Chance, bias and confounding - The INCLEN Trust

Chance, bias and confounding
• The observed statistical association between a
certain outcome and and the hypothesized
exposure could be a matter of chance
• Or it could be the result of systematic errors in
collection of data (sampling, disease and exposure
ascertainment) or its interpretation: the role of bias
• Or it could be due to the effect of additional
variables that might be responsible for the
observed association: the role of confounding
• Or it could be a real association

Confounder
• Is a factor that distorts the true relationship
between an exposure and the disease outcome on
account of its being associated with both the
exposure as well as the disease
• This distortion (over/underestimation) of the true
relation between exposure and disease can occur
only if this factor is unequally distributed between
the exposed and unexposed groups

Confounding
• A confounder is a third factor that is associated
with the exposure and independently affects the
risk of developing the disease
• It distorts the estimate of true relationship between
the exposure and disease: it may result in
association being observed when none in fact
exists; or no association being observed when a
true relationship does exist

Confounder
• A potential confounder must be predictive
of disease independently of its association
with the exposure under study
• This means that there must be an
association between the confounder and
disease even amongst the group unexposed
to the exposure under investigation

Confounder
• This third factor should not be merely an
intermediate step in the cause and effect
relationship between the exposure and the
disease outcome
• The association between the confounder and
the disease need not be causal. It may a
marker for for a risk factor other than the
one under investigation in a study.

Confounding
• Confounding can lead to the observation of
apparent differences between the study
groups when they do not truly exist, or
conversely, the observation of of no
difference when they do exist.

An example of confounding
• A number of observational epidemiological
studies have shown an inverse association between
the consumption of vegetables rich in β carotene
with the risk of cancer
• It is however possible that this association is
confounded by other differences between the
consumers and non-consumers of vegetables such
as fiber, which is known to reduce the risk of
cancer

Confounding: another example
• An observed association between the consumption
of coffee and the risk of MI could be due, at least
in part, to the effect of cigarette smoking, since
coffee drinking is associated with smoking , and
independent of coffee drinking, smoking is a risk
factor for MI
• The potential or true confounders are not always
as obvious as they are in the examples cited above

How to avoid confounding
• If a confounding factor does not vary between the
exposed and non-exposed, or those diseases and
non-diseased, then by definition, there can be no
confounding by that variable
• Thus if by design or analysis, the association
between disease and exposure is evaluated only
amongst those who are similar with respect to the
confounding factor, there can be no confounding

Controlling confounders
• Restriction of the study population
• Matching
• Randomization of exposure
• Stratification
• Multivariable analysis

Common confounders
• Age and sex are almost universal
confounders for all exposure – disease
associations
• This is because they are markers for a
whole lot of cumulative exposures. They
may not be causally related to disease, but
are markers for many other exposures
which might be truly related to disease.

Confounding: the intermediate
link
• Moderate consumption of alcohol is associated
with reduced risk of CAD
• HDL cholesterol also is protective for CAD
• Moderate alcohol consumption increases HDL
• If one controls for HDL, the association between
alcohol intake and the risk of CAD becomes weak
or statistically insignificant.
• Being an intermediate link between alcohol and
the risk of CAD, should HDL be considered a
confounder at all Should it be controlled

Positive and negative
confounding
• Tobacco smoking would be a positive
confounder in association between coffee
drinking and CAD
• The association between physical activity
and CAD would be negatively confounded
by gender, since women have lower risk of
CAD and they also exercise less than men.

Randomization
• Applicable only to interventional studies
• Most powerful method to control for
known, potential or unknown confounders
if the sample size is sufficiently large

Restriction
• Reduces the number of eligible subjects for
enrollment
• It limits generalizability of observations to
only the restricted population use for
drawing the random sample

Matching
• It includes elements of both design and analysis
• Mostly applicable to case-control study design
• It is expensive, difficult and time consuming
• By design, the effect of risk factor which has been
matched can not be studied
• Confounding is avoided not just by matching but
by special method of matched table analysis

Analysis
• Stratified analyses: Stratum specific estimates of
association are calculated, and the differences
amongst the strata are assessed by eyeballing, or
performing appropriate tests of statistical
significance
• Summary statistic for the pooled data is calculated
as per the method of Mantel and Haenszel
• The magnitude of confounding is assessed by
looking at the discrepancy between the crude and
adjusted estimates (without applying any tests of
statistical significance)

Confounding and effect
modification
• Confounding distorts the true relationship between
the exposure and disease and should be controlled
• Effect modification tells us that the association
between exposure and disease is modified by a
third factor. It should not be controlled for, the
magnitude of effect modification should be
reported and biological explanation for its
presence sought.

Bias
• The study must be designed and conducted in such
a manner that that every possibility of introducing
a bias is anticipated and steps are taken to
minimize its occurrence
• In spite of these precautions, the observed
association should be carefully examined to see if
it could be explained by bias.
• If indeed the study has elements of bias, it can not
be rectified at the stage of analysis (unlike
confounding)

Types of bias
• Selection bias: A particular problem in case
control and retrospective cohort studies
where both exposure and disease have
occurred at the time of selection of
individuals for the study
• Information bias

Selection bias
• Differential surveillance, diagnosis or
referral of individuals in the study: e.g.,
women using estrogen have uterine
bleeding more often, and seek medical
attention for this symptom. Hence they are
more likely to seek diagnostic evaluation
than those who are not on estrogens
resulting in more frequent diagnosis of
uterine cancer in women on estrogens

Multivariate regression analysis
• Several potential confounders can be
controlled; this is not easy in stratified
analysis
• It is an efficient method of data analysis
• Several models for regression exist. Choice
depends on the type of data to be analysed.