Dear colleagues
I don't want to extend this debate unnecessarily in the absence of
the information needed and I prefer to recommend a careful reading of
the discussion of these issues and others in Professor Hauer's book.
I agree with Professor Alsop that regression to the mean (RTM) is
almost certain to be a factor to be taken into account when high
accident rates have influenced the decision to make an
intervention. But that does not mean that it need not be considered
and accounted for where that was not the prime motive.
The subjective decision on intervention because of high accident
rates is highly indicative of a potential problem, but comparison
with other similar sites is needed to show whether regression to the
mean is likely to be a significant factor or not.
Regression to the mean can work both ways - unusually "safe" sites
can deteriorate for reasons of statistical fluctuation despite an
intervention with a potentially positive effect.
Professor Alsop's last comment about other sites is helpful and
absolutely right. To avoid any misunderstanding, I would only add
that it is the existence of data for other sites which are
sufficiently similar so that a similar intervention would be feasible
which is important, not the subjective matter of whether they were or
were not considered. I suspect that Dave du Feu's comments on the
Kensington scheme and its benefits relate to this point, but I cannot be sure.
I will drop out here. Once again, I recommend reading Hauer, who
considers these and other factors apart from RTM which may be more
important in Dr Bullas analysis.
Best regards
Robert Cochrane
At 15:26 10/01/2010, Richard Allsop wrote:
>Did the numer of accidents or casualaties for the months in
>2005-2007 influence the decision to make the intervention at this site?
>
>If so, you need to allow for regression to the mean if you can, or
>recognise that your result is subject to it if you cannot. If not,
>you do not need to allow for regression to the mean.
>
>If the result is subject to regression to the mean, then to allow
>for it using the empirical Bayes method, you need an estimate of the
>distribution of accident or casualty numbers in the before period at
>sites sufficiently similar to yours to have been candidates for the
>sam intervention if they had had enough accidents or casualties in
>the before period.
>
>If you do not have such an estimated distribution, you cannot
>estimate the effect of regression to the mean. This is not always
>recognised by those who advocate applying the empirical Bayes method.
>
>Richard Allsop
>
>
>Dr John C Bullas wrote:
>>As an earth scientist and road surface specialist I am outisdeof my
>>comfort zone with accident statistics
>>I have data for values for KSIs and slights for the months September
>>to December for 2008 when an
>>intervention was in place and for the same months for 2005-2007
>>when it was not
>>I believe since I cannot show the data is normally distributed, the
>>wilcoxon rank sum test might be
>>the best measure of whether 2008 is significantly lower than the other
>>years (as a group)
>>I do not have control data to hand nor traffic flows so will have to
>>state an assumption
>>Is this test (aka MANN -WHITNEY 'U' ?) a good choice?
>>Dr B
>
>--
>Richard Allsop
>Centre for Transport Studies
>University College London
>Gower Street
>London
>WC1E 6BT
>email [log in to unmask]
>www.cts.ucl.ac.uk
Robert A Cochrane
7 Lawn Terrace
Blackheath
London SE3 9LJ UK
Tel (44) 020 8297 1978
Mob (44) 07764 197 701
|