Dose-Response Relation between
Physical Activity and Health:
An Observational Epidemiologic
Perspective
I-Min Lee, MD, ScD
Clinical Trials vs.
Observational Studies
“The paradox of the clinical trial is that it is the best
way to assess whether an intervention works, but is
arguably the worst way to assess who will benefit
from it.”
“It must be recognised explicitly that, in order to
apply the results of the trials to individual patients,
there must be a parallel investment in observational
studies – both quantitative and qualitative.”
Mant; Lancet 1999;353:743-6
Research Direction
1950’s to
early 1990’s
Does PA decrease risk
of chronic diseases?
early 1990’s
onwards
We know PA decreases
risk of chronic diseases:
- how much?
- how intense?
- what duration?
- etc.
“dose-response”
Dose-Response: Definition
• “Describes the change in effect … caused by
differing levels of exposure to a stressor”; central
to determining safe and hazardous levels for
drugs and pollutants (Wikipedia)
• PubMed search for “dose response” and
“exercise” – studies as far back as 1967, but
mainly related to drug effects; e.g., effect of betablocker on exercise tolerance in angina patients
• In the mid-1980’s, we begin to see the term used
for the study of different doses of PA
• ? Influence of medical/epidemiologic principles
What Do We Know (2000)?
“Most of the evidence currently available
seems to be related to the effects
(benefits or risks) of regular physical
activity rather than to the relationship
between dose and response.”
Dose-Response Issues
Concerning PA and Health:
An Evidence-Based Symposium
(Hockley Valley, Canada, 2000)
What Do We Know (2007)?
“Participation in aerobic and muscle-strengthening
physical activities above minimum recommended
amounts provides additional health benefits ...”
It is well-documented that PA of longer duration or
higher intensity is associated with additional risk
decrements but the exact shape of the … curve
remains unclear and may vary depending on the
health outcome of interest and baseline PA level …”
Physical activity and public health:
Updated recommendation for adults
from ACSM/AHA. Haskell et al,
MSSE 2007;39:1423-34
Why the Interest in
Dose-Response?
1. Knowledge is useful in clinical and
public health setting
2. Dose-response is one criterion used to
judge cause-and-effect relation in
observational studies (most studies of
PA-chronic disease are observational)
3. Dose-response can help us understand
the biology underlying physical
activity-disease relation
Above
average
Immunosurveillance
Average
Infection risk
Below
average
Sedentary
Moderate
Very High
Exercise Workload
Nieman DC, MSSE 2000;32:S406-S411
Issues to Consider in
Studies of Dose-Response
Definition of Physical Activity
•
“Physical activity” is ambiguous;
need to specify which measure is of
interest (total volume? intensity?
duration?)
•
Analogy: I want to study the doseresponse between “diet” and health
(total calories? fat calories? meat
intake? etc.)
Translation of Physical Activity
from Epidemiologic Studies
•
Most epi studies use questionnaires
•
Questionnaire usually tested for
reliability (repeatability) and validity
(“correctness”)
•
But, note that high correlations for
validity don’t necessarily mean the
absolute values (e.g, comparing
questionnaire and doubly-labeled
water) agree
• E.g., Harvard alumni questionnaire:
comparing qx estimate and doublylabeled water, r = 0.67
• Qx estimate of PA = 716 kcal/day
• But, doubly-labeled water estimate of PA
(total EE minus RMR) = 802 kcal/day
• Not surprising, since qx estimates
walking, stairs, sports/recreation ONLY
• (Doubly-labeled water minus TEE)
calculates ALL physical activity
(including, e.g., brushing teeth, fixing
meals, etc.)
Inter-relations among PA components (1)
• Inter-relations among PA components
contribute to complexity of analyses
• Appropriate analyses need to be
conducted, to get “right” answer
• E.g., if we see these rates of disease
associated with different PA intensities:
no activity
15/10,000 (referent)
mod intensity
12/10,000 (p=0.32)
vig intensity
5/10,000 (p=0.02)
• What interpretations are possible, and
what are the implications?
Inter-relations among PA components (2)
• In randomized trials, consider this in
design of the intervention arms
• Not possible in observational studies,
so control for this in analyses. The
aim is to mimic the controlled
environment in a randomized trial.
Choice of Cutpoints (1)
• Continuous measure of PA, or
categories?
• Continuous measure:
- if non-linear relation, yields little
additional info
- interpretation less understandable
• Categorical measure:
- allows visualization of disease rates
in each category
Choice of Cutpoints (2)
• If categorical, how many groups?
• Quantiles (e.g., tertiles, quintiles)
• “Natural” breaks according to
research interest
• “Natural” breaks on numerical scale
• Driven by biology
Shape of the Curve (1)
Increasing rates of disease
C
A
B
Increasing doses of activity
Shape of the Curve (2)
• Most investigators test for linear trend
on log scale
• Note that a straight line on the log
scale translates to an exponential
curve on the natural scale
• However, for small effects (RR between
0.5 and 1, or 1 and 2), curves on log
and natural scales look similar
• Other shapes of the curve are possible
(e.g., regression splines)
Spline Regression Model
(cubic spline, 4 knots)
Hu FB et al, JAMA 2000;283:2961-7
Shape of the Curve (3)
• Individual studies have their own ways
of assessing and categorizing PA
• This makes it difficult to combine data
across studies to synthesize info on
the shape of the dose-response curve
for a particular health outcome
• Investigators have done the best they
can for meta-analyses, but may be
combining disparate categories
E.g., “quintiles” of PA do not mean the
same thing in different studies:
Finnish men
(twin study):
US women
(Nurses’ Health Study):
Q1: <4.0 MET-hr/wk
Q1: <2.5 MET-hr/wk
Q2: 4.1-9.0
Q2: 2.5-7.2
Q3: 9.1-17.4
Q3: 7.3-13.4
Q4: 17.5-34.1
Q4: 13.5-23.3
Q5: 31.5+
Q5: 23.4+
Men in Q5 expended 60% more energy
than women in Q5
Dose-response curves for physical activity
and physical fitness in relation to CVD
Williams PT, MSSE 2001;33:754-761
Population under Study
• Physical activity varies according to
age, sex, race, etc.
• Thus the range of activity in different
populations being studied may vary
• This may have an impact on the doseresponse curve observed in any
single study
• (Which is why combining data across
studies is needed to get a good idea
of the complete dose-response curve)
Hypothetical Spectrum of Response
105
100
Event Rate
P, trend <0.001
95
90
Sample 1
85
Sample 2
80
1
2
3
4
5
6
7
Physical Activity Doses
8
9
10
In Sample 1:
• Less active population (could be older
persons, women)
• We likely would choose Group 1 as our
reference (a very sluggardly group!)
• We would see nice dose-response
In Sample 2:
• More active population (could be
younger persons, men)
• We likely would choose Group 5 as our
reference (a pretty active group)
• We would see little dose-response
Dose-Response for Volume
of Physical Activity
Dose-Response for Volume
• Probably the simplest measure to
analyze
• Any assessment of PA, no matter how
simple, gives an idea of the volume of
PA even if estimation of the actual
energy expended cannot be made
• E.g., “none”, “a little”, “a lot”
Garfinkel et al, Cancer 1988;62:1844-1850
Age-Std Death Rate/10,000
Harvard Alumni Health Study
105
100
P, trend <0.001
95
90
85
80
<500
500999
10001499
15001999
2002499
25002999
30003499
3500+
Physical Activity, Kcal/Week
Lee et al, JAMA 1995;273:1179-1184
Mortality Rates by Tertiles of Free-Living Activity Energy Expenditure
Health ABC Study
302 healthy adults, 70-82 y, followed ~6 y, 55 deaths
Manini et al, JAMA 2006;296:171-9
Dose-Response for Intensity
of Physical Activity
Dose-Response for Intensity
• 2 issues to be concerned about:
confounding by volume of PA, and
confounding by activities of other
intensities
• Analogous to confounding by total
caloric intake and confounding by
other nutrients in diet studies
• Can borrow methods from diet studies
to control for these issues
Standard multivariate model
Health Professionals’ Follow-Up Study
Tanasescu et al, JAMA 2002;288:1994-2000
“Energy decomposition”
method
Harvard Alumni Health Study
Multivariate RR’s for Mortality:
Energy
expended in:
<150
kcal/wk
150-399
kcal/wk
400-750
kcal/wk
750-1499
kcal/wk
1500+
kcal/wk
Vigorous
activities
1.00
(ref)
0.88
(0.81-0.96)
0.91
(0.81-1.02)
0.87
(0.76-1.00)
0.86
(0.76-0.96)
Nonvigorous
activities
1.00
(ref)
0.98
(0.86-1.11)
1.09
(0.96-1.23)
1.08
(0.96-1.23)
1.05
(0.93-1.18)
P, (vig) –
(nonvig)
--
0.19
0.04
0.02
0.02
Lee et al, JAMA 1995;273:1179-1184
Notes:
• It is not sufficient to note that the
coefficient (RR) for vigorous activities is
significant, while that for non-vigorous
activities is non-significant
• What is important is: are the coefficients
(I.e., [vig] versus  [nonvig]) DIFFERENT
from each other?
Harvard Alumni Health Study
Vig
Non-vig
1.2
P, trend = 0.32
1.1
1.0
0.9
0.8
P, trend = 0.007
0.7
0.6
<150
150-399
400-749
750-1499
1500
Physical Activity, Kcal/Week
Lee et al, JAMA 1995;273:1179-1184
Restriction/Stratification
Women's Health Study
1.2
1.0
P, trend <0.001
0.8
0.6
0.4
0.2
0.0
Does not Walk
<1
1-1.5
2+
Walking, Hr/Wk
* multivariate; among women without vigorous activity
Lee et al, JAMA 2001; 285:1447-1454
Harvard Alumni Health Study
Non-vigorous activity, kcal/wk*:
Age-Std Mortality Rate (per 10,000)
<500
149
500-999
155
1000-1499
166
1500-1999
132
2000-2499
176
2500-2999
125
3000-3499
200
3500+
* Among 17,321 men, 1195 performed only
non-vigorous activities (380 deaths)
143
P, trend = 0.99
Lee et al, JAMA 1995;273:1179-1184
The updated physical activity
recommendation (2007 ACSM/AHA)
allows for combining moderate and
vigorous activities –
What data support this?
WHI Observational Cohort
~ 10 MET-hr
~ 12.5
MET-hr
Manson et al, NEJM 2002;347:716-25
Dose-Response for Duration
(Frequency) of Physical Activity
Dose-Response for Duration
• Same issue: we are concerned about
confounding by total volume of PA
• Close correlation between volume of energy
expended and duration or frequency of PA
• Longer duration, more frequent bouts result in
greater energy expended
• To isolate dose-response for duration or
frequency, analysis must control for volume of
energy expended
Dose-Response for Duration
• This answers the question: For two persons
expending the same volume of energy, does
the one exercising in long bouts/less
frequently benefit more than the one
exercising in short bouts/more frequently?
• Almost no data on this from observational
studies
• (In randomized clinical trials, by design,
investigator can control for total volume of
energy expended)
Harvard Alumni Health Study
Duration (min)
Age-Adjusted RR
(95% CI)
No sports/rec
1.00 (ref)
1-15
0.85 (0.55-1.31)
16-30
0.76 (0.57-1.03)
31-45
0.85 (0.58-1.24)
46-60
0.80 (0.59-1.10)
>60
0.78 (0.62-0.98)
p, trend = 0.04
Lee et al, Circulation 2000;102:981-986
Harvard Alumni Health Study
Duration (min)
No sports/rec
1-15
16-30
31-45
46-60
>60
Bivariate RR*
(95% CI)
1.00 (ref)
0.94 (0.61-1.47)
0.92 (0.65-1.30)
1.07 (0.68-1.67)
1.02 (0.69-1.50)
1.05 (0.73-1.49)
p, trend = 0.68
* adjusted for age and volume of PA
Summary
• We’ve become more sophisticated in the
questions we ask about PA-disease
associations, moving from “is there an
association?” to “what are the details of
the association?”
• Many complex issues arise from the latter
question related to dose-response
• Clear that there still are gaps in knowledge
to be addressed