American Journal of Epidemiology Vol. 142, No.

7
Copyright © 1995 by The Johns Hopkins University School of Hygiene and Public Health Printed in U.S.A.
All rights reserved

Effect of Risk Factor Values on Lifetime Risk of and Age at

First Coronary Event
The Adventist Health Study

Gary E. Fraser,1 Kristian D. Lindsted,1 and W. Lawrence Beeson1

The effect of traditional coronary heart disease risk factors on lifetime risk, age at onset, and survival free
of coronary disease has not been extensively studied. The authors have used the cohort data from 27,321
California Seventh-day Adventists who had no known heart disease in 1976 to investigate these questions.
Multiple decrement life tables incorporating non-parametric estimates of conditional probabilities for both
coronary disease and all other competing endpoints were used to estimate these survival outcomes. Variance
estimators are provided in an appendix. Persons characterized by being either past smokers, diabetic,
hypertensive, physically inactive, non-vegetarian, or infrequent consumers of nuts often showed substantial
differences in these survival outcomes. Statistically significant results include earlier age at onset of coronary
disease at between 4 and 10 years, reduced life expectancy free of the disease between 5 and 9 years, and
increased lifetime risk between 8% and 16%, when comparing groups with and without adverse values for
different risk factors. The presence of adverse levels of two risk factors predicted even greater differences in
these endpoints. These important effects are easily understood by the layman or non-epidemiologist profes-
sional, which is often not true of a relative risk. This should increase the effectiveness of such results when
promoting behavioral change. Am J Epidemiol 1995;142:746-58.

coronary disease; life table; risk factors; statistics

The effect of many physiologic and life-style vari-
ables to increase risk of coronary disease events is now
well known. However, the epidemiologic literature has
almost exclusively measured their impact on the dis-
ease in terms of relative risks. This is a useful concept
when assessing causality, but has some disadvantages
in other respects. For instance, a relative risk of 0.5 is
difficult for non-epidemiologists or the lay public to
interpret. Many may have a hazy notion that this
means that the risk of heart attack is halved, but they
do not realize that this is usually so only for short,
defined time periods (often one year), and that there
are summary exponential effects over a lifetime.
Clearly, the lifetime risk of heart disease is not in
general halved. For instance, our cohort of Seventh-
day Adventists is well known to have a reduction in
risk of coronary disease events and cancer when com-
pared with the general population, this being measured
Received for publication May 9,1994, and in final form December
23, 1994.
Abbreviation: NS, not significant.
1
Center for Health Research, Loma Linda University, Loma
Linda, CA.
Reprint requests to Dr. Gary E. Fraser, Loma Linda University,
Center for Health Research, Nichol Hall #2008, Loma Linda, CA
92350.
by relative risks substantially less than 1.0 (1). Yet,
despite this, cardiovascular diseases (International
Classification of Diseases, 9th Revision, codes 390-
459 as underlying or immediate cause on the death
certificate) account for 56.7 percent of deaths and
cancer for 20.9 percent of deaths, which is very similar
to proportions found in non-Seventh-day Adventists.
Actually, the average age at onset of a disease can be
markedly impacted by the effect of the exposure on
competing causes. For instance, if the exposure is
protective for the disease of interest but hazardous for
the aggregate of competing causes, the average age at
onset of the disease of interest could be reduced, as
few persons remain to be at risk at higher ages, due to
deaths from competing causes. If an exposure is haz-
ardous for the disease of interest, then the impact on
lifetime risk also depends on whether the exposure
affects competing causes. As an example, one could
imagine a situation where the effect on competing
causes was also hazardous, such that the proportion of
persons dying of the disease of interest (i.e., the life-
time risk) was not changed by the exposure. Clearly,
the effect of an exposure on outcomes such as lifetime
risk and average age at onset is complex and not

746

estimable by just knowing the relative risk for the
disease of interest.
The purpose of this paper is to apply and describe an
approach to the analysis of cohort study data that
allows the estimation of effects of particular practices
or risk factors on the lifetime risk of coronary heart
disease, and also how such effects may delay or ad-
vance the first expression of coronary disease, taking
into account competing risks from other sources of
mortality. The methods are illustrated by the use of the
risk factors hypertension, diabetes mellitus, exercise
habits, obesity status, past cigarette smoking, vegetar-
ian status, and consumption of nuts. While a number
of other investigators (2-8) have developed methods
for analyzing epidemiologic data in this fashion, we
know of few reports of such applications. We have
previously published relative risk analyses for coro-
nary disease from the Adventist Health Study (9, 10)
and now reexamine these data with new statistics.
MATERIALS AND METHODS
The Adventist Health Study is a cohort investigation
of 34,198 California non-Hispanic white subjects liv-
ing in Seventh-day Adventist households. The detailed
methods have been described elsewhere (11). Briefly,
the population included all volunteers over age 24
years in 1974 who responded to an invitation to par-
ticipate. Two extensive questionnaires were mailed to
participants in 1974 and 1976. The analyses below
used the 27,321 subjects who had no previous history
of coronary disease at entry to the study and excluded
non-Seventh-day Adventist members of study house-
holds. Exposure variables chosen for this report in-
clude traditional risk factors and two dietary variables
that showed significant associations with coronary dis-
ease risk in our previous relative risk analyses (9, 10).
The exposure variables were physician-diagnosed hy-
pertension or diabetes mellitus; self-reported height
and weight, from which the body mass index (weight
(kg)/height (m)2) was calculated; previous cigarette
smoking; exercise; and dietary habits. We note that
there were too few current smokers in this population
for these analyses. Exercise, categorized as low, me-
dium, and high, was a cross-categorization of two
questions pertaining to occupational and leisure activ-
ities (9). Dietary assessment was by a semiquantitative
food frequency questionnaire (10). Specifically, the
frequency of consumption of nuts was asked in eight
categories, ranging from "never consume" to "more
than once per day." These were then collapsed to less
than once a week, 1-4 times a week, and ^ 5 times a
week. A vegetarian was defined by eating flesh foods
less than once per week.
Am J Epidemiol Vol. 142, No. 7, 1995
Risk Factors and Survival Free of Coronary Disease 747
For any particular exposure variable, missing data
occurred in 0.8 to 8.8 percent of subjects depending on
the variable. Only subjects with missing data for the
exposure variable(s) necessary for that particular
analysis were excluded. Loss to follow-up occurred
in 2.5 percent of subjects during the 6 years, and
these were censored as described below in the
person-time calculations.
During 1976-1982, a brief annual questionnaire
was mailed to ascertain any hospitalizations during
that year. Study personnel then visited or communi-
cated with all relevant hospital medical record rooms.
Where a diagnosis of cancer or myocardial infarction
seemed possible, all histology reports, medical histo-
ries and examinations, electrocardiograms, and cardiac
enzyme results were microfilmed. A cancer diagnosis
required that the histology be confirmed. Fatal and
non-fatal coronary heart disease events were diag-
nosed using published international diagnostic criteria
(12). Deaths were ascertained by several mechanisms.
These included the use of church records, computer
matching with California state death tapes (1976-
1982), and matching with the National Death Index
(1979-1982).
Three outcome measures are estimated with the use
of multiple decrement life tables: 1) lifetime risk; 2)
mean age at onset (fatal or nonfatal coronary disease
combined); 3) life expectancy free of coronary dis-
ease, which allows either onset of the disease (fatal or
nonfatal combined) or death from a competing cause
as part of a composite endpoint.
Statistical methods
The basic parameters that are first estimated in these
analyses are 1) the age-specific conditional probabili-
ties of the disease of interest, and 2) the age-specific
conditional probabilities of all other causes of failure
(i.e., all other causes of death) that removed subjects
from risk of a coronary disease event.
The above probabilities were estimated as follows:
Using a non-parametric approach with attained age
as the time variable, the unit of time was one year of
age. If qd{k) and qc(k) are the conditional probability
estimates at age k for the events of the disease of
interest and the competing causes of mortality, respec-
tively, then qd{k) = Nd(k)/P(k), where Nd(k) is the
number of new events of disease d during the age
interval {k, k + 1), and P(k) is the number of subjects
at risk during the kth year of age. Similarly, qc(k) =
Nc{k)IP{k). We note that under this model, events are
mutually exclusive, and a particular subject either sur-
vives the age interval, withdraws, develops coronary
disease, or experiences a competing event. To accom-
modate subjects who were scheduled to start or end
748 Fraser et al.

follow-up at fractional ages, say k + rm(fc) for start of We have used the following equation:
follow-up, or k + sm(k) for end of follow-up, we used
w

2[qd(k)-S(k)-(Uk-y)]
k=y
P(k)= l(sm(k)-rm(k)), + y,
m=\

where n subjects contribute time to age k, where k=y

sm(k) = 1 except possibly for the final interval, and where S(k) is the survival to the beginning of the
rm{k) = 0, except possibly for the initial interval. In one-year interval represented by age k, and is esti-
the case of withdrawal or other censoring at k + tm(k), mated by
we calculated sm(k) = tm{k) in the above formula as
suggested by Hoem (13). i-l
The highest age at which any event is observed is
defined as w,. A final open-ended age interval denoted t=y
by w begins with age w0, where w0 is defined as the
smaller of 101 years or the greatest age k that is less Note that S(y) = 1 by convention.
than or equal tow,, such that the sum of both types of The second outcome measure, life expectancy free
of disease d, is estimated by
event at all ages greater than k — 5 exceeds four. This
last is to give a minimum of at least five events with w
which to estimate the mean survival time through w ed+c(y) - 2 {S(k) • (qd(k) + qc(k)) • (Uk - y)} + y.
for those surviving to w0. k=y
The life table was closed out using the method
described by Manton and Stallard (14), where data The third outcome measure, lifetime risk of disease
from ages w0 — 5 to w, years were used to estimate d, ld(y) is estimated by
conditional probabilities qc(w), qd(w), and C/w. Then w

id(y)= 2{s(k)-qd(k)}.
k=y
qd(w) = £ Nd(k)l Nc(k)),
Note that this is the denominator of ed(y) above.
Variance estimates for ed(y), ed+c(y), ld(y) are given in
and similarly for qc(w). In the formulae for ed(y) and the Appendix.
ed+c(y) below, the value of Uk = k + 0.5 when k < The non-parametric analyses are run separately for
w0, with the 0.5 reflecting the assumption of a uniform both sexes. An approximate test of effect modification
distribution of deaths during one year of age. When by sex is given by
(a\ ~ bi) - (a2 - b2)
(Nd(k) + Nc(k))
k=wo—5
k = w, Uw = w0 — I/log 1 - Vi=1

P(k) as a z statistic, where a and b are the outcome mea-

the latter expression being the estimated mean age at
death for those surviving to w0 using the inverse of the
estimated total event rate.
Then using the multiple decrement life table ap-
proach (14), it is possible to estimate ed(y), which is
the life expectancy free of disease d among those
persons who will eventually develop disease d. In all
that follows, we call ed(y), the mean age at onset,
where observation begins at age y. This can be written
algebraically in a number of ways.
sures for two different values of the risk factor, and
gender is symbolized by i (=1,2); V, = Var(a, — &,) =
Var(a() + Var(b(). Note that ai and bt are independent
as they result from analyses of separate populations
(i.e., those exposed or nonexposed).
If the test of effect modification is nonsignificant
(p > 0.10), differences in the outcome measure by risk
factor status are summarized (5) over the two sexes
using a weighted average of the sex-specific differ-
ences. The weights are the inverses of the respective
variance estimates.
Hence,

Am J Epidemiol Vol. 142, No. 7, 1995

 Risk Factors and Survival Free of Coronary Disease 749

These tables show that the predicted age at onset

S=
Assuming approximate normality, this implies that
tests the null hypothesis that 5 = 0.
Because ld(y) has a lower bound of zero and ex-
pected values sometimes within one or two standard
errors of this lower bound, it is evident that the distri-
bution is often moderately skewed. Thus, confidence
intervals for ld(y) were based on
± 1.96,/Var[logtt/y))]},
VarO/GO)
where Var[log(/rf(v))] was estimated by
RESULTS
The person-years at risk and the number of coronary
disease events by age are shown in table 1. The results
in tables 2-4 compare the effects of presence or ab-
sence, or relatively high or low values, of individual
risk factors and are thus univariate (apart from the
control of gender). Table 5 evaluates the effect of
selected pairs of risk factors operating together, where
those combinations selected were all the combinations
where there were at least 20 coronary disease events
and 20 complimentary events, so as to approximately
preserve the large sample normality of the outcome
measures.
among those who develop coronary disease is substan-
tially earlier in diabetics, hypertensives, past smokers,
the inactive, and males. This was possibly also true for
relatively obese women. The predicted number of
years of life expectancy free of coronary disease was
significantly less in males, the inactive, diabetics, low
nut consumers, and hypertensives. The predicted life-
time risk of the disease was significantly greater in
diabetics, hypertensives, and low nut consumers. This
was also true for obese males and non-vegetarian
males. In general, the pairwise combinations of the
variables showed qualitatively similar differences to
those seen with the constituent individual factors, but
are often of substantially greater magnitude, as might
have been expected.
The above results all pertain to subjects in whom the
nominated exposure status was present by age 30 years
and the life table begun at that age (i.e., y = 30 in the
equations of the "Statistical methods" section). The
corresponding results for subjects free of the disease
and starting the life table at age 60 years are as
follows. Compared with females, males develop cor-
onary disease 7.55 years earlier (p < 0.001) and have
a 3.06 percentage points greater remaining lifetime
risk (not significant (NS)). Past smokers compared
with never smokers develop the disease 4.03 years
earlier (p < 0.05), and have a 2.32 percentage points
greater remaining lifetime risk (NS). Persons in the
medium tertile of body mass index, compared with
persons in the highest tertile, develop coronary disease
3.99 years later (NS) and have a 3.30 percentage
points lower remaining lifetime risk (NS). In compar-
ison with men in the highest body mass index tertile,
men in the lowest tertile develop coronary disease 0.96
years earlier (NS), while women in the highest tertile
develop the disease 7.71 years later than women in the
lowest tertile (NS). Compared with men in the highest
body mass index tertile, men in the lowest tertile have
a 15.15 percentage points lower lifetime risk (p <

TABLE 1. Number of definite incident coronary heart disease events and person-years at risk by age: the Adventist Health
Study, 1977-1982

Males Females Total

Age (years)
No. of events Person-years No. of events Person-years No. of events Person-years

35-49 11 17,158 5 26,316 16 43,474

50-64 53 18,734 14 29,146 67 47,880
65-79 98 12,475 73 24,390 171 36,865
80-89 58 3,210 89 7,072 147 10,282
>90 10 554 37 1,298 47 1,852

Total 230 52,131 218 88,222 448 140,353

Am J Epidemiol Vol. 142, No. 7, 1995

750 Fraser et al.

TABLE 2. Predicted age at onset of coronary heart disease—non-parametric estimates: California Seventh-day Adventists
Males Females
Summary
Exposure variable
Age Age difference
(years)
95% Clt Difference
(years)
95% Cl Difference

Smoking
Past 73.1 68.9-77.4 83.4 77.1-89.7
Never 80.2 76.7-83.8 7.1** 87.9 85.5-90.2 4.5 6.0***
Body mass index (kg/m2):):
High 77.1 70.3-83.9 80.1 70.1-90.1
Medium 79.1 73.4-84.7 2.0 89.7 84.4-94.9 9.5* 4.8
Low 77.6 71.8-83.3 0.5 88.8 86.2-91.3 8.7* 4.0
Exercise
Low 75.6 72.0-79.2 84.5 82.0-86.9
High 78.4 72.7-84.1 2.8 91.6 83.7-99.6 7.1* 4.5*
Diabetes mellitus
Yes 68.6 60.4-76.9 76.0 69.8-82.3
No 78.3 75.3-81.3 9.7** 88.7 85.9-91.5 12.7**** 11.5****
Hypertension
Yes 75.0 71.4-78.6 83.2 79.5-86.9
No 77.7 73.9-81.6 2.7 89.3 85.8-92.9 6.1** 4.5**
Vegetarian
No 78.3 72.4-84.2 83.6 78.5-88.8
Yes 79.3 75.3-83.4 1.0 87.5 84.6-90.4 3.9 2.7
Nut consumption
Low 75.4 71.9-79.0 85.7 83.1-88.3
High 81.0 74.9-87.0 5.6 85.7 75.4-96.1 0.0 3.9
Sex 77.8 75.1-80.4 87.1 84.7-89.5 9.3****
* p =£ 0.10; * * p < 0.05; *** p < 0.01; * * * * p < 0.001. t Cl, confidence interval. % Differences are in comparison with high body mass
index.

0.05). However, women in the highest fertile com- DISCUSSION

pared with women in the lowest tertile have a 0.28
percentage points greater remaining lifetime risk of
coronary disease (NS). Compared with persons who
are more physically active, less physically active per-
sons develop coronary disease 4.17 years earlier (p <
0.10) and have a 5.08 percentage points greater re-
maining lifetime risk (NS). Diabetics compared with
nondiabetics develop coronary disease 8.03 years ear-
lier (p < 0.001) and have a 4.95 percentage points
greater remaining lifetime risk (NS). Compared with
nonhypertensives, hypertensives develop coronary
disease 2.82 years earlier (p < 0.10) and have a 10.2
percentage points greater remaining lifetime risk (p <
0.01). In comparison with vegetarians, non-vegetari-
ans develop coronary disease 1.77 years earlier (NS).
Among males, non-vegetarians versus vegetarians
have an 11.9 percentage points higher remaining life-
time risk (p < 0.05), but non-vegetarian females have
a 0.26 percentage points lower remaining lifetime risk
than female vegetarians (NS). Compared with persons
who eat nuts at least five times each week, those who
rarely consume nuts develop coronary disease 2.64
years earlier (NS) and have an 11.9 percentage points
greater remaining lifetime risk (p = 0.05).
One motivation for these analyses was our belief
that it is important to present results of risk factor
analyses in a way that the layman and the non-
epidemiologist professional can readily grasp. This
has not always been true with the relative risk ap-
proach, which we believe is usefully complemented by
analyses such as those presented here. The potential
for good compliance with medications or change in
behavior is enhanced if the patient has a clear under-
standing of the possible consequences of his/her ac-
tions. An extension of years free of coronary disease
or a reduction in lifetime risk of the disease are con-
cepts that can be readily understood by most people.
Such results can also be applied to populations. Then
it may be possible to estimate the average number of
years of deferral of onset of coronary disease in the
population, or the proportion of the population who
will develop the disease if a preventive policy is en-
acted. Clearly, this would be very useful information
for both the planning of health policy and the estima-
tion of the necessary public health resources. As usual,
a policy maker must also consider the effects of such
possible interventions on other clinical endpoints.
However, it is important to point out that these are
univariate results (aside from control of gender), and

Am J Epidemiol Vol. 142, No. 7, 1995

 Risk Factors and Survival Free of Coronary Disease 751

TABLE 3. Predicted life expectancy free of coronary heart disease (CHD)—non-parametric estimates in years: California
Seventh-day Adventists
Males Females

Ufe Ufe Summary

Exposure variable
expectancy expectancy difference
free of CHD 95% Clf Difference
free of CHD
95% Cl Difference
(years) (years)
Smoking
Past 80.2 76.0-84.3 83.7 72.8-94.6
Never 83.3 79.9-86.7 3.1 86.6 84.0-89.3 2.9 3.1
Body mass index (kg/m2)}:
High 81.7 75.3-88.0 86.5 81.1-91.8
Medium 83.3 77.9-88.7 1.9 88.4 81.4-95.4 1.9 1.9
Low 83.1 79.7-86.5 1.4 86.0 82.7-89.2 -0.5 0.3
Exercise
Low 78.3 75.2-81.5 83.7 81.2-86.2
High 83.6 78.8-68.4 5.3* 89.7 85.8-93.5 6.0** 5.7***
Diabetes mellitus
Yes 75.3 70.6-80.0 78.1 73.8-82.4
No 82.7 79.9-85.5 7.4*" 87.2 84.3-90.2 9.1"** 8 3***»
Hypertension
Yes 77.5 74.2-80.8 84.0 80.5-87.5
No 83.5 80.2-86.8 6.0** 87.7 84.0-91.3 3.7 4.9***
Vegetarian
No 80.7 75.9-85.5 85.7 82.6-88.8
Yes 84.0 80.1-87.9 3.3 86.5 83.1-69.9 0.8 1.7
Nut consumption
Low 79.4 77.0-81.8 84.7 81.2-88.2
High 85.3 80.9-89.8 5.9** 89.4 82.1-96.6 4.7 5.6**
Sex 82.1 79.6-84.7 86.4 83.8-88.9 4.3**
' p < 0.10; *• p s 0.05; **• p s, 0.01; **** p s 0.001. t Cl, confidence interval, t Differences are in comparison with high body mass
index.

that some confounding between variables is likely. It
can be stated that in this data set relative risk analyses
adjusting only for age and sex gave estimates for all of
these variables (except vegetarianism) quite close to
those found when all variables were included together
in the model. This suggests that confounding is minor
in relative risk analyses. Whether such confounding as
there is would have a greater effect on survival out-
comes compared with relative risks is unknown. Con-
sequently, our results as presented above should be
used cautiously for etiologic arguments, but are useful
for prediction. A multivariate approach is necessary to
effectively deal with confounding.
The results shown in tables 2-4 contain predictions
of substantial differences between exposure sub-
groups. For any particular endpoint, there is generally
good concordance between the sexes, with the excep-
tion of body mass index and vegetarian status. In
females, body mass index is not related to the lifetime
risk of coronary disease, but, among persons who are
predicted to develop the disease, the obese do so 8.7
years earlier (p < 0.10). However, obese males have a
16.3 percentage points greater lifetime risk than do
males in the lowest body mass index tertile (p < 0.05),
whereas age at onset in males does not depend on the
body mass index value. One implication may be that in
obese females the disease process in those who are
susceptible is accelerated, whereas in obese males a
higher proportion are sensitized (or initiated) but the
disease is not accelerated. As in relative risk analyses,
we again find that for men compared with women
vegetarian status appears to be more important for
coronary disease. The only survival outcome affected
by vegetarianism is lifetime risk, which is 13.1 per-
centage points greater in non-vegetarian males than
vegetarian males (p < 0.05), while in females, vege-
tarianism has no apparent impact. Such a difference
between the sexes in the effect of diet on coronary
disease risk has been postulated by other investigators
(15).
Our data predicted that, compared with females,
males have a 4.5 percentage points greater lifetime
risk of coronary disease (NS), and males who develop
the disease experience onset 9.3 years earlier than do
females who develop it (p < 0.001). Thus, despite the
fact that this is a relatively low risk population (16),
there still remains a substantial advantage in being
female with respect to risk of coronary disease.
Diabetes mellitus, as one might expect from com-
mon clinical experience, has major effects in acceler-
ating the expression of coronary disease in susceptible
persons, and also in increasing the lifetime risk of

Am J Epidemiol Vol. 142, No. 7, 1995

752 Fraser et al.

TABLE 4. Predicted lifetime risk of coronary heart disease—non-parametric estimates (%): California Seventh-day Adventists
Males Females
Summary
Exposure variable Lifetime
Ufetime difference
risk (%)
95% Clt Difference
risk (%)
95% Cl Difference

Smoking
Past 31.0 24.9-38.5 30.3 12.7-72.2
Never 26.6 21.5-32.8 4.4 22.9 19.0-27.6 7.4 4.7
Body mass index (kg/m2)^
High 34.0 24.0-48.3 21.5 10.0-46.2
Medium 26.1 18.6-36.5 7.9 25.4 16.1-40.2 -3.9 3.8
Low 17.7 11.7-26.9 16.3** 21.3 16.7-27.3 0.2 9.9**
Exercise
Low 29.7 23.1-38.4 23.4 19.4-28.3
High 24.0 17.9-32.3 5.7 25.9 8.8-76.6 2.5 4.8
Diabetes mellitus
Yes 41.1 28.4-59.5 29.2 21.1^0.4
No 27.9 23.7-32.9 13.2* 24.1 19.6-29.6 5.1 7.7**
Hypertension
Yes 37.8 30.7-16.6 29.1 23.5-35.6
No 24.2 19.9-29.5 13.6*** 20.3 15.1-27.4 8.7* 11.0****
Vegetarian
No 34.4 26.2-45.1 19.6 12.1-31.7
Yes 21.3 16.0-28.5 13.1" 23.2 18.3-29.4 -3.6 §
Nut consumption
Low 31.3 24.8-39.4 29.0 23.3-36.1
High 18.7 10.4-33.5 12.6* 17.0 7.8-37.4 12.0 12.4**
Sex 28.4 24.5-32.9 23.9 20.0-28.6 4.5
* p < 0.10; * * p < 0.05; * * * p =£ 0.01; **** p < 0.001. t Cl, confidence interval, t Differences are in comparison with high body mass
index. § Not summarized. Significant (p < 0.10) effect modification by gender.

coronary disease. Relatively large effects on all end-
points were found in our data. These effects are quite
consistent for both males and females. Knuiman et al.
(17) found in Australia that the life expectancy of type
II diabetics compared with non-diabetics at age 45
years was decreased by 11.4 years in females and by
3.5 years in males. While we did not report life ex-
pectancy data, this should not greatly differ from our
endpoint "life expectancy free of coronary disease." In
fact, our results for all diabetics were quite similar for
both sexes and fall between the male and female
results reported by Knuiman et al. (17).
Hypertensives also experience both an 11.0 percent-
age points greater lifetime risk of coronary disease (p
< 0.001) and an earlier expression of the disease by
4.5 years in susceptible persons (p < 0.05). In this
well-educated population (11), 49.6 percent of those
subjects who had "ever been told by a doctor" that
they had high blood pressure were also on antihyper-
tensive therapy in 1976. Presumably, untreated current
hypertensives in 1976 would have shown even greater
adverse effects. The survival experience of past smok-
ers was also of interest. Although about 19 percent of
our subjects had been past smokers, the average dura-
tion of church membership for those persons was 23.7
years. Hence, these data seem to indicate either a
long-term persisting effect of cigarette smoking
(which our multivariate risk ratio analyses also sug-
gested (10)), or that the past smokers have adverse
values of other risk factors.
We note the apparent salutary effect of frequent nut
consumption (>5 times per week). On average, per-
sons characterized by this simple dietary habit have a
12.4 percentage points lower lifetime risk of coronary
disease (p < 0.05), and males who develop the disease
do so 5.6 years later than males who consume few nuts
(NS). Whether this is causal or not needs further
investigation, but our previous multivariate risk ratio
analyses found little evidence of confounding (10) and
at least one other major cohort study has reported
similar risk ratio findings (18) for nut consumption.
Table 5 shows predicted survival experience based
on selected pairs of risk factors, where both factors
take either high or low risk values. These analyses
suffer from lower numbers and hence lower statistical
power. However, there are few surprises or evidence
of effect modification. The predicted differences in
survival experience between the contrasting subgroups
are often very substantial. However, such pairwise
combinations of such risk factors are not at all uncom-
mon in the population. Clearly, the analysis of effects
of several variables jointly will be more efficient and
is often only possible with multivariate techniques that
should improve statistical power but require more
assumptions.

Am J Epidemiol Vol. 142, No. 7, 1995

 Risk Factors and Survival Free of Coronary Disease 753

TABLE 5. Coronary heart disease (CHD) survival comparing contrasting values of selected! pairs of exposure variables:
California Seventh-day Adventists
Ufe
Age at
Lifetime expectancy
Exposure variables Sex
risk (%)
95% a * Difference 95% Cl Difference onset 95% Cl Difference
free of CHD
(years)
(years)

Smoking, hypertension
Past, yes Male 46.4 33.1-65.1 23.2*** 77.6 71.4-83.8 7.3* 77.6 69.8-85.3 3.3
Never, no 23.2 17.4-31.0 84.9 80.3-89.5 80.9 75.6-86.3
Smoking, exercise
Past, low Male 33.8 23.0-49.6 10.4 75.6 71.2-80.1 8.3** 71.9 65.8-78.1 7.8*
Never, high 23.4 16.2-33.7 83.9 78.1-89.8 79.7 72.8-86.6
Smoking, nuts
Past, low Male 37.1 27.3-50.3 21.6** 78.3 73.0-83.5 7.5* 74.6 67.9-81.3 5.3
Never, high 15.5 6.7-35.9 85.8 79.8-91.7 79.9 70.7-69.2
Smoking, vegetarian
Past, no Male 32.2 19.2-54.0 15.4** 80.2 74.8-85.5 4.8 73.7 64.0-83.5 6.9
Never, yes 16.9 11.1-25.5 85.0 80.0-90.0 80.6 75.9-85.2
Hypertension, exercise
Yes, low Male 34.8 23.1-52.4 15.1* 72.8 68.1-77.6 11.9*** 70.7 64.7-76.8 6.6
No, high 19.7 12.8-30.2 84.7 79.3-90.1 77.3 68.4-86.2
Hypertension, diabetes
Yes, yes Female 23.9 14.1-40.6 3.6 73.2 68.3-78.2 14.8**** 72.5 61.8-83.2 18.1***
No, no 20.3 14.7-28.1 88.0 84.2-91.9 90.6 86.9-94.3
Hypertension, nuts
Yes, low Male 38.9 28.0-53.9 23.5** 72.6 68.1-77.2 13.3**** 72.9 67.0-78.7 7.6
No, high 15.3 6.4-36.8 85.9 80.7-91.0 80.5 72.8-88.3
Hypertension, vegetarian
Yes, no Male 40.4 29.0-56.2 22.1*** 75.4 69.4-81.5 9.6** 71.1 64.7-77.4 8.1*
No, yes 18.3 12.3-27.3 85.0 80.0-90.1 79.2 72.8-85.5
Hypertension, vegetarian
Yes, no Female 24.1 15.2-38.2 3.1 84.4 77.8-90.9 2.8 81.9 73.5-90.2 7.7*
No, yes 21.0 14.7-30.0 87.2 82.5-91.8 89.6 86.1-93.0
Hypertension, BMI§
Yes, high Male 40.4 27.1-60.3 22.9** 76.8 69.9-83.6 7.1* 76.4 69.5-83.4 1.4
No, low 17.5 10.6-28.9 83.9 80.1-87.6 77.8 71.2-84.5
Hypertension, BMI
Yes, high Female 32.4 19.3-54.5 9.2 87.2 82.1-92.3 0.4 83.3 71.1-95.5 7.7
No, low 23.2 16.8-32.1 86.8 82.4-91.2 91.0 88.1-93.9
Exercise, diabetes
Low, yes Female 29.0 18.7-45.1 6.1 77.4 72.0-82.8 12.5**** 76.5 69.3-83.7 15.5***
High, no 22.9 6.6-79.2 89.9 85.7-94.2 92.0 84.4-99.5
Exercise, vegetarian
Low, no Male 31.8 21.4-^7.1 15.6* 77.0 71.7-82.3 8.6* 73.6 68.0-79.1 4.5
High, yes 16.2 8.1-32.4 85.6 77.7-93.6 78.1 68.4-87.8
Diabetes, nuts
Yes, low Female 30.5 18.8-^9.4 12.9 75.9 70.9-80.9 14.8*** 76.7 69.6-83.9 11.2
No, high 17.6 6.6-47.0 90.7 83.0-98.5 87.9 74.8-101.1
Vegetarian, BMI
No, high Female 17.3 5.9-51.1 -4.7 83.8 76.5-91.2 2.4 75.8 67.6-84.0 14.2***
Yes, low 22.0 15.9-30.2 86.2 82.3-90.1 90.0 87.0-93.0
• p s 0 . 1 0 ; " p < 0.05; *** p s 0.01; * * * * p < 0.001. f Pairs of exposure variables were selected only when there were at least 20 CHD
events and 20 complementary events, t Cl, confidence interval. § BMI, body mass index (kg/m2).

The Adventist Health Study cohort of 34,198 sub-
jects followed for 6 years has some advantages for
these types of analyses. Subjects entered the cohort
with no upper age limit, with the oldest subject being
105 years and with 309 other subjects aged at least 90
years at entry. Thus, we were able to make reasonable
estimates of conditional probabilities of disease right
through to the end of the life cycle. Other studies have
often not had this advantage (19-21). A problem with
most longitudinal studies (the Adventist Health Study
included) is the lack of repeated observations on risk
factor values during significant periods of the life-span
of each individual. We, for instance, evaluate condi-
tional probabilities of coronary disease in hyperten-
sives at all ages after the starting age. Hence, our
analyses test the effect of having hypertension contin-

Am J Epidemiol Vol. 142, No. 7, 1995

754 Fraser et al.

uously after this starting age, compared with its con- and only weakly negatively associated with the rela-
tinued absence. While this is of interest, the abnormal tive risk for coronary disease. The results hold when
risk factor value may develop only part way through qd(k) and qc(k) take values similar to those that we
the life-span. Thus, such results reflect the extreme, observed. It does assume a constant relative risk across
but not uncommon, circumstance of exposure to these all ages. Hence, our speculation in the introduction
conditions throughout adult life. An advantage in our that the effects of the exposure on competing causes
data was the general focus on habits (e.g., diet, exer- will also be important is borne out here.
cise) or obesity that typically remain fairly constant Table 6 shows the non-parametric (Mantel-
(22-24), at least in a relative sense, throughout long Haenszel) relative risks of both coronary disease and
periods. Results where the starting age for the life competing causes, for both sexes combined (when
tables was 60 years are generally very similar to those
starting at age 30 years, except for a modest reduction
in the magnitude of the differences. TABLE 6. Non-parametric relative risk estimates and 95%
confidence intervals (Cl) for coronary heart disease and
We also remind the reader that the current life table competing causes, for both sexes combined: California
approach to analysis requires the construction of a Seventh-day Adventists, 1977-1982
synthetic cohort. That is, it does not represent the Coronary heart disease Competing causes
actual life experience of any existing population, Exposure variable
Relative Relative
rather, the conditional probabilities found for persons risk
95% Cl
risk
95% Cl
at a particular age during the study are assumed to
Smoking
apply during that same year of age of the synthetic life Past 1.66 1.32-2.09 1.20 1.10-1.32
table population as it ages. There is no account taken Never 1.00"* 1.00"*
of probable secular trends in the conditional probabil- Body mass index
ities. Nevertheless, this approach does form a rational (BMI)t, *
High - 1.06 0.97-1.16
basis for comparison of the effects of different -
Medium 0.91 0.83-0.99
exposures. Low - 1.00***
The relation between risk ratios and life expectancy BMI, malest, X
is known to be complex. Keyfitz (25), Vaupel (26), High 2.02 1.33-3.08 -
Medium 1.47 0.97-2.22 -
and Tsai et al. (27) have shown that a fixed percentage
Low 1.00'" -
change in mortality that applies at all ages is related to BMI, femalest, t
the percent change in life expectancy by a multiplying High 1.53 1.09-2.14 -
factor, where this factor is a complex measure of the Medium 0.97 0.68-1.39 -
way that the underlying mortality changes with age in Low 1.00' -
the population. Thus, a different pattern of mortality Exercise
Low 1.00'" 1.00***
with age in some population may alter the way that a High 0.59 0.47-0.74 0.77 0.71-0.83
constant relative risk will affect life expectancy. These Diabetes mellitus
considerations do help us understand how different Yes 2.49 1.90-3.26 1.73 1.56-1.93
underlying patterns of change in mortality with age No 1.00"* 1.00"*
may produce different effects on life expectancy, with Hypertension
Yes 2.16 1.78-2.62 1.30 1.21-1.40
the same constant risk ratios. No 1.00*" 1.00"*
It is of interest to note in our data that, while in most Vegetarianf
cases a relative risk for coronary disease greater than No - 1.21 1.12-1.31
Yes - 1.00"*
one is associated with both a higher lifetime risk and
Vegetarian,
a lower expected age at onset, this is not always so. malest
Low nut consumption in females, non-vegetarianism No 1.78 1.28-2.48 -
in males, and high body mass index in males are all Yes 1.00*" -
apparently associated with a substantial increase in Vegetarian,
lifetime risk but little or no effect on age at onset. femalest
No 0.98 0.71-1.36 -
From theoretical considerations and confirmed by nu- Yes 1.00 -
merical exploration, it can be shown that 1) lifetime Nut consumption
risk of coronary disease is strongly positively associ- Low 1.00*** 1.00*"
ated with relative risk for the disease and strongly High 0.45 0.35-0.59 0.77 0.71-0.84
negatively associated with relative risk for competing • p < 0.05; " p < 0.01; * " p < 0.001. t Significant effect
modification by gender for the coronary disease endpoint. Results
causes, and 2) age at onset is strongly negatively are given separately for males and females, t For body mass index
associated with the relative risk for competing causes (kg/m2), p values are an overall test of significance.

Am J Epidemiol Vol. 142, No. 7, 1995


appropriate), according to exposure levels. Compari-
son of these relative risks with the survival outcomes
of tables 2-4 generally shows the expected correspon-
dence. As commented earlier, the survival outcomes
are influenced by effects of the exposure on both
coronary disease and competing causes. In most in-
stances, the relative risk effects (table 6) are greater for
coronary disease than the competing cause. Adding to
this already complex scenario, the Mantel-Haenszel
relative risks assume uniform effects among the four
age strata, whereas in the survival calculations, no
constraints were placed on the conditional probabili-
ties, which could change from year to year, or in a
more general sense may tend to be quite different in
older compared with younger persons. Such differ-
ences will somewhat obscure detailed comparisons
with the non-parametric relative risks in this table.
A few previous studies (19-21) have performed
analyses that have some similarity to those we present
here. Three reports have used age-specific Framing-
ham Study mortality and coronary disease incidence
data in life table analyses. These studies have evalu-
ated the effects of one or more of the traditional risk
factors (weight, blood pressure, blood cholesterol, and
cigarette smoking) or a lipid lowering diet. Typically,
the estimates of effect were rather smaller than those
that we report. Tsevat et al. (21) looked only at
changes in life expectancy predicted to result from
changes in the above risk factors. Grover et al. (19)
evaluated the predicted effects of treating hyperlipid-
emia on both life expectancy and life expectancy free
of coronary disease. The effect of a risk factor on life
expectancy would often be less than its impact on a
coronary disease survival endpoint. Taylor et al. (20)
focused on serum cholesterol reduction, but also pre-
dicted the effects of smoking cessation and blood
pressure reduction on life expectancy.
Direct comparisons between our data and these re-
ports are difficult for a number of reasons. These
include different definitions of exposure variables, the
use of extrapolation in other studies for the older ages,
mixing of data from two populations in some reports,
and the fact that our analyses assume continual expo-
sure and are non-parametric in nature. Our non-
parametric analyses suffer from reduced power com-
pared with a multivariate modeling approach, but are
not reliant on goodness-of-fit to such a model. Such
univariate results do not account for confounding, but
should still give an unbiased prediction of the experi-
ence of a group defined only by the presence or
absence of the exposure.
Can these results from a modestly lower risk
Seventh-day Adventist population (16) be applied to
the general population? We have reported (9) relative
Am J Epidemiol Vol. 142, No. 7, 1995
Risk Factors and Survival Free of Coronary Disease 755
risks for coronary disease from this same population
that are very similar to those found in the Framingham
Study and many other studies. However, as discussed
above, if the underlying mortality patterns with age are
different from those of the general population, the
changes that we predict in survival may differ some-
what from those predicted for the general population.
Nevertheless, numerical exploration over a moderate
range of conditional probabilities in unexposed sub-
jects, and keeping relative risk levels for both coronary
disease and competing events in the exposed constant,
suggests that although the absolute levels of the cor-
onary disease survival endpoints may change moder-
ately, the difference in values with and without expo-
sure (i.e., the effect) changes very little. Hence, our
results may be approximately applicable to non-
Adventist populations.
In summary, we have demonstrated that subgroups
of California Seventh-day Adventists characterized by
different values of coronary disease risk factors show
important differences in predicted lifetime risk of the
disease, age at onset, and life expectancy free of cor-
onary disease. These associations were generally con-
sistent for males and females. If these risk factors are
causal, there are important public and personal health
implications.
ACKNOWLEDGMENTS
This work was supported by grant no. 1 R01 AG08961-
01A2 from the National Institute of Aging.
The authors gratefully acknowledge the consultation of
Dr. Dennis Tolley and Eric Stallard. They also thank Cathy
Provencio for technical assistance with the manuscript.
REFERENCES
1. Phillips RL, Kuzma JW, Beeson WL, et al. Influence of
selection versus lifestyle on risk of fatal cancer and cardio-
vascular disease among Seventh-day Adventists. Am J Epide-
miol 1980;! 12:296-314.
2. Logue EE, Wing S. Lifetable methods for detecting age-risk
factor interactions in long-term follow-up studies. J Chronic
Dis 1986;39:707-17.
3. Benichou J, Gail MH. Estimates of absolute cause-specific
risk in cohort studies. Biometrics 1990;46:813-26.
4. Chiang CL. Introduction to stochastic processes in biostatis-
tics. New York: Wiley, 1968.
5. Cutler SJ, Ederer F. Maximum utilization of the life table
method in analyzing survival. J Chronic Dis 1958;8:699-712.
6. Elveback L. Estimation of survivorship in chronic disease: the
"actuarial" method. J Am Stat Assoc 1958;53:420-40.
7. Gray RJ. A class of ^-sample tests for comparing the cumulative
incidence of a competing risk. Ann Stat 1988;16:1141-54.
8. Matthews DE. Likelihood-based confidence intervals for
functions of many parameters. Biometrika 1988;75:139-44.
9. Fraser GE, Strahan TM, Sabate J, et al. Effects of traditional
coronary risk factors on rates of incident coronary events in a
756 Fraser et al.

low risk population. Circulation 1992;86:406-13.
10. Fraser GE, Sabate J, Beeson LW, et al. A possible protective
effect of nut consumption on risk of coronary heart disease.
Arch Intern Med 1992;152:1416-24.
11. Beeson WL, Mills PK, Phillips RL, et al. Chronic disease
among Seventh-day Adventists, a low risk group. Cancer
1989;64:57-81.
12. Gillum RF, Fortman SP, Prineas RJ, et al. International diag-
nostic criteria for acute myocardial infarction and stroke. Am
Heart J 1984;108:150-8.
13. Hoem JM. A flaw in actuarial exposed-to-risk theory. Scand
Actuarial J 1984;3:187-94.
14. Manton KC, Stallard E, ed. Chronic disease modelling. New
York: Oxford University Press, 1988: Chap 2.
15. Crouse JR. Gender, lipoproteins, diet, and cardiovascular risk.
Lancet 1989;333:318-20.
16. Fraser GE, Sabate J, Beeson WL. The application of results of
some studies of California Seventh-day Adventists to the general
population. (Letter). Arch Intern Med 1993; 153:533-4.
17. Knuiman MW, Welborn TA, Whittall DE. An analysis of excess
mortality rates for persons with non-insulin-dependent diabetes
mellitus in Western Australia using the Cox proportional hazards
regression model. Am J Epidemiol 1992;135:638-48.
18. Prineas RJ, Kushi LH, Folsom AR, et al. Walnuts and serum
lipids. (Letter). N Engl J Med 1993;329:359.
19. Grover SA, Abrahamowicz M, Brewer C, et al. The benefits of
treating hyperlipidemia to prevent coronary heart disease:
estimating the changes in life expectancy and morbidity.
JAMA 1992;267:816-22.
20. Taylor WC, Pass TM, Shepard DS, et al. Cholesterol reduction
and life expectancy: a model incorporating multiple risk fac-
tors. Ann Intern Med 1987;106:605-14.
21. Tsevat J, Weinstein MC, Williams LW, et al. Expected gains
in life expectancy from various coronary heart disease risk
factor modifications. Circulation 1991;83:1194-201.
22. Lindsted KD, Kuzma JW. Long-term (24-year) recall reliabil-
ity in cancer cases and controls using a 21-item food fre-
quency questionnaire. Nutr Cancer 1989;12:135—9.
23. Gordon T, Doyle JT. Weight and mortality in men: the Albany
Study. Int J Epidemiol 1988; 17:77-81.
24. Gordon T, Shurtleff D. Means at each examination and inter-
examination variation of specified characteristics. Framing-
ham Study. Exam 1 to Exam 10. (DHEW publication no.
(NIH) 74-478). Washington, DC: US GPO, 1974.
25. Keyfitz N, ed. Applied mathematical demography. New York:
Wiley, 1977.
26. Vaupel JW. How change in age-specific mortality affects life
expectancy. Population Studies 1986;40:147-57.
27. Tsai SP, Hardy RJ, Wen CP. The standardized mortality ratio
and life expectancy. Am J Epidemiol 1992; 135:824-31.
28. Fienberg SE, Neter J, Leitch RA. Estimating the total over-
statement error in accounting populations. J Am Stat Assoc
1977;72:295-302.

APPENDIX

Variances for the three survival outcomes were estimated by first order Taylor's series approximations. Two
of the outcome measures ld(y) and ed(y) were considered functions of both the qd(k) and qc{k), k = y, . .. , w , .
Recall that qd(w), qc(w), and Uw are also functions of qd(k) and qc{k), k = w0 — 5, . . . , w,. For notational
simplicity, consider F(-) to be a general function of the q's.
Hence,

Var[F(<fc(y), . . .,qd(w0 ~ 1), qM, • • •> <7cK - 1))]

KM , qc(k))],
k=y

as all other covariances are assumed to be zero, where

P(k)

^ qc(k)(l - qc(k))
q<&> p(lA

- qd(k)-qc(k)
Cov(qd(k),qc(k))± —

Then, Var(F) can be readily calculated using the following sum of quadratic forms:

), qc(k)l qJ

Am J Epidemiol Vol. 142, No. 7, 1995

 Risk Factors and Survival Free of Coronary Disease 757

The first derivative of F for F = ld(y) is given by

j=k+\ ek-S(w)-NcT-P(k)
S(k)~ + ^ , and
- qd(k) - qc(k))\

j=k+i €k'S(w)NdTP(k)
F
qc(k) ~
I - qM ~ qM)\
are as
where S(J), qd{j), qc(J) defined in the text; 8^ = 1.0 if k < w0 else 8^ = 0; e^ = 1.0 if k > w0 — 5, else

Nd(k),NcT= P(k).
k=wo~ 5

The first derivative of F for F = ed(y) is given by

1 ^
" y> ~ i\ — „ <b\ — „ (v\\ -^

l ek-S(w)NcTP(k)
8k\S(k)-
- qd(k) - qc{

and

- y)Nd
2 ) [50") • qM • (t/, -y)]\- ekS(w)P(k)
1(1 - qd(k) - qA Iog2(l -NT/PT)(PT-NT)

ekS(w)NdTP(k)
- qd(k) - qc(k))
j=k+\
+ •

For the endpoint ed+c{y), calculations are somewhat simpler as ed+c(y) can be considered as F(q(y), .. .,
<7(w,), where q(k) = qd(k) + qc(k).
Then,

Am J Epidemiol Vol. 142, No. 7, 1995

758 Fraser et al.

where

S(W)-P(k)
and = St - y) ~ - (*/, - y)]
\og\\-NTIPT){PT-NTY

Note that the variances of the conditional probabilities q(J) are estimated as functions of these observed
statistics rather than the underlying binomial parameters, 6(j) = E[q(j)]. Provided that numbers are reasonably
large, this should be satisfactory. However, even in a relatively large data set such as ours, with some subgroup
analyses, zero cases were sometimes observed for a particular age group. This would predict a variance (q) =
0 if the above approach is used. Clearly, this would be an underestimate. Hence, for variance estimation only,
we grouped our observations and person-years to 5-year age groupings and used [2Nd(k)]/['ZP(k)] and [2NC(£)]/
[2/>(&)] in place of qd(Jk) and qc(k) in the expressions above, where N is the number of events, T(k) represents
persons at risk during the year of age k, and summation is over the 5 years.
In the rare instance that there were no events observed for a whole 5-year age range (over the age of 50 years),
we elected to use a conservative estimate of 6d(k) or 6c(k). A value of, say, dd(k) was chosen such that there
would have been a probability of only 0.1 of observing a value as low as zero for all five consecutive years given
the observations for the c endpoint, under a binomial distribution. A similar approach has previously been
suggested for an accounting application (28). dd(-) were assumed to be constant during the 5 years, and
5

are labeled simply 6 in equations below. Therefore, U [Pr(OI 8, P(k))] = 0.10, where the values of k represent the
five individual ages within the 5-year range, i.e.,

o^M 1 - e)p{k)] = (i - = o.io.

k=\

Then 0 = 1 — (0.10)I/>w, but not to exceed 0.5, as the value of 0.5 maximizes the variance, estimated as
0(1 -8) .
Yax(q(k)) — , ,— for each k in the 5-year interval. This approach was used to estimate both dd(k) and Bc(k)
r\K),
in the few occasions where it is appropriate.

Am J Epidemiol Vol. 142, No. 7, 1995