ZORANA GLEDOVI´C
Institute of Epidemiology, School of Medicine,
University of Belgrade, Belgrade, Serbia
gledovic@sezampro.yu
Synonyms
Bias: Systematic error; Confounding: Bias due to confounding;
Interaction: Effect modification
Definition
The important issues in deriving causal inferences are:
bias, confounding and interaction.
“Bias can be defined as deviation of results or inferences
from the truth, or processes leading to such deviation”.
“Confounding is a situation in which the effects of two
processes are not separated”.
“Interaction is the interdependent operation of two or
more causes to produce or prevent an effect” (Last
2001).
Basic Characteristics
The goal of an epidemiological study is  accuracy
in measurement. Epidemiological studies are prone to
 error. Errors can be either random or systematic.
Since the errors can never be eliminated,much attention
is devoted to minimize them, and to assess their importance.
The principles of study design emerge from consideration
of approaches to reduce both types of errors.
Random error is the divergence, due to chance alone,
of an observation on a sample, from the true population
value, leading to lack of precision in themeasurement
of an association. Sources of random error are: individual
biological variation, sampling error and measurement
error (Bhopal 2002). Random error can never be
completely eliminated. The best way to reduce it is to
increase the size of the study.
Validity
The validity of a study is usually separated into two
components: internal validity and external validity.
Internal validity implies that the index and comparison
groups are selected and compared in such a manner, that observed differences between them on dependent variables
under study may be attributed only to the hypothesized
effect under consideration.
External validity concerns validity of the inferences as
they pertain to people outside the study population.
External validity depends on internal validity, which is
its prerequisite, but it depends also on the results of other
studies, theoretical knowledge of the disease process
and related factors, and biological considerations.
Internal validity is the degree to which the results of
an observation are correct for the particular group of
people being studied. Internal validity can be threatened
by all sources of systematic error (bias) but can
be improved by good design of a study (Rothman and
Greenland 1998).
There are three main types of biases: selection bias,
information bias and confounding.
Selection Bias
Selection bias occurs when there is a systematic difference
between characteristics of the people selected
for a study and the characteristics of those who are not.
There are a number of reasons for the occurrence of
this type of bias. Common feature for all of them is that
the relationship between the  exposure and disease
observed among those who participate in the study is
different from that for the individuals who would have
been eligible to participate but were unwilling or not
selected by the investigator (Rothman and Greenland
1998).
Example: Selection bias can occur if investigators
include hospital cases or cases under a physician’s care
and exclude those who die before admission to hospital
because the course of their disease was severe, those
with mild symptoms not requiring hospital care, cost of
hospital treatment or other factors.
Information Bias
Information bias occurs whenever the study subjects are
erroneously categorized with respect to either exposure
or disease. The effect of this bias depends on whether
this misclassification is differential or non-differential
(Rothman and Greenland 1998).
Differential misclassification occurs when the proportion
of subjects misclassified differ between the study
groups. It can occur when there is any systematic difference
in the soliciting, recording or interpreting of information
from study participants.
Example: Mothers whose children have had or have
died of leukemia are more likely than mothers of
healthy children (control group in a  case-control
study) to remember details of diagnostic x-ray examinations
to which these children were exposed in utero
(recall bias).
The effect of differential misclassification is overestimation
of an association even if it does not really exists,
or underestimation or lack of an association when it
really exists.
Non-differential misclassification occurs when inaccuracies
in the categorization of subjects by exposure
or disease are present in similar proportion in each
of the study group. Such misclassification is often
present because of inaccuracy of most measurements
in biomedicine.
Non-differentialmisclassification almost always results
in an underestimate of the true strength of the association.
Some degree of this misclassification is present
in almost all types of epidemiological studies and this
bias may account for some apparent differences in the
results of epidemiological studies (Hennekens and Buring
1987).
The control of potential biases must be accomplished
by careful study design. Some of design features that
can minimize potential biases are: carefully prepared
questionnaire (close-ended questions), clearly written
protocol, trained study personnel, the use of multiple
sources of data whenever possible etc.
Confounding
The word confounding is derived from a Latin word
meaning to mix up. The word’s meaning in everyday
language is to confuse or puzzle. Confounding mixes
up causal and non-causal relationships.
Confounding is a major cause of bias in epidemiology,
and the more difficult one to understand. The
potential for it to occur is whenever the cardinal rule’
compare like-with-like’ is broken. This rule is perhaps
never attained except in experimental research. Comparing
like-with-like may be achieved in experimental
studies where subjects can be randomly allocated to
one group or another, a technique which employs the
laws of chance to create comparable groups (Bhopal
2002).
Confounding is one of the most important problems in
observational studies.
Example: In a study of mortality rates, investigators
find that mortality rates in an English seaside resort are
much higher than in a country as a whole. Why might
this be so?
One possible explanation:
A holiday town attracts the elderly, so has a comparatively
old population.
What is confounding factor in this example?
Age, which is associated with both living in a resort and
with death.
The Control of Confounding
Several methods are available to control confounding,
either through study design or during the analysis of
the results (Hennekens and Buring 1987; Rothman and
Greenland 1998)
The Control of Confounding in Study Design
 Randomization is applicable only in experimental
studies. It is method which ensures that potential confounding
variables are equally distributed among the
groups being compared.
 Restriction is used to limit the study to people who
have particular characteristics. For example, in a study
on the effects of coffee on pancreatic cancer, participation
in the study could be restricted to nonsmokers,
thus removing any potential effect of confounding by
cigarette smoking.
Matching ensures that study participants are selected
so that potential confounding variables are evenly distributed
in the groups being compared. For example in
a case-control study each patient with a disease can be
matched with a control of the same sex and age group to
ensure that confounding by sex and age does not occur.
Control of Confounding in the Analysis of Results
 Stratification involves the measurement of the
strength of association in well-defined and homogeneous
categories (strata) of the confounding variable.
If the confounding variable is age, the association may
be measured in 10-year age groups. Stratification is
often limited by the size of the study and it cannot help
to control many factors simultaneously. In this situation,
mathematical modeling is required to estimate the
strength of the associations while controlling for a number
of confounding variables.
The multivariate modeling involves  logistic model
and analysis of covariance (Rothman and Greenland
1998).
Interaction
According to MacMahon interaction can be defined as
follows: “When the  incidence rate of disease in the
presence of two or more risk factors differs from the
incidence rate expected to result from their individual
effects” (MacMahon 1972). The effect can be grater
than that we would expect (positive interaction,  synergism)
or less than what we would expect (negative
interaction,  antagonism).
The problem is to determine what we would expect to
result from the individual effects of the exposures.
In exploring the possibility of interaction, the first question
is whether an association between exposure and
a disease exists. If it exist, is it due to confounding? If
it is causal, is it equally strong in each of the strata that
are formed on the basis of some other variable? (Gordis
2004).
Example: If the association of smoking and lung cancer
is equally strong in all strata formed on the basis
of degree of urbanization, there is no interaction. But if
the association is of different strength in different strata
formed on the basis of age, there is interaction.
Conclusion
Biases reflect inadequacies in the design or conduct of
a study and affect its validity. Because of that, biases
need to be assessed and, if possible, eliminated, while
confounding and interaction describe the reality of the
relationships between certain factors and a certain outcome
(Gordis 2004). Such relationships are particularly
important in investigating the role of various factors in
disease causation.
Cross-References
 Accuracy
 Antagonism
 Case Control Studies
 Error
 Incidence Rate
 Logistic Model
 Matching
 Precision
 Questionnaire
 Randomization
 Restriction
 Stratification
 Synergism
References
Bhopal R (2002) Concepts of Epidemiology: An integrated introduction
to the ideas, theories, principles and methods of epidemiology.
Oxford University Press, Oxford
Gordis L (2004) Epidemiology, 3rd edn. Elsevier-Saunders,
Philadelphia
Hennekens CH, Buring JE (1987) Epidemiology in Medicine.
Little, Brown and Company, Boston, Toronto, pp 4–8
Last JM (2001) A Dictionary of Epidemiology, 4th edn. Oxford
University Press, New York
MacMahon B (1972) Concepts of multiple factors. In: Lee DH,
Korin P (eds) Multiple Factors in the Causation of Environmentally
Induced Disease. Academic Press New York, New
York
Rothman KJ, Greenland S (1998) Modern Epidemiology, 2nd
edn. Lippincot-Raven, Philadelphi

Tags: , , , , , , , , , , , , , , , , , , ,

Related posts