Welcome to this presentation focused on effect modification.

1
After viewing this module, you will be able to
Discuss the distinction between confounding and effect modification
Use stratification to evaluate potential effect modifiers

2
We begin our discussion on effect measure modification with a few definitions and
examples.

When the measure of effect or measure of association changes over the value of
some other factor we say that that effect is modified or in other words we have
effect modification. You may also see the term “interaction” used to describe effect
modification.

As an example, a drug treatment effect on a particular health outcome may be

stronger in males than females, in which case, gender is acting as an effect
modifier. Or if we consider driving and alcohol consumption as risk factors for injury,
alcohol is likely to increase the impact of driving on the risk of injury. Alcohol is said
to modify the association between driving and injury.

Don’t confuse this with confounding. Effect modification doesn’t distort the data like
confounding.

Consider the example provided on this slide, if an association is stronger in older

people than in younger people, we say that age is acting as an effect modifier.

3
We defined effect modification in terms of measures of association such as odds
ratios or risk ratios.

But we want to point out that we can also think of interaction in terms of incidence
rates and comparisons of incidence rates such as cumulative incidence or incidence
density.

Another definition of effect modification is 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. For example, the effects are synergistic.

Another way to think of effect modification is that an interaction occurs when there is
a difference in the effect of driving on injury depending on the level of another factor,
such as alcohol.

4
Consider a numeric example of effect modification. In this study, investigators studied the
impact of smoking on deaths from lung cancer among individuals with and without
asbestos exposure.

We see that the risk of death from lung cancer is increased with cigarette smoking among
those with no asbestos exposure, but is even more dramatically increase with cigarette
smoking among those with asbestos exposure.

5
We can also assess the presence or absence of the interaction by examining
the association graphically.

In this example, we see the incidence of an outcome graphed according to

levels of exposure (positive and negative). In addition, we have separate risk
lines for men and women.

Let’s say that our outcome is incidence of high blood pressure and our
exposure would be something like obesity. The graph then summarizes the
incidence of high blood pressure for obese and non-obese men and women.

In this example, the two lines for each of our subgroups are parallel. This
means that the difference in the incidence of high blood pressure when
comparing obese to non-obese individuals is the same regardless of gender.
Even though we have different starting points if we were to calculate the
attributable risk or risk difference for men, we would have an attributable risk
among obese men of 30 (calculated as 60-30). Similarly, among women, we
would have an attributable risk of 30 (calculated as 40 – 10). Therefore, the
risk difference is the same regardless of the value of our third factor gender
so this would be a graphical example of no interaction.

6
Now, let’s consider the same relationship between the incidence of an
outcome and exposure status but now we consider subgroups defined by
age.

If we were to take a look at this relationship by age we can see that in this
example the lines are not parallel. What does this mean?

If we were to calculate the attributable risk for the older age group, we would
have an attributable risk of 60 (calculated as 90 minus 30) as compared to
the attributable risk among the younger age group which would be 40-10, an
attributable risk of 30.

We observe an attributable risk that is higher when our study subjects are in
the older age category.

In this example, we are observing a difference in the risk of the outcome in

the exposed and the non-exposed group depending on a third factor, age. In
this case, we see a greater difference among the older age category
compared to the younger age category. This is an example of the presence
of interaction.

7
We could say that the incidence of high blood pressure or the cumulative
incidence of high blood pressure differs between those who are obese and not
obese and it differs to a greater extent among those who are older compared
to those who are younger.

7
In this example we have a more extreme situation where our lines are
crossing or intersecting. So what's happening here?

This is another example of the presence of interaction. If we again consider

an outcome of high blood pressure now consider age to be the primary
exposure. We see that among the non-obese (non-exposed), the incidence
rate is slightly higher for young individuals than old individuals while among
obese (exposed), the incidence rate is higher among older than younger
participants. This is an example of an interaction. The effect of age on the
risk of high blood pressure differs depending on the obesity status of the
individual.

8
So just like we were able to describe the direction of confounding as either positive
or negative, we can also describe the types of interaction as either positive or
negative but we use the term synergistic and antagonistic.

A synergistic effect occurs when the effect modifier is accentuating the effect of the
exposure of interest. So the example we gave earlier concerning drinking and
driving on the risk of injury would be a synergistic effect. The effect of alcohol is
accentuating the effect of driving on risk of injury.

An antagonistic effect however would occur when the presence of the effect
modifier is diminishing or eliminating the effect of exposure. An example of this
would be the effect of folic acid levels during pregnancy on the effect of a certain
mycotoxin on the risk of neural tube defects. A mycotoxin called Fumonisins is a
type of mold that contaminates corn. It can be present in corn-based products like
tortillas and has been associated with an increased risk of neural tube defects in
infants. This association has been observed to exist; however, when the mother has
high folic acid levels, the effect of mycotoxin Fumonisins on nueral tube defects
appears to be eliminated. So the B-vitamin folic acid has an antagonistic effect on
the association between Fumonisins and neural tube defects.

In summary, synergistic factors magnify the effect and antagonistic factors weaken

9
the effect.

9
When assessing effect modification in our data, we can use the technique of
stratification to identify whether effect modification is present or absent.

We recall that this is the same method that we used to assess confounding.

We begin by calculating the crude overall association between exposure and

disease.

Then, we stratify the data by levels of the third factor that you suspect may
be an effect modifier. You then calculate the stratum-specific estimates and
compare those estimates to one another and to the crude measure of
association.

You are comparing the stratum-specific estimates to determine whether the

magnitude of the association is different across strata and if it is different, we
would conclude that the association differs according to the level of that third
factor. Therefore, there is evidence of effect modification.

Note that this approach differs from the approach we used to assess
confounding in that when we are comparing our stratum specific estimates

10
we compare them to one another first to ensure that they are similar or that
they are the same because that rules out effect modification. Then, if they are
the same, we can assess confounding by comparing those stratum-specific
estimates to the crude.

The main comparison for confounding is between the strata specific estimates
and the crude measure of association to determine if there's a difference after
adjusting for that factor whereas the primary comparison for assessing effect
modification is to compare the stratum specific estimates to one another to
determine if there is a difference.

10
Let’s now consider an example from a case‐control study of the association between
alcohol consumption and bladder cancer.

Based on these data, we estimate a crude odds ratio of 1.73 suggesting that the odds of
alcohol consumption are 73% higher among bladder cancer cases compared to controls
without bladder cancer.

11
Now, let’s determine if there is evidence that race modifies the association between
alcohol consumption and bladder cancer.

In this slide we see that within each race group, the odds ratio is similar with the exception
of the estimated odds ratio among Asian individuals where the odds ratio is 3.75 instead of
the value of 1.72 to 1.74 seen in the other race groups.

This higher odds ratio in the Asian group suggests that race modifies the association
between alcohol consumption and bladder cancer.

12
In this example, there was a clear interaction between race and alcohol consumption
relative to bladder cancer.

The association between alcohol consumption and bladder cancer is stronger among
Asians than among Whites or Black participants.

The association between alcohol consumption and bladder cancer is modified by race.

13
Now, let’s consider an example case‐control study investigating the association
between oral contraceptive use and breast cancer. We are concerned that smoking
may act as an effect modifier.

To investigate the impact of smoking status on the association between oral

contraceptive use and breast cancer, we estimate stratum‐specific odds ratios first
among smokers and then among non‐smokers.

The OR among smokers is 11.98 while the OR among non‐smokers is 1.1.

The OR values are very different between the smokers and non‐smokers suggesting
that the effect of oral contraceptive use on breast cancer is modified by smoking.
Smoking is an effect modifier.

14
When effect modification is present, we want to emphasize that our aim is to
identify it to describe it and to report it but not to control for it.

This approach differs from what we discussed with confounding.

Confounding is something that we want to remove and something we want to
control. That is not the case with effect modification. The presence of effect
modification or interaction represents a different association that exists
among subgroups of the population and so our goal then is to investigate
whether an interaction is present and if so, we want to describe it and we
want to report it.

After we find an interaction, it is appropriate to keep our data stratified to

assess our associations between exposure and disease separately among
those groups.

With effect modification you are not going to control it but instead, want to
describe it. We assess it by comparing the magnitude of association across
the strata specific estimates.

We can use stratification to evaluate and describe this association so the

method of stratification is utilized both to assess confounding and effect

15
modification. When considering stratified estimates, we compare across the
stratum-specific estimates to evaluate effect modification. If the stratum-
specific estimates are similar and we rule out effect modification, then we will
address confounding by comparing the stratum-specific estimates to the
crude estimate. If there are differences between the stratum-specific and the
crude measure, we would conclude that the third factor (the stratification
factor) is acting as a confounder and needs to be adjusted for in our analysis.

15
It is important to keep in mind that confounding and effect modification are different
concepts even though they both involve the influence of some third factor on an
association between exposure and outcome.

Effect modification is present if the estimate of the measure of association differs

according to the level of another variable.

Confounding is present when the observed effect is distorted by the influence of

another variable.

When assessing interaction and confounding in the same study, it is possible to

have one with or without the other.

We will now consider some numeric examples to distinguish effect modification

from confounding.

16
This slide includes three different scenarios where there is no confounding
and no effect modification or interaction.

You can see in each case that the stratum-specific estimates are nearly
identical to one another.

In practice, the stratum-specific and crude measures won’t be exactly equal.

Instead, we are looking for evidence that they differ substantially.

In these examples, the stratum-specific estimates are very similar in stratum

one to stratum two and of course they also do not differ from the crude
values. So we have no evidence of interaction because this stratum-specific
estimates are similar to one another and furthermore, there is no evidence of
confounding because the stratum-specific estimates are also similar to the
crude estimate.

17
In this example we have evidence of confounding but no interaction.

We begin by comparing the stratum-specific estimates to each other and find

in each case that they are similar to one another; however, the stratum-
specific estimates differ from the crude measure and therefore, would
conclude that confounding is present.

In the first case, the confounding results in bias away from the null; in the
second case, the confounder results in bias towards the null; and in the third
case, the apparent increased risk of the outcome with exposure is reversed
following adjustment for the confounder such that the exposure appears to
be protective after adjustment for the third factor.

18
In each of these examples, there is an indication of effect modification, because the
stratum‐specific estimates differ from each other, and furthermore, there is an indication of
confounding, because the stratum‐specific estimates differ from the crude relative risk. In
each case, we would want to report the stratified estimates.

19
In this series, the third factor again acts as both an effect modifier and a confounding
factor; however, the impact of effect modification is much greater than the impact of
confounding. Again, we would report the stratum‐specific estimates instead of the
adjusted or crude estimates.

20
Now, let’s consider a case‐control study that was conducted to investigate the relationship
between oral contraceptive use and ovarian cancer.

The crude OR was calculated as 0.77.

Given the data provided on the next slide, is there evidence that age is a confounder of the
relationship between oral contraceptive use and ovarian cancer?

21
The results have been stratified by age group. In order to investigate the impact of effect
modification, we will need to estimate the stratum‐specific odds ratio values.

22
We can use the observed data to estimate the stratum‐specific odds ratio values. In each
stratum, we find an estimated odds ratio that is roughly similar with values ranging from
0.61 to 0.65. Based on these data, there is no indication of effect modification.

However, the crude OR of 0.77 is meaningfully different from the stratum‐specific

estimates. Thus, there is evidence that the association between OC use and ovarian cancer
is confounded by age.

Is the confounding positive or negative confounding? Toward or away from the null value?

Answer: negative confounding, toward the null value

23
Let’s now consider some additional review problems.

A measles vaccine may be highly effective in preventing disease if given after a child is 15
months of age, but less effective if given before 15 months of age. What does this example
illustrate?

This is an example of effect modification; age modifies the association between vaccine
exposure and measles prevention.

24
A measles vaccine may be highly effective in preventing disease if given after a child is 15
months of age, but less effective if given before 15 months of age.

In this example, the exposure is VACCINATION, the outcome is MEASLES, and the effect
modifier is AGE OF THE CHILD.

25
Now, let’s consider three data sets and determine if an interaction is present and if
confounding is present.

In the first case, the stratum‐specific estimates are notably different from each other and
therefore, an interaction is present. Furthermore, the stratum‐specific estimates differ
from the crude estimate and therefore, the third factor also acts as a confounder.

In the second case, the stratum‐specific estimates are the same and therefore, an
interaction is NOT present. However, the stratum‐specific estimates differ from the crude
estimate and therefore, the third factor acts as a confounder.

In the third case, the stratum‐specific estimates are similar to each other and therefore, an
interaction is NOT present. Furthermore, the stratum‐specific estimates are similar to the
crude estimate and therefore, the third factor is not acting as a confounding factor either.

26
Let’s now consider an example of effect modification from the published study investigating
the association between coffee consumption and the risk of depression.

27
In the results, the authors note that smoking status may modify the association between
coffee consumption and depression and they provide a test for the interaction between
these factors in the model predicting depression.

We see that among current smokers, an increase in coffee consumption by two cups is
associated with a 22% reduction in the risk of depression (calculated as 1 – 0.78) and
furthermore, based on the confidence interval, this reduction is statistically significant
because the confidence interval does not include the value 1.

In contrast, when we consider never or past smokers, we see that there is no significant
association between coffee consumption and depression. An increase in coffee
consumption by two cups is associated with a 5% reduction in the risk of depression among
never or past smokers and this reduction is not statistically significant as indicated by the
confidence interval that includes 1.

28
In this module, we introduced effect modification, which occurs when the effect of a
particular exposure on an outcome differs depending on the level of a third variable. Note
that effect modification differs from confounding. We aim to control or adjust for
confounding factors, while we describe the impact of effect modification. Finally, we
introduced the use of stratification to evaluate the potential for effect modification. In
summary, our reporting of the association between an exposure and outcome will often
involve either adjustment or reporting of the influence on this association by a third factor
that is acting as a confounder or an effect modifier.

This concludes this series.

29