Get documents about "Gasoline Prices, Driving Ev ciency and Implications for Gasoline
Document Sample
Gasoline Prices, Driving E¢ ciency and
Implications for Gasoline Tax Policy
Jack Erb
University of California, Santa Barbara
Email: erb@econ.ucsb.edu
Job Market Paper
November 1, 2009
Abstract
We provide new evidence on how motorists adjust their travel speeds to
changes in gasoline prices. Previous studies have focused primarily on the
travel speeds of urban tra¢ c in major cities, …nding that tra¢ c speeds increase
with gasoline prices due to reduced congestion. In this paper, our focus is
on rural and suburban highways. We …nd that on uncongested, rural high-
ways, increases in gasoline prices lead to a uniform decrease in the distribution
of vehicle speeds. However, in more congested, suburban areas, increases in
gasoline prices lead to a widening of the speed distribution, as those previously
constrained by congestion are able to increase speeds. The results also have
implications for the use of gasoline tax policies that aim to reduce fuel con-
sumption and vehicle emissions. Overall, the "e¢ ciency" e¤ect can account
for up to 20% of the price elasticities of gasoline demand and emissions.
I would like to thank Doug Steigerwald, Jon Sonstelie, Olivier Deschenes, Kelly Bedard and
members of the UCSB Econometrics Research Group for helpful comments. I am especially grateful
to Ben Hansen for providing the speed data used in the analysis.
1
1 Introduction
In late 2008, the national average gasoline price fell from an all-time nominal high of
$4.17 to a December low of $1.67 –the lowest point in 5 years. Many environmental-
ists and policy makers saw this fall in gasoline prices as an opportunity to call for an
increase in gasoline taxes. Economic theory would suggest that increasing the price
of gasoline would impact not only the number of miles a motorist travels, but also
the e¢ ciency with which those miles are driven (e.g., driving at fuel-e¢ cient speeds,
avoiding periods of high congestion, etc.). However, recent research has found lit-
tle to no evidence of the latter behavior; the data typically show that individuals
will drive fewer miles in response to increases in gasoline prices, but they do not
appear to drive those miles at more e¢ cient speeds (at least when driver speeds are
unconstrained by congestion).
In this paper, we reexamine how travel speeds and tra¢ c volumes are a¤ected
by changes in gasoline prices, focusing exclusively on rural (and suburban) highway
tra¢ c outside of major city centers. The data employed come from the state of
Oregon during the years 2003-2007. In contrast to previous work, we …nd that
motorists on uncongested highways drive at more fuel-e¢ cient speeds when gasoline
prices are higher. However, consistent with previous work, we …nd drivers may
increase travel speeds when higher gasoline prices alleviate congestion. Our results
indicate that driving e¢ ciency may account for up to 15%-20% of the overall gasoline
price elasticities of fuel consumption and vehicle emissions. In addition, we …nd that
for certain speed/volume combinations the "e¢ ciency" e¤ect may be negative –that
is, higher gasoline prices will reduce the total number of miles driven, but driving
e¢ ciency may decrease if drivers increase speeds due to the reduced congestion.
The current U.S. gasoline tax rate of 18.4 cents per gallon is quite low relative to
2
tax rates in the rest of the world, and it has remained …xed since 1993.1 Attempts
to increase the federal gasoline tax since 1993 have been met with heavy political
and public opposition. However, proponents of increased gasoline taxes have once
again gained traction with the late-2008 drop in gasoline prices.2 While many of
these groups are advocating higher gasoline taxes primarily as a means of …nancing
highway construction and maintenance, the improved environmental quality is often
put forth as a secondary bene…t.
If the decision is to be made to increase federal or state gasoline taxes, it is
important to understand exactly how drivers respond to gasoline price increases.
Numerous studies of gasoline demand have focused on estimating how fuel prices
a¤ect gasoline consumption and vehicle miles traveled (VMT) (in addition to various
eet
characteristics of the vehicle ‡ such as turnover, scrap rate, stock, etc.). In a
meta analysis of 69 studies since 1992, Goodwin et al. (2004) estimate an average
short-run elasticity of gasoline demand of 0:25, with a long-run elasticity of gasoline
demand of 0:64. The short- and long-run elasticities of VMT with respect to the
price of gasoline are estimated at 0:10 and 0:29, respectively. The di¤erences
between the elasticities of demand for gasoline and VMT indicate that individuals
will reduce fuel costs as prices rise in ways other than simply deferring trips. Hinting
at this possibility, Goodwin et al. conjecture that "the reason why fuel consumed
falls by more than the volume of tra¢ c is probably because price increases trigger
1
In 2008, the average U.S. citizen paid roughly 40 cents per gallon in state and federal gasoline
s
taxes –the lowest gasoline tax rate among industrialized countries. In contrast, Britain’ tax rate
of 53.6 pence per liter –roughly $3.00 per gallon –was the highest (Parry and Small, 2005).
2
Two legislative commissions recently advanced proposals to signi…cantly increase the federal
gasoline tax. The National Commission on Surface Transportation Infrastructure Financing issued
a report to Congress in January 2009 which called for a 10 cent increase in the federal gasoline
tax rate, as well as a recommendation to tie the tax rate to in‡ ation. The National Surface
Transportation Policy and Revenue Study Commission also recommended a 25 cent increase in the
federal gasoline tax rate (at 5 cents per year over a …ve year period). In addition, politicians in many
states (such as California, Massachusetts, New Hampshire, Illinois, and Oregon) have introduced
bills that would increase their respective state gasoline taxes.
3
a more e¢ cient use of fuel (by a combination of technical improvements to vehicles,
more fuel-conserving driving styles and driving in easier tra¢ c conditions)." Indeed,
economic theory predicts that drivers should respond to increases in gasoline prices
by increasing driving e¢ ciency. However, the manner in which drivers alter their
travel speeds to achieve greater fuel e¢ ciency has gone largely unstudied. Although
driving e¢ ciency is likely to be a second-order e¤ect, it may constitute a non-negligible
fraction of the total change in gasoline demand as gasoline prices change.
In the …rst panel-data study of the impact of gasoline prices on vehicle speed,
Burger and Ka¢ ne (2009) …nd that tra¢ c speeds on congested, Los Angeles freeways
actually increase with increases in gasoline prices.3 They attribute this …nding to
reductions in congestion and estimate that a $1.00 increase in gasoline prices will
increase average freeway speeds by 3.37 mph during the traditional "rush" hours.
Estimating the e¤ect during uncongested periods, however, is more problematic as
Los Angeles tra¢ c rarely ‡ ow
ows at free-‡ speeds during the day. Consequently,
Burger and Ka¢ ne turn to the early-morning hours of 2 to 4 a.m. They …nd that
early-morning drivers do not alter their travel speeds with changes in gasoline prices,
and take this as evidence that drivers on uncongested highways are unresponsive to
prices. However, the result could also be due to the fact that drivers on the highway
between 2 and 4 a.m. are not representative of the "typical" or "average" driver.
We reexamine how drivers respond to gasoline prices during uncongested periods
by analyzing travel speeds on rural (rather than urban) highways. We …nd that a
$1.00 increase in gasoline prices reduces travel speeds on uncongested, rural highways
by about 1.25 mph on average. However, similar to Burger and Ka¢ ne, we …nd a
positive e¤ect for travel speeds on more congested, suburban highways, with average
3
In an early study, Dahl (1979) performs a cross-sectional analysis for one year of data at the
state level and …nds the elasticity of roadway speeds with respect to gasoline price to be -0.354.
4
speeds increasing by 0.38 mph for every $1.00 increase in price.
s
Our analysis relies on speed data gathered from Oregon’ rural and suburban
highways. The unique nature of the data allows us to make at least three contribu-
tions to the existing literature. First, previous studies have focused on the impact of
gasoline prices on mean driver speed. In addition to the mean, we are able to observe
the entire distribution of travel speeds, which allows us to measure how di¤erent per-
centiles of the distribution respond to changes in gasoline prices. The distributional
analysis gives us a clearer picture of the observed increases in mean speed on con-
gested roads when gasoline prices increase. Second, rural and suburban tra¢ c tends
ow ow
to ‡ at free-‡ speeds during most hours of the day, not just during the midnight
and early-morning hours. s
We therefore observe the driver’ unconstrained speed
choice at all hours of the day and can determine, for example, whether motorists
traveling at noon exhibit the same response to gasoline prices as motorists traveling
at midnight. Last, we can observe whether rural and suburban drivers respond to
gasoline prices in ways that di¤er from the responses of urban drivers by comparing
our results to the results of previous studies. The di¤erence between the behavior
of rural and urban drivers is important because the Federal Highway Administration
estimates that between 30 and 40 percent of total U.S. VMT occurs on rural roads.
Although this fraction is less than half, it still represents a signi…cant portion of total
U.S. vehicle travel. Any di¤erence between how rural and urban drivers respond to
gasoline prices would have implications for tax policy.
2 Gasoline Prices and Travel Costs
We begin with a discussion of some recent trends in U.S. gasoline prices and how
gasoline prices a¤ect the costs associated with motor vehicle travel. Figure 1 plots
5
Figure 1: US Average Gasoline Price
the trends in average retail gasoline prices in the U.S. since 1993.4 In 1993 –the last
year the federal gasoline tax changed – federal and state gasoline taxes accounted
for approximately 40% of the price of gasoline. This percentage fell to less than
10% when gasoline prices peaked in the summer of 2008. Overall, gasoline prices
have shown increased variability in recent years, spiking sharply during the summer
months.
Increased gasoline prices increase the costs of motor-vehicle travel, including the
marginal cost of driving at fuel-ine¢ cient speeds. Figure 2 depicts the relationship
between fuel consumption (measured in gallons per 100 miles) and vehicle speeds for
a sample of light-duty vehicles.5;6 The solid line graphically reproduces the composite
4
The monthly gasoline price data come from the Energy Information Administration and are
adjusted for in‡ ation using the Seasonally-Adjusted Consumer Price Index with base month Dec.
2008. The CPI data come from the FRED database of the Federal Reserve Bank of St. Louis.
5
The data for Figure 2 come from West et al. (1999), who measure fuel consumption for various
light-duty vehicles under simulated driving conditions. The plotted relationships are measurements
under a constant travel speed constraint (i.e., zero acceleration).
6
There are many other studies of how fuel consumption and vehicle emissions are related to
6
fuel-consumption estimates for light-duty vehicles from West et al. (1999), and the
dashed line gives the predicted values from a linear regression of fuel consumption on
a polynomial of vehicle speed.7 The average light-duty vehicle achieves fuel e¢ ciency
at about 50-55 miles per hour, and fuel consumption increases quite rapidly as speed
deviates from the fuel-e¢ cient speed. Traveling at 75 miles per hour increases fuel
consumption by about 30% relative to traveling at fuel-e¢ cient speeds. With an
average U.S. gasoline price of $2.44 per gallon between 2003 and 2008, this would be
equivalent to paying an additional $0.23 cents per gallon (an almost 10% increase)
for each 5 mph increase in speed above 60 miles per hour.8 Consequently, reducing
travel speeds is one method motorists could employ to defray the costs of highway
travel as gasoline prices rise.
Burger and Ka¢ ne (2009) analyze the relationship between vehicle speeds and
gasoline prices on Los Angeles highways. They …nd evidence of fuel-conserving
behavior during congested periods when speed is determined exogenously by the
total tra¢ c volume.9 However, they …nd no evidence of fuel-conserving behavior
when it might be most expected –during uncongested periods when individuals are
free to select their own speeds. This could be for a number of reasons. First, the
various driving characteristics such as speed, acceleration, breaking and speed variance (see e.g.,
Hansen et al., 1995; Jensen, 1995; Joumard et al., 1995). In addition, the EPA has developed
their own modeling software (MOBILE6) that simulates vehicle emissions under di¤erent driving
conditions (see http://www.epa.gov/OMS/m6.htm).
7
That is, the predicted values of the West et al. models come from estimating the regression
model
X6
Gallons = Ak Speedk + !;
k=0
where Gallons is the number of gallons consumed per 100 miles of travel. The R2 from this
regression is 0.976. The predicted values will be used in section 5 to estimate the slope of the
fuel-consumption function.
8
See http://www.fueleconomy.gov/feg/driveHabits.shtml.
9
Tra¢ c volumes fall as gasoline prices rise resulting in increases in average travel speeds during
rush hour. As the mean rush-hour speed in their sample is 46.4 mph, increases in speed are likely
to represent increases in fuel economy.
7
Figure 2: Fuel Consumption of Light-Duty Vehicles
4.4
4.2
Gallons per 100 miles
4
3.8
3.6
3.4
Source: West et al. (1999) Fuel Consumption
Predicted Consumption
3.2
20 30 40 50 60 70 80
Speed (mph)
8
congestion that is typical of L.A. freeways during the day necessitates that the authors
ow
use speeds during the 2-4am period as their measure of free-‡ speeds. However, the
drivers on the road from 2-4am may be quite di¤erent than the "average" or "typical"
L.A. driver. Second, as noted by Burger and Ka¢ ne, it may be the case that the
value of time for L.A. motorists, in general, is su¢ ciently high that travel speeds are
constrained by factors other than gasoline prices (such as speed-limit enforcement or
accident risk). If this is the case, motorist travel speeds on uncongested highways
would be una¤ected by changes in gasoline prices. The next two sections explore
the relationship between gasoline prices and travel speeds on uncongested highways
in more detail.
3 Model and Data
The goal of this paper is to empirically estimate how motorists adjust their driving
behavior with changes in the level of gasoline prices.10 We estimate this relationship
using the following regression model:
Speedit = 0 + 1 P Gast + x0it + h + d + m + y + i + uit (1)
where Speedit is the average travel speed on highway segment i at time t, and P Gast
is the price of gasoline in time t. The vector of controls, xit , includes the percentage of
total tra¢ c which are small, light-duty, passenger vehicles (as opposed to large, heavy-
duty, commercial vehicles) on roadway i at time t, and various weather variables to
account for road conditions on highway segment i at time t. The model also includes
10
s
A theoretical model of the driver’ speed decision (both with and without congestion) is de-
veloped in the working paper version of Burger and Ka¢ ne (2009). The model predictions are as
expected: Drivers will alter their travel speeds in order to increase fuel e¢ ciency as gasoline prices
increase (e.g., by reducing speeds on uncongested highways and increasing speeds when higher prices
alleviate congestion).
9
…xed e¤ects for each highway ( i ) and year ( y ), as well as seasonal indicators for
each hour-of-day ( h ), day-of-week ( d ), and month-of-year ( m ). The units on the
dependent variable are miles per hour, and the coe¢ cient of interest is 1.
It would be reasonable to assume that the average speed on the highway could also
depend on tra¢ c volume if the number of vehicles was su¢ cient to induce congestion.
Therefore, the coe¢ cient 1 is best thought of as the reduced-form e¤ect of the
gasoline price on average vehicle speed, including the e¤ect of tra¢ c congestion.
This reduced-form e¤ect may be signi…cantly di¤erent than the structural relationship
between gasoline prices and speed if congestion is signi…cantly heavy. On the other
hand, when tra¢ c volume is low, congestion is likely to have little impact on the
average speed, and the reduced form e¤ect may approximate the structural e¤ect.
For our analysis, we are primarily interested in the reduced-form e¤ect.
The econometric analysis employs a unique data set which brings together speed,
weather, and gasoline price data from the state of Oregon in the years 2003-2007.
Data on vehicle speeds were obtained from the Oregon Department of Transporta-
tion (ODOT). The Department of Transportation records information on individual
vehicle speeds at the hourly level by use of automatic tra¢ c recorders (ATRs) at 18
s
di¤erent locations along Oregon’ major rural highways. ATRs are embedded in the
roadway and record raw counts of the number of vehicles that pass over the recorder
within given speed intervals/bins over the course of each hour. From these data, we
construct the average travel speed (along with other moments of the speed distribu-
tion) of all vehicles passing over a given ATR for every hour of every day from June
2003 to December 2007.11
Figure 3 displays the locations of the 18 ATRs, with ATRs that are near to and
11
A few ATRs have occasional gaps in the data due to mechanical malfunction. However, the
longest such gap is less than 30 days in duration.
10
s
Figure 3: Locations of Oregon’ Automatic Tra¢ c Recorders
far from major cities (explained in more detail below) indicated by a square and
triangle, respectively. Seven ATRs collect data along Interstate 5, which runs in
a north-south direction in western Oregon from the California border to the city of
Portland. Seven ATRs collect data along Interstate 84 which runs from the Idaho
border (near Boise) in a primarily east-west direction along the Columbia River to
Portland. Of the remaining four ATRs, two are located on S.R. 22 which runs in an
east-west direction through Salem, one is located on US-97 which runs in a north-
south direction through central Oregon, and the last is located on Interstate 82 in
northeastern Oregon which connects the tri-cities area of Washington to I-84.
11
The weather data come from the Climate Data Online (CDO) database of the
National Climatic Data Center.12 Hourly surface data were obtained for a number
of weather stations in the state of Oregon (most often regional, national, and interna-
tional airports). The latitude and longitude of each weather station was used to link
hourly weather conditions to the nearest ATR.13 We employ three weather variables
in our analysis: an indicator for precipitation, an indicator for freezing temperatures,
and the interaction between the two.
Last, the gasoline price data come from the Energy Information Administration
(EIA) and are measured at the weekly level. Every Monday, retail gasoline prices are
collected from approximately 900 stations across the U.S. by the EIA.14 From this
sample, average weekly gasoline prices are computed for various geographic regions.
Unfortunately, the EIA does not directly compute a price series for the state of
Oregon. While various Oregon retail stations are sampled, their data are used to
construct estimates of more broadly de…ned regions (e.g., U.S., West Coast, and West
Coast less California).15 The EIA does, however, compute average gasoline prices
for the state of Washington and we use Washington retail gasoline prices as a proxy
for Oregon gas prices in our analysis.
12
These data can be found at the CDO website at http://cdo.ncdc.noaa.gov/CDO/cdo.
13
We use the "great circle" distance formula to calculate the minimum distance between two
points on a sphere. Speci…cally, the distance, D, between points A and B, is calculated as
1
D = cos [sin (latA ) sin (latB ) + cos (latA ) cos (latB ) cos (jlongA longB j)] ;
where lati and longi are the latitude and longitude of point i measured in decimal degrees.
14
Retail gasoline stations selected for the EIA sample are required by law to report (P.L. 93-275).
15
The West Coast less California index includes price data from Arizona, Oregon, Washington,
Alaska, and Hawaii.
12
3.1 Measurement Error and Instrumental Variables
Clearly, the use of Washington gasoline prices will introduce error to the primary
regressor of our analysis. If the structure of the Washington gasoline market is
systematically di¤erent from the Oregon market, the Washington gasoline prices may
be a poor proxy for Oregon gasoline prices, which could lead to inconsistency in our
estimator of 1. More importantly, the use of Washington (rather than Oregon)
prices will not be the only source of error in the gasoline price variable, and may
not even be the most signi…cant. Ideally, we would like to observe the price that
each individual motorist pays for gasoline, but even if we did have Oregon prices,
measurement error would still be likely. The reason is that we would only be able to
observe a single price for the entire state, but gasoline prices can vary substantially
among gas stations within the same city, and even more so at the state level. Gasoline
prices also tend to be higher in rural areas compared to urban areas, and, in addition,
we can not be certain that the individuals purchased their gasoline within the same
week as the computed average. All of these factors introduce error of some form or
another to the regressor of primary interest, and the error must be accounted for in
order to consistently estimate the e¤ect of gasoline prices on travel speeds.
One method of controlling for the measurement error in the gasoline price regressor
is through the use of an instrumental variable (see e.g., Griliches and Mason, 1972;
Blackburn and Neuman, 1992). We employ the U.S. free-on-board, crude oil spot
price as an instrument. The U.S. spot price for a barrel of crude oil is calculated
weekly by the EIA. The use of the crude price as an instrument for retail gasoline
prices will lead to a consistent estimate of 1 as long as it is correlated with gasoline
prices, but uncorrelated with both the error in equation (1) and the measurement
13
error in gasoline prices. Speci…cally, suppose
P Gast = P Gasit + vit ;
where P Gast is the observed value of the gasoline price regressor, P Gasit is the
"true" value of Oregon gasoline prices for drivers on highway i in time t, and vit is
the measurement error. The U.S. crude oil price (P Crudet ) is a valid instrument
for the observed gasoline price (P Gast ) if the following three conditions hold: (1)
Cov (P Crudet ; P Gast ) 6= 0; (2) E (P Crudet uit ) = 0; and (3) E (P Crudet vit ) = 0.
The …rst condition can be empirically veri…ed, and (as will be shown below) U.S.
crude prices are highly correlated with gasoline prices. Although the redundancy of
crude prices implied by condition 2 cannot be directly tested, it seems unlikely that
s
crude oil prices would factor into a motorist’ speed decision. One exception may
be during periods when crude prices peak and the record-high prices are reported in
the media. In these instances, crude prices may be highly visible and drivers may
anticipate corresponding peaks in gasoline prices. Nevertheless, for most of the year,
gasoline prices are likely more visible than crude prices for the majority of motorists,
and the remainder of this analysis maintains the assumption that crude oil prices are
excludable from (1). Condition 3 is also likely to hold. EIA calculates crude prices
using a di¤erent survey design and methodology than that used to calculate retail
gasoline prices, so crude prices should not be correlated with any survey-induced
measurement error in gasoline prices.16
16
Condition 3 rules out the use of other EIA gasoline averages as instruments (e.g., California, West
Coast, or U.S. average gasoline prices). Although these variables are likely to satisfy conditions 1
and 2, they are unlikely to satify condition 3. All EIA price averages are all computed from the same
overall sample, therefore, the Washington price average likely su¤ers from the same survey-sampling
error as the other price averages.
14
3.2 Descriptive Statistics
Table 1 contains descriptive statistics for the principle variables in our analysis. Our
sample consists of 580,191 observations measured at the hourly level for the 18 dif-
ferent Oregon highway segments. The average vehicle speed on these highways in
a typical hour is 63.3 miles per hour, with 1164 vehicles per hour passing over the
ATRs on average.17 About 70% of all recorded vehicles are classi…ed as "small"
(e.g., light-duty passenger vehicles) and the remaining 30% are classi…ed as "large"
(e.g., heavy-duty 18-wheelers, RVs, buses). Gasoline prices over the sample period
run from a low of $1.54 per gallon to a high of $3.46 per gallon, with an average
of $2.43 per gallon in 2006 dollars18 . The weather variables Precip it , T emp32it ,
and Precip32 it are binary. Precip it is equal to one if there was at least a trace of
recorded precipitation in time t at the weather station nearest to roadway segment i.
Likewise, T emp32it is equal to one if the recorded temperature in time t at roadway
segment i was below 32 degrees Fahrenheit. Approximately 16% of the roadway-
hours in our sample record at least a trace of precipitation and approximately 6%
record temperatures below 32 F. Last, the variable Precip32 it is the interaction be-
tween precipitation and freezing temperatures, which occurs less than 1% of the time
in our sample.
Columns 2 and 3 of Table 1 report the variables conditioned on whether the
tra¢ c recorder is located on the outskirts of a major city. In our data set, 4 of
the 18 ATRs are classi…ed as urban principle arterials under the National Highway
17
The total tra¢ c volume in a given hour represents the total number of vehicles passing over the
tra¢ c recorder for all lanes in both directions of travel. Every highway in our sample consists of
two tra¢ c lanes in each direction of travel (for a total of 4 lanes), with the exception of one location
near Portland which has three lanes of travel in each direction. Thus, the average tra¢ c volume of
1164 vehicles per hour indicates an average of approximately 4.85 vehicles per lane-minute passing
over the ATRs.
18
All prices are adjusted for in‡ ation (base year 2006) using the seasonally-adjusted consumer
price index from the FRED database of the Federal Reserve Bank of St. Louis.
15
Table 1: Descriptive Statistics
(1) (2) (3)
City i = 1 City i = 0
Variable Name Notation Mean Std. Dev. Mean Std. Dev. Mean Std. Dev.
Average Speed Speedit 63.33 3.52 62.89 3.04 63.69 3.82
Gasoline Price P Gast 2.43 0.53 2.45 0.53 2.42 0.52
US Crude Price P Crudet 48.60 15.78 48.02 15.65 49.21 15.91
16
Total Volume V olit 1163.96 1165.63 1729.94 1468.25 715.58 521.98
Percent Small Veh. Smallit 69.10 17.94 77.25 14.01 62.63 18.02
Precipitation > 0 Precip it 0.157 0.363 0.158 0.365 0.155 0.362
Temp. < 32 F T emp32it 0.063 0.244 0.055 0.227 0.070 0.256
Precip it T emp32it Precip32 it 0.007 0.084 0.008 0.090 0.006 0.079
Near City City i 0.442 0.497 – – – –
N=580,191 N=256,461 N=323,730
Functional Classi…cation system. The remaining 14 ATRs are classi…ed as rural
principle arterials. However, of the 14 "rural" highway segments, 5 lie within a few
miles of urban city limits and experience tra¢ c patterns that more closely resemble
the tra¢ c patterns of urban/suburban areas than those of rural areas. Even though
these ATRs lie in areas that are "technically" considered rural, it may be the case
that they are picking up the tra¢ c counts from those individuals who work in the
city, but choose to live far from the city center. This can be seen empirically, as
the rural ATRs that lie close to major cities (e.g., Portland, Eugene, Salem) exhibit
signi…cantly larger tra¢ c volumes, especially during traditional rush hours.
In order to better understand how tra¢ c congestion a¤ects the relationship be-
tween gas prices and travel speeds, we distinguish between the ATRs located near
cities and those that are not. Speci…cally, we consider a tra¢ c recorder as being
"near" a city if it lies within 5 miles of the city limits of a city with a population ex-
ceeding 25,000, or inside the city limits of a city with a population exceeding 15,000.19
These tra¢ c recorders are indicated by the binary variable City i . In Figure 3, ATRs
located near major cities are indicated by squares, and those located in rural areas
that are far from city limits are indicated by triangles. The near-city/suburban
designation allows us to view the distinction between congested and uncongested pe-
riods from a spatial angle, rather than the temporal angle used by Burger and Ka¢ ne
(2009). While Burger and Ka¢ ne looked at the di¤erence between rush-hour and
early-morning (2-4am) tra¢ c ‡ows within a major metropolitan area, we look at the
di¤erence between rural and suburban highway segments across all hours of the day.
Approximately 44% of all highway-hours in our sample are taken from ATRs lo-
19
The one exception to this classi…cation is the ATR that lies at the midpoint of the 30 mile seg-
ment of I-5 that connects the cities of Eugene (pop. 143,910) and Albany (pop. 43,600). Although
this ATR lies more than 5 miles from either city limit, tra¢ c ‡ows over this ATR during rush hour
are substantially higher, indicating we may be picking up commuter tra¢ c between the two cities.
17
cated near major cities (see Table 1). The data indicate that there is little di¤erence
between the weather patterns of suburban and rural ATRs.20 However, average
vehicle speed and total tra¢ c volume exhibit statistically signi…cant di¤erences be-
tween suburban and rural ATRs. The di¤erence is especially pronounced for tra¢ c
volume— where suburban highways average over 1,000 more vehicles in the typical
hour. The vehicle speeds near cities tend to be lower than the vehicle speeds in
rural areas by about 0.8 mph. Given that tra¢ c volumes are about 150% higher in
suburban areas, it is somewhat surprising that the di¤erence in speed is not larger.
However, the unconditional di¤erence in average vehicle speed does not account for
the di¤erences in vehicle composition, as the proportion of large vehicles on the road
is about 10% higher in rural areas. Controlling for vehicle composition increases the
di¤erence in average speed between the suburban and rural roads to 1.10 mph.21
4 E¤ect of Gasoline Prices on Travel Speeds
We report the 2SLS estimated coe¢ cients and standard errors for equation (1) in
Table 2. We estimate the model using the average gasoline price from Washington
as a proxy for the average gasoline price in Oregon. In addition, we employ the U.S.
crude oil spot price as an instrument for the gasoline price to ensure measurement-
error consistent estimates of the parameters. In our model, the dependent variable
is measured at the hourly level, but gasoline prices are measured at the weekly level.
In general, estimated standard errors will be biased downward when aggregate ex-
20
Gasoline prices di¤er across the city designations due to the unbalanced nature of the panel.
Recall that gasoline prices are measured at the weekly level, and take on the same value for every
observation in the sample within that week.
21
The di¤erence in average speeds between near-city and rural highways could also be due to
di¤erences in speed limits across highway segments. However, speed limits are constant across
time in our sample and highway-segment (ATR) …xed e¤ects will control for the di¤erences across
highway segments.
18
planatory variables are used to explain micro-level responses (see e.g., Moulton, 1990;
Donald and Lang, 2001; Wooldridge, 2003). Consequently, all standard errors are
clustered at the weekly level and are reported in brackets below the coe¢ cient es-
timates. The clustered standard errors will correct for any unobserved, week-level
correlation in our data, including correlation across highway segments. There are a
total of 237 weeks in our sample, and therefore 237 clusters. We indicate statistical
signi…cance at the 5% level by .
The …rst column of Table 2 contains the estimates of the impact of crude oil prices
on the average retail gasoline price. As expected, crude oil prices are highly correlated
with gasoline prices. The coe¢ cient on crude prices is very precisely estimated with
a corresponding F-statistic larger than 100.
The results of the second-stage estimation are included in columns 2 through 4
and are the principle results of our analysis. As theory predicts, drivers respond
to increases in gasoline prices by slowing down. Column 2 indicates that a $1.00
increase in the price of gasoline would cause average tra¢ c speeds on all highways to
decrease by about a half mile per hour. However, the results depend rather sharply
on whether or not the highway segment is near a large city. The two-stage least
squares estimates, conditional on city designation, are presented in columns 3 and
4.22 The coe¢ cient on gasoline prices drops to 1:249 for highway segments far
from city boundaries, indicating that a one dollar increase in prices causes motorists
to reduce their travel speeds by 1.25 miles per hour on average. The 95% con…dence
interval for this coe¢ cient is approximately [ 1:75; 0:75]. However, increases in
gasoline prices have the opposite e¤ect on the travel speeds of highway motorists near
major cities. The estimates suggest average tra¢ c speeds near major cities increase
22
The …rst-stage regressions for the results in columns 3 and 4 are also performed conditional on
near-city designation. The coe¢ cient estimates are highly signi…cant, but practically identical to
the estimates presented in column 1. Consequently, we choose not to report them here.
19
Table 2: Tra¢ c Speed Regression Results
First-Stage
Estimates Second-Stage Estimates OLS Estimates
(1) (2) (3) (4) (5) (6) (7)
Dep. Var.: P Gast Speedit Speedit
City i = 0 City i = 1 City i = 0 City i = 1
Regressor:
0:473 1:249 0:380 0:065 0:320 0:222
P Gast —
[0:191] [0:254] [0:176] [0:061] [0:088] [0:073]
0:016
P Crudet — — — — — —
[:001]
0:0004 0:091 0:095 0:073 0:092 0:096 0:074
Smallit
[:0001] [0:001] [0:002] [0:001] [0:001] [0:002] [0:001]
20
0:018 0:498 0:387 0:633 0:493 0:377 0:634
Precip it
[0:007] [0:021] [0:028] [0:022] [0:021] [0:027] [0:022]
0:006 0:370 0:256 0:391 0:368 0:257 0:393
T emp32it
[0:012] [0:057] [0:059] [0:074] [0:057] [0:057] [0:074]
0:012 0:751 0:935 0:629 0:761 0:968 0:626
Precip32 it
[0:009] [0:117] [0:164] [0:128] [0:118] [0:164] [0:128]
N 580,191 580,191 323,730 256,461 580,191 323,730 256,461
First-Stage
— 145.77 144.23 146.19 — — —
F-Stat
This table presents the 2SLS and OLS coe¢ cient estimates of equation (1). Each regression (including the …rst-
stage regressions) includes …xed e¤ects for each highway segment and year, as well as seasonal dummies for hour-
of-day, day-of-week, and month-of-year. Standard errors, clustered at the weekly level, are reported in brackets.
Statistical signi…cance at the 5% level is indicated by :
by 0.38 miles per hour for every dollar increase in the price of gasoline. As will be
discussed in more detail below, this increase in speed is likely due to reductions in
the level of tra¢ c congestion.
For comparison, we also include the ordinary least squares estimates of equation
(1) in columns 5 through 7. As is consistent with the classical errors-in-variables
model, the OLS coe¢ cient estimates appear to be attenuated toward zero. The
overall e¤ect of gasoline prices on travel speeds is not statistically (or practically)
di¤erent from zero (column 5). The estimates conditional on whether the road
segment is near a city are also both smaller than the 2SLS estimates in absolute
magnitude (although both remain statistically signi…cant at the 5% level). It appears
that the measurement error in the gasoline price regressor is a …rst-order issue in the
estimation of (1).
It should be noted that although these e¤ects di¤er substantially from the e¤ects
found in Burger and Ka¢ ne (2009), the di¤erences are not necessarily inconsistent.
We …nd a substantial negative e¤ect of gasoline prices on travel speeds on uncongested
(rural) highways, while Burger and Ka¢ ne …nd no e¤ect on uncongested (urban) high-
ways between 2 and 4 a.m. However, this di¤erence could be due to (1) di¤erences
in the characteristics of urban and rural drivers, (2) di¤erences in the typical driver
compared to those who drive at 2 a.m., or (3) some combination of the two. Likewise,
we …nd di¤erences in our estimates for congested highways. Our results indicate a
small, positive e¤ect of gasoline prices on travel speeds on congested (suburban) high-
ways, where Burger and Ka¢ ne …nd a much larger e¤ect (i.e., a coe¢ cient of 3.37) on
congested (urban) highways, but this di¤erence is likely due to the di¤ering levels of
congestion between the two samples. While congestion on Oregon highways near the
ow
outskirts of major cities may be heavy enough to restrict tra¢ c ‡ to speeds slower
ow
than free-‡ speeds, it surely pales in comparison to the gridlock on Los Angeles
21
highways during rush hour.
4.1 Speed Distributions
Average vehicle speed is only one of many measures we could use to assess how mo-
torists collectively respond to changes in gasoline price. In Table 3, we investigate
the e¤ects of changes in gasoline prices on other sample moments of the speed dis-
tribution. Each entry in Table 3 represents a separate (2SLS) regression, where we
change only the dependent variable in equation (1). The numbers reported are the
estimated coe¢ cients on the gasoline price variable ( ^ 1 ) and indicates signi…cance
at the 5% level. The …rst row of both Panel A and B reports the e¤ect of changes
in gasoline prices on the variance of vehicle speeds for rural and near-city highways,
respectively. On rural roads, the variance in vehicle speeds appears to decrease
slightly as gasoline prices increase, but the result is not statistically di¤erent from
zero. However, when the variances of small and large vehicles are regressed sepa-
rately, the results are strikingly di¤erent. The change in the variance of small-vehicle
speeds is e¤ectively zero as gasoline prices change, but the variance of large vehicles
drops by about 3.26 mph2 for each dollar increase in gasoline prices. On suburban
highways, the speed variances of both large and small vehicles increase as gasoline
prices increase. But the e¤ect on small vehicles is slightly smaller than the e¤ect on
large vehicles, with coe¢ cients of 4.75 and 6.28, respectively.
The e¤ect of changes on gasoline prices on the distribution of travel speeds can
be more easily seen by looking at percentiles of the speed distribution. On rural
highways, the distribution of travel speeds appears to decrease at every percentile for
both large and small vehicles as gasoline prices rise. For small vehicles, the shift
is quite uniform with e¤ects ranging between -1.54 and -1.11. The vehicles at the
22
upper tail of the distribution react similarly to those at the center and lower tail of
the distribution. On the other hand, large vehicles at the upper tail of the speed
distribution reduce their speeds signi…cantly more than those at the lower tail (which
leads to a reduction in the variance). The e¤ect of a one dollar increase in gasoline tax
on large vehicles ranges from -2.22 to -0.52 mph. Overall, these results are consistent
ow
with driver responses under free-‡ tra¢ c conditions in which individuals are free
to drive their "optimum" speed.
Panel B shows how suburban motorists respond to changes in gasoline prices.
Those at the lower tail of the distribution appear to slightly reduce their travel speeds
as gasoline prices rise for both small and large vehicles (with an overall e¤ect on all
vehicles of -0.72). However, those at the center and upper tails of the distribu-
tion increase speeds substantially as gasoline prices increase, and the e¤ect becomes
progressively larger in the upper tails of the speed distribution (especially for large
vehicles).
The e¤ect of gasoline prices on the distribution of suburban travel speeds is con-
sistent with the speed choices of drivers constrained by tra¢ c congestion. Those at
the lower tail of the distribution are likely to be traveling at their optimal speeds,
even under congestion. These are the drivers who end up setting tra¢ c speeds when
roads are slightly congested. As gasoline prices rise (and the marginal cost of driving
increases), the motorists traveling at their optimum speeds will likely respond by
slowing down (i.e., driving more e¢ ciently). In addition, as gasoline prices rise,
some motorists are likely to forego their trips completely, and this reduction in tra¢ c
volume alleviates congestion. As congestion eases, those previously traveling below
their optimal speeds can now more easily navigate around slower tra¢ c, and speeds
at the upper end of the distribution increase. The relationship between gasoline
prices and tra¢ c volume is discussed in more detail below.
23
Table 3: Speed Variance and Percentile E¤ects
Panel A: Cityi = 0
Dep. Var. All Veh. "Small" "Large"
Speed Variance -0.942 0.146 -3.264
Speed Percentile
1 -0.879 -1.119 -0.523
10 -1.308 -1.539 -1.024
25 -1.213 -1.315 -1.330
50 -1.299 -1.409 -1.254
75 -1.217 -1.290 -1.287
90 -1.307 -1.319 -1.540
99 -1.664 -1.310 -2.221
Panel B: Cityi = 1
Speed Variance 4.994 4.749 6.278
Speed Percentile
1 -0.716 -0.718 -0.658
10 -0.001 -0.070 0.198
25 0.157 0.201 0.043
50 0.437 0.376 0.702
75 0.566 0.586 0.651
90 0.768 0.669 1.224
99 0.835 0.538 1.989
This table presents 2SLS estimates of 1 in equation (1)
where the dependent variable is replaced. Each entry
represents a separate regression. Standard errors are
clustered at the weekly level and represents signi…cance
at the 5% level.
24
4.2 Robustness Checks
As a …nal check on the robustness of the estimates presented in Tables 2 and 3,
we report the results of alternative speci…cations of equation (1) in Table 4. Once
again, reported are the estimated coe¢ cient and (clustered) standard error for the
gasoline price regressor in a 2SLS regression, with denoting signi…cance at the 5%
level. The …rst column reports the results for rural highways far from city limits;
the second column reports the results for the near-city/suburban highways; the last
column indicates how the regression di¤ers from the baseline regression in (1).
The …rst row of Table 4 reports the gasoline price coe¢ cient from a weighted
regression of equation (1) where the weights are total hourly tra¢ c volume. This
may give us a better sense of how the "average" driver is impacted by gasoline price
changes, as time periods with more vehicles are given more weight. The coe¢ cient
on rural highways decreases in absolute magnitude by about 20% to -1.03, while the
coe¢ cient for suburban highways increases in magnitude by more than 100% to 0.88,
indicating that there are larger gasoline price e¤ects during periods of higher tra¢ c
volume.
Rows 2 and 3 allow us to more closely compare our estimates to those found
in Burger and Ka¢ ne (2009) by strictly examining what happens on the highways
during rush-hour and early-morning hours. The response of rural drivers di¤ers
very little between the early-morning and rush hours. Early-morning rural drivers
are slightly more sensitive to gasoline prices than those driving at rush hour with
coe¢ cients of -1.41 and -1.28, respectively. On the other hand, the impact of gasoline
prices on near-city/suburban highways varies substantially from early morning to
rush hours. Suburban highway speeds increase during rush hour by about a half
mile per hour for each dollar increase in gasoline prices, but speeds are virtually
25
Table 4: Robustness Checks of E¤ect of Gasoline Prices on Speed
Coe¢ cient Estimates Notes on Di¤erences from Equation (1)
Cityi = 0 Cityi = 1
1:028 0:877
(1) Weighted by Total Tra¢ c Volume
[0:194] [0:170]
1:276 0:512
(2) Rush Hour (4 - 6 p.m.)
[0:282] [0:257]
1:408 0:040
26
(3) Early Morning (2 - 4 a.m.)
[0:274] [0:234]
0:060 0:017
(4) log-log Regression
[0:013] [0:008]
P Gast P Gas2t P Gast P Gas2 t
(5) 4:475 0:982 0:083 0:079 Quadratic in Gasoline Price
[1:183] [0:221] [0:888] [0:161]
1:09 0:355
(6) Instrument: P Crudet 1
[0:228] [0:164]
Table 4 (Continued): Robustness Checks of E¤ect of Gasoline Prices on Speed
Coe¢ cient Estimates Notes on Di¤erences from Equation (1)
Cityi = 0 Cityi = 1
1:047 0:346
(7) Instruments: P Crudet & P Crudet 1
[0:222] [0:161]
P Gast P Gast 1 P Gast P Gast 1
(8) 3:723 3:003 0:805 0:517 Includes P Gast & P Gast 1
[1:146] [1:212] [0:723] [0:762]
27
1:193 0:326
(9) ATR-year …xed e¤ects
[0:253] [0:177]
1:025 0:156
(10) Cubic Trend Replaces Year Fixed E¤ects
[0:195] [0:158]
1:015 0:096 Cubic Trend for Each ATR
(11)
[0:199] [0:156] Replaces Year Fixed E¤ects
This table presents the 2SLS coe¢ cient estimates of the e¤ect of gasoline prices on average speed.
Each entry represents a separate regression where equation (1) has been modi…ed according to the
description in the last column. The quadratic model in row (5) is estimated using PCrudet and
PCrude2 as instruments. Standard errors, clustered at the weekly level, are reported in brackets.
t
Statistical signi…cance at the 5% level is indicated by a *.
una¤ected by changes in gasoline prices in the early-morning hours. Interestingly,
the impact of gasoline prices on early-morning vehicle speeds varies substantially
between rural and near-city highways. This could be due to di¤erences in the value
of time between rural and urban drivers or di¤erences in other factors that constrain
driver speed such as the likelihood of speed limit enforcement by the highway patrol.
Regardless, our estimated e¤ect on uncongested, suburban highways is e¤ectively
zero and corresponds quite closely to the results found by Burger and Ka¢ ne for
early-morning, Los Angeles highways.
We report estimates of the log-log model in row 4 and estimates from a quadratic
gasoline price relationship in row 5. The coe¢ cients from the log-log model indicate
that a ten percent (10%) increase in the price of gasoline will reduce travel speeds
on rural highways by 0.60% and increase speeds on suburban highways by 0.17%.
The elasticity for rural highways is about 27% larger in absolute magnitude than
that obtained by computing the elasticity directly from the coe¢ cient in the linear
R
model (i.e., ^ 1 Price
Speed
= 0:047). However, the elasticity for suburban highways
is quite similar to that obtained by computing the elasticity from the coe¢ cient
S
in the linear model (i.e., ^ 1 Price
Speed
= 0:015). Estimating the gasoline impact on
vehicle speeds using a quadratic relationship (row 5) produces a speed marginal e¤ect
that is substantially smaller than the marginal e¤ect in the linear model. When
evaluated at the average gasoline price, the marginal e¤ect on rural roads is 0:298
(= 4:475 2 0:982 2:43). But once again, the marginal e¤ect on near-city roads
is very similar to the linear model (0:301 = 0:083 + 2 0:079 2:43).23
Rows 6 and 7 of Table 4 investigate the use of lagged crude prices as both an
23
The quadratic relationship on rural roads reaches a maximum at a gasoline price $2.28 per
gallon, indicating that the marginal e¤ect will be negative for most of the gasoline prices at the
higher end of the gasoline price distribution. On the other hand, the quadratic e¤ect on near-city
highways reaches a minimum at about $0.53 per gallon, leading to a positive marginal e¤ect for all
gasoline prices in the relevant range of data.
28
alternative or additional instrument for gasoline prices. For the reasons discussed
above, it may be the case that individuals adjust their driving behavior in response
to changes in crude oil prices, especially when those prices are high. Therefore,
lagged crude oil prices may be a more plausible instrument if current crude prices
are not excludable from (1). The estimated coe¢ cients using lagged crude oil prices
as an instrument for current gasoline prices are qualitatively similar to the baseline
speci…cation in Table 2 (though somewhat smaller in absolute magnitude). The
impact of a one dollar increase in gasoline prices on rural and near-city highways is
1:09 and 0:36 mph, respectively. Row 7 uses both current and lagged crude prices
as instruments for gasoline prices, and we …nd the results to be economically and
statistically similar to the estimates in row 6.24
We explore possible dynamic e¤ects in row 8 by including the current and lagged
price of gasoline (for which we instrument using current and lagged crude oil prices).
The coe¢ cient on P Gast indicates that a 10 cent week-over-week increase in gasoline
prices (holding P Gast 1 constant) leads to 0:37 mph reduction in speed on rural
roads. The e¤ect on near-city roads is 0:08 mph per 10 cent increase in prices, but
is not statistically signi…cant. The longer-term response to a change in the level of
gasoline prices can be found by summing the two coe¢ cients on P Gast and P Gast 1 .
Therefore, a one dollar increase in gasoline prices leads to a 0:72 mph reduction in
vehicle speeds on rural highways, and a 0:29 mph increase in vehicle speeds on near-
city highways. These e¤ects are smaller in absolute magnitude than those estimated
in Table 2. Also, it appears that individuals are substantially more sensitive to the
week-to-week change in gasoline prices than the overall level of gasoline prices.
The last three rows of Table 4 alter the manner in which we control for the
24
An added bene…t from using both current and lagged crude prices as instruments is that we can
perform an instrument over-identi…cation test. Using the Hansen J-Statistic, we fail to reject the
hypothesis that the instruments are valid for both the rural and near-city regressions.
29
trends in average roadway speeds. The baseline model includes year and highway
…xed e¤ects, which assumes the trend exists at the yearly level and is the same
for all highways in the sample. However, it may be the case that the trends in
average speeds are di¤erent for di¤erent stretches of highways and/or vary over the
course of the year. Row 9 allows for di¤erent highway trends at the yearly level
through the use of highway-year …xed e¤ects. Although this speci…cation only slightly
decreases the absolute magnitudes of the gasoline price coe¢ cients, the coe¢ cient in
the suburban regression is no longer signi…cant at the 5% level. Row 10 replaces
the year …xed e¤ects in the baseline model with a cubic time trend at the daily level.
This speci…cation causes the coe¢ cients in both the rural and suburban regressions
to fall in absolute magnitude to 1:03 and 0:16, respectively. The last speci…cation
in row 11 replaces the year and highway …xed e¤ects with a daily, cubic trend that is
allowed to di¤er for each highway segment. The coe¢ cient on gasoline prices remains
signi…cantly di¤erent from zero with a value of 1:02 for rural highways. However,
the coe¢ cient for near-city highways drops to 0:10 and is not statistically di¤erent
from zero.
Overall, the marginal e¤ect of changes in gasoline prices on uncongested, rural
highway speeds is fairly robust to di¤erent model speci…cations, with most of the
estimates lying in the range 1:41 to 0:72 (all of which are statistically di¤erent
from zero). The one exception to this is the quadratic speci…cation in gasoline
prices which leads to an average marginal e¤ect of approximately 0:30 mph per
dollar increase in gas prices. On the other hand, the e¤ect of gasoline prices on
near-city/suburban travel speeds varies quite substantially from one speci…cation to
another, with marginal e¤ects ranging from about 0:04 to 0:88 (some of which are not
signi…cantly di¤erent from zero). The suburban e¤ect appears especially sensitive to
how we account for the time trend.
30
4.3 Gasoline Prices and Tra¢ c Volume
From Table 1, we see that the most obvious di¤erence between highways near major
cities and those that are not, is the di¤erence in tra¢ c volumes. Consequently, one
could speculate that tra¢ c congestion is the likely reason for the opposite signs in
the e¤ects of gasoline prices on travel speeds across the two types of roads. To better
understand the role congestion plays in the di¤erences between suburban and rural
highways, we estimate the following model:
0
V olit = 0 + 1 P Gast + zit + h + d + m + y + i + "it : (2)
Once again, we use the U.S. crude oil price as an instrument for the retail gasoline
price to control for the e¤ects of measurement error. The vector of controls, zit ,
includes the weather indicators discussed above. Similar to equation (1), we include
…xed e¤ects for ATRs ( i ) and years ( y ), and seasonal dummies for hour-of-day
( h ), day-of-week ( d ), and month-of-year ( m ).
t
Unfortunately, total tra¢ c volume isn’ a perfect indicator of congestion, as the
number of cars that pass over the ATR in a given hour is itself dependent upon
congestion. For example, there may be a period of low congestion (with substan-
tial spacing between vehicles) in which 800 vehicles pass over the tra¢ c recorder at
ow
free-‡ speeds. However, there may also be a period of high congestion (with thou-
sands of vehicles tightly spaced on the highway) in which only 800 vehicles pass over
the ATR at low speeds. This type of hyper-congestion is unlikely on the two-lane
interstate highways in our sample, however, the example illustrates that tra¢ c vol-
ume alone is not su¢ cient to distinguish between periods of high and low congestion.
ow
Nevertheless, at speeds su¢ ciently close to the free-‡ speed, volume may give some
information about the level of congestion on the road.
31
Table 5: Tra¢ c Volume Regression Results –2SLS
First-Stage Second Stage Estimates
Estimates
(1) (2) (3) (4) (5) (6) (7)
Dep. Var.: P Gast V olit V olCarit V olT ruckit
City i = 0 City i = 1 City i = 0 City i = 1 City i = 0 City i = 1
Regressor:
44:63 49:15 43:52 47:69 1:11 1:46
P Gast —
[29:29] [47:33] [33:19] [52:42] [11:13] [17:61]
0:016
P Crudet — — — — — —
[0:001]
0:018 6:14 37:21 2:16 31:82 3:98 5:39
32
Precip it
[0:007] [3:77] [7:82] [3:58] [7:22] [1:08] [1:83]
0:007 39:328 52:97 33:30 31:66 6:03 21:30
T emp32it
[0:012] [6:852] [20:46] [6:63] [18:95] [2:02] [3:26]
0:012 80:86 55:88 66:49 35:07 14:37 20:81
Precip32 it
[0:009] [12:30] [34:58] [10:81] [33:93] [4:10] [4:51]
N 580,191 323,730 256,461 323,730 256,461 323,730 256,461
First-Stage
— 144.41 145.80 144.41 145.80 144.41 145.80
F-Stat
This table presents the 2SLS estimates of equation (2). Each regression (including the …rst-stage regressions) includes
…xed e¤ects for each highway segment and year, as well as seasonal dummies for hour-of-day, day-of-week, and month-
of-year. Standard errors, clustered at the weekly level, are reported in brackets. Statistical signi…cance at the 5% level
is indicated by .
Columns 2 and 3 of Table 5 display the estimated coe¢ cients of the e¤ect of
gasoline prices on tra¢ c volumes. The coe¢ cient on gasoline prices is similar for
both rural and suburban highway segments, with estimates of -44.6 and -49.2, re-
spectively.25 Unfortunately, both estimates are measured imprecisely (and are not
statistically di¤erent from zero) so it is di¢ cult to come to a de…nitive conclusion
on the role that congestion plays in motorist responses to gasoline prices. But it is
interesting to note that if the slope estimates are converted to elasticities, the e¤ect
of gasoline prices on tra¢ c volume is similar to the -0.10 short-run VMT elasticity
found in the meta-analysis by Goodwin et al., with our estimates being -0.15 for
rural highways and -0.069 for suburban highways.26 Although the point estimates
in Table 5 suggest that the average change in tra¢ c volumes in response to changes
in gasoline prices is similar for motorists near and far from city limits – both de-
creasing about 45-50 vehicles per dollar increase in prices – the resultant e¤ect on
road congestion may still be quite di¤erent. A 45 vehicle reduction in tra¢ c volume
when average volume on the highway is 1730 vehicles per hour may increase vehicle
mobility substantially more than a 45 vehicle reduction when the average volume is
only 715 vehicles per hour.
Burger and Ka¢ ne (2009) use an indirect approach to estimate the elasticity
of VMT demand – …nding elasticities in the range of -0.12 to -0.29, depending on
the speci…cation. t
Although they don’ directly look at the impact of gas prices
on tra¢ c volume, they do …nd that L.A. public transportation ridership increases
during periods of higher gasoline prices. They take this as evidence that individuals
are substituting away from motor-vehicle tra¢ c when gas prices are high, and as the
25
The analysis is nearly identical if volume per lane is used as the dependent variable rather than
total tra¢ c volume.
26
Direct comparison of the elasticities requires the further assumption that the average trip length
is constant across highways and over time, and that it does not change with changes in gasoline
prices.
33
mechanism by which travel speeds on highways increase. For our study of rural tra¢ c,
alternative modes of travel (such as public transportation on light-rail train or bus)
may be poor substitutes for motor vehicle travel on the fringes of major cities. More
importantly, the alternatives may not exist at all, as public transportation routes
rarely extend to city boundaries, let alone beyond them. Therefore, it seems unlikely
that public transportation ridership would give us much information about tra¢ c
on rural highways outside city boundaries. Therefore, while we suspect reduced
congestion is the mechanism driving the positive coe¢ cient in the suburban, average
speed regressions (especially when the percentile results of Table 3 are considered),
t
we cannot be certain as we don’ know the exact point where congestion begins to
impact travel speeds.
In columns 4 through 7, we see the e¤ect of gasoline prices on the tra¢ c volumes of
small and large vehicles separately. We denote the tra¢ c volumes of small and large
vehicles by the dependent variables V olCarit and V olT ruckit , respectively. Once
again, there appears to be little di¤erence in the response of total tra¢ c volume
between rural and suburban highways, and none of the estimates are signi…cantly
di¤erent from zero. However, there appear to be large di¤erences in the point
estimates of small and large vehicles. Nearly all of the reduction in total tra¢ c
volume in response to higher gasoline prices is due to light-duty, passenger vehicle
travel. The change in heavy-duty, commercial vehicle tra¢ c volume is e¤ectively
zero, with coe¢ cients of less than -1.5 vehicles per dollar increase in prices. While
commercial vehicles will adjust their travel speeds in response to changes in gasoline
prices, it appears that they are not responding by adjusting trip frequency.27
27
This could be due to the structure of the commercial trucking industry. For example, if
trucking companies are locked into long term contracts with clients, they may be required to deliver
goods regardless of the price of gasoline, and consequently forced to absorb the increase in cost.
Alternatively, if there are no feasible shipping substitutes, trucking companies may be able to pass
the increased gasoline costs directly to the client.
34
5 Gas Prices, Fuel Consumption, and Emissions
The previous two sections illustrate how drivers on di¤erent stretches of highway
respond di¤erently to gasoline price changes. As economic theory suggests, motorists
on rural highways (far from city limits) choose to reduce their travel speeds as gasoline
prices rise, on average. However, as motorists approach city limits, they encounter
ow
increased congestion which may limit their free-‡ travel speeds. In these situations,
higher gasoline prices will lead to increases in travel speeds as congestion is eased,
and our results (as well as the results of Burger and Ka¢ ne) indicate that this e¤ect
becomes progressively larger as highway congestion gets heavier.
The manner in which motorists respond to increases in gasoline prices has im-
portant implications for the e¤ectiveness of tax policies that aim to reduce petrol
dependence and emissions in the U.S. Most studies of gasoline demand and vehicle
emissions focus on the total number of miles traveled, and the e¢ ciency with which
the miles are traveled is largely ignored (see e.g., Goodwin et al., 2004; Bento et al.,
2008). Although travel e¢ ciency may indeed be of second-order importance, our
results suggest that it should not be ignored altogether, as it may comprise a non-
negligible fraction of the overall change in gasoline demand and vehicle emissions.
To illustrate the possible impact of driving e¢ ciency, we calculate a simple ex-
ample using fuel consumption and emissions data from West et al. (1999). Let C
represent gasoline consumption measured in gallons for some …xed period of time.
Assume gasoline consumption is proportional to the number of miles driven so that
C = g (s) V M T , where the factor of proportionality, g (s), is the number of gallons
consumed per mile driven and is a function of speed. The derivative of gasoline
35
consumption with respect to the gasoline price can be written as
dC @g (s) ds dV M T
= V M T + g (s) : (3)
dp @s dp dp
If we further assume that the average length of a trip does not vary with the price of
gasoline, then V M T = m V ol, and (3) becomes
dC @g (s) ds dV ol
=m V ol + m g (s) ; (4)
dp @s dp dp
where V ol is total tra¢ c volume and m is the average number of miles per trip.28
The …rst term on the right side of (4) represents the marginal change in gasoline con-
sumption that comes from drivers adjusting their fuel e¢ ciency when gasoline prices
change. The second term represents the marginal change in gasoline consumption
from changes in the number of trips taken (and therefore miles driven) when prices
change.
Using Equation (4), we combine our estimates from Section 4 to get a crude
measure of the impact of increasing the gasoline tax on both rural and suburban
highways. Figure 2 shows a U-shaped relationship between fuel consumption and
vehicle speeds, with the minimum being reached at about 50-55 mph. Table 1
s
(Column 3) shows that the average speed and total tra¢ c volume on Oregon’ rural
highways is about 63.89 mph and 715 vehicles per hour. At an average speed of 63.89
mph, the estimates of West et al. imply that g (63:89) = 0:037 and @g (63:89) =@s =
0:000445. Tables 2 and 5 give point estimates for ds=dp and dV ol=dp of 1:25 and
45, respectively. Plugging these numbers into (4) gives the total change in fuel
28
It may be the case that the average distance of trips is a function of the gasoline price. For
example, as gasoline prices rise, motorists may choose to defer longer trips …rst. If this is the case,
the number of gallons saved from reduced tra¢ c volume may be larger.
36
consumption for the average highway segment in the average hour:
dC
= ( 0:398) m + ( 1:665) m
dp
= 2:063 m:
The amount of fuel saved due to increased fuel e¢ ciency is about 0.398 m gallons
and the fuel saved due to reduced tra¢ c is about 1.665 m gallons. This indicates
that almost one-…fth (19%) of the change in gasoline demand comes from changes in
driving behavior along the intensive (rather than extensive) margin.
Similarly, these numbers can be calculated for the average suburban highway in
our sample, giving
dC
= (0:266) m + ( 1:79) m
dp
= 1:524 m:
The fuel saved from reduced tra¢ c volume is approximately 1.79 m gallons, but
the increased speeds cause fuel consumption to increase by about 0.266 m gallons,
or 15%. These numbers imply that in suburban areas, increases in gasoline prices
have less of an impact on fuel consumption than the total reduction in tra¢ c volume
would suggest.
However, increases in speed need not always lead to reductions in fuel e¢ ciency, as
light-duty vehicles in heavily congested areas (such as Los Angeles during rush hour)
usually travel at speeds below the fuel-e¢ cient speed. In these cases, the speed
and volume impacts in fuel consumption will work in the same direction once again.
From a gasoline consumption standpoint, it appears as if the intensive changes in
vehicle speeds will typically enhance the e¤ectiveness of gasoline tax policy on most
37
segments of highway, as they work in the same direction as the change in VMT.
Unfortunately, the overall impact of a gasoline tax on vehicle emissions is a little
less clear. Estimates of the relationship between speed and emissions vary across
studies. For example, in their controlled, laboratory experiment, West et al. …nd
that emissions rates are, for the most part, monotonically increasing in speeds (see
Figure 4 below). But in many "on-road" studies of emissions rates (e.g., Kean et
al., 2003) the speed-emissions relationship is similar to the U-shaped relationship of
speed and fuel e¢ ciency (see Figure 5). Emissions are obviously tied to the burning
of fuel, but the emissions rates of many pollutants are also tied to other factors, such
as engine load or temperatures of the combustion process, which increase as speeds
increase (see Krishnamurthy et al., 2007, pg. 673). Consequently, the emissions rates
of these pollutants are most likely U-shaped, but could reach a minimum at speeds
other than the fuel-e¢ cient speed. If the minimum is su¢ ciently low, it is possible
that changes in speed could increase emissions, even though speeds are moving closer
to fuel-e¢ cient levels.
As a simple example of how gasoline price changes may impact changes emissions,
we employ the CO2 emissions rates used by the UK Highways Agency in the "Design
Manual for Roads and Bridges." These data were also used to plot the emissions
rates for CO2 , CO, non-methane hydrocarbons, nitrogen oxides, and PM10 in Figure
5. We assume once again that CO2 emissions are proportional to total miles traveled,
where the factor of proportionality is a function of speed: CO2 = e (s) V M T . As
with fuel consumption, we can decompose the marginal change in emissions with
respect to gasoline prices into an "e¢ ciency" e¤ect and a "volume" e¤ect:
dCO2 @e (s) ds dV ol
= V ol m + e (s) m :
dp @s dp dp
38
Figure 4: Light-Duty Vehicle Emissions Relative to Emissions at Idle
40
Carbon Monoxide
Source: West et al. (1999)
Hydrocarbons
Nitrogen Oxides (NO, NO )
2
35
Emissions Rate (mg/sec) Relative to Rate at Idle
30
25
20
15
10
5
0
0 10 20 30 40 50 60 70 80
Speed (mph)
39
Figure 5: Emissions Rates of the "Typical" On-road Vehicle
450 2.5
400
(g/m)
CO (g/m)
350 2
2
CO
300
250 1.5
20 40 60 80 20 40 60 80
mph mph
0.5 2.6
NMHC (g/m)
2.4
(g/m)
0.4
2.2
x
NO
0.3
2
0.2 1.8
20 40 60 80 20 40 60 80
mph mph
0.12
(g/m)
0.1
0.08 Source: "Des ign Manual for Roads
10
and Bridges" -- UK Gov. (2005)
PM
0.06
0.04
20 40 60 80
mph
40
For the rural highway segments in our sample, the impact of a change in gasoline
prices is
dCO2
= 2; 049:6 m 9; 515:1 m
dp
= 11; 564:7 m;
Similarly, the impact of gasoline prices in suburban areas is
dCO2
= 1; 505:5 m 10; 478:8 m
dp
= 8; 973:3 m:
Decreases in speed amplify the marginal e¤ect of gasoline prices on CO2 emissions in
rural areas, comprising about 18% of the total e¤ect. However, increases in speed in
suburban areas will mitigate the marginal impact of gasoline prices by approximately
14%. While these are back-of-the-envelope calculations, they illustrate that small
changes in the behavior of a large number of drivers can serve to o¤set large changes
in the behavior of a small number of drivers.
Further evidence of the short term fuel-e¢ ciency and emissions impact of gasoline
prices in the U.S. is illustrated in Table 6. The columns under heading (1) report
the coe¢ cient and standard error estimates from a linear regression model where
the dependent variable is annual transportation-related CO2 emissions (measured in
teragrams). The regressors are annual US VMT (in billions of miles), the average
annual US retail gasoline price (in 2009 dollars), and a yearly time trend.29 The
29
s
Emissions data are taken from the EPA’ "Inventory of U.S. Greenhouse Gas Emissions
and Sinks." Gasoline price data come from the EIA (and are in‡ ation adjusted). The
data on Fuel Consumption and VMT were obtained from "National Transportation Statis-
tics" published by the Bureau of Transportation Statistics (see Tables 1-32 and 1-33 at
http://www.bts.gov/publications/national_transportation_statistics/).
41
results indicate that CO2 emissions are highly correlated with total miles traveled,
with emissions increasing by 1.67 Tg for each additional billion miles traveled. The
coe¢ cient on the yearly time trend indicates that the emissions rate of the vehicle
eet
‡ is falling over time as vehicles become "cleaner". However, the coe¢ cient on
price indicates that (after controlling for VMT) gasoline prices and vehicle emissions
are positively correlated. While the coe¢ cient is imprecisely measured, there is at
least some evidence that the increased vehicle speeds that result from reduced tra¢ c
congestion are increasing the average emissions rate of roadway vehicles.
The next two columns under heading (2) report the results from a similar regres-
sion where the dependent variable is U.S. annual gasoline consumption (measured
in thousands of barrels per day). Average gasoline consumption has been trending
upward signi…cantly since 1993. However, once we control for VMT (and a yearly
trend), the price of gasoline is negatively correlated with gasoline consumption. The
relationship between gasoline consumption, VMT and gasoline prices is likely en-
dogenous, and the coe¢ cients are not likely to represent the structural parameters.
Nevertheless, the results indicate that gasoline prices continue to have explanatory
power even after controlling for VMT, and the negative correlation indicates that not
only do drivers reduce their mileage as gasoline prices rise, they also improve their
driving e¢ ciency.
Over the long run, increases in gas prices will induce individuals to purchase more
fuel-e¢ cient vehicles. Therefore, it is likely that higher gas prices (from increased
taxes or otherwise) will lead to long-run pollution reductions as drivers substitute out
of older cars in favor of newer, more e¢ cient vehicles. However, the scrappage rate
of light-duty vehicles in the U.S. is rather low. For example, the scrappage rate used
by the EPA in the MOBILE6 emissions modeling program is 6% per year, indicating
s eet
that it would take longer than 10 years to reduce today’ vehicle ‡ by 50% (see e.g.,
42
Table 6: E¤ect of Gasoline Prices on Annual CO2 Emissions and
Fuel Consumption (1993 - 2006)
(1) (2)
Dep. Var.: Dep. Var.:
Transportation-Related US Avg. Gasoline Consumption
CO2 Emissions (Tg) (Thousands of Barrels per Day)
Regressors: Mean S.E. Coef. S.E. Coef. S.E.
VMT (billions) 2692 240 1.67 [0.55] 0.12 [0.66]
Price 1.49 0.48 64.51 [53.63] -159.53 [55.84]
43
Year 1999.5 4.18 -71.36 [36.57] 150.77 [42.45]
N 14 14 14
Mean and S.E. 1717.6 132.4 8417.1 596.6
of Dep. Var
This table presents the coe¢ cient and robust standard error estimates from the
regression model y = 0 + 1 VMT + 2 Price + 3 Year +" where the dependent
variable is as listed in columns (1) and (2) above. Statistical signi…cance at the
5% level is indicated by .
Jackson, 2001; Chen and Niemeier, 2005). Consequently, the pollution bene…ts from
vehicle turnover take many years to realize. In the short run, pollution reductions
from increased gasoline taxes will largely be a factor of how motorists reduce their
travel miles and adjust their travel e¢ ciency. Considering over 60% of total miles
traveled in the U.S. occurs on urban roads, the e¤ect of increased gasoline taxes could
have a smaller e¤ect on emissions in the short run than the number of deferred tra¢ c
trips would suggest.
6 Conclusion
In April 2009, the Environmental Protection Agency issued a proposed …nding that
certain greenhouse gasses (including CO2 ) "endanger the public health and welfare
of current and future generations" (U.S. EPA, 2009). If the proposal is approved,
s
greenhouse gas emissions would fall under the EPA’ regulatory authority granted by
the Clean Air Act. With carbon dioxide being the most prevalent (and the principle
anthropogenically sourced) greenhouse gas, regulators will surely look for ways to
limit transportation-related CO2 emissions, and an increase in the federal gasoline
tax is likely.
However, the results of this paper suggest that the resulting change in emissions
will depend, in part, on how drivers adjust their driving e¢ ciency in response to fuel
price increases, at least in the short run. We …nd that drivers on rural highways
drive more e¢ ciently, on average, as gasoline prices increase. These reductions in
travel speed will improve fuel e¢ ciency and reduce vehicle emissions. On the other
hand, drivers on congested, suburban freeways tend to increase their travel speeds
as gasoline prices increase. These increases in travel speed could either increase or
decrease fuel e¢ ciency depending on how they compare to the fuel-e¢ cient speed.
44
The impact of increased speeds on emissions is also slightly ambiguous. Although
the net e¤ect of changing the gasoline price on fuel consumption and emissions will
be dominated by the changes in total VMT, the "e¢ ciency" e¤ect of driver speed
adjustments is not negligible.
Our estimates di¤er from the estimates of Burger and Ka¢ ne (2009), who …nd a
larger e¤ect of gasoline prices on the travel speeds of vehicles on "congested" high-
ways, and a smaller e¤ect (in absolute magnitude) on the travel speeds of vehicles
on "uncongested" highways. However, our results should be viewed as complemen-
tary (rather than contradictory) to their …ndings, as we look at speeds in rural areas
and Burger and Ka¢ ne analyze speeds in urban Los Angeles. Taken as a whole,
it appears that drivers generally respond to changes in gasoline prices as predicted
by economic theory, as long as travel speeds are not constrained by congestion. In
addition, progressively heavier congestion leads to progressively larger increases in
speeds as the congestion is alleviated by increases in gasoline prices.
45
7 References
1. Bento, A., L. Goulder, M. Jacobsen and R. von Haefen, 2009, “Distributional
and E¢ ciency Impacts of Increased U.S. Gasoline Taxes,”American Economic
Review, Forthcoming.
2. Blackburn, M. and D. Neuman, 1992, “Unobserved Ability, E¢ ciency Wages,
and Interindustry Wage Di¤erentials,” The Quarterly Journal of Economics
107, 1421-1436.
3. Burger, N. and D. Ka¢ ne, 2009, “Gas Prices, Tra¢ c and Freeway Speeds in
Los Angeles,”The Review of Economics and Statistics 91, 652-657.
4. Chen, C. and D. Niemeier, 2005, “A Mass Point Vehicle Scrappage Model,”
Transportation Research Part B 39, 401-415.
5. Dahl, C., 1979, “Consumer Adjustment to Gasoline Tax,” The Review of Eco-
nomics and Statistics 61, 427-432.
6. Donald, S. and K. Lang, 2007, “Inference with Di¤erence-in-Di¤erences and
Other Panel Data,”The Review of Economics and Statistics 89, 221-233.
7. Fullerton, D. and L. Gan, 2005, “Cost-E¤ective Policies to Reduce Vehicle
Emissions,”The American Economic Review 95, 300-304.
8. Goodwin, P., J. Dargay, and M. Hanly, 2007, “Elasticities of Road Tra¢ c and
Fuel Consumption with Respect to Price and Income: A Review,” Transport
Reviews 24, 275-292.
9. Griliches, Z. and W. Mason, 1972, “Education, Income, and Ability”The Jour-
nal of Political Economy 80, S74-S130.
10. Hansen, J., M. Winther and S. Sorenson, 1995, “The In‡uence of Driving Pat-
terns on Petrol Passenger Car Emissions,” The Science of Total Environment
169, 129-139.
11. Jackson, T., 2001, “Fleet Characterization Data for MOBILE6,”Technical Re-
port EPA420-R-01-047: M6.FLT.007, U.S. Environmental Protection Agency.
46
12. Jensen, S., 1995, “Driving Patterns and Emissions from Di¤erent Types of
Roads,”The Science of Total Environment 169, 123-128.
13. Joumard, R., P. Jost, J. Hickman and D. Hassel, 1995, “Hot Passenger Car
Emissions Modelling as a Function of Instantaneous Speed and Acceleration,”
The Science of Total Environment 169, 167-174.
14. Kean, A., R. Harley and G. Kendall, 2003, “E¤ects of Vehicle Speed and Engine
Load on Motor Vehicle Emissions,” Environmental Science & Technology 37,
3739-3746.
15. Krishnamurthy, M., D. Carder, G. Thompson and M. Gautam, 2007, “Cost
of Lower NOx Emissions: Increased CO2 Emissions from Heavy-Duty Diesel
Engines,”Atmospheric Environment 41, 666-675.
16. Moulton, B., 1990, “An Illustration of a Pitfall in Estimating the E¤ects of
Aggregate Variable on Micro Units,” The Review of Economics and Statistics
72, 334-338.
17. Parry, I. and K. Small, 2005, “Does Britain or the United States Have the Right
Gasoline Tax,”American Economic Review 95, 1276-1289.
18. Puentes, R. and A. Tomer, 2008, “The Road...Less Traveled: An Analysis of
Vehicle Miles Traveled Trends in the U.S.,” Brookings Report, Metropolitan
Policy Program at Brookings, Washington, D.C.
19. U.S. Environmental Protection Agency, 2008, “Inventory of U.S. Greenhouse
Gas Emissions and Sinks: 1990-2006.”
20. U.S. Environmental Protection Agency, 2009, “Proposed Endangerment and
Cause or Contribute Findings for Greenhouse Gases Under Section 202(a) of
the Clean Air Act,”Federal Register, Docket ID: EPA-HQ-OAR-2009-0171, Vol.
74, No. 78, 18886-18910.
21. Virley, S., 1993, “The E¤ect of Fuel Price Increases on Road Transport CO2
Emissions,”Transport Policy 1, 43-48.
22. West, B., R. McGill, J. Hodgson, C. Slunder and D. Smith, 1999, “Develop-
ment and Veri…cation of Light-duty Modal Emissions and Fuel Consumption
Values for Tra¢ c Models” Technical Report, Oak Ridge National Laboratory,
Publication No. FHWA-RD-99-068.
47
23. Wooldridge, J., 2003, “Cluster-Sample Methods in Applied Econometrics,”The
American Economic Review 93, 133-138.
48
Related docs
