1. Lecture 2: Bias Jeffrey E. Korte, PhD
BMTRY 747: Foundations of Epidemiology II
Department of Public Health Sciences
Medical University of South Carolina
Spring 2015

2. Observational research
Observational research studies are subject to bias!
This must be addressed early.
Bias can only be avoided (or reduced) when designing studies and collecting data.
After data are collected, we can only assess bias and try to estimate its impact on results.

3. Precision and validity
Random error
Threatens precision
Systematic error
Threatens validity

4. Precision and validity (vs. random and systematic error)
High precision, high validity
Low precision, high validity
High precision, low validity
Low precision, low validity

5. Precision (lack of random error)
Random error is always present
Is it partly due to pure chance?
Or is everything causally deterministic?
We cannot tell the difference
Causation is too complex
Example: flipping a coin

6. Precision Sources of random error:
Sampling error (imperfect subset of entire population – any epidemiologic study)
Measurement of important study variables
Confounding by unmeasured variables
Residual confounding from measured variables

7. Precision Three ways to reduce random error: Increase study size:
Improve measurement (e.g. finer distinctions)
Increase the size of the study
Increase the study efficiency (change study design)
Increase study size:
Statistical power depends on sample size, and standard error (or distribution) of key variables

8. Precision Increase the study efficiency:
Example 1: cohort study with 100,000 men
Question: daily aspirin use and CVD mortality
If only 100 men take aspirin daily, very few cases will occur in this group
This will lead to an imprecise estimate of the association

9. Precision Increase the study efficiency:
Example 2: cohort study with 100,000 men
Question: daily aspirin use and CVD mortality
50,000 men take aspirin daily
But all men are aged 30-39, so very few cases will occur in this entire cohort
This will lead to an imprecise estimate of the association

10. Precision Increase the study efficiency:
Example 3: cohort study with 100,000 men
Question: daily aspirin use and CVD mortality
50,000 men take aspirin daily
All aspirin takers are aged 40-49, but all others are aged 50-59, so if we try to stratify by decade of age, we cannot estimate any association at all
Other analysis strategies are viable, but results will be problematic (must control for age)

11. Systematic error (error not distributed randomly)
Lack of internal validity
Selection bias
Affects internal and external validity
Confounding
Unmeasured variables
Residual confounding
Information bias
Lack of external validity (generalizability)

12. Selection bias

13. Selection bias
Individuals with different exposure/outcome combinations have different probabilities of being included in the study sample
Exp
No exp
Dis ?
No dis

14. Selection bias Cohort study Case-control study
Exposed population is systematically ascertained or identified in a different way from unexposed population (impacting likelihood of experiencing the outcome)
Case-control study
Case population is systematically ascertained in a different way from control population (impacting likelihood of being exposed)

15. Selection bias Example 1: healthy worker survivor effect
Occupational cohort study
“Exposed” group composed of workers
“Unexposed” group: includes non-workers, or a sample of a different group of workers
Usual result of including non-workers in “unexposed” group: attenuation of associations between chemical exposures and disease/death

16. Healthy worker survivor effect (cohort study)
Dis
Not dis
Exp 50
450
500
No exp 21
479 71
929
1000
Dis
Not dis
Exp 35
465
500
No exp 21
479 56
944
1000
Experiment (RR=2.4)
Occupational cohort study (RR=1.7)

17. Selection bias Example 2: medical surveillance bias
Subclinical disease, if present, is more likely to be detected among individuals with frequent physician visits
This could result in a spurious association observed (for example) in a case-control study of oral contraceptive use and diabetes
Diabetes cases identified in a clinic (more likely to be on OC than general population)
Diabetes controls recruited from general population

18. Selection bias Example 2: medical surveillance bias
Any factor that increases the likelihood of screening will increase the likelihood of subclinical disease detection
Screened population may be overrepresented in case group of case-control study
One solution: recruit controls representative of the population giving rise to cases

19. Medical surveillance bias (case-control study)
Dis
Not dis
Exp 95
190
No exp
405
810
500
1000
Dis
Not dis
Exp
170 95
265
No exp
330
405
735
500
1000
No bias (OR=1.0)
Biased (OR=2.2)

20. Selection bias Example 3: hospital controls
Hospital-based case-control study
Cases come from hospital
Controls are patients in the same hospital
Might have different type of cancer
Might have very different disease from disease of interest

21. Selection bias Example 3: hospital controls
Most groups of hospital patients are different from the general population
More risk behaviors (smoking, diet, exercise)
Lower socioeconomic status
Risk factors may be similar across diseases
Result: attenuate association when comparing cases and hospital-based controls
Solution: recruit controls to be representative of population giving rise to cases

22. Selection bias
Hospital controls may be different from cases in other ways, depending on disease
Example: aplastic anemia cases being treated in a major referral hospital
Usually there for a bone marrow transplant
Therefore more likely to have large family (genetically matched sibling donor)
Also more likely to have high income/wealth

23. Hospital controls (case-control study of smoking and heart attack)
Dis
Not dis
Exp
250
125
375
No exp
625
500
1000
Dis
Not dis
Exp
250
225
475
No exp
275
525
500
1000
Population controls (OR=3.0)
Hospital controls (OR=1.2)

24. Selection bias Selecting controls
Equalize selection forces (bias) for cases and controls (they should be from the same subset of the population)
Example: cases identified through screening program; controls taken from same screening program
Cases and controls may have similar above-average risk factors, leading to screening
This may allow valid comparison of case/control

25. Testing for selection bias
Recruitment rates into case-control study
Differences between cases and controls?
Recruitment rates into cohort study
Differences between exposure groups?
Note: low recruitment rates can result in low generalizability (separate issue)

26. Testing for selection bias
Collect minimal information on refusers (who were not recruited into study)
Age, sex, ethnicity, education, etc.
Maybe behavioral or health information
After recruitment is complete: can compare basic measures between enrolled and non-enrolled, and quantify bias

27. Sensitivity analysis
Test how much your results change if you could have included those who dropped out (under various scenarios, if information is missing)
With complete information on factors affecting selection, selection bias can be quantified and adjusted

28. Selection bias
(go through concrete example of how bias works, and how compensating bias can help solve the problem – tables 4-1 through 4-4, page )

33. Information bias

34. Information bias
Systematic tendency for some individuals to be misclassified on exposure or outcome
Continuous variable: may be systematically overestimated, or systematically underestimated
Misclassification may be based on self-report, or study observations

35. Cohort study: exposure bias
Exposure information quality is unlikely to differ much by eventual disease status.
Neither investigator nor participant knows what the eventual disease status will be
Usual result: any misclassification is nondifferential (bias towards null)

36. Cohort study: outcome bias
Outcome information may be biased:
Due to interviewer/observer
Example: classification of pathology slides (e.g. alcoholic cirrhosis)
Or due to participant
If outcome is obtained by self-report (e.g. migraine headache)

37. Case-control study: exposure bias
Exposure information is gathered after disease status is known. Can result in differential misclassification
Recall bias
Cases may be more likely to recall exposures
Cases may be more likely to overestimate exposures
Interviewer bias
Possible if non-blinded to case/control status

38. Cross-sectional studies
Exposure bias: same problems as case-control studies, because disease status is known
Recall bias
Interviewer bias
Outcome bias: same problems as cohort studies, because exposure status may be known before outcome is assessed

39. Prevention of recall bias (exposure)
Verification of responses
Using medical charts, pharmacy records, etc.
Have to make judgment about resolving discrepancies
Use “diseased” control group
Try to equalize “rumination” about exposures between case group and control group
e.g. symptomatic patients from the same clinic who were found not to have the disease
May not ruminate the same amount as cases
e.g. patients with a different disease
Same problems as hospital controls discussed earlier

40. Prevention of recall bias (exposure)
Choose objective markers of exposure
Self-report may be more consistent, compared to subjective factors
Some biomarkers may reflect cumulative exposures, partially reflecting exposures occurring before disease
Genetic polymorphisms obviously can be considered stable throughout life

41. Prevention of recall bias (exposure)
Use a prospective design!
Cohort study
Case-control study within a cohort study
Case-cohort
Nested case-control

42. Interviewer bias (exposure)
Can occur when:
interviewer is not masked to case/control status AND interviewer does not fully standardize the questionnaire
Clarification of questions
OK if there is a standardized protocol
Probing for more information
Must always follow established skip patterns

43. Prevention of interviewer bias
Reliability and validity substudies
Laboratory tests
Validation with medical records
Questionnaire data: may be difficult to validate
Repeat measurements
Answer may vary for different interviews
Participant may remember previous answer
Interviewer may remember previous answer

44. Prevention of interviewer bias
Mask interviewer to case/control status, or exposure status in cohort studies
May be feasible to blind them
Or mislead them about the hypothesis/exposure of interest
Or mislead them about participants’ case/control status

45. Outcome identification bias
Observer bias
May occur when:
Outcome assessment is subjective, AND
Exposure status is known by assessor
Results in differential misclassification
Can bias towards or away from the null

46. Prevention of observer bias
Mask observer to exposure status
Stratify on certainty of diagnosis
“possible” disease: more vulnerable to bias
Employ multiple observers
e.g. 2 of 3 must agree on diagnosis

47. Outcome identification bias
Respondent bias
May occur when:
Outcome is based on self-report (e.g. migraines), AND
Respondent knows exposure status
Results in differential misclassification
Can bias towards or away from the null

48. Prevention of respondent bias
Confirm self-reported outcome data
Medical records
Collect detailed information about outcome to characterize it as well as possible

49. Randomized controlled trial
Outcome information may be biased
Observer bias: example: interpretation of signs of disease in a vaccine trial (discount possible signs among placebo participants)
Participant bias possible if trial is unblinded, and if the outcome is obtained by self-report (e.g. migraine headaches)

50. Outcome-dependent follow-up
All follow-up studies (cohort or RCT)
Outcome-dependent follow-up creates an opportunity for bias in outcome information
Example:
Study of sexually transmitted infections
Participants encouraged to return for “problem visits” if they think they have symptoms
Result: factors related to vigilance will be related to apparent risk of sexually transmitted infection

51. Nondifferential misclassification: bias towards null
Exp
No exp
Dis 50 10 60
No dis
250
290
540
300
600
Exp
No exp
Dis 70
( ) 38
( )
108
No dis
230
( )
262
( )
492
300
600
NO BIAS
RR=5
BIAS
(10% outcome error)
RR=1.8

52. Nondifferential misclassification: bias towards null
Bias towards null: conservative bias
Therefore considered “safe”
If you observe a significant association, you can be sure that it is valid (true association is even stronger)

53. Differential misclassification
Can result in bias towards null (conservative)
Can result in bias away from null (inflated association) (may produce spurious association)

54. Differential misclassification: bias away from null
Exp
No exp
Dis 50 10 60
No dis
250
290
540
300
600
Exp
No exp
Dis 70
( ) 10 80
No dis
230
( )
290
520
300
600
NO BIAS
RR=5
BIAS
(10% outcome error among exposed)
RR=7

55. Differential misclassification: bias away from null
Exp
No exp
Dis 50 10 60
No dis
250
290
540
300
600
Exp
No exp
Dis 75
(50+25) 10 80
No dis
225
(250-25)
290
520
300
600
NO BIAS
RR=5
BIAS
(outcome specificity among exposed)
RR=7.5

56. Differential misclassification: bias towards null
Exp
No exp
Dis 50 10 60
No dis
250
290
540
300
600
Exp
No exp
Dis 45
(50-5) 10 55
No dis
255
(250+5)
290
545
300
600
NO BIAS
RR=5
BIAS
(outcome sensitivity among exposed)
RR=4.5

57. Differential misclassification: bias towards null
Exp
No exp
Dis 50 10 60
No dis
250
290
540
300
600
Exp
No exp
Dis 50 39
(10+29) 89
No dis
250
261
(290-29)
511
300
600
NO BIAS
RR=5
BIAS
(outcome specificity in unexposed)
RR=1.3

58. Discussion in class next time
Selection bias
Healthy worker survival effect
Two articles
Compensating bias, control selection: example of coffee and pancreatic cancer