Comments on problems with small effects in case-control and cohort studies

Abstract

In his contribution D. Kromhout discussed a phenomenon to which we may return repeatedly in the following, namely regression dilution or, as we used to call it in the old days, regression towards the mean. I met this as an analytical problem in 1967, when J. Cornfield and W. Haenszel gave a course in cancer epidemiology in Oslo. Among their exercises was one on the relationship between serum cholesterol and coronary heart disease. The simplest way to adjust, if you have a linear relationship and can approximate the intraindividual variation, is to turn the slope. With our Norwegian data from around 1960 we ended up with roughly a 40% increase in the relative risks. However, it is not easy to measure such an increase directly. It is more difficult today than in the past. D. Kromhout spoke of taking the average of 11 consecutive measurements. In Norway in the seventies and eighties we have been conducting very large cardiovascular surveys, in which 22,000 men and 22,000 women were examined twice at an interval of 3–5 years, with a 3–9 year mortality follow-up after the second screening. The improvement in the slope by using the mean of the two predictor measurements, rather than a single one, turned out to be much less than prescribed in theory. With such a long interval between the measurements they no longer represent random samples from a constant level — quite apart from the attention given by clinicians to the top measurements, which may well distort any risk-curve more seriously than was the case 30 years ago.