Good Calories, Bad Calories

In 1984, when Keys and his colleagues published their report on the data after fifteen years of observation, they explained that “little attention was given to longevity or total mortality” in their initial results, even though what we really want to know is whether or not we will live longer if we change our diets. “The ultimate interest being prevention,” they wrote, “it seemed reasonable to suppose that measures controlling coronary risk factors would improve the outlook for longevity as well as for heart attacks, at least in the population of middle-aged men in the United States for whom [coronary heart disease] is the outstanding cause of premature death.” Now, however, with “the large number of deaths accumulated over the years,” they realized that coronary heart disease accounted for less than one-third of all deaths, and so this “forced attention to total mortality.”

Now the story changed: High cholesterol did not predict increased mortality, despite its association with a greater rate of heart disease. Saturated fat in the diet ceased to be a factor as well. The U.S. railroad workers, for instance, had a death rate from all causes lower—and so a life expectancy longer—than the Finns, the Italians, the Yugoslavs, the Dutch, and particularly the Japanese, who ate copious carbohydrates, fruits, vegetables, and fish. Only the villagers of Crete and Corfu could still expect to live significantly longer than the U.S. railroad workers. Though this could be explained by other factors, it still implied that telling Americans to eat like the Japanese might not be the best advice. This was why Keys had begun advocating Mediterranean diets, though evidence that the Mediterranean diet was beneficial was derived only from the villagers of Crete and Corfu in Keys’s study, and not from those who lived on the Mediterranean coast of Yugoslavia or in the cities of Italy.

In discussions of dietary fat and heart disease, it is often forgotten that the epidemiologic tools used to link heart disease to diet were relatively new and had never been successfully put to use previously in this kind of challenge. The science of epidemiology evolved to make sense of infectious diseases, not common chronic diseases like heart disease. Though the tools of epidemiology—comparisons of populations with and without the disease—had proved effective in establishing that a disease such as cholera is caused by the presence of micro-organisms in contaminated water, as the British physician John Snow demonstrated in 1854, it is a much more complicated endeavor to employ those same tools to elucidate the subtler causes of chronic disease. They can certainly contribute to the case against the most conspicuous determinants of noninfectious diseases—that cigarettes cause lung cancer, for example. But lung cancer was an extremely rare disease before cigarettes became widespread, and smokers are thirty times as likely to get it as nonsmokers. When it comes to establishing that someone who eats copious fat might be twice as likely to be afflicted with heart disease—a very common disorder—as someone who eats little dietary fat, the tools were of untested value.

The investigators attempting these studies were constructing the relevant scientific methodology as they went along. Most were physicians untrained to pursue scientific research. Nonetheless, they decided they could reliably establish the cause of chronic disease by accumulating diet and disease data in entire populations, and then using statistical analyses to determine cause and effect. Such an approach “seems to furnish information about causes,” wrote the Johns Hopkins University biologist Raymond Pearl in his introductory statistics textbook in 1940, but it fails, he said, to do so.

“A common feature of epidemiological data is that they are almost certain to be biased, of doubtful quality, or incomplete (and sometimes all three),” explained the epidemiologist John Bailar in The New England Journal of Medicine in 1980. “Problems do not disappear even if one has flawless data, since the statistical associations in almost any nontrivial set of observations are subject to many interpretations. This ambiguity exists because of the difficulty of sorting out causes, effects, concomitant variables, and random fluctuations when the causes are multiple or diffuse, the exposure levels low, irregular, or hard to measure, and the relevant biologic mechanisms poorly understood. Even when the data are generally accepted as accurate, there is much room for individual judgment, and the considered conclusions of the investigators on these matters determine what they will label ‘cause’…”