How much is too much exercise?

And how do we know?
You'd had to have been buried under a rock to have missed the headlines a few weeks back: "Excessive Jogging will Kill You"! etc. All of them traced back to this article on the Copenhagen City Heart Study. When I looked at the article, the purported U shaped risk curve looked to me like anything but -- for every group past the first two, the confidence limits included 1, the "no effect" level, and usually also included the average risk of the level below. I get it. A non-linear relationship is just so much cooler than a linear one. I should know, I keep pushing graduate students to find them. But, please. Confidence limits interpreted correctly please! Please! Pretty please?

But it turns out there are more articles on the badness of too much exercise! Vox summarized quite a few, which turn out to have mixed outcomes, with some articles saying yes and some no to the increased risk with increased exercise.

Then, only a day or two after the Copenhagen study splashed across the interwebs, this study on women in the UK came out. Here's their first figure:

So the x-axis is frequency of physical activity, and the left column is the category of "strenuous exercise", while the right hand column is "any physical activity". Strenuous activity is "any work or exercise causing sweating or a fast heartbeat", whereas any physical activity included housework, walking, gardening, cycling. So there could be overlap in the responses; basically the "strenuous activity" is more restrictive. Three different endpoints, coronary heart disease, Cerebrovascular disease and Venous Thromboembolism. The confidence limits are generally much smaller than in the Copenhagen study, reflecting the much larger sample size (~100 times bigger). That's much more convincing that the curve is turning upwards. Different endpoints have different turning points, but it's there.

I would be much more convinced however, if they had kept "hours of strenuous exercise" as a continuous variable and fit a spline. Although with only a few points of support that would be pretty uninteresting. With a continuous variable they could compare a linear and quadratic model (1 and 2 parameters respectively) with this 4 parameter model. I'll ask it again, why don't epidemiologists do model selection?

Note that all these articles support one very important point. No exercise is much worse than some exercise. Even in the UK study daily strenuous exercise is equivalent to no exercise for only one endpoint. Otherwise, any level of exercise is better than none.

Oh well, off to the gym! Only 4 times a week though! And daily housework increases your risk of strokes and heart attacks!