Cardiovascular risk factors predict age at death in 60-year follow-up of the Seven Countries Study

To study age at death (AD) and its determinants in cohorts of middle-aged men followed-up until extinction. A total of 9063 middle-aged men enrolled in 10 cohorts of 6 countries (USA, Finland, the Netherlands, Italy, Greece and Japan) within the Seven Countries Study were examined and then followed up for 60 years until extinction. AD was computed and a small number of risk factors were tested through multiple linear regression as possibly related to attained AD. AD ranged across cohorts from 71.8 years in East Finland and 80.5 years in Crete with levels roughly lower in the USA and Northern Europe and higher elsewhere. Across cohorts, the correlation coefficients of systolic blood pressure (R = −0.58) and of CVD prevalence (R = −0.65) versus average AD were the only significant ones. At the individual level in the pool of all cohorts, a multiple linear regression model showed that age, vigorous physical activity, never and ex-smokers were favorably related to AD, while the reverse was true for systolic blood pressure, heart rate, serum cholesterol, CVD prevalence and silent ECG abnormalities. BMI had a parabolic relationship with AD. The predicting power of single risk factors, expressed in years gained or lost, was relatively small, but arbitrary combinations of several of them produced large differences in AD. A small number of CVD risk factors were strongly associated with AD in a life-long follow-up.


Introduction
Age at death (AD) is a metric of mortality trends that has recently been re-evaluated. It is widely used in demography to describe evolution in mortality across populations [1][2][3][4] while it is almost never used in classical field epidemiology, probably because of the relative rarity of population samples followed up until extinction. In fact, AD has little or no meaning if many subjects of a group are still alive.
We do not completely understand the molecular basis that modulates the aging process, but we have gained remarkable insights into the plasticity of life span and health span from behavioral life-style, genetics and pharmacology of longevity that may open the way toward longer age in good health. Our research group has already used these metrics having the opportunity to exploit data of cohorts close to extinction [5][6][7][8][9]. The use of AD as an end-point becomes a must when the populations are extinct since in that case death rates from all-cause are the same in all of them and no comparison is liable.
On this occasion we could use data on 10 cohorts of middle-aged men studied in the Seven Countries Study of Cardiovascular Diseases (SCS) that reached practical extinction after an exceptional 60-year follow-up. We study midlife risk factors and behaviors in relation to the distribution of AD.

Populations and measurements
The analysis was run on 10 of the 16 cohorts of the SCS that reached a follow-up for life status and mortality of 60 years. They were the US Railroad workers; two rural cohorts in Finland (East and West Finland); one sample of men from the commercial town of Zutphen-the Netherlands; two rural cohorts in Italy (Crevalcore and Montegiorgio); two rural cohorts in Greece (Crete and Corfu); one cohort from a rural village and one from a fishing village in Japan (Tanushimaru and Ushibuka). All men were aged 40 to 59 year at the entrance examination held between 1958 and 1961 and the participation rate of those invited was on average around 95%. More details can be found elsewhere [10,11].
Cardiovascular risk factors measured in all cohorts and used in this analysis were: (a) age, in years approximated to the nearest birthday; (b) physical activity at work, derived from a few questions combined with the reported occupation, and classified as sedentary, moderate or vigorous (in percent %); in one country this classification was indirectly validated by ergonometric measurements [12] and energy intake derived from dietary history [13]; (c) smoking habits derived from a standard questionnaire, and classified as never smokers, ex-smokers and current smokers (in percent %); (d) body mass index (BMI), derived from height and weight measured following the procedure of the WHO Cardiovascular Survey Methods Manual (WHO Manual), in kg/m 2 [14]; (e) systolic blood pressure measured in supine position at the end of a physical examination using mercury sphygmomanometers, following the procedure described in the WHO Manual [14] (in mmHg); the average of two measurements taken one minute apart was used for analysis; (f) heart rate, derived from a resting ECG (in beats/min); (g) serum cholesterol measured on casual blood samples following the technique of Anderson and Keys (expressed in mmol/L) [15]; (h) prevalence of cardiovascular diseases (CVD) as defined by the CVD prevalence criteria of the SCS [10] expressed as 1 = yes; 0 = no; (i) prevalence of silent major ECG abnormalities in subjects without a diagnosis of CVD, including any of the following codes of the Minnesota Code, edition 1968 [14] Life status and mortality were periodically checked for 60 years including the date of occurrence. Out of 9063 men, only 3 men were still alive and 68 lost to follow-up at defined dates and all of them were censored (0.8%).

Statistical analysis
The analysis was based on 10 cohorts but for some pieces of analysis we considered 6 countries combining the cohorts belonging to the same country. We could trust this operation since cohorts of the same country have rather similar characteristics. About 4 per 1000 baseline measurements were missing and were imputed by multivariate normal procedure.
Baseline risk factor measurements were described as means or proportions for each cohort and on total cases and compared across cohorts by ANOVA or chi-squared. Demographic data of follow-up were described for each cohort and tabulated including cases of death, alive and lost to followup (censored men) and AD in years was computed, including its percentile distribution.
Multiple linear regression models (MLR) were computed for each of the 6 countries, combining cohorts belonging to the same country, with AD as a dependent variable and 12 risk factors as independent variables (plus 2 references). The main MLR model was computed combining all cohorts with AD as a dependent variable and 12 risk factors as independent variables (plus 2 references) and adding dummy variables for the identification of cohorts (Crete used as a reference). In fact, a quadratic term of BMI was added to explore the possible parabolic relationship of BMI with AD. Findings included standardized coefficients, given by coefficient times standard deviation of the risk factor divided by the standard deviation of a dependent variable, used to provide a comparable rank of risk factors role for their predictive power. From this computation, we excluded the two terms of BMI and the dummy variables for cohorts.
As men lost to follow-up and those still alive at 60-year follow-up were 71, the analysis conducted on proper AD covered 99.2% of men enrolled at baseline while those not used were 0.8%. However, for men not recorded as dead we computed AD as the age at the time when last seen alive and censored. To evaluate possible distortion on MLR coefficients of the main MLR model due to abnormalities of baseline and follow-up data, we made the following preliminary comparisons: (1) coefficients of the main model with AD of deaths only versus coefficients of a model including also cases alive or lost to follow-up (with AD being represented by that at the time when last seen alive): no significant differences; (2) coefficients of the main model including dummy variables for the identification of cohorts versus coefficients of a model not including those dummy variables: only one risk factor with a significantly different coefficient (vigorous physical activity); (3) coefficients of the main model versus coefficients of a model excluding the Japanese cohorts (being those with the greatest losses to follow-up): no significant differences.
Cox proportional hazards models were solved with allcause mortality as the end-point and the same risk factors used in the MLR as predictors. Two versions were computed, i.e. one including and the other excluding people censored due to survival of lost to follow-up.
Using intercept and coefficients of the main MLR model a few theoretical estimates of AD were made, combining in an arbitrary way different levels of risk factors.

Results
In this analysis, we included all men enrolled in the entry examination to describe the natural history of these cohorts and, therefore, prevalent cases of major CVD were used as risk factors (and not excluded).
Mean levels of risk factors at baseline examination showed large differences across cohorts as suggested by the test of ANOVA and chi-squared test reported in Table 1. In general, adverse higher levels of some risk factors (serum cholesterol, blood pressure, CVD prevalence) were more common in North American and Northern European cohorts, compared to the others. However, potentially favorable levels of risk factors (vigorous physical activity) were common in Finland while adverse levels of smoker prevalence were common in the two Japanese cohorts.  During 60 years, 99.2% of men died but death rates in Japan were definitely lower since almost 2/3 of all lost to follow-up of the 10 cohorts were located in those two cohorts. In particular, lost to follow-up was 4.3% in Japan versus 0.3% in the pool of all other cohorts. The 43 men lost to follow-up in Japan had a mean age of 86.6 years with a range of 70.7 to 110 years. Overall, the 71 men censored because alive or lost to follow-up had an average age of 88.5 years with a range of 66.8 to 110 years. The only cohort that was fully extinct was that of Crete in Greece.
Mean AD (Table 2) was largely different across cohorts with extremes in East Finland (71.8 years) and Crete (80.5 years). In general, they were lower in North America and Northern Europe and higher in Southern Europe. Japanese results should be taken with caution due to the problem explained above. As a consequence, the difference between AD (on dead people) and that including also the age for men alive or lost to follow-up were relatively large in Japan and minimal elsewhere. Another way to describe AD is given in Table 3 with its percentile distribution that suggests similar conclusions.
Simple preliminary test was made producing a correlation matrix using the 10 cohorts as statistical units (in a kind of ecological analysis) and from this attempt only systolic blood pressure (R = −0.58) and prevalence of CVD (R = −0.66) were significantly correlated with AD.
Then, six MLR models were solved, with AD as a dependent variable and 12 risk factors (plus 2 references) predictors, that is one for each country, combining cohorts of the same country when proper (details not reported). The performance of these models was not so good with R squared ranging from 0.05 in the Netherlands to 0.13 in Finland. Only three risk factors were significant in all 6 countries i.e. never smokers, systolic blood pressure and CVD prevalence. Test of heterogeneity of coefficients of these three risk factors across countries was highly significant with p < 0.0001. Their impact on AD for each country is reported in Fig. 1.
The MLR model reported in Table 4 includes all cohorts, the 12 risk factors plus 2 references, dummy variables identifying the 10 cohorts (Crete as reference) and AD as endpoint. All risk factors except moderate physical activity carried significant coefficients. Favorable variables were age, vigorous physical activity, never and ex-smokers, while adverse variables were systolic blood pressure, heart rate, serum cholesterol, CVD prevalence and ECG abnormalities. BMI had a significant parabolic relationship with AD and its shape is visually given in Fig. 2. The highest AD was associated, everything being equal, with about 27 units of BMI.
The R-squared of MLR was 0.11. The rank of predictive power defined by the standardized coefficients was led by systolic blood pressure, never smokers and CVD prevalence, incidentally the same three factors that were significantly predictive in the MRL models run separately in the 6 countries.
In the MLR model, the coefficient corresponds to the years of AD gained or lost (depending on its algebraic sign). For discrete variables, the indication is immediate, while for continuous variables it may be necessary to multiply the coefficient by the difference in risk factor levels for which the estimate is requested. Systolic blood pressure was the top-ranking risk factor following the standardized coefficients and therefore its relationship with AD was visualized in Fig. 3. One of the curves was based on crude data from the pool of all cohorts, while the other one was derived from the multivariate-adjusted coefficient of systolic blood pressure. There is an impressive regular decline of AD as a function of increasing systolic blood pressure levels, although the adjusted line is a little less steep than the other one.
In parallel, Cox proportional hazards models were solved with all-cause mortality as end-point and the same risk factors used in the MLR models. Two models were computed, Table 4 Multiple Linear Regression model with age at death as end-point, 12 risk factors (plus 2 references) as predictors and 9 dummy variables of areas as confounders (plus 1 reference)

CI confidence intervals
Delta of continuous variables for computation of change in age at death roughly corresponding to 1 standard deviation. In this MLR model, the coefficient corresponds to the years of age at death gained or lost (depending on its algebraic sign) for a unit of measurement of each risk factor   Table 5 where it appears that the significant risk factors were the same as Fig. 3 Age at death as a function of increasing levels of systolic blood pressure. Estimate 1 was made using crude data from all cohorts. Estimate 2 was made using the multivariate coefficient of systolic blood pressure from the MLR model  -129 130-139 140-149 150-159 160-169 170-179 180-189 190-199   found in the MLR model, but carrying an inverse algebraic sign. Theoretical estimates of AD as a function of risk factor levels using the coefficients of the MLR model offer almost infinite combinations to test. Three examples reported in Table 6 provide different situations, including the role of mean risk factors levels (valid for a group), very favorable and very adverse risk factors levels (valid for single individuals) showing large differences of AD up to 20 years or more. These individual estimates were based on the experience of the cohort of Crete and therefore to obtain an estimate valid for other locations the outcome should be added algebraically with the coefficient of the chosen cohort.

Discussion
In this analysis, we could use only a small set of risk factors, all of CVD type, since they were the only ones available from all cohorts. Some caution should be taken in the interpretation of the Japanese mortality data due to the large loss of men during the follow-up. We are more confident in the multivariate MLR model as a consequence of the comparisons across different models where the role of the Japanese problem is likely diluted into the larger contribution of other cohorts. This analysis has shown that a few CVD risk factors, including prevalent CVD, measured in middle-aged men were strongly associated with AD that corresponds to the extinction of the original cohorts in a 60-year follow-up. In general, the predicting power of single risk factors, expressed in years gained or lost, is relatively small, but arbitrary combinations of several of them may result in large differences in AD.
Findings suggest that CVD risk factors might have influenced CVD mortality but some of them have probably a multiple role being related also to other fatal conditions. Among them, smoking habits and levels of physical activity are the obvious candidates but also blood pressure and BMI (with its parabolic relationship) probably play a strong role.
In another analysis of the SCS we used a larger set of cohorts (13 instead of 10), a shorter follow-up period (45 instead of 60 years), a minimal set of risk factors (4 instead of 12), focusing on AD of specific CVD fatal endpoint and showing the difference in AD for specific CVD mortality groups and the major role of those four CVD risk factors. [7]. Instead, in an analysis run only on the Italian Areas of the SCS, 32 risk factors could be tested and 21 of them proved to be significantly related to AD for all causes, suggesting that the few CVD risk factors used in this analysis cover only a part of the possible determinants [8] which is anyhow demonstrated also here by the relatively restricted size of R-squared of the MLR. Findings provided by the Cox model with all-cause deaths as an end-point offered the same conclusion as the MLR model but the latter has the definite advantage to express the coefficients as years gained or lost as a function of the correspondent risk factors which is not the case for the Cox models.
Our main interest was that to relate baseline risk factor measurements with the end-point represented by AD and the documented relationships were still strong. However, we acknowledge the fact that during the long follow-up levels of some risk factor might have changed possibly influencing the outcome. In this material, long-term data on risk factor changes were available for about 30-35 years only in the European cohorts (except Corfu) where we found a generalized physiological reduction of smoker prevalence while different trends were observed for other major risk factors. A risk factor change score [16] involving cigarette smoking, serum cholesterol and systolic blood pressure declined during 30-35 years in Finland and the Netherlands, slightly increased in Italy and strongly increased in Crete-Greece with a significant association with acceleration or deceleration of CHD death rates expressed by the Weibull distribution shape. However, these were partial data limited to a subgroup of cohorts and could not be used in this analysis still probably being responsible for some aspects of the final outcome.
Specific literature on AD based on classic epidemiological cohort study is practically not existing. On the other hand, there were contributions that tested the predictive power of CVD risk factors versus all-cause mortality, but in the majority of cases the follow-up was temporarily limited and never reached the stage of cohort extinction [17][18][19][20][21][22][23].