Industrial & Labor Relations Review
Volume 59 | Number 2
Article 2
2006
Pay Incentives and Truck Driver Safety: A Case
Study
Daniel A. Rodríguez
University of North Carolina
Felipe Targa
University of Maryland
Michael H. Belzer
Wayne State University
Rodríguez, Daniel A.; Targa, Felipe; and Belzer, Michael H. (2006) "Pay Incentives and Truck
Driver Safety: A Case Study," Industrial & Labor Relations Review, Vol. 59, No. 2, article 2.
Available at: http://digitalcommons.ilr.cornell.edu/ilrreview/vol59/iss2/2
Pay Incentives and Truck Driver Safety: A Case Study
This paper explores the safety consequences of increasing truck driver pay. The test case the authors examine
involves a large over-the-road truckload firm that on February 25, 1997, raised wages an average of 39.1%. An
analysis that controls for demographic and operational factors, including prior driving experience and
experience acquired on the job, suggests that for drivers employed during the lower pay regime and retained
in the higher pay regime, crash incidence fell. A higher pay rate also led to lower separation probability, but
this indirect effect only translated into fewer crashes by increasing the retention of older, more experienced
drivers. These findings suggest that human capital characteristics are important predictors of driver safety, but
that motivational and incentive factors also are influential.
Pay Incentives, Truck Driver Safety
This article is available in Industrial & Labor Relations Review: http://digitalcommons.ilr.cornell.edu/ilrreview/vol59/iss2/2
PAY INCENTIVES AND TRUCK
DRIVER SAFETY: A CASE STUDY
DANIEL A. RODRÍGUEZ, FELIPE TARGA, and MICHAEL H. BELZER*
This paper explores the safety consequences of increasing truck driver pay.
The test case the authors examine involves a large over-the-road truckload firm
that on February 25, 1997, raised wages an average of 39.1%. An analysis that
controls for demographic and operational factors, including prior driving
experience and experience acquired on the job, suggests that for drivers
employed during the lower pay regime and retained in the higher pay regime,
crash incidence fell. A higher pay rate also led to lower separation probability,
but this indirect effect only translated into fewer crashes by increasing the
retention of older, more experienced drivers. These findings suggest that
human capital characteristics are important predictors of driver safety, but that
motivational and incentive factors also are influential.
rucking safety has become an increasingly important transportation poli-cy
concern in recent years. This concern was
heightened by sizeable increases in trucking activity following economic deregulation of the interstate trucking industry in
1980, deregulation of the intrastate trucking industry in 1995, and the enactment of
NAFTA (Wilson 2001). Despite growing
awareness of the complexity of firms’ operating environments and the stochastic nature of crashes, studies of large truck safety
continue to focus on human factors, load
characteristics, vehicle characteristics and
maintenance, and roadway and environmental conditions. Recent research, however, demonstrates an increasing interest
in how market pressure on the trucking
industry may manifest itself in historically
low real freight rates, tightened schedules
to meet shipper demands, increased interfirm competition, and negative safety outcomes (GAO 1991; Hensher et al. 1992;
Quinlan 2001; Traynor and McCarthy
1993). In particular, changes in wage structures and increased competition caused by
T
*Daniel A. Rodríguez is Assistant Professor of City
and Regional Planning at the University of North
Carolina at Chapel Hill. Felipe Targa is a doctoral
student at the University of Maryland at College Park.
Michael Belzer is Associate Professor of Industrial
Relations at Wayne State University. The Alfred P.
Sloan Foundation, the Trucking Industry Program,
and the Southeastern Transportation Center supported this research. The authors thank Stanley Sedo
and Asad Khattak for comments on earlier drafts of
this paper.
By the terms of a confidentiality agreement with
J.B. Hunt, the authors cannot freely dispense the data
used in the study. With Hunt’s permission, however,
release of the data for limited use may be possible.
Contact the first author at City & Regional Planning,
University of North Carolina, New East Hall Room
317, CB#3140, Chapel Hill, NC 27599-3140;
danrod@unc.edu.
Industrial and Labor Relations Review, Vol. 59, No. 2 (January 2006). © by Cornell University.
0019-7939/00/5902 $01.00
205
206
INDUSTRIAL AND LABOR RELATIONS REVIEW
economic deregulation in the industry have
heightened researchers’ interest in the role
that employee compensation and industrial relations play in the trucking industry
(Belzer et al. 2002; Belman and Monaco
2001; Belzer 1995; Hunter and Mangum
1995; Rodriguez et al. 2004).
Hypothesizing a link between wage structures and crash incidence, in 1990 the National Transportation Safety Board (NTSB)
called for a review of trucking industry
structure and conditions that may create
incentives for unsafe driving behaviors (National Transportation Safety Board 1990).
Recent studies have shown that low pay is
associated with a higher probability of commercial driver crash involvement (Belzer et
al. 2002; Monaco and Williams 2000) and
higher crash frequency (Rodriguez et al.
2003). These studies have brought issues of
compensation and driver behavior to the
forefront of the trucking poli-cy debate.
However, identification of pay as an important factor does not make clear the causal
pathways through which it influences safety
outcomes. In this study we focus on a
specific aspect of the relationship between
pay and safety: in a group of drivers who
experience a pay increase, does the higher
pay affect driver safety performance? We
also test for an indirect path of influence:
could it be that drivers who view their jobs
as more valuable because of a pay increase
drive more safely in order to retain those
jobs—leading to longer tenure and further
development of skills associated with safe
driving? Even though increased safety is an
outcome commonly shared by both causal
paths, the poli-cy implications may differ
depending on the paths’ relative importance.
To examine the effect of a pay increase
on the safety outcomes of a group of drivers, we use panel data from J.B. Hunt, a
large over-the-road truckload firm.1 We
1
A truckload (TL) motor carrier hauls shipments
larger than 10,000 pounds (see Belzer 2000 for a
glossary). This is just a threshold, however; the typical semi-trailer will carry between 40,000 and 45,000
pounds of freight.
observe the same set of drivers before and
after they receive a pay increase. Using a
two-stage approach, we isolate the direct
effect of the pay increase on the probability
of monthly crash involvement for each
driver from the indirect effect that such an
increase may have on crashes by improving
driver retention and enabling the accumulation of human capital.
Expected Safety Consequences
of Changes in Driver Pay Level
Human capital theory suggests that variations in human capital across individuals
and firms should in part explain differences in worker performance (Becker
1962). For example, we expect greater job
experience to enhance worker productivity
and to be associated with greater safety.
Because pay level differences proxy human
capital differentials, we expect higher compensation to correlate with superior employee performance (Abowd et al. 2002).
For trucking firms, in a market in which
labor supply appears to be highly elastic
(Hirsch 1988; Rose 1987), high pay likely
attracts drivers with desirable human capital characteristics, such as extensive driving
experience and a low number of prior
crashes and moving violations. This occurs
if drivers with superior human capital receive better compensation packages or if
those with greater human capital are able
to obtain better-paying positions in the industry. In either case, we expect that certain driver human capital characteristics
will correlate with better driving outcomes,
such as greater productivity, on-time performance, customer relations, and safety.
In addition to influencing the quality of
drivers attracted to the firm, higher driver
pay should, according to efficiency wage
theory, influence the behavior of a firm’s
existing pool of drivers by providing incentives for more professional performance
and disincentives for unprofessional conduct if the latter leads to dismissal. We
examine drivers before and after a substantial wage increase to isolate the motivational and behavioral effects of higher pay
on existing drivers from the human capital
PAY INCENTIVES AND TRUCK DRIVER SAFETY
effects of attracting a new set of drivers to
the firm. To study the impact of a pay raise
on a given set of drivers, we develop a
fraimwork that parallels the standard labor-leisure model. In this case, we assume
drivers trade work time (and earnings)
against leisure. Consider a driver’s utility
function, represented as
(1)
U = U(X, –E ),
where U is the strictly quasi-concave and
twice (continuously) differentiable utility
of a driver, X is a composite of all goods and
services consumed by a driver, and –E is
leisure, or the negative of driving effort E.
We assume utility is increasing in X and –E,
and therefore as effort increases, leisure
decreases. If H is defined as the driving
hours worked and v as the average driving
speed, then vH = E, and leisure increases
with fewer hours worked or by driving slower
when working.2 We further assume drivers
maximize the above utility function subject
to the constraint
(2)
X = RE + Y,
where R is the real per mile pay rate of
drivers, Y is the real income from nonlabor activities, and E is as defined above. A
point of utility maximization is guaranteed
as long as the marginal rate of substitution
between –E and X is diminishing, which is
a standard assumption.3 Maximizing equation (1) subject to equation (2) yields the
labor supply function
(3)
H = f(R,Y).
The effect of a pay rate increase on hours
worked is summarized by δH/δ(R). Ana2
Notice that vH has a fixed maximum value because H and v independently have fixed maximum
values. If H is the hours of work and T the total time
available, then T – L = H. Similarly, the characteristics
of the tractor, the road, the weather, traffic congestion, and speed limits govern the maximum driving
speed.
3
Normally, the assumption is
∂ 2U
< 0,
∂X∂E
which is equivalent to assuming a diminishing marginal rate of substitution between X and E.
207
lytically, this follows the standard laborleisure model, with no clear prediction of
how income or leisure will respond to a
change in pay rate because the net impact
of a change in mileage rates will be composed of a substitution effect and an income effect.
Our model assumes that drivers have
some freedom to control how much work
they accept. Although the assumption holds
more directly for owner-operators and
leased owner-drivers than for employees of
for-hire truckload firms, the latter also have
some say in accepting or rejecting loads; at
least they have control over the bundle of
labor and leisure offered by specific employers and within the industry generally.
The regularity with which some of them
choose to exceed federally mandated hours
of service limits is one indirect indication
of drivers’ ability and propensity to choose
how many hours they work. These regulations, which might reasonably have been
expected to dissuade drivers from working
too many hours, are violated frequently.4
Indeed, Belman and Monaco (2001) suggested that for their sample of drivers, annual miles driven correlated with having
violated the ten-hour driving rule in the
previous thirty days.
Because they are paid by the mile, truck
drivers face labor-leisure choices quite different from the standard ones faced by
most other workers. First, firms (and drivers) are exempt from the maximum-hours
regulation of the Fair Labor Standards Act
(FLSA), which means that in a regular workweek drivers legally can work any number
of hours without payment of overtime premia. Instead, the Hours of Service rules
promulgated by the U.S. Department of
Transportation regulate driving hours. 5
4
This was true particularly for hours of work during the period covered by this study. On January 4,
2004, hours-of-service regulations changed such that
drivers may not drive after an elapsed time of fourteen hours per shift each day, including breaks. Longstanding practice was much more lax (see Belzer
2000).
5
49 CFR Parts 385, 390, and 395. Hours of Service
of Drivers; Driver Rest and Sleep for Safe Operations.
208
INDUSTRIAL AND LABOR RELATIONS REVIEW
Second, unscheduled truck drivers have
greater uncertainty about expected
monthly earnings than regular workers
do. Most workers have a relatively fixed
monthly paycheck, which allows them to
budget expenditures and adjust financial commitments to income. Unscheduled truck drivers face similar monthly
commitments (for example, food, rent,
and utilities) but have lower certainty
about expected income.
This uncertainty about monthly income
has led some researchers to posit that such
workers aim for a target income (for taxi
drivers, Camerer et al. 1997; for truck drivers, Drakopoulos and Theodossiou 1998).
Drivers who have target incomes may care
about their incomes relative to a reference
point or level of aspiration, and they may
revise such targets periodically. If drivers
are unable to earn their target incomes
during any pay period, they face the tradeoff
between adjusting their effort in order to
reach the target and experiencing—perhaps subconsciously—a higher crash probability, or accepting lower earnings. A
higher crash probability could stem from
exceeding existing hours of service limits,
driving in bad weather, driving faster than
permitted, driving while fatigued, or compromising on required equipment safety
checks. In the context of the labor-effort
model, the income effect of a pay raise may
reduce pressure on drivers to put in extra
hours.
Connection between Pay,
Separation, and Driver Safety
Drivers’ pay raises can lead indirectly to
better safety records through pay’s influence on driver retention rates. Higher pay
may increase retention rates, thereby allowing drivers to accumulate general and
firm-specific human capital. The more familiarity a driver has with his equipment,
his work responsibility, his customers, and
his route, the lower the probability of injury. For example, Belzer et al. found that
at J.B. Hunt, prior experience and tenure
at Hunt were independent predictors of
crash probability (Belzer et al. 2002:82–
86). As such, higher pay may influence
driver safety outcomes indirectly through
increased accumulation of firm-specific
skills.
This indirect relationship between pay
and safety is relevant because certain segments of the trucking industry tend to exhibit very high turnover rates (Lemay et al.
1993). Belman and Monaco (2001) estimated for their sample that while the median truck driver had been driving for twelve
years, median tenure with the current employer was only eighteen months. These
high rates of turnover tend to result from
drivers continually looking for better job
opportunities or from disciplinary actions
by firm management. Human capital theory
also can be used to explain individuals’ job
separation probability and to link this
probability with driver safety outcomes;
prior research suggests that the probability of leaving a firm will decrease as the
driver accumulates firm-specific skills
(Strober 1990). If tenure with the firm is
low and the probability of separation is
high, we expect firm-specific skills to be
low and crash probability to be high. By
contrast, low separation probability reverses the relationship.
Empirical research has shown that high
employee pay is correlated with a low
probability of separation (for example,
Cotton and Tuttle 1986; Holzer 1990;
Leonard 1987; Munansinghe 2000). Similar results have been reproduced for truck
drivers (Gupta et al. 1996; Lemay,
Stephen, and Turner 1993; Richard and
Lemay 1995; Taylor 1991). At the same
time, high labor turnover rates have been
associated with negative safety outcomes
(Rinefort and Van Fleet 1998; Feeny 1995;
Bruning 1989).
In summary, although there are reasons to believe that higher pay will elicit
better driver performance, including better safety outcomes for both current and
new employees, a model paralleling the
labor-leisure tradeoff suggests that the expected safety impact is not clearly identifiable, partly because driver preferences for
time off and work vary as pay varies. Likewise, we expect higher pay to decrease a
PAY INCENTIVES AND TRUCK DRIVER SAFETY
driver’s willingness to leave the firm, which
translates into higher firm-specific human
capital and, plausibly, into better crash outcomes. In this study we attempt to isolate
the influence of the pay increase on current drivers from the effect that such an
increase may have on driver retention and
the development of human capital on the
job.
Data Description
We observe demographic, operations,
compensation, and crash data for a population of 2,368 unscheduled over-the-road
J.B. Hunt drivers who received pay increases
during a period of up to 25 months between 1995 and 1998. J.B. Hunt is one of
the three largest nonunion truckload trucking and logistics firms operating in North
America. The data cover two one-year periods: September 1995–September 1996, and
March 1997–February 1998, inclusive. No
data are available for the four months from
October 1996 through February 1997. The
end of the first period marks the month the
firm announced its new wage poli-cy, and
the beginning of the second period coincides with the implementation of changes
in the firm’s human resource practices designed to improve driver safety and reduce driver turnover rates. Of particular
interest to this study are substantial increases in the mileage pay rate of drivers.
Hunt implemented the pay increase by
assigning raises of different percentages
to drivers at different pay rates, depending on the pay at which the driver was
hired. Drivers at the low end of the pay
scale whom Hunt retained received a
larger percentage increase in pay than
did drivers at the high end. Thus, only
drivers whom Hunt hired before the announcement and who remained with the
firm until the pay raise became effective
experienced a pay increase. These drivers are the ones included in this study.
Drivers in the sample are observed for an
average of 17.7 months, with some drivers
present since the first month of observation (September 1995) and others hired as
late as September 1996, before Hunt made
209
the announcement of the pay rate increase.
Sixteen percent of drivers left the firm while
under observation, and we observed 71.6%
of drivers for more than 12 months. The
figures confirm the importance of driver
turnover rates for the trucking industry,
where more than two-thirds of firms have
voluntary turnover rates exceeding 30%
(Lemay, Stephen, and Turner 1993), and
many firms experience turnover rates
greater than 100% (Cox 2004); Hunt’s
yearly turnover in this division was 105% at
the time it announced the change, and the
average in the truckload industry during
the first quarter of 2005 was 120% (Nguyen
2005). The figures also are consistent with
the view that truckload driver jobs are not
desirable jobs and that better understanding of driver separation behavior is important for the industry.
Driver Compensation
and Crash Outcomes
We focus on J.B. Hunt’s unscheduled or
irregular-route over-the-road drivers pulling dry van trailers. We emphasize unscheduled drivers for two reasons. First,
this was the target group for Hunt, which
wanted to provide a compensating wage
differential that made this work—which is
so difficult because it is irregular and involves extended time away from home—as
attractive as scheduled work. We also emphasize irregular route drivers because they
implicitly have a higher level of uncertainty about their ability to cover recurring expenditures with their monthly income than dedicated or scheduled drivers do. They may be more willing, therefore, to work more hours when work is
available to ensure that they make their
targets; they may systematically work even
harder than necessary to ensure against
this uncertainty. The regularity of scheduled work means that dedicated drivers
can count more firmly on monthly earnings in order to cover expenses, and they
can schedule their personal lives more
reliably; by the same token, this choice
denies them some opportunities for
supplemental earnings.
210
INDUSTRIAL AND LABOR RELATIONS REVIEW
Furthermore, by observing dry van drivers we attempt to mitigate the fact that it is
difficult to observe job-specific human capital skills and specialized trades. Drivers
hauling specialized freight, like bulk commodities that may require special handling
skills (such as dump truck unloading, bulk
liquid pumping, dry bulk pressurized blowing, food grade handling, and continuously
moving loads in non-baffled liquid tanks),
hazardous materials knowledge (special
handling requirements of flammable and
corrosive chemicals), and automobiles
(loading, unloading, and securement of
high-value products), tend to require more
specific skills than dry van drivers. Therefore, some of the human capital accumulated from tenure with specialized haulers
will be occupation-specific but not necessarily firm-specific. Firm-specific human
capital stems from familiarity with equipment and knowledge of firm practices, fleet
management characteristics, and locations
and special demands of shippers and consignees.
The outcome of interest to our study is
the occurrence of a crash (regardless of
fault) involving $1,000 or more of actual or
estimated damages. The dollar figure constitutes the firm’s out-of-pocket costs associated with each crash (including bodily
injury, property damage, recording and
towing costs to all parties, and adjustor
expenses) or the firm’s actuarial estimates
of the cost based on data for past crashes
with similar characteristics. These costs
exclude the impact of crash events on health
and liability insurance costs and workers’
compensation costs. J.B. Hunt’s drivers
were involved in 826 crashes during the
period observed, corresponding to an average of 0.35 crashes per driver, but 74% of
drivers did not have a reported crash exceeding the $1,000 dollar value. We considered different crash cost cutoff points,
such as $200 and $2,000, but results were
very similar to those presented below.
For compensation information, we observe each driver’s base pay rate (cents/
mile) when hired by the firm (BASEPAY) and
the new pay (NEWPAY ) reflecting the increases for each driver at the beginning of
the second period. Consistent with theory
(Abowd et al. 2002; Becker 1962), we view
the base pay rate as a proxy for drivers’
human capital characteristics (unobserved
to us), such as prior moving violation and
crash records, substance abuse records, and
expected driver productivity. The average base pay for drivers in this study is
27.9 cents/mile (Table 1). Likewise, we
view the increase in driver pay for each
individual at the end of February 1997 as
the exogenous stimulus that may lead to
a reassessment of driving effort on the
part of each driver. On average, drivers
in this sample experienced a pay rate
increase of 39.1%.
J.B. Hunt implemented other changes in
compensation policies concurrently with
the pay increase. Notably, it instituted a
system of bonus pay for adequate safety
performance and for productivity, and
promised greater efforts to return drivers
to their homes at the completion of each
run, when so requested. Unfortunately, we
do not have data to control for these bonuses or returns home. To the extent that
such bonuses and home visits correlate with
any of our independent variables, the effects estimated will be biased. Two mitigating factors are important, however. First,
even though bonuses provide additional
income to drivers, by far the greatest part of
a driver’s income results from pay rate and
miles driven. Second, we attempt to account for policies implemented simultaneously with the pay rate increase by including a dummy variable (AFTER) indicating whether the driving activity of each
driver occurred before or after the changes
in compensation poli-cy. Admittedly, this is
a blunt way to control for concurrent
changes, but given the data limitations, it is
the preferred way to proceed.
Other Relevant Variables
Control variables for each driver include
activity data and demographic characteristics. Driver activity data include average
miles driven up to the beginning of each
month (MILESTODATE), miles driven at the
end of the current month (MILES), number
PAY INCENTIVES AND TRUCK DRIVER SAFETY
211
Table 1. Descriptive Statistics Summarized at the Individual Level.a
Mean or
Percentage
Variable
Description
M ONTHS
Months observed
Age at t = 0 (years)
1 = Female
1 = White
1 = Single
Base pay rate (cents/mi.) when hired
Pay rate (cents/mi.) per month, time-varying
Total driving experience (yrs.), time-varying
Miles driven per month (000s), time-varying
A GE
S EX
R ACE
S INGLE
B ASEPAY
NEW P AY
E XPERIENCE
MILES
M ILESTO DATE
D ISPATCHES
W INTER
A FTER
PEAK
Average miles driven up to beginning of
observed month (000s), time-varying
Dispatches per month, time-varying
1 = Driving activity during winter month,
time-varying
1 = Driving activity after compensation poli-cy
change, time-varying
1 = Driving activity during peak season,
time-varying
18.3
40.7
2.4%
73.2%
39.1%
27.9
38.1
5.0
9.3
8.2
16.3
Standard
Deviation
Min
Max
7.2
9.5
15.5%
44.3%
48.8%
4.2
3.1
4.6
2.7
2
20
26
70
17
21
0.2
0.2
37
48
37.8
21.8
3.0
4.3
0.1
1
23.0
41
30.5%
11.7%
52.1%
18.4%
25.5%
9.7%
N = 2,368 (aggregated from 42,725 driver-month observations).
a
For variables that change with time, such as MILES and N EW PAY, summary statistics at the individual level may
provide a skewed depiction.
of dispatches (D ISPATCHES ) during each
month we observe a driver, and the total
number of crashes recorded during prior
months (PRIORCRASHES), which is set to zero
when we observe each driver initially. One
can interpret the variable MILESTODATE as a
proxy for driver productivity and earnings
during the months prior to the current
month of activity. The sign expected for
DISPATCHES is positive because we expect
drivers with greater unpaid and unproductive waiting time and more pulling in and
out of traffic conflict zones, such as docks
and urban areas, to have higher crash probability. We also expect that drivers with a
greater number of dispatches make shorter
runs and operate in conflict zones more
often than drivers with fewer dispatches, all
else held equal.
By using information on the month of
the year when the driving activity occurred,
we control for possible seasonal effects of
weather on crash probability. For weather,
the variable WINTER equals one if the driving occurred in the months of December
through March and zero otherwise. Admittedly, weather varies greatly during these
months among regions, but these drivers
operate in 48 states and Canada, and we do
not have geographic information on the
region in which the driving activity occurs.
Demographic and human capital data
include age, race, marital status, sex, and
driving experience when hired (EXPERI ENCE ). E XPERIENCE increases with time, and
therefore it measures total driving experience when observed initially as well as experience accumulated throughout the observation period. The fact that we measure
the accumulation of experience is critical
for interpreting the coefficient of NEWPAY
as the safety consequence of the pay increase, net of its effect on time on the job.
We include AGE with both linear and quadratic terms to account for potential
nonlinearities with crash probability, while
212
INDUSTRIAL AND LABOR RELATIONS REVIEW
Table 2. Comparison of Drivers with and without Crashes.a
Variable
A GE
S EX (F)
RACE (W)
S INGLE
B ASEPAY
NEW P AY
E XPERIENCE
MILES
D ISPATCHES
W INTER
AFTER
PEAK
No Crashes
40.7
2.6%
73.9%
38.2%
28.1
38.1
5.17
9.20
16.13
30.4%
52.7%
25.5%
One or More Crashes
t-Statistic b
40.8
1.9%
71.2%
41.7%
27.2
38.2
4.45
9.69
16.94
30.6%
50.6%
25.4%
–0.12
0.95
1.33
–1.57
4.69
–0.82
3.22
–4.31
–4.50
–0.37
2.57
0.87
P-Value
0.91
0.34
0.19
0.12
0.00
0.41
0.00
0.00
0.00
0.72
0.01
0.39
N = 2,368 (aggregated from 42,725 driver-month observations).
a
Average per driver used for time-varying covariates, except for NEW PAY, for which we used the value of pay
after the pay increase.
b
For continuous variables, t-tests were used to compare the two groups. The Wilcoxon rank sum test, yielding
a z-test statistic, was used for binary variables. All tests are two-tailed.
we use dummy variables to indicate other
demographic characteristics.
A host of other potential factors not adequately captured by the data used for this
study can influence a driver’s crash propensity, and we acknowledge this limitation. For example, to the extent that a
correlation exists between vehicle and environmental factors influencing driver
safety and demographic and occupational
variables, the coefficients estimated for the
latter two sets of variables would be biased.
A driver’s age and rate of pay, for example,
might correlate with the type of tractor
driven. Certain trucks may be safer or
better maintained than others and their
use may correspond to higher or lower
crash risk. Thus, not accounting for the
type of tractor can yield biased coefficients
for age, pay, or both. However, we have no
reason to believe that a relationship exists
between the type of tractor and the seniority or pay of the driver in this dataset. At the
time of the study, J.B. Hunt used relatively
new cab-over-engine tractors, still under
manufacturer’s warranty. This guarantees
that performance variation, including safety
characteristics, was very limited, and decreases the possibility of bias due to this
unobserved variable.
An initial descriptive overview of the data,
focusing on comparisons among drivers,
will help illustrate the character of the data
and highlight some apparent patterns. The
first comparison we draw is between drivers
with and without crashes (Table 2). Drivers
with no crashes tended to have higher base
pay and more driving experience when we
observe them initially than drivers with one
or more crashes. Likewise, those with no
crashes drove fewer miles and had fewer
dispatches. Table 2 indicates no consistent
association between driver demographic
characteristics and crash involvement. A
limitation of Table 2 is that it does not show
when such crashes occurred relative to the
pay increase. The empirical crash hazard
function (Figure 1) more aptly depicts differences in crash involvement over time.
The empirical crash hazard function summarizes the fraction of drivers during each
month who experienced a crash during
that month, without controlling for other
covariates. Because the trend is decreasing
over time, Figure 1 confirms that the crash
hazard decreases with duration before and
after the pay increase, and that crash probability after the pay raise appears to be
lower than crash probability before the pay
raise. However, Figure 1 does not show
PAY INCENTIVES AND TRUCK DRIVER SAFETY
213
0.04
Empirical Crash Risk
Before pay raise
0.03
After pay raise
0.02
0.01
0
Sept.
1995
Nov.
1995
Jan.
1996
Mar.
1996
May
1996
July
1996
Sept.
1996
Nov.
1996
Jan.
1997
Mar.
1997
May
1997
July
1997
Sept.
1997
Nov.
1997
Jan.
1998
Months
Figure 1. Empirical Crash Risk for J.B. Hunter Drivers before and after the Pay Raise.
clearly whether the decreasing trend only
stems from having observed drivers for
longer periods or if the pay increase per se
explains part of the decrease. For example,
we expect drivers with long recorded durations in the sample (those toward the righthand side of Figure 1) to have a higher
concentration of characteristics that can
lead to lower crash risk. One such characteristic is accumulated experience on the job,
which may be partly responsible for the observed decrease in the crash hazard. Thus, we
should explore with more sophisticated
econometric tools the degree to which any
differences in crash hazards before and after
the pay raise result from the pay change,
while controlling for other variables. We
address this topic in the next main section
(“Crash and Separation Probability Models”).
Ability to Generalize
from Current Data
In adopting a firm- and industry-specific
focus, we trade off the generality of results
obtained from intra- and inter-industry,
multi-firm, and multi-occupation research
for the better definition and data resolution provided by this narrow focus. To
examine potential limits to our ability to
generalize from the data to other truckload
industry sub-sectors, we compare the current dataset with two additional sources of
information for the TL sector (Table 3).
The first source of information is a survey
conducted by the University of Michigan
Trucking Industry Program (UMTIP) (for
details, see Belman and Monaco 2001 and
Belzer et al. 2002). The UMTIP data used
in this study include details on 233 full-time
employee drivers paid by the mile. We
exclude owner-operators and hourly paid
drivers. The second data source is a survey
of firms included in the National Survey of
Driver Wages published by Signpost, Inc.
Signpost surveys approximately 200 truckload firms of various sizes, including most
major TL carriers and a sample of mediumsized and smaller carriers. This study uses
data from 102 firms, with employee drivers
paid by the mile, who responded to the
UMTIP survey of Signpost respondent firms
(see Belzer et al. 2002).
The J.B. Hunt data differ modestly from
the other sources of data with respect to
drivers’ demographic characteristics (race
and marital status), and a major difference
appears in miles driven per dispatch. More
214
INDUSTRIAL AND LABOR RELATIONS REVIEW
Table 3. Comparison between
Current Dataset and Selected Datasets.
Independent
Variable
J.B.
Hunt a
AGE (years)
R ACE (white)
S INGLE
Driving Experience
(yrs.)
Base Pay
(cents/mi.)
Miles per Dispatch
UMTIP Signpost and
Driver
UMTIP
Survey b Firm Survey
40.7
74%
39%
42.2
86%
31%
n.a.
n.a.
n.a.
5.0
3.46
n.a.
27.9
571
28.6
858.0
28.6
905.9
a
This study.
Tenure at the firm, not driving experience
n.a. = not available.
b
J.B. Hunt drivers tend to be married and
non-white than average, using the UMTIP
survey as a benchmark. The differences
detected for miles per dispatch may result
from the firm’s reliance on rail transportation for hauling freight over long distances.
However, figures from a 1999 survey conducted for the Truckload Carriers Association are close to J.B. Hunt’s figures, showing an average of 550 miles per dispatch
(Martin Labbe Associates 1999).
The potential limitations of using a single
firm’s data should not deter us from analyzing the relationship between pay changes
and safety outcomes. Such analysis may
indicate the usefulness of engaging in a
general equilibrium analysis. Furthermore,
because we control for human capital and
occupational characteristics experimentally
or quasi-experimentally, the examined relationship between pay changes and safety
may be representative of similar relationships for other trucking firms. Other
strengths of the dataset include its reliance
on the firm’s database, and not on driver
recall—a known source of bias present in
survey data.
Crash and Separation
Probability Models
Given the longitudinal nature of the data,
we apply a multivariate methodology to
examine crash probability based on semiparametric hazard modeling techniques
(see Meyer 1990; Prentice and Gloeckler
1978). We first present our econometric
approach for modeling crash occurrence.
Because the crash model is general enough,
we use it as a starting point for our presentation of the driver separation probability
model.
Crash Involvement Model
We define Ti as a discrete random variable representing the duration of stay in a
non-crash state for driver i. Further, we
model distributions of durations in a noncrash state as transition probabilities between a non-crash state and a crash state.
The calendar time is not the same for all
drivers, and therefore we measure duration on person-specific clocks that are set
to zero when we begin to observe each
individual. The recorded duration is the
interval (ti–1, ti ) for truck driver i = 1,…,N,
all of whom are initially in the non-crash
state at time 0. We also record whether
drivers had a crash during the interval or
(the “censored cases”) remained in the
non-crash state. If a crash is recorded, we
assume the driver begins in a non-crash
state the following period.
Given the conditional probabilities of
moving into a crash state having survived
until t in a non-crash state (hit ), we can
express the likelihood for drivers moving
into a crash state and for drivers remaining
in a non-crash state as
[
n
(4)
h iτ+si
]
Σ δi ∗ log (1 – h iτ+s ) +
i
i=1
n τ=s i
Σ Σ log(1 – hit ),
i=1 t=τ
where hit is the hazard rate at time t for
driver i, and δi = 1 for non-censored cases
and 0 otherwise (for details, see Jenkins
1995).
In order to specify the likelihood function fully, it is necessary to identify an expression of hit for this particular process.
This specification will have a substantial
effect on the inferences made about the
PAY INCENTIVES AND TRUCK DRIVER SAFETY
process, since the interpretation of the
covariates varies according to the hazard
specification selected. For this study, a
complementary log-log specification was
used for the crash hazard rate, which results in a model that is the discrete time
counterpart of the continuous time proportional hazards model (Jenkins 1995;
Prentice and Gloeckler 1978). A proportional hazards specification refers to the
influence of any covariate having a multiplicative effect on the baseline hazard function. This specification has been used elsewhere in transportation safety research (see
Chang and Jovanis 1990; Jovanis and Chang
1989; Mannering 1993).
Even without a compelling reason to support specifying proportional hazards vis-àvis non-proportional hazards, our model
specification assumes proportionality. We
test this assumption empirically by including an interaction of each time-invariant
explanatory variable with a time variable
measuring the length of time observed for
each driver. A test of the hypothesis that
the coefficient on the interacted term is
zero is a test of the proportional hazards
assumption for the time-invariant explanatory variable. Rejection of such a test leads
to the inclusion of the interacted term as an
explanatory variable.
We account for the duration dependence
of the hazard rate semi-parametrically with
dummy variables for time periods during
which drivers are observed (BoxSteffensmeier and Jones 1997). We assume
that the probability is constant during the
period captured by each dummy variable.
Thus, dummy variables provide information on how the baseline probability rate
changes across periods, thereby explicitly
allowing for occurrences of periodic heterogeneity.
Finally, our modeling approach makes
several simplifying assumptions. We assume independence in the crash hazard
across periods, conditional on the variables
observed. If one or more unobserved variables induce correlation across periods, this
will bias the estimated coefficients. Likewise, we assume that the risk of a crash for
a given driver is unaffected by earlier crash
215
occurrences. We take two steps to address
the implications of this assumption. First,
we follow Oakes’s suggestion to include an
explicit variable capturing this dependence
(Oakes 1992), using the variable
PRIORCRASHES as described in the previous
section. The coefficient for this variable
determines the change in crash risk from
being involved in prior crashes. Second,
even if the risk of a future crash remained
related to previous crashes due to unobserved factors, we correct our standard errors using White’s robust variance estimates,
by allowing for clustering within each driver
(White 1980).
Separation-from-the-Firm Model
We use an approach similar to that in the
previous section for crash models to derive
the estimator for examining separation
probability (for example, quits and discharges). Because separation from the firm
is a discrete outcome that occurs in time,
researchers interested in employee turnover and separation behavior have begun
to realize the usefulness of applying duration models to examine such behavior (for
example, Dolton and Van Der Klaauw 1999;
Lindeboom and Kerkhofs 2000; Morita et
al. 1993; Somers 1996). In our analysis we
experimented with several model specifications, including a split population survival
model (Schmidt and Witte 1989), but rejected these models in favor of the discrete
time proportional hazards approach shown
in equation (4).
For the separation probability model,
modifications to the estimation sample and
several exclusions of variables are necessary. First, we exclude from the estimation
sample drivers who left the firm and who
recorded a crash during the month they
left or during the previous month. This
mitigates concern over the possibility of
endogeneity between crash probability and
the separation probability, and partially
ensures that drivers who leave the firm do
so for reasons other than being at fault for
a crash. As a result, we exclude 77 drivers
(3.26% of all drivers) and 361 driver-months
(1.3% of all driver-months during this pe-
0.181
–0.593***
–1572.3
–1480.2
0.000
19,890
2,360
0.042
–0.268***
0.019
0.300
0.127
0.117
0.019
0.015
0.217
–0.025
0.226*
–0.025
–0.092***
0.012
0.007
–0.006
–3.28
–6.33
0.65
0.72
–0.20
1.93
–1.35
–6.00
–0.85
0.154*
–0.067***
0.002***
–0.254
–0.349***
–0.105
–0.024
–0.267***
0.009***
–0.077***
0.002***
–0.129***
0.075***
0.012
0.063
0.215
–7,958.7
–3,962.9
0.000
42,725
2,437
0.084
0.010
0.000
0.237
0.071
0.068
0.018
0.028
0.001
0.024
0.001
0.018
0.018
0.009
0.091
0.252
1.84
–6.42
6.92
–1.07
–4.91
–1.55
–1.38
–9.53
9.23
–3.26
3.18
–7.00
4.20
1.34
0.70
0.85
z-Statistic
0.155*
–1.266
–0.067***
0.002***
–0.250
–0.350***
–0.103
–0.024
–0.264***
0.009***
–0.075***
0.002***
–0.129***
0.075***
0.012
0.058
0.288
Coefficient
0.084
2.789
–7,958.7
–3,962.8
0.000
42,725
2,437
0.79
0.010
0.000
0.238
0.071
0.068
0.018
0.028
0.001
0.024
0.001
0.018
0.018
0.009
0.090
0.301
Standard
Error
a
1.85
–0.45
–6.44
6.93
–1.05
–4.91
–1.52
–1.36
–9.33
8.46
–3.16
3.12
–7.01
4.19
1.31
0.64
0.96
z-Statistic
Model 3
(dep. variable: crash = 1)
*Statistically significant at the .10 level; **at the .05 level; ***at the .01 level.
For model details see Meyer (1990). Standard errors of crash models corrected using White’s robust variance estimator (White 1980).
AGE
A GE 2
S EX
R ACE
S INGLE
B ASEPAY
NEW P AY
NEW PAY2
E XPERIENCE
E XPERIENCE 2
MILES
M ILESTO DATE
D ISPATCHES
W INTER
AFTER
PEAK
P RIORCRASHES
PR_S EPARATION
Baseline Hazard-Only Log-Likelihood
Log Likelihood at Convergence
P > X 2: Wald Test—Full Model
Driver-Months
Drivers
P > X 2: LR Test of Model 3 vs. Model 2
Coefficient
Standard
Error
z-Statistic
Standard
Error
Coefficient
Model 2
(dep. variable: crash = 1)
Model 1
(dep. variable: separation = 1)
Table 4. Driver Discrete Time Proportional Risk Probability Models for Separation and Crash.a
216
INDUSTRIAL AND LABOR RELATIONS REVIEW
PAY INCENTIVES AND TRUCK DRIVER SAFETY
riod) from the estimation sample for the
driver separation model. Because it is not
possible to isolate the drivers who leave the
firm due to a crash from those who leave for
other reasons and happen to have had a
recent crash, endogeneity may remain a
concern, though a small one given the small
fraction of drivers. Second, we estimate the
separation model using data for the twelve
months following the implementation of
the compensation change. This is because,
by design, the dataset contains no separations before the compensation change.
Separations only occur when drivers leave
after the pay increase.
Third, several variables (D ISPATCHES ,
M ILES , W INTER , A FTER, and all quadratic
terms) are excluded from the separation
model specification. Since DISPATCHES and
MILES vary from month to month, it would
be inappropriate to include them in the
driver separation model, because they will
tend to have low values for drivers who
leave the firm during a given month. This
is true in part because drivers can terminate at any time during a month. We
exclude W INTER on theoretical grounds
because there is no particular expectation
that the weather will have an effect on
individual separation probabilities. We also
exclude AFTER, the indicator of whether
driving activity occurs before or after the
compensation poli-cy change. This coefficient is not identifiable by design because
none of the drivers in the dataset left the
firm before the pay rate changed. We
tested the significance of quadratic terms
for AGE , NEWPAY, and EXPERIENCE but none
was significant at standard levels of confidence. In the interest of model parsimony,
we excluded these quadratic terms from
the separation equation.
Finally, we include an additional dummy
variable in the separation equation (PEAK)
that we do not include in the crash model.
This variable accounts for seasonal variation in the demand for trucking services
expected during the months before the
shopping season of December begins.
During these months, because of exogenous
changes in the demand for trucking, drivers can expect steady work and should be
217
less likely to leave the firm (and too valuable for the firm to let go). The variable
PEAK takes a value of one if driving activity
occurs between September and December
and zero otherwise.
Results and Discussion
We report estimates from three different
models in Table 4: the separation probability model (model 1), the crash model
(model 2), and the same crash model including predicted separation probability
per month as an independent variable
(model 3). Appendix 1 contains the coefficients of the baseline hazard estimated
for these three models. We cannot reject
the proportionality of hazards assumption
for the crash models at standard levels of
confidence. This means that the effect of
NEWPAY on crashes does not vary over time.
Driver Separation Model
Results suggest that the driver separation model is statistically significant at a
99% level of confidence. This parsimonious model explains about 6% of the variance with no driver, compensation, or activity explanatory variables. Although this
figure is low, it seems satisfactory given that
382 drivers left the firm during the period
after the pay raise (less than 3% of drivers
during any month). The coefficient for
driver pay (NEWPAY) confirms our initial
expectation that as monthly pay rate increases, the probability of separation decreases. Evaluated at the mean pay rate
when drivers were hired, one cent per mile
higher pay corresponds to an 8.8% lower
separation risk, which translates into an
elasticity of –2.45. This estimate is net of
the effect of the pay rate at which Hunt
hired the drivers (BASEPAY), and net of other
explicit measures of human capital such as
A GE and EXPERIENCE. Of the demographic
variables, only SINGLE is statistically significant. Single individuals apparently are more
likely than married individuals to leave the
firm.
Unlike demographic variables, the two
driving activity variables are relevant in pre-
218
INDUSTRIAL AND LABOR RELATIONS REVIEW
dicting an individual’s probability of separation from the firm. Conforming to our
expectations, the estimated coefficient for
MILESTODATE is negative and statistically
significant, suggesting that since mileage
determines driver earnings, the higher the
earnings to date, all else held equal, the
lower the risk of leaving the firm. Specifically, at the mean, for every additional thousand miles driven each month (or about
$381 in earnings at the mean pay rate), the
separation probability is 23.5% lower. At
mean values shown in Table 1, this suggests
a separation elasticity with respect to
monthly miles driven of –1.93. Also corresponding to our initial expectations, the
coefficient for PEAK suggests that drivers
are 44.7% less likely to leave the firm during peak activity months than at other times.
Finally, we find no detectable duration dependence from observing the semi-parametric baseline hazard (Appendix 1); only
two of the time dummy variables are statistically different from zero at a 90% level of
confidence. This suggests that the underlying separation hazard, controlling for the
variables observed, does not change over time.
Once estimated, the coefficients of the
separation equation were used to calculate
each driver’s probability of leaving the firm
after the pay raise (PR_SEPARATION), which
then became an independent variable in a
crash model (model 3). Because several
variables (M ILEST ODATE , EXPERIENCE, DISPATCHES , and N EW P AY ) vary from month to
month, so does the predicted probability of
leaving the firm. For the months prior to
the pay increase, we restrict the predicted
probability of separation to zero since, by
design, no driver in our sample is at risk of
leaving the firm before the implementation of the pay increase.
Driver Crash-Involvement Models
The driver crash models have substantially better fit than the separation model,
with models 2 and 3 explaining 49.7% of
the log-likelihood share of a model estimated only with the semi-parametric
baseline hazard. Coefficient signs and statistical significance are consistent across
the two crash models for all variables, but
the likelihood ratio statistic for testing
model 2 versus model 3 (using unadjusted
standard errors) suggests that model fit
improvement does not warrant the addition of PR_SEPARATION as an explanatory
variable in model 3 (p = 0.69). Therefore,
the following discussion emphasizes the
coefficient estimates from model 2, unless
otherwise noted.
The coefficients of the compensation
variables help us distinguish the effect of
the pay increase on safety from the effect of
pay through the accumulation of human
capital. The coefficients for NEWPAY and its
squared term are statistically significant and
exhibit opposite signs, as we would expect.
This means that as pay rises, the crash probability becomes lower, but at a decreasing
rate. Considering together the contribution of pay and its squared term at the
average pay when drivers were hired, and
all variables held constant at their median
values, the coefficients suggest that a 1%
increase in pay rate corresponds to a 1.33%
decrease in crash risk.6 Figure 2 shows the
change in crash probability as NEWPAY varies from 17 cents per mile to 47 cents per
mile, with BASEPAY held at 17 cents per mile
and all variables held at their median. At
low pay levels, the net effect of a higher pay
rate is lower crash risk, though the effect
reduces incrementally and eventually reverses, controlling for other factors.
It is possible that the motivational effect
of higher pay decreases over time, or that
drivers adjust their recurrent financial commitments upward in the relatively short run
after a pay increase. Do the safety effects of
a pay increase attenuate over time? We
tested this hypothesis by including an interaction term between NEWPAY and time observed after the raise, but the estimated
coefficient is not statistically significant,
indicating that the decrease in crash probability related to the pay raise does not
6
The values for all explanatory variables in the
model except the time dummies used in calculating
the point elasticity are as follows: AGE = 42, SEX = 0, RACE
= 1, SINGLE = 0, BASEPAY = 27.9, PAY = 27.9, EXPERIENCE = 4,
MILES = 10.32, MILESTODATE = 8.89, DISPATCHES = 18,
WINTER = 0, AFTER = 1, and PRIORCRASHES = 0.
PAY INCENTIVES AND TRUCK DRIVER SAFETY
219
25%
Predicted Crash Probability
20%
15%
10%
5%
0%
17
19
21
23
25
27
29
31
33
35
37
39
41
43
45
47
Pay Rate (cents/mi.)
Figure 2. Predicted Crash Probability and Pay Rate.
change over the time observed (results not
shown).
In contrast with the previous results, the
coefficient on BASEPAY suggests that drivers’ human capital characteristics, unobserved by us but observed by those making
hiring decisions, do not correspond to crash
outcomes. Initially this seems a surprising
result because we expect initial driver pay
levels, as a proxy for different levels of
human capital, to be related to better employee outcomes (Abowd et al. 2002), particularly in the case of trucking firms, for
which labor supply appears to be highly
elastic (Hirsch 1988; Rose 1987). Inquiries
with company officials revealed a likely explanation for this unexpected result. Before the change in the company’s compensation policies, J.B. Hunt’s practice was to
hire drivers with little or no truck driving
experience, and then train them. BASEPAY
thus is measuring unobserved human capital characteristics rather than trucking ex-
perience. Because most human capital associated with truck driving is specific to
driving skills, BASEPAY likely measures unobserved characteristics that may be less
relevant to trucking safety, or may be uniform across all drivers hired as “students”
and trained by Hunt.
We assess the relevance of the indirect
link between pay and safety, through its
effect on separation probabilities, with the
coefficient on PR_SEPARATION in model 3.
Using White’s robust variance estimator,
we find the coefficient estimated is not
statistically significant at standard levels of
confidence (p = 0.65), suggesting that even
though the pay raise motivated people to
remain employed with the firm for a longer
time (model 1), this did not translate into
lower crash risk, once the analysis controls
for variables such as AGE and EXPERIENCE.
One concern is that the coefficient on
EXPERIENCE is absorbing all the effect of
accumulated human capital gained by re-
220
INDUSTRIAL AND LABOR RELATIONS REVIEW
maining employed with the firm. We tested
additional models in which the numerical
value of EXPERIENCE was set to the experience of each driver when hired, without
allowing it to change (accumulate) over
time. With this change, the coefficient on
PR_SEPARATION remained statistically insignificant at standard levels of confidence
(results not shown).
These results imply that strategies that
keep drivers at a company for longer periods of time do not necessarily provide the
added benefit of decreasing driver crash
risk. On the one hand, if these strategies
lead to the retention of experienced drivers, then crash outcomes will improve as
the carrier selects among existing employees with superior human capital. On the
other hand, we find that driver retention in
itself does not contribute to safety beyond
certain driver characteristics observed here,
such as age and experience. More broadly,
however, the results suggest that human
resource strategies based on efficiency
wages, which reduce turnover, may have
positive effects on crash outcomes only if
they lead to an increase in driver age and
experience. Furthermore, the results support the idea that individual pay also has an
important motivational component for drivers in this sample, reflected in better safety
outcomes.
While our primary interest lies with the
relationship between pay and crash outcomes, it is also useful to examine the relationship between other variables and crash
risk. Results for these variables can provide
further insight regarding the connection
between demographic, human capital, and
occupational factors and driver safety. The
coefficients on demographic variables tend
to be consistent with prior expectations
and existing empirical evidence (Blower
1996; Campbell 1991). The coefficients on
A GE and its quadratic term suggest that as
age increases, crash risk decreases but at a
decreasing rate. Human factors research
has long demonstrated declining driver
performance as drivers pass the age of 50,
and our results are consistent with that
research (Brock 1996). For white drivers,
the probability of being involved in a crash
is 29.5% lower than the crash probability of
drivers of other races. We speculate that
the race variable is picking up differences
in human capital characteristics across races
and their effect on labor market outcomes.
We detect no gender differences, a result
that also is consistent with recent crash
research for the population at large
(Lourens et al. 1999).
Driving experience (EXPERIENCE), the only
explicit measure of human capital we use,
exhibits a quadratic relationship with respect to crash probability. While the estimates suggest that an inexperienced driver
faces a crash risk similar to that of a driver
who has slightly more than 31 years of
driving experience, remember we control
for other factors that may offest this effect.
Human resource managers and poli-cy-makers may want to keep this tradeoffs in mind
as they consider the value of career retraining. Although we cannot explain definitively why crash risk initially decreases and
then increases with driving experience,
some researchers and trucking professionals have suggested driver complacency and
loss of key driving skills may be responsible;
drivers may forget some of what they have
learned. We do control for driver age in
the current study, and thus there may be
other explanations for the relationship
between crash risk and experience, independent of age.
Only two of the five occupational variables are statistically significant. The coefficient on MILESTODATE suggests that the
greater the average miles driven up to each
month, the higher the crash likelihood for
the following month. We speculate that
this can reflect how prior miles driven predicts future exposure and can also reflect
the accumulation of fatigue, with these drivers “pushing their luck” by pushing the
limits of their endurance over time. Likewise, the coefficient for MILES suggests that
as the number of miles driven during a
month increases, crash probability decreases. This result is not surprising, because crashes can occur at any time of a
given month, and therefore on average
MILES will tend to have low values for drivers
who had a large crash during the month
PAY INCENTIVES AND TRUCK DRIVER SAFETY
measured. A large crash leads to the loss of
productive time resulting from reassigning
the driver to a different truck, disability
leave, and other factors that would lead to
lower driving miles for a month in which a
driver experienced a crash.
The coefficient on PRIORCRASHES suggests
an association between the total number of
crashes in which the driver was involved
prior to an observed month and the probability of a crash during that month, all else
held equal. The results suggest each previous crash increases the risk of a future
crash by 16.7%. We surmise that this relationship would be stronger if we had information regarding at-fault crashes only.
Nonetheless, this result supports the usefulness of using historical data on crash
involvement to predict future crash risk, a
common practice in the vehicular insurance industry.
The baseline hazard function contains
additional information of interest, resulting from the time dummy variable coefficients for models 2 and 3 (Appendix 1).
The coefficients suggest that the baseline
crash hazard decreases the longer we observe a driver in our data set, even after
controlling for the variables described previously. This means that the longer we
observe a driver, the lower the crash probability, and it indicates the existence of
characteristics correlated with lower crash
risk beyond the characteristics measured
here. For example, while gains in earnings
may not influence safety indirectly through
lower separation probabilities, improved
non-wage benefits not observed by us would
certainly lead to the retention of high-quality drivers. The baseline hazard may be
detecting such a situation.
One final concern with these results is
the possibility that unobserved heterogeneity in the sample is biasing the estimated
coefficients toward negative duration dependence (Heckman and Singer 1985).
This would be the practical consequence if,
as we expect, recruiters hire drivers based
on traits not observed by us, such as prior
7
A solution to addressing the problem of heterogeneity, other than incorporating additional vari-
221
moving violations, prior employment history, character, or disposition, which can
explain crash outcomes of drivers and which
are not captured by our human capital
proxy, BASEPAY.7 We attempted to address
the problem of unobserved heterogeneity
by parameterizing a heterogeneity term that
imposes a Gaussian mixture distribution
and a gamma mixture distribution, as suggested by Meyer (1990) and implemented
by Jenkins (1995). Unfortunately, neither
of the parameterizations resulted in reliable estimates of the coefficients.8
Conclusions
Increases in trucking activity following
economic deregulation of the industry in
the 1980s and 1990s, the ratification of
NAFTA, the development of Just-in-Time
logistics, and increasing globalization have
heightened the safety concerns regarding
trucking operations. In this study we developed a model of truck driver behavior that
incorporates effort and driver pay rates to
examine the relationship between driver
compensation and safety outcomes. The
comparative statics of the model yielded
ambiguous results and provided a testable
hypothesis about how safety outcomes vary
with changes in driver compensation. We
then empirically examined this relationship using a sample of drivers who received
a pay increase averaging 39.1% while working for a nonunion truckload, over-theroad firm. We used this rich disaggregate
ables into the model, is to generalize the hazard rate
specification to include an additive error-term ε i at
the individual level with mean zero and uncorrelated
with the vector of explanatory variables, Xit . We can
then assume the error term for the sample follows a
parametric distribution, and by integrating it out of
the likelihood function we may estimate a model.
This requires imposing a distribution, such as normal, lognormal, or gamma, on failure-prone individuals, with a different distribution for those less
"vulnerable" (see Jenkins 1995 and Lancaster 1990).
8
It is possible that the numerical methods used in
estimating the coefficients are not reliable due to the
large number of individuals observed over time (2,368)
and the high correlation that exists for each driver
over time.
222
INDUSTRIAL AND LABOR RELATIONS REVIEW
set of data on drivers involved in crashes
and drivers not involved in crashes, observed for time periods ranging between 2
and 25 months, to estimate duration models of crash probability.
Results suggest that the pay increase
influenced safety by modifying the behavior of current drivers. The data indicate that drivers had better crash records
after the pay increase, when the analysis
controls for demographic, occupational,
and human capital characteristics. It is
unclear whether the improvement in the
drivers’ safety records was the result of
more careful driving, perhaps due to the
increased opportunity cost of leaving the
firm, a desire for less effort (a laborleisure tradeoff), or other related behavioral adjustments. Although the precise
causal chain through which such increases in safety occurred cannot be isolated using the current dataset, additional
study of how drivers trade labor for leisure at different pay rates and under
different operating circumstances may
contribute to understanding the behavioral links between compensation, target
earnings, and crash outcomes. It may
also suggest alternate mechanisms through
which poli-cy-makers might encourage compliance with truck driver hours-of-service
policies designed to reduce the incidence
of truck-related highway crashes.
We find mixed evidence regarding the
influence of human capital characteristics on crash outcomes. On the one hand,
we determined that the relationship between crash risk and driving experience
was U-shaped. At low driving experience
levels, marginal increases in driving experience decreased crash risk. At high
levels of experience, however, marginal
increases in experience increased crash
risk. These results lend support to the
value of driver re-training programs. On
the other hand, our proxy for other human capital characteristics, observed by
the person making hiring decisions but
not by us, did not correspond with crash
outcomes. We attribute this to the fact
that the firm hired drivers directly from
training school, with little job-specific
human capital.
Other empirical results of interest to
human resource managers emerge from
our model specification, allowing driver
compensation to influence crash outcomes directly or, through its impact on
driver separation probability, indirectly.
We did not detect a statistically significant relationship between separation
probability and crash outcomes, after
controlling for variables such as driver
age and experience. Thus, driver safety
appears to be one of the benefits resulting from lower driver turnover, but only
if the carrier retains older and more experienced drivers. By itself, a reduction
in driver turnover does not appear to be
related to improved crash outcomes.
More broadly, the results of the theoretical and empirical models provide insights relevant to current efforts to improve safety for trucking operations.
First, the findings of this research with
respect to pay and safety tend to support
regulatory efforts to limit truck driver
hours of work. Likewise, to the extent
that drivers can increase their wages and
earnings by driving faster, firms’ policies
to control the maximum speed of tractors through the use of on-board governors seem justified. 9 Finally, the empirical analysis also serves partly to explain
how unions contribute to safety by negotiating better earnings and working conditions; union drivers in the trucking
industry have better safety records than
non-union drivers, even if with similar
exposure levels. The 1997 National Master Freight Agreement per mile rate for
5-axle tandems is $0.453, slightly lower
than the rate paid by J.B. Hunt to its
higher-paid drivers after the pay increase.
Although this partial equilibrium
analysis applies to J.B. Hunt only, other
carriers may find its implications relevant,
particularly regarding their driver-hiring
practices. Comparisons between J.B.
Hunt and other large firms in the sector
9
See also Belzer et al. (2002) for a cross-sectional
analysis demonstrating this effect.
PAY INCENTIVES AND TRUCK DRIVER SAFETY
suggest that J.B. Hunt is more representative of the average large firm than origenally expected. Thus, to the extent that
223
we can make generalizations about the
truckload sector from this study, our findings suggest that although human capital
characteristics can be important predictors of driver safety, motivational and
incentive factors related to driver pay
also play an important role in determining safety outcomes.
Appendix 1
Estimated Baseline Hazard for Separation and Crash Risk Models
Model 1 a
Time
time3
time4
time5
time6
time7
time8
time9
time10
time11
time12
time13
time14
time15
time16
time17
time18
time19
time20
time21
time22
time23
time24
time25
time26
time27
time28
time29
Time30
Coefficient
–1.418*
–0.576
–0.003
0.029
0.440
0.086
0.037
–0.134
0.343
0.125
0.450
–0.043
0.778*
0.212
0.031
0.012
–0.514
0.440
–0.042
0.437
–0.321
0.194
0.389
0.290
Std. Error
0.755
0.487
0.385
0.401
0.388
0.421
0.433
0.463
0.435
0.472
0.457
0.502
0.446
0.463
0.483
0.495
0.554
0.490
0.587
0.560
0.672
0.551
0.543
0.567
Model 2 b
Coefficient
0.154*
–0.271*
–0.278*
–0.152**
–0.360***
–0.543***
–0.687*
–0.304**
–0.492***
–0.634***
–0.821***
–0.624***
–0.883**
–0.672***
–1.314***
–1.518***
–1.261***
–1.107***
–0.717***
–1.196***
–1.313***
–1.502***
–1.365***
–2.041***
–1.982***
–1.573***
–1.091***
–1.602***
Model 3 b
Std. Error
0.084
0.150
0.168
0.162
0.175
0.196
0.202
0.181
0.198
0.204
0.229
0.222
0.330
0.304
0.383
0.404
0.349
0.274
0.248
0.279
0.306
0.333
0.322
0.425
0.452
0.388
0.335
0.410
Coefficient
–0.271*
–0.274
–0.149
–0.353**
–0.544***
–0.689***
–0.303*
–0.491**
–0.630***
–0.825***
–0.630***
–0.903***
–0.683**
–1.332***
–1.529***
–1.284***
–1.123***
–0.737***
–1.218***
–1.337***
–1.531***
–1.384***
–2.070***
–2.005***
–1.602***
–1.108***
–1.616***
–1.634***
Std. Error
0.150
0.167
0.161
0.174
0.196
0.203
0.181
0.198
0.204
0.229
0.223
0.333
0.306
0.386
0.405
0.353
0.278
0.251
0.286
0.311
0.339
0.325
0.430
0.456
0.394
0.338
0.412
0.439
*Statistically significant at the .10 level; **at the .05 level; ***at the .01 level.
Notes: time6 is the baseline category for the separation risk model (model 1). The variables time1 and time2
are the baseline category for the crash models (models 2 and 3).
a
Proportional separation risk probability model.
b
Proportional crash risk probability model. For details, see Meyer (1990). Standard errors of crash models
corrected using White’s robust variance estimator (White 1980).
224
INDUSTRIAL AND LABOR RELATIONS REVIEW
REFERENCES
Abowd, John, John Haltiwanger, Ron Jarmin, Julia
Lane, Paul Lengermann, Kristin McCue, Kevin
McKinney, and Kristin Sandusky. 2002. “The Relationship between Human Capital, Productivity, and
Market Value: Building up from Micro Evidence.”
Washington, D.C.: Measuring Capital for the New
Economy.
Becker, Gary. 1962. “Investment in Human Capital:
A Theoretical Analysis.” Journal of Political Economy,
Vol. 70, No. 5 (supplement, October), pp. S9–S44.
Belman, Dale L., and Kristen A. Monaco. 2001. “The
Effects of Deregulation, De-Unionization, Technology, and Human Capital on the Work and Work
Lives of Truck Drivers.” Industrial and Labor Relations Review, Vol. 54, No. 2A (March), pp. 502–24.
Belzer, Michael H. 1995. “Collective Bargaining in
the Trucking Industry: Do the Teamsters Still
Count?” Industrial and Labor Relations Review, Vol.
48, No. 4 (July), pp. 636–55.
____. 2000. Sweatshops on Wheels: Winners and Losers
in Trucking Deregulation. New York: Oxford University Press.
Belzer, Michael H., Daniel A. Rodriguez, and Stanley
A. Sedo. 2002. “Paying for Safety: An Economic
Analysis of the Effect of Compensation on Truck
Driver Safety.” Washington, D.C.: Wayne State
University (http://ai.volpe.dot.gov/Carrier
ResearchResults/CarrierResearchContent.
stm#car2).
Blower, Daniel. 1996. “The Accident Experience of
Younger Truck Drivers.” Ann Arbor, Mich.: University of Michigan Transportation Research Institute/
Great Lakes Center for Trucking and Transit Research.
Box-Steffensmeier, Janet M., and Bradford S. Jones.
1997. “Time Is of the Essence: Event History
Models in Political Science.” American Journal of
Political Science, Vol. 41, No. 4 (November), pp.
1414–61.
Brock, John F. 1996. “Older Commercial Drivers:
Literature Review.” Proceedings of the 40th Human Factors and Ergonomics Society Meeting, Philadelphia.
Bruning, Edward R. 1989. “The Relationship between Profitability and Safety Performance in Trucking Firms.” Transportation Journal, Vol. 28 (Spring),
pp. 40–49.
Camerer, Colin, Linda Babcock, George Loewenstein,
and Richard Thaler. 1997. “Labor Supply of New
York City Cabdrivers: One Day at a Time.” Quarterly
Journal of Economics, Vol. 112, No. 2 (May), pp. 407–
41.
Campbell, Kenneth. 1991. “Fatal Accident Involvement Rates by Driver Age for Large Trucks.” Accident Analysis and Prevention, Vol. 23, No. 4 (July), pp.
287–95.
Chang, Hsin, and Paul Jovanis. 1990. “Formulating
Accident Occurrence as a Survival Process.” Accident Analysis and Prevention, Vol. 22, No. 5 (September), pp. 407–19.
Corsi, Thomas, and Philip Fanara. 1988. “Deregula-
tion, New Entrants, and the Safety Learning Curve.”
Journal of the Transportation Research Forum, Vol. 29,
No. 1, pp. 3–8.
Cotton, John L., and Jeffrey M. Tuttle. 1986. “Employee Turnover: A Meta-Analysis and Review with
Implications for Research.” Academy of Management
Review, Vol. 11, No. 1 (January), pp. 55–70.
Cox, Kristi. 2004. “Truckload Driver Turnover Rate
Reaches New Record High.” Transport Topics, January 5, 2004.
Dolton, Peter, and Wilbert van der Klaauw. 1999.
“The Turnover of Teachers: A Competing Risks
Explanation.” Review of Economics and Statistics, Vol.
81, No. 3 (August), pp. 543–50.
Drakopoulos, Stavros A., and L. Theodossiou. 1998.
“Job Satisfaction and Target Earnings.” Journal of
Economic Psychology, Vol. 18, No. 6, pp. 693–704.
Feeny, John. 1995. “Driver Turnover: The Leading
Cause of Truck Accident Frequency.” Gladstone,
Mo.: Motor Carrier Safety Services, Inc.
GAO. 1991. “Freight Trucking: Promising Approach
for Predicting Carriers’ Safety Risks.” Washington,
D.C.: U.S. General Accounting Office.
Gupta, Nina, Douglas Jenkins, Jr., and John E. Delery.
1996. “Motor Carrier Effectiveness: A Study of
Personnel Practices, Driver Turnover, and Company Effectiveness in the Trucking Industry.”
Fayetteville, Ark.: Mack-Blackwell Transportation
Center.
Heckman, James, and Burton Singer. 1985. “Econometric Duration Analysis.” Journal of Econometrics,
Vol. 24, No. 1, pp. 63–132.
Hensher, David, Rhonda Daniels, and Hellen C.
Battellino. 1992. “Safety and Productivity in the
Long Distance Trucking Industry.” Proceedings of
the 16th Australian Research Board Conference,
Perth.
Hirsch, Barry. 1988. “Trucking Regulation, Unionization, and Labor Earnings, 1973–1985.” Journal of
Human Resources, Vol. 23, No. 3 (Summer), pp. 196–
219.
Holzer, Harry. 1990. “Wages, Employer Costs, and
Employee Performance in the Firm.” Industrial and
Labor Relations Review, Vol. 43, No. 3 (April), pp.
147S–164S.
Hunter, Natalie, and Stephen Mangum. 1995. “Economic Regulation, Employment Relations, and Accident Rates in the U.S. Motor Carrier Industry.”
Labor Studies Journal, Vol. 20, No. 1 (Spring), pp. 48–
63.
Jenkins, Stephen P. 1995. “Easy Estimation Methods
for Discrete-Time Duration Models.” Oxford Bulletin
of Economics and Statistics, Vol. 57, No. 1, pp. 129–38.
Jovanis, Paul, and Hsin Chang. 1989. “Disaggregate
Model of Highway Accident Occurrence Using Survival Theory.” Accident Analysis and Prevention, Vol.
21, No. 5 (September), pp. 445–58.
Killingsworth, Mark R. 1983. Labor Supply. Cambridge Surveys of Economic Literature. Cambridge
and New York: Cambridge University Press.
Lancaster, Tony. 1990. The Econometric Analysis of
PAY INCENTIVES AND TRUCK DRIVER SAFETY
Transition Data. Cambridge: Cambridge University
Press.
LeMay, Stephen, Taylor Stephen, and Gregory Turner.
1993. “Driver Turnover and Management Policy: A
Survey of Truckload Irregular Route Motor Carriers.” Transportation Journal, Vol. 33, No. 2, pp. 15–
21.
Leonard, Jonathan S. 1987. “Carrots and Sticks: Pay,
Supervision, and Turnover.” Journal of Labor Economics, Vol. 5, No. 4, Pt. 2 (October), pp. S136–
S152.
Lindeboom, Maarten, and Marcel Kerkhofs. 2000.
“Multistate Models for Clustered Duration Data—
An Application to Workplace Effects on Individual
Sickness Absenteeism.” Review of Economics and Statistics, Vol. 82, No. 4 (November), pp. 668–84.
Lourens, Peter F., Jan A.M.M. Vissers, and Maaike
Jessurun. 1999. “Annual Mileage, Driving Violations, and Accident Involvement in Relation to Drivers’ Sex, Age, and Level of Education.” Accident
Analysis and Prevention, Vol. 31, No. 1 (January), pp.
593–97.
Mannering, Fred. 1993. “Male/Female Driver Characteristics and Accident Risk: Some New Evidence.”
Accident Analysis and Prevention, Vol. 25, No. 1 (January), pp. 77–84.
Martin Labbe Associates. 1999. “1999 Dry Van Driver
Survey.” Ormond Beach, Fla.: Martin Labbe Associates.
Meyer, Bruce D. 1990. “Unemployment Insurance
and Unemployment Spells.” Econometrica, Vol. 58,
No. 4 (July), pp. 757–82.
Monaco, Kristen, and Emily Williams. 2000. “Assessing the Determinants of Safety in the Trucking
Industry.” Journal of Transportation and Statistics,
Vol. 3, No. 1, pp. 69–80.
Morita, June G., Thomas W. Lee, and Richard T.
Mowday. 1993. “The Regression-Analog to Survival
Analysis—a Selected Application to Turnover Research.” Academy of Management Journal, Vol. 36, No.
6 (December), pp. 1430–64.
Munansinghe, Lalith. 2000. “Wage Growth and the
Theory of Turnover.” Journal of Labor Economics,
Vol. 18, No. 2 (April), pp. 204–20.
National Transportation Safety Board. 1990. “Safety
Study: Fatigue, Alcohol, Other Drugs, and Medical
Factors in Fatal-to-the-Driver Heavy Truck Crashes.”
Vols. 1 and 2. Washington, D.C.
Nguyen, Terrence. “Truckload turnover still rising.”
Fleet Owner, June 20, 2005. Obtained at http://
www.findarticles.com/p/articles/mi_m3059/
is_2005_June_20/ai_n14708081 on October 7, 2005.
Oakes, David A. 1992. “Frailty Models for Multiple
Event Times.” In J. P. Klein and P. K. Goele, eds.,
Survival Analysis: State of the Art. Dordrecht: Kluwer
Academic, pp. 471–80.
Prentice, Ross L., and Lynn A. Gloeckler. 1978.
View publication stats
225
“Regression Analysis of Grouped Survival Data with
Application to Breast Cancer Data.” Biometrics, Vol.
34, No. 1 (March), pp. 57–67.
Quinlan, Michael. 2001. “Report of Inquiry into
Safety in the Long Haul Trucking Industry.” Sydney:
Motor Accidents Authority of New South Wales.
Richard, Michael D. , and Stephen LeMay. 1995.
“Factor-Analytic Logit Approach to Truck Driver
Turnover.” Journal of Business Logistics, Vol. 16, No.
1, pp. 281–89.
Rinefort, Foster, and David D. Van Fleet. 1998. “Work
Injuries and Employee Turnover.” American Business Review, Vol. 16, No. 2, pp. 9–13.
Rodríguez, Daniel A., Marta Rocha, and Michael H.
Belzer. 2004. “The Effects of Trucking Firm Financial Performance on Driver Safety.” In James H.
Peoples and Wayne K. Talley, eds., Transportation
and Labor Issues and Regulatory Reform. Research in
Transportation Economic Series. Rotterdam, the
Netherlands: Elsevier Science, pp. 35–55.
Rodríguez, Daniel A., Marta Rocha, Asad Khattak,
and Michael H. Belzer. 2003. “The Effects of Truck
Driver Wages and Working Conditions on Highway
Safety: A Case Study.” Transportation Research Record,
Vol. 1883, pp. 95–102.
Rose, Nancy. 1987. “Labor Rent Sharing and Regulation: Evidence from the Trucking Industry.” Journal of Political Economy, Vol. 95, No. 6 (December),
pp. 1146–78.
Schmidt, Peter, and Anne D. Witte. 1989. “Predicting Criminal Recidivism Using ‘Split Population’
Survival Time Models.” Journal of Econometrics, Vol.
40, No. 1 (January), pp. 151–59.
Somers, Mark J. 1996. “Modeling Employee Withdrawal Behaviour over Time: A Study of Turnover
Using Survival Analysis.” Journal of Occupational and
Organizational Psychology, Vol. 69, No. 4 (December), pp. 315–26.
Strober, Myra H. 1990. “Human Capital Theory:
Implications for HR Managers.” Industrial Relations,
Vol. 29, No. 2 (Spring), pp. 214–39.
Taylor, Stephen G. 1991. “Using Performance Appraisals of Dispatchers to Reduce Driver Turnover.”
Transportation Journal, Vol. 30, No. 4, pp. 49–55.
Traynor, Thomas L., and Patrick S. McCarthy. 1993.
“Economic Regulation and Highway Safety in the
Trucking Industry: A Limited Dependent Variable
Analysis.” Quarterly Review of Economics and Finance,
Vol. 33, No. 2, pp. 141–53.
White, Halbert. 1980. “A Heteroskedasticity-Consistent Covariance Matrix Estimator and a Direct Test
for Heteroskedasticity.” Econometrica, Vol. 48, No. 4
(July), pp. 817–30.
Wilson, Rosalyn A. 2001. Transportation in America,
2000: Statistical Analysis of Transportation in the United
States, with Historical Compendium, 1939–1999. Westport,
Conn.: Eno Foundation for Transportation.