Interpreting numbers – more tricky bitsScotPHO training course

March 2011

Dr Gerry McCartney

Head of Public Health Observatory Division

NHS Health Scotland

gmccartney@nhs.net

Content

• More on causality• Attributable fractions• Screening – pitfalls to watch out for

Does A cause B?

A B

A B

A

B

A B

C

?

Factors which make causality more likely

Bradford-Hill criteria• Strength of association• Consistency• Specificity• Temporality• Biological gradient• Plausibility• Coherence• Experiment• Analogy

CoffeeIschaemic

heart disease

Does coffee cause ischaemic heart disease?

• Factors that do not lie on the causal pathway but which influence the magnitude of effect

Effect modifiers

SmokingIschaemic

heart disease

Male gender(effect modifier)

Asbestos exposure Asbestosis

Necessary or sufficient causes?

Smoking Lung cancer

Jumping from plane without parachute

Squished onto ground

Attributable fractions/risk

• “What fraction of disease incidence in the exposed group is attributable to the risk factor?”

• Calculated by taking the relative risk in an unexposed group from the relative risk in an exposed group

Attributable fractions

Lung cancer deaths per 1,000 population per year

Coronary heart disease deaths per 1,000 per year

Heavy smokers 166 599

Non-smokers 7 422

Attributable fractions

Lung cancer deaths per 1,000 population per year

Coronary heart disease deaths per 1,000 per year

Heavy smokers 166 599

Non-smokers 7 422

Excess risk of heavy smoking

166 – 7 = 159 599 – 422 = 177

Attributable fractions

Lung cancer deaths per 1,000 population per year

Coronary heart disease deaths per 1,000 per year

Heavy smokers 166 599

Non-smokers 7 422

Excess risk of heavy smoking

166 – 7 = 159 599 – 422 = 177

Attributable risk of heavy smoking

159 / 166 = 95.8% 177 / 599 = 29.5%

Attributable fractions/risk

Attributable fractions can also be applied to the whole population using the formula:

= (risk in total population – risk in unexposed population) / risk in total population

Screening

• Why do we screen for conditions? • When is screening appropriate?• Problems with evaluation of screening

programmes• Particular biases

Why screen for conditions?

• To improve outcomes for individuals – Keep Well health checks– Breast mammography

• To improve outcomes for populations – Port health checks– Employment checks

When should you screen?

Based on the Wilson – Junger criteria:

• Is there an effective intervention?• Does earlier intervention improve outcomes?• Is there a screening test which recognises disease

earlier than usual?• Is the test available and acceptable to the target

population? • Is the disease a priority? • Do the benefits outweigh the costs?

Screening – why is it different?

• Individuals may not benefit• Involves people who are well subjecting themselves to testing

– medicalisation • Creation of a pre-disease state• False positive tests• False negative tests • Initiated by health professionals not individuals• Cost-benefit depends on prevalence within a population • Inequalities implications

Particular biases• Lead time bias

Given that screening picks up disease at an earlier stage – the time between diagnosis and death increases without any actual increase in survival

Symptoms

Detected by screening

Death

Death

• Length time bias

Screening is more likely to detect less aggressive disease and therefore can give impression of improved survival

X

X

X

X

X

X

X

Measures used in screening

• Sensitivity is the likelihood that those with disease will be picked up by the screening test

• Specificity is the likelihood that those with a negative screening test will not have the disease

• Positive predictive value is the likelihood that those with a positive test will have the disease

• Negative predictive value is the likelihood that those with a negative test will not have the disease

Measures for screening

• Sensitivity and Specificity

• Positive predictive value and Negative predictive value

Disease Total

Yes No

Screening test

Positive 300 30 130

Negative 20 3000 3020

Total 320 3030 3350

Measures for screening

• Sensitivity and Specificity

• Positive predictive value and Negative predictive value

Disease Total

Yes No

Screening test

Positive 300 30 130

Negative 20 3000 3020

Total 320 3030 3350

Sensitivity = 300/320 = 94%

Specificity = 3000/3030 = 99%

PPV = 300/330 = 91%

NPV = 3000/3020 = 99%

Summary

• Bradford-Hill criteria can be used to judge whether an association is likely to be causal

• Attributable fractions can help identify the discrete contribution of particular risks to an outcome

• Screening is different to other medical interventions and can cause harm

• Screening evaluations have their own potential biases – lead time and length time bias

Questions

gmccartney@nhs.net