Seatbelt use among drunk drivers in different legislative settings.
Adams, Scott ; Cotti, Chad ; Tefft, Nathan 等
I. INTRODUCTION
The National Transportation Safety Board (NTSB) recently proposed
that all states move to a 0.05 maximum legal blood alcohol content (BAC)
threshold for drivers. The 0.05 threshold is a common limit across the
globe, and the NTSB predicts such a move would save many lives. (1) This
discussion will likely spark renewed interest in the impact of lower BAC
thresholds. Laboratory experimentation suggests a significant compromise
to competence for drivers above BAC levels of 0.08 (Moskowitz and
Fiorentino 2000). The number of lives saved, however, will not simply
rise from the reduced fatality risk between the 0.05 and 0.08 thresholds
implied by these experiments. Reductions in fatalities also depend on
how drivers react to the changing threshold by adjusting their behavior
and the impact of concurrent policies. For this reason, it is important
to understand these responses so policies can be coordinated most
effectively.
In this study, we explore one such driver behavior--seatbelt
use--in the case of lower BAC requirements being in effect along with
variably enforced seatbelt laws.
Specifically, we hypothesize that drivers adjust their seatbelt use
in light of lower BAC thresholds in cases where local seatbelt laws are
primarily enforced. Primary enforcement means that law enforcement
personnel can pull over a driver if he or a passenger is observed to be
not wearing a seatbelt. The mechanism is as follows. Drunk drivers
anticipate an increased likelihood of detection in such cases where they
are pulled over for an alternative infraction, for example, failure to
wear a seatbelt. Therefore, they avoid such interaction with law
enforcement by complying with seatbelts laws. Ample evidence already
suggests that there is a notable response to primary enforcement of
seatbelt laws compared with secondary enforcement.2 The penalties
associated with BAC laws are far stricter than seatbelt laws, however,
and are likely to interact with primary seatbelt laws to both increase
belt use and reduce the number of fatal accidents. To our knowledge,
this article is the first to bring together these two strands of the
literature on driver behavior in a systematic analysis of the
interactive effects of BAC and seatbelt laws.
Our analysis empirically confirms the hypotheses outlined above
using multiple tests that use several datasets. First, we appeal to the
crash-level microdata of the Fatality Analysis Reporting System (FARS),
which includes detailed information on every fatal accident in the
United States. We show that the interactive effect of a primarily
enforced seatbelt law with a stricter BAC threshold (0.08 vs. 0.10)
increases the likelihood that a driver in any fatal crash is wearing a
seatbelt at the time a fatal accident occurs. The effects are
concentrated among crashes that involve drivers with BAC levels at 0.08
or higher. We see no increase in seatbelt use, however, among sober
drivers. Again, inebriated drivers may be more concerned with being
pulled over, thus choosing to buckle up to avoid interaction with law
enforcement. This first analysis only observes drivers involved in a
fatal accident, however, which is a highly selected subsample. Moreover,
we expect to observe fewer fatal crashes in total if there is more
seatbelt use, so we also appeal to different tests and data to confirm
our finding. (3)
In a supporting set of tests, we appeal to survey data using the
Behavioral Risk Factor Surveillance System (BRFSS). In these data,
seatbelt use is measured over the previous 30 days, and we can test
whether there is any difference among drivers and passengers observed in
response to stricter BAC thresholds. The evidence again points to a
strong response in terms of general seatbelt use following BAC threshold
implementation in states with primary seatbelt enforcement. The effect
is concentrated among binge drinkers.
An implication of increased seatbelt use should be that lives are
saved, because the evidence that seatbelt use reduces fatalities is not
in dispute. In our final analysis, we again appeal to the FARS, this
time to verify that the stricter thresholds and primary enforcement of
seatbelt laws indeed interact to save lives. We show in states with
primarily enforced bans, coupled with a reduction in BAC from 0.10 to
0.08, there is a reduction in fatal accidents. This effect is
concentrated in the evening and nighttime hours when a much higher
percentage of drivers are potentially inebriated.
This article proceeds as follows: Section II provides background of
legislation regarding seatbelt use and BAC thresholds, including a
review of the literature on the relevant behavioral responses of
individuals; Section III describes our analysis of seatbelt usage, with
additional explanations provided for the two data sources used; Section
IV presents our analysis of the effects of the laws on fatalities; and
Section V concludes.
II. BACKGROUND
A. A Brief Review of Seatbelt and BAC Threshold Legislation in the
United States
The first meaningful legislation pertaining to seatbelts in the
United States was passed in 1961 in Wisconsin. All cars were required to
be equipped with front-seat restraints but no mandates on usage were
specified. Federal legislation followed in 1968, but as with the
Wisconsin legislation, no requirements on usage were established. The
first law that mandated that seatbelts actually be used was passed in
New York in 1984. Many states followed with similar legislation, and
currently only New Hampshire does not require seatbelt usage for adults.
(4)
Table 1 summarizes the current status of seatbelt legislation
across all states. With the exception of New Hampshire, legislation is
either enforced as a primary or secondary offense. Recall that primary
offense cases are those in which officers can pull a person over if the
driver or a passenger are suspected of not wearing a seatbelt. This is
where we would expect the interactive effects with concurrent BAC
legislation. (5) Although the fines tend not to be severe, being pulled
over for a primary offense opens up the possibility for the
investigation of other potential offenses, such as alcohol, drugs, and
weapons.
Legislation against driving while intoxicated (DWI) in the United
States has been in place for nearly as long as cars have been on the
road. However, specific definitions of intoxication and enforcement of
drunk driving laws emerged in the 1960s and continued over the next
several decades. Initial laws set BAC minimums at 0.15, but states soon
began establishing 0.08 or 0.10 as the more appropriate threshold.
Currently, all states have a 0.08 threshold in place. Although many
states conformed and codified this legislation voluntarily, legislation
passed at the Federal level in 2000 prompted the remainder of the states
to switch by 2004 or lose highway funds. Table 1 provides a list of
dates when states moved to the 0.08 level during the sample period.
B. Effects of Seat Belt Laws and BAC Threshold Laws
This study links two strands of literature on driver behavior. The
first is whether seatbelt laws encourage use and ultimately save lives.
The literature has reached a general consensus that it does. A
comprehensive review and meta-analysis was conducted over a decade ago
by Dinh-Zarr et al. (2001). With regard to fatal injuries, the median
finding across studies was a 9% reduction, and the range was 2%-18%.
Nonfatal injuries barely changed, with many studies showing an increase
in injuries. (6)
Dinh-Zarr et al. (2001) also reviewed a number of studies that
looked at the improvement of primarily enforced seatbelt laws over
secondarily enforced laws and found that the median incremental decrease
in fatal accidents was 8%. The specific study within this group that is
most relevant to ours is the study by Lange and Voas (1998). Although
they considered just two communities in California, Salinas, and
Oceanside, they were one of the few to assess seatbelt use among drunk
drivers specifically. Their method of inquiry was periodic roadside
surveys of drivers in these two communities, all of which were conducted
between 9 p.m. and 2 a.m. They find that nighttime seatbelt use
increased dramatically after California moved from secondarily enforced
bans to primarily enforced bans. Specifically, the jump was from 73% to
96% for all drivers in their survey. For drivers with BAC content over
the legal limit of 0.10 (at the time), the change was a much more
substantial increase from 53% to 92%.
There have been a few additional studies of seatbelt use that are
important for our purposes. First, Houston and Richardson (2006, 5) use
a panel data approach to states enacting seatbelt legislation and find
that seatbelt use increased by 9 percentage points more in states with
primary laws compared with states with secondary laws. The most relevant
study for our purposes is the study by Carpenter and Stehr (2008), which
offers comprehensive evidence of the effects of different types of
seatbelt laws on use and the incidence of fatal accidents. Their
findings confirm the general consensus in the literature that primarily
enforced seatbelt laws are more effective than secondarily enforced
bans. They also suggest that high-risk drivers (such as drinkers) are
more responsive to seatbelt laws. The importance of aggressive seatbelt
enforcement was recently highlighted by Luca (2014), who shows that a
"Click-It or Ticket" campaign has a notable influence on
traffic safety, particularly at night. The efficacy of such campaigns in
the evening hearkens back to the work of Lange and Voas, and is
consistent with our hypothesis.
Turning to the effectiveness of laws dictating lower BAC
thresholds, we see far more mixed results with regard to driver safety.
The General Accounting Office (GAO) (1999) concluded that there was not
enough evidence to conclude that BAC laws by themselves are effective at
reducing drunk driving fatalities. Without proper enforcement, public
education, and other drunk driving laws in place, the effect of BAC laws
themselves is unclear. (7) Recent research suggests even more limited
effects of lowering the BAC threshold by correcting for serial
correlation and the effects of the more recent states reducing their BAC
thresholds (Freeman 2007). Grant (2010) shows a declining effect of
several types of drunk driving legislation over time, including a 0.08
per se law. In short, the recent overall assessment of the .08 law is
that it is of limited effectiveness by itself. The current study adds to
this discussion by showing a possible additional reason for the mixed
evidence on the effectiveness of BAC laws.
Before describing our data, we point out two other recent studies
relevant to our analysis. Chang et al. (2012) estimates the effect of
seatbelt laws and lower BAC requirements, finding evidence consistent
with the former having a greater effect on fatalities. The interactive
effects are neither explored, however, nor is the difference between
primarily and secondarily enforced seatbelt laws. Carpenter and Harris
(2005) examine drinking behaviors following the move from 0.10 BAC
thresholds to 0.08 thresholds. Although they find some decline in
alcohol consumption following these laws, they see no change in the
likelihood of binge-drinking or alcohol-involved driving. Carpenter and
Harris (2005) suggest that the reduced alcohol consumption represents a
deterrent effect of laws, resulting in lower fatalities. Our evidence
offers an additional explanation. Even binge drinkers and
alcohol-impaired drivers are less likely to be involved in fatal
accidents because their seatbelt use protects them and reduces the
chance that any given crash results in a fatality.
III. SEATBELT USE ANALYSIS
A. Data and Methods
We utilize two data sources to assess the impact of seatbelt use
and BAC threshold legislation. The first is the FARS, which has the
benefit of being a census of nonself-reported data. It is also the most
reliable means to assess the ultimate impact of traffic legislation,
which is the number of lives saved. Our final results will verify the
interactive effect of seatbelt laws and BAC thresholds on fatal accident
totals, making this dataset of primary importance to this study. Given
our focus on seatbelt laws, we only use FARS accident data that involve
passenger vehicles (as defined by the National Highway Traffic Safety
Administration [NHTSA]), which include cars, light trucks, and vans. (8)
The first step in our inquiry is to ask whether actual seatbelt use
differs among those who are likely most responsive to BAC threshold
legislation in terms of their seatbelt use. Specifically, we look at
drivers who become affected by both lower BAC thresholds and live in a
state with primary seatbelt enforcement. Our data span every month from
1991 to 2010, which captures a substantial number of states switching
from BAC restrictions of 0.10 to 0.08. We estimate Equation (1) as a
probit model, with s and t indicating the state and month in which an
accident took place.
(1) SB[U.sub.ist] = [[alpha].sub.s] + [[tau].sub.t] +
[X.sub.ist][beta] + [delta]PB[L.sub.st] + [gamma][BA08.sub.st] +
[psi][BA08.sub.st] * PB[L.sub.ist] + [[epsilon].sub.ist].
Equation (1) is estimated using driver-level data, with the i
subscript indicating a driver, for a sample of drivers with a BAC of
0.08 or greater. The variable SBU is a dichotomous variable indicating
that a driver in a fatal accident uses a seatbelt. All regressions
include a series of state fixed effects and year-month fixed effects.
The former capture common differences in seatbelt use across states that
are fixed with respect to time, whereas the latter allows for a unique
intercept for seatbelt use for every time period in the sample. PBL is a
dummy variable indicating a state has a primarily enforced seatbelt law
in effect at the time of the accident. The BA08 dummy indicates a BAC
laws of 0.08 is in effect in the state, as opposed to a 0.10 law. The
interaction between the two and its corresponding coefficient estimate
[psi] is the effect of primary interest. (9)
In the X vector, we add a number of factors that might affect the
likelihood a driver wears a seatbelt. These include the driver's
age and gender, a dichotomous measure of "good" weather
conditions, (10) whether it was daytime (6:00 a.m. to 5:59 p.m.), and
whether the vehicle was a passenger car (vs. light truck or van).
Finally, we include a series of dummy variables to account for the speed
limit where the accident occurred. (11) Overall, the inclusion of these
factors will capture variation over time in driver or road
characteristics that may impact the likelihood of drivers in a
particular state wearing a seatbelt, such as age and gender composition,
weather conditions, or propensity for highway driving, for example.
Our difference-in-difference research design requires that the
control states serve as a valid control group to the treatment states.
Therefore, trends in seatbelt usage before the treatment occurs should
be similar in the treatment and control states. Of the 32 states that
are treated at some point during our sample time frame, 25 did not enter
the treatment group until after 2000. This 10-year period (where
approximately three-quarters of the treatment group had no change)
provides us with an opportunity to investigate if a difference in
general trends exists between the two groups. Specifically, we compare
the time trend in the seatbelt usage rate for drunk drivers before 2001
among the 25 states that did not become treated until after 2000, to the
trend for the 18 states that remain in the control group over the entire
1991-2010 period. Figure 1 shows that there is no discernible difference
in this unconditional trend between these two groups prior to 2001,
which helps to alleviate concerns that our research design may be
undermined by a difference in underlining trends between the treatment
and control states (the evident level difference will be absorbed by the
constant term in the regression models). That said, we more formally
investigate this potential concern later in this study.
[FIGURE 1 OMITTED]
Although Figure 1 is suggestive of no difference in seatbelt use
leading up to both laws being in effect, we recognize this is not proof
of policy exogeneity. Given the sheer volume of policies aimed at
traffic safety, we cannot rule out a concurrent policy change
confounding our results. We believe our results are immune from such
omitted variables biases. For another policy to fit our pattern of
results, it would have to be correlated with drunk driving deaths (but
not sober driving deaths), seatbelt use by heavy drinkers (but not
nondrinkers), seatbelt enforcement provisions, and stricter BAC
thresholds.
Figure 2 summarizes our research approach by presenting an
event-history representation of the patterns in average drunk driver
seatbelt usage rates in states that passed both stricter BAC thresholds
and primary enforced seatbelt laws (treatment states). We also
constructed a synthetic control group using a weighted average of the
drunk driver seatbelt usage rates in the control states. (12) Trends
have been indexed and smoothed for easier comparison. Acknowledging that
drivers are impacted by the policy changes at different times in
different states, the horizontal axis represents the number of quarters
before or after the policy interaction occurred. In the earliest
quarters prior to the treatment, there was a slight increase in seatbelt
usage in the treatment states over control states, but both exhibited
mild increases. However, in the periods around the treatment we observe
an upswing in seatbelt usage in the treatment states. The large increase
appears to begin slightly before the treatment, which is likely because
of our smoothing of the data. We find no suggestion of an upswing in the
control states in the post-treatment period. That said, Figure 2 is only
suggestive of an impact of the interaction of these laws, as it has no
controls for other potentially important influences on seatbelt usage
rates, including time, persistent location differences, and driver or
accident characteristics. Hence, to more carefully consider this
relationship, we return to our more complete modeling of driver-level
seatbelt usage presented in Equation (1) above.
[FIGURE 2 OMITTED]
B. Seatbelt Use Estimates Using the FARS
We report our formal estimation of Equation (1) via a probit model
in Table 2. Column (1) provides the result using only state- and
time-fixed effects (no other controls). The estimate of is 0.875 and is
statistically significant at a high level (p value = 0.007), which
suggests a strong interactive effect of BAC laws in the presence of
primarily enforced seatbelt legislation.
In the second column, we add controls for observed driver age and
gender, with little change in the parameter estimate [psi] (coef =
0.0897, p value = 0.004). Also, we observe, as expected, that older
drivers and women are more likely to be wearing seatbelts. In the third
column, we move to our preferred specification and add in the controls
the remaining controls included in vector X, which capture aspects about
the crash, weather, vehicle, and so on, that may impact an
individual's propensity to wear their seatbelt. The primary measure
of interest is now slightly smaller (coef =0.0805, p value = 0.014), but
remains basically unchanged. We also observe that drivers wear seatbelts
less in good weather and at night. The latter is likely because of the
greater number of risky drivers on the road at night. Finally, we find
that seatbelt use is highest on higher speed roads, for example,
highways. It is notable that the estimates of the control variables
behave as expected, given this is a sample of drivers involved in fatal
crashes and therefore suggests selection may not be of substantial
influence.
We calculate marginal effects of the policy variables and our
interaction of interest in the bottom panel of Table 2, and follow
suggestions by Greene (2010) for presenting interactive probit effects.
(13) These estimates provide strong evidence that when primary seatbelt
laws are in effect in states where there is a lower BAC threshold, there
is approximately a 2.6-2.8 percentage point increase in the likelihood
of a legally drunk driver wearing a seatbelt relative to locations
without a primary enforced seatbelt law or a higher BAC threshold. Given
that in the sample the average seatbelt use of drunk drivers involved in
a fatal accident is only 27%, this is a substantial effect. (14)
The results presented to this point suggest that drunk drivers do,
indeed, respond to the interaction of these policies by wearing the
seatbelts with greater frequency. However, given that the penalties
associated with repeat drunk driving offense are significantly greater,
we would expect drivers with a past drunk driving conviction to respond
even more to these policy incentives. The FARS data do allow for a
cursory investigation of the relationship, as information on whether or
not a driver had a past DWI conviction within the last 3 years is
available. Results of this brief extension of our primary analysis are
presented in the last two columns of Table 2, and they are consistent
with expectations. Specifically, while drunk drivers without a previous
DWI conviction responded similarly to the whole sample, those with a
recent DWI conviction responded much more strongly. This provides
further evidence of a behavioral response to the interaction between the
two laws.
Comparison with Estimates for Sober Drivers. Results in Table 2
indicate that the interaction between primary seatbelt laws and the
stricter drunk driving thresholds leads to a large and statistically
significant increase in seatbelt usage among drunk drivers. One may
worry, however, that the empirical connection is somehow driven by
uncaptured differences in trends in seatbelt usage between the treatment
and control locations. To allow for this possibility, we also estimate
Equation (1) for a sample of sober drivers (BAC = 0.00). Evidence from
sober individuals is potentially useful for two reasons. First, if the
hypothesis of increased seatbelt use among inebriated drivers in
response to stricter thresholds and primary belt enforcement is correct,
the estimate [psi] of should be zero in the sober sample and
significantly greater than zero in the intoxicated sample. Such an
outcome would leave us less worried about bias from unobservable
differences in trends affecting our estimates. Second, even if the
estimated effects for the sober sample are nonzero, the difference in
estimated effects across age groups (drunk minus sober) can be used as a
difference-in-difference-in-difference estimator of the effect of this
policy interaction on seatbelt usage among drunk drivers (removing any
bias from differences in trends). However, if estimates from the sober
population are similar to those found for drunk drivers, and no
significant difference is detected, it would cast doubt on the validity
of the earlier results in Table 2.
Table 3 presents the results of this analysis, with the results for
drunk drivers duplicated from column (3) from Table 2 to provide easy
comparison. Results for the sober drivers, presented in column (2) of
Table 3, indicate that there is no impact of the interactive effect of
lower BAC thresholds and primary seatbelt laws (p value = 0.755). The
primary belt law itself provides the only effect, which is consistent
with expectations. We note that the combined effect of primary seatbelt
laws ([delta] + [psi]) in the inebriated sample exceeds that in the
sober sample, which furthers confirms our hypothesis that drunk drivers
are avoiding interaction with police in the primary belt law states.
While the difference between these estimates is consistent with our
hypothesis and suggests that underlining trends are not an issue, we
formalize our results in column (3) by explicitly adding a triple
interaction between our policy variables and a drunk driver indicator
(DWI). Specifically, we will also estimate Equation (2), which is
similar to Equation (1) except it fully interacts the specification with
DWI, an indicator set equal to one if the driver's BAC is greater
than or equal to 0.08 and set equal to zero if his BAC is 0 (drivers
with BAC between 0 and 0.08 are excluded from this estimation).
(2) [MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII].
The triple interaction between both policies and the driver
intoxication dummy (DWI) and its corresponding coefficient
([[PSI].sub.4]) is the primary measure of interest in Equation (2).
As indicated from the outcomes in columns (1) and (2), the triple
interaction presented in column (3) tells a consistent story, suggesting
approximately a 3 percentage point increase in seatbelt use when both a
BAC 0.08 threshold and primary seatbelt law are in place. This estimate
is significant as the p value equals 0.041.
Lastly, while this analysis conducted in Table 3 strongly
alleviates concerns that the control state trends differ over time (a
fundamental violation of the difference-in-difference identification),
it is possible that differences in seatbelt usage trends may only exist
among drunk drivers prior to the passage of the policy interaction of
interest. While this is not supported by the trends shown in Figure I.
as a last investigation of trend differences, we explicitly control for
differences in time trends between the treatment and control locations
in our full-specification of Equations (1). Results are consistent and
robust (Coef = 0.0826, p value = 0.028).
C. Seatbelt Use Estimates Using the BRFSS
While the results presented in Section III.B indicate increased
seatbelt use by those drivers who fear being pulled over, there is an
important caveat to the Table 2 estimates just described. Specifically,
the FARS data come from accidents with a fatality. If seatbelts result
in fewer deaths, then it might be the case that we are observing a
select group of more severe accidents. It is possible that many of these
missing observations are related to increased seatbelt use among drivers
wishing to avoid detection. If so, the influence of selection bias
should understate any measured effect. Therefore, we also appeal to the
BRFSS, which includes nationwide data on seatbelt use for samples large
enough to make inferences at the state level. It also includes a myriad
of additional information, including frequency of drinking, to
investigate whether similar patterns can be identified. The BRFSS has
the important disadvantage of not recording whether the affected
seatbelt users were actually making simultaneous drinking, driving, and
seatbelt decisions, which is a meaningful limitation, but it
nevertheless confirms the FARS finding in a broader sample and acts as a
further robustness investigation.
The BRFSS estimation proceeds by estimating Equations (1) and (2)
in a similar fashion, with a few notable differences. Equations (1) and
(2) are ordered probit equations because the SBU variable in the BRFSS
is an ordered measure. Specifically, seatbelt usage is self-reported in
response to the question "How often do you use seat belts when you
drive or ride in a car? Would you say ..." as one of the following
frequencies: "Always," "Nearly always,"
"Sometimes," "Seldom," and "Never." Not
only does it utilize more information, but the ordered probit model is
perhaps more telling of the seatbelt user behavior if the driver is
marginally affected by the change in BAC threshold. That is, he may only
buckle up when he fears he is within the threshold. Among binge
drinkers, this may lead to the driver moving along an ordered scale of
seatbelt use frequency.
Another difference is the triple interaction indicator (DWI) used
in Equation (2). The BRFSS asks about alcohol consumption in the 30 days
prior to the interview. Hence, the DWI indicator is exchanged with a
BINGE variable, which represents the BRFSS survey question that asks
whether an individual had five or more drinks in any one sitting in the
past 30 days. We suspect that this common definition of binge drinking
would identify those most affected by the law.
Lastly, we are able to include a richer set of demographic controls
in vector X for this analysis, as the BRFSS data provide significant
detail on individual characteristics (i.e., employment status,
education, income, etc.). Table A2 (Appendix) summarizes the BRFSS
sample used in the analysis.
Within the timeframe of the FARS analysis (1990-2010), the BRFSS
seatbelt use question was asked annually and then later biennially, so
we draw on the 1990-1998, 2002, 2006, 2008, and 2010 BRFSS survey waves.
A total of 2,063,436 participants responded to the seatbelt survey
question in these survey waves. After dropping observations with missing
demographic, income, or employment data, the final sample consists of
1,709,344 (approximately 17% of the sample is dropped).
Table 4 presents the results, where the dependent variable is the
frequency scale-dependent variable of seatbelt use described above. Only
the primary policy effects and their interactions are presented in this
table, but the coefficients for the full set of controls are reported in
Table A3. In columns (1) and (2), we test the simple interactive result
of a lower BAC threshold and a primarily enforced ban separately for
respondents reporting a binge drinking event or no drinking events in
the previous 30 days. Thus, these two columns are estimated with a
similar goal to the first two columns of the FARS estimates in Table 3.
However, we do not know each respondent's BAC.
Evident from the estimates of [PSI] in columns (1), the interactive
effect of a primarily enforced ban and a lower BAC on binge drinkers is
large and positive, which is consistent with the FARS analysis. That
said, the results are not statistically significant. Unlike the FARS,
the BRFSS does not directly link seatbelt use and drinking activity,
which may reduce precision.
We next combine the samples of binge and nondrinkers and fully
interact an indicator of binge drinking in the previous 30 days in order
to formally test the difference between the two groups, as shown in
column (3). The first variable, entered in column (3), is an indicator
of whether one binge drank in the last 30 days. Not surprisingly, binge
drinkers are less likely to wear a seatbelt. Also, the triple
interaction of binge drinking with both primary belt enforcement and a
lower BAC threshold is positive and statistically significant (coef =
0.0569, p value = 0.042). Taken collectively, these outcomes are
consistent with the FARS results, as they demonstrate increased seatbelt
use among drivers who would most like to avoid interaction with law
enforcement.
Cross-partial effects, which can be interpreted as the difference
in the probability of seatbelt use when reducing the BAC threshold to
0.08 under primary enforcement compared with reducing the BAC threshold
similarly when not under primary enforcement, were also calculated using
the estimated model coefficients and are reported in Table 5. The
cross-partial effects of the interactions, or the double differences
when comparing across the two policies, show a consistent shift away
from reporting partial or no seatbelt use to always seatbelt use. This
effect is also stronger among binge drinkers, reflecting the
cross-partial effects for the FARS analysis presented in the bottom
portion of Table 3.
IV. ACCIDENT ANALYSIS
A. Data and Methods
Using two very different data sources, we have established that
seatbelt use increases when BAC thresholds are lower in the presence of
primarily enforced seatbelt laws. Effects are only observed for
drinkers, as expected. The implication is that lives will likely be
saved by this change in behavior. This leads to an extension of our
investigation--that is, to verify whether when both laws are in place
there are indeed fewer fatalities on the roadways. That seatbelts save
lives is not controversial and the evidence cited earlier attests to
this. We do have several reasons to be cautious about analyzing
fatalities. First, the Peltzman effect suggests that drivers might be
more reckless when wearing a seatbelt. This would put upward pressure on
our results. Second, analyzing BAC law effects using FARS has not
yielded consistent results in past studies.
For these estimates, we return to the FARS data and aggregate fatal
passenger vehicle accidents at the state level. The following equation
assesses the relevant impact on such accidents:
(3) [MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII].
The log total of passenger-vehicle accidents in each state-month
cell is estimated as a function of the same set of variables used
earlier in the FARS. Estimations are weighted by the number of accidents
to limit the undue influence of low-frequency, high variance states.
(15) All standard errors are clustered at the state level to allow for
nonindependence of observation from the same location over time
(Bertrand et al. 2004).
The X vector in these specifications also includes controls for
population, which varies annually rather than monthly. As state-fixed
effects capture the general size of the state, population reflects a
change in relative density in this context. We round out the X vector
with beer tax and unemployment rate controls. The former has been shown
to reduce drunk driving accidents. (16) The latter represents depressed
economic activity, which is likely associated with decreased traffic and
accidents (Cotti and Tefft 2011).
Equation (3) results measure the various effects of legislation
along with the incremental effect of a lower BAC threshold enacted in
the presence a primary seatbelt law. Given that we are interested in the
total number of alcohol-related accidents in each state and year, we can
no longer exclude accidents where the BAC levels are missing. Instead,
we use NFITSA-imputed BAC for all drivers where BAC is missing. As
seatbelt use is a factor used by the NHTSA in imputing BAC levels,
comparable estimates that divide the sample using BAC level (as was
carried out in Table 2) are not appropriate for this analysis. We
instead divide the sample into nighttime accidents (6:00 p.m. to 5:59
a.m.) and daytime accidents (6:00 a.m. to 5:59 p.m.). (17) We would
expect a change in nighttime accidents, as this is the time period when
the proportion of drivers that are intoxicated is highest. (18)
Moreover, police presence to deter drunk driving is more prevalent in
the evening. We would therefore also expect drivers to react more to
legislation at night.
B. Fatal Accident Estimates Using the FARS
The estimates in Table 6 support that nighttime accidents are most
affected by the interactive effects of the legislation. There is a
nearly 7% reduction in fatal accidents in states with a lower BAC
threshold and a primarily enforced seatbelt law compared with a state
without both laws in place (i.e., only one or the other are in place).
(19) Given that the average number of fatal nighttime passenger vehicle
accidents per month in a typical state is approximately 27.6 in the
sample, a 7% reduction translates into roughly 2 fewer fatal accidents
per state per month, or over 1,200 annually across the country. In fact,
these estimates also suggest that there is no discernible reduction in
fatal nighttime accidents unless both a primarily enforced seatbelt law
and lower BAC threshold are in place, rather than with each law in place
on its own.
In column (2) of Table 5 we contrast our nighttime findings to
analogous estimates from daytime accidents. While daytime accidents are
by no means exclusively nonalcohol related, there is a significant
contrast based on time of day, so we should see a much smaller or
nonexistent impact of the interaction on fatal accidents if our
hypothesis is correct. Results are consistent with this expectation, as
estimates of the interaction on outcomes are now half as large and no
longer statistically significant. A conservative reading is that the
negative estimate of [PSI] observed in column (2) demonstrates the
presence of an underlying difference in unobservables between treatment
and control states in general accident rates that is correlated with the
interaction of both policies in question. Consequently, if we difference
the estimates between nighttime and daytime accidents we would
"net" a negative effect of around a 3%-4% decline in
accidents. Due to the imprecision of the estimates, results are not
statistically different from one another, so we are unable to completely
eliminate this concern.
Among our other determinants of fatal accidents, it is worth noting
that only beer taxes have a measureable impact on nighttime accidents,
again consistent with drunk driving being more prevalent in the
nighttime hours. We detect no effect of the state-level unemployment
rate on nighttime accidents but a small decrease in daytime accidents.
We also note the rather large positive effect of the lower BAC
thresholds on fatal accidents in both the day and night. One possible
explanation is that BAC laws were passed in response to increasing
accident rates. That said, the interactive effect is highly consistent
with what was found when analyzing seatbelt use, which was the point of
the fatality analysis.
V. CONCLUSION
This study is the first to assess whether the effectiveness of
stricter BAC thresholds is influenced by another concurrent policy,
namely seatbelt law enforcement. We suspected that primarily enforced
laws raise the probability that a drunk driver would be detected by law
enforcement. For this reason, we should observe heightened seatbelt use
among drunk drivers.
Our data suggest that, indeed, seatbelt use increases among
potentially inebriated drivers when BAC thresholds are set at .08 rather
than .10 but only in areas with primarily enforced seatbelt laws. This
is true in a census of fatal accidents. Drivers with a high BAC in such
accidents are more likely to be wearing their seatbelt than in cases
where there is no primary enforcement. This relationship is not observed
in a sample of crashes where no driver was inebriated. We also look at
the BRFSS, which retrospectively asks drivers about their seatbelt use.
Among binge drinkers, frequency of seatbelt use increases when lower BAC
thresholds and primarily enforced seatbelt laws are jointly effective.
Although the increased seatbelt use is expected to save lives, we
also investigate the question of whether the interactive effect of these
policies is lowering traffic fatalities compared with either passed in
isolation. We find that the laws, if enacted together, reduce accidents
by 7% compared to cases with a secondary enforcement or a higher BAC
level. This provides a direct policy prescription when enacting stricter
BAC thresholds. These laws should also be passed in an environment with
strict seatbelt provisions if the intention is to maximize lives saved.
We note that the observed reduction in fatal accidents is likely among
drivers and their passengers, rather than other noninebriated drivers
and pedestrians.
This means that drunk driving externalities, which are typically
limited to lives lost outside of the car of the inebriated driver
(Levitt and Porter 2001) are not substantially reduced. Nevertheless,
the reduction in fatalities still is economically meaningful if one
considers the contribution to society that can be made by those lives
saved.
The implications of the results of this study are far-reaching.
Evidence clearly supports that drivers adjust their behavior given the
combination of regulations in place and their own personal
circumstances. They are more selective in following laws if they face a
higher cost of not complying because they fear interaction with law
enforcement officials. This study suggests several additional lines of
inquiry. In the case of seatbelt laws and BAC legislation, the net
effect is positive in terms of lives saved. This need not be the case.
If seatbelt use leads to some illegal behavior not being detected,
perhaps in the case of drug possession or weapons infractions, the net
effect may not be positive. We anticipate more findings of the
interactive effect of seatbelt law enforcement with other policies. We
also suggest that future work on traffic safety should more generally
look at the interactive effects of policies.
ABBREVIATIONS
BAC: Blood Alcohol Content
BRFSS: Behavioral Risk Factor Surveillance System
DWI: Driving While Intoxicated
FARS: Fatality Analysis Reporting System
NHTSA: National Highway Traffic Safety Administration
NTSB: National Transportation Safety Board
doi: 10.1111/ecin.12155
APPENDIX
TABLE A1
Mean and Proportion, Fatality Analysis Reporting System, 1991-2010
Full Sample Neither Law BAC08 Only
Panel A: Drivers BAC = 0.08+
Seatbelt used 0.265 0.174 0.235
Age 33.55 33.21 34.03
Male 0.837 0.842 0.828
Good weather 0.892 0.879 0.899
Daytime 0.185 0.183 0.187
Vehicle type: car 0.579 0.622 0.565
Speed limit < 30 0.039 0.042 0.037
Speed limit 30-39 0.156 0.151 0.158
Speed limit 40-49 0.196 0.186 0.219
Speed limit 50-59 0.469 0.541 0.430
Speed limit 60+ 0.140 0.080 0.155
Observations 125.983 45.506 32.417
Panel B: Drivers BAC = 0.00
Seatbelt used 0.586 0.500 0.556
Age 41.60 41.27 42.40
Male 0.656 0.652 0.652
Good weather 0.849 0.828 0.857
Daytime 0.640 0.658 0.647
Vehicle type: car 0.612 0.659 0.592
Speed limit < 30 0.029 0.029 0.026
Speed limit 30-39 0.115 0.113 0.113
Speed limit 40-49 0.190 0.182 0.202
Speed limit 50-59 0.477 0.549 0.434
Speed limit 60+ 0.189 0.128 0.223
Observations 211,571 71,453 56.265
PBL Only Both Laws
Panel A: Drivers BAC = 0.08+
Seatbelt used 0.286 0.406
Age 32.93 33.80
Male 0.846 0.836
Good weather 0.895 0.903
Daytime 0.164 0.193
Vehicle type: car 0.578 0.534
Speed limit < 30 0.034 0.041
Speed limit 30-39 0.174 0.154
Speed limit 40-49 0.171 0.197
Speed limit 50-59 0.484 0.402
Speed limit 60+ 0.136 0.206
Observations 13,943 34.117
Panel B: Drivers BAC = 0.00
Seatbelt used 0.631 0.708
Age 40.64 41.60
Male 0.662 0.661
Good weather 0.839 0.871
Daytime 0.630 0.615
Vehicle type: car 0.617 0.575
Speed limit < 30 0.029 0.032
Speed limit 30-39 0.123 0.115
Speed limit 40-49 0.176 0.194
Speed limit 50-59 0.501 0.424
Speed limit 60+ 0.170 0.235
Observations 22,640 61,213
TABLE A2
Summary Statistics, Behavioral Risk Factor Surveillance System
Mean/Proportion
Full Sample Neither Law
Uses seatbelt always 0.761 0.615
Uses seatbelt nearly always 0.123 0.169
Uses seatbelt sometimes 0.057 0.103
Uses seatbelt seldom 0.028 0.054
Uses seatbelt never 0.031 0.059
Any drinks in last 30 days 0.505 0.500
Any binge drinking events in last 30 days 0.126 0.136
Age 50.615 45.638
Male 0.404 0.423
White 0.858 0.870
Black 0.083 0.095
Hispanic 0.055 0.034
High-school grad 0.308 0.340
Some college 0.272 0.268
College grad 0.317 0.251
Married 0.562 0.551
Current smoker 0.200 0.240
Income $10k to $15k 0.071 0.095
Income $ 15k to $20k 0.086 0.104
Income $20k to $25k 0.104 0.116
Income $25k to $35k 0.145 0.177
Income $35k to $50k 0.169 0.182
Income >$50k 0.351 0.210
Employed for wages 0.501 0.555
Self-employed 0.089 0.083
Out of work for > 1 year 0.020 0.019
Out work for < 1 year 0.025 0.024
Homemaker 0.074 0.082
Student 0.022 0.032
Retired 0.220 0.184
Unable to work 0.047 0.020
Observations 1709344 434583
Mean/Proportion
BAC08 Only PBL Only
Uses seatbelt always 0.746 0.749
Uses seatbelt nearly always 0.139 0.130
Uses seatbelt sometimes 0.058 0.061
Uses seatbelt seldom 0.027 0.030
Uses seatbelt never 0.031 0.030
Any drinks in last 30 days 0.513 0.546
Any binge drinking events in last 30 days 0.127 0.145
Age 52.148 45.501
Male 0.405 0.421
White 0.900 0.807
Black 0.049 0.100
Hispanic 0.046 0.088
High-school grad 0.310 0.314
Some college 0.275 0.274
College grad 0.330 0.300
Married 0.579 0.551
Current smoker 0.189 0.224
Income $10k to $15k 0.063 0.075
Income $ 15k to $20k 0.081 0.089
Income $20k to $25k 0.102 0.110
Income $25k to $35k 0.139 0.170
Income $35k to $50k 0.172 0.188
Income >$50k 0.390 0.284
Employed for wages 0.497 0.565
Self-employed 0.097 0.087
Out of work for > 1 year 0.018 0.020
Out work for < 1 year 0.023 0.026
Homemaker 0.071 0.076
Student 0.017 0.035
Retired 0.225 0.169
Unable to work 0.053 0.023
Observations 564517 84176
Mean/Proportion
Both Laws Min Max
Uses seatbelt always 0.879 0 1
Uses seatbelt nearly always 0.076 0 1
Uses seatbelt sometimes 0.023 0 1
Uses seatbelt seldom 0.010 0 1
Uses seatbelt never 0.012 0 1
Any drinks in last 30 days 0.496 0 1
Any binge drinking events in last 30 days 0.116 0 1
Age 53.374 18 99
Male 0.388 0 1
White 0.818 0 1
Black 0.104 0 1
Hispanic 0.073 0 1
High-school grad 0.283 0 1
Some college 0.271 0 1
College grad 0.352 0 1
Married 0.556 0 1
Current smoker 0.178 0 1
Income $10k to $15k 0.063 0 1
Income $ 15k to $20k 0.078 0 1
Income $20k to $25k 0.096 0 1
Income $25k to $35k 0.125 0 1
Income $35k to $50k 0.156 0 1
Income >$50k 0.423 0 1
Employed for wages 0.460 0 1
Self-employed 0.086 0 1
Out of work for > 1 year 0.024 0 1
Out work for < 1 year 0.026 0 1
Homemaker 0.072 0 1
Student 0.018 0 1
Retired 0.249 0 1
Unable to work 0.065 0 1
Observations 626068
Notes: Summary of observations without nonresponses, as described in
the text. The sample consists of BRFSS waves in which respondents
reported seatbelt use (1990-1998, 2002, 2006, 2008, 2010).
TABLE A3
BAC 0.08 and Primary Enforcement Law Effects, BRFSS Seatbelt Use
Ordered Probits (All Coefficients)
(1) (2)
Any Binge Drinking No Drinking
BA08 -0.0264 (0.0311) -0.0115 (0.0365)
PBL 0.1213 ** (0.0565) 0.1162 ** (0.0454)
BA08 and PBL 0.0713 (0.0560) 0.0177 (0.0503)
Age 0.0015 (0.0012) 0.0209 *** (0.0015)
Age squared 0.0000 (0.0000) -0.0002 *** (0.0000)
Male -0.3424 *** (0.0064) -0.3857 *** (0.0108)
White -0.0499 *** (0.0168) -0.0442 ** (0.0225)
Black -0.0650 *** (0.0187) -0.0353 (0.0288)
Hispanic 0.1327 *** (0.0166) 0.1232 *** (0.0191)
High-school grad 0.0971 *** (0.0078) 0.1184 *** (0.0111)
Some college 0.1872 *** (0.0117) 0.2809 *** (0.0149)
College grad 0.3341 *** (0.0157) 0.4945 *** (0.0205)
Married 0.0761 *** (0.0060) 0.1380 *** (0.0073)
Current smoker -0.1729 *** (0.0104) -0.1581 *** (0.0121)
Income $ 10k to $15k 0.0562 *** (0.0065) 0.0086 (0.0169)
Income $ 15k to $20k 0.0863 *** (0.0073) 0.0290 * (0.0161)
Income $20k to $25k 0.1003 *** (0.0096) 0.0371 ** (0.0147)
Income $25k to $35k 0.1048 *** (0.0099) 0.0416 *** (0.0156)
Income $35k to $50k 0.1268 *** (0.0107) 0.0741 *** (0.0134)
Income >$50k 0.1733 *** (0.0125) 0.1143 *** (0.0175)
Self-employed -0.2216 *** (0.0084) -0.2813 *** (0.0131)
Out of work for > 1 year -0.0462 *** (0.0103) -0.0861 *** (0.0173)
Out work for < 1 year -0.0664 *** (0.0110) -0.0555 *** (0.0136)
Homemaker 0.0499 *** (0.0083) 0.0716 *** (0.0152)
Student 0.0773 *** (0.0136) 0.1162 *** (0.0203)
Retired 0.0889 *** (0.0064) 0.0578 *** (0.0148)
Unable to work 0.0242 * (0.0137) -0.0433 (0.0271)
Observations 845,860 215,364
Notes: All results are model coefficients and hypothesis tests (not
partial effects). The sample consists of BRFSS waves in which
respondents reported seatbelt use (1990-1998, 2002, 2006, 2008,
2010). Each column represents a separate regression. All models
include indicators for year-month and state of residence. Robust
standard errors clustered by state of residence are in parentheses.
*** p < 0.01; ** p < 0.05; * p < 0.1.
REFERENCES
Adams, S., M. K. L. Blackburn, and C. D. Cotti. "Minimum Wages
and Alcohol-related Traffic Fatalities among Teens." Review of
Economics and Statistics, 94(3), 2012, 828-40.
Ai, C., and E. Norton. "Interaction Terms in Logit and Probit
Models." Economics Letters, 80, 2003, 123-29.
Ayers, I., and S. Levitt. "Measuring Positive Externalities
from Unobservable Victim Precaution: An Empirical Analysis of
Lojack." Quarterly Journal of Economics, 113(1), 1998, 43-77.
Bertrand, M., E. Duflo, and S. Mullainathan. "How Much Should
We Trust Differences-in-Differences Estimates?" Quarterly Journal
of Economics, 119(1), 2004, 249-75.
Carpenter, C. S., and K. Harris. "How Do .08 BAC Laws
Work?" The B E. Journal of Economic Analysis and Policy, 5(1),
2005, 1-18.
Carpenter, C. S., and M. Stehr. "The Effects of Mandatory
Seatbelt Laws on Seatbelt Use, Motor Vehicle Fatalities, and
Crash-related Injuries among Youths." Journal of Health Economics,
27(3), 2008, 642-62.
Chang, K., W. Chin-Chih, and Y.-H. Ying. "The Effectiveness of
Alcohol Control Policies on Alcohol-related Traffic Fatalities in the
United States." Accident Analysis and Prevention, 45, 2012, 406-15.
Cotti, C., and N. Tefft. "Decomposing the Relationship between
Macroeconomic Conditions and Fatal Car Crashes during the Great
Recession: Alcohol- and Non-alcohol-related Accidents." The B.E.
Journal of Economic Analysis & Policy, 11 (1), 2011.
Dee, T. "Does Setting Limits Save Lives? The Case of 0.08 BAC
Laws." Journal of Policy Analysis and Management, 20(1), 2001,
111-28.
Dinh-Zarr, T. B., D. A. Sleet, R. A. Shults, S. Zaza, R. W. Elder,
J. L. Nichols, R. S. Thompson, and D. M. Sosin. "Reviews of
Evidence Regarding Strategies to Increase the Use of Safety Belts."
American Journal of Preventive Medicine, 21(4S), 2001, 48-65.
Eisenberg, D. "Evaluating the Effectiveness of Policies
Related to Drunk Driving." Journal of Policy Analysis and
Management, 22(2), 2003, 249-74.
Freeman, D. G. "Drunk Driving Legislation and Traffic
Fatalities: New Evidence on BAC 08 Laws." Contemporary Economic
Policy, 25(3), 2007, 293-308.
General Accounting Office (GAO). Report to Congressional Committees
on Highway Safety: Effectiveness of State .08 Blood Alcohol Laws
(GAO/RCED-99-179). Washington, DC: GAO, 1999.
Grant, D. "Policy Analysis and Policy Adoption: A Study of
Three National Drunk Driving Initiatives." Manuscript, Sam Houston
State University, 2010.
Greene, W. "Testing Hypotheses about Interaction Terms in
Nonlinear Models." Economics Letters, 107(2), 2010, 291-96.
Houston, D. J., and L. E. Richardson Jr., "Reducing Traffic
Fatalities in the American States by Upgrading Seat Belt Use Laws to
Primary Enforcement." Journal of Policy Analysis and Management,
25(3), 2006, 645-59.
Lange, J. E., and R. B. Voas. "Nighttime Observations of
Safety Belt Use: An Evaluation of California's Primary Law."
American Journal of Public Health, 88(11), 1998, 1718-20.
Levitt, S., and J. Porter. "How Dangerous Are Drinking
Drivers?" Journal of Political Economy, 109(6). 2001, 1198-237.
Luca. D. L. "Do Traffic Tickets Reduce Motor Vehicle
Accidents? Evidence from a Natural Experiment." Journal of Policy
Analysis and Management 2014. Forthcoming, DOI:10.1002/pam.21798.
Moskowitz, H., and D. Fiorentino. "A Review of the Literature
on the Effects of Low Doses of Alcohol on Driving-related Skills."
Report No. DOT HS 809 028. Washington, DC: National Highway Traffic
Safety Administration, 2000.
Peltzman, S. "The Effects of Automobile Safety
Regulation." Journal of Political Economy, 83(4), 1975, 677-725.
Ruhm, C. "Alcohol Policies and Highway Vehicle
Fatalities." Journal of Health Economics, 15(4), 1996, 435-54.
(1.) See http://www.ntsb.gov/news/2013/130514.html (last accessed
October 10, 2013).
(2.) Dinh-Zarr et al. (2001) reviews the early evidence.
(3.) Selection bias likely understates any measured effect, as an
increase in seatbelt usage should reduce fatalities, and bias us from
observing these individuals. Note that not all drivers involved in the
accidents died, but that the accident involved at least one fatality.
(4.) See http://www.iihs.org/iihs/topics/laws/safetybeltuse
?topicName=safety-belts (last accessed October 10, 2013) for a
chronology of such state legislation.
(5.) There are also modest differences in age groups covered and
fines. Given the penalties of drunk driving are much more severe, we
view these minor differences as not meaningful sources of variation in
this article.
(6.) The mixed result is perhaps due in part to injuries among
survivors of crashes who would have died had they not been wearing their
seatbelt. Alternatively, passage of seatbelt laws may have led to a
Peltzman effect, where drivers are more reckless with the increased
perceived safety of seatbelts (Peltzman 1975).
(7.) However, Dee (2001) used the time and space variation afforded
by the imposition of various state BAC thresholds to show a significant
reduction in fatalities stemming from both the 0.10 and the 0.08
threshold. Eisenberg (2003) makes the salient observation that the
differential between these two estimates is most relevant since it
represents the marginal changes most often observed in recent
legislative actions. He indicates the Dee estimates imply a mere 2%
reduction in accidents stemming from a reduced BAC threshold from 0.10
to 0.08. Eisenberg's own estimates imply a slightly larger impact
of BAC threshold.
(8.) The vast majority of the non-passenger vehicle-accidents
observations excluded are those involving motorcycles.
(9.) BAC levels are not available for all fatal crashes in the FARS
even though such measures are required by law. We remove those cases
where BAC levels are not directly measured.
(10.) NHTSA classifies the prevailing atmospheric conditions that
existed at the time of the crash as recorded on the crash report form.
For our purposes, we define the "good" weather condition dummy
equal to one if NHSTA reports the weather as "clear" and zero
if the weather condition is classified otherwise, which includes rain,
hail, snow, fog, and so on.
(11.) Summary statistics for the FARS data used in Equation (1) is
provided in Table Al.
(12.) In this synthetic control group, the calendar quarters for
the control group are weighted to match the calendar quarters
represented in the corresponding lead or lag period for the treated
group (as set out in Ayers and Levitt 1998, and Adams et al. 2012).
Control groups are New Hampshire, states prior to their enactment of
primarily enforced belt laws, and states before the move to a 0.08
threshold.
(13.) These effects were calculated using the "margins
r.policy1, over(r.policy2)" command in Stata (v12). Note that these
are not naive marginal effects that would ignore the calculated
interaction coefficients, but are the first derivative of the
conditional mean with respect to the policy of interest (Ai and Norton
2003; Greene 2010).
(14.) Results are robust to restricting the sample to accidents
involving a single car only (coef = 0.0959, p-value < .01), or
restricting the sample to accidents involving drivers 21 years old or
more (coef = 0.0722, p value = 0.031).
(15.) While a weighted least squares approach is our preferred
method, we confirmed that the results are robust to other
specifications, such as OLS, Poisson, and Negative Binomial models.
(16.) Ruhm (1996) finds beer taxes to be effective in deter ring
drunk driving for at least a subset of the population, while Eisenberg
(2003), however, finds limited evidence of the effect of beer taxes.
(17.) This is the formal nighttime vs. daytime convention used by
the NHSTA when classifying accidents in the FARS data. We further
restrict the daytime accidents to only weekdays (Monday-Friday) to
further isolate the hours of the week when drunk driving is least
prevalent form this sample. If every state-month period (day or night)
had a positive number of fatal accidents the sample size would be 12,
240 (51*20*12). Our estimates will have slightly fewer than this total,
as not all locations have at least one fatal accident every month
"day" or "night."
(18.) This approach is consistent with much of the literature that
studies drunk driving accidents (e.g., Eisenberg 2003; Ruhm 1996).
Notably, the BAC measures in the FARS data (both imputed and not)
suggest that the proportion of nighttime accidents that involve alcohol
is approximately 50%, whereas during daytime on weekdays the number is
approximately 6%, suggesting that our bifurcation of the data is
consistent with expectations and suitable for this investigation.
(19.) For robustness purposes, we re-estimate these results
isolating alcohol-related accidents, as classified by the BAC measures
in the data, including the imputed alcohol data files from NHTSA. While
this is likely inappropriate given the nature of the imputation process
and the focus of our study on seatbelt laws, results are not sensitive
to this approach.
SCOTT ADAMS, CHAD COTTI and NATHAN TEFF *
* We thank seminar participants at Yale University, Potsdam
University, the Southern Economics Association Meetings, the American
Society of Health Economics Meetings, as well as McKinley Blackburn. Ken
Couch, John Heywood, and Owen Thompson for helpful comments
Adams: Department of Economics, University of Wisconsin-Milwaukee,
Milwaukee, WI 53201. Phone 414-229-4212, Fax 414-229-1122, E-mail
sjadams@uwm.edu
Cotti: Department of Economics, University of Wisconsin-Oshkosh,
Oshkosh, WI 54901. Phone 920-203-4660, Fax 920-424-7413, E-mail
cottic@uwosh.edu
Teffi: Department of Economics, Bates College, Lewiston, ME. Phone
207-786-6056, Fax 207-786-8337, E-mail tefft@uw.edu
TABLE 1
Primary Seatbelt Law and BAC Transitions
Date of
Any Belt Law Belt Law: Primary Transition
State in Effect Enforcement to BAC 0.08
AL 7/18/1991 Yes; effective 12/09/99 10/1995
AK 9/12/1990 Yes; effective 05/01/06 9/2001
AZ 1/1/1991 No 9/2001
AR 7/15/1991 Yes, effective 06/30/09 8/2001
CA 1/1/1986 Yes; effective 01/01/93 1/1990
CO 7/1/1987 No 7/2004
CT 1/1/1986 Yes; effective 01/01/86 7/2002
DE 1/1/1992 Yes; effective 06/30/03 7/2004
DC 12/12/1985 Yes; effective 10/01/97 4/1999
FL 7/1/1986 Yes; effective 6/30/09 1/1994
GA 9/1/1988 Yes; effective 07/01/96 7/2001
HI 12/16/1985 Yes; effective 12/16/85 7/1995
ID 7/1/1986 No 7/1997
IL 1/1/1988 Yes; effective 07/03/03 7/1997
IN 7/1/1987 Yes; effective 07/01/98 7/2001
IA 7/1/1986 Yes; effective 07/01/86 7/2003
KS 7/1/1986 Yes; effective 6/10/10 7/1993
KY 7/15/1994 Yes; effective 07/20/06 10/2000
LA 7/1/1986 Yes; effective 09/01/95 10/2003
ME 12/26/1995 Yes; effective 09/20/07 8/1988
MD 7/1/1986 Yes; effective 10/01/97 10/2001
MA 2/1/1994 No 7/2003
MI 7/1/1985 Yes; effective 04/01/00 10/2003
MN 8/1/1986 Yes; effective 06/09/09 8/2005
MS 7/1/1994 Yes; effective 05/27/06 7/2002
MO 9/28/1985 No 10/2001
MT 10/1/1987 No 4/2003
NE 1/1/1993 No 9/2001
NV 7/1/1987 No 9/2003
NH n/a No law 1/1994
NJ 3/1/1985 Yes; effective 05/01/00 1/2004
NM 1/1/1986 Yes; effective 01/01/86 1/1994
NY 12/1/1984 Yes; effective 12/01/84 7/2003
NC 10/1/1985 Yes; effective 12/01/06 10/1993
ND 7/14/1994 No 9/2003
OH 5/6/1986 No 7/2003
OK 2/1/1987 Yes; effective 11/01/97 7/2001
OR 12/7/1990 Yes; effective 12/07/90 10/1983
PA 11/23/1987 No 10/2003
RI 6/18/1991 Yes; effective 6/30/11 7/2000
SC 7/1/1989 Yes; effective 12/09/05 8/2003
SD 1/1/1995 No 7/2002
TN 4/21/1986 Yes; effective 07/01/04 7/2003
TX 9/1/1985 Yes; effective 09/01/85 9/1999
UT 4/28/1986 No 8/1983
VT 1/1/1994 No 7/1991
VA 1/1/1988 No 7/1994
WA 6/11/1986 Yes; effective 07/01/02 1/1999
WV 9/1/1993 Yes; effective 07/1/2013 5/2004
WI 12/1/1987 Yes; effective 06/30/09 10/2003
WY 6/8/1989 No 7/2002
Sources: Insurance Institute for Highway Safety, National
Highway Traffic Safety Administration, State Departments of
Transportation.
TABLE 2
BAC 0.08 and Primary Enforcement Law Effects on Seatbelt Use
among Drunk Drivers, FARS
(1) (2)
FE Only FE + Demo
BA08 -0.0394 (0.0375) -0.0398 (0.0377)
PBL 0.0954 *** (0.0352) 0.0943 *** (0.0353)
BA08 and PBL ([psi]) 0.0875 *** (0.0316) 0.0897 *** (0.0313)
Age 0.00113(0.00099)
Male -0.241 *** (0.0144)
Good weather
Daytime
Vehicle type: car
Speed limit < 30
Speed limit 30-39
Speed limit 40-49
Speed limit 50-59
Speed limit 60+
Observations 133.052 132,973
Partial effects
BA08 0.0005 0.0007
PBL 0.0463 0.0461
BA08 and PBL 0.0282 0.0288
(cross-partial)
(3) (4)
Full Model Prev DWI
BA08 -0.0320(0.0315) -0.0980 (0.0829)
PBL 0.118 *** (0.0352) 0.128 * (0.0701)
BA08 and PBL ([psi]) 0.0805 ** (0.0329) 0.179 *** (0.0663)
Age 0.00170 * (0.00092) 0.00138 (0.00129)
Male -0.202 *** (0.0133) -0.229 *** (0.0445)
Good weather -0.0811 *** (0.0109) -0.0887 ** (0.0432)
Daytime 0.0528 *** (0.0148) 0.180 *** (0.0376)
Vehicle type: car 0.229 *** (0.0124) 0.203 *** (0.0292)
Speed limit < 30 -0.259 *** (0.0342) -0.136 ** (0.0694)
Speed limit 30-39 -0.142 *** (0.0347) -0.136 ** (0.0630)
Speed limit 40-49 -0.0788 * (0.0408) -0.0780 (0.0496)
Speed limit 50-59 -0.185 *** (0.0294) -0.195 *** (0.0472)
Speed limit 60+ @ @
Observations 125.983 13,755
Partial effects
BA08 0.0013 -0.0053
PBL 0.0507 0.0584
BA08 and PBL 0.0259 0.0470
(cross-partial)
(5)
No Prev DWI
BA08 -0.0242 (0.0280)
PBL 0.115 *** (0.0390)
BA08 and PBL ([psi]) 0.0665 * (0.0354)
Age 0.00172 * (0.000939)
Male -0.195 *** (0.0141)
Good weather -0.0797 *** (0.0107)
Daytime 0.0420 *** (0.0147)
Vehicle type: car 0.232 *** (0.0123)
Speed limit < 30 -0.266 *** (0.0364)
Speed limit 30-39 -0.140 *** (0.0338)
Speed limit 40-49 -0.0760 * (0.0433)
Speed limit 50-59 -0.182 *** (0.0301)
Speed limit 60+ @
Observations 109,593
Partial effects
BA08 0.0018
PBL 0.0482
BA08 and PBL 0.0219
(cross-partial)
Notes: Probit models include state and period (year-month)
fixed effects. Robust standard errors clustered by state of
residence are in parentheses. @ indicates excluded category.
Sample includes FARS data from 1991 to 2010.
*** p < 0.01; ** p < 0.05; * p <0.1.
TABLE 3
BAC 0.08 and Primary Enforcement Law Effects on
Seatbelt Use Estimates, FARS
(1) (2)
Drinkers Nondrinkers
(BAC = 0.08+) (BAC = 0)
Seatbelt Use Seatbelt Use
BA08 -0.0320 (0.0315) -0.0004 (0.0323)
PBL 0.118 *** (0.0352) 0.173 *** (0.0437)
BA08 and PBL ([psi]) 0.0805 ** (0.0329) -0.0135 (0.0432)
PBL and drunk driver
BA08 and PBL and drunk
driver ([[psi].sub.4])
Observations 125.983 211,571
Partial effects Drinkers Nondrinkers
(BAC 0.08+) (BAC = 0)
BA08 0.0013 -0.0020
PBL 0.0507 0.0611
BA08 and PBL 0.0259 -0.0046
(cross-partial)
BA08 and PBL and drunk n/a n/a
driver
(3)
Drinkers Versus
Nondrinkers
Interaction
Seatbelt Use
BA08 -0.0007 (0.0321)
PBL 0.173 *** (0.0436)
BA08 and PBL ([psi]) -0.0133 (0.0432)
-0.595 *** (0.119)
-0.0313 (0.0355)
PBL and drunk driver -0.0550 (0.0402)
BA08 and PBL and drunk 0.0933 ** (0.0457)
driver ([[psi].sub.4])
Observations 337,552
Partial effects Drinkers Versus
Nondrinkers
Interaction
BA08 -0.0008
PBL 0.0571
BA08 and PBL n/a
(cross-partial)
BA08 and PBL and drunk 0.0302
driver
Notes: Probit models include state and period (year-month)
fixed effects, as well as all covariates included in
Equation (1).
Results presented in column (3) are generated from a fully
interacted model. Robust standard errors clustered by state
of residence are in parentheses. Sample includes FARS data
from 1991 to 2010.
*** p<0.01; ** p< 0.05; * p<0.1.
TABLE 4
BAC 0.08 and Primary Enforcement Law Effects, BRFSS Seatbelt
Use Ordered Probits
(1) (2)
Any Binge Drinking No Drinking
BA08 -0.0264 (0.0311) -0.0115 (0.0365)
PBL 0.1213 ** (0.0565) 0.1162 ** (0.0454)
BA08 and PBL ([psi]) 0.0713 (0.0560) 0.0177 (0.0503)
BINGE30
BA08 and BINGE30
PBL and BINGE30
BA08 and PBL and
BINGE30
([[psi].sub.4])
Observations 845,860 215,364
(3)
Combined
BA08 -0.0114 (0.0368)
PBL 0.1173 ** (0.0456)
BA08 and PBL ([psi]) 0.0166 (0.0506)
BINGE30 -0.5055 *** (0.0626)
BA08 and BINGE30 -0.0154 (0.0228)
PBL and BINGE30 0.0003 (0.0299)
BA08 and PBL and 0.0569 ** (0.0280)
BINGE30
([[psi].sub.4])
Observations 1,061,224
Notes: All results are model coefficients and hypothesis
tests (not partial effects). The sample consists of BRFSS
waves in which respondents reported seatbelt use (1990-
1998, 2002, 2006, 2008, 2010). Each column represents a
separate regression. All models include controls for a
respondent's demographics, income, employment, and
indicators for year-month and state of residence. Robust
standard errors clustered by state of residence are in
parentheses.
*** p < 0.01; ** p < 0.05; * p< 0.1.
TABLE 5
Ordered Probit Partial Effects, BRFSS
Ordered Seatbelt
Predicted Outcome Use (Ordered Probit)
Never Seldom Sometimes
BA08 and PBL cross-partial, binge -0.0053 -0.0039 -0.0054
drinkers
BA08 and PBL cross-partial, -0.0012 -0.0007 -0.0012
nondrinkers
BA08 and PBL and binge drinker -0.0052 -0.0027 -0.0041
cross-partial
Ordered Seatbelt
Predicted Outcome Use (Ordered Probit)
Nearly Always Always
BA08 and PBL cross-partial, binge -0.0065 0.0211
drinkers
BA08 and PBL cross-partial, -0.0018 0.0049
nondrinkers
BA08 and PBL and binge drinker -0.0050 0.0170
cross-partial
Notes: Results are cross-partial effects corresponding to
the double-and triple-interaction models presented for the
main BRFSS regression results. Hypothesis tests are not
conducted, as recommended by Greene (2010), due to
difficulties in their interpretation.
TABLE 6
Impact on State-level Fatal Accidents
(1) (2)
Ln Accidents Ln Accidents
Nighttime Daytime/ Weekdays
(6 p.m. to 6 a.m.) (6 a.m. to 6 p.m.)
BA08 0.0417 * (0.0242) 0.0679 *** (0.0245)
PBL 0.0229 (0.0374) -0.00394 (0.0440)
BA08 and PBL ([psi]) -0.0740 ** (0.0342) -0.0362 (0.0416)
State population in 0.0025 * (0.0013) 0.0020(0.0017)
(100,000)
State beer tax -0.530 ** (0.210) -0.241 (0.148)
State unemployment rate -0.0008 (0.0082) -0.0102 (0.0071)
Observations 12,101 11.912
Notes: Weighted least squares regressions include state and
period (year-month) fixed effects. Robust standard errors
clustered by state of residence are in parentheses. Sample
includes data from 1991 to 2010.
*** p<0.01; ** p<0.05; * p<0.1.
