Validity of Epi Studies 2019

Validity of Epidemiologic Studies and Causal
Inference
Fessahaye Alemseged (MD, MPHE)
04/13/2020 1
Session objectives
• Discuss the sources of error
• Explain criteria for establishing causation
04/13/2020 2
References
• Ann Aschengrau. Essentials of epidemiology in public

health, 2020.
• Rothman K, et al. Modern Epidemiology, 3rded, 2008.
04/13/2020 3
Validity of epidemiologic studies
• refers to correctness of study findings
• improved by reducing error from various sources (“Beasts of
epidemiology”):
– Sampling error
– Systematic error
– Confounding
– Interaction?
04/13/2020 4
Sampling error/random error/chance
• Occurs when using sample to infer about population
• Refers to lack of precision or power
• How to minimize sampling error:
– By increasing the sample size
• Degree of sampling error has to be estimated
– Values of level of precision (CI) and power

Systematic error/Bias
• Occurs due to deviation from plan
• more dangerous and more difficult to measure

• Types
– Selection bias
– Non-participation
– Information bias
Selection bias
• Results from non-representative sample selection
– Due to non-random selection
• There are several types

Selection bias: Types
• Biases related to institution-based researches

– Diagnostic bias: when using controls form hospitals
– Healthy worker effect: when taking cohorts from worksites
• Self-selection bias – volunteer bias
• Survival bias
• Temporal bias
Solution – population-based studies using random selection

Non-participation
• Commonest known source of bias
• Results from non-response or loss-to-follow-up (LTF)
– Non-response could be due to rejection or unavailability
– LTF – major problem in cohort studies
• Non-participation leads to serious bias when

– non-participants are systematically different from participants
Non-participation: Solutions
• Conducting pre-test
• Adding contingency in the sample
• Diligent efforts to make

– Revisits and
– Make participation as convenient and painless as possible – E.g.

mobile examination
• Evaluate non-respondents; Do extreme analysis

Information bias/measurement error
• Sources of measurement error
– Observer/Interviewer bias
– Interviewee bias, Recall bias
– Invalid instrument/tool
• Solutions:
– Standardization of procedures
– intensive training, supervision, duplicate observation
– use of memory aids
– Pre-testing
Confounding
• The distortion of measure of effect of exposure

– By other factors that are known as confounder variables
• For a factor to be a confounder it has to be

– associated with the exposure and
– independently associated with the outcome
• A confounder is not an intermediate variable or

consequence of the disease
Illustration of confounder
Coffee association of interest

Cancer
Drinking
association association
Smoking
Smoking can confound association between coffee and cancer
04/13/2020 13
Confounding: solutions
• Solutions:
– prevent during design: restriction, matching or randomization
– Control during analysis: stratified analysis
• Restriction: conducting the study in only one group of the confounder
variable
• Randomization: prevents confounding by known and unknown factors
• Stratified analysis? will be discussed in the last session
04/13/2020 14
Confounding: solutions
• Matching: selecting comparison groups that are
similar with regard to values the confounder variable

• Types:
– Individual matching: for continuous variables like
age
– Frequency matching: for categorical variables like
sex
04/13/2020 15
Interaction/Effect modification
• Refers to varying effect of exposure in presence and

absence of another factor
• If present, the adjusted measure can’t adequately
summarize the association
– Assess association separately for each group
• E.g. Asbestos and cigarette smoking

Establishing Causation
• Approaches for making causal inference

– Classic approach: using Bradford Hill’s criteria
– Potential outcome approach
• Bradford Hill’s criteria:

– Guides, not rules – help to rule out erroneous association and make
causal inference
– Include: strength, consistency, temporality, dose-response relationship,

biologic plausibility, experimental evidence
04/13/2020 17
Temporality
• Refers to the necessity that the cause occurs preceding the
effect
• Which type of observational study is best for documenting time
sequence?
• Such time sequence does not in itself establish causation

– could be confounded
• Reverse time order does not rule out causal relationship
04/13/2020 18
Strength of association
• Refers to magnitude of measure of association
• Strong association is more likely to be causal

– If it was due to confounder, it would, likely, have been obvious
• Weak associations could be causal

– E.g. smoking and cardiovascular disease
• Even strong associations might not be causal

– E.g. – down syndrome Vs birth order is confounded by maternal age
04/13/2020 19
Dose-response relationship
• Presence of outcome incidence gradient with

increase in level of exposure
– E.g. As BP level increases and risk of IHD increases
• But some causes have all-or-none or non-linear effect

– E.g. – DES and adenocarcinoma of vagina
• Dose-response relation could be due to confounding

– E.g. gray hair Vs death by age
04/13/2020 20
Consistency
• Refers to presence of similar findings in different
populations
– From systematic reviews and meta-analysis
• Lack of consistency does not rule out causal
association
– It might be due to missing unknown component cause
04/13/2020 21
Biologic Plausibility (1)
• Coherence with current scientific knowledge on
pathogenesis of the disease
– E.g. 1– smoking and lung ca; substances in the cigarette
smoke have been shown to cause histopathological changes
in respiratory epithelium of animals
– E.g. 2- presence of viral genome in liver cells supports

association between Hepatatis B and liver cancer
04/13/2020 22
Biologic Plausibility (2)
• However, lack of plausibility may reflect lack of current
scientific knowledge on pathogenesis
• Doubts about acupuncture and homeopathy is partly due to
absence plausible biological mechanism
04/13/2020 23
Reversibility
• removal of the factor results in a reduced disease risk
– E.g. – cessation of smoking reduces risk of HTN
• for causes that leads to rapid irreversible changes
– reversibility cannot be a condition for causality
04/13/2020 24
Study Design
Relative ability of different types of study to “prove”

causation: Hierarchy of Causation
Type of study Strength of
evidence
Systematic review and
Meta-analysis
Randomized controlled
Cohort
Case-control
Cross-sectional
Ecological
04/13/2020 25
Judging the evidence
• There are no completely reliable criteria, uncertainty remains
• Evidence is stronger when several criteria are met
– More weight is given to plausibility, consistency and dose–response
relationship; correct temporality is essential
• Based on evaluation criteria, evidence is ranked as:

• Sufficient, Suggestive but not sufficient, Inadequate, Suggestive of
no causal relationship
04/13/2020 26

Validity of Epi Studies 2019

Uploaded by

Copyright:

Available Formats

Validity of Epi Studies 2019

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Validity of Epi Studies 2019

Uploaded by

Copyright:

Available Formats

Validity of Epidemiologic Studies and Causal

Fessahaye Alemseged (MD, MPHE)

• Discuss the sources of error

• Explain criteria for establishing causation

• Ann Aschengrau. Essentials of epidemiology in public

• Rothman K, et al. Modern Epidemiology, 3rded, 2008.

• improved by reducing error from various sources (“Beasts of

• Occurs when using sample to infer about population

• Refers to lack of precision or power

• How to minimize sampling error:

– By increasing the sample size

• Degree of sampling error has to be estimated

– Values of level of precision (CI) and power

• more dangerous and more difficult to measure

• There are several types

• Biases related to institution-based researches

– Healthy worker effect: when taking cohorts from worksites

• Self-selection bias – volunteer bias

Solution – population-based studies using random selection

– LTF – major problem in cohort studies

• Non-participation leads to serious bias when

• Adding contingency in the sample

• Diligent efforts to make

– Make participation as convenient and painless as possible – E.g.

• Evaluate non-respondents; Do extreme analysis

– Interviewee bias, Recall bias

– intensive training, supervision, duplicate observation

– use of memory aids

• The distortion of measure of effect of exposure

• For a factor to be a confounder it has to be

– independently associated with the outcome

• A confounder is not an intermediate variable or

Coffee association of interest

Smoking can confound association between coffee and cancer

– Control during analysis: stratified analysis

• Restriction: conducting the study in only one group of the confounder

• Randomization: prevents confounding by known and unknown factors

• Stratified analysis? will be discussed in the last session

• Matching: selecting comparison groups that are

similar with regard to values the confounder variable

• Refers to varying effect of exposure in presence and

• E.g. Asbestos and cigarette smoking

• Approaches for making causal inference

– Potential outcome approach

• Bradford Hill’s criteria:

– Include: strength, consistency, temporality, dose-response relationship,

• Which type of observational study is best for documenting time

• Such time sequence does not in itself establish causation

• Reverse time order does not rule out causal relationship

• Strong association is more likely to be causal

• Weak associations could be causal

• Even strong associations might not be causal

• Presence of outcome incidence gradient with

• But some causes have all-or-none or non-linear effect

• Dose-response relation could be due to confounding

• Lack of consistency does not rule out causal

– E.g. 2- presence of viral genome in liver cells supports

Relative ability of different types of study to “prove”