1. Epidemiology II University of Cologne Germany P. Morfeld
Institute for Occupational Epidemiology and Risk Assessment (IERA)

2. Epidemiology II: e-learning Epidemiology II OBJECTIVES EPI 1 :
You should be able to explain...
Concepts/Definitions
conception of epidemiology
notion of stochastics
Descriptive Approach
study group (source/target population)
responses
covariables (exposure)
statistic
Analytical Approach
regression techniques
Bias (Ecological Fallacy, Confounding)
notion of biased comparisons
solutions with multiple regression
Epidemiology II:
e-learning
P. Morfeld, IERA, / University of Cologne

3. Epidemiology II In Epidemiology I, we followed the famous analysis
of John Snow and identified the Broad Street pump
as the cause of the cholera mortality in Soho,
September 1854.
See the next slide to convince yourself again.
P. Morfeld, IERA, / University of Cologne

4. Epidemiology II The Broad Street pump is obviously the „cause“ – but
Snow, Sept, 1854 (John Snow, 1813 – 1858)
The Broad Street pump is obviously
the „cause“ – but
not the Brewer Street pump !
C.F. Cheffins Map of Soho (London),
modified by Snow (1850 –
1854)
BROAD STREET pump
• death from cholera
 public water pump X
BREWER STREET pump
P. Morfeld, IERA, / University of Cologne

5. Epidemiology II Another obvious finding: The Brewer Street pump
is not related to cholera mortality.
Can we derive this finding with regression analysis
as applied to the Broad street pump data in
Epidemiology I?
Let us do a parallel analysis of both pumps.
P. Morfeld, IERA, / University of Cologne

6. Epidemiology II Cholera Epidemic in London, Sept. 1854 50 100 150 20050
100
150
200
250
300
350
400 20 40 60 80
120
140
160
Cholera Epidemic in London, Sept. 1854
Data acc. to J. Snow, 1855
number of deaths from cholera
per grid square
distance from pump in yards
BREWER STREET
P. Morfeld, IERA, / University of Cologne

7. Epidemiology II Cholera Epidemic in London, Sept. 1854 50 100 150 20050
100
150
200
250
300
350
400 20 40 60 80
120
140
160
distance from pump in yards
number of deaths from cholera
per grid square
Cholera Epidemic in London, Sept. 1854
Data acc. to J. Snow, 1855
BROAD STREET
BREWER STREET
P. Morfeld, IERA, / University of Cologne

8. Epidemiology II
x: distance to pump „BREWER Street“ y: number of deaths from cholera : residual
model: y = bO + b1x + 
estimation of parameters:
p = 0.11
decrease,
but
not
significant:
<
<
P. Morfeld, IERA, / University of Cologne

9. Epidemiology II Cholera Epidemic in London, Sept. 1854 50 100 150 20050
100
150
200
250
300
350
400
-20 20 40 60 80
120
140
160
Cholera Epidemic in London, Sept. 1854
Data acc. to J. Snow, 1855
number of deaths from cholera
per grid square
BREWER STREET
distance from pump in yards
P. Morfeld, IERA, / University of Cologne

10. Epidemiology II Cholera Epidemic in London, Sept. 1854 50 100 150 20050
100
150
200
250
300
350
400
-50
BROAD STREET
number of deaths from cholera
per grid square
Cholera Epidemic in London, Sept. 1854
Data acc. to J. Snow, 1855
distance from pump in yards
BREWER STREET
P. Morfeld, IERA, / University of Cologne

11. Epidemiology II Although the effect estimate is more pronounced
for the Broad Street pump we observe a
downward trend with the Brewer Street pump also.
Thus, this regression analysis indicates
that the Brewer street pump may also be related to
cholera mortality.
P. Morfeld, IERA, / University of Cologne

12. Epidemiology II But we know that the Brewer street pump is not
related to death from cholera (see slide 4).
This is confusing! Why is the regression analysis
misleading? Why does the regression analysis
not return the conclusion that the Brewer
Street pump is unrelated to cholera mortality?
P. Morfeld, IERA, / University of Cologne

13. Epidemiology II Let us discuss this important phenomenon
called “confounding” in a different setting:
Coffee consumption and heart attacks
P. Morfeld, IERA, / University of Cologne

14. Epidemiology II CONFOUNDING superficial coffee risk of
consumption heart attack,
risk of
lung cancer
Epidemiological studies showed a positive association of coffee consumption and risk of heart attack, as expected by some of the discussants based on mechanistic considerations - but the studies also showed an association with lung cancer risk. The lung cancer result was counterintuitive and shed doubt on the heart attack finding!
P. Morfeld, IERA, / University of Cologne

15. Epidemiology II CONFOUNDING superficial coffee risk of
consumption heart attack,
risk of
lung cancer
more accurate
coffee
consumption
tobacco
risk of
heart attack
risk of lung cancer
A more detailed analysis revealed a complicated role of tobacco consumption.
P. Morfeld, IERA, / University of Cologne

16. Epidemiology II It was shown that
coffee consumption was associated with tobacco consumption (“smokers drink more coffee than non-smokers”)
smoking was linked with lung cancer (and cardiac diseases)
The association between coffee consumption and heart attacks was illusory and due to uncontrolled smoking effects (“confounding”)
P. Morfeld, IERA, / University of Cologne

17. Epidemiology II CONFOUNDING a covariable is a CONFOUNDER only if
association covariable – response and
association covariable – exposure
 whether a covariable is a COUNFOUNDER or not depends on the concrete study
e.g., if 1) is valid but 2) is not valid:
despite 1) the covariable is no confounder!
P. Morfeld, IERA, / University of Cologne

18. Epidemiology II A pronounced form of confounding is
known as “Simpson’s paradox”
On average mortality rates are lower in Alaska
than in Florida but in each age group mortality
rates are higher in Alaska than in Florida
(see next slide)
P. Morfeld, IERA, / University of Cologne

19. Epidemiology II
SIMPSON´S PARADOX: change in direction of effect due to confounding
[first described by K. Pearson 1899, published by E.H. Simpson 1951]
Number of inhabitants, number of deaths and mortality rate (per 10,000 inhabitants and year)
by age and US state, 1974.
P. Morfeld, IERA, / University of Cologne

20. Epidemiology II Background of “Simpson’s paradox”
People are older in Florida than in Alaska:
Thus, age is associated with “exposure”
(exposure = Alaska vs. Florida)
And age is a strong risk factor for death.
It follows that age is a potential confounder!
P. Morfeld, IERA, / University of Cologne

21. Epidemiology II Geometrical explanation of confounding
(see next slide):
If tobacco consumption is associated with coffee
consumption we will observe an inclining trend of
lung cancer risk across coffee consumption (upper
lines). This phenomenon will not occur if coffee
and tobacco consumption are unrelated (lower
lines).
P. Morfeld, IERA, / University of Cologne

22. Epidemiology II lung cancer risk tabacco consumption
coffee consumption
tabacco consumption
P. Morfeld, IERA, / University of Cologne

23. Epidemiology II The geometrical explanation provides a
solution how to control for confounding: fit the full
plane and not single straight lines.
The problem is three-dimensional: single lines
between two variables deal with the data as if the
problem were of only two dimensions. The plane
relies on all data simultaneously and shows both
effects: an increase in lung cancer risk with
smoking but no effect across coffee consumption.
P. Morfeld, IERA, / University of Cologne

24. Epidemiology II How to do this? What is a plane algebraically?
The answer is simple: Just add a further
x-variable to the usual straight line equation.
And all standard statistical programs will do the
analysis for you.
P. Morfeld, IERA, / University of Cologne

25.   Epidemiology II < <
x1: distance to pump „BROAD Street“ x2: distance to pump „BREWER Street“
y: number of deaths from cholera : residual
model: y = b0 + b1x1 + b2x2 + 
estimation of parameters:
„BROAD Street“
„BREWER Street“ 
< 
<
P. Morfeld, IERA, / University of Cologne

26. Epidemiology II The Broad Street pump is linked to cholera
mortality in this extended analysis. The result is
very similar to the finding we got from the simple
regression analysis in Epidemiology I.
But the Brewer Street pump is not related to
cholera mortality in this extended analysis, very
different from the findings in the simple regression
analysis.
P. Morfeld, IERA, / University of Cologne

27. Epidemiology II This extended approach is called
- multiple regression or
- multivariable regression
and can tackle the problem of confounding.
P. Morfeld, IERA, / University of Cologne

28. Epidemiology II To produce confounding in the simple Brewer
Street analysis we need a confounder,
i.e., a variable that
is linked with the Brewer Street data and
is a risk factor for cholera mortality
P. Morfeld, IERA, / University of Cologne

29. Epidemiology II Cholera Epidemic in London, Sept. 1854 Note:50
100
150
200
250
300
350
400
p= 0.10
distance to pump "BREWER Street" in Yards
Cholera Epidemic in London, Sept. 1854
Data acc. to J. Snow, 1855
distance to pump "BROAD Street" in Yards
Note:
even a non- significant association can produce a strong confounding!
P. Morfeld, IERA, / University of Cologne

30. Epidemiology II The Broad Street pump is a risk factor and
distances from the Broad Street pump are
correlated with distances from the Brewer Street
pump.
Thus, the Broad Street pump is a confounder for
the association of the Brewer Street pump and
death from cholera.
P. Morfeld, IERA, / University of Cologne

31. Epidemiology II CONFOUNDING
another example: Armstrong et al 1994, Armstrong and Theriault 1996
Investigation of the effect of benzo(a)pyrene (BaP) on lung cancer risk in aluminium production plant workers in Canada
BaP is an important polynuclear aromatic hydrocarbon that is produced and released during the production of aluminium (Soderberg process)
BaP can produce lung cancer in experimental animals and has been shown to cause lung cancer in humans
Aim of the study: estimate the effect of BaP on lung cancer quantitatively taking a possible confounding by smoking habits into account
P. Morfeld, IERA, / University of Cologne

32. Epidemiology II smoking: STRONG risk factor!
P. Morfeld, IERA, / University of Cologne

33. Epidemiology II smoking: NO confounder!
P. Morfeld, IERA, / University of Cologne

34. Epidemiology II CONFOUNDING a covariable is a CONFOUNDER only if
association covariable – response and
association covariable – exposure
 whether a covariable is a COUNFOUNDER or not depends on the concrete study
e.g., if 1) is valid but 2) is not valid:
despite 1) the covariable is no confounder!
Examples: - Broad Street pump confounds the effect of the Brewer Street pump - Smoking does not confound the effect of BaP
P. Morfeld, IERA, / University of Cologne

35. University of Cologne
P. Morfeld, IERA, / University of Cologne