Epid 600 Class 8 Bias

EPID 600; Class 8 Bias University of Michigan School of Public Health

1

Bias Systematic error in the design, conduct or analysis of a study that results in a mistaken estimate of an exposure’s effect on disease

2

Bias Systematic error in the design, conduct or analysis of a study that results in a mistaken estimate of an exposure’s effect on disease

Wrong study design! Wrong sampling strategy!

3

Bias Systematic error in the design, conduct or analysis of a study that results in a mistaken estimate of an exposure’s effect on disease

Problems in enrollment of cases, of controls! Loss to follow-up! Poor collection of data! 4

Bias Systematic error in the design, conduct or analysis of a study that results in a mistaken estimate of an exposure’s effect on disease

Wrong modeling assumptions! Miscategorization of variables! 5

Rothman KJ. Epidemiology: An Introduction. Oxford, 2002.

6

Evaluating bias 1. 

Why did it occur?

2. 

What effect does it have on the observed association?

3. 

What can be done to control for bias in this study and to prevent it in future studies?

7

Types of (important) bias 1. 

Selection bias Error in selection of study participants

2. 

Information bias Errors in procedures for gathering relevant information

8

1. Selection bias Systematic error in selecting subjects into one or more of the study groups, such as cases and controls, or exposed and unexposed

9

Study question Does coffee drinking cause pancreatic cancer?

10

Selection Bias: in a case-control study Cases: patients hospitalized with a diagnosis of pancreatic cancer Controls: patients hospitalized for other reasons by the same gastroenterologist who had hospitalized the case Results: found a strong relationship between coffee drinking and pancreatic cancer

11

What happened?

POPULATION

Persons who do not drink coffee are more likely to be controls

Cancer No

Yes No Cancer Yes

Coffee

Coffee

Yes

No

Yes No STUDY SAMPLE 12

Study question Is there a relation between occupational exposure to asbestos and lung cancer?

13

Selection Bias: in a cohort study Exposed: workers who handle asbestos (100% participation) Unexposed: workers in other areas of the factory who agree to participate (50% participation) Results: found NO relationship between asbestos and lung cancer

14

What happened? UNEXPOSED workers who participate are those at high risk for lung cancer, so unexposed with disease are overrepresented

POPULATION Cancer No

Yes No

Cancer

Asbestos

Asbestos

Yes

Yes

No

Yes No STUDY SAMPLE 15

2. Information Bias Systematic error in obtaining information regarding subjects in the study Examples: bias in recall, in collecting data, in interview, in reporting

16

Study question Is perinatal infection associated with a risk of congenital malformation?

17

Information Bias in a case-control study: Example 1 Cases: newborns with congenital malformations Controls: healthy newborns Results: found a strong relationship between mother’s recall of infection during pregnancy and malformation

18

What happened? Recall bias Parents of children with congenital malformations were more likely to report infection during pregnancy than parents of children without congenital malformations

19

What happened?

Yes No

Congenital Malformation Yes Infection during pregnancy

Infection during pregnancy

POPULATION Congenital Malformation Yes No

No

Yes No STUDY SAMPLE

20

What happened?

Misclassification of unexposed as exposed is more common in cases than in controls  DIFFERENTIAL MISCLASSIFICATION

Yes No

Congenital Malformation Yes Infection during pregnancy

Infection during pregnancy

POPULATION Congenital Malformation Yes No

No

Yes No STUDY SAMPLE

21

What happened?

Yes

Misclassification of unexposed as exposed is more common in cases than in controls  DIFFERENTIAL MISCLASSIFICATION

No

Congenital Malformation Yes Infection during pregnancy

Infection during pregnancy

POPULATION Congenital Malformation Yes No

No

Yes No STUDY SAMPLE

22

What if there is misclassification and it is similar in both cases and controls ?

Infection

Case Non-Case Yes No

Non-differential misclassification Usually biases estimate of association towards 1 (the null) 23

“Toward the null”

3

2 ”the null” 1 0.5 0 24

Study question Is smoking associated with an increased risk of myocardial infarction (MI) ?

25

Information Bias in a case-control study: Example 2 Cases: hospitalized cases of MI in elderly adults Controls: elderly adults, randomly selected from the community, who have never been hospitalized for MI Results: found a weak relationship between smoking and MI

26

What happened? Many true cases of MI are misclassified as non-cases, and are included in the controls (they were not hospitalized and had no symptoms)

27

What happened?

Misclassification of cases as controls is similar in smokers and nonsmokers  NON-DIFFERENTIAL MISCLASSIFICATION

Yes No

Myocardial Infarction Yes Smoke

Smoke

POPULATION Myocardial Infarction Yes No

No

Yes No STUDY SAMPLE

28

What happened?

Misclassification of cases as controls is similar in smokers and non-smokers  NONDIFFERENTIAL MISCLASSIFICATION

Yes No

Myocardial Infarction Yes Smoke

Smoke

POPULATION Myocardial Infarction Yes No

No

Yes No STUDY SAMPLE

29

What happened?

Misclassification of cases as controls is similar in smokers and non-smokers  NONDIFFERENTIAL MISCLASSIFICATION

Yes No

Myocardial Infarction Yes Smoke

Smoke

POPULATION Myocardial Infarction Yes No

No

Yes No STUDY SAMPLE

30

Study question Is use of oral contraceptives (OC) associated with an increased risk of venous thrombophlebitis (blood clots)?

31

Information Bias: in a cohort study Exposed: women who use OC Unexposed: women who do not use OC Results: found a strong relationship between OC use and thrombophlebitis

32

What happened? Detection bias (also called surveillance bias) Women who are on oral contraceptives are more likely to receive a diagnosis of thrombophlebitis

33

What happened?

POPULATION

Yes No Thrombophlebitis Yes OC Use

OC Use

Thrombophlebitis Yes No

No

Yes No STUDY SAMPLE

34

What happened?

Misclassification of non-disease as disease is different in exposed and unexposed persons  DIFFERENTIAL MISCLASSIFICATION

POPULATION

Yes No

Thrombophlebitis Yes OC Use

OC Use

Thrombophlebitis Yes No

No

Yes No STUDY SAMPLE

35

What happened?

Misclassification of non-disease as disease is different in exposed and unexposed persons  DIFFERENTIAL MISCLASSIFICATION

POPULATION

Yes No

Thrombophlebitis Yes OC Use

OC Use

Thrombophlebitis Yes No

No

Yes No STUDY SAMPLE

36

Putting numbers to the differential vs. nondifferential examples, 1 Misclassification of non-disease as disease is different in exposed and unexposed persons  DIFFERENTIAL MISCLASSIFICATION RESULTING IN BIAS AWAY FROM THE NULL

POPULATION

OC Use

Thrombophlebitis Yes No Yes No

50 50

25 100

Thrombophlebitis Yes No

(100*50)/ (50*25)=4

OC Use

REAL OR = Yes No

70 50

5 100

STUDY SAMPLE

BIASED OR = (100*70)/ (50*5)=28 37

POPULATION Congenital Malformation Yes No

50 No 50

Yes

REAL OR = (100*50)/ (50*25)=4

Misclassification of unexposed as exposed is more common in cases than in controls  DIFFERENTIAL MISCLASSIFICATION RESULTING IN BIAS AWAY FROM THE NULL

25 100

Congenital Malformation Yes

Infection during pregnancy

Infection during pregnancy

Putting numbers to the differential vs. nondifferential examples, 2

75 No 25

Yes

No

25 100

STUDY SAMPLE

BIASED OR = (100*75)/ (25*25)=12 38

Putting numbers to the differential vs. nondifferential examples, 3 Misclassification of exposed as unexposed is more common in cases than in controls  DIFFERENTIAL MISCLASSIFICATION RESULTING IN BIAS TOWARDS THE NULL

POPULATION Disease Exposure

Yes

50 No 50

Yes

No

25 100

Disease

(100*50)/ (50*25)=4

Yes Exposure

REAL OR =

Yes No

25 75

No

25 100

STUDY SAMPLE

BIASED OR = (100*25)/ (25*75)=1.3 39

Putting numbers to the differential vs. nondifferential examples, 4 Misclassification of cases as controls is similar in smokers and non-smokers  NONDIFFERENTIAL MISCLASSIFICATION RESULTING IN BIAS TOWARDS THE NULL

Smoke

POPULATION Myocardial Infarction Yes No

50 No 50

Yes

25 100

Myocardial Infarction

REAL OR = (50*25)=4

Smoke

(100*50)/

Yes

25 No 25

Yes

No

50 125

STUDY SAMPLE

BIASED OR = (125*25)/ (25*50)=2.5 40

Accuracy of weight/height reports Obesity is acknowledged as a critical health problem internationally Studies often use reported (as opposed to measured) data to estimate the prevalence of overweight and obesity at the population level There have been investigations regarding the “truth” of these reported values in adults and adolescents; the validity of parent-reported weight and height was studied by a team in Canada.

Dubois and Girad. Accuracy of maternal reports of pre-schoolers’ weights and heights as estimates of BMI values. Int J Epid. 2007; 36: 132-138.

41

Height/weight reports

1) Mothers asked to report on height and weight of children aged 4

2) Within 3 months, children’s weight and height were directly measured

3) Investigators examined the prevalence of obesity based on reported values versus prevalence of obesity based on measured values

Dubois and Girad. Accuracy of maternal reports of pre-schoolers’ weights and heights as estimates of BMI values. Int J Epid. 2007; 36: 132-138.

42

Height/weight reports The cohort: 4-year old children in 2002, who were part of a regional stratified sample of children born in Quebec in 1998 Height/Weight report: One care-giver, usually the mother, reported height and weight to an interviewer; the caregiver was not told that subsequent measurement would be taken. Interviewers made sure that mothers recalled these values rather than measuring them on the spot Height and weight measurement: Within three months of the interview, nutritionists followed a standardized protocol and measured height and weight of children

Dubois and Girad. Accuracy of maternal reports of pre-schoolers’ weights and heights as estimates of BMI values. Int J Epid. 2007; 36: 132-138.

43

Height/weight report: is it the same for all? Is any group of people consistently overreporting BMI of children? Odds ratios among boys: BMI>95th Percentile SES

Reported

Measured

Highest

1

1

Middle

1.8

1.7

Lowest

2.2

1.9

Dubois and Girad. Accuracy of maternal reports of pre-schoolers’ weights and heights as estimates of BMI values. Int J Epid. 2007; 36: 132-138.

44

Height/weight reports In this figure, the measured weight is 17 kg for a 51month-old child who is 1.03m tall. This child ranks at the 71st percentile if the child is a girl and at the 65th percentile if the child is a boy. If the mother reports the weight as being 2 kg less than the actual value, the child would be classified as being below the 15th percentile.

Dubois and Girad. Accuracy of maternal reports of pre-schoolers’ weights and heights as estimates of BMI values. Int J Epid. 2007; 36: 132-138.

45

Height/weight report: findings Heights were reported more accurately than weights (there was no difference in the means of reported vs. measured heights) A greater proportion of mothers overestimated boys weights; a greater proportion of lower SES mothers misreport 12% of the children were classified as overweight based on the reported data; 9% were classified as overweight using measured data  3% overestimation of overweight in this population

Dubois and Girad. Accuracy of maternal reports of pre-schoolers’ weights and heights as estimates of BMI values. Int J Epid. 2007; 36: 132-138.

46

Special biases Non-respondent bias Persons who do not participate in a particular study may be different than those who do e.g., in telephone surveys, women are more likely to answer surveys than are men; if the exposure of interest is differentially distributed between women and men and if gender is associated with the outcome of interest bias will result

47

Other special biases Unmasking (detection signal) bias Membership bias Diagnostic suspicion bias Exposure suspicion bias Recall bias Family information bias Neyman bias Berkson bias etc 48

Evaluating Bias 1. 

Why did it occur?

2. 

What effect does it have on the observed association?

3. 

What can be done to control for bias in this study, and to prevent it in future studies?

49

Preventing Bias Careful attention to sampling Minimize non-response Standardization of measurements Training and quality control Blinding

50

51