Middle East Fertility Society Journal, Vol. 10, No. 3, 2005, pp. 258-261
EVIDENCE-BASED MEDICINE CORNER
Bias in RCTs: confounders, selection bias and
allocation concealment
Abdelhamid Attia, M.D.
Professor of Obstetrics & Gynecology
and secretary general of the center of evidence-Based Medicine, Cairo University,
Egypt
Correspondences: Dr. Abdelhamid Attia, 18 El-Ghaith St., El-Agouza, Cairo,
Egypt. Email: amattia@link.net
Code Number: mf05044
The double blind
randomized controlled trial (RCT) is considered the gold-standard in clinical
research. Evidence for the effectiveness of therapeutic interventions should
rely on well conducted RCTs. The importance of RCTs for clinical practice can be
illustrated by its impact on the shift of practice in hormone replacement
therapy (HRT). For decades HRT was considered the standard care for all
postmenopausal, symptomatic and asymptomatic women. Evidence for the effectiveness of HRT relied always on observational
studies mostly cohort studies. But a single RCT that was published in 2002 (The
women's health initiative trial (1) has changed clinical practice all over the
world from the liberal use of HRT to the conservative use in selected
symptomatic cases and for the shortest period of time. In other words, one well
conducted RCT has changed the practice that relied on tens, and probably
hundreds, of observational studies for decades.
But what is the appeal
of RCT and why does it have such a place at the very top of the hierarchy of
evidence? It is because it is the least design of clinical research that can
be
affected by bias if it has been conducted properly. Conducting a RCT allows
investigators to control many types of bias that are hardly, if ever,
controllable in other study designs such as the non-randomized controlled
trials, cohort and case-control studies. Thus adequate knowledge of the
different types of bias that may distort a RCT results and how to avoid
them is mandatory for researchers who seek conducting proper research of high
relevance and validity.
What
is bias?
Bias is the lack of
neutrality or prejudice. It can be simply defined as "the deviation from
the truth". In scientific terms it is "any factor or process that
tends to deviate the results or conclusions of a trial systematically away from
the truth2". Such deviation leads, usually, to over-estimating the effects
of interventions making them look better than they actually are.
Bias can occur and
affect any part of a study from its planning phase to its publication. It
arises mainly due to the adoption of an inadequate design, misconduct of the
research methodology or the inadequate analysis of data. As research is
important for determining whether a new intervention is effective or not and
if
effective what is the magnitude of its effectiveness, bias is obviously
detrimental to research and hence to clinical practice.
There are
many types of bias that affect scientific research. Sackette has identified
35 types of bias2. However, there are major types that drastically affect the
conclusion of a study and there are others that are minor ones. Selection bias
is one of the major types of bias that can impair the results of a RCT but
due
to the nature of the design of a RCT it can, and should be, avoided.
Confounders
Research aims primarily
at measuring the association between two variables; an intervention (or
exposure) and an outcome. This can be achieved by designing a comparative
research with at least two groups; one receiving the intervention under
investigation (study group) and another either receiving a placebo or another
intervention (a control group). The outcomes in both groups are then compared.
But to study the effect of interventions properly one important pre-requisite
is that participants in both groups (the study group and the control one)
should be similar in all characteristics except for the intervention being
studied.
Suppose that a study has
been conducted to compare pregnancy rates in patients with anovulatory
infertility that are subjected to intrauterine insemination (IUI) after ovarian
stimulation either with FSH injections in one group or clomiphene citrate (CC)
in the other group. Thus, participants were distributed into two groups; a
study group that received FSH and IUI and a control one that received CC and
IUI. After 6 months of therapy there was a 32% pregnancy rate in the FSH/IUI
group and 24% in the CC/IUI group but during the analysis of data there was a
significant difference in the body mass index (BMI) between the two groups with
more obese females in the CC/IUI group than in the FSH/IUI one. The question
is: "Was the lower pregnancy rate in the CC group due to the inferior
effect of CC compared to FSH or due to the higher BMI that is known to affect
ovulation and pregnancy rates unfavorably?" We can not precisely know as
both factors could be responsible for the results of the research. The same
problem would occur if we found a significant difference in the mean age
between the two groups with older females in the CC/IUI group than in the
FSH/IUI one as age is inversely proportionate to fertility. So, if any factor,
that has an effect on the study outcome other than the one studied, prevails in
one group more than the other it would negatively affect the study result. Such
factors, as obesity and age in regard to fertility in our example, are known as
"confounders".
A confounder is defined
as "a variable, other than the one studied, that can cause or prevent the
outcome of interest." For any outcome in research there are many
confounders that should be considered in the planning phase of the trial,
reported in the results section (usually table 1 in the manuscript), and
analyzed for significant differences between the groups. Any confounding
variable should be equally distributed in the two groups to give balanced
groups. Some other examples of confounders are the effect of smoking, life style,
and dietary habits on bone mineral density and the frequency of sexual
intercourse, duration of sexual activity, and number of partners on cancer
cervix.
Selection
bias
Interferences from
researchers to divide patients into groups (select which patient goes to which
group) will result in dissimilar or unbalanced groups and would introduce bias
into the study. Such type of bias is known as "selection bias."
If investigators
"thought wrongly" that they can equally distribute or balance all the
basic characteristics and risk factors or confounders between the groups, they
definitely can not ensure balancing unknown risk factors or unknown
confounders. The best way of eliminating selection bias, then, is by
randomizing patients properly into groups.
Randomization
is achieved by using any method that gives every participant an equal chance to
be allocated into any of the study groups. In other words, after consenting to
participate in the study, every participant should have a 50% chance to be
allocated to the study group and 50% chance to be allocated to the control
group. Such randomization can be achieved by many methods as simple as coin
tossing or rolling a dice or better by using random numbers tables or computer
generated random numbers. A scheme is then chosen defining what numbers lead to
which group (randomization code or sequence). The most important is that once
the randomization method and sequence have been determined, they should never
be changed and randomization should never be repeated for the same participant
for whatever reason. The effect of randomization as a protection against
selection bias was studied in a Cochrane systematic review in which control
group patients in non-randomized controlled trials were frequently found to
have a worse prognosis than patients in the study group3. This, of course, lead
to exaggeration of the treatment effect of studied interventions.
Allocation
concealment
Unfortunately,
using a perfect randomization method alone does not ensure avoidance of
selection bias due to human interference in the procedure. Suppose that during
the conduct of the FSH vs. CC trial the researcher who is responsible for
randomizing patients into groups found an eligible patient (a patient who meets
inclusion criteria) who can not afford for the cost of FSH injections or who
refused to take injections and preferred oral CC. On randomizing her, the
randomization process directed her to the FSH group! Here the investigator
may try to help this participant and solve this problem either by excluding
her
from the study (and prescribing CC to her) or by repeating the randomization
method till she is directed to the CC group. Thus knowing the randomization
sequence or code that directs patients to the study or the control groups can
affect selection of patients and allows for re-directing them to desired
groups. Hiding the randomization sequence or code from those performing the
randomization achieves neutrality and ensures that the randomization process
is
properly applied and not repeated to direct certain participants to certain
group in the study. Hiding the allocation sequence from those performing
randomization is known as "allocation concealment". Here, after
randomization, the randomization code is sent with the patient name to the
principal investigator or better, for more neutrality, to a third party who
has the
randomization codes to decide whether this code directs the patient to the
study group or to the control one preventing it from being changed. Failure
to apply an adequate method of allocation concealment exaggerates treatment
effect
by 40%4. Thus, allocation concealment is another important pre-requisite in
RCTs to prevent selection bias.
There are methods of
randomization that look perfect but in reality can lead to bias as they can
never be concealed. Such methods lead to pseudo-randomization or what is known
as quasi-randomization. The use of hospital admission numbers, date of birth or
day of enrollment into the study as a method of randomization is inadequate as
the randomization sequence can not be hidden in such situations and patients
can be excluded from the study based on the knowledge of their group assignment
or can easily be re-directed to another group. Using such methods makes the
trial falls to the category of non-randomized controlled trials.
Moher et al, in 1998,
reported that allocation concealment was reported in less than 10% of articles
describing RCTs published in prominent journals in five different languages5.
Thus randomization and concealment of the randomization sequence became
pre-requisites in the CONSORT statement that aims for improving the reporting
of randomized controlled trials enabling readers to understand a trial's
conduct and to assess the validity of its results (6).
CONCLUSION
In conclusion, selection
bias is detrimental to randomized controlled trials. To prevent selection bias,
investigators should anticipate and analyze all the confounders important for
the outcome studied. They should use an adequate method of randomization and allocation
concealment and they should report these methods in their trial. Editors and
peer reviewers should enforce the importance of use and reporting these methods
before accepting RCTs for publication.
REFERENCES