Bias

Confounding
Interaction
Dr Mostafa El Houssinie
Professor
Department of Community,
Environmental and Occupational
Medicine
Ain Shams University
 Study of the variables from the functional
point of view
Independent Dependent
Predictor Predicted
Exposure Response
Covariate/Factor Outcome

• The relationship between an exposure and a response is the main objective

of Epidemiology
• This relationship depends on Measurements of Exposure and Response
• The process of measurement may be subjected to error
• Error : discrepancy between the observed result and the true value
 Study Validity

• Internal validity
• The degree to which the observed results of the study are true.
• Inferences from the sample can be generalized on the study population
• External Validity
• Generalizability of the result to populations outside the study population.
 Study Validity

• Internal validity
• A study of Maternal morbidities and mortalities in a sample of Women (19-
50 yeas old) selected randomly from Menoufeya Governorate.
• Generalization of sample results: Estimates and Relationships to the
study population = Menoufeya women aged 19-50 years depends on
avoiding bias and confounding in the design and analysis of the study.
 Study Validity

• External Validity
• Generalizability of the result to populations outside the study population.
How much did the Estimates represent parameters in other
Governorates?
• Not statistical job!!!!!
• Maternal morbidity and mortality seems to reciprocally related to women
education and accessible easy transportation methods.
• Upper Egypt governorates, then, are expected to have higher values for
maternal morbidity and mortality
 Study Validity: An
Approach
• The observed results occurred because:
• Chance
• Random error
• Bias
• Systematic error
• Confounding
• Truth
• If the role of chance is small, if bias can be reasonably
excluded, and if confounding is addressed, then the study is
internally valid
 Sources or error in studies
• Sampling or random error = an error resulting from the process of
sampling so that
– Sample statistic = population parameter + random error
– Can be measured and controlled
– Usually of small size and in ANY DIRECTION
– Repeated samples form rural lower Egypt results in HCV prevalence in adults 40+
ranging from 8% to 16%

• Bias or systematic error = error due to design, measurement or analysis

– Sample statistic = population parameter + systematic error + random error
– Can not be measured nor controlled
– Of any size but in ONE direction
– Prevalence of HCV infections in Mansheyet Nasser is 70% taken as an estimate of HCV
prevalence in Cairo
 Bias
• Bias is any systematic error in the design, conduct, or
analysis of a study that results in a mistaken estimate of an
exposure’s effect on the risk of disease
• Once introduced, can never be measured nor controlled for
• Better to avoid ….
• May be present without investigator being aware
• Sources may be difficult to identify
• Influence may be difficult to assess or remove
 Bias
• Selection Bias
• The process of recruiting and enrolling participants into the
study distorts the relationship between the exposure and the
outcome
• The relation between exposure and outcome is different for
those who participate and those who would be eligible for the
study but do not participate. Rothman
• Information
• Systematic errors in measurement
• Misclassification
• Differential
• Non-differential
 How selection bias works
Asthma Emotional Free from emotional Total Incidence RR
problems problems
Asthma
Free from asthma
100
1000
900
9000
1000
10000
0.10
0.10
1 Village

Asthma Emotional Free fro emotional Total Incidence RR

problems problems

Asthma 70 180 250 0.28

2
Clinic
Free from asthma 150 900 1050 0.14

Probability of Hospital Admission

Asthma + emotional problems Emotional problems alone Asthma alone Free from both

70% 15% 20% 10%

 Selection Bias
• Volunteer bias
• Membership bias-self selection bias
• Industrial workers are usually healthier than general population, resulting in the healthy worker
effect..
• Use of prevalent cases, instead of incident cases, in case-control studies may result in biased
results if the risk factor is also a prognostic factor, called Neyman bias. Prevalent cases reflect
both risk factors and survivorship factors.
• Use of hospitalized cases of the disease may produce a wrong association of two conditions
unrelated in the general population as both conditions may be related to the hospital admission or
getting under clinical observation, called Berkson’s bias.
• Non-responders or refusals, the characteristics that are associated with non-response may
interfere in the association.
• Loss during a follow-up study is another example of selection bias affecting cohort studies and
clinical trials
• Referral bias when severe and late complicated cases are referred to the major health care
facilities, e.g., university and specialized hospitals.
 Information Bias
• Recall bias
– Disease occurrence enhances recall about potential exposures … Fabrication
….Denials … who is to blame!!!!!! (our memory is SELECTIVE)
– Recall bias is a misinformation resulting from selective recall of previous
activities and exposures of cases and/or controls prior to the occurrence of the
disease condition. Patients, especially when educated, will search very hardly
their memories to identify causes of the disease, a process that may be affected
by believes and attitudes.
• Reporting bias
– Exposure may be under-reported because of attitudes, perceptions, or beliefs…
sexual behavior, drug use, violence in childhood …..(sensitive issues)
 Information Bias

• Interviewer bias
• Taking sides by the interviewers when asking about exposures in case-control
studies.
• They may probe more deeply for responses among cases. This may take the
form of asking ‘extra questions’, use of voice tone, body language and face
expression to suggest a preference to certain responses.
• They are exploiting the phenomenon of ‘social acceptance’ i.e., people try to
please the interviewer by giving him what he likes!
• Interviewer bias, to occur, depends upon knowledge of the interviewer of the
examined hypothesis and the status of the interviewed (case or control).
 Information Bias

• It is well-known that diagnostic tests are not 100% valid

• There will be a proportion of false positives and false negatives that
would result in the misclassification; cases reported as controls and/or
controls reported as cases.
• Measurements by scales or instruments may also introduce bias due
to human of mechanical errors. A vitalograph measuring lung volumes
and capacities or a scale measuring weights tend to over or
underestimate the measurements as the machine ages, so frequent
calibration is necessary.
 Information Bias

Information bias leads to Misclassification

• Non-differential
• Error in assessing exposure or disease is similar between study groups
• Measure of effect tends toward the null value
• Differential
• Error in assessing exposure (or disease) differs in different study
groups
• May increase or decrease measure of effect
  Truth Study
  Fat intake Fat intake
  High Low Misinform High Low
24 Misinform
MI 60 40 36 64
(40%)
16 Misinform
Free from MI 40 60 24 76
(40%)
  Odds ratio = 2.3   Odds ratio = 1.8
  Truth Study
  Fat intake Misinform Fat intake
  High Low High Low
36 Misinform
MI 60 40 24 76
(60%)
16 Misinform
Free from MI 40 60 24 76
(40%)
  Odds ratio = 2.3   Odds ratio = 1.0

  Truth Study
  Fat intake Misinform Fat intake
  High Low High Low
3 Misinform
MI 60 40 57 43
(5%)
16 Misinform
Free from MI 40 60 24 76
(40%)
  Odds ratio = 2.3   Odds ratio = 4.2
 How to prevent Bias

• Approaches to bias
• Random error
– increase sample size
– change design
– improve instrumentation
• Systematic error
• Careful planning of measurements
• Formal assessments of validity
• Regular calibration of instruments
• Training of data collection personnel
The third Variable‫لث‬HH‫لثا‬HH‫لرجل ا‬HH‫ا‬ Maternal deprivation (=poverty) is an
independent variable in delivering a
low birth weight baby (the dependent
variable). Diet plays the role of a
mediator as poverty results in poor
insufficient diet which results in low
birth weight baby. Age (A confounder),
however, may be associated with
woman’s poverty (being older with
large family size) and is associated
-biologically-with low birth weight.
Smoking act as an effect modifier
(Moderator) as deprived mothers who
smoke produce babies even smaller
than the separate expected effects of
smoking and deprivation. Maternal
height has an effect on birth weight but
has nothing to do with maternal
deprivation.
 Confounding
• The Idea:
• Confounding is a confusion of effects.

• Definition:
• The apparent effect of the exposure of interest is distorted
because the effect of an extraneous factor is mistaken for
or mixed with the actual exposure effect.
 Confounding
Properties of a Confounder:
• A confounding factor must be a risk factor for the
disease.
• The confounding factor must be associated with
the exposure under study in the source
population.
• A confounding factor must not be affected by the
exposure or the disease.
• The confounder cannot be an intermediate step in the
causal path between the exposure and the disease.
 Confounding -Causal Diagram
Non-causal
Causal
Confounder Smoking

Exposure Outcome Coffee drinks Lung cancer

Confounder Diabetes

Exposure Outcome Hypertension Coronary HD

Example of confounding –Stratified analysis
Coffee No coffee Coffee No coffee
Cancer 60 35 Smoker 800 300
Total 1000 1000 Total 1000 1000

Coffee No coffee CI Coffee No coffee RR

Smokers Cancer 56 21 7% 7% 1
Total 800 300

Nonsmokers Cancer 4 14 2% 2% 1
Total 200 700

All Cancer 60 35 6% 3.50% 1.71

Total 1000 1000
 Confounding
Case-control study to determine whether vitamin C intake is associated with
colon cancer.

Diet/lifestyle

Vitamin C Cancer

Non-causal
People who take vitamin C may eat a healthier
Causal diet and live a healthier lifestyle
 • Do seatbelts reduce crash injuries?

Risk averse

Wear seatbelts ↓Injured in a crash

?
 Dietary Coronary
Serum
intake of Heart
triglycerides
animal fat Disease

Serum triglycerides is A Mediator between dietary animal fat and

coronary heart disease….
as it is in the causal pathway in the E D relationship
 Confounding-Detection
• Biologic model or underlying theory should allow you to specify
potential confounders in advance of study/analysis

• Assess for confounding in a systematic way

• Known or potential confounding factors
• Other factors not previously known to be confounding factor

• Age and sex are known to affect morbidity and mortality from most
health related condition – Universal Confounders - so if they are
distributed differently between the two compared groups a
confounding effect is inevitable.
 Confounding - Remedies
• Design
• Restriction
• Matching
• Individual matching
• Group matching
• Randomization in Clinical and field experiments
• Analysis
• Stratified analysis
• Adjustment
• Age-adjustment
• Regression analysis
Interaction-Effect Measure Modification
• “Interaction is present 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.”

• Effect Measure Modifier is A factor that modifies the measure of effect of a

putative causal factor under study. There is effect modification when the
selected effect measure for the factor under study varies across levels of
another factor.

• But what effect measure: Risk Difference or Relative Risk (or its
equivalent Odds Ratio)?
• If there is Additive Interaction then the effect of combined exposures ≠
the sum of the effect of either exposure measured by Risk Difference
• If the effect of combined exposure > the sum of RD of both exposures =
Synergism
• If the effect of combined exposure < the sum of RD of both exposures =
Antagonism

If there is Multiplicative Interaction then the effect of combined exposures ≠ the
product of the effect of either exposure measured by the Relative Risk or Odds Ratio
If the effect of combined exposure > the product of RR (OR) of both exposures =
Synergism
If the effect of combined exposure < the product of RR (OR) of both exposures =
Antagonism
 No additive interaction between smoking and Positive additive interaction between smoking
Asbestos exposure and Asbestos exposure
50 50

45 45 10

40 40

35 35

30 20 30 20

25 25

20 20 20 20

15 10 10 15 10 10

10 10

5 10 10 10 10 5 10 10 10 10

0 0
 Example of Interaction –Stratified analysis
No Non- Smokers Non-smokers
Asbestos Asbestos Smokers smokers
Asbestos Cancer 40 30
Cancer 70 30 Cancer 60 40
Total 1000 1000 Total 1000 1000 Total 1000 1000

No Asbestos Cancer 20 10
Total 1000 1000

Asbestos Smoking Absolute Risk Risk Difference Risk Ratio

No No 10 0 1
No Yes 20 10 2
Yes No 30 20 3
Yes Yes 40 30 4
 Example of Interaction –Stratified analysis
No Non- Smokers Non-smokers
Asbestos Asbestos Smokers smokers
Asbestos Cancer 50 30
Cancer 80 30 Cancer 70 40
Total 1000 1000 Total 1000 1000 Total 1000 1000

No Asbestos Cancer 20 10
Total 1000 1000

Asbestos Smoking Absolute Risk Risk Difference Risk Ratio

No No 10 0 1
No Yes 20 10 2
Yes No 30 20 3
Yes Yes 50 40 5
The following is the incidence of upper respiratory tract infections in relation
to parents’ smoking behavior: we have 4 groups: mother and father are
nonsmokers, only the father is a smoker, only the mother is a smoker and
both are smokers

Reference group

Mother Nonsmoker Smoker

Father Non-smoker 14.9 17.1
smoker 16.4 46.1
Attributable Risk for low BW in smoking mothers
Father Non-smoker 0 (17.1-14.9) = 2.2 Additive
Expected joint effect
Smoker (16.4-14.9) = 1.5 (46.1-14.9) = 31.2 AR= 1.5+2.2 = 3.7
Relative Risk mother's smoking
Father Non-smoker 1.0 (17.1/14.9) =1.1 Multiplicative
Expected joint effect
Smoker (16.4/14.9) =1.1 (46.1/14.9) = 3.1 RR= 1.1x1.1 = 1.2
In case-control studies we are looking for multiplicative
interaction only as we can not calculate the Risk Difference
and only the Odds Ratio cal be calculated

The relationship of smoking and oral contraceptive use to the odds of myocardial infarction in
women
Heavy smoking OC use Cases Controls Odds Ratio Explanation
No No 10 60 1 Reference group
No Yes 15 20 4.5 Effect of OC in nonsmokers
Yes No 12 10 7.2 Effect of smoking in non-OC users
Yes Yes 13 2 39.0 Effect of OC in smokers
 Interaction and Confounding
• Relative risk of cigarette smoking and development of oral
cancer with and without exposure to alcohol intake

Condition Alcohol No alcohol

  Intake intake Both groups confounding Interaction
A 4 4 4 no confounding no interaction
B 4 4 1 confounding no interaction
C 4 2 2.8 no confounding Interaction
D 4 2 1 confounding Interaction

C : Weighted average of RR of both strata

Interaction fallacy-The two strata have VERY different
baseline risk of the disease and so the use of OR may give
a false impression of a large difference of association in the
two strata while RR shows no such difference

Individuals with high-risk COX-2 Individuals without high-risk COX-2

genotype genotype

Low aspirin intake 75 25 30 70

High aspirin intake 25 75 10 90
RR 3.0 3.0
OR 9.0 3.9
Incidence of disease 50% 20%
Approach to Interaction and Confounding
Is there interaction?

Yes No

Report separate Is there confounding?

measures for levels
of covariate
Yes

No

Adjust for No need for

confounder adjustment
 Any
Question???

39