Accid. Anal. and Prev., 26, No. 5, pp. 647-654.

1994
Copyright 6 1994 Efsevier Science Ltd
Pergamon P&&d in the USA. All rights reserved
~~-4S75/94 $6.00 + .oO zyxwvutsrq

THE GRAND RAPIDS DIP REVISITED

PAUL M. HURST,' DAVID HARTE,~ and WILLIAM J. FRITH~

iConsultant, Alcohol and Traffic Safety, Pukerua Bay, New Zealand; *Victoria University of Wellington,
New Zealand; 3Land Transport Safety Authority, Wellington, New Zealand zyxwvutsrqponmlkjihgfedc

(Accepted 23 October 1993)

Abstract-This analysis addresses an issue that has concerned road safety authorities for some 28 years:
the celebrated “Grand Rapids Dip.” This, most readers will recognise, is the below-baseline excursion,
which occurs in the blood alcohol concentration (BAC) interval of .Ol%-.04%, of the relative risk curve for
accident risk versus blood alcohol, derived from the 1964 Grand Rapids Study data. The present analysis
has its starting place in the explanation advanced by Allsop, who noted that the case/control comparisons
were biased due to the disproportionate representation of demographic subgroups in different blood alcohol
concentration class intervals. Indeed, when relative-risk curves are derived separately for subgroups of
differing drinking habits, the resulting separate risk curves all show monotonic increases at al1 blood alcohol
concentration ranges. Such separate relative risk curves are unpopular, and most of the road safety community
pays them little heed. Thus, the original concept of the “dip” remains with us. For this reason, we have
derived, using a simple but realistic statisticat model, a single relative-hazard curve from the Grand Rapids
data, one that is free from the distortion introduced by unequal representation of different demographic
subgroups in different blood alcohol concentration class intervals. This curve indicates that accident risk
increases with increased blood alcohol concentration regardless of self-reported drinking frequency. However
more frequent drinkers have less risk at all blood alcohol concentration levels, including zero, than less
frequent drinkers at the times and places sampled.

Keywords-Alcohol, Risk, Crashes, Accidents, Drink driving

INTRODUCTION mathematical proof, extending comparisons beyond

The Grand Rapids curve (Borkenstein et al. 1964) the region of the “dip”. He showed that self-re-
estimates the relative risk of accident of drinking ported drinking frequency, the strongest of the sub-
drivers at increasing blood alcohol concentration group differentiators referred to by Allsop, was suf-
(BAC) with the relative risk at 0.00% BAC set to ficiently powerful that the daily drinkers had a case/
one. This curve, when aggregated over all subclasses control ratio at blood alcohol concentration of .09%
with available data, drops below one for blood alco- that was approximately the same as those obtained
hol concentration counts of .Ol%-.04% and then for the yearly and monthly drinkers at zero BAC.
starts to increase exponentially. (See Fig. 1.) However (and this is the basis of our current discus-
Such a dip is not present in similar curves pro- sion) every subgroup showed its lowest case/control
ratio at the lowest blood alcohol concentration inter-
duced by alternative methods zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQPONMLKJIHGFEDCBA
suc h as those given
in Zador (1989), where the relative risks were esti- val (zero to .Ol%). (See Fig. 2.)
mated using data from the U.S. Fatal Accident Re- In other words, the more frequent drinkers were
porting System in conjunction with roadside breath good enough drivers (for demographically associ-
testing survey data. The “dip” in the curve at ated causes) that they still had relativefy good safety
.Ol-.04% has become known as the Grand Rapids records in the zero to .09% BAC range, despite the
Dip. Allsop (1966) showed that the Grand Rapids impairment involved.
Dip could be attributed to disproportionate repre-
sentation of driver population subgroups in the vari-
THE RATIONALE BEHIND THIS PAPER
ous blood alcohol concentration class intervals rep-
resenting levels measured on the road. More Although Allsop made the original observation
generally this is known as Simpson’s paradox (see more than 25 years ago, it seems to have attracted
Freeman 1987). Hurst (1973) elaborated on Allsop’s surprisingly little notice. There still seem to be many
648 P. M. HURST et al.

Y
cn
.-
a4
0
>
.-

-a3

i I / 1 / I

0 zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQPONMLKJIHGFEDCBA
00x 0 04%
01-C 0 05-O 07x 0 08-O IQx 0 Ilii

Blood Alcohol Concentration

Fig. 1. The Grand Rapids Dip,

who believe that the Grand Rapids Dip is actually only self-reported drinking frequency but also age,
evidence, of at least tentative quality, that a little sex, education, economic status, and driving expe-
alcohol improves driving. Perhaps this is because rience. However, this would have required the
people prefer to think in terms of “a” curve, or original raw data, which are no longer available.
relative hazard function, rather than a number of Accordingly, we chose the strongest single attri-
separate ones. For this reason, we thought it desir- bute, self-reported drinking frequency. The data
able to derive, using a simple but realistic statistical available to us from the Grand Rapids study con-
model, a single relative-hazard curve from the Grand sisted of a table of case and control counts for
Rapids data, without the distortion previously five self-reported drinking frequency classes (daily,
caused by the inadequate representation of different thrice-weekly, weekly, monthly, and yearly or less)
subpopulations in different blood alcohol concentra- and five blood alcohol concentration classes (.OO%.
tion class intervals. .Ol-.04%, .05-.07%, .08-.lO%, .ll+%). Unfortu-
nately , in the cases of “yearly or less” and
“monthly” drinkers, accident counts for the three
OUR METHOD
higher blood alcohol concentration classes had in
It would have been most elegant to adjust the each case been accumulated into just one cell
relative hazard function for simultaneous variations (Table 1). In our analysis we treated the two cells
in a number of moderating influences, including not that result as a blood alcohol concentration of
 The Grand Rapids Dip revisited 649

I
7
.
,’ I
1 ,’
/’
!’ ii
,’
,’ i:
Y ,’
(I)
I
.-
,’

I
,’
ii
a ,’
,’
5
.’ P ii
al ,’ /’
>
.- ,’
,’
/’

/’
/’
I’
I
/I
,’
,I 8’

//’

1
m /’
,’ I
/’

I”’
/
a ,’ I’ I
a zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQPONMLKJIHGFEDCBA
,’ ,’
,’ /
/’ ,’
,’ ,/ /i
,’ , / /
/ /
I’ ,
/

;___~.~~.~~.~
,’ ,’
,’ ,I
/
/

0 00X 0 01-O 04x 0 05-O 071 0 08-O I3Z 0 llil

Blood Alcohol Concentration

Fig. 2. Relative risk by blood alcohol concentration and self-reported drinking frequency.

Table I. Grand Rapids accident and control data

Self-reported Blood alcohol concentration (%)

drinking
frequency .oo .Ol-.04 .05-.07 .08-. IO .11+

Yearly or less Control drivers 2783 88 4 zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQPO

0
Accident drivers 2485 94 a* /7*
Monthly Control drivers 1010 57 8 I
Accident drivers 934 55 d* e*
Weekly Control drivers 1375 I62 39 I7
Accident drivers 959 I22 50 51 97
3 times per week Control drivers 616 103 32 18 I3
Accident drivers 340 64 29 25 52
Daily Control drivers 505 136 43 27 I6
Accident drivers 160 63 22 27 67

Note: The counts marked * are presented as an aggregate for all cases over .05% in Borkenstein et al. (19641,
such that (I + zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQPONMLKJIHGFEDCBA
b + c = 34 and d + r + f‘ = 49.
650 P. M. HURST et al.

.08-.lO%, being the “average” cell over which and observed values will be denoted by lower-case
the accumulation had taken place. characters.
A brief description of the way the data were Further, let ‘pijbe the probability that a journey
sampled in the original study follows. Accident data in the ijth class results in an accident and define Zijk
over a one-year period were classified according to as such:
each involved driver’s self-reported drinking habits
and measured blood or breath alcohol at the time of
0 if the kth journey in class ij
the accident. results in no accident
Using an historical set of accident data, previ- z,, =
ous accident sites, dates, and times of day were 1 if the kth journey in class ij
1 results in an accident,
extracted and tabulated. This information was used
to furnish guidelines for the selection of a “control”
then
group of drivers. About four drivers were stopped
at each historical accident site at an anniversary date
and time. They were asked about drinking fre- wy = ‘pjj (1)
quency, among other matters, and given breath
tests. and
We can therefore think of the situation as af-
fording two contingency tables involving accident Var (Z,,) = E[Z$.,] - E[Z,J2
and control drivers. Both tables have the same for-
mat, with rows representing reported drinking fre- = cpjj(l - cpij). (2)
quency and columns the estimated blood alcohol
Using these definitions
equivalent resulting from breath tests.
The control table consists of the total number
of drivers stopped, interviewed, and tested. The
margins of the control table will be random, but
the total number of drivers is predetermined. The Then
control table provides us with relative exposure by
types of drivers (based on reported drinking fre- -wijl = zyxwvutsrqponmlkjihgfedcbaZYXWVU
a&jJmqjl
quency) at times and sites at which accidents have
previously occurred. It must be carefully noted that = PijE[wijl, (3)
this control group is not typical of the general driver
population, but is rather typical of the population of i.e. the expected number of accidents equals the
drivers using the roads at the times and places of probability of an accident times the expected num-
the accidents recorded a year earlier. Therefore, the ber of journeys. Also
absolute number of drivers in each cell is not so
much of interest as the ratio of numbers in different E[ Y;] = E[WJE[Z;.J + E[Wij (WV - l)]zz[Z,,]?
cells (relative exposure). By performing a chi-square
= E[W,,](E[Z&] - E[Z~J’> + E[W$E[Z,J2
test on the control table, one can test the hypothesis
that self-reported drinking frequency is independent = E[ WJ Var (Z,,)
of blood alcohol concentration. The test statistic is
+ _f?[W$E[Z,,]* (from eq 2). (4)
very large and this hypothesis is clearly rejected.
In fact it can be easily seen that blood alcohol is
positively correlated with self-reported drinking Now, consider a sample of drivers taken at
frequency. some other time and let Xi be the number of these
sampled drivers (journeys) in the ijth class. Then
an estimator of the number of journeys in the “acci-
dent” interval observed is given by rXij, where Y >
MODEL DESCRIPTION 0 is some scaling constant, and so
Let Y, be the number of accidents, during some
period of time, of drivers in the ith self-reported E[W,,] = rE[X& (5)
drinking frequency class and the jth BAC interval.
Let Wi, be the total number of journeys in the ijth and from eq 3,
class in the same period of time. Note that random
variables will be denoted by upper-case characters E[Y,] = cpijrE[XJ. (6)
 The Grand Rapids Dip revisited 651

Therefore an estimator for the number of accidents lf!lu = exp (a; + pj + 8,)

during the interval of interest is rcpGXij.In order to

use this as an estimator, we need to know the ex- where (Y;is a parameter relating to the ith level of the
pected variability between the actual numbers of blood alcohol concentration factor, pj is a parameter
accidents, Yu, and that predicted by the estimator. relating to thejth level of the self-reported drinking
That is, we consider E[( Yi, - r~Xij>*]. Between the frequency factor, and 6, describes a possible interac-
first and second equals signs below we assume that tion between the two. The interaction between the
E[XqYJ = E[XJE[YJ. This will be the case if X0 factors could be estimated only if we assumed some
and Ygare independent. Recall that Yiiare the acci- functional relationship within the two main effect
dent counts in some interval of time and Xij are factors (e.g. risk increases exponentially with blood
numbers of sampled journeys at some other time. alcohol concentration). This is because there are
Thus, we assume that the accident counts are statis- insufficient degrees of freedom to allocate a parame-
tically independent of the latter sampled journeys. ter for each level of both factors and each level of the
Thus interaction term. Assuming a functional relationship
would relinquish a sufficient number of degrees of
freedom to estimate the interaction. This would pre-
sumably defeat the purpose of the exercise in some
= E[Y$ + r2#[X$ - 2rcpijE[XijYJ way, as the line no longer has the ability to assume
= E[WJ Var (Z,,) + E[ W$E[Z,]* “dips” if one has assumed some monotonic func-
tion. We have therefore retained a separate parame-
+ r*p@[X$ - 2u2~~E[X,]’ (from eqs 4 and 6) ter for each factor level and assumed that 6, = 0.
The validity of this assumption can be checked by
= &[X&&l - Q) + &{Var (WJ + E[W,]‘}
the amount of residual variation remaining after fit-
+ r*q$ Var (XJ - &E[W,]* ting the main effects model. This is done in the next
section.
(from eqs 1, 2, and 5)
The parameter estimates were calculated using
iteratively reweighted least squares, the weight be-
Assuming that Xij and W, are Poisson random vari- ing the inverse of the variance, estimated at each
ables (and therefore E[X,,] = Var (X,) and iteration by
E[WJ = Var (WJ), then it follows from eq 5 that
Var (W,) = r Var (X,) = rE[XJ. Thus

-WY, - v&J*1
k[(y.. - r(p..X_)2]
rJ rJ rJ
=
( !Q11 +
_Q
6..
rJ’

= rEIXJcpij(l - CpJ + Y@[XJ + r*$O;EIXjj] This is based on eq 7. PROC NLIN in the SAS
(SAS Institute Inc., Cary, NC) package was used
= rt&S[XJ( 1 + r(pij) to achieve this. A number of iterations are required
before convergence is achieved. The parameter esti-
=(I+%) E[ Y,,] (from eq 6). (7) mates could also be calculated with other packages
such as Genstat and GLIM. For a general descrip-
tion of this class of models see McCullagh and
Therefore, the actual value of r, which is not known, Nelder (1990).
is not needed. There is then sufficient information to The risk ratios were calculated as exponentials
calculate the relative risks and the relevant standard of the corresponding parameter estimates.
errors.
The expected number of accidents during the re-
cording period is modelled by eq. 6. We used the RESULTS
control survey results, xij, as estimates for E[XJ. The accident risk was assumed to be of the form
Thus the estimated number of accidents, fig, is
$6 = exp (a; + pj),

pij = x&
This model assumes that the interaction term, 6,, is
where I& is an estimate of rpij. Note that I$~/&~~is zero.
an estimate of (pijl(pili.or the relative accident risk. An assessment of the validity of our model can
We assume that the risk is multiplicative, thus be achieved by considering the Pearson residuals.
652 P. M. zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQPONMLKJIHG
HURSTet al.

Table 2. Estimated relative risks goodness-of-fit of the model. Our model gives
Self-reported Blood alcohol concentration (70) 11.171/(21 - 9) = 0.931, indicating that the ob-
drinking served variability is comparable to that predicted by
frequency .oo .Ol-.04 .os-.07 .08-.I0 .l I+ our model (i.e. close to one). The largest residual
Yearly or less 1.00 1.17 1.71 3.93 10.68 value (1.84) originates from the cell: self-reported
Monthly 1.03 1.20 1.I7 4.05 11.01 drinking frequency of yearly or less with a blood
Weekly 0.78 0.91 1.33 3.05 8.29 alcohol concentration of O.OS%-0.10%. This is not
3 times per week 0.59 0.70 1.02 2.34 6.35
Daily 0.37 0.43 0.63 1.45 3.93 unexpected, as our knowledge of blood alcohol con-
centration for this cell is more uncertain than for
other cells. However, such a value is still within
The mean sum of squares of the Pearson residuals, acceptable limits.
i.e. The risk in the (i, j)th cell relative to the (1, 1)th
cell is given by

where &(Y, - rcpijX,)‘] is given by eq 8 and zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQPONMLKJIHGFE

df Estimated values are tabulated in Table 2. Given
is the residual degrees of freedom. measures the the satisfactory fit of the main effects model (exclud-

1 I T zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQPONMLKJIHGFEDCBA
1

IC

Y 7
cn
.-

lx
6
Q)
>
.-

- 5
a
-
0
m 4

0 1 / I I I , I

0 00 0 02 0 04 0 06 0 08 0 10 0 12 a 14 0 16

Blood Alcohol Concentration

Fig. 3. Relative risk by blood alcohol concentration.

 The Grand Rapids Dip revisited 653

Accident risk appears to decrease generally

with increasing self-reported drinking frequency.
The estimate for monthly drinkers shows a slightly
greater risk than for the yearly ones, but the differ-
ence is not significant. It seems reasonable to con-
clude that the risks for monthly and yearly drinkers
are much the same. Weekly and more frequent
drinkers have a significantly (p < .05) lower accident
risk for a given blood alcohol concentration than
monthly and yearly drinkers at times and places
surveyed.
These conclusions are simply relative risks
shown by a cross tabulation. It would be going far
beyond the data to infer a causal relationship.based,
e.g. on acquired tolerance to alcohol’s pharmacologi-
cal effects. Such a tolerance may well be involved, but
there are other potential contributors as well. Note,
first, that the more frequent drinkers were safer at all
blood alcohol concentrations, including zero.
Second, note that certain demographic attri-
butes (viz. age and reported annual mileage) were
associated with blood alcohol concentration in the
Grand Rapids control group in such a way as to load
the zero BAC subsample with large proportions of
very young, very old, and low-annual-mileage driv-
ers. It would, indeed, be enlightening to have all of
the cross-tabulations at hand. The original Grand
Rapids data represent a lost treasure trove: a tragic
loss, since no subsequent case-control study has
been able to muster the resources needed to gather
the same wealth of information.
Were the original raw data available, it would
-20.i5-IO-05 00 05 IO I5 20
be appropriate to include in the model a number of
Pearson Residual interrelated variables such as age, sex, education,
socioeconomic level, and reported driving habits
Fig. 4. Histogram of Pearson residuals. (annual mileage). Since this information is lost, we
have had to content ourselves with a simplified anal-
ing the interaction term, 8,) then, given a specific ysis based on blood alcohol concentration and re-
driver, the relative risk of an accident at different ported drinking habits. The new relative risk curve,
blood alcohol concentration levels is independent of as shown in Fig. 3, removes the misleading dip from
that driver’s self-reported drinking frequency. Our the Grand Rapids curve. It will now be possible for
model predicts that these relative risks are given by Ministries and Departments of Transport to illus-
exp @jYexP (Pi), trate their drinking-driving publications without ed-
iting the “dip” region of the Grand Rapids curve.
for each of the five blood alcohol concentration cate- This might lead, in turn, to greater public acceptance
gories. These estimated relative risks for increasing of official pronouncements concerning relative risks.
blood alcohol concentration, based on our model, Although the “dip” has vanished, it is still ap-
are plotted in Fig. 3. A histogram of the Pearson parent that the Grand Rapids results are not directing
residuals is shown in Fig. 4. our attention to the left-hand part of the curve.
DISCUSSION CONCLUSIONS
Accident risk appears to increase monotoni- Accident risk appears to increase with blood
cally with blood alcohol concentration. All non-zero alcohol concentration. All non-zero BACs have sig-
blood alcohol concentration levels have significantly nificantly more risk than a zero BAC at the .05
more risk than a zero BAC (p < .05). level.
654 P. M. HURST et al

Accident risk appears generally to decrease DISCLAIMER

with increasing self-reported drinking frequency.
This paper is the work of the authors and
The estimate for monthly drinkers produces a
does not necessarily reflect the views of their
slightly greater risk than for yearly, though this is
employers.
by no means significant. It would be reasonable to
assume that the risk of yearly and monthly drinkers
are much the same. Weekly and more frequent REFERENCES
drinkers have a significantly (at .OS% level) lesser Allsop, R. E. Alcohol and road accidents. A discussion
accident risk for a given blood alcohol concentration of the Grand Rapids study. RRL Report No. 6 Ham-
than monthly and yearly drinkers. mondsworth: Ministry of Transport, Road Research
This does not, of course, imply that regular in- Laboratory; 1966.
gestion of ethanol confers a degree of prophylaxis Borkenstein, R. F.; Crowther, R. F.; Shumate, R. P.;
Ziel, W. B.; Zylman, R. The role of the drinking
against accident risk. As can be seen in the original driver in traffic accidents. Bloomington, IN: Depart-
Grand Rapids Report (Borkenstein et al. 1964), there ment of Police Administration, Indiana University;
are many suggestive demographic correlates of self- 1964.
reported drinking frequency, including age and so- Freeman, D. H. Applied categorical data analysis. New
cioeconomic status. In general, these self-selective York: Marcel Dekker; 1987.
Hurst, P. M. Epidemiological aspects of alcohol in driver
features tend to favour regular drinkers as opposed crashes and citations. J. Safety Res. 5:130-148;
to the less frequent users. 1973.
The conclusions above are not a statement McCullagh, P. & Nelder, J. A. Generalized Linear Models
about frequency of accident involvement of driver (2nd Edition). London: Chapman & Hall, London;
types. They are the relative risks, given a certain 1990.
Zador, P. L. Alcohol related relative risk of fatal driver
blood alcohol concentration and self-reported drink- injuries in relation to driver age and sex. Arling-
ing frequency, and the Grand Rapids demographic ton, VA: Insurance Institute for Highway Safety;
pattern. 1989.