You are on page 1of 1

Seminar: “Strategies to minimise bias in analysis of randomised trial evidence”

Date: June 14th, 2018


Student: Cintia Juliana Torres Galvis (Student 6324452)

Epidemiology studies the distribution and factors that determine events related to health. In order to
do this, epidemiology relies on studies conducted in defined populations. Since the results are the
major source of information to analyse, it is critical that these studies eliminate bias as much as
possible. In this seminar, the presenter invited to reflect on the question: “what does it mean to have
bias-free studies?”.

In individual randomised trials there are at least four sources of bias including selection, performance,
detection, attrition and reporting bias. However, one should not lose trust in studies when there are
signs of bias, since it is almost impossible to eliminate it in its totality. One can, however, rely on
studies that do not contain all types of bias in its design. Cochrane has developed a tool that allows
people to filter studies depending on the risk of bias. For researches that design and carry out the trial,
it is also important to know that there are ways to minimise this problem. In the case of selection bias,
the most valuable strategies are the allocation concealment and randomization. In the performance
and detection phases, the researchers can blind the participants and the researchers in the outcome
assessment. To solve (in part) the attrition problem, follow-up could be a solution to eliminate as much
of this bias as possible. Also, when publishing the results of the trial, it is important to avoid selective
reporting of outcomes.

As stated before, not all biased is only visible at the end of the study. When designing the study, it is
smart to predict that some changes will be done along the way and the fact that they could induce
some bias. Many studies do not report any changes in the original protocol and that is not realistic. At
a certain point when researchers have to make a lot of modifications in the trial design, they have to
ask themselves if it is valid to continue the study until the end. For instance, if some modifications
need to be performed, the unblinded statistician should not be involved in the decision of changing
the protocol or if the study should be stopped. Also, the researchers should also ask if the modified
questions are still relevant and plausible to interpret.

Another not-so-used option is to consider prospective meta-analysis. The aim of doing this is to specify
the research questions in advance to the results, thus overcoming the issue of selecting the trials
based on the outcomes that were obtained. Most of the studies for which this is possible is for
randomised controlled trials. In order to do this, researchers should register their trials so it is
accessible to learn the original research questions. The traditional way of conducting meta-analysis its
been retrospective in nature.

For those who make decisions based on trial results, for instance, global health organizations, official
health advisors, health ministries, regulatory agencies etc., it is important to consider that many
studies contain publication bias. It is known that studies that get published are more likely to have
positive results than those that did not get published. To reduce the impact of publishing bias, it is
smart to consider to analyse only from large scale studies. In order to fill the gap of those small-sized,
one can always use “funnel plots” which are graphical options to visualize their effect on the
conclusion. Also, the Rosenthal’s method is valuable when estimating the number of studies that
render a statistically significant result. To avoid concluding on false positives, some adjustment for
multiple comparisons and systematic review of all evidence should be carried out. Also, seeking
independent confirmation of the results and the analysis is important. Personally I consider that
precautions should be taken in order to minimise bias as much as possible. Now that the field of
epidemiology has advanced in many ways, there are many literature and tools available to do so.

You might also like