by their driver. They produce so many of these externalities because they don’t pay for their full cost.

So instead of expansion, we should instead think about ways to use that scarce right-of-way (and our scarce time and energy) more efficiently.

  1. Where there is congestion, we should price roads to encourage use in the off-peak and discourage use in the peak. This revenue should be used for the operations and maintenance of roads and should largely replace existing funding sources (fuel taxes, vehicle taxes, property taxes). Prices need to be systematic, not just on specific routes, to maximize system efficiency. We should avoid having a random set of underutilized toll roads, while free roads remain congested.
  2. Pollution and noise and crash risk should have their own externality charges.
  3. We should encourage narrower vehicles and, for instance, provide two six-foot lanes for narrow vehicles in place of one twelve-foot lane where we can, and promote use of driverless cars so that cars can use less space. Lane widths are standard, and changing them would require changing standards. Ultimately we should move to a model where we don’t need pre-defined lanes, but rather have vehicles move as near each other as possible without colliding, such as we do when walking in crowds.
  4. Cars should be lighter. If all cars were lighter, everyone would be safer. The greatest risk is when big car/truck meets small car. In particular we should encourage use of neighborhood cars that are specialized for local, lower speed travel. This may or may not require private vehicle ownerships as opposed to vehicle rental/sharing.
  5. We should promote technology to enable real-time, ad-hoc ride-sharing (with compensation for the ride provider) to better utilize excess capacity within vehicles. (In many places the compensation is illegal, as it looks and smells like a taxi, which are highly regulated).
  6. We should narrow up roads where we can, and use strategies so that people can share more cars, so we need fewer of them, so we don’t need to spend as much road space for vehicle storage. Road widths are again set by standards, often determined by the fire department (which does not want to back up their trucks).
  7. We should be able to eliminate many traffic signals with appropriate use of roundabouts (and later with driverless vehicles). Given the vacant space available on roads, if vehicles and inter-vehicle communications were better, we should be able to arrange real-time coordination of vehicle movements and have as a goal eliminating almost all stopped delay at undersaturated intersections that are today signalized. Pedestrian/vehicle conflicts might still remain, and require controls.

This is hardly a complete catalog of what we can and should do, but I hope the key point, there is plenty of pavement already, we just need to use it more wisely, comes through.


[1] If we follow the “2 second rule” (2 seconds between two successive vehicle’s front bumpers) (or 1800 vehicles per hour), at 60 mph we have a vehicle density of 30 vehicles per mile, or 176 ft per vehicle. Obviously with congestion, we are wasting time because we don’t increase throughput and we decrease speed, though we increase density. Vehicles are typically 26 ft, so we are using about 7 vehicle lengths for every vehicle we are moving at free flow speed near maximum stable throughput on a pipeline section without a bottleneck.

Network Structure and Travel Time Perception


Recently Published:

  • Parthasarathi P, Levinson D, and Hochmair H (2013) Network Structure and Travel Time Perception PLOS ONE: 8(10): e77718. doi:10.1371/journal.pone.0077718 :
  • “The purpose of this research is to test the systematic variation in the perception of travel time among travelers and relate the variation to the underlying street network structure. Travel survey data from the Twin Cities metropolitan area (which includes the cities of Minneapolis and St. Paul) is used for the analysis. Travelers are classified into two groups based on the ratio of perceived and estimated commute travel time. The measures of network structure are estimated using the street network along the identified commute route. T-test comparisons are conducted to identify statistically significant differences in estimated network measures between the two traveler groups. The combined effect of these estimated network measures on travel time is then analyzed using regression models. The results from the t-test and regression analyses confirm the influence of the underlying network structure on the perception of travel time.”

The Traffic is Falling

Published at the CTS Conversations Blog as Why are Twin Citians taking fewer trips?

The latest summary of the Twin Cities Travel Behavior Inventory is out, and it says the total number of trips in 2011 is lower than in 2001.

This is consistent with a lot of evidence we have been seeing from various sources on “Peak Travel”. Nationally, passenger miles on highways traveled today is lower than 1999, and miles per capita lower still. This is in stark contrast with the trend from 1900 to 2000, when vehicle miles traveled increased almost every year, with a few exceptions due to economic downturns or energy shocks. Why might this be? There are a number of hypotheses:

  • Employment – If fewer people are working, fewer people are traveling for work, and fewer discretionary trips are made by both workers (nervous about spending money) and the unemployed (who have little or no money to spend). Unemployment increased sharply beginning in 2008, and though it has declined, employment participation are still much lower (hovering around 58.6% of the population, down from 62.7% for the period before the recession). Demographics are part of this, and many people have just dropped out of the labor market, as their skills have been devalued by the economy, and either chosen early retirement (if near retirement) or deferred entry into the workforce (if young). The rise in female labor force participation from the 1950s through the 1990s has run its course, as labor force participation is roughly at parity by gender.
  • Gas prices – The price of fuel increased sharply in the run-up to the Great Recession – and this certainly discouraged car travel. Interestingly, it also reduced car crashes by more than the reduction in VMT, which we attribute to worse than average drivers (especially the young) being more likely to be priced off the road.
  • Changes in driver license regulations – It is harder for young people to get drivers licenses, and less valuable since they now need more supervision. More defer licensure and auto acquisition.
  • Social networking – It is now easier to virtually communicate with friends, in real-time and asynchronously, where-ever you are, so it may be less necessary to actually visit them. Time online just continues to rise, especially mobile.
  • At home work – Telecommuting continues to rise, more as a complement to office work than a substitute (check your email when you get up, check your email before bed), but even as the occasional substitute, so I can work from home either sometimes or regularly.
  • At home shopping – The rise of online retailing allows us to substitute delivery for fetching, and reduce the amount of shopping trips.
  • Kids today just don’t like cars – The culture just may be different, and the desire for mobility, especially auto-mobility has dropped, as kids would prefer to spend their disposable income on the latest internet-connected gadget.

There may certainly be other explanations as well. No one of these explanations is the whole story, and some are certainly more likely than others, but they all work in the direction of reducing auto travel. Importantly, travel by other modes has not made up for the large drop in car use. While transit, e.g., is up nationally due to the large investments in rail lines, that 20% increase in transit use in the decade is a little more than 1% of the 20% drop in passenger miles by motor vehicles.

Colleagues (Jason Cao, Yingling Fan, Michael Iacono, Greg Lindsey) and I have a study currently underway, Travel Behavior Over Time, with support from the Minnesota Department of Transportation and the Metropolitan Council to examine not just the top-line number, but what goes on underneath. We plan to determine the factors that explain travel demand and see if they are the same as in previous decades. We hope to understand whether the factors are the same, but the conditions are different, or whether underlying travel preferences have changed.

People and their Paths 5: How to get people to behave optimally

A shortened version of this post was adapted for Symposium Magazine‘s article Understanding the Irrational Commuter, which appeared in the September 2013 issue.

At the beginning of this series, we described Wardrop’s First Principle, of User Equilibrium. He also had a second principle, of System Optimality, which says: “At equilibrium the average journey time is minimum.” To achieve this requires every traveler to act in accordance with society’s best interest, which as we noted in part 2, is not generally calculable by an individual. This ratio of the total system travel time associated with a user equilibrium traffic pattern and the system optimal travel pattern has been dubbed “The Price of Anarchy” by Tim Roughgarden, who has applied this to computer networks. This number indicates the inefficiency of autonomous (or selfish) control in a system, compared to a theoretically best central control.

UE SO Flow

When choosing a route, selfish users see the costs they incur, but not the costs they impose on others. This is analogous to the difference between average and social marginal costs in economics. If we somehow persuaded travelers to make route decisions considering the cost they impose on others, their marginal cost, we could achieve a minimal total cost for the system. In economics, the classic theoretical mechanism for this is called a Pigouvian Tax, which charges the polluter for the negative externalities imposed on the pollutee (the difference between the social marginal cost and social average cost). In this case the externality is congestion, or travel time imposed by a vehicle on all other vehicles in excess of what would be borne in the vehicle’s absence. The Pigouvian Tax gains its name from Arthur Pigou, a British economist from the 1920s, who discusses the idea in his text The Economics of Welfare.

Travelers facing travel times and a Pigouvian Tax might choose a route that satisfies both of Wardrop’s Principles. The User Equilibrium (UE) solution would equal the Social Optimal (SO).

Using traffic assignment models we compared system optimal and user equilibrium flows and travel times for the Minneapolis – Saint Paul regional planning network, assuming total traffic flow between origins and destinations were fixed (i.e. unaffected by our distortion of route prices). We found the SO assignment had a 1.7% overall time savings, and a slightly higher average speed (63.2 km/h vs. 61.8 km/h). Perhaps surprisingly, it also had somewhat more total vehicle kilometers traveled (9.37M vs. 9.33M), as drivers had to take longer routes to avoid imposing congestion on others.

While the SO result is better than the UE (it cannot be worse), we might ask “SO What?”. The price of anarchy, letting drivers choose their own routes rather than being centrally directed, is relatively small, under 2 percent. It turns out it is much more important to get people to choose an efficient time of day than to worry about micro-managing which route they select.

We could post time-varying prices (just like the HOT lanes of Part 3, or many transit systems which have peak and off-peak fares) to discourage demand when it is highest, and encourage demand at off-peak periods. This is done on some toll facilities now, and other schemes, like the London Congestion Charge, have two prices: free or tolled, depending on time of day. But this can be as refined as we want it, with prices changing every hour, every five minutes, or even continuously. The prices might change in real-time, or change according to a fixed and posted schedule.

Nobel-winning Economist William Vickery developed the first version of the bottleneck model, which showed how varying prices would allow people to trade-off being on-time (at a higher toll) or being early or late (at a lower toll, but a higher cost in what transportation researchers call “schedule delay”).


The simplest version of this has two players1. Imagine two boats racing for a canal lock, or, as in the image, two weightlifters trying to get through a narrow door on the London Underground. When they arrive at the same time, only one can make it through first, the other has to wait. The one who makes it through imposed schedule delay on the one who waited. But if they arrived at different times, there would be no direct schedule delay, though one might not get into the canal (or through the door) at their preferred time. If we appropriately price simultaneous arrivals, we will discourage them. While with two players this may be feasible to coordinate with direct communication by saying don’t arrive when the other guy arrives, and negotiating, for 2000 people instead of 2, coordination is better through posted price signals than conversation and negotiation. Prices varying by time-of-day is what congestion pricing is about, putting a higher price on times which are most desired, and lower prices on the less desired times.

There are perhaps other ways to achieve this end. On most roads, it is assumed no one owns the travel time, and so we get congestion. If there were some kind of property in the right to travel at a given time, we could auction off this right to the highest bidder, and similarly avoid congestion. This would more closely follow a strategy of establishing property rights to avoid externalities, as suggested by British-American economist Ronald Coase (who is still talking about economics at the age of 102). In the transportation literature, this has come to be known as reservation pricing. Just as one does not expect to be seated when showing up unexpectedly at a popular restaurant that takes reservations, one should not expect to use a high-demand bottleneck facility on the transportation network without making arrangements in advance. Of course it is much more complicated with a real-time system like transportation, and to maximize throughput, it is likely that some queueing is required. This queue ensures there is someone waiting to take advantage of the next gap that opens. The alternative would be that the facility remains under-utilized for part of the time, which has its own costs. Even restaurants that reserve tables sometimes make you wait a little bit, for their immediate convenience, not yours, maximizing the productivity of their staff.

Unfortunately congestion pricing in any form remains more in the realm of theory than practice. While there are a few Congestion Charging programs: notably Singapore, London, and Stockholm, they are not over a large enough area, or variable enough in prices, to produce an end to congestion. Once many of these are implemented, I expect many cities will look at their peers and copy them, and it will become standard in all large metropolitan areas. But to date, the cases are fairly exceptional: Central Singapore, a city-state governed by a strong Prime Minister, whose family has been in power for five decades; Central London, a city governed at the time by “Red Ken” Livingstone, a radical thinker who was willing to take the political heat for the decision; and Stockholm, which conducted a trial experiment before holding an election to allow residents to vote up or down. Technically the systems all work well, and certainly do reduce congestion compared to the unpriced alternative. Politically they have been difficult to emulate. New York City tried and failed2, and no other US city has been willing to do something quite so radical.

Another possible deployment path for congestion pricing is through what is variously called a Vehicle Mileage Tax, or a Mileage-based User Fee. Gas tax revenues, which provide a large share of road funding, have been declining for a long time in the US, both due to leveling off of demand for driving, as well as better fuel economy. The simplest solution is to raise the gas tax, which solves an immediate problem, but not the longer term one. While hybrid gasoline-electric vehicles (like the Toyota Prius) still pay some gas tax, plug-in electrics (like the Tesla, Chevy Volt, or Nissan Leaf) pay almost none. Yet they still use the roads. Although they are presently a small share of the market, that share is likely to grow. Some states are beginning to think about how to charge EVs for the use of roads, just as gasoline-powered vehicles are charged based on a gas tax. Once a device is placed in cars tracking miles traveled (basically just the odometer, though possibly with some locational data to allow prices to vary by location (urban vs. rural) (although it technically feasible to ensure privacy, by not tracking which specific miles are traveled, no one will believe government protestations isn’t tracking them anymore, anyway), that device can also track when those miles are traveled, and vary the rate by time-of-day. The State of Washington now taxes EVs $100 per year to offset the lack of gas tax revenue. Oregon is conducting a large scale test of the Vehicle Mileage Tax, allowing 5000 volunteers to pay by the mile and have their gas tax rebated.

We are getting to the point where we can provide incentives and disincentives to efficiently manage road use. The technology exists, it is probably accurate enough. The cost of collecting a new road fee is non-trivial (especially compared with the gas tax, which simply requires an annual check of refinery sales), but the costs should drop with widespread deployment. The benefits are a significant improvement in the management of road use, so that drivers who do not need to travel when roads are congested, will have incentives to avoid those times.

If applied correctly, the resulting changes in route choices it will reveal where roads are overbuilt, and where demand, even after pricing, is sufficient to justify new capacity. The most cost-effective thing we can do in transportation is to get the prices right, all else will follow. This requires above all else, field experiments where different strategies are tested and evaluated, and deployment by replicating the successful experiments.




People and their Paths 4: Time and its Mis-perception

A shortened version of this post was adapted for Symposium Magazine‘s article Understanding the Irrational Commuter, which appeared in the September 2013 issue.

Bishop Berkeley said “Esse es Percipe“, meaning “to be is to be perceived”.

Not all is not all time created equal (as discussed yesterday), people systematically misperceive time. This means, sometimes they think places are farther away than they really are, and other times they are closer. Freeways seem to take shorter than they really area, local streets longer. We believe this in part has to do with task complexity, or the “mental transaction costs” involved in traveling.

When I need to make a lot of small driving and navigation decisions, like on a signalized route with lots of turns, I need to focus on driving more times. Each time I am engaging my conscious brain in traveling decisions. More brain-space is occupied by traveling thoughts.

(Other factors include temporal relevance (is the trip important?), temporal expectancies (what do I think the travel time will be?) temporal uncertainty (how reliable is my estimate of travel time?), affective elements (what is the emotional state of the traveler?), absorption and attentional deployment (am I paying attention to the task at hand?) and arousal (how physically activated am I, am I on drugs?).)

When I can drive on an uncongested freeway, I can avoid many such traveling thoughts. Driving is less salient. Time passes faster. As the expression goes, “time flies when you are having fun”

Vierordt’s Law, named for German physician Karl for Vierordt who published “Der Zeitsinn nach Versuchen” — “The experimental study of the time sense” in 1868, says people are more likely to over-estimate short times and under-estimate long times.

Time Perception at Traffic Signals

In our first study considering travel time perception, we wanted to compare how people perceive and value travel time waiting at red lights compared to moving on surface streets. The graduate student working on this, now Professor Xinkai Wu at Cal-Poly Pomona, created a simulation where drivers would see a traffic signal ahead, and get stuck behind a car which waited for the red light. They wait, and they wait, and they wait, and they get annoyed, and they keep waiting; for up to two minutes.

Travel perception experiment at traffic signals
Travel perception experiment at traffic signals

We had a set of scenarios. For instance, in one scenario, they would be waiting 120 seconds on the minor route but then they wouldn’t have any delay at two subsequent traffic signals. In another, they would only be waiting 30 seconds at the first light, but 60 seconds at second traffic light and 60 seconds at a third.

We thus tested (and did not corroborate) Vierordt’s Law. Perceived and actual waiting time were virtually identical for the first 30 seconds. But for times greater than 30 seconds, actual waiting time was higher than perceived waiting time, up to 120 seconds. At 120 seconds, the trend was for perceived time to over-take actual time, but that was the cut-off for the experiment, so perception findings in this situation require more information. However, the annoyance level at 120 seconds of waiting was much higher than the annoyance of waiting 30 seconds. Further people hated stops.

Of course, with all of this, it depends on how you frame the question, what you ask, and what travelers were expecting. Recall that comparing a computer-administered stated preference with one in which travelers were in a driving simulator completely flipped preferences for ramp meters.

In travel surveys we have a common phenomenon of rounding reported times, and times are usually rounded up. So if a trip was 14 minutes, it would be rounded to 15 minutes. If it were 22 minutes, it might be rounded to 25 or even 30 minutes. This makes self-reported times significantly biased in travel analysis. Until recently, that was the only data available. But now with the advent of GPS devices and cheap sensors tracking traffic across networks, we can get much better speed and travel time estimates.

How Network Structure Affects Time Perception

In another study we considered whether network structure affects how people report travel time (which we take to be perceived time, as we have no reason to assume people intentionally lie about  their perceived travel times). Looking at the network in downtown Minneapolis there is a very tight grid of streets, so the block sizes are relatively small compared to, for instance, suburban Woodbury. The older Minneapolis network is a fine-scaled grid while Woodbury is very circuitous and less well connected. Although there is an arterial grid in Woodbury, following the pattern set by the Northwest Ordinance, it’s at a much larger scale (about 1 mile rather than 0.1 mile).

Minneapolis and Woodbury
Minneapolis and Woodbury

We measured network structure along the actual route travelers pursued and compared reported times with our best estimate of measured travel times on their actual routes using GPS data.

We then stratified travelers into two groups, those who underestimated their travel time and those who overestimated their travel time, to see if there were any difference in the network structure experienced by each of these populations.

We measured network continuity (how often you change routes), positing that if you have more discontinuity in your network you’re more likely to overestimate travel time, because you spend more time thinking about.

Similarly, if there is a higher intersection density, more intersections per linear mile, you are more likely to overestimate the time. So each time you have to stop, or think about stopping, because there is potentially oncoming traffic, that’s going to be a mental transaction cost that increases how long you think about traveling, and thus how long you think you are traveling.

However, when your shortest path is along freeways, which have fewer decision points, you are more likely to underestimate travel time.

The accuracy of travel time perception on traffic signal wait duration, network structure, and what kind of route you’re taking. Undoubtedly it depends on many other factors yet to be discovered.


People misperceive the network systematically and we can predict how this misperception works, but of course we can’t predict any one person’s individual perception. On average we can see that in certain conditions some people are more likely to overestimate (or underestimate) their time. We need behaviorally-based route choice procedures. Transportation analysts should think about route choice not (only) as a mathematical problem of how to calculate the shortest path in the network but also about the things that people value, and what are the things that people perceive about the network, both of which will affect individual decisions.

People and their Paths 3: Not all time is created equal

A shortened version of this post was adapted for Symposium Magazine‘s article Understanding the Irrational Commuter, which appeared in the September 2013 issue.

One of the hypotheses which might explain why people don’t choose the shortest travel time path is that not all travel time is created equal.

We’ve done a series of experiments looking at how people perceive and value travel time under various conditions. There are two basic ways to ascertain valuations in travel behavior research, one is called revealed preference which examines the decisions people actually make and infers the causes using statistical tests. The second is called stated preference, where people are presented a set of hypothetical scenarios, and asked to choose between them. The advantage of revealed preference is that it is based on real decisions. Unfortunately, it is unable to provide insight to alternatives that don’t exist yet (differences of kind rather than quantity, or that are beyond the ranges currently observed), or that travelers have not themselves experienced.

For an extreme example, suppose you wanted to predict ridership on the Hyperloop. While we have data on people’s willingness to drive or take airplanes (or intercity buses) we don’t really have information about the willingness to tolerate particular specification of Hyperloop, responses to the comfort/discomfort in conditions of rapid acceleration and deceleration. So instead of relying on observed behavior alone, we would try to get estimates using other kinds of experiments. The simplest would try to describe the alternatives in words (either orally or printed) and ask people which they would choose (e.g. 30 minutes, high deceleration, $100 vs. 1 hour, airplane ride, $50). But as you might imagine, people might answer these questions differently in a survey than they would if they were faced with the real choice.

The experiments described below use hybrids between stated and revealed preference, where the subjects gained some experience with alternatives beyond a simple word-based description.

Experiment 1: Bicycle Route Choice and Multi-media Enhanced Stated Preference

Bicycling has seen a resurgence in many parts of the United States, from a being a fringe mode to gaining a significant number of users. In the City of Minneapolis, which is perhaps surprisingly at the forefront of this trend, bicycle mode shares to work are on the order of 4% (compared to about 0.5% nationally), and much higher in some areas (Downtown and near the University of Minnesota) and lower in others. Some studies have put the number as high as 11% of all trips. Some of this increase is due to greater investments in bicycling infrastructure. (And of course, this investment in bicycle infrastructure is due in part to increasing numbers of bicyclists lobbying for it).

As part of his Master’s Thesis, Nebiyou Tilahun, now an Assistant Professor at the University of Illinois at Chicago, Prof. Kevin Krizek (now at the University of Colorado) and I examined the weights people assign to factors when choosing bicycle routes1. While we couldn’t use a full-fledged driving simulator for this experiment (such devices are not cheap), we did the next best thing, we used a computer-based study with conditions shown as first person videos of riding a bicycle in different conditions. This presentation of alternatives is based on Kevin Krizek riding a bicycle, hands-off, taking a video camera and videotaping each condition. He’s an avid bicyclist and for people who don’t know him, his bicycles are probably worth more than my car.

Off road bike trail

Bike lane next to parking lane

So the question is essentially: Imagine you are commuting by bicycle and Routes 1 and 2 are the available choices. Route 1 is a nicer, off-road facility, but takes 40 minutes, while Route 2 is adjacent to traffic, and takes 20 minutes. Which do you prefer?

Then based on your choice you’d get another presentation where the travel times would change. One travel time would become higher or lower depending on what you answered. This allowed us to determine an indifference point, where you were indifferent to Route 1 or Route 2.

While everyone prefers the high quality route, all else equal, the indifference point varies by person. It systematically differs by gender. Women report they are willing to pay (in travel time) more for a higher quality off-road route than males are, though the general patterns are the same. It also varies by season: in the (cold, Minnesota) winter travelers are not willing to pay as much time to get a better quality route than in the summer.

Experiment 2: Ramp Meters and Virtual Experience Stated Preferences

Ramp Meters (sometimes called merge lights or flow signals) are traffic lights at freeway entrance ramps that ration the number of cars entering onto the freeways in order to smooth out traffic. This ensures that “platoons” of cars, a series of cars all following a leader (many early traffic engineers had a military background) are not all entering at once. Meters have also been used to ensure that the total number of cars on the freeway is below the capacity of downstream bottlenecks. Now common in many cities, they were first deployed in 1963 in Chicago and then Saint Paul starting in 1969.

In the year 2000, controversy arose about the effectiveness of the ramp metering system in the Minneapolis-Saint Paul region. In particular, some people were highly concerned about long waits at metered entrance ramps, which anecdotally people reported to be as high as 20 minutes. Unfortunately, at the time, there wasn’t any way to systematically determine the actual duration of waits.

The Minnesota Department of Transportation (MnDOT) was instructed by the state legislature to turn the ramp meters off for at least four weeks as an experiment. In fact, MnDOT continued the experiment for over 8 weeks. MnDOT ultimately concluded that ramp meters were valuable, but changed the metering strategy to reduce maximum queues at the ramps to no more than 4 minutes.

In light of this ramp metering shutdown, Kathleen Harder, John Bloomfield, Kasia Winiarczyk, and I conducted two experiments with the same framework where the essential difference was in how we asked the question2. The first is a classical Stated Preference experiment, administered on a computer, which to distinguish we called Computer Administered Stated Preference (CASP). Noting the weakness of stated preference is that people may respond differently to hypothetical situations, especially scenarios they have never experienced, we developed a second experimental method that we labeled a Virtual Experience Stated Preference (VESP).

In the Virtual Experience Stated Preference we put subjects in a driving simulator. This car, a 2002 Saturn Coupe, is on the third floor of the HumanFIRST Laboratory at the University of Minnesota’s Mechanical Engineering building. It is enveloped in a set of screens, and the driver is surrounded by animations which make it feel like driving. The simulator has speakers, and it vibrates, to add to the verisimilitude. When the driver turns the car, then the point of view changes. It’s a very sophisticated set up. (However, the 10 percent of the population sensitive to motion sickness are advised not to use driving simulators, and were excluded from the study. Subjects who completed the VESP said they reported no discomfort).

There were many different conditions that were tested. In one set of conditions, subjects had to rank four alternatives, which were exactly the same for the virtual and computer experiences. In the CASP, this was presented in a bar graph showing total travel time and the travel time of each component, with clear text labels.

  1. 0 minutes of ramp delay, 20 minutes of stop-and-go traffic at 30 MPH
  2. 2 minutes of ramp delay, 15 minutes of congested traffic at 40 MPH
  3. 4 minutes of ramp delay, 12 minutes of moderately congested traffic at 50 MPH
  4. 6 minutes of ramp delay, 10 minutes of freeflow traffic at 60 MPH

CASP Experiment

If you were strictly rational and cared only about minimizing total travel time you would prefer 10 miles in 16 minutes. But travelers had a mental model of what waiting at a ramp was like and a mental model of what driving in traffic was like, and only 1 of 44 subjects preferred to minimize total travel time in this CASP presentation.

In the VESP people nominally experienced the same conditions. However 15 of 17 subjects preferred the total travel time minimizing condition.

We estimated statistical models for the CASP experiment and found ramp time was about 1.6 times as onerous as freeway time. In contrast, when we estimated the model in the VESP experiment we obtained exactly the opposite result, that ramp time is preferred to freeway time. What explains these differences? Some explanations:

  • Simultaneity vs. Sequencing. In the CASP they were looking at all the four options simultaneously and they’re doing this in about 15 minutes, however long it takes them to read the screen for multiple presentations of similar questions. They’re evoking from their memory previous experiences about what travel conditions are like. In the VESP it is a longer time, the whole experience takes 90 minutes because we gave them those four presentations in sequence. They were sitting at the ramp meter and then driving through congested traffic, and we asked them questions at the end.
  • Recency. In the VESP there is a recency effect. The thing that you remember the most at the end of the trip is that you have just been through stop-and-go traffic. And you think to yourself, ‘I don’t like stop and go traffic’; whereas the ramp meter was a long time ago.
  • Simulator Realism. The stop-and-go traffic in the simulator may be more (or less) intense than how any particular traveler experiences traffic. In the CASP they’re remembering their actual work trips while.
  • Goal-directedness. In VESP travelers are in a driving simulator because the researcher tells them to be and not because they have any real goal at the end of it, except collecting the $25 for participating in this study. In the CASP, subjects may be recalling the goal of their own work trip.

Experiment 3: Information and Field Experiment Stated Preferences

In theory, nothing beats the real world, so what could better simulate different road conditions but driving in actual traffic? Kathleen Harder and I conducted a study to try and ascertain certain value of information and route choice. We asked people, to drive from McNamara Alumni Center on the University of Minnesota Minneapolis campus to the Cathedral in St. Paul by different routes. Each person was asked to go down one route and come back on another route and go down on a third route and back on a forth route, and after each of these rate those routes. The routes in the study were I-94, a depressed (and perhaps depressing) freeway connecting Minneapolis and St. Paul, University Avenue – an auto-oriented commercial, industrial, and retail street that had seen better days (this took place before recent construction to install a LRT along the corridor), Grand Avenue (a pedestrian oriented retail street), Summit Avenue (a beautiful boulevard lined with nice homes and mansions), and Marshall and Selby Avenues (a mix of residential and industrial uses). We had students collecting data and giving instructions at each end of the experiment. We provided some drivers the “expected” travel time of the route. The data collection was led by Brendan Nee (now a principal at BlinkTag and designer of a number of traveler information tools).

Map of route choice experiment

We had a GPS unit in the car so we knew exactly how long it took them. We randomized who took which route in which sequence. We estimated models that would predict which route people would prefer; we asked what route would you prefer for commuting, which route would you prefer for shopping trips or going to entertainment, and so on, with the idea that commuting trips would present a different set of preferences than for discretionary trips without the rigorous time constraints. Part of Lei Zhang’s dissertation work estimated this model and used a machine learning algorithm to identify the explanatory factors3. His dissertation goes on to develop a Behavior User Equilibrium, in contrast to the Wardropian User Equilibrium discussed in the previous post. The idea is that users are not minimizing travel time, but instead making choices that satisfy a set of heuristic rules which embed preferences for a larger set of factors, and that these behavioral factors need to be discovered empirically, through Search, Information, Learning, and Knowledge acquisition. This SILK Theory presents a very different paradigm for modeling route choices than traditional User Equilibrium, since it is a positive, empirical approach which describes what people actually do instead of a normative approach which ascribes to people what we believe they should do (in their own interest).

As part of this research, we discovered a set of decision rules that explain driver preferences. Drivers will switch to a new route if the difference in time is great, or if the difference in time isn’t so great but the pleasure obtained from the route is higher. Drivers prefer familiar routes. They consider aesthetics, e.g. Summit Avenue is a nicer road than University or Marshall Avenues to drive down. People don’t like stopping at stop signs, people like driving on routes that they’re familiar with, and of course people like to have information about how long it’s going to take (even if the information has only a passing resemblance to reality).

Experiment 4: HOT Lanes and Actual Commute Experience Revealed Preference

Running due west from Minneapolis, I-394 was the last Interstate Highway opened in the Minneapolis-Saint Paul region. As one of the last highways, it face controversies in development, which resulted in some innovations once built. One of those was the use of a reversible High Occupancy Vehicle (HOV) lanes on the eastern part of the route and regular (peak period, non-reversible) HOV lanes on the western part of the route. As with HOV lanes in many cities, they are designed to provide a travel time advantage to vehicles that carry more than one passenger, and of course to buses. However, it turns out that there is insufficient demand from HOVs in this corridor to fully utilize the capacity available in the HOV lanes, while the parallel (General Purpose) lanes remain congested during the peak periods. This is inefficient.

One solution would have been to open up the lanes to all vehicles, but this would remove any time savings for HOV vehicles and bus passengers. It would also make it difficult to restore HOV status to the lanes in the future, as that would be perceived as a taking-away capacity, and as Kahneman and Tversky point out in their work on prospect theory (corroborating experience in the transportation field from previous unsuccessful attempts to take-away general purpose capacity for HOV lanes), people are very loss-averse, more than gain-seeking. The implemented solution transformed the HOV lanes to High Occupancy/Toll (HOT) lanes.

HOT lanes allow vehicles equipped with transponders to use what were the HOV lanes for a price. This toll varies with traffic conditions, but aims to ensure that the HOT lanes maintain lower travel times than the general purpose lanes. Travelers in a hurry might be willing to pay a premium to guarantee they can avoid congestion. Other travelers won’t.

Kathleen Harder and I conducted a study on the I-394 corridor, with graduate students Shanjiang Zhu and Carlos Carrion (now a post-doc at SMART: Singapore MIT Alliance for Research and Technology), using a methodology we call Actual Commute Experience Revealed Preference. The aim was to combine the best of these studies, use each commuter’s actual origin and destination, but make sure they experience alternative routes and that they’re not just bound to the route that they normally travel on 4.

We recruited subjects who live along the I-394 corridor, people who lived out in Plymouth and Minnetonka and other western suburbs. They were required to have a daily route of at least 20 minutes so that the alternative commutes would make sense, that I-394 had to be a plausible alternative for them, they commuted on a regular basis, they worked near downtown Minneapolis, and that they drove a car alone. We obtained permission to install a GPS unit in their vehicle, and they had to follow our instructions about which route to take which weeks. In the end, we discarded the results from a number of subjects because they didn’t follow instructions.

We gave them three sets of alternative routes: one was the HOT Lane on 394; one would be the general purpose (un-tolled) lanes; and one would be a parallel arterial, which depended on where they lived, and which would be most feasible. We asked them questions every week about what they were doing, we installed a GPS unit in their vehicle, gave them two weeks of free choice in the beginning (to establish their baseline preference), for the middle six weeks we told them for each two week period please drive on a particular route. For the final two weeks, travelers were again free to pick a route, which allowed us to see if their behavior changed after the experience.

The official aim of this study was to ascertain the Value of Travel Time Reliability. This is, how much of a premium are drivers willing to pay to have a low variation in their travel time. The HOT Lanes offer a lower travel time, but more importantly, they offer a much more reliable or predictable travel time, it is much less likely to vary than the congested general purpose lanes, which might have an average delay of 2 minutes, but some days might be much much higher.

Depending on how we measure variability we get estimates of value of time ranging from $9 to $20 an hour. We also get a value of reliability ranging from $3.80 to $18.23 an hour. Other people have tried to estimate a reliability ratio in the literature, using a variety of methods, and the numbers vary widely, typically however they are on the order of 1.0, indicating that avoiding a minute of standard deviation is as valuable as avoiding a minute of expected travel time.


So people are choosing routes not just based on travel times, not just based on ramp time being different than freeway time, not just based on aesthetics and familiarity of routes, but also considering the travel time reliability – how predictable is the time and they might choose one route not because it’s the shortest time on average but because they have a high probability of not being late more than, say 5 percent of the time.

One of the things you learn when doing travel behavior experiments is that people don’t always respond the way you think they will. Travelers may not follow instructions on their daily commute, even though they agreed to. As the experiments with the driving simulator show, the methodology can greatly shape, and even reverse the result. So all experiments provide information, but none can be definitive. As Carl Sagan put it “Extraordinary claims require extraordinary evidence”.






People and their Paths 2: Why aren’t people taking the shortest path?

A shortened version of this post was adapted for Symposium Magazine‘s article Understanding the Irrational Commuter, which appeared in the September 2013 issue.

Recall in yesterday’s post we considered Wardrop’s User Equilibrium Principle, which says: the journey times on all routes actually used are equal and less than those which would be experienced by a single vehicle on any unused route. We then asked if people were taking the shortest travel time path, and came to conclude no.

This leads us to the next question:

Why aren’t people taking the shortest path?

Here are a few conjectures:

  1. Selflessness: The principle assumes that people are selfish, but perhaps they are selfless. We assume they aim to minimize their own travel time rather than society’s. We will come back to this point in Episode 5, but for now, let it suffice to say that people cannot know what decision will minimize society’s travel time, because of computational and informational issues, discussed below. Perhaps if they had that information, they might selflessly choose a different route. In the absence of that information, they are, at best, left guessing whether what they are doing is best for everyone else, even if at some self-sacrifice. (This assumes they are still making the trip at that time. In general, if there is congestion, it would be better for everyone else from a travel time perspective to avoid the trip altogether).
  2. Rationality: The principle assumes that people are rational, but maybe people aren’t rational, or at least not rational all the time. In one sense that is of course true, people react emotionally and intuitively, employing what Nobel Prize winner Daniel Kahneman calls System 1 in Thinking Fast and Slow 1, based on heuristic rules. They don’t have time for rational assessment. In another sense, for a repeated decision like commuting back and forth to work daily, it costs a significant amount of travel time, a scarce resource, to systematically behave irrationally. We thus assume people are behaving rationally (engaging Kahneman’s System 2) when they can.The idea of Bounded Rationality, developed by Herbert Simon, also a Nobel Prize winner, has been applied to route choice problems by many researchers, including my Master’s Advisor, University of Maryland Prof. Gang-len Chang in his dissertation work with Prof. Hani Mahmassani at the University of Texas. We can build models with bounded rationality assuming or estimating the bounds to this rationality due to information, cognitive limits, and time available to make a decision. We discuss some of the factors described below.
  3. Perception: It might be that people think they have the shortest travel time on their route, but they misperceived the travel time on the network.There are perception or cognition limits. On a 24 minute trip, are you going to know what the travel time is to the nearest 30 seconds or minute? I would because I’m a transportation geek, but most people aren’t going to measure their time that precisely. When you look at how people report travel times in surveys, they typically round to 5 minutes and sometimes they round to the nearest 15 minutes. If people are only dealing with time perception in 5 or 15 minute chunks, saving a minute or two isn’t going to show up on their radar as something that is important to them.There are many other aspects to the perception of travel time, which we will discuss more in episode 4.
  4. Computation: Sadly (for us modelers), people are not computers. They cannot accurately add travel times across different road segments, they can’t systematically compare the travel times over alternative routes even if they had a complete data set.
  5. Information: Not only are people not computers, they are not GPS systems either. People don’t have complete maps of the network; people often have good mental maps of the local street network around where they live and a little bit around where they work and where they travel frequently, but if they live far from where they work, they tend not to know the detailed network in-between. There are limits to people’s ability to navigate. People’s cognitive or mental maps are far from complete. They only have the experience of the routes they have actually used. They can test other routes to gain experiences but they don’t have those innately.
  6. Valuation: Maybe people are minimizing the weighted sum of travel time, where time spent in different conditions is valued differently. We know, for instance, from the transit literature that time spent waiting for a bus is much more onerous than time while on-board a vehicle in motion making progress towards its destination, especially if the arrival time of the bus is uncertain. In tomorrow’s post we will discuss valuation of travel time under various conditions.
  7. Objective: We assume that people care about minimizing only travel time. It might be that people are rational but they care about things besides or in addition to travel time.We have evidence from other transportation choices that people aren’t minimizing travel time. When you choose a place to live, you are not choosing to minimize your commute time to work. In fact there have been studies that have considered a hypothetical relocation of everybody’s place of residence in order to be in a house that was equivalent to the structure in which they currently reside, but was as close to their work place as possible (given everyone else was similarly moved), average commutes fell from about 24 minutes to 8 or 10 minutes. There is a significant amount of “excess travel” from a strict travel-time minimizing perspective. There are lots of reasons for excess travel, but the most obvious is that it is not excess from the point of view from people who are making it. They’re making home location decisions for a variety of reasons; the journey to work isn’t the only thing in their mind. (Travel time must be a consideration though, otherwise cities wouldn’t exist). It might be when choosing where to reside people might underestimate the amount of time that will be spent traveling, and probably underestimate the pain associated with long commutes, and are thus unhappier than expected. Stutzer and Frey call this phenomenon The Commuting Paradox. A major source of time estimation error arises because most people search for homes on the weekend, but tend to commute on weekdays.Some candidate factors for route choice are given below:
    • Search cost: How long does it take to figure out what the travel time is on alternative routes? Are you willing to spend ten minutes exploring the network in order to save 30 seconds of travel time every day for the rest of your career? Rationally it might be worth doing so, the payback is only 20 days. People often will discount the possibility of saving time, worrying that this short-cut will actually be longer, or maybe they’re afraid of getting lost. Fear of the unfamiliar is a major deterrent to exploration.
    • Route quality: Many factors describe the quality or condition of a route and its environment. Is it potholed or newly paved? Does it run through a pleasant or unpleasant neighborhood? We have evidence that some people prefer a longer route if it’s an attractive boulevard or parkway rather than drive through a freeway trench.
    • Reliability: The likelihood of arriving on time, and not just the expected travel time, affects willingness to select a route. There is the old parable of the man who drowned in an average of one inch of water. Similarly, it might not matter to me that the average travel time is 20 minutes if one day a week (but never knowing in advance which day) I can expect a travel time of 60 minutes. I don’t want to leave 40 minutes earlier to avoid the occasional bad outcome. I might be willing to take a slower but more reliable route. I might even have a mixed strategy, or portfolio, combining different routes in order to achieve a personally satisfactory trade-off between expected time and reliability. In practice this means some people might take surface streets, which are generally slower, but more reliable, instead of freeways, which are faster, but subject to more catastrophic breakdowns of traffic flow.
    • Pleasurability of travel: Maybe people are rational, but they like traveling a little bit more than being at work or home, and so choose longer routes to prolong the experience.Many people want to commute; Redmond and Mokhtarian find there’s a positive value to some amount of commuting, that the preferred commute length is not typically zero. However, it appears that many commutes are longer the desired amount. However for some people, the longer route, which provides some psychological buffer between the stresses of work and the stresses of home, is desired.

The next two posts will consider issues of valuation and perception in more detail.

1. The terminology “System 1″ and “System 2″ apparently derives from Stanovich, K E.; West, R F. (2000). “Individual difference in reasoning: implications for the rationality debate?“. Behavioural and Brain Sciences 23: 645–726. The idea itself is much older.