Professional Documents
Culture Documents
Bayesian Updating Explained - 1547532565
Bayesian Updating Explained - 1547532565
Bayesian Updating Explained - 1547532565
Recipients: XXXXXXX
nd
Date: 22 November, 2000
a. Estimation:
a. Under estimation of supplier costs
b. Under estimation of delays
c. Underestimation of cost-over-runs
d. Overestimation of probabilities of unlikely events
e. Under estimation of very large impact £ values
This discussion paper looks at one of the problems areas – estimates in conflict with
the views of the quantity surveyors. However, the other areas, mentioned above,
have all been subject to extensive academic study and I will try, subject to time and
motivation, to write something about them for them.
If you have time please send me comments on any of these issues as well as about
Bayesian Revision.
To tackle the problem of estimation error, one possible solution is the use of
‘Bayesian Revision’ to bridge the gap between the differing perspectives in
determining the necessary contingency funding.
The National Audit Office report that concludes that all major government
sourced civil/engineering projects have cost over-runs of 40% on average.
A consortium of companies assess that the impact of cost over-runs from all
sources will only amount to 15% at worst on this government project.
The QSs state that in their experience of similar projects the cost-over-runs
have been between –5% to 29% and averaging 20%.
Sample mean of m = 20
By Bayes Theorem:
{ 9 x 1/4 x 20 1/10 x 259 x 1/4 1/10 ) ; 9 x 1/4 + 1/10} x { 20, 259 x 1/4
x 1/10) / (9 x 1/4 + 1/10)}
{ 47.252.35 ) ; 2.35} x { 20, 252.25 x 0.1) / (2.25 + 0.1)} {20, 25, .225
/ 2.35}
= N (20.21, 0.43)
4. Confidence Interval
Bayes was the son of a non-conformist clergyman, and he went into the
ministry, ending his life at the Presbyterian Chapel in Tunbridge Wells,
Kent.
But after his death in 1761, Richard Price, a friend, found among his
papers “Essay towards solving a problem in the doctrine of chances”,
work produced as a result of a curious hobby – he used to while away his
time in the local court trying to assess whether a defendant was guilty on
the basis of his appearance and how the subsequent evidence altered
this perception. It was published by the Royal Society in 1763 and
became the basis of a statistical technique, now called Bayesian
estimation, for calculating how true a proposition is likely to be. Unlike
classical statistics, it enables prior judgement to be factored into the
equation.
His conclusions fell into disuse earlier this century, and Bayesians have
always felt despised by orthodox statisticians because of their guesswork.
However, now they are growing in numbers and confidence. Bayesian
methods have been used in court cases, in analysing the results of drug
trials and in improving the service of banks.
Put simply this means that whatever our initial knowledge state the
addition of new information will change our perceptions in some way to a
new altered state – our posterior belief. In risk management terms we can
use the Bayesian approach to produce a risk weighting of new
information using the precision (reciprocal of the standard deviation) to
derive a new perspective from impact of diverse data sources.
C
A B
P ( A B) = P ( A ) x P ( B ) = P ( B ) x P ( A )
Now including the conditional notation “ | “ which means ‘given that’ or ‘conditional
on’ and “, C “ where C represents the state of the world and need not be included in
the formula because it is implicit.
P ( B A, C) = P ( A | B , C ) x P ( B | A , C ) = P ( B | A , C ) x P ( A | B , C)
P ( B | A) = P ( B A ) / P ( A )
P ( B | A ) = P ( A | B ) x P (B) / P ( A )
7. The general example 2
Here is another quite simple way of doing the calculation that is given in Statistics: A
Bayesian Perspective by D. A. Berry. (It doesn’t need sample sizes but it does
require that you have some idea of the variance attached to your ‘beliefs’ – where
might it come from and how might it be elicited from respondents?) It is in fact just a
weighted averaging process using the reciprocal of the variance i.e. the precision.
2
Prior precision: c0 = 1 / h 0 = 0
Sample Precision: c = n / h2
(I have had some problems with this method when there are more than
three sources of new data – the variance and the mean don’t change as
expected; I have yet to work out what is going on.
What happens when the new data increases the uncertainty e.g. In a Court
case there is some evidence that is we believe to be compelling, the
Council for the plaintiff will present it forcefully so that the judge and jury
increase their belief in it, whereas the council for the defence will do the
reverse. In addition the quality of the respective QCs in their presentation,
of which we have had previous experience, will affect the perception of the
evidence by virtue of their performance.
In the influence diagram below I have tried to capture the general risk/business
process, however, for the purpose of entertainment I have included two
hypothesised authoritative risk research studies that may influence the way the
Lenders’ Technical Advisors (LTA) due diligence team view Metronet’s risk
assessment.
c. The consultant has worked on 6 rail projects and has observed that cost
over-runs amounted to 23% on average.
d. Assume that Metronet have allocated 10% contingency for all cost over-
runs.
In the attached spreadsheet the calculation is made for the Bayesian revision.