Source:
Fernando Soares Schlindwein
FFT and AR are both estimates of the spectrum. Their results depend on how close the assumptions built in the 2 different techniques hold. FFT assumes periodicity. AR assumes that future values can be predicted from a linear combination of previous values (that is, that there will be no discontinuities on the signal and a certain number of derivatives). For the same data length AR has a better spectral resolution than FFT, especially for short data segments. The 'difficulties' associated with AR are that (I) it requires the choice of a 'model order' p and (ii) it is slower than FFT for the same data set. Also AR can only 'see' a maximum of p/2 spectral peaks for a model order p. Fortunately there are several good techniques for finding the 'optimum order' of the AR model. Check the literature. In my experience an order of around p=16 should produce good results. Also beware that AR does not require a sampling frequency which is too high. Oversampling in AR is not a good idea. Just respect Nyquist, (fsam>2fmax) but don't go to more than 4fmax, as doing that will not improve your spectral estimation and will demand an artificially higher model order (and consequently longer processing time).
I would not dismiss AR spectral analysis. It is a very interesting approach.
The following 2 publications might be useful: and .
Antonio Carlos Silva-Filho
The basic difference between FFT and AR is the data treatment. The AR method "smooths" the frequency curve, which may lead you to misinterpretation of data, taking out outliers that can be part of the pattern, thus strongly interfering in the final analysis.
James A J Heathers
On a practical level, it isn't worth losing much sleep over - IF you're interested in reporting the PSD of broad frequency bands. The results are closely equivalent. There may be further factors to consider with your stated "correction for respiration", but we'd need more information.
On a more detailed level, there are a number of differences which have been outlined in great length in previously published work. I have attached two comprehensive treatments of this issues, bear in mind there are a lot of others:
Henrique Soares
As mentioned by the colleagues, the two methods are quite different in data processing. In a recent paper from our group that has been accepted, we sought to assess whether using FFT or AR would influence the association of the HRV parameters with some cardiovascular risk factors (systolic BP, BMI, waist circumference) in a large cohort of adolescent boys. Our results confirmed the differences between AR and FFT by providing different values for all variables (LF, HF, LF/HF, total power).
however, the magnitude of the association between HRV and the cardiovascular risk factors was similar between methods (FFT and AR). Although further studies are needed to confirm these results, it is plausible to suppose that both methods could be used for assessing the association between cardiovascular risk and HRV, at least in adolescent boys.