Participants were 7744 men (20-89 yr) initially free of CVD who returned a mail-back survey during 1982. Time spent watching TV and time spent riding in a car were reported. Mortality data were ascertained through the National Death Index until December 31, 2003.
They noted that they adjusted for age. I presume other factors like the economic status correlate also with the time they watched tv. Don't know if they've adjusted for that as well.
In other words, if you went back and looked at your data and found the best cutoffs to maximize difference in risk, is it any surprise that you found a big difference in risk? This is why they need to validate these cutpoints in a separate study.
Your thinking could stem from an affliction brought on by high schools teaching of science. They really only teach one of the scientific methods, that of hypothesis, experimentation, observation and conclusion in that order. However depending on the field of science and the subject at hand that isn't necessarily the order thing are done in. Science isn't a simple linear process like we are lead to believe in high school.
There is nothing wrong with "my thinking" but you appear to think so because your understanding is incorrect. Post hoc ergo propter hoc is the fallacy of thinking that, because something comes before another, that first something causes the second. I'm not talking about that at all. Post hoc analysis is something totally different - viz, it's performing non-prespecified analyses. It's more closely coupled to multiple testing than it is to the other fallacy you are referring to. It certainly is not limited to small sample sizes or sample sizes of one, in contrast to what you state. You can do science improperly with samples of any size.
But in this case the general hypothesis is pre-chosen and fixed, basically "men who watch TV more hours [whatever] more than men who watch TV fewer hours", and the choice of threshholds is more of a reporting step, pulling out some binned tails at either extreme for emphasis and ease of discussion, while the real correlation is full-curve-to-full-curve.
And I absolutely agree that a sort of nonparametric post hoc approach that you poo-pooed would be even more highly suspect. But I don't see the choice of thresholds as simply a reporting step that can be manipulated ad lib without adjustment of your belief in the association and magnitude of effect. (I don't mean to imply that you think that you can adjust cutoffs willy-nilly - just stating my opinion clearly here.)
Either they controlled for occupation (doubtful) or sitting in an office didn't matter or the thing whole is just off-base.
One way that sitting in an office is different from watching TV involves complete passivity whereas in an office setting, your actively engaged in some task.
But driving a car is probably closer to sitting in an office than watching TV. So the whole things seem problematic.