Biomedical Engineering Reference
In-Depth Information
The only secure basis for scientific progress is in the reproduction of
results, both by the original investigator and by others. Until several
different experiments have given similar answers, it is unwise to be
convinced. When they have, conviction may yet be reversed, but
it is reasonable. Experiments need to be independently repeated,
over and above what would be considered necessary from statistical
considerations alone.
Summarizing
In summary, data may be used for a variety of purposes over a wide
range of levels of confidence. Most particularly, there should be quite
different requirements for entertaining or postulating hypotheses -
where hints or trends are quite appropriate and for confirming theories
or implementing therapies or tests in general clinical practice where
the requirement of 95% confidence or better, and of conviction in the
sense used above, is entirely appropriate. We should always bear in
mind the level at which a result has been established, and be willing
to use it for purposes consistent with the confidence we can properly
have in it.
H
YPOTHESIS
T
ESTING AND
M
EASUREMENT
When, say, two therapies are compared, one may interpret the results
of an experiment in two rather different ways.
Let us assume that the outcomes in a trial comparing two therapies are
result
1
and
result
2
. (I leave until later the issue of what is
meant by a “result” in this setting.) First, one can consider the
experiment as a
measurement
of the difference in outcomes,
diff
,
where
diff
=
result
1
-
result
2
. A statistical analysis then
estimates the uncertainty in
diff
based on the uncertainties in
result1
(standard deviation =
SD
1
) and
result2
(standard
deviation =
SD
2
). This allows one to make a statement of the range of
values within which the true value
of
diff
is expected to lie, at
some stated level of confidence. This is the '
confidence interval
'. At
due to some experimental artifact; perhaps to an instrumental failure.
Had we not repeated the experiment, we would have been highly
embarrassed to have published a statistically significant but quite false
result. (Parenthetically, the original observation of a violation of
quantum electrodynamics was also wrong and the error was eventually
attributed to problems of instrumentation.)
