The value of time and reliability: measurement from a value pricing experiment

Abstract

We measure values of time and reliability from 1998 data on actual behavior of commuters on State Route 91 in Orange County, California, where they choose between a free and a variably tolled route. For each route at each time of day and for each day of the week, the distribution of travel times across different weeks is measured using loop detector data. The best-fitting models represent travel-time by its median, and unreliability by the difference between the 90th percentile and the median. We present models of route choice both alone and combined with other choices, namely time of day, car occupancy, and installation of an electronic transponder. In our best model, containing all these choices except time of day, value of time (VOT) is $22.87 per hour, while value of reliability is $15.12 per hour for men and $31.91 for women. These values are 72%, 48%, and 101%, respectively, of the average wage rate in our sample.

Introduction

Recent policy innovations regarding highway congestion underscore the importance of knowing how travel-time and its reliability are valued by travelers. These innovations include experiments with congestion pricing and its cousin, value pricing,¹ as well as applications of information technology. All of them require, for their design and evaluation, knowledge of how travelers will react to time-varying toll schedules and/or how they react to and evaluate changes in the extent and predictability of congestion.

Although the value of time (VOT) has been thoroughly studied, full consensus on many issues has not been achieved.² Furthermore, only a few empirical studies of VOT make use of information on road users' reactions to tolls, and even fewer do so with actual as opposed to hypothetical tolls. The value of reliability (VOR) has received much less attention.³ Virtually all the work on it has used data related to hypothetical scenarios, for two reasons: measuring the variability of travel times facing actual travelers is difficult, and travel-time variability is highly correlated with mean travel time.

In this paper, we simultaneously measure VOT and VOR using data on actual travel behavior in a real pricing context. We observe people who face a choice between two parallel routes, one free but congested and the other with time-varying tolls. We do this by taking advantage of a nearly unique experiment in Orange County, California, on a major commuting highway known as State Route 91 (SR91).

In late 1995, a new privately constructed set of toll lanes in the center of SR91 (also known as the Riverside Freeway) opened, with tolls varying over time according to a preset schedule. This corridor connects fast-growing residential suburbs in Riverside and San Bernardino Counties to job centers in Orange and Los Angeles Counties. There are five free lanes and two toll lanes (called “express lanes”) in each direction; we refer to the choice between them as “route choice” although in fact they are part of the same highway.

By summer 1998, when our data were collected, the toll had evolved to a highly sophisticated one; for example, people driving west (toward job centers in Orange and Los Angeles Counties) faced 12 different toll levels applying to different time periods, all identified on a published schedule. (This does not include the zero toll level, which continues to apply to the public lanes during all time periods.) Vehicles with three or more occupants are charged half the published toll. Our data come from mail surveys of people making work trips on the corridor for the entire 10-mile (16 km) length of the demonstration project (hence not using the few intermediate entrances and exits).⁴

There are two main difficulties with estimating VOT and VOR from such data. First, the main variables of interest – differences across the two routes in time, reliability, and cost – vary across times of day and days of the week, but in a highly correlated manner by design. Second, the survey responses must be supplemented by other data in order to accurately measure travel time and its uncertainty.

Kazimi et al. (2000) faced comparable difficulties with data from a value-pricing experiment in San Diego, and we follow a similar strategy. We overcome the second difficulty using data laboriously extracted from loop detectors placed by the California Department of Transportation (Caltrans) in both the free lanes and the express lanes. A serious limitation is that inability to obtain satisfactory data for 1998 has forced us to use data from the same months in 1997, then apply an adjustment factor to account for growth in travel congestion in the free lanes between 1997 and 1998. Even so, these data are superior to either of the two most commonly used data sources on travel times in studies of actual behavior, namely survey-based estimates (which are subject to serious perception errors)⁵ and network-based estimates (which usually cannot describe fine variation by time of day). Furthermore, we are able to measure the distribution of travel time across weeks, for a given 15 min time-of-day interval and a given day of the week, for up to 10 weeks during the summer months.

We overcome the first difficulty, that of correlation among variables, in several ways. First, by measuring travel times for relatively narrow (15 min) time-of-day intervals, we take advantage of substantial variation in the degree of congestion on the free route across the 4 h peak period (5–9 am), during which tolls on the express lanes are at a nearly constant level. Second, the ratio of toll to travel-time savings is higher in the shoulders of the peak (4–5 am, 9–10 am) than in the peak itself, and higher still in the mid-day and night periods; by including work trips that occur during these off-peak periods, we obtain additional independent variation. Third, our measurements reveal that mean (or median) travel time and variability in travel time are only imperfectly correlated across time-of-day intervals; for example, variability is especially high late in the peak period. Fourth, carpooling introduces additional variation into the cost per person of the toll, both because the toll can be shared among occupants and because carpools of three or more receive a 50% discount. Finally, one expects the VOT and VOR to depend on certain measurable socio-economic and travel characteristics; by specifying interactions between travel variables and these characteristics, we obtain additional independent variation in the variables entering the model.

The result is that we have found many model specifications that fit well and result in statistically significant coefficients on all three of the key travel variables. We thereby obtain credible estimates of VOT and VOR which vary in plausible ways with traveler characteristics.

In Section 2, we describe our data more fully. Subsequent sections report the results of various models that include route choice, first by itself and then simultaneously with related decisions – namely time of day, car occupancy, and installation of an electronic transponder (which is required to use the toll route). As it turns out, our estimates of VOT and VOR are fairly robust across different assumptions about the simultaneity of such decisions, with the exception of time of day which we are not able to treat very satisfactorily.