Causal Inference, Bias and Confounding
Causal Inference
- Causal Inference: Inferring or concluding the cause of a disease from evidence.
- Evidence of a causal relationship between exposure and disease includes:
- Temporality: The cause must precede the effect in time.
- Consistency: The relationship is consistently seen across numerous studies.
- Strength of Association: A strong relationship between exposure and disease increases the likelihood of causality.
- Biological Coherence: Biological evidence supports the exposure causing the disease (e.g., mesothelioma caused by asbestos).
Consistency Example
- Studies showing second-hand cigarette smoke increases lung cancer risk for non-smokers.
- Takeshi Hirayama's 1981 study demonstrated increased lung cancer risk in non-smoking Japanese women married to smokers, compared to those married to non-smokers.
- Since Hirayama’s original paper, 37 studies of passive smoking and lung cancer have been published.
Strength of Association Example
- A dose-response relationship exists between latitude of sunlight exposure and skin melanoma.
- Skin melanoma risk is greatest at the equator and decreases towards the North and South poles.
Relationships May Only Appear Causal
- Reasons why a relationship between exposure and disease may appear causal, but in fact is not:
- Chance variations between samples of a population.
- Bias: Influence or prejudice of the data.
- Selection bias: Participants selectively opt out of a study.
- Measurement bias: Inaccurate data measurement (e.g., from faulty equipment).
- Information bias: Data collected inconsistently (e.g., varying questions).
- Recall bias: Incomplete or inaccurate recall by study participants.
- Confounding: A third factor distorts the association between the studied cause and effect.
- Example: Age is a confounder in carcinogen studies because cancer likelihood increases with age.
Prevent Making False Associations
- Methods to prevent false associations between exposure and disease by reducing chance, bias, and confounding:
- Reduce error by chance: Increase sample size.
- Minimize bias:
- Improve questionnaire response rates (e.g., by making questionnaires less time-consuming).
- Ensure quality control of monitoring equipment (e.g., comparing results to a standard).
- Use a standardized set of questions in outbreak investigations.
- Aid recall by asking questions promptly after an event.
Confounding May Be Minimized By
- Matching: Select controls with similar characteristics to the cases.
- Example: Selecting controls with the same age as the cancer cases.
- Stratification: Create subsets of case and control groups to separate out the third factor effects.
- Example: Studying lung cancer due to radon exposure in underground miners, where smoking is a confounder.
- Create subsets:
- Non-smoking miners with lung cancer.
- Smoking miners with lung cancer.
- Separate out the confounder (smoking); lung cancer among non-smokers is likely due to radon.
- Randomization: Ensure samples are randomly selected.