Page 36 - Contributed Paper Session (CPS) - Volume 1

CPS694 Ordak Michal
I cumulative probability. The more independent comparisons of pairs of levels of
a given independent variable are made, the greater the risk that the statistically
significant result obtained is a matter of chance. If many tests are carried out,
each of which is affected by an error (e.g. p=0.05), the type I cumulative error can
be very large. To avoid this, an analysis of variance should be performed. When
a statistically significant result of the F test is obtained from the analysis of
variance, explanatory analyses are performed, i.e. a posteriori (post-hoc) or a
priori (contrasts). This makes it possible to determine which pairs of levels of a
given factor have statistically significant differences (McHugh, 2011). A large
number of people performing statistical analyses in medicine do not understand
why they often do not obtain statistically significant differences when they
perform explanatory analyses, but the situation is just the opposite when they
conduct a series of t-tests. The great number of statistical analysis and reviews
that I carried out between 2006 and 2018 led to one conclusion. Namely, there
was a tendency to publish results that could be type 1 cumulative errors.
Researchers admitted that they performed a series of over a dozen t-tests in
order to obtain statistically significant results instead of performing explanatory
analyses. This was to increase the chances of their articles being accepted for
publication in medical journals. My long term research in a group of 14,000
researchers (including physicians, graduate students, PhD students, PhDs and
professors) in various fields of medicine has allowed me to state that we have
reasons for concern. Namely, as many as 76% of respondents stated that they
did not know what type 1 cumulative error is; 46% of people admitted that they
often performed several or over a dozen t-tests instead of conducting an analysis
of variance. While 10% of them did so due to a lack of knowledge, others wanted
to increase the chance of obtaining a statistically significant result. Moreover,
52% of respondents chose a wrong post-hoc test. This is one of the reasons why
I have written this article. Its aim is to make medical researchers aware of the very
important problem, repeated testing of means, and thus obtaining results that
are a matter of chance.
Another problem that I observed between 2006 and 2018 is the growing
lack of knowledge among people performing statistical analyses in medicine
regarding the use of parametric or non-parametric equivalents of statistical
tests used. The high relevance of parametric tests is related to numerous
assumptions that should be met (Fagerland, 2012). One of the most common
questions researchers ask is: ‘Can I use a parametric equivalent if the
assumption about the normality of distribution is not met?’ Unfortunately, my
research indicates that many of the assumptions are poorly known and rarely
used. Although assumptions are not met, numerous authors use parametric
equivalents of statistical tests in order to increase the chances for their articles
to be accepted in peer-reviewed journals. Unfortunately, this is a big mistake.
One such assumption concerns the equinumerosity of groups studied and the

25 | I S I W S C 2 0 1 9