1. Bias
Lecture notes
Sam Bracebridge

2. By the end of the lecture fellows will be able to
Define bias
Identify different types of bias
Explain how bias affects risk estimates
Critique study designs for bias
Develop strategies to minimise bias

3. What do epidemiologists do? Measure effects Attempt to define a cause
Epidemiologic Study
What do epidemiologists do?
Measure effects
Attempt to define a cause
an estimate of the truth
Implement public health measure
An epidemiologic measure could be, for example, a rate or risk.

4. Estimated effect: the truth?
Mayonnaise Salmonella
RR = 4.3
True association?
Bias?
Chance?
Confounding?

5. Chance and confounding can be evaluated quantitatively
Warning!
Chance and confounding can be evaluated quantitatively
Bias is much more difficult to evaluate
Minimise by design and conduct of study
Increased sample size will not eliminate bias
Despite all preventive efforts, bias should always be considered among alternative explanations of a finding.
It has to be remembered that bias:
   - may mask an association or cause a spurious one
   - may cause over or underestimation of the effect size
Increasing the sample size will not eliminate any bias. A study that suffers from bias lacks internal validity

6. Definition of bias
Any systematic error in the design or conduct of an epidemiological study resulting in a conclusion which is different from the truth
Measurements

7. Errors in epidemiological studies
Random error (chance)
Systematic error (bias)
Study size
Source: Rothman, 2002 7

8. Main sources of bias
Selection bias
Information bias

9. Selection bias Two main reasons: Selection of study subjects
Factors affecting study participation
association between exposure and disease differs between those who participate and those who don’t

10. Types of selection bias
Sampling bias
Ascertainment bias
referral, admission
Diagnostic/surveillance
Participation bias
self-selection (volunteerism)
non-response, refusal
survival
Sampling bias is systematic error due to a non-random sample of a population,2 causing some members of the population to be less likely to be included than others, resulting in a biased sample
Ascertainment bias
Surveillance -
referral, admission – tertiary referral centre
Diagnostic – those who are tested are diagnosed, pre-screening of trial participants
Participation bias
self-selection (volunteerism) – new screening test likely to be taken up by those who are health concious, or worried well,
non-response, refusal – moved out of area, tobacco smokers
survival - discounting trial subjects/tests that did not run to completion

11. Selection bias in case-control studies

12. Estimate association of alcohol intake and cirrhosis
Selection of controls
Estimate association of alcohol intake and cirrhosis
OR = 6
How representative are hospitalised trauma patients of the population which gave rise to the cases?

13. Selection of controls a b c d
OR = OR = 36
Higher proportion of controls drinking alcohol in trauma ward than non-trauma ward a b c d

14. Some worked examples Work in pairs In 2 minutes:
Identify the reason for bias
How will it effect your study estimate?
Discuss strategies to minimise the bias

15. Oral contraceptive and uterine cancer
You are aware OC use can cause breakthrough bleeding
Overestimation of “a”  overestimation of OR
Diagnostic bias
OC use  breakthrough bleeding  increased chance of testing & detecting uterine cancer
Alternative:
Prospective cohort study with clinician blinded to the exposure a b c d

16. Asbestos and lung cancer
Prof. “Pulmo”, head specialist respiratory referral unit, has 145 publications on asbestos/lung cancer
Lung cancer cases exposed to asbestos not representative of lung cancer cases
Alternative:
Strict admission criteria?
Strict testing criteria?
Study in another hospital?
Other lung disease controls?
Overestimation of “a”  overestimation of OR
Admission bias a b c d

17. Selection bias in cohort studies

18. Association between occupational exposure X and disease Y
Healthy worker effect
Association between occupational exposure X and disease Y
General population includes those who do not work due to illness
Source: Rothman, 2002

19. Healthy worker effect
Source: Rothman, 2002

20. Prospective cohort study- Year 1
lung cancer
yes no
Smoker
Non-smoker

21. 50% of cases that smoked lost to follow up
Loss to follow up – Year 2
lung cancer
yes no
Smoker
Non-smoker
50% of cases that smoked lost to follow up

22. Minimising selection bias
Clear definition of study population
Explicit case, control and exposure definitions
CC: Cases and controls from same population
Same possibility of exposure
Cohort: selection of exposed and non-exposed without knowing disease status

23. Sources of bias
Selection bias
Information bias

24. Information bias During data collection Differences in measurement
of exposure data between cases and controls
of outcome data between exposed and unexposed 24

25. Information bias
Arises if the information about or from study subjects is erroneous

26. Information bias 3 main types: Recall bias Interviewer bias
Misclassification

27. Cases remember exposure differently than controls
Recall bias
Cases remember exposure differently than controls
e.g. risk of malformation
Overestimation of “a”  overestimation of OR
What to do?
Check other sources of information
Questionnaire
Mothers of children with malformations remember past exposures better than mothers with healthy children

28. Interviewer bias
Investigator asks cases and controls differently about exposure
e.g: soft cheese and listeriosis
Cases of
Controls
listeriosis
Overestimation of “a”  overestimation of OR
Eats soft cheese a b
Does not eat c d
soft cheese
Investigator may probe listeriosis cases about consumption of soft cheese (knows hypothesis)

29. Misclassification Exposure Outcome
Measurement error leads to assigning wrong exposure or outcome category
Exposure
Outcome

30. Misclassification Systematic error
Missclassification of exposure DIFFERS between cases and controls
Missclassification of outcome DIFFERS between exposed & nonexposed
=> Measure of association distorted in any direction

31. Misclassification
OR = ad/bc = 3.0; RR = a/(a+b)/c/(c+d) = 1.6

32. Misclassification
OR = ad/bc = 1.5; RR = a/(a+b)/c/(c+d) = 1.2

33. Minimising information bias
Standardise measurement instruments
questionnaires + train staff
Administer instruments equally to
cases and controls
exposed / unexposed
Use multiple sources of information

34. Summary: Controls for Bias
Choose study design to minimize the chance for bias
Clear case and exposure definitions
Define clear categories within groups (eg age groups)
Set up strict guidelines for data collection
Train interviewers

35. Summary: Controls for Bias
Direct measurement
registries
case records
Optimise questionnaire
Minimize loss to follow-up

36. The epidemiologist’s role
Reduce error in your study design
Interpret studies with open eyes:
Be aware of sources of study error
Question whether they have been addressed

37. Bias: the take home message
Should be prevented !!!!
At PROTOCOL stage
Difficult to correct for bias at analysis stage
If bias is present:
Incorrect measure of true association
Should be taken into account in interpretation of results
Magnitude = overestimation? underestimation?

38. Questions?

39. References
Rothman KJ; Epidemiology: an introduction. Oxford University Press 2002,
Hennekens CH, Buring JE; Epidemiology in Medicine. Lippincott-Raven Publishers 1987,