How to define the knee frequency for separating low- and high-frequency slopes in PSD?

[Eunggyu]

Hello, I am currently attempting various approaches to analyze the differences in the 1/f slope between two groups.

I would like to divide the slopes into low-frequency and high-frequency components based on the point where the knee occurs.

My expectation was that the knee frequency in the PSD would be around 1.3 (20 Hz). Therefore, I planned to calculate and analyze the slopes for 1–20 Hz and 20–50 Hz separately.

However, the average value of the knee frequency I obtained from each PSD, calculated as "knee^(1/slope)", turned out to be around 1.11 (13 Hz), which was much earlier than I had expected. I double-checked using kernel density estimation, but found no error.

My questions are as follows:

Does the calculated knee frequency truly reflect the actual bending point in the PSD? If this value is not accurate, would it be acceptable to simply define the bending point visually at 20 Hz?

I have not found any papers that aperiodic fitting only the 1–15 Hz range (or below 15 Hz). Is this because the fitting range is too narrow to produce reliable results?

Are there alternative methods available for determining the bending point in the PSD?

Thank you in advance to anyone who can provide insights.

[voytek]

It's more complex than the visual "bending point" due to so many interacting factors – including the oscillatory "bumps" – so it's hard to give you any concrete advice. The math is finding it at ~13 Hz.

But whenever we run into situations where we're just not sure, I tell folks to run a sensitivity analysis: iterate a bunch of times using different fitting ranges and fitting settings and see if there's a common trend.