Intermediate methods in observational epidemiology

2008

Measures of Association

1) Measures of association based on ratios

Cohort studies
Relative risk (RR)
Odds ratio (OR)

Case control studies

OR of exposure and OR of disease
OR when the controls are a sample of the total population

Prevalence ratio (or Prevalence OR) as an estimate of

the RR

2) Measures of association based on absolute

differences: attributable risk

Cohort studies
Hypothetical cohort study of the one-year incidence (q) of acute myocardial
infarction for individuals with severe systolic hypertension (HTN, 180 mm Hg) or
normal systolic blood pressure (<120 mm Hg).
Severe
Number
Myocardial infarction
Systolic
Present
Absent
Probability (q) Probability oddsdis
HTN
Yes
10000
180
9820
0.0180
0.01833
No

10000

30

9970

0.0030

0.00301

180
0.0180
10000
RR

6.00
30
0.0030
10000

OR dis

q
0.0180
180
1 q 1.0 0.0180 9820 0.01833

6.09
q
0.0030
30
0.00301
1 q 1.0 0.0030 9970

Severe
Systolic
HTN
Yes

Number

No

Myocardial infarction
Present

Absent

Probability (q) Probability oddsdis

10000

180 (a)

9820 (b)

0.0180

0.01833

10000

30 (c)

9970 (d)

0.0030

0.00301

The OR can also be calculated from the crossproducts ratio if the table is organized exactly as
above :
a

OR disease

a
ab
ab
q
a
b
a
1

1 q
a b a b b ad

q
c
c
c bc
1 q
cd
cd
d
d
c
1

cd
c

OR disease

180 9970
6.09
9820 30

When (and only when) the OR is used to estimate the

RR, there is a built-in bias:

q
q
1 q
q 1 q
1 q
OR

q
1 q
q
q 1 q
1 q
RR
bias
Example:
Severe
Systolic
HTN
Yes

Number

No

Myocardial infarction
Present

Absent

10000

180 (a)

9820 (b)

0.0180

0.01833

RR=6.0

10000

30 (c)

9970 (d)

0.0030

0.00301

OR=6.09

OR dis

Probability (q) Probability oddsdis

1 0.003
6.0
6.09
1 0.018

IN GENERAL:
The OR is always further away from 1.0
than the RR.
The higher the incidence, the higher
the discrepancy.

Relationship between RR and OR

when probability of the event (q) is low:
q
q
1 q

1q
1 .0
or, in other words, (1-q) 1, and thus, the built-in bias term,
1q
and OR RR.
Example:
Severe
Systolic
HTN
Yes

Number

No

Myocardial infarction
Present

Absent

10000

180

9820

10000

30

9970

OR 6.0

180
RR 10000 6.00
30
10000

1 0.003
0.997
6.0
6.09
1 0.018
0.982

180
OR 9820 6.09
30
9970

Relationship between RR and OR

when probability of the event (q) is high:
Example:
Cohort study of the one-year recurrence of acute myocardial infarction
(MI) among MI survivors with severe systolic hypertension (HTN, 180
mm Hg) and normal systolic blood pressure (<120 mm Hg).
Severe
Systolic
HTN
Yes

Number

No

Recurrent MI
Present

Absent

10000

3600

6400

10000

600

9400

OR 6.0

q
0.36
0.06

3600
RR 10000 6.00
600
10000

1 0.06
0.94
6.0
8.81
1 0.36
0.64

3600
OR 6400 8.81
600
9400

OR vs. RR: Advantages

OR can be estimated from logistic regression.
OR can be estimated from a case-control study

Case-control studies

A) Odds ratio of exposure and odds ratio of disease

Hypothetical cohort study of the one-year incidence of acute
myocardial infarction for individuals with severe systolic hypertension
(HTN, 180 mm Hg) and normal systolic blood pressure (<120 mm Hg).

Severe
Systolic
HTN
Yes

Number

No

Myocardial infarction
Present

Absent

10000

180

9820

10000

30

9970

Odds dis exp

OR dis
Odds dis non-exp

same

Hypothetical case-control study assuming that all members of the

cohort (cases and non cases) were identified

Severe Syst HTN

Cases

Controls

Yes

180

9820

No

30

9970

OR exp

180
9820 6.09
30
9970

Odds exp cases

Odds exp non-cases

180
30 6.09
9820
9970

Retrospective (case-control) studies can estimate the OR of disease

because:

ORexposure = ORdisease
Because ORexp = ORdis, interpretation of the OR is always prospective.

Calculation of the Odds Ratios: Example of Use of

Salicylates and Reyes Syndrome
Past use of
salicylates
Yes

Cases

Controls

26

53

No

87

Total

27

140

Odds Ratios

(26/1) (53/87) =
43.0

Preferred Interpretation: Children using salicylates have an odds (risk) of

Reyes syndrome 43 times higher than that of non-users.
Another interpretation (less useful): Odds of past salicylate use is 43
times greater in cases than in controls.
(Hurwitz et al, 1987, cited by Lilienfeld & Stolley, 1994)

Cohort study:

O R

d is

O dds
O dds

d is e x p
d is u n e x p

180
9820

6 .0 9
30
9970

In a retrospective (case-control) study, an unbiased sample of the cases and controls yields an unbiased OR

It is not necessary that the sampling fraction be the same in both cases and
controls. For example, a majority of cases (e.g., 90%) and a small sample of controls
(e.g., 20%) could be chosen (assume no random variability).
(As cases are less frequent, the sampling fraction for cases is usually greater than that
for controls).

OR exp

Oddsexp in cases
Odds

exp in cntls

162
27 6.09
1964
1994

Case-control studies
B) OR when controls are a sample of the total population

OR exp

Risk factor

CASES

NON-CASES

Present
Absent

a
c

b
d

Odds exp cases

Odds exp non -cases

a
c
b
d

OR exp

TOTAL
POPULATION
a+b
c+d

Odds exp cases

Odds exp population

a
c
ab
cd

a
a b RR
c
cd
In a case-control study, when the control group is a sample of the total
population (rather than only of the non-cases), the odds ratio of
exposure is an unbiased estimate of the RELATIVE RISK

Example:
Hypothetical cohort study of the one-year recurrence of acute myocardial
infarction (MI) among MI survivors with severe systolic hypertension (HTN,
180 mm Hg) or normal systolic blood pressure (<120 mm Hg).
Severe
Systolic
HTN
Yes

Recurrent MI

Total
population

Present

Absent

3600

6400

10000

No

600

9400

10000

3600
RR 10000 6.00
600
10000

Example:
Hypothetical cohort study of the one-year recurrence of acute myocardial
infarction (MI) among MI survivors with severe systolic hypertension (HTN,
180+ mm Hg) or normal systolic blood pressure (<120 mm Hg).
Severe
Systolic
HTN
Yes

Recurrent MI

Total
population

Present

Absent

3600

6400

10000

No

600

9400

10000

Using a traditional casecontrol strategy, cases of

recurrent MI can be compared
to non-cases, i.e., individuals
without recurrent MI:

OR exp

3600
600 8.81
6400
9400

3600
RR 10000 6.00
600
10000

Example:
Hypothetical cohort study of the one-year recurrence of acute myocardial
infarction (MI) among MI survivors with severe systolic hypertension (HTN,
180+ mm Hg) or normal systolic blood pressure (<120 mm Hg).
Severe
Systolic
HTN
Yes

Recurrent MI
Present

Absent

3600

6400

10000

No

600

9400

10000

Using a traditional case-control

strategy, cases of recurrent MI are
compared to non-cases, i.e.,
individuals without recurrent MI:

OR exp

3600
600 8.81 ORdis
6400
9400

Total
population

3600
RR 10000 6.00
600
10000

Using a case-cohort strategy,

the controls are formed by the
total population:

OR exp

3600
3600
600 10000 6.00 RR
10000
600
10000 10000

Note that it is not necessary to have a total group of cases and non-cases or the total
population to assess an association in a case-control study. What is needed is a sample
estimate of cases and either non-cases (to obtain the odds ratio of disease) or the total
population (to obtain the relative risk). Example: samples of 20% cases and 10% total
population:

O R

exp

720
1 2 0 6 .0 R R
1 000
1 000

Thus RR= unbiased exposure odds estimate in cases divided by

unbiased exposure odds estimate in the total population.

To summarize, in a case-control study:

What is the control
group?

What is calculated?

Sample of
NON-CASES

OR exp

Sample of the
TOTAL POPULATION

OR exp

Odds exp cases

Odds exp non-cases

Odds exp cases
Odds exp total pop

To obtain ...
ORDisease

RR

How to calculate the OR when there are

more than two exposure categories
Example:
Univariate analysis of the relationship between parity and
eclampsia.*
Parity
2 or more
1
Nulliparous

Cases
11
21
68

Controls
40
27
33

* Abi-Said et al: Am J Epidemiol 1995;142:437-41.

OR

1.0 (Reference)
(21/11)(27/40)=2.9
(68/11)(33/40)=7.5

How to calculate the OR when there are

more than two exposure categories
Example:
Univariate analysis of the relationship between parity and
eclampsia.*
Parity
2 or more
1
Nulliparous

Cases
11
21
68

Controls
40
27
33

OR
1.0
2.9
7.5

* Abi-Said et al: Am J Epidemiol 1995;142:437-41.

Correct display:
Log
scale

Baseline is 1.0

12 for linear trend 29.215,

p 0.0001

A note on the use of estimates from a

cross-sectional study (prevalence ratio, OR) to estimate the RR
P
P I D
1 - P I D


If the prevalence is low (~5%)
Prevalence Odds=
P
I D
P I D
1 - P

P I

P I
If this ratio= 1.0
Duration (prognosis) of the disease after onset is
independent of exposure (similar in exposed and
unexposed)...
However, if exposure is also associated with shorter survival (D+ < D-), D+/D- <1 the
prevalence ratio will underestimate the RR.

P I

P I
Example?

Smoking and emphysema

Measures of association based on absolute differences

(absolute measures of effect)

The excess risk (e.g., incidence)

among individuals exposed to a certain
risk factor that can be attributed to the
risk factor per se:
AR exp q q 20

1000

10

1000

10 / 1000

Or, expressed as a proportion

(e.g., percentage):
%AR exp

q q
20/1000 - 10/1000
100
100 50%
q
20/1000

Alternative formula for the %ARexp:

%AR exp

RR - 1
2.0 - 1.0
100
100 50%
RR
2.0

Incidence (per 1000)

Attributable risk in the exposed:

20/1000

ARexp
10/1000

Unexposed Exposed

 Population attributable risk:

The excess risk in the population that can be attributed to a given risk
factor. Usually expressed as a percentage:

%PopAR exp

qpop q
qpop

100

The Pop AR will depend not only on the RR, but also on the prevalence
of the risk factor (pe).
p e (RR 1)
100
Levins formula %PopAR exp
p e (RR 1) 1
(Levin: Acta Un Intern Cancer 1953;9:531-41)

Pop AR

Pop AR

ARexp

Unexposed Population Exposed

Incidence (per 1000)

High exposure prevalence

Incidence (per 1000)

Low exposure prevalence

ARexp

Unexposed Population Exposed

Chu SP et al. Risk factors for proximal humerus fracture. Am J Epi 2004; 160:360-367
Cases: 448 incident cases identified at Kaiser Permanente. 45+ yrs old, identified through
radiology reports and outpatient records, confirmed by radiography, bone scan or MRI.
Pathologic fractures excluded (e.g., metastatic cancer).
Controls: 2,023 controls sampled from Kaiser Permanente membership (random sample).
Dietary Calcium (mg/day)

Odds Ratios (95% CI)

Highest quartile (970)

1.0 (reference)

Third quartile (771-969)

1.36 (0.96, 1.91)

Second quartile (496-770)

1.11 (0.81, 1.52)

Lowest quartile (495)

1.54 (1.14, 2.07)

What is the %AR in those exposed to the

lowest quartile?

More or less 1.0

Percent ARexposed

RR - 1
OR - 1
1.54 1
100 ~
100
100 35%
RR
OR
1.54

Interpretation: If those exposed to values in the lowest

quartile had been exposed to other values, their odds
(risk) would have been 35% lower.
What is the Percent AR in the total population due to exposure in the lowest quartile?
Levins formula for the Percent ARpopulation
Percent Population AR

p
p

exp

exp

( R R 1)

( R R 1) 1

Pexp (RR 1)
Pexp (RR 1) 1

1 0 0 ~

p
p

exp

exp

(O R 1)

(O R 1) 1

100
100

RR estimate ~ 1.54
Pexp ~ 0.25

0 . 2 5 ( 1 .5 4 1 )
1 0 0 1 1 . .9 %
0 . 2 5 ( 1 .5 4 1 ) 1

Interpretation: The exposure to the lowest quartile is responsible for about 12% of the total
incidence of humerus fracture in the Kaiser permanente population