You are on page 1of 9

A CLARIFYING CORRELATION …provides a new tool for

traffic safety analysis


(as submitted to Traffic Technology International in the UK)
Because his FISITA 2004 paper (Barcelona, May 24-27) followed directly
from his work on the article Solar Power (pp. 80-81, Aug/Sept ’03) Al (and
we) felt it important to review the paper here …

Diligent readers of TTi will by now be aware that my research has indicated a dominant
role for mental distraction in collision causation. Indeed, as I have often said, all
‘accidents’, whether in the home, on the job or on the road, occur because “the mind is
not on the motion”
The quest for yet more proof that mental distraction is the fundamental cause of traffic
collisions has led me through volumes of traffic safety reports, through thickets of
economic acronyms and through libraries literally the world around. Throughout that
research the good safety record of three countries has caught and held my curiosity.
The AMPS Theory of ‘Accident’ Causation: The conventional wisdom is that there are
many ‘causes’ of collisions. Eight years of research and six traffic safety papers have
unequivocally demonstrated that ‘economic distraction’ must be among those causes.
Moreover, mathematics suggests strongly that, if a change in “one of many” causes can
produce such a significant change in the overall collision rate, that cause must be a
dominant cause (i.e. responsible for a very large fraction of all collisions).
BUT, humans can be mentally distracted on many subjects. In fact , without prior
knowledge of that dominance of economic distraction, one may have predicted that the
love life of the young driver, and the family life of the middle-aged, were the ‘dominant’
mental distractions. Following that thought soon leads one to the conclusion that the
‘causes’ cited by the public authorities (speeding, following too closely, etc.) are not
‘causes’ at all. They are merely symptoms of an underlying disease … the disease of
mental distraction.
The importance of this ‘disease’ in collision genesis is argued convincingly by Frank et
al (FISITA 2002-Helsinki). In their analysis of “ … deficiencies in the … Vehicle
Control Process (VCP) i.e. the sequence (of) perception, decision/evaluation, and
execution.” they point to the perception phase as the source of almost all the variation in
the reaction time to any traffic incident. And 100+ years of international traffic police
experience (Europe & North America) fully supports that conclusion.
Discussions with those BMW engineers over the last half of ’00 during preparations for
a “BMW Showcase” at ITS Torino, together with the reaction of the Showcase audience,
have shown that the best way to communicate the kind of deep distraction that is the root
cause of all traffic ‘accidents’ is to call it the Absent-Minded Professor Syndrome or
AMPS* (*trademark applied for). We thus arrive at … the AMPS Theory of ‘Accident’
Causation:
‘Accidents’, whether in the home, on the job or on the road,
occur because “the mind is not on the motion”.
As I reported in the Aug/Sept 2003 issue of TTi, the SUNflower trio’s (Sweden, UK,
Netherlands) superior performance has also intrigued officers of the European Union who
have sponsored the ‘SUNflower’ study as a tripartite effort by VTI, Sweden, TRL, UK,
and SWOV, Netherlands.
That high-quality study shows that the Top Trio are doing a very good job on all of the
‘standard’ traffic safety programs. They are however not significantly better on those
programs than several other European countries, notably Germany, France and
Switzerland. Thus, although the study has not (yet) yielded a positive result, we do have
some confidence in an important negative result: none of the ‘standard’ programs can
explain the Top Trio’s superior safety record. Moreover, the study also provides the
data to explore …

… some ‘non-standard’ explanations.


Many ‘local area’ studies have shown that the fatality rate decreases in high-density
traffic - with the ‘conventional wisdom’ concluding that this is due to the lower average
speeds observed in such traffic. (It is perhaps important to note here that such ‘micro’ studies were
almost exclusively carried out in Europe where, to North American eyes, the comparison was always being
made between dense traffic and very dense traffic.) No country maintains a sufficiently large
traffic monitoring system that such a ‘micro’ parameter could be measured accurately at
the macro level, i.e. as an annual country statistic. Nonetheless, use of such a density
parameter held so much promise that researchers were willing to try a ‘reasonable
surrogate’- Annual Average Daily Traffic (AADT, best understood, as ‘vehicles per
day’ … averaged over every kilometre of roadway in the country!)
Brühning and Berns, of BASt, made early (1995) use of IRTAD to show that, indeed,
there was a strong inverse relationship between AADT and fatalities/billion VKmT on
motor ways (MFR). However, because their data covered the years 1980 to 1993, they
were limited to the 8 most densely populated European countries plus the USA. Thus,
again, their data covered only the range from 25,000 (dense) to 55,000 AADT (very
dense) but, within that limited range, the correspondence was very good. And two of our
Top Trio are in the bottom right with the largest AADT and the lowest fatality rates.
The study thus provided some support for the then-developing theory that the 3P, the
‘Plausible Physical Path’, connecting denser traffic with lower fatality rates is the lower
average speed of that denser traffic. However, with data now available for the third
member of our Top Trio, Sweden becomes …

… the exception which disproves the rule.


After a decade of good progress in data gathering (for which all credit is due the
OECD/IRTAD/BASt group) we now have several more countries to add to that graph …
and one of them is Sweden. The IRTAD shows Sweden with the lowest motorway
AADT (of all countries reporting), and the SUNflower study shows it has the expected
higher speeds. However, its MFR is fully comparable with the two countries with the
highest AADT.
Now it is a fundamental principle of science that, although a theory may have been
supported by hundreds of tests, that theory is negated by a single, scientifically-valid
experiment or, as in this case, a valid epidemiological observation.
Although the IRTAD can never be at the level of precision expected in ‘a scientific
experiment’, it is rapidly approaching a level from which ‘valid epidemiological
observations’ can be made. Moreover, the data for our Top Trio has been extensively
‘peer-reviewed’ in the SUNflower study. The data for Sweden thus calls into serious
question the ‘Speed Kills’ theory of the mechanism by which AADT influences the
traffic fatality rate. In the following we will update the AADT graph and show that …

…there is nothing more practical than a valid theory.


Math is a tool to help us understand the physical world. That 'tool' is itself blind and
operates on a numerical representation of the various elements of the physical world
which we provide to it. If we have overlooked an important element the 'tool' cannot
search for and insert it. The calculated result will thus be deficient ... and that deficiency
will be in direct proportion to the importance of the missing element.
In examining the relationship between two variables it is often of great practical value
to confine an investigation to the range of immediate interest. Such ‘limited range
models’, while ‘getting the job done’, will often make non-credible predictions outside
that range. Such predictions are usually discarded as irrelevant. They should, however, be
carefully examined for a more insidious danger: that the developing theory of the
interaction between the variables is not valid … and thus that any corrective action
indicated by the invalid theory may instead make matters worse.
In this case we are using math to 'model’ the relationship between the MFR and AADT.
We will use the resulting 'model', i.e. the regression line, to pass judgement on the traffic
safety programs of various countries according to their deviation from the 'average' line.
If that judgement is to be accurate that ‘average line’, i.e. the ‘model’, must be correctly
positioned all along its length.
Now the 0,0 data point is not only the most accurately measured relationship between
the two variables, it is also completely free of the confounding variables (roadway
infrastructure, economics, weather etc.) which are 'known unknowns' in all the other data
points. Moreover, forcing the intercept through that 0,0 point will materially affect the
position of the line throughout its length. Thus, in the search for the true "model",
including this "important element" in the math 'tool' is not only useful, it is imperative!
Similar considerations apply with equal force at the other end of the data range. A
model which makes non-credible predictions for points near the end of the data range
(either inside or outside that range) must be rejected in favour of a model which makes
more reasonable predictions, without regard to the respective correlation coefficients. If
these ‘boundary conditions’ are taken into account …

… choosing between models is a snap


In the AADT/MFR graph (facing page) the 2nd degree regression (upper line) is for the
original nine countries available to Brühning and Berns (but now with updated data). It
shows, on their limited data range, the same good correlation … but also with the same
non-credible predictions at both boundaries. [Note that, as indicated by the arrows, some points for
I and A are ‘off the graph’]
For the 3rd degree regression (lower line) I, A & B were excluded (not only because of their
'outlier' position but also because they have the lowest seatbelt usage and among the highest per-capita

Numbers sometimes lie …


Scientific analysis relies heavily on “least squares regression” as an objective evaluation
of the ‘goodness of fit’ of the observed values of two parameters with a ‘regression line’,
i.e. an assumed relationship between them. The ‘goodness’ is usually judged by the
‘correlation coefficient’ (R2) with a range of 0-for no connection whatever, to 1-for
perfect correlation.
Social scientists are pleased with an R2 of 0.5 but physical scientists don’t accept
anything less than 0.7 as being a ‘good fit’. However, even if the ‘objective arithmetic’
produces a good fit the scientific analysis is only half done. The investigator must still
find a 3P (Plausible Physical Path) linking the two in a causal relationship … and that
‘theory’ must hold throughout the full range of possible values for the two parameters.
Thus, in the MFR/AADT graph, if one pays no attention to the left end of the
regression line (where the fatality rate soars off to infinity as the traffic density goes to
zero) the correlation calculation for the original 9 countries produces some truly
phenomenal numbers.
The R2 starts off at a ‘good’ 0.73 and, after normalization for seat belt usage, it jumps to
the 0.85 shown on the graph. Moreover, if a ‘numbers person’ looks at the graph and
decides that Belgium is so far away from the others that it can legitimately be excluded
(as an ‘outlier’) from the correlation calculation, the R2 jumps to a ‘too good to be true’
0.95!
Clearly the numbers are lying to us. The size of the R2 must be due to the ‘accidental’
position of Italy and Austria in practically a straight line with the six countries with much
higher AADT … and the very high MFRs for the former two probably have nothing
whatsoever to do with AADT. Nor, see next paragraph, do those high MFRs have
anything to do with higher average traffic speed (which is the 3P purportedly causing the
MFR to rise as average speed increases with decreasing AADT in the latter six
countries).
The author has put over 25,000kms, mostly motorway driving, on rental cars on 6 trips
to Europe since ’96. With several thousand kilometres in the eastern US as well, the only
country of the ‘original 9’in which I have not recently driven is the UK. There is very
little observable difference in motorway traffic conditions between countries. In peak
periods, of course, there is no difference at all but even in ‘free-flow’ all motorway traffic
appears much the same … and with only a barely noticeable increase in average speed
from that prevailing on North American motorways!
and all countries now reporting (S, FIN, CDA & DK) have been
alcohol consumption)
added. Moreover, for better accuracy, the fatality rates of all countries were ‘normalized’
to 95% seatbelt usage. [The normalized fatality rates are shown as ‘open’ diamond markers below the
original country markers. Four countries, F, D, UK and S, were already at 95% usage and thus the two
markers are coincident.] In spite of being ‘forced through 0,0’ this ‘full-range’, 10 country
model retains the good correlation of the ‘limited range’ model but with good credibility
at both boundaries.
Italy is way off
the graph with Motorway Fatality Rate (pe
fatality rates of:
8 before 13.4
after 9.32 A'01
normalization for (9.2, 24.6)
seat belt usage
7

USA'99
5 CDN'99(est.)
Killed/10^9 VKmT

F'01
FIN'01
4 DK'01
The lower trendline thus becomes a useful model for the average MFR (i.e. normal or
‘expected’) for a modern, industrialized nation. The model equation is
MFR(Expected) = 0.00004033x3 - 0.00685x2 + 0.3099x CH'01
(where ‘x’ is AADT in thousands).
In this3PAST (Program Assessment for Safer Traffic) model a country’s position relative
to the line shows whether they have a better (below) or worse (above) safety record than
the average, given the existing traffic density in that country.
S Regression equation for Final
NOTE: this author is neither the only, nor the first, TTi correspondent to suggest using
AADT as a normative tool in research SUN2000 (w/o I,DOT
on traffic safety. The Colorado A(p.20,
& B) after normaliz
2
Aug/Sept ‘03) uses AADT as a measure of “traffic exposure”. forThey95%
consider that the
seat
“normal level” of safety is represented by the, “ … expected number of accidents at a
belt usage
specific level of AADT (and) then the degree of deviation from ythe
= norm
4E-05x3
can be - 0.0069x2 + 0.3
stratified to represent specific levels of safety.”
R2 = 0.72
The SUN countries provide an interesting example of the effect of the different
1
perspective. Based on MFR alone they are very close and in the order (best to worst) U-
N-S. Based on the PAST/AADT model the order is exactly reversed … and Sweden is
best by far!

0
0.00 10.00 20.00 30.00 40.00

AADT-1000 VKmT per


A Clarifying Correlation
The 3rd degree model thus extends the correspondence between MFR and AADT to
include countries representing the full range of motorway traffic density while retaining
the very good correlation of the original, limited range, model. This ‘epidemiological
observation’ is fatal to the Speed Kills Theory but is completely in keeping with the
AMPS™ Theory of ‘Accident’ Causation … as can be fully explained by the ‘tibii’ - time
interval between interaction incidents.
In simple terms, denser traffic is safer (per vkmt) because each driver has frequent
'interaction incidents' and thus is intensely aware of the need to keep her/his mind 'on the
road' rather than on business or family matters.
In average densities on the other hand, in which all vehicles are moving quickly and
(apparently) safely, there is a sense of routine rather than of danger. The tibii is much
longer and much more of the driver's mental activity is devoted to non-driving tasks …
and often doesn't get back to driving in time to avoid a collision.
In low density traffic, although much time is spent in AMPS™, there is more distance
and time between vehicles. Thus each vehicle which comes within a possible conflict
range is sufficiently unusual that it 'catches the drivers attention' before the conflict
becomes a collision.
Returning to the right side of the line, the lower speeds observed in denser traffic can
also be accommodated within the AMPS™ Theory. As the driver has ‘the mind’
frequently brought back to ‘the motion’ the speed is lowered by the driver, either
consciously or (most often) sub-consciously, in accordance with a, probably sub-
conscious, perception of greater risk.
Thus, a valid theory has produced (Section 12, F2004V297) a very practical
recommendation for safer traffic: “3. Traffic authorities should refocus their ITS
development programs on 'active safety', i.e. helping drivers maintain their attention on
traffic when it matters most.”
And that’s not the only thing this new model does for us. It also provides …

…a new tool for traffic safety analysis.


In trying to learn from each other the PAST/AADT model improves greatly on the simple
question, “How many died on your roads last year?” That question, even if answered
accurately, will make no progress beyond mutual sympathy. The first step to a useful
comparison of traffic safety between countries is to account for the effect of traffic
volume … and that IRTAD has done by reporting the statistic “fatalities per billion
kilometres travelled”. They have also made substantial progress on the second step,
accounting for the effect of traffic density (discussed above), with many countries
reporting both total roadway length and total motorway length.
In a static world answering these two questions (How much are the vehicles used? and,
How many are on the same road at the same time?) would give us a useful comparison of
differing fatality levels between countries. However, although motorway infrastructure is
relatively constant within and between countries, the passive safety technology on the
vehicles has been changing dramatically, and for the better, over the past thirty-five
years. Moreover, the rate of progress also differs greatly between countries.
Fortunately, the simple provision, and use, of a seatbelt accounts for almost all of the
fatality rate improvement. Unfortunately, although all modern countries require the
installation of seat belts as standard equipment in new motor vehicles, the usage rate
varies widely between countries. Various studies have, however, provided similar
estimates of the fatality rate reductions obtainable with increased usage of seatbelts. The
PAST/AADT model has thus been ‘normalized’ for 95% seat belt usage in all countries.
As noted above this new tool enables any country to compare its motorway safety level
with other countries. They simply calculate their AADT and MFR and then check
whether they are above or below the ‘average’ line. However, the mathematical theory
behind that regression line suggests a much more important use for it.
In brief (TTi readers don’t need the details; only, perhaps, a reminder) the theory states
that the correlation coefficient – R2 – of 0.72 indicates that 72% of the variability in the
MFRs for these 10 countries is explained by their AADT values. (If this seems strangely
high, remember that the MFRs were first normalized to 95% seat belt usage.) Of more
importance, the theory goes on to say that all the other parameters one might think had an
influence on differences in safety among the countries must be accommodated within the
remaining 28% … and that the influence of those ‘other’ parameters can be determined
by regressing each of them against the ‘residuals’ of the PAST/AADT model.
A further reminder: a residual for a country is the vertical distance from the MFR for
that country to the regression line. It is usually expressed as a percent + or- of the
corresponding value of the regression equation at the AADT value of that country. For
example the residual for Sweden is -28.9% and that for France is +16.3%, being
respectively, of the 10 countries studied, the furthest below and above the model line.
Thus, for example, one can examine the PAST/AADT residuals to determine …

… whether General Speed Limits (GSL’s) are beneficial … or not.


The author has long been interested in the effect of GSL’s on traffic safety and thus I had
the motorway GSL’s for all 10 countries in my files. Those limits ranged from 100 to 130
kmph for 9 countries and the author’s experience (over 25,000kms on European
motorways in the past eight years) made it possible to estimate an ‘effective limit’ for
Germany’s autobahns. That is, in all countries, a GSL is ”honoured more in the breach
than the observance”. (see also the “triple tragedy” discussion on p.100 of the April/May
2002 TTi) Police officers everywhere recognize that it would be folly, and perhaps even
increase the collision rate, to try to strictly enforce the dysfunctional limit in the face of
such massive ‘civil disobedience’. Instead they, again everywhere, adopt an ‘unofficial’
ticketing limit about 20kmph higher. Thus, to have the Autobahn ‘limit’ correspond to
the observed speed distribution in other countries, 140kmph was used for Germany in the
calculations as "that speed which most traffic is travelling 20kmph over".
With the PAST model residuals now also available (from the previous section) it
became possible to examine whether that 30kmph range of the GSL’s was correlated with
the observed difference in safety levels in these ten countries. Before going on to do that
regression the reader is reminded that, within an R2 range of 0 to 1, the higher numbers
indicate a strong correlation whereas 0 indicates the relationship is absolutely random,
i.e. no connection whatsoever between the two data sets. In this case the two sets are: the
GSL levels and the fatality rate residuals (i.e. those differences which have remained
unexplained after consideration of traffic volume, traffic density and seatbelt usage).
Fig. 5 - Motorway Speed Limits (GSL's, in kmph) an
PAST Model Residuals (as % +or - of fatality rate n
50.0
All 10 Countries
Linear (All 10 Countries)
40.0 Linear (w/o Sweden)

30.0
PAST Model Residuals

20.0

CDA FIN
10.0
Fig.5 (from F2004V297, corrected) shows the result of the regression calculationUK for USA
speed limit level against safety level (Note that, here again, a lower position indicates a
better safety level). For scientists the completely random result, even if Germany is
0.0
excluded (see note in graph), completely negates the theory that General Speed Limits
are beneficial80 to traffic safety and90
would normally need100no further comment 110
here. 120
However, some of the TTi readers may want to show this graph to politicians … and few
of the latter group have any scientific background. DK
-10.0 a few random observations (double-entendre intended), which might
Accordingly
prove useful in they political
= 0.2317x - 29.671
arena, should be made here: CH
1. Germany, without a GSL 2 (but making good use of TripleSL’s -Situation Specific
R = 0.0289
Speed Limits) lies exactly in the middle of the vertical range - not, as the ‘Speed Kills
-20.0
Theory’ would have it, at the top; NL
2. Canada (100) and Germany (140) are at the two ends of the GSL (horizontal) range but
have similar safety level (numerically, Germany is even slightly better);S
3. Switching to the vertical direction, and excluding Sweden (because of very low levels
-30.0
of ‘drink driving’), a 3-country group shows the best safety level (about 15% better than
average) and all have a GSL of 120kmph. If the Speed Kills Theory were correct the
others would be all above (worse safety) and scattered along a rising line to the right
(higher limits) of these three. Although the remaining 6 countries are in fact all above the
best -40.0
three, they are scattered along a level line and include the full range of the speed
limits (i.e. both higher and lower limits than the ‘best three’). It is perhaps timely to
General Speed L
remind the ‘political’ readers that, with reference to scientific theories, one-third right
equals all wrong.
In conclusion then … a valid theory has had a very practical result. The AMPS
Theory of ‘Accident’ Causation fully explains the observed relationship between traffic
density and traffic safety … and does that across the full range of traffic density. At the
same time the regression calculation itself shows that the ‘density effect’ explains about
70% of the observed differences in Motorway Fatality Rates between countries. The
density parameter, calculated as AADT, is thus revealed as second only to traffic volume
in determining how many of us will die in traffic each year.
Of even more practical importance, the 10-country set of regression residuals (from the
PAST/AADT model) has provided a never-before-available tool to explore which other
parameters might explain the remaining differences between countries.
As an example of the efficacy of that tool Fig. 5 (above) has been able to show that the
level at which the speed is limited (as a General Speed Limit) is absolutely irrelevant to
the observed differences in safety levels on motorways. Moreover, the presence of
Germany’s Autobahns in more or less the exact middle of the range of safety levels of the
ten countries means, to the scientific mind at least, that whether or not there is any GSL
is irrelevant to the safety level.

You might also like