Confounding and effect

modification

Module 6

Dr Dana Bliuc
PHCM9794: Foundations of Epidemiology
2022

Based on the 2022 lecture developed by Dr Falster, A/Professor Heywood, A/Prof. Turner.
 Acknowledgement of
Country

• Office of the Pro Vice-Chancellor

Indigenous, UNSW
Aboriginal History of the Main UNSW Campus
 Module 6 Learning objectives

▪ Describe the criteria for confounding;

▪ Understand and describe methods to control for confounding in

epidemiological studies;

▪ Identify and describe methods for avoiding or minimising confounding in

different epidemiological study designs

▪ Describe effect modification and methods to identify it.

 Study Accuracy

Internally valid? Precise?

Systematic error: Random error:

▪ Selection bias ▪ Sampling error
▪ Measurement bias ▪ Measurement error
▪ Confounding

(in the context of the target

Externally valid? population)

Adapted from Figure 8-1, W. Oleckno. Epidemiology concepts and methods. Waveland Press Inc. 2008.
 Confounding

?
Yellow Fingers Lung cancer

Do you think the crude association between

yellow fingers and lung cancer means that
yellow fingers cause lung cancer?
 Confounding

Yellow Fingers Lung cancer

Cigarette
smoking
 Confounding

Confounding occurs when a measure of

effect is biased because of the
association of the exposure (study
factor) with other factors that influence
the outcome.
 We need to think about confounding
(and how to minimise it) when we ask causal
questions.
We need to think about confounding when we want to
identify ‘causes’ of health outcomes
Examples:
▪ Does smoking increase the risk of lung cancer, compared with not
smoking, among people aged 40 years and older?
▪ Does initiating hormone therapy increase the risk of breast cancer,
compared with not initiating hormone therapy, among women aged 45
years and older?

• The aim is to make causal inferences about the effect of the exposure on the
health outcome in the target population

• We measure the average measure of association of an exposure on the health

outcome is measured in the study population, as an estimate of effect.
 We need to think about confounding when we want to
understand if an intervention improves health outcomes
Examples:
▪ Does initiating statins reduce cardiovascular disease outcomes,
compared with not initiating statins, among people aged 60 years and
older?
▪ Does bowel cancer screening decrease the incidence of bowel cancer
mortality within 5-years of diagnosis, compared with no bowel cancer
screening, among people aged 50 years and older?

• The aim is to make causal inferences about the effect of the intervention
on the health outcome in the target population
• The average treatment effect (measure of association) of an intervention
on the health outcome is measured in the study population
 Confounding is ….

▪ Often referred to as the ‘mixing’ or ‘muddling’ or ‘confusion’ of effects

▪ When the effect of the exposure/intervention we are interested in is

confused by the effect(s) of one or more other factors.
If you want to infer that an exposure is a ‘cause’
of the health outcome,

what would be the ideal ‘make-up’ of the exposed

and unexposed groups?

HINT: THINK BACK TO MODULE 2a LECTURE…

 Ideally the exposed and unexposed groups would be the
same, except for the exposure
Real world

Exchangeable groups:

The same on the

measured and
Parallel universe

unmeasured
characteristics that may
confound the exposure-
outcome association we
are interested in
measuring.
Time zero (baseline)
 Ideally the exposed and unexposed groups would be the
same, except for the exposure
Potential health outcomes: e.g.
coronary heart disease (CHD) vs. no CHD

Exposed (smoking) Exposed

Real world

Measure of
causal effect of
smoking on CHD
→
no confounding if
Parallel universe

Unexposed Unexposed exchangeable

groups

Time zero (baseline) End of study period

 When is exchangeability important?

▪ When we ask causal questions

▪ When the aim is to quantify the causal effect of an exposure/intervention

on an outcome.
▪ e.g. the effect of smoking on risk of coronary heart disease.

▪ To do this, the exposed and unexposed need to be as similar as possible

(or ’exchangeable) on all characteristics other than smoking

Why is this important?

 Why is exchangeability important?

▪ When the exposed and unexposed groups are not comparable on

characteristics that may confound the exposure-outcome relationship →
then these factors may explain (at least in part) differences in the
outcome between the two groups.

▪ That is, the other factors may ‘confound’ the measure of association.

▪ Confounding is a major methodological challenge in studies answering

causal questions, particularly observational studies.
 Confounding on a simple causal diagram

Direct acyclic graph (DAG)

Study factor Outcome factor

Confounder

The measures of association may be  or  than the true value if confounding has
not been accounted for in the design and/or analysis of an epidemiological study.

17
 Example: Does alcohol consumption increase the risk of
lung cancer?

(Exposure) (Outcome)
Alcohol Lung cancer

Example of simple causal diagram or Directed Acyclic Graph (DAG)

 Characteristics of a confounder:

1. It is a (causal) ‘risk factor’ for the study outcome;

2. It is associated with the exposure; and,
3. It is not on the causal pathway between the study factor and outcome
factor
3
Study factor/exposure Outcome

2 1
Confounder

19
Webb, Bain and Page (2020) Essential Epidemiology, Ch 8.
 Example: variables on the causal pathway

High blood lipid

High fat diet Heart disease
levels

Study factor Intermediary Outcome factor

High blood lipid levels are on the causal pathway between high fat diet and heart
disease (i.e. high fat diet leads to high blood lipid levels).
 Example: Does alcohol consumption increase the risk of lung
cancer?

(Exposure) 3. Not on causal pathway (Outcome)

Alcohol Lung cancer

1. Established
cause of outcome

Smoking
(Confounder)

Example of simple causal diagram or Directed Acyclic Graph (DAG)

 Is smoking associated with alcohol?

Exposed Unexposed
Is there imbalance in the
proportion of smokers in the Alcohol = Yes Alcohol = No
exposed vs unexposed groups?

Yes → smoking and alcohol are

associated in this cohort

N=50,000, 40,000 smoke N=50,000, 10,000 smoke

Proportion smokers = 80% Proportion smokers = 20%
= non-smoker = smoker
22
Example: Does alcohol consumption increase the risk of lung
cancer?

(Exposure) 3. Not on causal pathway (Outcome)

Alcohol Lung cancer

2. Alcohol and smoking are 1. Established cause

associated in this cohort of outcome

Smoking
(Confounder)

Example of simple causal diagram or Directed Acyclic Graph (DAG)

 Is smoking a potential confounder?
Characteristic Evidence for smoking
1. Established cause of the ✅ Based on published literature
outcome
NB. If no existing literature, check if potential
confounder associated with outcome in the
unexposed group in study data.
2. Associated with the exposure ✅ Smoking unequally distributed between
(may or may not be causal effect on alcohol and non-alcohol drinkers (or exposed
exposure) and unexposed).

NB. There may be existing evidence from the

literature.
3. Not on the causal pathway ✅ Smoking is not caused by the exposure, so
between exposure and outcome it does not lie on the causal pathway between
alcohol and lung cancer.
24
 Is smoking a potential confounder?
Characteristic Evidence for smoking
1. Established cause of the ✅ Based on published literature
outcome
NB. If no existing literature, check if potential
confounder associated with outcome in the
unexposed group in study data.
2. Associated with the exposure ✅ Smoking unequally distributed between
(may or may not be causal effect on alcohol and non-alcohol drinkers (or exposed
exposure) and unexposed).

NB. There may be existing evidence from the

literature.
3. Not on the causal pathway ✅ Smoking is not caused by the exposure, so
between exposure and outcome it does not lie on the causal pathway between
alcohol and lung cancer.
25
 Is smoking a potential confounder?
Characteristic Evidence for smoking
1. Established cause of the ✅ Based on published literature
outcome
NB. If no existing literature, check if potential
confounder associated with outcome in the
unexposed group in study data.
2. Associated with the exposure ✅ Smoking unequally distributed between
(may or may not be causal effect on alcohol and non-alcohol drinkers (or exposed
exposure) and unexposed).

NB. There may be existing evidence from the

literature.
3. Not on the causal pathway ✅ Smoking is not caused by the exposure, so
between exposure and outcome it does not lie on the causal pathway between
alcohol and lung cancer.
26
Yes! Smoking is a confounder of the alcohol-lung cancer relationship.
Is smoking a confounder ‘in practice’ in this
cohort?
 Calculate the crude measure of association

Lung Cancer Incidence

Alcohol Yes No Total (per 1,000)

Yes 1,230 48,770 50,000 24.6

No 420 49,580 50,000 8.4

Total 1,650 98,350 100,000

Ie = 1230 / 50,000 = 24.6 per 1,000 RR = 24.6 / 8.4

Iu = 420 / 50,000 = 8.4 per 1,000 = 2.9
Example: Does alcohol consumption increase the risk of lung
cancer?

(Exposure) Unadjusted RR = 2.9 (Outcome)

Alcohol Lung cancer

Let’s stratify the data by

smoking status to test for
confounding!
Smoking
(Confounder)

Example of simple causal diagram or Directed Acyclic Graph (DAG)

 Stratification: a practical test for confounding
SMOKERS
Lung Cancer Ie = 1200/40,000 = 30 per
Alcohol Yes No Total 1,000
Yes 1,200 38,800 40,000 Iu = 300/10,000 = 30 per
No 300 9,700 10,000 1,000

Total 1,500 48,500 50,000 RR = 1.0

NON-SMOKERS
Lung Cancer Ie = 30/10,000 = 3 per
Alcohol Yes No Total 1,000
Yes 30 9,970 10,000 Iu = 120/40,000 = 3 per
No 120 39,880 40,000 1,000

Total 150 49,850 50,000 RR = 1.0

30
 Stratification: a practical test for confounding

Did smoking confound the association between alcohol and lung

cancer in this study?
✅ The stratum specific measures of association are similar
▪ RR of lung cancer among alcohol drinkers (smokers) = 1.0
▪ RR of lung cancer among alcohol drinkers (non-smokers) = 1.0
✅ The stratum specific measures of association are different to the crude
measure of association
▪ Crude RR of lung cancer among alcohol drinkers, compared with
non-alcohol drinkers = 2.9

Yes!! There is no association between alcohol consumption and lung

cancer after stratifying by smoking.

31
 Identify and assess confounding
▪ Identify potential for confounding:
▪ At the study design stage
▪ Factors that have characteristics of a confounder → measure these when
collecting data!
▪ Useful to draw relationships between variables (NB. causal diagrams are
covered in more depth in Epidemiological Methods PHCM9518 course)

▪ Stratification to assess confounding in practice:

▪ Stratification can be used to assess and control for confounding for one
confounder at a time
▪ Confounding in practice when:
▪ stratum-specific measures of association are similar
▪ stratum-specific measures of association are different to the crude
measure of association
32
 How can we minimize confounding in the
design and analysis stages of epidemiological
studies?

The goal is to achieve ‘exchangeability’ between the exposed and unexposed

groups to estimate ‘causal’ effects of an exposure on an outcome.

33
Design and analysis methods to minimise confounding
When designing a study: When analysing results:
• Identify and measure potential • Stratification
confounders • Restriction
• Matching methods
• Design the study to address confounding, • Multivariate statistical methods
e.g.:

➢ Randomization (in RCTs)

In observational studies:
➢ Restrict eligibility (confounder not absent)
➢ Stratify selection
➢ Identify groups where confounders can be
measured accurately
➢ Match cases and controls on several
confounders (e.g. age, sex) 34
 Randomisation to exposure

▪ Randomisation aims to evenly distribute both

measured and unmeasured pre-exposure
characteristics that are prognostic of the outcome
between exposure groups.

▪ Always assess balance of pre-exposure

characteristics in published paper (usually Table 1)

▪ If not balanced, measure of effect may be

confounded.

▪ NB. There will not be any association between the

potential confounders and the study factor in the
study data (i.e. characteristic #2 of a confounder) if
randomised.

35
In observational studies…
 Restriction

Entire cohort
▪ Restrict study population to
individuals with absent confounder
Excluded:
▪ Can be used to control for one Cohort
members who
confounder, but not multiple ever smoked

Cohort
members who
never smoked

To analyse alcohol – lung cancer association,

without smoking as a confounder 37
 Measurement

▪ Identify populations where the confounder(s) can be measured

accurately

▪ Who might be willing to have measures that are needed?

▪ e.g. saliva tests to measure cotinine level for smoking
▪ e.g. blood tests for lipid measures

38
 Matching: design stage

▪ Individual matching: 1:1 match of individual cases and controls for one or
more potential confounders (e.g. age, sex)

▪ Group (or frequency) matching: aim for the same proportion of potential
confounders among exposed and unexposed groups.

NB. There will not be any association between the confounder and study factor
(characteristic #2 of a confounder) in the study data after matching.

39
 Matching: analysis methods

▪ Calculate matched pairs measures of association if matching

undertaken at the design stage (e.g. matched pairs RR)

▪ Multivariate statistical methods can be used to match subset of

individuals with similar characteristics from existing data

Covered in elective courses:

• Epidemiological Methods
• Advanced Biostatistics and Statistical Computing

40
 Stratification methods

▪ Can be used for a single confounder

▪ Stratify 2x2 tables by potential confounder and calculate measures of

association

▪ Adjust for confounder using statistical methods

▪ e.g. Mantel-Haenszel method Covered in Epidemiological Methods (PHCM9518)

▪ Overall weighted measure of association based on stratum-specific

measures of association x weights for each stratum

41
 Multivariate statistical methods

▪ Necessary for multiple confounders

▪ Generate adjusted measure of association by including multiple
covariates in the model, in addition to the exposure (independent
variable) and outcome variable
▪ e.g. Linear regression (continuous data)
▪ e.g. Logistic regression (categorical data) Covered in elective courses:
• Epidemiological Methods
• Advanced Biostatistics and Statistical
Computing

The adjusted measure of the association for the exposure and outcome
controls for measured confounding variables included in the model.

42
 Residual confounding

Common reasons for residual confounding:

1. Unknown/unmeasured confounders
▪ can’t assess or adjust for confounders in the analysis if not measured

2. Control of confounding not adequate

▪ E.g. use of broad age groups rather than age as continuous variable

3. Measurement error in measurement of confounders (e.g.

misclassification)
 Example: Case-control study

Confounder

Blood lead levels Attention Deficit

Hyperactivity Disorder

44
Full paper
 Matching to control for (some) confounding
Design stage: “We conducted a pair-matching case–control study ….
matched on the same age, sex, and socioeconomic status.”

Full paper 45
 Is matching on 3 variables adequate?
…to control for all potential confounding?

Confounder

Attention Deficit
Blood lead levels
Hyperactivity Disorder

46
Full paper
 Additional methods to control for confounding
Design stage:
▪ Measure potential confounders:
▪ “In addition, we considered multiple covariates and potential
confounders for the association of lead exposure and ADHD in our
study. They were based on established predictors of child
behavioral problems and those widely used in studies of
pediatric lead exposure…” (pg. 1402)
▪ Covariates measured: family history of ADHD, household
composition, maternal smoking and drinking during pregnancy, a
range of birth complications, parent’s age at birth, parent’s education.

Analysis stage:
▪ Adjust for measured confounders in statistical models
47
Full paper
 48
Full paper
 Analysis: adjust for measured confounders

Adjusted Odds Ratios shows an increased odds of exposure to

higher lead levels among cases than controls, after adjusting
for measured confounders.
Does this mean the authors have addressed all potential confounding of the
lead-ADHD relationship? 49
 Summary: confounding

▪ Confounders distort the association between the Study Factor and the
Outcome Factor

▪ Design stage:
▪ need to identify and measure potential confounders
▪ Randomisation (RCTs)
▪ Restriction
▪ Matching (case-control studies)

▪ Analysis stage:
▪ Stratification to assess confounding
▪ Adjust for confounders in statistical analysis

50
 Critical appraisal: design stage
Broad considerations: ▪ More detailed considerations:
Did the authors identify ▪ Did the authors refer to a causal diagram/model?
potential confounders? ▪ Did the authors list the measured confounders?
▪ Did the authors discuss unmeasured
confounders?
▪ Do you think there are important unmeasured
confounders?
Did the authors measure ▪ How accurately/precisely were confounders
potential confounders? measured?
▪ Is it likely that measurement bias an issue for any
measured confounders?
Did the authors use any other ▪ Randomisation?
aspects of study design to ▪ Matching?
address confounding? ▪ Restriction?
51
 Critical appraisal: analysis stage
Broad considerations: More detailed considerations:
Are the exposed and unexposed • Did the authors include a Table showing the distribution
groups ‘exchangeable’ (or similar) (i.e. N, %) of measured confounders in the exposed and
in terms of pre-exposure unexposed groups?
characteristics/confounders? • Is there an imbalance in the distribution of confounders
between the exposed and unexposed groups?

Is there likely to be residual • Did the authors use appropriate methods to adjust for
confounding? confounding in the analysis? e.g. multivariate regression
models.
• What confounders were not measured?
• What confounders were not measured in detail?
• What confounders may have been impacted by
measurement bias?

Discuss likely sources and scale of residual confounding and

possible impact on results (internal validity of study).
52
Effect modification
 Effect modification

▪ Effect modification occurs when a 3rd factor modifies the measure of

association (effect) between the study factor and outcome factor.
Example:
Does screening for osteoporosis in people aged 60 + reduces the
risk of fracture compared to no screening.

▪ Average screening effect in whole study population: RR = 1 (no

difference).

▪ Does sex modify the effect of screening on fracture risk?

▪ RR = 0.8 in women (screening reduces the risk of fracture)
▪ RR = 1.2 in men (screening has no effect on fracture risk)
54
 Effect modification
▪ The exposure’s ‘effect’ on the outcome differs by population sub-groups
when effect modification is present.

▪ Subgroups of interest may be defined by any characteristic, e.g.

▪ Age
▪ Sex
▪ Genetic markers

55
 Practical test for effect modification
▪ Stratification

▪ The measure of association will be different in the two strata of the ‘effect
modifier’ variable if effect modification is present

▪ Case-control study example

▪ Does smoking modify the (causal) effect of asbestos exposure on the

development of mesothelioma?

56
 Practical test for effect modification

Calculate crude odds ratio for average ‘causal’ effect

Mesothelioma
Asbestos Yes No Total
Yes 100 18 118
No 100 182 282
Total 200 200 400

OR = (a*d) / (b*c) = (100*182) / (18*100)

OR = 10.1
57
Stratification: a practical test for effect modification

SMOKERS
Mesothelioma
Asbestos Yes No Total OR = (a x d) / (b x c)
Yes 82 3 85 = (82 x 122) / (3 x 93)

No 93 122 215 OR = 36
Total 175 125 300

NON-SMOKERS
Mesothelioma
Asbestos Yes No Total OR = (a x d) / (b x c)
Yes 18 15 33 = (18 x 60) / (15 x 7)
No 7 60 67
OR = 10
Total 25 75 100
58
 Stratification: a practical test for effect modification

Did smoking modify the association between asbestos and

mesothelioma in this study?
✅ The stratum specific measures of association are different
▪ OR of asbestos exposure among cases (smokers) = 36
▪ OR of asbestos exposure among cases (non-smokers) = 10

Yes!! Smoking modifies the association between asbestos and

mesothelioma.

59
Overweight/Obesity Hypertension

Birth weight
From: Effect Measure Modification by Birth Weight on the Association Between Overweight or Obesity and
Hypertension in Children and Adolescents
JAMA Pediatr. 2023;177(7):735-737. doi:10.1001/jamapediatrics.2023.0799

Table Title:
Effect Measure Modification by Birth Weight (BW) on the Association Between Overweight or Obesity and Hypertension (N = 14 615)
 When is effect modification important?

▪ To identify a subset of population who is at higher risk of an exposure

▪ Average “causal effect” in different population groups is of interest to

policy makers

▪ For targeted early intervention for groups “at-risk”

 Confounding vs Effect modification

Confounding Effect Modification

No true association True association between

between exposure and exposure and outcome
outcome
Different effect in different
strata
Does not provide any
insight on the hypothesis Increased knowledge/ public
under investigation health intervention

63
Ch 4 Effect Modification, from ‘Causal Inference: What If’ by Hernán and Robins
 Stratification: a practical test
for confounding and effect
modification

From Webb, Bain and Page (2020) Essential Epidemiology, Ch 8 (Fig 8.7, page 201).
 Summary
Confounding = Distortion in the measure of association between the study
factor and outcome factor due to another (nuisance) factor

▪ Aim to remove by design and/or analysis to estimate the ‘causal’ effect of

the exposure on the outcome.

Effect modification = Modification in the measure of association between the

study factor and outcome due to a third factor

▪ Report effect modification → important to understand if the effect of an

outcome varies between different groups (e.g. males vs females)

▪ Generalisability of study findings may be affected by different distribution of

effect modifiers in other population considering intervention

▪ (e.g. study/target population may have had less participants from poor
socioeconomic circumstances, compared to population considering
implementing intervention).
65