Causal Inference using Instrumental
Variables in an Epidemiological
Application
Nuala Sheehan & Vanessa Didelez
University of Leicester & University College London
Causal Inference in Epidemiology
One important reason is to find and assess the size of the
effect of modifiable risk factors (e.g. diet) on diseases so
that public health interventions can be informed.
Example:
Observational studies consistently show positive association
between homocysteine levels and coronary heart disease
(CHD).
Homocysteine levels are reduced by folate intake.
If the relationshiop is causal, we can reduce CHD risk by
adding folate to the diet.
Problem: Association
6= Causation
We might find an association but the intervention turns out
to be useless.
Example: Beta-carotene and lung cancer
• Peto et al. (1981): increased intake of vitamin
beta-carotene “reduces” risk of smoking related
cancers
• Could not be reproduced in randomised controlled
trials (1994)
Need to distinguish between association and causation so
that we know whether an intervention will be useful.
Problem: Association
6= Causation
Fisher (1926): Randomised experiments render reverse
causation and confounding highly unlikely.
Randomised/controlled experiments not always
possible—ethical, practical or financial problems.
Require causal inferences from observational data.
Confounding problems—exposures and diseases of
interest often related to socioeconomic or behavioural
factors. We can try to adjust for confounding but need to
know and measure the confounding factors.
Mendelian Randomisation
Katan(1986)—letter to the Lancet:
Hypothesis under debate in mid-1980s: low serum
cholesterol increases risk of cancer.
Have to satisfactorily eliminate
1. Reverse causation: Does presence of hidden
tumours induce a lowering of cholesterol in future
cancer patients?
2. Confounding: Are other factors such as diet and
smoking affecting both cholesterol levels and cancer
risk?
Mendelian Randomisation
Katan(1986)—letter to the Lancet:
Rare disease abetalipoproteinaemia −→ prac