here are three good reasons for these requirements. First
and foremost, this ensures a high level of scientific
quality for papers published in EBR. Second, because
genetic engineering is politically charged, critics and
proponents must have the opportunity to evaluate
independently the quality of the available research.
Finally, the scientific community needs to be able to
build on the published literature. A sound meta-analysis
of the accumulated results of many publications requires
knowledge of the sufficient statistics of each of those
experiments (e.g., Arnqvist and Wooster, 1995). For
example, 5 non-significant results could combine into a
statistically significant result via meta-analysis, or
significant results might melt into non-significance under
the weight of multiple studies. Such meta-analyses would
be valuable contributions to the scientific literature.
There are two types of error in any experiment. Type I
error occurs if the null hypothesis is erroneously rejected
when in actuality it is true. Typically the Type I error rate
is 0.05, i.e., a 1 in 20 chance that the null hypothesis is
mistakenly rejected. This kind of error is routinely
handled by conventional scientific practices. Type II
error occurs when the null hypothesis is not rejected
when in actuality it should have been rejected. Negative
data suffers from the possibility of Type II error. Type II
error is problematic, because as scientists we have been
trained to minimize Type I errors and not be as concerned
with Type II errors. Type II error is measured by
statistical power. An experiment with high power has a
low Type II error rate and an experiment with low power
has a high Type II error rate.
In risk related problems, however, Type II errors can
be more serious than Type I errors (e.g., Hill and
Sendashonga, 2003). For example to answer the question,
what amount of GM-maize can be introduced without
harming a non-target species, a relevant null hypothesis
is that a certain quantity of Bt maize does not adversely
affect non-target species. If an experiment with low
statistical power were conducted, the probability of
rejecting the null hypothesis will be low, whether or not
the true effect had been biologically significant (Marvier,
2002). If Bt maize were introduced in quantity, this
kind of Type II error would result in adverse non-target
effects when none had been expected. Thus, for risk
related problems, Type II error must be considered
explicitly.
RETROSPECTIVE POWER ANALYSIS
A