This analysis shows that some classic risk factor baseline levels had a relatively constant relation with the occurrence of CHD deaths during a 35 year follow up. This is true for age, systolic blood pressure, serum cholesterol, and cigarette smoking directly and for physical activity inversely. Body mass index was never significantly related with the outcome and its time relation with an event but rather fluctuated around the level of no relation.
The curves are to be interpreted as the description of the strength and possible stability of the relation of risk factors with the occurrence of events at each subsequent time interval, although in this particular analysis, the interval is relatively long—five years. Among other things, this analysis tests the proportionality hazards assumption of the Cox model.
There were some irregularities in the curves describing the relation between risk factors and events during the follow up period, but they are partly justified by the small number of events in each five year block. On the other hand, the straight line fitted on data have large correlation coefficients, giving strength to the relatively monotonic and regular relation between risk factor levels and CHD deaths. Some relations are not linear along the whole time period. This can be interpreted as a non-linear relation or as the consequence of selection of cases due to the influence of risk factors potentially important also for non-coronary conditions.
There are technical alternatives to the partitioned Cox model, such as the Aalen model,8
which is based on a linear function and evaluates each single event, one by one, during the follow up and then uses a smoothing technique to cumulate the hazard functions. It can be shown that for relatively small effects, such as those of few events every five years, the two approaches produce similar results.
The strength of the association of baseline measurements with early and late events seems to be reinforced when changes of risk factor levels are added to the predictive models. However, the shapes of the plotted curves are unchanged, although their steepness is greater at least for systolic blood pressure and serum cholesterol, and for the first 15 years of follow up when data were reliable for the purpose.
The use of baseline measurements only in longitudinal studies tends to underestimate the associations between usual risk factors and CHD rates because of regression dilution effects.10–12
In fact baseline measurements are often observed to be of lesser value in predicting long term events than they actually are, and adjustment for regression dilution can give a better estimate of their associations. The regression dilution effects can be due to random error, short term variation, and long term change. Our analysis with the inclusion of risk factor changes should be considered in the broader context of the regression dilution phenomenon, which can be used to improve the predictive ability of events after long periods of time.
In general, a single measurement in middle age retains its association with events and predicts events with similar strength for a long period of time. The strength of the association is somewhat improved by taking into account possible changes in risk factor levels due to aging or other reasons. The additional information of changes in risk factor levels has been described elsewhere for the first 25 years of follow up of this study2
and there is no contradiction with these findings.
The problem of predicting early versus late events has been addressed in other studies, although using different techniques.
As mentioned above, by using the same data for a shorter follow up period, we have shown that the predictive power of coronary risk factors was similar for CHD incident events during the first 10 years and the subsequent 15 years of follow up.2
An analysis of the Honolulu heart study of middle aged men segregated early events from late events on the basis of age at the time of the event (< 60 years and 60 or older).13
Systolic blood pressure, alcohol intake, serum cholesterol, serum glucose, and cigarette smoking were independent predictors for early and late onset coronary events. On the other hand cigarette smoking had a smaller effect on late events, whereas serum triglycerides were predictors only of early events.
In the 25 year follow up of the Finnish cohorts of the seven countries study, the association of baseline cholesterol with coronary mortality was similarly strong during the first 10 years and the subsequent 15 years of follow up.14
The extension of the follow up to 30 years, with an analysis of three decades in the same cohorts, suggested that baseline serum cholesterol and smoking habits were strong predictors of CHD death occurring either early or late during the follow up, while baseline systolic blood pressure was no longer associated with CHD deaths after 20 years. Only after that time did baseline body mass index became directly associated with events.15
In our data the coefficients for systolic blood pressure become much smaller, and on two occasions non-significant, after 20 years of follow up.
In the Whitehall study in the UK, it was shown that, for a given age at death, the longer the gap between cholesterol measurement and death the more predictive the cholesterol concentration for CHD and all cause mortality, confirming that a single measurement taken earlier in life still holds predictive value after many years.16
In the Chicago Heart Association detection project in industry,17
late follow up events were predicted with the same strength as early follow up events by baseline measurements of blood pressure, cholesterol, and smoking habits.
Our findings and those reported in the literature help in interpreting some misunderstanding about the predictive power of risk factors in the elderly. The association of serum cholesterol concentrations with CHD events has been long debated and its interpretation is uncertain.18
The same applies to blood pressure with contrasting evidence about its ability to predict events when measured in old age.19–21
Our findings clearly raise two entirely different questions: firstly, whether a single measurement in adulthood can predict events far in the future and even in to very old age; and secondly, whether measurements taken in the elderly can still predict the occurrence of events. It seems now that the first question can be answered by stating that some risk factors are associated for a very long time even with very distant events that occur in advanced ages. For example, in our sample, CHD deaths recorded during the last five years of follow up were strongly predicted among men aged 70–94 years whose risk factors were measured when they were aged 40–59 years.
The interpretation of this phenomenon is not clear. Personal characteristics in middle age may change over time, but at that age they are probable indicators of the degree of damage already done and maintain their strength and predictive value for long time. Again, this does not negate the possible benefits of changing levels of modifiable risk factors but is a warning that the association of past levels with the occurrence of events may last for life.