Confounding Variables in Epidemiologic Studies: Basics and Beyond

Abstract
This article discusses the importance, definition, and types of confounders in epidemiology.
Methods to identify and address confounding are discussed, as well as their strengths and
limitations. The article also describes the difference among confounders, mediators, and
effect modifiers.
Introduction
in epidemiology, like other fields of science, we look for causes of diseases, by which we
mean exposures that change the risk of diseases. For example, by “smoking causes lung
cancer”, we mean smoking increases the risk of lung cancer; lifetime risk of lung cancer is
17% in male smokers versus 1% in male non-smokers.1 Once we find that smoking causes
lung cancer, people are encouraged not to smoke and public policies are made. Whereas
causation always results in a change in risk, the converse is not necessarily true. Increased risk
of a health outcome in the presence of an exposure doesn’t necessarily simply a causal
relationship between the exposure and outcome. One reason for such non-causal associations
is the presence of a third variable called confounder or confounding variable. See the example
below.
Example 1: Some epidemiologic studies have found that poor oral health and/or tooth loss is
associated with an increased risk of esophageal cancer.2,3 But does this mean that poor oral
health causes esophageal cancer? Maybe yes. But maybe there are other factors (e.g., smoking)
behind the scene. Smoking causes poor oral health and it also causes esophageal cancer.
Therefore, an association between tooth loss (the exposure) and esophageal cancer (the
outcome) may be due to smoking (a confounder).

In this article, we discuss the following topics:

1) Criteria for confounding;
2) Types confounders;
3) Surrogate confounders;
4) stratification as a method to understand confounders;
5) Confounders versus other “third” variables (mediators and effect modifiers);
6) Confounding versus selection bias;
7) Confounding by indication;
8) How to identify potential confounders;
9) Methods used to address confounders;
10) Deficiencies of methods used to address confounders;
11) Overadjustment; and
12) How strongly can the confounders distort the associations.

In the final part, summary and conclusions, we tie these 12 topics together and provide a
framework for thinking about and handling confounders.

1. Criteria for confounding

A confounder is a variable that distorts the association between two other variables (the
exposure and the outcome). Often the exposure is what is being studied as a potential cause of
the outcome, such as tooth loss in Example 1. Statistical adjustment for the confounder results
in a change of relative risk. For a variable to be a confounder, it must have three
characteristics: 1) it must be associated with the exposure (causally or not); 2) it must be a
cause, or a surrogate of the cause, of the health outcome; 3) it should not be in the causal
pathway between the potential risk factor and outcome.4 See Example 2.
Example 2: Research shows that higher parity (mother’s number of pregnancies) is associated
with higher risk of Down syndrome. For example, on average, the tenth pregnancy is more
likely to result in a child with Down syndrome than the first pregnancy. However, we know that
this association is not because of parity, but it is because of the age of the mother, as the tenth
child is on average born to an older mother than the first child. In fact, the tenth child of a
mother who is 26 years old at pregnancy may have a lower risk of Down syndrome that the
first child born to a mother who is 39 years old. In this case age (the confounder) is associated
with both the exposure (parity) and the outcome (Down Syndrome) but doesn’t come in
between them (figure 1).

Figure 1. The association between parity and Down Syndrome is confounded by maternal
age. In this figure, double sided arrow denotes association (causal or non causal), one-sided
arrow means causal association, and the dashed arrow denotes a potential causal association
that is under investigation.
Example 3: Ginseng is an herb, mainly cultivated in China and Korea, which is used for
medicinal purposes. Some people believe that it can strengthen the body and prevent diseases.
A cohort study that investigated the association of ginseng with gastric cancer in China found
that, contrary to initial expectations, ginseng increased the risk of gastric cancer by 40%
(relative risk of 1.40).5 However, after adjusting for age, the association completely
disappeared and ginseng neither increased nor decreased the risk. In this case, age was the
confounder; older people were more likely to use ginseng and they were more likely to
develop gastric cancer (figure 2).

Figure 2. Increased risk of gastric cancer associated with ginseng intake is explained by age.

It is important to evaluate the above-mentioned three requirements before we consider a

variable as a confounder. Consider a study of alcohol consumption and breast cancer. Smoking
is not a confounder in this study. Smoking is related to alcohol consumption but not to the risk
of breast cancer.6 So it does not satisfy all the three requirements.
2. Types of confounders
X on founders may be classified into two categories Y qualitative and quantitative. After adjusting
for qualitative confounders, the association between exposure and outcome completely
disappears or even reverses direction, meaning that the quality or nature of the association
changes. In examples 2 and 3, the association disappeared after adjusting for age, which was
the confounder in both cases. See example 4 for a confounder that reverses the direction of
association. Unlike that for qualitative confounders, adjusting for quantitative confounders only
changes the magnitude of the association but not its nature. See example 5 below for a
quantitative confounder.
Example 4: Obesity, sedentary life-style, air pollution, and smoking all make life shorter. So,
why is it that people had a much shorter life span 500 years ago, when they were much leaner,
were more physically active, were breathing cleaner air, and smoked less? The answer lies in
the confounding effect of advances in modern life, such as better hygiene, and development
of vaccines and antibiotics. This is an example of confounding that reverses the real
association. Had we not adjusted for the advances, comparing now with 500 years ago might
have led to a conclusion that was exactly opposite the truth.
Example 5: The results of a cohort study in Iran showed that opium consumption was
associated with an increased risk of death with a relative risk of 2.26 (126% increased risk).7
One potential confounder was tobacco use, as tobacco users are more likely to use opium
(association with the exposure) and also more likely to die (association with the outcome). Age,
sex and other factors may also act as confounders in this association. In fact, after adjusting for
smoking, age, sex, and some other potential confounders, this association was less strong
(relative risk of 1.86, or 86% increased risk) but it did not disappear. Here, the confounders
resulted in only in a change in relative risk. Therefore, they are quantitative confounders.
Figure 3. The association between opium use and mortality is to some extent
confounded by smoking.

Quantitative confounders can further be classified into positive and negative confounders.
Positive confounders are those that magnify the association beyond its real size – i.e., make
the association seem to be bigger than it is. Negative confounders are those that make the
association seem to be smaller than it is. Adjustment for positive confounders results in a
relative risk that is closer to one, and adjustment for negative confounders results in a
relative risk that is further from one. Example 5 describes a positive confounder, as
adjustment reduced the relative risk from 2.26 to 1.86.
figure 4 summarizes the classification of confounders.

Figure 4. Classification of confounders.

While this terminology (positive versus negative confounders) is sometimes used in

epidemiology papers and textbooks,8 we introduce it here mostly to emphasize that
confounders can act in various directions. Learning the concept is more important than the
terminology. It is also important to pay attention to the magnitude of change in the relative
risk. It is one thing if after adjustment relative risk changes from 6.0 to 5.6, and another
thing if relative risk changes from 6.0 to 1.5, although these are both examples of
quantitative positive confounders.
3. Surrogate confounders
At times we cannot adjust for the causal confounder itself. In such cases, we may be able
to alleviate the problem by adjusting for a variable or a number of variables that together
act as a surrogate for the causal confounder. These are surrogate confounders. For example,
assume that wealth is a confounder in the relationship between a risk factor called R and an
outcome called O. Study participants may have not been asked about their wealth, but they
have been asked about their education, their residence zip code, and their profession. A
combination of these factors may work as a surrogate for their wealth. In the study of opium
and mortality (Example 5),7 adult height was adjusted for as a surrogate for socioeconomic
status during childhood.

4. Stratification as a different method to understand confounding

When a variable acts as a confounder, stratifying the results on the levels of the confounder
produces seemingly paradoxical results. The relative risks for each stratum may be different
from that seen for the overall association. In the examples below, we illustrate the effect
of confounding.
Example 6: Assume there is a rare congenital disease called congenia. We conduct a case-
control study recruiting mothers of 200 cases of congenia and mothers of 400 control children.
A potential risk factor is father’s smoking. The table below shows the results for this case-
control study.

Congenia Controls
Father smoker 140 160
Father not smoker 60 240

From this table, the odds ratio for the association between father being a smoker and congenia
is 3.50 (OR = (140 × 240) / (160 × 60) = 3.50).

However, a potential confounder may be mother’s smoking, as mother’s smoking is related to

this disease and men who smoke are more likely to have wives who smoke. When we stratify the
results by the two levels of mother’s smoking, in neither group do we see an association; i.e.,
in both groups the odds ratio is 1.00.
In smoking mothers (n = 300)

Congenia Controls
Father smoker 135 90
Father not smoker 45 30

OR = (135 × 30) / (90 × 45) = 1.00

In non-smoking mothers (n = 300)

Congenia Controls
Father smoker 5 70
Father not smoker 15 210

OR = (5 × 210) / (70 × 15) = 1.00

Whereas we introduce stratification as a method to better understand confounding, it can
initially confuse or “confound” the readers. It may take some time and practice before the
uninitiated understands this. One may ask “how is it possible that the overall OR is 3.00, but
when we stratify the results by whether or not the mother is a smoker, the odds ratio for each
group is 1.00? Figure 5 shows this phenomenon.

Figure 5. When there is confounding, the overall odds ratio is different from odds ratios in each
stratum of the confounder (here, mother’s smoking status).

This indeed does confound uss Confounding is part of what is called Simpson’s Paradox. The
results seem paradoxical, but there is no trick, and it is what happens.
One needs to know that when we stratify the results by the two levels of the confounder, it is
unlikely that the two ORs are exactly the same, but as long as they are not statistically
significantly different from each other, we usually take a weighted average of the two (using
various methods, such as Mantel-Haenszel method). This weighted average is the adjusted
odds ratio.
Example 7: The overall OR for the association between X and Y is 3.00. When we stratify
the results by the two levels of sex (men and women), the ORs for men and women are 2.20
and 1.90, respectively. Assume that these two numbers are not statistically significantly
different from one another (P value = 0.84). If the Mantel-Haenszel weighted average of these
two numbers is 2.10, the adjusted OR is 2.10.

5. Differentiating confounders from other “third” variables (mediators and effect

modifiers)
The exposure (the potential risk factor) and the outcome are the two main variables of each
association. A confounder is a “third” variable that affects the association of the exposure with
outcome. In addition to confounders, there are other “third” variables of interest that play a role
in an association. The two most important of such variables are mediators and effect modifiers,
and it is important to distinguish confounders from mediators and effect modifiers.

5.1. Mediators versus confounders

One of the characteristics of a confounder, in addition to being associated with the exposure
and the outcome, is that it should not be in the pathway between the two. If the third variable
is in the pathway, it is called a mediator or intermediate factor. See the example below.
Example 8: Poverty is a risk factor for many diseases including myocardial infarction, stroke,
diabetes, HIV/AIDS, esophageal cancer, and gastric cancer, to name a few. Let’s take the
example of poverty and diabetes. Is this association real, or is it confounded by an unhealthy
diet? If poverty leads to poor choice of diet or limited access to healthy food, then poverty is a
real cause of diabetes and unhealthy diet is a mediator. see figure 6.

Poverty  Limited access to healthy foodDiabetes

Figure 6. The association between poverty and diabetes is mediated via limited access to
healthy food.

In this case, unhealthy diet is associated with both poverty and diabetes. Also, adjusting for
unhealthy diet results in a change in relative risk estimates. But it is not a confounder because
it is in the pathway between poverty and diabetes.
A mediator is conceptually different from a confounder. A confounder may result in a non-
causal association between the exposure and the outcome, such that an exposure that doesn’t
cause the outcome is associated with it. Take Example 3. Taking ginseng will not decrease or
increase risk of gastric cancer; all of the apparent association is because of age. So ginseng is
not a real cause of gastric cancer. However, a mediator simply explains part or all of the reason
why the exposure causes that outcome. In this latter case, the exposure really does cause the
outcome. For example, poverty causes diabetes. If poverty is eliminated, then unhealthy eating
can be reduced, and thus risk of diabetes would be lower. See another example of a mediator
below.
Example 9: Having multiple sex partners is a cause of cervical cancer. This causal
relationship is mediated through exposure to human papillomavirus (HPV). see figure 7.

Multiple sexual partners Increased risk of HPV infectionCervical cancer

Figure 7. HPV mediates the causal relationship between having multiple sexual partners and
risk of cervical cancer.

Figure 8. If the effect of poverty on myocardial infarction is mediated via

three factors, after fully adjusting for one of these factors, the relative risk
shows the effect for the other two factors.

Figure 9. Obesity works both as a confounder and as a mediator in this relationship.

Should we adjust for mediators, as we do for confounders? The answer is that we can, but the
meaning of this adjustment is different. Before adjusting for the mediator, we have the total
effect of the potential risk factor on the health outcome, whereas after adjusting for the
mediator, we have the remaining effect of the risk factor after the partial effect of that mediator
is considered. See examples 10 and 11.
Example 10: Assume poverty results in myocardial infarction through three mechanisms:
eating more unhealthy food; increasing anxiety; and lower birth weight. Here we have three
mediating factors between poverty and myocardial infarction. If we do not adjust for birth
weight, then the relative risk (say 2.40) shows the overall effect of poverty on myocardial
infarction. However, if we adequately adjust for low birth weight, then adjusted relative risk
(say 1.60) will show the effect of poverty on myocardial infarction through the other two
mechanisms, i.e., higher anxiety and eating unhealthy food. see figure 8.

Example 11: A prospective cohort study of approximately 10,000 civil servants living in
London, England, found that those in the lowest socioeconomic position had a 60% increased
risk of total mortality (relative risk = 1.60) compared to those in the highest socioeconomic
position.9 After adjusting for several potential mediators, i.e., smoking, alcohol consumption,
physical activity, and unhealthy diet, the lowest socioeconomic group were only 14% at higher
risk of total mortality (relative risk = 1.14). Therefore, the authors concluded that a substantial
fraction of the effect of socioeconomic status on mortality is mediated via these factors.
While this distinction between confounders and mediators initially seems to be
straightforward, in reality, it may be very difficult to determine whether a variable acts as a
confounder or as a mediator. Often, it works as both, particularly when there are vicious or
virtuous cycles.
Example 12: We want to examine the association between family wealth and risk of overall
mortality. Should we adjust for education as a potential confounder? On the one hand,
education is associated with both wealth and mortality, and it may not be entirely in the causal
pathway. On the other hand, wealthier people are more likely to receive a better education.
Therefore, education may be both a confounder and a mediator. Please note that the relationship
between education and wealth is one of a virtuous cycle; one leads to the other, and vice versa.
Example 13: When investigating the association between physical inactivity and
cardiovascular outcomes, obesity can act partially as a confounder and partially as a mediator
(Figure 9). Obesity due to overeating can make one be less physically active. Also, physical
inactivity may lead to obesity and in turn to cardiovascular outcomes. The relationship
between obesity and physical inactivity is one of a vicious cycle; one leads to the other, and
vice versa. see figure 9.

5.2. Effect modifiers versus confounders

Effect modifiers are also third variables that affect the relationship between the exposure
and the outcome. A detailed discussion of effect modifiers is beyond the scope of this article.
However, we provide a brief treatment here. Effect modifiers are variables that modify the
strength of the association between the exposure and the outcome. stratification is a method
to identify effect modifiers. When we stratify the results of the association of a potential
risk factor and a health outcome by the two levels of the third variable, if the two relative
risks (or the two odds ratios) are statistically significantly different from each other we will
conclude that there is effect modification (interaction). see example 14.
Example 14: In the study mentioned in Example 5, the overall adjusted relative risk for the
association between opium and overall morality was 1.86. This association was stronger
for women (relative risk = 2.43) than for men (relative risk = 1.63); P value < 0.001. This
means that while opium increased the risk of death in both men and women, it did so more
strongly for women.
Please note that to learn about confounding, we compare the adjusted relative risk with the
unadjusted relative risk. In contrast, to learn about effect modification (interaction), we
compare the relative risks across strata. See examples 15 to 18 and the associated figures.
Example 15: In unadjusted analyses, X is associated with Y with a relative risk of 2.00. When
we stratify by sex, the relative risks are 3.00 for women and 1.00 for men (p-value for interaction
= 0.001), and the weighted average of these two numbers is 2.00. Here, the effect of X on Y
depends on sex; it increases risk in women but not in men. This is a clear example of effect
modification by sex. However, the average result is 2.00, after and before adjustment. So, there
is not much evidence for confounding by sex. see figure 10.
Example 16: In unadjusted analyses, X is associated with Y with a relative risk of 2.00. When
we stratify by sex, the relative risks are 1.40 for women and 1.50 for men (p-value for
interaction = 0.78), and the weighted average of these two numbers is 1.46. Here, the effect of
X on Y does not depend on sex; it increases risk in both women and men to almost the same
extent. So there is no effect modification by sex. However, the adjusted relative risk (1.46) is
different from the unadjusted one (2.00). This is a clear example of confounding by sex. see
figure 11.

Figure 10. Relative risks across strata of sex differ (effect modification by sex) but the
adjusted relative risk is the same as the unadjusted relative risk (no confounding).

Figure 11. Relative risks are similar across strata of sex (no interaction by sex) but the
adjusted relative risk is different from the unadjusted relative risk (confounding).

Figure 12. Relative risks are similar across strata of sex (no interaction by sex), and the
adjusted relative risk is similar to the unadjusted relative risk (no confounding).

Figure 13. Relative risks depend on sex (interaction by sex) and the adjusted relative risk is
different from the unadjusted relative risk (confounding).
Example 17: In unadjusted analyses, X is associated with Y with a relative risk of 2.00. When
we stratify by sex, the relative risks are 2.05 for women and 1.94 for men (P-value for
interaction = 0.66), and the weighted average of these two numbers is 2.01. Here, the effect
of X on Y does not depend on sex, so there is no interaction by sex. Also, adjusted and
unadjusted relative risks are similar, so there is no confounding by sex. see figure 12.
Example 18: In unadjusted analyses, X is associated with Y with a relative risk of 2.00. When
we stratify by sex, the relative risks are 1.62 for women and 1.00 for men (p-value for
interaction = 0.01), and the weighted average of these two numbers is 1.46. Here, the effect of
X on Y depends on sex, so there is interaction by sex. Also, adjusted and unadjusted relative
risks are substantially different, so there is confounding by sex. see figure 13.

6. Confounding versus selection bias

Some forms of selection bias, such as the difference between the exposed and unexposed in the
baseline of a cohort, can be alternatively classified as confounding.
Example 19: In a study of oropharyngeal cancer patients, White patients had a much better
survival than Black patients.10 This difference, however, was shown to be the result of higher
prevalence of HPV-induced tumors in White patients with oropharyngeal cancer. On the one
hand, this is confounding, because HPV-induced tumors are associated with being White and
concomitantly HPVinduced oropharyngeal tumors have better prognosis than other forms of
this cancer. On the other hand, this can be considered selection bias, as there is a systematic
difference in the type of tumor that Whites and “lacks have.
7. Confounding by indication
Confounding by indication is a form of selection bias. This term is used to describe a type of
confounding encountered in observational epidemiologic studies of drugs. Since in
observational studies the treatment is not dictated randomly – it is rather based on indication
for treatment (hence the name confounding by indication)
– those who take the drug may be substantially different from those who don’t with respect to
several characteristics. For example, those who have a severe disease may be more likely to
receive the treatment. Therefore, if one finds an association between treatment and higher
mortality, it may be due to confounding by indication rather than the adverse effect of the
drug. Confounding by indication is most often seen in for drugs that are rarely used but also
seen for commonly used drugs such as acetaminophen (Tylenol). Several excellent examples
are provided by Signorello et al.11

8. How to identify and select potential confounders

One of the major difficulties in epidemiologic studies, particularly in observational studies,
is to determine what the potential confounders are, or what to adjust for. There are multiple
methods to select confounders12–16 but these methods can be classified into two broad
categories: a priori selection methods based on our knowledge of the field, and selection
methods based on statistical analysis of the data.

8.1. A priori selection of confounders

In Example 5 (opium and overall mortality), there are a few obvious confounders that one
can determine a priori. For example, the locals know that in that particular area of Iran older
people and men are more likely to use opium, and those are the same people who are at
higher risk of death. So adjusting for sex and age is a must. Similarly, tobacco users are more
likely to use opium and more likely to die. So the need to adjust for tobacco use seems to be
obvious too. But it is unlikely that we know all the potential confounders or to have collected
data on all of them.
Associations vary geographically and change through time. Therefore, confounders may vary
by time and population, which may make a priori selection of confounders difficult. see the
example below.
Example 20: The results of a large prospective cohort study, published in the New England
Journal of Medicine,17 showed that coffee drinkers were less educated than non-drinkers and
coffee drinking was associated with a number of unhealthy behaviors, including smoking,
drinking large amounts of alcohol, physical inactivity, and consuming less fruits and
vegetables. Therefore, the association of coffee drinking with total mortality was confounded
by these factors. Whereas the unadjusted relative risks showed an increased risk of death
associated with coffee drinking, after adjustment the association changed qualitatively and
coffee was shown to reduce risk of death. However, this pattern could easily change in 20
years. Educated people are more likely to read the results of new research on health. If more
papers like this are published, perhaps in the future we will see that educated people are more
likely (rather than less likely) to drink coffee, and coffee drinking may become associated with
healthy habits.
This variation in pattern could pose a challenge for some observational studies. For example,
if in some countries tomatoes are heavily exposed to pesticides, while in others they are not,
assessing the effect of tomatoes on health could be different across countries because of
variation in confounding patterns. If the researchers have the data on all important
confounders, in theory, they can adjust for them. However, often this is not the case.

8.2. Statistical selection of confounders

In these methods, the researchers look into the association between a large number of
variables (as potential confounders) with both the potential risk factor and the health outcome.
If one finds an association with both, then that factor may be a confounder. In the example
of opium and mortality, one may explore socioeconomic status, body mass index, ethnicity,
marital status, and a number of other variables as potential confounders. A variety of statistical
criteria have been used to select the confounders, including change in estimates and statistical
significance tests.12 These methods are deficient in several ways, too. first, in addition to
confounders, mediators are associated with both the potential risk factor and the outcome, and
it may be difficult to distinguish between the confounders and mediators using statistical
methods. Second, testing for a large number of variables may lead to chance findings. Third,
there are always a number of unknown or unmeasured potential confounders. Fourth, there is
no universal agreement on which of these statistical methods performs best. fifth, it is difficult
to determine the cutpoint for considering a variable a confounder; 5%, 10%, and 20% change
in the estimates have all been used, but choosing between these is somewhat arbitrary.
The truth is that neither the a priori method nor the statistical methods work perfectly well,
and it is impossible to determine all the relevant confounders. However, we can do our best
through experience, educated guesses, and trying a combination of variables. Also, as
explained above, a combination of several surrogate confounders in the adjustment model may
work reasonably well.
Methods used to address confounders
Several methods are available to address confounders. Where possible, randomization with
large sample sizes is the strongest method to minimize the effect of confounders. As
discussed below, randomization with large enough sample sizes can balance the different
arms of the study for both known and unknown confounders. In observational studies, where
randomization is not possible, a variety of other methods are used to control for
confounders. These methods include matching, restriction, stratified analyses, and
regression methods. Some of these methods are used during the design of the study
(matching and restriction) and some during analysis (stratified analyses and regression
methods). This is not an exhaustive list; other methods, such as propensity score matching
may also be used.18 As discussed below, none of the methods used in observational studies
are entirely satisfying.

8.3. Randomization
Confounders are major issues in analyzing validity of observational studies, such as case-
control or cohort studies. In a cohort study of dietary vitamin E and cancer, for example,
individuals who take more vitamin E may be different from those who don’t in many ways.
In one such cohort study in the United States, those who took more dietary vitamin E were
more likely to be female, abstain from smoking, use less alcohol, be physically active, and
possess at least an undergraduate college degree.19
Compared to observational studies, randomized trials are less subject to confounders,
particularly when the sample size is very large. Given that the assignment of the exposure
in randomized trials is done at random, it is independent of all characteristics of the
participants, such as their age, sex, wealth, etc. Therefore, large randomized trials potentially
adjust for both known and unknown confounders. For example, in a large randomized trial
vitamin E and cancer, the group that received vitamin E was very similar to the group that
did not receive it with respect to age, body mass index, cigarette smoking, exercise, alcohol
consumption, aspirin use, parental history of cancer, and self-reported history of cancer.20
This list is only a sample of variables that are similar between the two arms of the study, and
one could be highly confident that the two arms are balanced for nearly all other
confounders.
However, we should add that small randomized trials are at risk of confounding, as the two
groups may be substantially different by chance.21 Nevertheless, if there are several such
small trials, a meta-analysis of them may effectively handle the problem. When it comes to
adjustment, for most practical purposes, trials with over 1000 subjects in each arm can be
considered large, those with fewer than 100 subjects in each arm may be considered small,
and those with 100 to 1000 in each arm are intermediate size.

8.4. Matching
In cohort studies, one can match exposed and unexposed individuals for the potential
confounder. For example, if we are assessing the effect of opium on total mortality and sex is
a potential confounder, one can match a male opium user to a male opium non-user and a
female opium user to a female opium non-user. This way users and non-users will be exactly
the same for sex, and thus sex could not confound the association. By extension, one can match
for more than one variable, such as by age and sex. For example a 56-year-old male opium
user can be matched to a 56-year-old male non-user. However, there is a practical limit to the
number of variables that we can match for; it is often difficult to find a good match based on
age, sex, ethnic group, education, wealth, total intake of fruits and vegetables, etc. This makes
matching a less useful method than regression (see below).

Box 1. The most commonly regression models in epidemiologic studies

The three most commonly used regression models in epidemiology are linear regression,
logistic regression, and Cox proportional hazards regression. Linear regression is used mainly
when the outcome is continuous (e.g., weight). Logistic regression is mostly used when the
outcome is binary (e.g., breast cancer; either the study subject gets breast cancer or not).
Extensions of logistic regression, such as polytomous logistic regression and ordinal logistic
regression, are also sometimes used. Polytomous logistic models are used for outcomes with
three or more categories treated as nominal variables, in which each category is compared to
a reference category. For example, risk of three different cancer types (lung, stomach, and
esophagus) can be simultaneously compared to a control group. Ordinal logistic regression
models are used when the outcome has three or more categories but is treated as an ordinal
variable. Cox proportional hazards models are used when the outcome is “time to event”, for
example time from entering the study to be being diagnosed with lung cancer. In a logistic
model, two people who got lung cancer during the follow-up contribute similarly to the
outcome, whereas in Cox models, if the first person got lung cancer in one year and the other
got lung cancer in 10 years, their contribution is different, as time to event differs. Measuring
time to event needs follow-up, therefore Cox regression is often used in prospective studies
with follow-up such as cohort studies and randomized clinical trials.
8.5. Restriction
In this method, we restrict the study population to one level of the potential confounder. For
example, if a researcher wants to study the association of opium use and total mortality and is
highly concerned about confounding by tobacco use, she can restrict the study population to
those who have never used any form of tobacco. This is indeed an extreme form of matching.
Restriction can also be done during analysis. The statistician can limit the analysis to a subgroup
of all study participants, such as to never-tobaccousers only. Problems with restriction are even
more severe than matching. It is difficult to restrict the study population to a group based on
several variables (age, sex, ethnic group, education, etc.), as sample size becomes small and
generalizability of the results will be limited. Again, using regression methods is favorable.

8.6. Stratified analysis

In this method, the statistician stratifies the analysis on different levels of the potential
confounder to examine if there is evidence for confounding. For example, if sex is a potential
confounder, the statistician can analyze the results separately by male and female. ”ike the two
previous methods, stratification has a practical limit for the number of variables chosen to
stratify. For example, if we choose sex (two levels: male and female) and race (four levels:
White, “lack, Asian, Other), the data need to be stratified into 4
× 2 = 8 strata. If we additionally stratify based on 10 categories of age, data will have 4 × 2 ×
10 = 80 strata. Again, this favors using regression methods. However, as explained above,
stratification is important in understanding confounding and to distinguish it from effect
modification.

8.7. Regression
Multipredictor regression models adjust for confounders by modeling the exposure and
potential confounders in relation to the outcome. These regression models estimate the effect of
the exposure while keep the levels of the confounder constant. For example, if the potential
confounder is sex, regression models act as if they estimate the effect of the exposure for men
and women separately and take a weighted average of the results. This is intuitively similar to
stratification but offers several advantages over stratification. See below. Depending on the type
of outcome, several regression models can be used. See Box 1.
Choice of method of adjusting for confounders
Where possible, randomization with large sample sizes is the most effective method for
dealing with confounders, as it balances the different arms of the study for both known and
unknown confounders. However, due to ethical and logistic problems, randomization is often
not feasible. For reasons mentioned above, the most commonly used method to handle
confounders in observational studies is using multi-predictor regression methods. These
methods are able to control for several confounders at a time and put relatively little
restriction on the study design and participant recruitment.
In theory, matching can be used in small randomized studies and cohort studies to control for
confounding, but this is rarely done in such studies. Matching is commonly done in case-
control studies. However, as Rothman and coauthors have shown, in case-control studies
matching may not only be ineffective in dealing with confounders, it may actually cause a
new form of confounding.4 (We understand that this is counter-intuitive. See the reference4 for
further information.) When the matching factor is strongly associated with the exposure but not
with the outcome (hence not a confounder), matching may cause confounding. Thus, it is
 Confounding Variables in Epidemiologic Studies

highly recommended that the matching factor be adjusted for in case-control analyses.4
stratified analyses may be illustrative at times and they can help researchers distinguish
between confounding and effect modification (discussed above).
Restriction in design is rarely used to control for confounders. When study participants are
restricted to a group, it is often for reasons other than confounding, such as for efficiency or
ethical reasons. Restricting the analysis to a certain subgroup, such as nonsmokers, which is in
fact a form of stratified analysis, is often done. In a study of oral health and esophageal cancer,
for example, the researchers restricted the analysis to never smokers and still found an
association.3 This was done to avoid any potential for residual confounding from smoking.
See below for further information on residual confounding.

9. Deficiencies of methods to address confounders

As mentioned earlier, other than randomization, other methods are not capable of adequately
handling confounders. One major problem with all of these methods is that there may be
several unknown or unmeasured confounders in each study. Another problem is that even
when confounders are known or measured, they may have been measured poorly, leading to
inadequate adjustment for them. This latter case is also called residual confounding. See
examples 21 and 22.
Example 21: Income is a potential confounder in a study and thus participants are asked to report
their income, but they fail to do so accurately. In this situation, errors in the recorded values of
income lead to imperfect adjustment for it, and hence residual confounding.
Example 22: Diet is usually measured using food frequency questionnaires in epidemiologic
studies. However, answers to these questionnaires are subject to substantial misclassification
and measurement errors.22 People barely remember how many tomatoes they have eaten each
week during the past year. Even if they do, their diet might have changed over the years. So,
the responses do not fully re–ect the life-long exposure to that dietary factor. Therefore, although
many epidemiologic studies claim that they have adjusted for dietary factors, that adjustment is
often quite inadequate. At times potential confounders are measured in broad categories, for
example, education may be measured as illiterate, elementary, middle or high school, and
higher education. This may also lead to residual confounding. Finally, poor modeling in
regression analyses may result in inadequate adjustment and residual confounding.
See the Figure 14.

Figure 14. Deficiencies in methods used to adjust for confounders may results in inadequate
adjustment.

10. Overadjustment
While we adjust for potential confounders, it may be counterproductive to adjust for too many
variables. First, if we adjust for mediators, rather than confounders, the adjusted result will not
be a correct estimate of the entire effect of the exposure on the outcome. Second, if one
predictor variable in the regression model is highly correlated with another predictor variable,
or a combination of variables, we may face a problem called collinearity. When collinearity
exists, standard errors of estimates will be very large and estimating the effect of collinear
variables will be very imprecise. To maintain precision, sometimes the statistical software
drops one of the collinear variables. Extreme cases of collinearity are unusual but they do
happen. For example, hemoglobin and hematocrit are highly correlated and therefore
collinear, and putting both in the regression model as predictors may cause a problem.
Likewise, assume a researcher wants to learn about predictors of infant mortality. If she puts
weight, height, head circumference, and abdominal circumference at birth plus gestational
age, there will likely be collinearity, as weight at birth should be strongly correlated with a
combination of the other four variables. Third, adding a large number of variables (that are
not real confounders) to the model for adjustment may slightly reduce power or make the
model unstable, particularly if some of the variables are categorical with sparse numbers in
some of their categories, and when the ratio of the number of variables included to the sample
size is large.

11. How strongly can confounders distort the association?

Confounders should be strongly associated with both the exposure and the outcome to have a
material effect on the relative risk estimates. For example, if the risk factor increases the risk
of the outcome by 2-fold (relative risk = 2.00), the confounder should be associated with a 5-
fold increased risk of both risk factor and outcome to completely negate the association after
adjustment. By comparison, in this same example if the confounder is associated with the risk
factor with a relative risk of 1.2 and with the outcome with a relative risk of 1.4, it is unlikely to
have a material effect on the relative risk; it may just change the risk from 2.00 to say 1.95. In
Example 6, after adjustment for mother’s smoking, the unadjusted OR of 3.50 for father’s
smoking and congenia changed to
1.00. This is a substantial change in OR. However, please note that the association between
mother’s smoking and congenia (OR = 21.00) and mother’s smoking with father’s smoking
(OR = 9.00) were very strong, both much higher than 3.50, otherwise we wouldn’t have seen
such a substantial change. These are shown in the tables below.
Congenia Controls
Mother smoker 180 120
Mother not smoker 20 280

OR = (180 × 280) / (20 × 120) = 21.00

Mother Mother not

smoker smoker
Father smoker 225 75
Father not 75 225
smoker
OR = (225 × 225) / (75 × 75) = 9.00

Siemiatycki and colleagues conducted an empiric investigation of occupational exposures

and various cancers to determine the effect of inclusion and exclusion of three potentially
 Confounding Variables in Epidemiologic Studies

important confounders, i.e., smoking, ethnicity, and socioeconomic status. Of the 75 OR’s
examined in this study, only eight OR estimates were distorted by more than 20%, of which
seven involved lung cancer, a disease very strongly associated with smoking. Therefore,
these investigators concluded that “relative risks between lung cancer and occupation in
excess of 1.4 are unlikely to be artifacts due to uncontrolled confounding. For bladder
cancer and stomach cancer, the corresponding cut point may be as low as 1.2”.23 While
not everyone may agree with this latter conclusion, the findings of this study corroborate the
fact that not-so-strong confounders are unlikely to change the results substantially.
The level of residual confounding depends on the number of unmeasured or incorrectly
measured confounders, their strength of association with exposure and outcome, the
prevalence of confounders, and the correlation among confounders. A simulation study
showed that under reasonable circumstances, for example when two independent
confounders each increase the risk of the outcome by 2-fold, unadjusted odds ratios of
approximately 2.00 can be generated (unmeasured or unadjusted confounders).24 This
study showed that under similar circumstances odds ratios of 1.50 can be generated even
after adjustment for the two confounders if the confounder is poorly measured (residual
confounding). However, the results of this study implied that odds ratios of 2.50 or higher
are unlikely to be due to confounding alone.
Change in the magnitude of associations due to confounding is a very important point in
discussions of epidemiologic findings. When the initial studies of smoking and lung cancer
were published in the 1950s, Ronald Fisher, a world-renowned geneticist and the most
prominent statistician of his time, argued that these associations may be confounded by
genes;25 some genes could cause you to smoke and those same genes could cause lung
cancer. Today we know that Fisher, who was a life-time smoker, was wrong in this case.
The relative risk for the association is between smoking and lung cancer can be as high as 30
for long-time smokers. If Fisher were to be right, then the relative risks for the association
between genes and lung cancer (and genes and smoking) should be substantially higher than
30. This is clearly not the case, as the relative risk for most common polymorphisms
associated with lung cancer is very low, around 1.25.26

Summary and conclusions

Confounding factors pose a major problem in identifying the real causes of diseases.
Confounders may erroneously increase or decrease the magnitude of an association, or even
invert the direction of the association. The steps outlined below may help in thinking about and
addressing confounders in epidemiologic studies.
Step 1. Do we need to be concerned about confounders? If the study is a large randomized
trial, confounding is not a major con cern. If it is not, then we need to identify, select, and
adjust for potential confounders.
step 2. How do we find the potential confounders? Potential confounders are found based on
a priori reasons or statistical reasons. Not all “third variables” are confounders. Assess whether
a variable meets the three criteria for confounding. If it does, then also consider how strongly
it is associated with the exposure and the outcome and how much it changes the relative risk.
Confounders should be strongly associated with both the exposure and the outcome to have a
material effect on the results. Otherwise, they may not be of major concern. When confounders
cannot be m When confounders cannot be measures and or are poorly measure we need to find
surrogates confounders and adjust for them. Step 3. How do we adjust for confounders?
Regression analysis is the most common method to adjust for confounders in observational
studies. We should use the most appropriate statistical regression method (e.g., linear
regression, logistic regression, Poisson regression, or Cox proportional hazards regression) to
adjust for confounders. This choice often depends on the type of the outcome. Step 4. How
do we interpret the results after adjustment? Even after adjusting for confounders, we need to
keep in mind that the results may not have been fully adjusted for due to unmeasured or poorly
measured confounders (residual confounding). However, often adjusting for surrogate
confounders may alleviate the problem. We should be reasonably cautious but not over-
critical. Whether or not confounders have been adequately dealt with is a matter of opinion.
But experience helps in making educated guesses about the presence and magnitude of residual
confounding.