Attributable fractions

(AS11)
EPM304 Advanced Statistical Methods in Epidemiology

Course: PG Diploma/ MSc Epidemiology

This document contains a copy of the study material located within the computer
assisted learning (CAL) session.
If you have any questions regarding this document or your course, please contact
DLsupport via DLsupport@lshtm.ac.uk.
Important note: this document does not replace the CAL material found on your
module CDROM. When studying this session, please ensure you work through the
CDROM material first. This document can then be used for revision purposes to
refer back to specific sessions.
These study materials have been prepared by the London School of Hygiene & Tropical Medicine as part of
the PG Diploma/MSc Epidemiology distance learning course. This material is not licensed either for resale
or further copying.
London School of Hygiene & Tropical Medicine September 2013 v1.0

Section 1: Attributable fractions

Aim

To learn how to estimate the fraction of disease occurrence attributable to an

exposure in a population.

Objectives
By the end of this session you will be able to:

Estimate and interpret population attributable fractions (PAF)

Obtain a PAF from different types of epidemiological studies
Obtain a PAF adjusted for confounding factors
Evaluate the PAF for the joint effect of multiple exposures

This session should take you between 1 and 2 hours to complete.

Section 2: Planning your study

In this session you will learn how to obtain estimates of the impact of an exposure in
a population (i.e. population attributable fractions), for different studies, adjusted for
confounding factors, and for multiple risk factors, in simple situations.
To begin, you will review the calculations of attributable risks and fractions. For a
more detailed review, refer to 'Measures of impact' in study unit EP101.
You will also use logistic regression for estimating attributable fractions when there
are multiple risk factors, so you may wish to review the relevant sessions in
Statistical methods in Epidemiology, EP202.
Measures of impact
Logistic regression

FE06
SM07, SM08, SM09

Interaction: Hyperlink: FE06:

New session opens from FE06
Interaction: Hyperlink: SM07:
New session opens from SM07
Interaction: Hyperlink: SM08:
New session opens from SM08
Interaction: Hyperlink: SM09:

New session opens from SM09

2.1: Planning your study

To illustrate the calculation of attributable measures of effects and impact you will
consider a study from Mwanza, Tanzania.

Section 3: Population attributable fractions

This session considers how to measure what risk is attributable to an exposure.
If a causal relationship between an exposure and an outcome can be assumed we
can calculate attributable risk for:
exposed individuals
or
the total population
To begin with, we will consider the simplest situation and examine the impact of a
single binary exposure on the incidence of a disease in a defined study population,
then look at more complex situations.
Attributable risk
Attributable risk in an exposed group (ARE) is a measure of the effect of an

exposure in an exposed group.

Population attributable risk

Population attributable risk (PAR) is a measure of the impact of an exposure on a

total population.

3.1: Population attributable fractions

If we let
p = Proportion exposed in the study
population
1 = Incidence rate in the exposed
0 = Incidence rate in the unexposed
T = Incidence rate in the total population
What is the rate ratio, , for the effect of exposure?
Interaction: Button: clouds picture (pop up box appears):
The rate ratio is the ratio of the incidence rate in the exposed and the incidence rate
in the unexposed, = 1/ 0.

On the following pages we will review the attributable measures of exposure effect
and impact on a population.

3.2: Population attributable fractions

Attributable risk
The attributable risk in an exposed group measures the rate of disease attributable
to exposure in that group.
ARE = 1 - 0

Note: The attributable risk is independent of the prevalence of the exposure in the
population.

3.3: Population attributable fractions

The attributable risk is generally used when there is a positive causal association
between the exposure and the outcome. However, if the exposure prevents the
outcome, the attributable risk will be negative. In this situation you can interchange
the exposed and unexposed in the attributable risk formula. The resulting measure
gives the protective effect of exposure.

3.4: Population attributable fractions

Attributable fraction
The attributable fraction in the exposed group is the proportion of the risk in the
exposed group that is due to exposure.

1 - 0
AFE =

This can also be written

-1
AFE =

where = the rate ratio, 1/ 0.

3.5: Population attributable fractions

Again, if the rate in the unexposed is greater than that in the exposed then instead
of an attributable fraction we have a preventable fraction. The preventable fraction
is the protective effect as a proportion of the unexposed risk.

3.6: Population attributable fractions

Population attributable risk
The population attributable risk estimates the disease rate attributable to the
exposure in the whole population.
It is a measure of the impact of the exposure in the population and is dependent on
the prevalence of the exposure in the population.
PAR = T 0
T = incidence rate in the population
0 = incidence rate in the unexposed
Interaction: Button: More (card appears on right handside):
If you don't have information on the total risk in the population, this can be
estimated using the risk in the exposed and unexposed and the prevalence of
exposure.
T = p 1 + (1 - p) 0
where p = prevalence of exposure.

3.7: Population attributable fractions

The population attributable risk is given by:

PAR = T 0
Substituting for T:
PAR = p1 + (1 p) 0 0
PAR = p(1 0)

3.8: Population attributable fractions

The population attributable fraction estimates the proportion of cases of disease
in the entire population that is due to exposure.
PAF = PAR = (T 0) / T
T
again using the substitution
T = p1 + (1 p) 0, this can be rewritten as:

PAF =

p( - 1)
[p( - 1) + 1)]

where = the rate ratio.

Interaction: Button: More (card appears on right handside):
The PAF increases with the rate ratio , but also with the prevalence of exposure p.
It will therefore vary between populations, depending on how common the exposure
is.

Interaction: Button: Show graph (card appears):

The PAF is a valuable public health concept, since it measures the contribution of a
single exposure to the overall rate of disease in a population.

3.9: Population attributable fractions

There is an alternative formulation for the PAF that is more useful when taking
account of confounders and joint effects.
Interaction: Button: More (text appears on bottom LHS):
Suppose that d cases of the disease occurred in the population, of which a proportion
p' came from the exposed group.
Subset of cases = d p'
Removing the exposure could only have an effect on this subset of cases. What
proportion of these cases would have been avoided if the exposure were removed?
Interaction: Button: clouds picture (pop up box appears and cad appears on the
right handside):
This proportion of cases that would be avoided by removing the exposure is the
attributable fraction in the exposed, given by:
AFE = ( 1) /

Therefore the total number of cases that would have been avoided in the subset
would be this fraction of the subset:
d p'AFE = d p' ( 1) /
So the fraction avoided in the total population is given by dividing this by the total
number of cases d:
PAF = p' ( 1) /
This formula is more easily generalised to more complex situations.

3.10: Population attributable fractions

Estimates of PAF need to be interpreted with care.
Now, suppose an observational study of risk factors for cardiovascular disease gives
a PAF of 20% for a high cholesterol level.
Could you conclude that the drug that successfully lowered such levels to normal
would reduce the rate of disease by 20%?
To answer this question, you need to consider the potential problems shown
opposite.
Interaction: Tabs: 1:
The association may not be causal. For example, a high cholesterol level may simply
be a secondary manifestation of the disease process leading to cardiovascular
disease.
Interaction: Tabs: 2:
There may be confounding factors. Subjects with high cholesterol are likely to differ
from other subjects with respect to other risk factors for cardiovascular disease, so
that their rate of disease would still be different even if their cholesterol level were
reduced to normal. As you will see, you can take account of confounding factors, but
only those that you are able to measure adequately.
Interaction: Tabs: 3:
Even if the drug is effective in reducing cholesterol levels to normal, this may not
immediately reduce the disease rate in these subjects to that of subjects whose
cholesterol levels have been constantly low. Similarly, if all smokers stopped
smoking today, the rate of lung cancer would not immediately be reduced to that of
lifelong non-smokers.
Interaction: Tabs: 4:
The computation of the PAF assumes that all other risk factors remain constant. But
the reduction in cholesterol achieved by the drug may be accompanied by changes

(decreases or increases) in other risk factors. Moreover, side-effects and other

adverse effects of the drug would also need to be considered.

Section 4: PAF for different epidemiological studies

On the following pages we will look at how to obtain the PAF for study types listed
opposite.
Cohort studies
Cross-sectional studies
Unmatched case-control studies
Matched case-control studies

4.1: PAF for different epidemiological studies

Cohort studies
This is the simplest situation, since disease rates can be measured directly.
If the cohort is a random sample of the entire study population you can obtain
estimates of:
p - the proportion of subjects exposed, or
p' - the proportion of cases exposed
The rate ratio, = 1 / 0, can be obtained directly.
However, if exposed and unexposed cohorts have been sampled separately for the
study (with differing sampling fractions), then can be estimated directly, but a
separate estimate of p or p' will be required
If p is known:

PAF =

p( - 1)
[p( - 1) + 1)]

If p' is known:

PAF =

p'( - 1)

4.2: PAF for different epidemiological studies

Cross-sectional studies

In a cross-sectional study, the prevalence of a disease state is measured, rather than

its incidence.
To estimate the population attributable fraction of prevalent cases you can use the
formula opposite.

This may be a useful indicator of the effect of exposure on the current burden
of disease, but is subject to the usual problems of interpretation of prevalent rathe
than incident cases.
If p is known:

PAF =

p( - 1)
[p( - 1) + 1)]

If p' is known:

PAF =

p'( - 1)

Note: Here, is the ratio of prevalences, rather than rates.

4.3: PAF for different epidemiological studies

Unmatched case-control studies
Consider a simple unmatched case-control study in which there are a representative
sample of cases, and a random sample of controls from the source population.
can be estimated by the odds ratio
p can be estimated as the proportion of
controls exposed
p' can be estimated as the proportion of
cases exposed

a
c

Unexpos
ed
b
d

m1

m0

Exposed
Cases
Control
s
Total

PAF =

p( - 1)
[p( - 1) + 1)]

where = ad/bc and p = c/n0, or:

Total
n1
n0
N

PAF =

p'( - 1)

where p' = a/n1

4.4: PAF for different epidemiological studies

These formulae can be rearranged to give the more simple formula:
PAF = 1 (b / n1) / (d / n0)
Note: The second half of the formula is just the ratio of the proportion of cases
unexposed to the proportion of controls unexposed.

4.5: PAF for different epidemiological studies

An unmatched case-control study looked at risk factors for HIV among women in
Mwanza, Tanzania.
The table opposite shows the proportion of cases and controls who reported recent
travel to a large town outside the region. A larger proportion of cases (14%)
reported recent travel compared to the controls (5%).
Can you calculate the odds ratio from the table? Give your answer to 2 decimal
places.
Interaction: Calculation: Can you calculate the odds ratio from the table? Give your
answer to 2 decimal places.
Correct Response 3.43:
Correct
Yes, the odds ratio is given by:
= ad / bc = (26 x 548) / (160 x 26) = 3.43.
The corresponding confidence intervals and
P-value are:
95% CI: 1.95 to 6.03; P < 0.0001.
Incorrect Response:
Sorry, that's not right. The odds ratio is calculated as follows:
= ad / bc = (26 x 548) / (160 x 26) = 3.43.
The corresponding confidence intervals and

P-value are:
95% CI: 1.95 to 6.03; P < 0.0001.

Cases
Controls

Travel to large town

Yes
No
n
%
n
%
26
14
160
86
26
5
548
95

Total
186
574

4.6: PAF for different epidemiological studies

The PAF can be calculated directly from the frequencies in the table using:
PAF = 1 (b / n1) / (d / n0)
Calculate the PAF for recent travel, giving your answer as a percentage to one
decimal place.
PAF =

Interaction: Calculation: PAF = _____%

Correct Response 9.9:
Correct
That's right, the PAF is given by:
PAF = 1 - (160 / 186) / (548 / 574) = 1 - 0.901 = 9.9%
Incorrect Response:
Sorry, that's not right. The PAF is calculated as follows:
PAF = 1 (b / n1) / (d / n0)
= 1 (160 / 186) / (548 / 574)
= 1 0.901
= 9.9%

4.7: PAF for different epidemiological studies

How would you interpret the estimate of PAF of 9.9% in this study with respect to
recent travel to a large city?

Remember PAFs assume that

exposure is causal.

Select an option opposite that reflects your interpretation.

9.9% of the cases of HIV in this

study can be attributed to recent
travel to a large city

10% of women who recently

travelled to a large city develop
HIV

No cases of HIV in this study can

be attributed to recent travel

Interaction: Hotspot: 9.9% of the cases of HIV in this study can be attributed
to recent travel to a large city:
Correct Response:
Correct
Yes, we are looking at what proportion of HIV cases were due to the exposure of
travelling to a large city. In this study, 9.9% of the cases of HIV can be attributed to
recent travel to a large city.
Interaction: Hotspot: 10% of women who recently travelled to a large city
develop HIV:
Incorrect Response:
No, it is the total proportion of cases of HIV and exposure to travelling to a large city
that we are interested in, not just the women who were exposed. The PAF tells us
that 9.9% of HIV cases in this study were attributed to recent travel.
Interaction: Hotspot: No cases of HIV in this study can be attributed to recent
travel
Incorrect Response:
No, in this example we are looking at the proportion of cases attributable to recent
travel, assuming exposure is causal. The PAF tells us that 9.9% of HIV cases in this
study were attributed to recent travel.

4.8: PAF for different epidemiological studies

Confidence intervals for PAF
It is possible to obtain an approximate confidence interval for the PAF in this simple
case. It's easiest to work with
(1 PAF) = (b / n1) / (d / n0) to do this.
To obtain 95% confidence intervals, you can multiply and divide (1 PAF) by the
error factor below, and subtract each limit from 1 to obtain a confidence interval for
the PAF.
Error factor =
exp{1.96 x [a / (bn1) + c / (dn0)]}
95% confidence interval for the PAF =
1 [(1 PAF)

Error Factor]

4.9: PAF for different epidemiological studies

Using the error factor formula opposite, can you calculate a 95% confidence interval
for the PAF of 9.9% that you obtained for recent travel?
First calculate the error factor. Give your answer to 3 decimal places.
Error factor =
Interaction: Calculation: Error factor =_____:
Correct Response 1.062:
Correct
That's right, using the formula opposite gives:
Error factor = exp{1.96 x [26 / (160 x 186) + 26 / (548 x 574)]} = 1.062
Incorrect Response:
Sorry, that's not right. Substituting values from the table into the equation for error
factor should give you:
Error factor
= exp{1.96 x [26 / (160 x 186) + 26 / (548 x 574)]}
= exp(0.0605)
= 1.062

Travel to large town

Yes
No
%
n

Total

Cases
Controls

26
26

14
5

160
548

86
95

186
574

PAF = 9.9%
95% Cl = 1 [(1- PAF)

Error factor]

Error factor =
exp {1.96 x [a / (bn1) + c / (dn0)]}

4.10: PAF for different epidemiological studies

Finally, you multiply and divide (1 PAF) by the error factor to obtain the lower and
upper limits for (1 PAF), and subtract these from 1 to obtain the confidence
interval for PAF.
This has been done opposite for the lower limit of the PAF confidence interval. Click
the box to see how the value was calculated. Then calculate the value for the upper
limit, and enter it as a percentage, to one decimal place.

Lower limit for PAF =

4.3%

Upper limit for PAF =

_____%

Interaction: Button:
Output:

4.3%

Start with (1 PAF) = 0.901. Then multiply by the error factor to get the upper limit for (1 PAF):
Upper limit for (1 PAF) = 0.901 x 1.062 = 0.957
Then transfer back to PAF:
Lower limit for PAF = 1 0.957 = 0.043 = 4.3%
End interaction
Interaction: Calculation: Upper limit for PAF = ____:
Correct Response 15.2%:
That's right, the lower limit for (1 PAF) is 0.901 / 1.062 = 0.848, which means the
upper limit for PAF is given by:

Upper limit = 1 0.848 = 0.152 = 15.2%

Incorrect Response:
Sorry, that's not right. You should have started with (1 PAF) = 0.901. Then divide
by the error factor to get the lower limit for (1 PAF):
Lower limit for (1 PAF) = 0.901 / 1.062 = 0.848
Then transfer back to PAF:
Upper limit for PAF = 1 0.848 = 0.152 = 15.2%

4.11: PAF for different epidemiological studies

Matched case-control studies
In a matched case-control study we do not have the proportion of exposure in the
population and cannot estimate it from the control group. The controls are
individually matched to cases and are therefore not representative of the population.
Therefore you can only calculate the PAF based on the proportion of cases exposed
and the matched odds ratio.
PAF = p' ( 1) /
p' = proportion of cases exposed
= matched odds ratio

Section 5: Exposures with multiple levels

So far we have considered the simplest situation of a single binary exposure. Now
let's examine a single exposure with several levels.
For such an exposure we can estimate the proportion of cases attributable to each
level of exposure. That is, the proportion of cases that would be avoided if the rate of
disease in each exposure group were reduced to that in the unexposed group.
The notation for exposure level, number of cases, and proportion exposed is shown
below.
Exposu
re level
0
1
...
k
...
K

Numbe
r of
cases
d0
d1
...
dk
...
dK

Proport
ion of
cases
p0'
p1'
...
pk'
...
pK'

RR
1
1
...
k
...
K

5.1: Exposures with multiple levels

The attributable fraction among cases exposed at level k is

AF EK = ( k -1)/ k
and the population attributable fraction for exposure at level k (i.e. proportion of
cases in the population attributable to exposure at level k) is:

PAF K = p k '( k -1)/ k

These level-specific PAFs can be added to give the overall proportion of cases
attributable to the overall exposure:

PAF = PAFK
= pk '(k-1)/k
You will see how this works on the following pages.

5.2: Exposures with multiple levels

The table below shows the numbers of lifetime sex partners reported by the cases
and controls in the HIV case-control study. The attributable fraction can be
calculated for each level of exposure using:

AF EK = ( k -1)/ k
The estimated AFE in the group with 2 to 4 partners is 0.531. Click here to show the
calculation of this.
Interaction: Hyperlink: show the calculation (pop up box appears and the table on
the bottom centre changes as shown below):
AFE1 = (1 1)/1
= (2.13 1) / 2.13
= 0.531
Number of
partners
0-1 (k = 0)

Cases
27 (14.8%)

2-4 (k = 1)

92 (50.3%)

Controls
173
(31.3%)
277
(50.2%)

OR

AFEk
1

2.13

0.531

5-9 (k = 2)

40 (21.9%)

10+ (k = 3)
Total

24 (13.1%)
183

Number of
partners
0-1 (k = 0)

Cases
27 (14.8%)

2-4 (k = 1)

92 (50.3%)

5-9 (k = 2)

40 (21.9%)

10+ (k = 3)
Total

24 (13.1%)
183

83
(15.0%)
19 (3.4%)
552

Controls

3.09
8.09

OR

173
(31.3%)
277
(50.2%)
83
(15.0%)
19 (3.4%)
552

AFEk
1

2.13
3.09
8.09

The estimated AFE in the group with 10+ partners is 0.876. Click here to show the
calculationof this.
Now, can you calculate the AFE for the group with 5-9 partners? Give your answer to
3 decimal places.
Interaction: Hyperlink: show the calculation (pop up box appears and the bottom
centre table changes as shown below) :
AFE3

= (3 1)/3
= (8.09 1) / 8.09
= 0.876

Number of
partners
0-1 (k = 0)

Cases
27 (14.8%)

2-4 (k = 1)

92 (50.3%)

5-9 (k = 2)

40 (21.9%)

10+ (k = 3)
Total

24 (13.1%)
183

Controls
173
(31.3%)
277
(50.2%)
83
(15.0%)
19 (3.4%)
552

OR

AFEk
1

2.13

0.531

3.09
8.09

0.876

Interaction: Calculation: Now, can you calculate the AFE for the group with 5-9
partners? Give your answer to 3 decimal places.:
Correct Response 0.676 (pop up box appears and the bottom centre table changes
as shown below:
Correct
Yes, the attributable fraction for group 2 is given by:

AFE2

= (2 1)/2
= (3.09 1) / 3.09
= 0.676

Number of
partners
0-1 (k = 0)

Cases
27 (14.8%)

2-4 (k = 1)

92 (50.3%)

5-9 (k = 2)

40 (21.9%)

10+ (k = 3)
Total

24 (13.1%)
183

Controls

OR

173
(31.3%)
277
(50.2%)
83
(15.0%)
19 (3.4%)
552

AFEk
1

2.13

0.531

3.09

0.676

8.09

0.876

Incorrect Response:
Sorry, that's not right. To calculate the attributable fraction you should use the
following formula:
AFE2 = (2 1)/2
= (3.09 1) / 3.09
= 0.676

5.3: Exposures with multiple levels

Similarly, we can calculate the PAF for each level of exposure above the baseline
using:

PAF K = p k '( k -1)/ k

The estimated PAF in the group with 2-4 partners is 0.266. Click here to show the
calculation of this.
Interaction: Hyperlink: show the calculation (pop up box appears and table on
centre bottom changes as shown below):
PAFE1 = p1' (1 1) / 1
= 0.503 (2.13 1) / 2.13
= 0.266
Number of
partners
0-1 (k = 0)

Cases
27 (14.8%)

2-4 (k = 1)

92 (50.3%)

5-9 (k = 2)

40 (21.9%)

10+ (k = 3)

24 (13.1%)

Controls
173
(31.3%)
277
(50.2%)
83
(15.0%)
19 (3.4%)

OR

PAFk
1

2.13

0.266

3.09
8.09

Total

183

Number of
partners
0-1 (k = 0)

Cases
27 (14.8%)

2-4 (k = 1)

92 (50.3%)

5-9 (k = 2)

40 (21.9%)

10+ (k = 3)
Total

24 (13.1%)
183

552

Controls

OR

173
(31.3%)
277
(50.2%)
83
(15.0%)
19 (3.4%)
552

PAFk
1

2.13
3.09
8.09

The estimated PAF in the group with 5-9 partners is 0.148. Click here to show the
calculationof this.
Now calculate the PAF for the group with 10+ partners, giving your answer to 3
decimal places:
Interaction: Hyperlink: show the calculation (pop up box appears and the table on
bottom centre changes as shown below):
PAFE2 = p2' (2 1) / 2
= 0.219 (3.09 1) / 3.09
= 0.148
Number of
partners
0-1 (k = 0)

Cases
27 (14.8%)

2-4 (k = 1)

92 (50.3%)

5-9 (k = 2)

40 (21.9%)

10+ (k = 3)
Total

24 (13.1%)
183

Controls
173
(31.3%)
277
(50.2%)
83
(15.0%)
19 (3.4%)
552

OR

PAFk
1

2.13

0.266

3.09

0.148

8.09

Interaction: Calculation: Now calculate the PAF for the group with 10+ partners,
giving your answer to 3 decimal places::
Correct Response 0.115 (pop up box appears and bottom centre table changes as
shown below):
Correct
That's right, the PAF is given by:
PAFE3 = p3' (3 1) / 3
= 0.131 (8.09 1) / 8.09

= 0.115
Number of
partners
0-1 (k = 0)

Cases
27 (14.8%)

2-4 (k = 1)

92 (50.3%)

5-9 (k = 2)

40 (21.9%)

10+ (k = 3)
Total

24 (13.1%)
183

Controls

OR

173
(31.3%)
277
(50.2%)
83
(15.0%)
19 (3.4%)
552

PAFk
1

2.13

0.266

3.09

0.148

8.09

0.115

Incorrect Response:
Sorry, that's not right. To calculate the population attributable fraction you should
use the following formula:
PAFE3 = p3' (3 1) / 3
= 0.131 (8.09 1) / 8.09
= 0.115

5.4: Exposures with multiple levels

We can now calculate the overall fraction of cases in the population attributable to
any level of exposure by simply adding the PAFs across all levels.

PAF = PAFK
= pk '(k-1)/k
Click the 'add' button to do this.
Interaction: Button: Add (graph on centre bottom changes to the following):
Number of
partners
0-1 (k = 0)

Cases
27 (14.8%)

2-4 (k = 1)

92 (50.3%)

5-9 (k = 2)

40 (21.9%)

10+ (k = 3)
Total

24 (13.1%)
183

Controls
173
(31.3%)
277
(50.2%)
83
(15.0%)
19 (3.4%)
552

OR

PAFk
1

2.13

0.266

3.09

0.148

8.09

0.115
0.529

How can you interpret the overall PAF?

Interaction: Button: clouds picture (pop up box appears):

The overall PAF is approximately 53%. This represents the proportion of cases that
would have been avoided if the risk in all women were as low as in those reporting
no more than one lifetime partner.
Note: This analysis ignores the effects of all other factors, and is not adjusted for
confounding.

Section 6: Adjusting PAF for confounders

To illustrate the calculation of PAF in the presence of confounders, we will consider
the effect of marital status in the HIV case-control study.
The table below shows an unstratified analysis for divorced/widowed women
compared to those currently married or never married.
Marital status
Married/Single
Divorced/Widowe
d
Total

Cases
n
143
38
181

Controls
%
79
21

n
515
48

%
91
9

Odds
ratio
1.00
2.85

PAF

0.136

663

The crude OR is 2.85 with strong evidence for an association with HIV status (P <
0.001), and the unadjusted PAF is 13.6%.
Can you think of any potential confounders for the effect of marital status?
Interaction: Button: clouds picture (pop up box appears):
Marital status is closely related to age, which is also strongly associated with HIV
infection. Age therefore may be a confounding factor.
To adjust the PAF estimate for age we need to first stratify by age group.
Note: You can also adjust for age in a logistic regression model.

6.1: Adjusting PAF for confounders

The tabs opposite show the effect of marital status on HIV infection stratified into 3
age groups, coded 0 to 2.
The Mantel-Haenszel age-adjusted OR is 3.43 (P < 0.001), indicating that
confounding resulted in under-estimation of the effect of marital status (crude OR =
2.85).
A test for heterogeneity gave P = 0.8, indicating the data are compatible with no
interaction across strata.

Interaction: Tabs: Agegroup 0:

Age group 15 to 24 years
Marital status
Married/
Single
Divorced/Widowe
d
Total

Cases
63

Controls
193

Total
256

10

68

198

266

Cases
53

Controls
154

Total
207

14

14

28

67

168

235

Cases
27

Controls
168

Total
195

19

29

48

46

197

243

Odds ratio = 3.06

P = 0.07
Interaction: Tabs: Agegroup 1:
Age-group 25 to 34 years
Marital status
Married/
Single
Divorced/Widowe
d
Total
Odds ratio = 2.91
P = 0.007
Interaction: Tabs: Agegroup 2:
Age-group 35 to 54 years
Marital status
Married/
Single
Divorced/Widowe
d
Total
Odds ratio = 4.08
P = 0.0001

6.2: Adjusting PAF for confounders

Using the adjusted odds ratio we can calculate the PAF for the effect of marital status
adjusted for the confounding effect of age.
PAF = p' ( 1) /
where p' = proportion of cases exposed
= 38 / 181 = 0.210
= Mantel-Haenszel adjusted

odds ratio estimate

= 3.43
Marital status
Married/Single
Divorced/Widowe
d
Total

Cases
n
143
38

Controls
%
79
21

181

n
515
48

%
91
9

Adjusted
odds ratio
1.00
3.43

663

Can you calculate the PAF, giving your answer to 3 decimal places?
PAF =
Interaction: Calculation: PAF =____:
Correct Response:
Correct
Yes, the PAF is given by
PAF
= p' ( 1) /
= 0.210 (3.43 1) / 3.43
= 0.149
So, after adjusting for age, you can say that approximately 15% of HIV cases are
attributable to being divorced or widowed.
Incorrect Response:
No, that's not correct. The PAF is calculated as follows:
PAF
= p' ( 1) /
= 0.210 (3.43 1) / 3.43
= 0.149
So, after adjusting for age, you can say that approximately 15% of HIV cases are
attributable to being divorced or widowed.

6.3: Adjusting PAF for confounders

As you know, stratified analysis has limitations when you want to adjust for many
exposures. In general, regression models are used at this stage. In our example
logistic regression can be used.
p6c4rhs
To obtain the PAF for an exposure adjusted for other exposures, you use the
adjusted parameter estimate from the regression analysis in the formula below. You
also need to know the total proportion of cases exposed.
PAF = p' ( 1) /

p' = proportion of cases exposed

= adjusted odds ratio estimate

6.4: Adjusting PAF for confounders

Let's illustrate this with the Mwanza case-control study on risk factors for HIV. Say
you wanted to estimate the PAF for marital status adjusted for age and education.

Age is in 3 groups, modelled

as a linear term:
Group 0: 15 to 24 years,
Group 1: 25 to 34 years,
Group 2: 35 to 54 years.

Education is in 2 groups:
Group 0: no formal
education,
Group 1: some formal
education

Estimates from logistic regression

Odds
Standa z
ratio
rd
error
Msta1
3.6899
0.9440
5.103
35
35
Age1
0.9213
0.1120
0.674
63
24
Ed1
2.5918
0.5517
4.474
85
20

P > |z|

95% confidence
interval

< 0.001

2.2348
43
0.7260
01
1.7077
54

0.501
<0.001

6.0924
30
1.1692
96
3.9337
46

Log likelihood = 388.94571

The table below shows estimates from a logistic regression analysis.
What is the OR estimate for the effect of marital status adjusted for age and
education? Give your answer to 2 decimal places.
Interaction: Calculation: What is the OR estimate for the effect of marital status
adjusted for age and education? Give your answer to 2 decimal places.
Correct Response 3.69 (pop up box appears):
Correct

Yes, the odds ratio can be read from the table and it is 3.69. So, after adjusting for
age and education the odds of HIV infection for women who are widowed or divorced
is 3.69 times greater than the odds of HIV infection for married or single women.
Incorrect Response: (pop up box appears):
No, that's not right. In fact you can just read the odds ratio for the effect of marital
status from the table - it is 3.69. So, after adjusting for age and education the odds
of HIV infection for women who are widowed or divorced is 3.69 times greater than
the odds of HIV infection for married or single women.

6.5: Adjusting PAF for confounders

We can use this adjusted estimate to obtain an adjusted PAF for marital status.
PAF = p' ( 1) /
As we saw earlier, the proportion of cases who are divorced or widowed (exposed) is
38 / 181= 0.210, and the odds ratio = 3.69.
Therefore,
PAF = 0.210 (3.69 1) / 3.69 = 0.153
In other words, after adjusting for age and education, approximately 15% of HIV
cases can be attributed to being divorced or widowed

6.6: Adjusting PAF for confounders

The table below shows the frequency of cases and controls by formal education. Click
'swap' to see the logistic regression estimates again.
Calculate the PAF for education, adjusted for the effects of marital status and age.
Give your answer as a proportion (not a percentage), to 3 decimal places.
PAF =
Interaction: Calculation: PAF =____:
Correct Response 0.455:
Correct
Yes, the PAF is given by
PAF
= p' ( 1) /
= 0.741 (2.59 1) / 2.59
= 0.455
Incorrect Response:
No, that's not correct. From the table of cases and controls you can see that the
proportion of cases exposed to formal education is p' = 140 / 189 = 0.741. From the

table of results of logistic regression, the adjusted odds ratio is 2.59. Therefore,
the PAF is calculated as follows:
PAF
= p' ( 1) /
= 0.741 (2.59 1) / 2.59
= 0.455
Interaction: Button: Swap (graph on bottom centre changes to the following):
Estimates from logistic regression
Odds
Standa
z
ratio
rd
error
Msta
3.689935 0.94403
5.103
1
5
Age1
0.921363 0.11202
0.674
4
Ed1
2.591885 0.55172
4.474
0

P > |z|
< 0.001
0.501
<0.001

95% confidence
interval
2.2348
43
0.7260
01
1.7077
54

6.0924
30
1.1692
96
3.9337
46

Log likelihood = 388.94571

Formal
Education
None/adult
only
Some
Total

Cases
n

49

140

Controls
%

26
74

189

263
311

46
54

574

How can you interpret this estimated PAF?

Interaction: Button: clouds picture (pop up box appears):
After adjusting for marital status and age, 45.5% of HIV cases can be attributed to
exposure to some formal education.

6.7: Adjusting PAF for confounders

Computing PAFs using STATA:
The STATA commands ir, cs and cc can be used to obtain PAF estimates for incidence
rates (cohort studies), risks (cohort or cross-sectional studies) and for unmatched
case-control studies. Only unadjusted estimates are given, and no confidence
intervals are provided. However, calculation of adjusted PAFs from logistic regression
or mhodds output, using the equations given previously, is straightforward.

There is a user-written program called aflogit which will calculate adjusted PAFs and
confidence intervals based on asymptotic approximations, following logistic
regression or poisson regression, using the logistic and poisson commands.

Section 7: Joint effect of multiple exposures

So far you have considered the estimation of the PAF for a single exposure variable,
with or without adjustment for the effects of confounders. Let's now consider how
you can obtain an estimate of the fraction of cases attributable to two or more
variables.

The joint effect of two exposures is

not simply the sum of the individual
PAFs.

7.1: Joint effect of multiple exposures

To illustrate how the joint effect is obtained, consider two exposures from the
Mwanza case-control study of HIV risk factors.
The categories are as follows:
Number of lifetime sexual partners:
Group 0: 01 partners,
Group 1: > 1 partner.
Formal education:
Group 0: none/adult only,
Group 1: some formal schooling
Note: These variables are not clearly interpretable as causal, but we will put this
issue aside for now.
Interaction: Tabs: No. of partners:
Number of partners

Control
Case
Total

Number of
partners
0
1
173
379
27
156
200
535

Odds ratio = 2.64

(95% CI: 1.64 to 4.12)
P < 0.001

Total
552
183
735

NB: The totals are different between tables because the data used for the 'number of
partners' table contains missing values.
Interaction: Tabs: Formal Education:
Formal education

Control
Case
Total

Formal education
0
1
263
311
49
140
312
451

Total
574
189
763

Odds ratio = 2.42

(95% CI: 1.67 to 3.50)
P < 0.0001
NB: The totals are different between tables because the data used for the 'number of
partners' table contains missing values.

7.2: Joint effect of multiple exposures

The results of fitting a logistic regression model with number of lifetime sexual
partners and education are shown below.
Click 'swap' to see the distribution of cases between the two exposures.
Interaction: Button: Swap (the graph in the bottom centre changes to the
following):
Frequency of cases by number of partners and formal
education
Formal education
Total
Number of
0
1
partners
0
6
21
27
1
42
114
156
Total
48
135
183

Estimates from logistic regression

Odds
Standard
z
ratio
error
NPa
2.780940 0.641290 4.43
1
5
Ed1
2.421478 0.463038 4.62
5

P>
|z|
<
0.001
<
0.001

95% confidence
interval
1.769713 4.369988
1.664611

3.522477

Using the case frequencies and the odds ratio estimates from the logistic regression,
calculate the PAF for each exposure. Give your answers to 3 decimal places.

Number of partners: PAF =

Formal education:

PAF =

Interaction: Calculation: Number of partners: PAF =____:

Correct Response 0.546:
Yes, the PAF is given by
PAF
= p' ( 1) /
= (156 / 183) (2.781 1) / 2.781
= 0.546
This is the adjusted effect for number of partners.
Incorrect Response:
No, that's not right. Remember that the PAF is given by p' ( 1) / . Now, p' is the
proportion of cases exposed to the higher number of partners. From the table you
can see that p' = 156 / 183. The other table gives you the odds ratio for number of
partners, = 2.781. Therefore the PAF is:
PAF = (156 / 183) (2.781 1) / 2.781
= 0.546
This is the adjusted effect for number of partners.
Interaction: Calculation: Formal education: PAF =____:
Correct Response 0.433:
Correct
Yes, the PAF is given by
PAF
= p' ( 1) /
= (135 / 183) (2.421 1) / 2.421
= 0.433
This is the adjusted effect for number of partners.
Incorrect Response:
No, that's not right. Remember that the PAF is given by p' ( 1) / . Now, p' is the
proportion of cases exposed to some formal education. From the table you can see
that p' = 135 / 183. The other table gives you the odds ratio for formal education,
= 2.421. Therefore the PAF is:
PAF
= p' ( 1) /
= (135 / 183) (2.421 1) / 2.421
= 0.433
This is the adjusted effect for formal education.

7.3: Joint effect of multiple exposures

In order to calculate the PAF for the joint exposure of number of lifetime sexual
partners and formal education you need to look at the four combinations of the two
variables and the odds ratio at each level.
Interaction: Button: Note (pop up box appears):
Since there was no interaction you can assume the effects of the two exposures
combine multiplicatively. Therefore, for individuals with both exposures is given by
2.78 x 2.42 = 6.73. If there is an interaction, you should fit the model with the
interaction and adjust for the four combinations of the exposures.

No. of
partners
01
01
>1
>1
Total

Education

Cases

None
Some
None
Some

6
21
42
114
183

OR

PAF
1.00
2.78
2.42
6.73

0.073
0.135

Interaction: Hotspot: 0.073 (pop up box appears):

PAF

= p' ( 1) /
= (21 / 183) (2.78 1) / 2.78
= 0.073

Interaction: Hotspot: 0.135 (pop up box appears):

PAF

= p' ( 1) /
= (42 / 183) (2.42 1) / 2.42
= 0.135

Interaction: Calculation: yellow box in table:

Correct Response 0.530:
Correct
Yes, the PAF is given by:
PAF
= p' ( 1) /
= (114 / 183) (6.73 1) / 6.73
= 0.530
Incorrect Response:
No, that's not right. From the table, the proportion of cases exposed p' is 114 / 183,
and the odds ratio is 6.73. Therefore:
PAF
= p' ( 1) /
= (114 / 183) (6.73 1) / 6.73
= 0.530

Consider the table below. You can click on the highlighted PAF cells to see how each
one was calculated.
Can you calculate the missing value in the table, for women with > 1 partner and
some formal education?

7.4: Joint effect of multiple exposures

The joint effect of the two exposures is given by the sum of the individual PAFs as
shown in the table. PAF = 0.738 (73.8%) How do you interpret this?
Interaction: Button: clouds picture (pop up box appears):
PAF = 73.8% indicates that if both risk factors are removed then 73.8% of cases
might have been avoided.
No. of
partners
01
01
>1
>1
Total

Education

Cases

None
Some
None
Some

OR

6
21
42
114
183

PAF
1.00
2.78
2.42
6.73

0.073
0.135
0.530
0.738

Interaction: Hotspot: 0.073 (pop up box appears):

PAF

= p' ( 1) /
= (21 / 183) (2.78 1) / 2.78
= 0.073

Interaction: Hotspot: 0.135 (pop up box appears):

PAF

= p' ( 1) /
= (42 / 183) (2.42 1) / 2.42
= 0.135

7.5: Joint effect of multiple exposures

We previously calculated the PAFs of the exposures separately:
For number of partners:
PAF = 54.6%
For formal education:
PAF = 43.8%
Sum = 54.6% + 43.8% = 98.4%
not 73.8% as we have in the table!
In fact, it is common that the individual PAFs for several risk factors sum to more
than 100%. Of course, avoiding more than 100% of cases would be impossible!

You cannot simply assume that the proportion of cases not attributable to one or
more causes (for example, the 100 74.5 = 25.5% of HIV positive) are the only
ones for which other causes may play a part.
Interaction: Button: Example (pop up box appears):
This is an error frequently made with smoking and lung cancer, for which it is
incorrectly argued as, say, 90% of cases are attributable to smoking, other causes
can play a part in only 10%.

7.6: Joint effect of multiple exposures

This results from the fact that each individual PAF assumes that the values of all
variables apart from the exposure under consideration remain constant. When we
look at the joint PAF of two variables, this is not the case. We must identify the PAF
of one variable when the other is present, then the additional cases that could be
avoided after the effect of the first variable has already been removed.
There is no general formula for the joint PAF of two or more exposures, although if
the exposures are distributed in the population independently of one another, it can
be approximated by the formula below.
Joint PAF = 1 [(1 PAF1) x (1 PAF2)
x ... x (1 PAFK)]
This is shown opposite for the joint effect of number of partners and formal
education.
Number of partners:

PAF = 0.546

Formal education:

PAF = 0.438

Joint PAF = 1 [(1 0.546) x (1 0.438)]

= 1 [0.454 x 0.562]
= 0.745
= 74.5%
This is approximately the joint PAF calculated looking at the 4 combinations of the
two exposures.

Section 8: Probability of causation

Where legal liability for the effect of an exposure is involved, epidemiology is
sometimes called upon to answer the question:
"What is the chance that this case was
caused by a certain exposure?"
This is sometimes called forensic epidemiology.

Most often this question has arisen for cancers, with respect to ionising radiation,
asbestos, and other environmental exposures.
As AFE is the proportion of cases that would be avoided if the exposure were
removed, it is a natural extension to interpret it also as the probability that a specific
exposed case was caused by that exposure.
With this interpretation, it is called the probability of causation (PC), or assigned
share of causation.

8.1: Probability of causation

This interpretation and use of AFE depends on certain philosophical and social
assumptions. This can be particularly difficult if the effect of an exposure on
mortality is so strong that it affects a substantial fraction of total deaths among the
exposed.

Section 9: Summary
The main points of this session will appear below as you click through the step card
opposite. Click on any of the list entries below to go back to that page.
What do AF and PAF measure?
The attributable fraction in the exposed (AFE) is the proportion of cases in the
exposed group that is due to exposure.
The population attributable fraction (PAF) estimates the proportion of cases in
the entire population that is due to exposure.
Estimating PAF in different types of studies
The PAF can be estimated in different types of epidemiological studies.
In general, you need to have estimates of:
p the proportion of the population
exposed, or
p'
the proportion of cases in the
population exposed.
and
the rate ratio (or odds ratio of exposure in a case-control study)
Adjusting for confounders

To adjust a PAF for confounding factors you use the adjusted rate ratio, and the
formula using p'.

PAF K = p k '( k -1)/ k

The adjusted rate ratio (or odds ratio of exposure in a case-control study) is obtained
using stratification or regression methods.
p9c4rhs
Joint effect of multiple exposures
The fraction of cases attributable to the joint effect of multiple exposure is not simply
the sum of individual PAF for each exposure.
The joint effect can be estimated using the sum of the PAFs for different
combinations of the exposures.