Pawin Numthavaj M.D. ([email protected])
SSection for Clinical EpidemiologyFaculty of Medicine Ramathibodi Hospital
Mahidol Universityhttp://www.ceb-rama.org
• Association and cause– Causal criteria
• Bias & Confounders– Types of Bias– Controlling for Bias
• Interaction– Types of Interaction
2
3
Association
Cause4
• An association is a statistical relationship between two or more events, characteristics or other variables.
5
Alcohol breath test
(mg%)
Case (%)
N = 43
Control (%)
N = 113
OR(95% CI)
P-value
Alcohol breath test
> 00
11 (44)14 (56)
9 (8)103 (92)
8.99 (3.17-25.52)1
<0.001
6
7
• Artifact from bias ?• Chance ?• Confounding ?• Causation ?
8
9
AAssociation• An association is a
statistical relationship between two or more events, characteristics or other variables.
Cause• A cause of disease event
is an antecedent event, condition, or characteristic that was necessary for the occurrence of the diseaseat the moment it occurred, given that other conditions are fixed.
Kenneth J. Rothman & Sander Greenland 10
• Both suspected cause and effect must be associated if they are causally related, but not all associations are causal.
• When statistical associations emerge from clinical research, they do not necessary imply causal associations.
11 12
13
1. The specific organism should be shown to be present in all cases of animals suffering from a specific disease but should not be found in healthy animals.
2. The specific microorganism should be isolated from the diseased animal and grown in pure culture on artificial laboratory media.
3. This freshly isolated microorganism, when inoculated into a healthy laboratory animal, should cause the same disease seen in the original animal.
4. The microorganism should be reisolated in pure culture from the experimental infection.
Robert Koch, Discovered the TubercleBacillus, 1882
14
15By [mike jones] - Own work, CC BY-SA 3.0, https://commons.wikimedia.org/w/index.php?curid=24819700 16
• Proposed by Hill, AB• 1965 paper• 9 elements that support
the strength of association to be “cause”
SSir Austin Bradford Hill(8 July 1897 – 18 April 1991)
17
• More strength – more possibility of cause• Relative Risk / Odds Ratio• Cholera epidemic (Snow 1855)
– Polluted water supply by Southwark and Vauxhall Company: 71 death/10,000 house
– Lambeth Company: 5 death/10,000 house
18
19 20
21 22
• Has the observed association been repeatedly observed by different persons, in different places, circumstances and times?
23
• If the association is limited to specific workers (patients) and to particular sites (area) and types of disease and there is no association between the work and other modes of dying, then clearly there is a strong argument in favor of causation
• Does exposure lead only to outcome?• Weak criterion: some exposure (Ex.
smoking) leads to multiple outcomes
24
• Time Order• Exposure always precede outcome• Essential criterion• Study design: Cohort, RCT, Case-control
Cause Effect
Time25
• Dose-response curve• Rate of lung cancer rises linearly with
number and duration of of cigarettes smoked daily
26
0
5
10
15
20
25
30
35
40
45
<20 20-29 30-39 40-49 50+
Rel
ativ
e ris
k
Years of smoking
<10 cigarettes perday
Figure 1.1: Relative risk of lung cancer, according toduration and intensity of smoking, men
http://www.cancerresearchuk.org/cancer-info/cancerstats/causes/lifestyle/tobacco/tobacco-and-cancer-risk
• Biologically plausible• Does the association make sense?• Depend on our lack of knowledge in the
field
27
SmokingPolycyclic
aromatic carbonCarcinogen
DNA adductsMutation and other changes
Lung Cancer
28
• “Internal consistency”• Is the association consistent with other
available evidence?• Consistently evident within subgroups• If smoking cause lung cancer
– It should cause in men, women, those finished high school, those finished university
29
• Has a RCT been done?
30
• Is the association similar to others?• Indirect information• Ex:
– Oral contraceptive cause MI in smoker– Oral contraceptive cause stroke
31 32
• Precision (↓ Random error)• Validity (↓ Systematic error)
– External Validity• Generalization• How to generalize study result to general
population– Internal Validity
• How close the study findings are to the truth
Population of patients with condition of interest
Sample
Sample
Group 2Group 1
Conclusion
1 G
Conclusionclus
Bias
Chance
Internal Validity
• Measurements are based on a sample.• Number of selected samples is important for
the precision of the measurements
Number of sample == Precision
Population: prevalence = 20%
Sample 1: prevalence = 0%
Sample 2: prevalence = 20%
Sample 3: prevalence = 40%
• Larger sample size provides more narrow interval
• N=1000 : 30% (95%CI 25 -36%)• N=100 : 30% (95% CI 9 -58 %)
N = 1000
N = 100
30%
• For comparison of 2 or more• Tell the probability (percentage) of a
wrong conclusion of study result• Ex: Association between
–Number coffee drinking per day–Developing CA stomach–RR = 3.6– X2
, P value = 0.35
• External validity of Generalization• Internal validity
– Bias• Selection bias• Measurement bias
– Confounder
40
• “Deviation from the truth”• Undermines iinternal validity of research
41
• Sackett: 35 different classifications• Feinstein:
– Susceptibility bias (difference in baseline)– Performance bias (different proficiencies of
treatment)– Detection bias (different measurements of
outcome)– Transfer bias (differential losses to follow-up)
42
1. Selection bias2. Information bias3. Confounding
43
• Are the groups similar in all important respects?
• MMembership bias– Member of group differ from others– Ex: – Objective: Jogging & prevention of MI– Design: Cohort from Joggers in club vs general
student– Joggers may be different in smoking history, diet
44
• BBerkson bias– Different rates of admission for case and
control– Ex: – Objective: Intrauterine device & salphingitis– Design: Case control where case is from
patient who was admitted with salphingitis, control is general patient in gynecologic ward
– Doctor are prone to admit more patient with history of IUD use
45
• NNeyman bias– Disease that is quickly fatal– Ex:– Objective: Snow shoveling & MI– Design: Case control where case is from
patient who was admitted with MI, control is general patient in ward
– More snow shoveling cases died at the site and never reach hospital – so less case in the hospital
46
• UUnmasking bias– Exposure unmasked the outcome– Ex:– Objective: Estrogen therapy & Entrometrium
cancer– Design: Cohort in patient with estrogen/no
estrogen– Estrogen therapy cause symptomless cancer
to bleed, thus showing sign of cancer even if it is already there
47
• NNon-respondent bias– Nonresponders are different than responders– Ex: – Objective: Smoking & vitamin taking habit– Design: Cross sectional with food
questionnaire– Smokers are less likely to return
questionnaires
48
• Define criteria of selection of case and control independent of exposure in case-control study
• Define criteria of selection of exposed and non-exposed independent of disease outcome in cohort study
• Use RCT
49
• Has information been gathered in the same way?
• Synonym: “Observation” “Classification” “Measurement” bias
• Outcomes should be obtained in the same way for exposed and unexposed
50
• AAscertainment bias– Gathering information in different ways– Ex: bedside interview for case but telephone
interview for control• Diagnostic suspicion bias
– Knowledge of exposure leads to intensive search for disease
– Ex: IV drug users are likely to have HIV tested by their physicians
51
• RRecall bias– Cases tends to remember better than healthy
controls– Ex: breast cancer patients may recall their
abortion history better than controls
52
• Information bias might not leads to one direction (increase/decrease RR) but may also leads to obscured difference
• Ex: Ambiguous questionnaires leads to error in data collection in both case and control found no association
53
• SStandardized protocol for data collection• Methods of data collection similar• Interviewers and study personnel are
unaware of exposure/disease status• Strategy to assess potential information
bias
54
• Is extraneous factor blurring the effect?
EffectEffect
55
• Associated with the exposure• Affects the outcome• But not an intermediate link in the chain of
causation between exposure and outcome
56
Cigarette smoking as a confounder of the coffee drinking-cancer of pancreas relationship.
Zaccai J H Postgrad Med J 2004;80:140-147
Copyright © The Fellowship of Postgraduate Medicine. All rights reserved.57
• Confounding can be corrected if it was anticipated and requisite information gathered
• Possible approaches:– Restriction– Matching– Stratification
58
• RRestriction– Enroll only people without confounder– Ex: Enroll only non-smokers– Cons:
• Hinders recruitment• Cannot extrapolate to those with confounder e.g.
cannot extrapolate to smokers (poorer external validity
59
• MMatching– Match case and control with smoking status– Cons
• Recruitment• Cannot examine effect of matched variables
• Stratification– Post hoc analysis– Results are calculated separately for smokers
and non-smokers to see if the same effect arises independent of smoking
60
Figure 1. Example of confounding in a hypothetical cohort study of intrauterine device use and salpingitis. When the crude relative risk is controlled for the confounding effect of number of sexual partners, the raised risk disappears.
David A Grimes , Kenneth F Schulz BBias and causal associations in observational research The Lancet Volume 359, Issue 9302 2002 248 – 252 http://dx.doi.org/10.1016/S0140-6736(02)07451-2
61 62
IUD Multiple Sexual Partner
Salphingitis
• MMultivariate techniques– Mathematical modeling– Similar to stratification but can control for
multiple variables
63 64
• Effect between 2 risk factors and outcome • Effect of one risk factor is different within
strata defined by the other
65
• Effect of aflatoxin in chronic hepatitis B on the development of liver cancer– RR for CA liver from HBV = 7.3– RR for CA liver from aflatoxin exposure = 3.4– RR for CA liver from both = ? 59.4!
66
• If the observed risk for having both A and B is equal to the expected, then there is nno iinteraction
• If the observed risk for having both A and B is greater than expected risk, then there is synergistic interaction
• If the observed risk for having both A and B is less than expected risk, then there is antagonistic interaction
67
Relative RiskFactor A
- +
Factor B- RR00 RR10+ RR01 RR11
No interaction : RR11 = RR10 RR01Synergistic Interaction : RR11 > RR10 RR01
Antagonistic interaction : RR11 < RR10 RR01
68
Relative RiskSmoking
No Yes
AlcoholNo 1.00 1.53Yes 1.23 5.71
Expected RR for both smoking and alcohol = 1.23 1.53 = 1.88In reality it is 5.71 (>1.88)So it is “Synergistic Interaction"
69
• EEffect modification occurs when a factor (an effect modifier) modifies the causal relationship between a risk factor and the outcome.
• For example, immunization status is a strong effect modifier of the relationship between exposure to a specific pathogen and subsequent disease from that pathogen.
• As an effect modifier, immunization status then modifies the biologic response of the person.
• The presence of effect modification supports a causal relationship because it implies a biologically plausible process.
• Interaction: Statistics term– Joint effect found from data analysis – Not necessary for biological/mechanism
explanation • Effect modification: Epidemiology term
– Joint effect found in population-base analysis– Not necessary for biological/mechanism
explanation
• When there is interaction in terms of the ratio measures of association (in this case, the risk ratio), we call this mmultiplicative interaction
• Interaction in the risk difference is called additive interaction
• Use stratification analysis (subgroup) to identify
• Interfere the finding • Prevent, control, role out• Asymmetry
• Use stratification analysis (subgroup) to identify
• Show real finding • Understand, explain,
report• Symmetry
CConfounding Interaction
74