That's what I get for running my mouth without looking at the paper. Skimming it now, the "dancing [was] associated with a reduced risk of dementia" result smells like multiple testing throwing up spurious crap. There's a dementia hazard ratio listed for 9 physical activities, and the ratio for dancing was the only one that was "significant" — but it's only barely significant as the confidence interval goes from 0.06 to 0.99! And the hazard ratios for the other physical activities are all over the place, half of them less than 1.00, half of them at least 1.00. There doesn't seem to be any real evidence of an association between exercise & dementia, taking this paper as a whole.