Conclusions
| Home | | Pharmacovigilance |Chapter: Pharmacovigilance: Statistical Methods of Signal Detection
The validity of the method has been demonstrated. It detects existing problems; it finds new signals and prioritises them for the benefit of assessors; it is very simple, transparent, and objective; and it can be auto-mated very easily indeed.
CONCLUSIONS
The
validity of the method has been demonstrated. It detects existing problems; it
finds new signals and prioritises them for the benefit of assessors; it is very
simple, transparent, and objective; and it can be auto-mated very easily
indeed.
The
‘Bayesian Data Mining’ approach used by the FDA (DuMouchel, 1999) is very
similar, but offers a better statistical analysis when very small numbers are
involved. It emphasises ranking of the equivalent of the logarithm of PRR. It
uses an ‘Empirical Bayes’
method,
which shrinks log towards zero, and the shrinkage is important if E is small,
but gives very similar results when observed or expected numbers of reactions
are reasonably large. It is slightly more complex in the calculations, and
consequently less transparent.
The
WHO has a new approach (Bate et al.,
1998) based on a Bayesian confidence propagation neural network, but again is
very similar to a PRR. It uses the log (to the base 2) of the PRR based on the
same 2 × 2 table as used
with PRRs. Its use of Bayes’ theorem in a 2 × 2 table is not controversial and
does not incorporate prior belief. The cut-off for a signal is based on the
confidence interval around their statistic. The method has the ability to scan
the whole database relatively rapidly, forming all tables for combinations of
drugs and reactions that occur together and is used routinely.
The major issues are the potential
for misinterpre-tation of the signals and over-reliance on automation. The
statistical methods are a first stage of assessment, and careful evaluation
using medical scientific knowl-edge is still required. At the same time, the
potential contributions of statistical methods and of statisti-cians have not
been fully realised. The advances in the past few years seem to have been greater
than since the mid-1980s, although it is recognised that there has been some
re-invention. Further statistical creativity is possible, particularly in
integrating spon-taneous reporting with epidemiological methods and randomised
trial data.
Related Topics
