Bias and Confounding Play or Chance Measure of Association

1
Bias and ConfoundingPlay or ChanceMeasure of
Association

• Introduction to Epidemiology
• Fall, 2000

2
Objectives - Bias

• Define the following
• Bias
• Selection bias
• Information bias
• Recall bias
• Interviewer bias

3
Objectives - BIAS

• Describe the effects of bias on measures of
  association
• Describe procedures for controlling selection and
  information bias
• Learn to identify possible sources of bias,
  effects of bias, and means of reducing bias

4
Definition

• Bias is introduced by any systematic error in the
  design, conduct, or analysis of a study that
  results in a mistaken estimate of the exposures
  effect on the risk of disease.
• Schlesselman Stolley, 1982

5
Accuracy and Sources of Error

• Purpose of epidemiologic study
• To estimate the effect of an exposure on an
  outcome
• Main objective
• To measure the exposure and outcome accurately
• That is, to measure without error

6
Measures of Association

• Epidemiologist tend to view cause and effect as
  binary variables
• Either you are exposed (or diseased)
• Are you arent exposed (or diseased)
• How we measure these variables can have a
  profound influence on our results

7
Validity

• Validity the degree to which the data measure
  what they were intended to measure
• Bias systematic error
• Selection bias
• Information bias
• Confounding
• Precision random error

8
Logistical factors

• Measures are often chosen because they are
  inexpensive
• More important might be
• Likelihood of compliance
• Ease of entering and evaluating data

9
Ethical issues

• High risk procedures cannot be preformed on all
  subjects
• Coronary angiography
• Laporoscopy
• Prostate biopsy

10
Importance

• Death is certainly important to the individuals
  involved
• But it occurs infrequently
• Surrogate measurements are used
• Fall in blood sugar
• Improvement of blood pressure
• Regression of tumor

11
Sensitivity

• The measured variable needs to be related to the
  real exposure of interest
• Measured blood pressure ? end organ damage ?
  death
• Measured blood glucose ? end organ damage ? death
• Serum cholesterol ?CVD ?death

12
Types of variables

• Categorical
• dead / alive
• diseased / normal
• white / black / Hispanic
• Continuous
• diastolic blood pressure
• hemoglobin level
• pain, mood, disability

13
Are these results valid?

• We know how to measure associations (RR, OR, AR,
  EF)
• We can explain these associations with words

HOWEVER
14
Validity

• Do these results tell us what really happened?
• Do OCs increase risk of uterine cancer?
• Does streptokinase reduce cardiac mortality?
• Do education programs reduce the prevalence of
  obesity?

15
Random or Systematic Errors

• Random error refers to imprecision
• Governed by chance
• Systematic error refers to mistakes
• Also called bias
• Is not random

16
Random Errors

• Random governed by chance
• small sample size
• biological variability
• instrument variability
• chance variation
• Can often be fixed by increasing the number of
  study subjects

17
Random Error

• effect size (magnitude of association) from
  repeated samples are likely to be distributed
  around the true effect

18
Systematic Errors

• BIAS is determined by
• Errors in
• Selection of subjects
• Collection of information
• Classification of exposure or outcome
• Drawing of conclusions

19
Precision

• Reliability of Observations
• Intra-observer (does the same observer get the
  same results given the same situation)
• Inter-observer (do different observers get the
  same results given the same situation)

20
Precision

• Errors attributable to the observer
• inexperience
• carelessness
• previous knowledge or beliefs
• inconsistencies in procedures
• coding errors
• human biases

21
Precision

• Errors attributable to instruments
• unreliable measuring scales
• faulty laboratory equipment
• inappropriate instruments
• surveys not validated for this population

22
Precision

• Errors attributable to observed
• biologic variation over time
• regression to the mean
• those who start with high values are likely to
  have lower values on second reading
• those who start with low values are likely to
  have higher values on second reading

23
Random Error (imprecision)
Systematic Error (Bias)
High
Low
Low
High
24
Threats to validity

• Internal validity
• do these results represent what is really
  happening in the study population.
• are the results due to
• Bias
• Confounding
• Chance

25
Threats to validity

• External validity
• are these results generalizable to a larger
  population.
• how well does the study population reflect the
  general population?

26
Threats to Validity

• External population
• Target population?
• Study population
• Study Comparison
• Participants Group

27
Evaluate Validity

• Absence of systematic errors
• Findings represent the study sample
• Findings are generalizable to larger populations
• Internal validity is the primary objective
• Without internal validity
• there is no reason to generalize

28
Evaluate Validity

• GIGO

Garbage in garbage out
29
BIAS

• Bias has to do with research design
• Bias results from systematic flaws in
• study design
• data collection
• analysis
• interpretation

30
BIAS

• Bias is the difference between the expected value
  of an estimate and the real population
  parameter it purports to estimate
• Bias is an attribute of methodology

31
BIAS

• Two major types to consider
• selection bias non-comparable
• criteria used to enroll participants
• information bias non-comparable
• information obtained due to
• interviewer or recall bias

32
BIAS

• If the study population is selected in a way to
  represent the target population in terms of the
  distribution of the variables of interest, and
  the data is collected in a way to reflect the
  real status of the individual in terms of the
  presence or absence of the variables of interest,
  then the bias is minimized in the study.

33
BIAS

• Telephone survey at 10 a.m. Monday morning.
• Interview the first 100 people who answer the
  phone.
• Is this a representative sample?

• What groups would be systematically excluded from
  the sample?

34
Selection Bias

• A distortion in a measure of disease frequency or
  association resulting from the manner in which
  subjects are selected for the study

35
Selection Bias

• When the sample is not representative of the
  target population
• When selection was related to either exposure or
  disease

36
Selection Bias

• Alf Landon was predicted to win the election
  against Franklin Roosevelt
• Interviews by phone
• Few people had phones
• Rich people had phones
• Rich people were more likely to be Republicans

37
Selection Bias

• Contraceptive failure rate of IUDs versus OCs
• Previous studies
• IUD failure due to early expulsion
• OC failure due to not taking pills as prescribed

38
Selection Bias

• Selection 10,000 women each exposure group
• with IUDs for at least 1 year
• OC users
• Results
• OC failure 8
• IUD failure 5

39
Selection Bias

• The study population was not representative of
  all individuals who could have been included
• Factors making these women different could have
  affected the results

40
Selection Bias
41
Selection Bias
42
Selection Bias

• How to minimize selection bias
• always try to avoid human choice in the selection
  of a sample
• (depends on your study design)
• whenever possible, use random sampling mechanisms

43
Selection Bias

• Berksons bias or hospital admission rate bias
• hospitalized people are more likely to suffer
  from
• multiple illnesses,
• have more severe illnesses, and
• have less healthy lifestyles

44
Berksons Bias

• Is coffee associated with pancreatic cancer
• Case / controls from MDA
• Coffee drinking may be associated with other
  forms of cancer. The prevalence of coffee
  drinking in cancer patients may be higher than in
  the target population.

45
Selection Bias
46
Selection of Controls

• Hospitals
• Special groups friends, neighbors, relatives
• General populations
• Multiple comparison groups may solve some of the
  problems with using hospital-based controls.

47
Selection Bias

• Survival bias
• Exposed cases do not have the same survival as
  non-exposed cases
• Non-response bias
• Participants are different than non-participants
• Publicity bias
• News media may effect behavior

48
Selection Bias

• Healthy worker effect
• ill and chronically disabled people are excluded
  from the work-force
• Time or place bias
• health events or exposures may not occur
  symmetrically over time

49
Selection Bias

• Selection Bias involves errors in determining
• who to select and
• how they will be selected

50
Selection Bias

• On WHO to select
• We would want to select groups from the diseased
  and non-diseased populations that do not have a
  particular distribution of exposure that is
  different from that in the target population.

51
Selection Bias

• On HOW to select
• The choice of the study population might be
  valid, but the way we choose to sample from the
  study population might introduce bias

52
Selection Bias - WHO

• In a hospital based case-comparison study of the
  association of CHD and alcohol consumption the
  comparison group should consist of those without
  CHD. A poor choice of a non-CHD group would be
  patients admitted for cirrhosis of the liver,
  because of the known high alcohol intake level
  among that group.

53
Selection Bias - WHO

• The results from such a study will tend to show
  no association between alcohol intake and CHD
  simply because the comparison group that was
  chosen to had a high exposure level.

54
Selection Bias - WHO

• In studying the relationship between phlebitis
  and OC users, women who are OC users are subject
  to more medical surveillance and to more thorough
  examination.
• Identification of cases among these women - in
  whom the exposure of interest is high - is more
  likely than among women who are not OC users.

55
Selection Bias - WHO

• The results from such a study will tend to show
  a high association between OC use and phlebitis,
  simply because the cases were chosen in a way to
  have higher exposure than the cases in the target
  population.

56
Selection Bias - HOW

• In a community based case-comparison study of
  the association of CHD and alcohol consumption
  comparisons are recruited by placing ads in all
  the local community papers, including the Baptist
  Weekly Crier, and the Women's Temperance
  Newsletter.

57
Selection Bias - HOW

• The problem with such a sampling method, is that
  the volunteers who would respond to the ads might
  have different prevalence of exposure than that
  of the general population
• This type of bias is referred to as "volunteer
  bias"

58
Selection Bias

• Etiology of homosexuality (1962)
• Three questionnaires sent to members of the New
  York-based psychoanalytic society
• Psychiatrist complete forms on homosexual
  patients
• If fewer that 3 - use remaining forms for male
  heterosexuals as controls

59
Information Bias

• Assume your initial decision on who to select as
  diseased individuals is correct (i.e. your
  non-diseased individuals really do represent all
  non-diseased individuals in regard to exposure).
• However, you incorrectly divide them into exposed
  or non-exposed because you do not accurately
  measure the exposure (e.g. your information on
  exposure is faulty).

60
Information Bias

• If this happened to a different extent in the
  diseased and non-diseased groups then bias is
  introduced.

61
Information Bias

• Non-differential misclassification in a
  case-comparison study in regard to exposure will
  bias the odds ratio towards the null (towards
  1.0).
• misclassification occurs in the exact same
  proportion among the diseased and the non-diseased

62
Information Bias
63
Information Bias

• 10 of Diseased are misclassified as exposed
• 10 of Non-Diseased are misclassified as exposed
• 20 Diseased truly non-exposed
• 10 of 20 2
• 20-2 18
• 80 2 82

64
Information Bias

• Misclassification
• Interviewer
• Recall

65
Information Bias

• Reproducibility or precision
• The probability that multiple measurements of
  exposure or outcome will yield the same results
• Systematic
• Random

66
Information Bias

• An association between cervical cancer and
  circumcision of primary sexual partner was
  described in 1954

67
Information Bias

• The original study was criticized because it did
  not take into account religion.
• Jewish and Muslim men are more likely to be
  circumcised - and their religious beliefs may
  influence their sexual practices

68
Information Bias

• The first study also asked women about the
  circumcision status of their sexual partners
• A second study was conducted and information was
  collected on religion, and on circumcision from
  females and from their sexual partners

69
Information Bias

• Also - men were asked to confirm their
  circumcision status with a physical examination

70
Interviewer Bias

• Interviewer Bias Example
• An interviewer might ask the comparisons
• INTERVIEWER On the average how many cups of
  coffee did you drink per day when you were 25
  years old?
• COMPARISON About two
• INTERVIEWER Thank you

71
Interviewer Bias

• INTERVIEWER On the average how many cups of
  coffee did you drink per day when you were 25
  years old?
• CASE About two
• INTERVIEWER Are you including coffee from
  coffee breaks, what about decaffeinated coffee is
  that included?
• CASE OH! OOPS, No, No, No, Not two. Three
  cups all decaffeinated

72
Interviewer Bias

• Obviously information about exposure was
  prompted more thoroughly from the case than from
  the control, possibly leading to
  misclassification of exposure more among the
  controls than among the cases.

73
Recall Bias

• Additionally if there has been publicity on the
  adverse effects of coffee, particularly in regard
  to cancer who do you think is more likely to
  overestimate, or perhaps recall more accurately,
  their past coffee consumption?
• Cases or controls?
• Why?

74
Bias and Measures of Association

• Depending on how they operate in specific
  circumstances, selection bias and information
  bias can distort the true association in every
  conceivable way They can
• Create a positive or negative association where
  none exists.
• Change an association from positive to negative,
  or vice versa.

75
Bias and Measures of Association

• Make an association appear stronger than it truly
  is.
• Make an association appear weaker than it truly
  is, or eliminate it entirely.

76
Controlling BIAS

• Prepare a manual that describes in detail the
  procedures for selecting participants. Avoid
  letting the interviewer choose who will be
  selected.
• Thoroughly train study personnel in these
  procedures
• Standardization of procedures, including tight
  control over the conduct of these procedures

77
Controlling BIAS

• In a hospital-based study, consider the
  possibility of obtaining a second control from
  the general population
• Select a population that can be followed with
  little or no loss to follow-up
• Choose study groups to be representative of the
  target groups

78
Controlling BIAS

• Prepare a detailed manual of operations that
  covers all aspects of data collection. No room
  for individual interpretation of procedures
  should be permitted.
• Thoroughly train all study personnel in those
  procedures. Establish minimum criteria for
  performance in key areas.

79
Controlling BIAS

• In multi-center projects, use central facilities
  for interpreting and analyzing data, e.g., a
  central laboratory for doing blood chemistries
• Maintain tight quality control, e.g., by sending
  blind replicates to your laboratory, retesting
  technicians, holding retraining sessions, and
  collecting data on reliability and validity
• Keep morale high for participants and study
  personnel

80
Controlling BIAS

• Sources of data and methods for collecting data
  should be the same for all participants
  regardless of exposure status or disease
• Participants should be unaware of specific
  hypotheses under investigation. Sometimes study
  personnel should also be kept unaware of specific
  hypotheses (but sometimes it is difficult to do
  this and keep morale high).

81
Controlling BIAS

• Whenever feasible, data on exposure in should be
  obtained by study personnel who are unaware of a
  participant's outcome status
• Data on occurrence of outcomes should be obtained
  and evaluated without knowledge of exposure
  status.
• Consider the possibility of collecting data that
  may help determine whether information bias has
  occurred, and, if so, its direction and magnitude.

82
Confounding

• Objectives
• Define Confounding
• Describe the effects of confounding on magnitude
  of association
• Describe procedures for controlling confounding

83
Confounding

• a mixing of effects
• between the exposure, the disease, and other
  factors associated with both the exposure and the
  disease
• such that the effects the effects of the two
  processes are not separated.

84
Confounding

• A bias due to the association of a third variable
  with both the exposure and the disease
  independently and the failure to disassociate the
  third variable from the association under study

85
Confounding

• A situation in which the effects of two processes
  are not separated. The distortion of the
  apparent effect of an exposure on the risk is
  brought about by the association with other
  factors that can influence the outcome.
• "Admixture of effects"

86
Confounding

• What is a confounding variable?
• A variable which distorts an association wholly
  or partially due to its association with both the
  outcome (disease) and the exposure under study
  independently.

87
Confounding

• the variable must be associated with the disease
  (i.e., the confounder itself may be a risk
  (factor).
• the variable is associated with the exposure
  independently of the disease
• the results of the association under study must
  be confounded (i.e., the result achieved is
  false)

88
Confounding

• IT IS NOT NECESSARY THAT THE CONFOUNDING VARIABLE
  BE CAUSALLY OR SIGNIFICANTLY ASSOCIATED WITH THE
  DISEASE OR EXPOSURE

89
Confounding
Coffee Observed Association
Cancer Presumed causation
Smoking, Alcohol, other Factors
90
Confounding
Low SES
Hypertension
Race/Ethnicity
91
Confounding
Obesity
Hypertension
Age
92
Confounding
Gambling
Cancer
Smoking, Alcohol, other Factors
93
Confounding

• HYPOTHESIS Is the incidence of coronary heart
  disease greater among men who drink coffee than
  among men who do not drink coffee
• DISEASE Coronary heart disease
• EXPOSURE History of coffee drinking
• POTENTIAL CONFOUNDER Smoking

94
Confounding

• To assess whether or not smoking confounds the
  association between coronary heart disease and
  coffee drinking three questions must be answered.
• What are these three questions?

95
Confounding

• 1) Is smoking associated with coffee
  drinking? exposure
• 2) Is smoking associated with coronary heart
  disease? disease
• 3) Are the stratified odds ratios for the
  association between the exposure and the disease
  different than the crude odds ratio?

96
Confounding
Risk Factor Independent Variable Coffee
Disease Dependent Variable CHD
Covariable Confounder Smoking
97
Confounding

• Are the results of the original analysis actually
  confounded by the potential confounder?
• Examine the analysis stratified by the potential
  confounder
• If the association are different by strata than
  confounding has been demonstrated

98
Confounding

• Detecting and removing spurious associations
  related variables can be done at
• the design stage, and/or
• the analysis stage

99
Control of Confounding

• Design stage
• restriction
• matching
• Analysis stage
• stratification
• multivariate techniques

100
Restriction

• Confounding cannot occur if the factor does not
  vary.
• For example if the study is limited to black
  women, race and gender cannot be confounding
  variables.
• However if restriction is carried to extremes the
  study may have a limited number of eligible
  participants

101
Restriction

• Restriction also limits the interpretation of the
  study.
• Often partial restriction is used.

102
Matching

• Matching is used mainly in case-comparison
  studies.
• Application of restraints to the comparison
  group to make it more similar to the case group
  is respect to one or more potential confounding
  variables.

103
Matching

• How close should matching be? Matching may be
  done on an individual basis (pair-matching) or on
  a group basis (frequency matching)
• If a pair-matched design is used, then matching
  must be taken into account in the analysis.

104
Randomization

• Randomization is used in experimental studies to
  allocate individuals to treatment groups by
  chance with the purpose of ensuring that all
  potential confounders are equally distributed
  among the groups. It is not haphazard
  assignment. Randomization does not always
  achieve its purpose.

105
Stratification

• Examine the association within strata of the
  potential confounder.
• These strata-specific estimates can be combined
  together using weighted averages to give an
  unconfounded overall estimate of effect.
• direct age-adjustment
• SMRs
• Mantel-Haenszel procedure

106
Multivariate Analysis

• Similar to stratification, but permits the use of
  continuous independent variables. The models
  rest on certain assumptions and do not always
  give the right answer if the assumptions are
  violated.
• logistic regression
• Cox (proportional Hazard) model

107
Selection Bias

• If the exposure is oral contraception use and the
  outcome is uterine cancer - how can you ensure
  that those who use OCs are similar in all ways to
  those who do not.
• Where would you look for non-exposed subjects

108
Effect Modification

• Separate from confounding
• Extraneous factor that modifies the effect of an
  exposure
• Statistically described as interaction
• Difference in effect of one factor according to
  the level of another factor

109
Effect Modification

• Direct Biological and Public Health Relevance
• Synergy - the combined effect of individual
  factors has an impact that is not entirely
  predicted by the sum of their parts

110
Effect Modification

• In the presence of smoking
• oral contraceptive use is risky
• In the absence of smoking
• oral contraceptive use is safe
• Without information on smoking status
• OC risk cannot be generalized

111
Effect Modification

• A variable may a confounder an effect modifier -
  neither or both
• You want to reveal confounding and interaction -
  these are usually data driven phenomenon - not
  design flaws

112
Measures of association

• relative risk
• odds ratio
• attributable risk
• also called risk difference
• attributable risk percent
• Also called etiologic fraction

113
Measures of association
114
Relative Risk
The association between cardiac deaths and
treatment with cholestyramine. JAMA 251351-374,
1984.
115
Relative Risk

• Disease Occurrence Among Exposure and Non-Exposure

116
Relative Risk

• Mortality among cholestyramine group was 0.79
  times that of the placebo group.
• Mortality among cholestyramine group was reduced
  21

117
Odds Ratio
Tobacco smoking as a possible etiologic factor in
bronchogenic carcinoma a study of 648 proved
cases. JAMA 143329-336, 1950
118
Odds Ratio

• Odds of Exposed vs Non-Exposed Among Disease and
  Non-Disease Cases

119
Odds Ratio

• Individuals with bronchogenic carcinoma were 9.33
  times more likely to have been smokers than
  individuals without bronchogenic carcinoma.

120
Odds Ratio

• Odds of Exposed vs Non-Exposed Among Disease and
  Non-Disease Cases

121
Attributable Risk (Risk Difference)

• absolute rather than relative
• AR Ie Io
• AR 38/1906 30/1900
• 4.1/1000
• 4.1/1000 cardiac deaths are attributable to
  untreated high cholesterol levels

122
Attributable Risk Percent or Etiologic Fraction

• AR (Ie Io) / Ie
• AR (30/1900 38/1906) 30/1900
• 20.8
• 20.8 of all cardiac deaths were due to untreated
  high cholesterol levels.

123
Common Pitfalls in Research

• Failing to evaluate accuracy
• Drawing spurious conclusions
• Generalizing to inappropriate populations
• Failing to evaluate the role of chance
• Assuming causality based only on statistical
  significance

124
Bias in a Case Series

• no comparison group
• selection of study group cannot described
• no way of ascertaining confounding

125
Bias in a Case Control Study

• do the controls represent the population from
  which the cases were drawn
• are controls at similar risk of being exposed?
• is case status / control status similar
• survival bias
• volunteer bias
• information bias

126
Bias in a cross sectional study

• survival bias
• migration out of exposure
• cart before the horse bias
• confounding

127
Bias in a cohort study

• exposed and non-exposed from same base population
• Internal comparisons start with a
  cross-sectional study of a population sample
• External comparisons try to ensure that the
  non-exposed are similar in all ways to the
  exposed group.