Another social scientist (sociology) chiming in - the concept of an "ecological fallacy" might be useful here.
Here's a short definition: "False conclusions made by assuming relationships found through research with groups can be attributed to individuals."
And a longer one: "The ecological fallacy is the logical fallacy of interpreting general data too particularly or minutely. An example would be projecting to the level of individuals the generalizations that apply to a population. This fallacy, and the opposite fallacy of generalizing from the particular, have been responsible for some misguided health policies. For example, many epidemiological studies have demonstrated an increased risk of heart disease associated with high-fat diets, cigarette smoking, and lack of exercise; but not everyone who exhibits these behaviors necessarily dies of a heart attack—and it is a mistake to blame such people if they experience a heart attack because many other factors could precipitate such an event."
Most of us can easily observe a correlation between eating habits and exercise for ourselves - if I eat more and don't exercise, I gain weight. So it's easy for me to assume that people who weigh less than me must be eating much less or exercising a lot more, and that people who weigh more than me must be eating much more and exercising less. This would be overgeneralizing from my own experience.
Based on what everyone has said (I skimmed through the beginning of the book, but didn't read this chapter carefully) the research that Kolata is summarizing seems to say that on average, people who are overweight are not eating much more or don't have different eating behaviors than those who are not overweight. But there's a lot of individual variation contained within those overarching patterns - to assume that it applies to all individuals would another kind of overgeneralization. The anecdotes that this thread has generated - thin people who eat a lot and don't gain weight, thin people who carefully monitor their weight and exercise, overweight people who eat alot, and overweight people who eat moderately, etc. etc., - are all examples of that individual variation, which could easily be contained within the overarching pattern.
Just as statistical patterns don't apply to all individuals, anecdotes don't trump statistical patterns. But it also sounds as though Kolata isn't defining what terms like "overweight" very carefully, and as though she herself may be committing an ecological fallacy, by assuming that the statistical patterns can be interpreted as true for individuals.
I didn't read very far into the book because I wasn't very impressed by her approach or tone. I've enjoyed her articles in the New York Times, but the book seemed much less thoughtful and nuanced than I would have expected from a science journalist of her stature.