Causation and Bias

Prof. Joseph Ntaganira, MD, PhD

Learning Objectives
When you have completed this session you
will be able to:
Understand the basis and use of Kochs
Postulates
Define the issues relevant to causality
Describe the impact of bias
Define types of bias
Describe methods of reducing bias

2
Causality
What do we mean by 'cause'?

"Money is the root of all evil.
Smoking causes lung cancer.
Drinking cold soda can give you flue
3
Causal Association
A change in the frequency or quality of an
exposure or characteristic results in a
corresponding change in the frequency of
the disease of outcome of interest.
4
Three Types of Causality
1. Sufficient cause
'X Y' (X implies Y; if X occurs then Y will follow)
This means that if X happens, then Y will also happen.
Medical example:
Iron deficiency causes anaemia.
(However, the reverse is not true: all cases of anaemia
are not caused by iron deficiency - bleeding, vitamin
B
12
-deficiency, leukaemia.)
5
Types of Causality
2. Necessary cause
'X Y' (Y implies X; if Y occurs then it must have
been preceded by X)
Medical example:
Exposure to the tubercle bacillus causes tuberculosis.
(Yes, you cannot get TB without being exposed to the
bacillus. However, most people exposed will never
develop disease.)
6
Types of Causality
3. Additive cause
Several risk factors are needed to cause disease.
Medical example:
Coronary infarction (heart attack)
Smoking
High blood pressure
Diet
Lack of exercise
7
Types of causation
A1
A2
A3
A4

B

C

A1 + A2 + A3 B C
A1, A2, A3 (2 of 3) B C
8
Seldom as simple as one risk factor being at the
same time a necessary and a sufficient cause of a
disease.
Example?
9

How do you determine that a factor causes
disease (in either of the three ways above)?

One of the first attempts: Lffler's postulates
(1883):

"The fulfilment of these postulates is necessary in order to
demonstrate strictly the parasitic nature of a disease:

1. The organism must be shown to be constantly present in
characteristic form and arrangement in the diseased
tissue.

2. The organism which, from its behaviour appears to be
responsible for the disease, must be isolated and grown
in pure culture.

3. The pure culture must be shown to induce the disease
experimentally."

(Sometimes called Koch's postulates.)


What type of cause do these describe?
10
How do these postulates work in public
health?


Best way to prove Kochs is experimentation.

However, in most situations in public health, we
cannot perform experiments.

Instead, we look at associations between a risk
factor and a disease.

Such as smoking and lung cancer or high fat
intake and cardiovascular disease.

(or - an economic system and living standard)


(or - a legal system and crime rate)


(or - a health care system and general health)

11
Causality in Public Health
Strength of association
Consistency of association
Specificity of the association
Temporal relationship
Dose-Response
Biologic plausibility
Experimental confirmation
12
We can never be 100% certain. However, Bradford
Hill also pointed out (1965):
"All scientific work is incomplete -
whether it be observational or
experimental. All scientific work is
liable to be upset or modified by
advancing knowledge. That does not
confer upon us a freedom to ignore
the knowledge we already have, or to
postpone the action it appears to
demand at a given time."
In everyday life - beware of regarding temporal
associations as causal!
13
Validity
External validity
Does sample represent the population
Measure the phenomena of interest
Can we generalize our findings
Can we generalize the status of HIV infection
in Rwanda from a study in the East Province?
14
Validity
Internal Validity
How well was the study done?
Are the measurements accurate?
Are the groups comparable?

Example:
Study of circumcision preventing HIV infection
that takes volunteers for the circumcision group
15
Error
The difference between the true value and
the recorded value
Random Error
Role of chance
Sample size, measurement accuracy
Systematic error
Related to study design
Bias and confounding
16
Chance
Chance may influence the resultsdue to
random variation from sample to sample
17
Chance ExampleObesity Study
Weigh 5 people
40, 60, 70, 90, 110 Kilogram
Great variability


Kilograms
40 60 70 90 110
18
Chance ExampleObesity Study
Weigh 100 people
40-110 pounds
Less variability, more reliability

Kilograms
19
Random error
valid but not precise
x
x
x x
x x
x
x
x
x
x
20
Systematic error
may be precise, but not valid
x
x
x
x
x
x x
x
x
x
x
21
Bias
Any systematic error in an epidemiologic
study that results in an incorrect estimate
of the association between exposure and
risk of disease
Hennekens and Buring, 1986

Not even, unbalanced, preferred
Miss the target every time in the same way
22
Stages of study at which bias occurs
Reading literature
Selecting sample
Execution of study
Measuring exposures or outcomes
Analyzing data
Interpreting analysis
Publishing results
EVERYWHERE
23
Selection Bias
Systematic error in the specification of the
study sample(s) or in the enrollment of
study participants that results in a mistaken
estimate of an exposures effect on the risk
of disease.
OR
A problem with who gets into your study.
24
Selection Bias
Enrolment criteria differs between groups of comparison AND is related
to exposure and outcome
Exposure
+
--
Disease status
Controls Cases
A
B
C D
a
c
b
d
25
Selection Bias
Comparison group must had the
opportunity to have been exposed to
the causative factors(s) as were the
cases.


26
Selection Bias
Diagnostic Bias
How define a case
Knowledge of exposure
Is it a real non-case

Example: Selecting controls for a measles
vaccine study:
Controlsnever had measles, but could have
had subclinical measles
27
Selection Bias
Comparison group
Hospitalized patients
Friends
Neighbors
Random selection from community from
which cases come
28
Selection Bias
Self-selection
Soldiers to be part of a study of Gulf War
syndrome
Diagnostic bias
Often use hospitalized patients to make
conclusions
Many studies made on hospital patients
Case fatality rate
Vaccination status
Cases more thoroughly investigated than
controls.
29
Selection Bias
Non-response bias
Low response rate
Bias introduced if response less than approximately 80% of
those selected to be in study. Ex: Study of sexual practices
Non-participation bias
People who volunteer may be very different from those who
don't.
modern sexual behaviour studies
any postal questionnaire
Survivorship bias
More likely to study people with less severe disease
Study of men with prostate cancer in general medicine clinic
30
Selection Bias
Big problem in case control studies
Control group not representing the true
background population.
Controls are different from cases
Every control, if he had been sick, should
have been eligible as a case.
31
70% of women married more than
five years are having sex outside of
marriage.
- The New Hite Report
Based on 4,500 responses
32
Information Bias
Systematic error in the collection of
exposure or disease data that results in a
mistaken estimate of an exposures effect
on the risk of disease.
OR
A problem with the information you get
from the people in your study.
33
Information Bias
Questionnaire Design
Unclear questions
Inappropriate questions

Interviewer Bias
Asks questions differently
34
Information Bias
Respondent Bias
The respondent may give false answer in order
to please the interviewer or withhold data

Recall Bias
Influenced by time between cause and effect

Reporting or Ascertainment Bias
Results from publicity
35
Misclassification or Information Bias

Exposure
+
--
Disease status
Controls Cases
A
B
C D
a
c
b
d
36
Information bias
Differential misclassification
Misclassification in one direction related to exposure
or outcome status
Controls for cholera cc study claim to treat water more
than cases
Bias OR/RR in one direction
Nondifferential misclassification
Misclassification in both directions, not related to
exposure or outcome, random
Cases and controls equally likely to forget exposure
Bias OR/RR towards null
37
Information bias
Recall bias
Mothers of babies with birth defect remember
drugs they took during pregnancy better
Cases of aflatoxin remember eating rotten
maize more than controls
Interviewer bias
Interviewer tries harder to find exposure in
cases than controls
38
Control of Bias
Bias cannot always be controlled in the analysis
Prevention of bias in the design phase of an
investigation is crucial to the validity of the study
results.
Choice of study population
Comparability (hospital controls)
Reduce loss to follow-up
Methods of data collection
Data collection instruments
Administration of those instruments, Training important
No prompting for responses
Give prompts to help recall (calendars, sample of pill)

39
Recap
Now that you have completed this session you
should be able to:
Understand the basis and use of Kochs
Postulates
Define the issues relevant to causality
Describe the impact of bias
Define types of bias
Describe methods of reducing bias

40
Information Bias
Even with a perfect sample a subjects, we may not
be getting the information we think. (Relates mostly
to interview situations.)
Questionnaire faults
Do others mean what we mean?
pre-test
reverse translation
cut down
Questionnaire design is a science in itself.
Interviewer bias
Is everyone asked the same questions in the same way?
hepatitis C in blood donors
different interviewer for cases and controls
41
Recall Bias
Cases may be able to (want to) remember
exposures much better than controls
food items at an outbreak
abortion and risk of breast cancer
Greatest problem when hypothesis tested is
already known
Creutfeldt-Jakob and beef consumption
When there may be economic compensation
involved.
42
Biases can give rise both to apparent
associations when there aren't any, and
also hide true associations. Impossible
to say which.

A bias introduced into a study can
never be removed by clever
manipulation in the analysis of data.

43