Post on 12-Nov-2014
EPID 600; Class 9 Confounding
University of Michigan School of Public Health
1
Is drinking coffee causally associated with doing well in EPID 600?
Coffee Course grade
Coffee drinkers Mean course grade
None 72
A little 77
Some 84
A lot 92
2
Coffee Course grade
Studying
Non-causal association
Causal association
3
Summary…
Those who study drink more coffee Studying is associated with a higher course grade in EPID600 So, the apparent association between coffee drinking and doing better on this course is explained by the different studying habits of those who drink coffee and those who do not
4
Sick
Healthy
Observed
Counterfactual (parallel universe)
Counterfactual thinking
5
Sick
Sick
Observed
Counterfactual (parallel universe)
Counterfactual thinking
6
Confounding
Confounding centrally is a violation of exchangeability, that is, when the presence of variables make the counterfactual scenarios non-comparable When confounding occurs the apparent effect of exposure on outcome is distorted because the effect of an extraneous factor is mistaken for an actual exposure effect The concept of confounding is central to epidemiology If we want to make inferences regarding causation we need to take into account the possibility of confounding Confounding not an “all or none” phenomenon
7
Violation of exchangeability, i.e., confounding
Truth Study
Exposed Counterfactual exposed if they were not exposed
Exposed Not exposed
Disease
8
Confounding
When a non-causal association between a given exposure and an outcome is observed as a result of the influence of a third variable, that third variable is usually designated a confounding variable or confounder
9
Is urbanicity a risk factor for developing coronary heart disease?
Incidence of CHD
Urban 9 per 1000
Non-urban 7 per 1000
10
Age and cities
Urban Rural
Old 20% 5%
Middle-age 40% 45%
Young 40% 50%
11
Age and CHD incidence
CHD incidence per 1,000
Old 15
Middle-age 10
Young 5
12
Urban CHD incidence
Age distribution
Non-causal association
Causal association
13
Summary...
In urban areas you have more older people Older people have a higher incidence of coronary heart disease than younger people So, the apparent association between living in urban areas and CHD is explained by the different age distribution between urban and non-urban areas
14
Is drinking alcohol a risk factor for lung cancer?
Lung cancer
Drinkers 90/100,000
Non-drinkers 80/100,000
15
Alcohol Lung cancer
Smoking
Non-causal association
Causal association
16
General rule
A variable is a confounder if
1. It is a risk factor for the outcome
2. It is associated with the exposure
17
Schematically...
EXPOSURE OUTCOME
OBSERVED ASSOCIATION
18
Schematically...
EXPOSURE OUTCOME
“THIRD VARIABLE”
19
Example: male gender as a risk factor for malaria
Cases Controls Total
Male 88 68 156
Female 62 82 144
Total 150 150 300
Adapted from Szklo & Nieto, 1999 20
Example continued...
88 / 62
68 / 82 OR = =
88*82
62*68 = 1.71
21
Is a confounder responsible for the observed association?
MALE GENDER MALARIA ?
22
Is a confounder responsible for the observed association?
MALE GENDER MALARIA
OUTDOOR OCCUPATION
?
23
Two questions to be asked...
Is working outdoors a risk factor for malaria? Is male gender associated with working outdoors?
24
Is working outdoors a risk factor for malaria?
42% of the cases work mostly outdoors
12% of the controls work mostly outdoors
Cases Controls
63 18 87 132
Mostly Outdoor
Mostly Indoor 150 150
OR = 18 / 132
63 / 87 =
87*18
63 * 132 = 5.3
25
Is working outdoors associated with being male?
Males
68
88
Females
13
131
Mostly Outdoor
Mostly Indoor
Total
81
219
% Male
84%
40%
Those who work outdoors are more likely to be male than those who work indoors.
26
Back to the questions
Is working outdoors a risk factor for malaria?
Is male gender associated with working outdoors?
27
Back to the questions
Is working outdoors a risk factor for malaria? YES Is male gender associated with working outdoors? YES
28
Male Malaria
Outdoor occupation
Non-causal association
Causal association
29
So...
Working outdoors may be a confounder of the observed relation between male gender and malaria...
But is it?
30
One question to be asked...
Is there still a relationship between male gender and malaria when we account for potential confounders?
31
Stratification
1. Hold occupation (the potential confounder) constant 2. Examine the relationship of interest within strata of
occupation (i.e., indoors vs. outdoors)
32
Adapted from Szklo & Nieto, 1999
Cases
Males 53 15
Females 10 3
Controls
Total 63 18
Mostly Outdoor Occupation
OR = 10*15
53*3 = 1.06
Cases
Males 35 53
Females 52 79
Controls
Total 87 132
Mostly Indoor Occupation
OR = 52*53
35*79 = 1.00
Stratified analyses of the association between gender and malaria according to whether individuals work mainly outdoors or indoors
33
The principle...
IF working outdoors did NOT explain the relationship between male gender and malaria, then men SHOULD have higher risk of malaria whether they worked outdoors or not
34
Summary
Crude OR: 1.7
Outdoor work adjusted odds ratio: 1.0
35
Back to the question...
Is there still a relationship between male gender and malaria when we account for potential confounders?
36
Back to the question...
Is there still a relationship between male gender and malaria when we account for potential confounders?
NO
37
So...
Working outdoors is a confounder of the observed relation between male gender and the risk of malaria.
38
Summary: is a covariate a confounder?
QUESTION 1: Is it associated with exposure? QUESTION 2: Is it causally associated with outcome?
YES
STEP 1: Calculate Crude Measure of Association STEP 2: Calculate Measure of Association within strata of potential confounder
39
Assess measure of association within strata
Are stratum specific measures same?
No Yes
Yes No
Confounding report measure
adjusted for confounding
Apparently unconfounded Report crude measure
Next lecture Crude measure=stratum specific?
Caveat 1: Remember that confounding is not an “all-or-none” phenomenon; therefore association does not have to “go away”, simply “change”
Caveat 2: All this is assuming no bias
40
Are taller people happier?
It is believed that mental illness is shaped by factors throughout the life course Some hypothesize that childhood illness, socioeconomic adversity, and diet may play a role in the development of mental illness as an adult In the absence of direct measures for these childhood , variables, investigators have looked at birthweight, height, and BMI as proxies. In Norway a prospective cohort of 74,332 men and women was used to investigate the association of height, and body mass index, with suicide, anxiety, and depression.
Bjerkeset et al. Association of adult body mass index and height with anxiety, depression, and suicide in the general population. Am J Epidemiol. 2007; 167(2):193-202
41
1984– 1986: (HUNT 1) participants aged 20 years or more
How many suicides occur? What is the relative height of those who commit suicide?
Height and suicide
Participants in the Nord-Trøndelag
Health Study
Participants followed up until December 31,
2002.
Bjerkeset et al. Association of adult body mass index and height with anxiety, depression, and suicide in the general population. Am J Epidemiol. 2007; 167(2):193-202
42
Height and suicide: total cohort
Bjerkeset et al. Association of adult body mass index and height with anxiety, depression, and suicide in the general population. Am J Epidemiol. 2007; 167(2):193-202
43
Height and suicide
Adjusted only for age and sex, it appears that increasing height is in fact related to a lower risk of suicide
Heightquar,le Hazardra,o(adjustedforage
andsex)Lowestquar,le 1.00Quar,le2 1.01Quar,le3 0.65Highestquar,le 0.69HeightperSD(sexspecific)
0.83
Bjerkeset et al. Association of adult body mass index and height with anxiety, depression, and suicide in the general population. Am J Epidemiol. 2007; 167(2):193-202
44
Height and suicide: subsample with confounder information
Bjerkeset et al. Association of adult body mass index and height with anxiety, depression, and suicide in the general population. Am J Epidemiol. 2007; 167(2):193-202
45
Height and suicide: un-confounded
After adjusting for education, marital status, smoking history, frequency of alcohol use, physical activity, vigor, nervousness, calmness, cheerfulness, and frequency of using tranquilizers, the association is highly attenuated
Height quartile Hazard ratio (adjusted for age and sex)
Lowest quartile 1.00 Quartile 2 1.17 Quartile 3 0.84 Highest quartile 0.97 Height per SD (sex specific)
0.97
Bjerkeset et al. Association of adult body mass index and height with anxiety, depression, and suicide in the general population. Am J Epidemiol. 2007; 167(2):193-202
46
Handling confounding
Analysis Stratification Adjustment Restriction
Design Matching Randomization (experimental) studies
47
Stratification: another example
Look at the association within strata of the confounder (i.e., holding the confounding constant)
Is alcohol consumption related to lung cancer in SMOKERS?
Is alcohol consumption related to lung cancer in NON-SMOKERS?
Alcohol Lung cancer
Smoking
?
48
Adjustment
Use a statistical technique to estimate what the association would be IF the confounder was not associated with the exposure
49
Standardization, particularly age standardization, has historically been very important for this purpose
Urban Non-urban
Crude incidence
9 per 1000 7 per 1000
? ?
How would the incidence rates of CHD between urban and non-urban areas compare if both areas had the same age-distribution?
50
Standardization
Takes into account different underlying population structures, i.e., accounts for possible confounding due to different other variables Typically applied to age Essentially calculates weights against another population to make different rates comparable Allows direct comparison of mortality data in different populations, controlling for age (but not other potential confounders, of course)
51
Matching in case control studies
For each case of malaria that works outdoors, choose a control that works outdoors For each case of malaria that works indoors, choose a control that works indoors Make cases and controls similar in terms of where they work and take potential confounder into account in analysis 1 of the 2 conditions for confounding is NOT met
Outdoor Work
Malaria Gender ?
52
Aside...advantages and disadvantages of matching
Advantages Some control over confounding in a ‘relatively straightforward’ way If we match on strong confounders we may increase study power
Disadvantages Difficult to find controls if matching on too many variables Variables used in matching cannot be assessed as either independent or dependent variables If we match on variables related to exposure, we are overmatching Need to use special techniques to analyze matched data
53
Randomization in experimental studies
“Assign” people at random to alcohol intake vs. no alcohol intake On average proportion of smokers will be similar in both groups 1 of the 2 conditions for confounding is NOT met
Alcohol Lung Cancer
Smoking
?
54
A detail
A variable is a confounder if: 1. It is a risk factor of the outcome 2. It is associated with the exposure
55
A detail...
Diet Heart disease
Cholesterol
?
Cholesterol is a RESULT of the exposure Cholesterol is then NOT a confounder However (and here’s where it gets tricky), cholesterol can be both a mediator and a confounder
56
Concluding confounding
Confounding is a violation of exchangeability; the extent to which we introduce error (bias) in our study relates to how well we deal with the confounding
57