Hello, and thank you for the comment. What I am looking for is some kind of formal way to make that decision. How do I measure whether or not the form of my model fits well enough? Do i just decide based on measures of fit like r-squared, or are there any bootstrap specific ways to make that decision? The case based results differ enormously from the residual based ones

Now that is just willfully misinterpreting of what I said, isn't it? I clearly wrote that the sample size can only be correctly estimated if you know these kind of informations. Judging from his question I assume that he doesn't know these. I never wrote that 30 is fine for everything, just that it is quite common to see this kind of sample sizes for studies that seem similar with the one he is planning.

Oh, that's quite embarrassing. On my first quick glance , I thought the author would use the term autocorrelation to describe the time-lagged models featured in that link, so I assumed he is using the term differently than I do. but he obviously doesn't do that. I have to edit my answer in that regard

I know the common interpretation of autocorrelation. But if you look into the first link I provided, some econometricians seem to use the term autocorrelation to refer to something that looks more like common regression, with some time-lagged terms. That's why I added the sidenote, because the method in the link is called autocorrelation and doesn't fit the normal definition of autocorrelation.

That is a bit like objecting the common 5% threshold for statistical significance just because there is no statistical reason for it.... the op seems to be asking for practical advice and I am giving it based on common samples sizes I observed in similar studies as the one he is describing. But yeah, I will edit my answer if it's not clear

Of course you are right, that there is no statistical basis to say someone he is safe. But I don't think I did that. I hope he understands my commentary as "30 being safe, regarding to common practices of his research community"

another sidenote: Aren't those the kind of questions, that your supervisor is there to discuss with you?

It's a one hour episode of "talking machines", where they have a discussion with a researcher practicing machine learning in Africa (ca. middle to end) iTunes link: itunes.apple.com/de/podcast/talking-machines/…