Idea Transcript
Main objective of Epidemiology Inference to a population Example: Treatment of hypertension: Research question (hypothesis): Is treatment A better than treatment B for patients with hypertension?
Statistical Inference Study sample: 143 female nurses (30-50 years of age, with hypertension. Target population: female nurses (30-50 years of age, with hypertension). External population: males, females, all age groups, with hypertension.
External population
Study design: Clinical study: select a sample; 143 female nurses (30-50 yrs of age, with hypertension), administer randomly 2 treatments A and B.
Target Population
Statistical analysis on data.
Study Sample
Interference on population
Statistical Inference: Example Evaluate the performance of a drug before its release?? External population Target Population
Statistical Inference: Example Evaluate the performance of a drug before its release?? Based on the new drug’s performance among a sample of patients with the disease, a conclusion is drawn regarding the drug’s performance were it will be used among a population of patients with the disease.
Study Sample
Concerns in Epidemiologic studies
Concerns in Epidemiologic studies
I.
Internal validity: Was the study carefully designed and analyzed?
I.
Internal validity: Was the study carefully designed and analyzed?
II.
External validity (Generalizability): Are the results applicable to the external population?
II.
External validity (Generalizability): Are the results applicable to the external population?
III.
Reproducibility of the results (precision): Are the results reproducible?
III.
Reproducibility of the results (precision): Are the results reproducible?
Internal validity
Internal validity
The extent to which the analytic inference derived from study sample is correct for the target population
The extent to which the analytic inference derived from study sample is correct for the target population. External population
1. 2. 3. 4.
Target Population Study Sample
Sample selection bias
Sample selection bias
Sample should be representative of the target population. If the relation between the exposure and the outcome is different for those who do participate and those who would be eligible but do not participate, then we have selection bias.
OC
Hospital
Thrumbo.
if the sample is based from a referral of a specific community
Thrumbo.
20
15
OC
80
85
100
100
Thrumbo.
II. Self Selection Bias: OR=1.4
Thrumbo.
OC
50
OC
50
85
100
100
Total
Referral Bias:
Sample selection bias
OC Total
I.
Thromboembolism
Referral Bias Population
Sample selection bias Information bias Confounding bias Reverse causality
15
OR=5.7
subjects self select themselves into a study
Sample selection bias
Healthy Worker Effect
III. Healthy Worker Effect (HWE):
Lifting heavy weight pain
Back BP
F-up
Forestry workers
healthy workers remain for the study (un-employed, retired are less healthy)
BP BP
F-up
General population BP
Results showed NO association :WHY???
Sample selection bias
Sample selection bias
IV Loss to follow up: in Cohort studies
Loss to follow up: in Cohort studies
Dust
Dust
Chronic Lung Disease
Dust
CLD 1,000 (10%)
CLD 9,000 (1%)
Total 10,000
Dust
500 (2%)
9,500 (0%)
10,000
RR=2
Minimizing Selection Bias
To minimize selection bias: Select a representative sample of the target population.
Chronic Lung Disease
CLD
CLD
Total
Dust
900
8,910
9,810
Dust
490
9,500
9,990
RR=1.8
Information Bias
(Misclassification Bias)
Errors in the classifying of the exposure or outcome.
Misclassification of Exposure
Misclassification of Exposure If misclassification of exposure is related to outcome: Differential Misclassification If misclassification of exposure is not related to outcome: Non-Differential Misclassification
Example: Recall bias Interviewers bias Error during data entry
Misclassification of Outcome
Minimize Information Bias
Example Diagnostic bias If misclassification of outcome is related to exposure: differential misclassification If misclassification of outcome is not related to exposure: non-differential misclassification
To minimize information bias:
Standardize detection procedures Blind subjects and observers
Confounding Bias: Example 1
Confounding Bias: Example 2
Example 1: Does gambling cause cancer?
Example 2: TIME Magazine 1978: “Sonic DoomCan jet noise kill?”
Nevada state (gambling is legal)
Random sample Matched on age, 10 years Follow-up sex, residence, family income
Utah state (gambling is illegal)
Random sample Matched on age, 10 years sex, residence, Follow-up family income
If we eliminate gambling, prevent 86,000 cancer deaths a year????
The study was conducted in Los- Angeles. The study concluded: People living in areas where jet plane noise >90 decibels had a significant increased death rate. WHY???
Confounding Bias: Definition Is present when the association between an exposure and an outcome is distorted by an extraneous third variable (referred to a confounding variable).
Confounding Bias: Example 3 Example 3: Study the association between coffee drinking and lung cancer LC
Confounding Bias: Definition For a variable to be a confounder, it must be:
1. A risk factor the disease 2. Associated with the exposure being studied.
Confounding Bias: Example 3 Is smoking a confounding variable? Coffee
Coffee
Yes
No
Yes
80
15
No
20
85
OR= (80x 85)/ (15 x 20)= 22
Lung cancer
Smoking Smoking is a confounding variable.
What would you conclude????
Confounding Bias: Minimize bias
Data Analysis: Stratification Multivariate statistical techniques
Example 1: Association between place of residence & Chronic bronchitis Chronic bronchitis Residence
Research Design: Use of randomized clinical trial Restriction Matching
Confounding Bias: Stratification
Yes
No
Urban
194
2219
Rural
69
1208
RR= 1.48
Confounding Bias: Stratification
Example 1:
Example 1: Stratified by Smoking:
Residence
Smokers
Non-smokers
Bronchitis
Bronchitis
Yes
Yes
No
But 1.18 ≅ 1.17 = 1.48 (crude)
No
Urban
167
1094
27
1125
Rural
53
417
16
791
RR=1.18
Confounding Bias: Stratification
Smoking is a confounding variable!
RR=1.17
Confounding Bias: Stratification
Confounding Bias: Stratification
Example 2: Study the association between Alcohol consumption & Myocardial Infarction
Example 2: Stratified by smoking: Smokers
MI No
71
52
No
29
48
Alcohol
Alcohol
Yes Yes
Non-smokers
MI
No MI
MI
No MI
Yes
8
16
63
36
No
22
44
7
4
OR=1
OR=2.2
OR=1
Confounding Bias: Stratification
Confounding Bias: Stratification
Example 2:
Example 3: Study the association between Treatment 1 & Treatment 2 with survival But 1 = 1 = 2.2 (crude) Smoking is a confounding variable!
Alive
Dead
Total
T1
40
60
100
T2
60
40
100
100
100
200
Total
RR (Crude) = 0.67
Confounding Bias: Stratification
Example 3:
Example 3: Stratified by gender: Females
Confounding Bias: Stratification
Males
Alive
Dead
Alive
Dead
T1
24
3
16
57
T2
58
30
2
10
RR=1.35
But 1.35 ≅ 1.32 = 0.67 (crude) Gender is a confounding variable!
RR=1.32
Mantel-Haenzel: Calculating an adjusted RR
Mantel-Haenzel: Example Example:
ith stratum:
Females
Outcome No outcome Total E No E Total RR
MH
=
ai
bi
ci
di
Males
Success Failure Success Failure
Ni ∑ ai(ci + di) /Ni ∑ ci(ai + bi) /Ni
T1
24
3
16
57
T2
58
30
2
10
RR= 1.35
RR= 1.32
RRAdjusted= RRMH= 24(58+30)/115 + 16 (2+10)/85 58(24+3)/115 + 2 (16+57)/85 = 1.34
Concerns in Epidemiologic studies I.
Internal validity: Was the study carefully designed and analyzed?
II.
External validity (Generalizability): Are the results applicable to the external population?
III.
Reproducibility of the results (precision): Are the results reproducible?
External validity Also Known as generalizability. The extent to which the analytic inference derived from study sample is correct for the external population. External population
Target Population Study Sample
External validity
Concerns in Epidemiologic studies
External validity is dependent on internal validity.
I.
Internal validity: Was the study carefully designed and analyzed?
II.
External validity (Generalizability): Are the results applicable to the external population?
III.
Reproducibility of the results (precision): Are the results reproducible?
Maximize external validity by selecting study subject from a target population as similar as possible to the external population.
Reproducibility
Reproducibility
Subjects in a study are always a sample of a population.
Target Population
Repetitive sampling results in a range of estimates for different samples: Sampling variation. The smaller the sample, the less reproducible will be the sample estimate.
Sample 2 Sample 1
Study sample Sample 3
To decrease sampling error, increase sample size.
Effect Modification (Interaction)
Effect Modification (Interaction)
Example 1: Association between severe injury & death.
Example 1: Stratified by age:
Death
No Death
Severe injury
45
55
No Severe injury
6
94
RR = 0.45/0.06 = 7.5
65years
Death
No Death Death
No Death
15
35
30
20
28
4
66
No Severe injury 2
RR= 4.48
RR= 10.5
Effect Modification (Interaction)
Effect Modification (Interaction)
Example 1:
When the exposure- outcome relationship is different for the different levels of a third variable, we have interaction (effect modification). The crude RR is hiding important effects. 4.48 = 10.5 = 7.5 (crude) Age is an effect modifier!
Effect Modification (Interaction) Example 2:
Effect Modification (Interaction) Example 2:
D
No D
Total
E
200
1800
2000
No E
400
3600
4000
Total 600
5400
6000
RR= (200/2000) / (400/4000) =1
Female
1.69 = 3 = 1 (crude)
Males
Gender is an effect modifier!
D
No D
Total D
No D Total
E
110
390
500
90
1410
1500
No E
380
2620
3000
20
980
1000
Total 490
3101
3500
110 2390
RR= 1.69
2500
RR= 3
Effect Modification (Interaction) A Confounder or Effect modifiers? RR=2
Males
Hip Fracture
RR= 0.7
Old Males RR= 3
Effect Modification (Interaction) A Confounder or Effect modifiers?
Hip Fracture
Young Males
Solution: Report the RR separately for each category of the variable (for the different levels of the effect modifier).
RR=1.6
Obesity
Pre-menopausal Hip Fracture
Obesity
Breast cancer
RR= 1.1
Breast cancer
Post-menopausal Obesity
Breast cancer
RR= 2
Effect Modification (Interaction) A Confounder or Effect modifiers? Smoking
RR=3
Low birth weight (LBW)
Old
Young Smoking
LBW
RR= 2
Smoking
LBW
RR= 4
Effect Modification (Interaction) Example 2: A cohort study was conducted to evaluate the association between air pollution and a specific lung disease. From the information below would you conclude that gender is a confounder or an effect modifier??? D
No D Total
E
200
1800
2000
No E
400
3600
4000
Total 600
5400
6000
Biases: Review
Biases: Review
T (True) or F (False):
T (True) or F (False):
To ensure that the study is internally valid we need to check for the 4 main biases: Selection bias, information bias, confounding and effect modification.
Biases: Review
External validity depends on ______________
3500 individuals were females 500 females were exposed 490 females had the disease 110 females had disease & were exposed 1410 males were exposed & did not have the disease 20 males were not exposed & had the disease
Generalizability is the extent to which the inference derived from the study sample is correct for the target population.
Biases: Review
If you increase the sample size you increase _______________
Biases: Review
Biases: Review
T (True) or F (False):
T (True) or F (False):
Blinding the interviewer minimizes observation bias.
Biases: Review A cohort study is planned to investigate the association between maternal alcohol consumption during pregnancy and fetal alcohol Syndrome (a disease that is difficult to diagnose) in newborn children. What are the possible biases?
Biases: Review What are the characteristics of a confounder?
Recall bias is common in case control studies while loss to follow up is common in randomized clinical trials.
Biases: Review In a closed cohort study, most of those who were exposed and developed the disease died before the end of the study. What will happen to the RR?
Biases: Review T (True) or F (False): The crude relative risk for smoking and myocardial infarction was 2.8. When stratified by gender, the relative risk for males was 5.6 and 1.5 for females. The study investigators concluded that results are biased by gender which is an effect modifier and hence they adjusted for gender in the analysis.