Hampel Matlab vs R

I am struggeling rewriting matlab code in R as the hampel implementation in matlab seems to be different to the pracma package in R which, according to the window length leaves the lower and upper ends of times series untouched, whicle matlab seems to modify them. I have highlighted the problem already on stackoverflow: https://stackoverflow.com/questions/64631709/hampel-matlab-vs-r

As I would really like to know where the discrepancies are coming from, help is highly appreciated!

With kind regards

CV

Answer 1

[Bjorn Gustavsson]

You seem to use different filter lengths in your call to hampel in matlab (30 - which I take to mean you will get a window of 2*30+1 samples, i.e. 30 samples before and 30 samples after the central point) and R (15 which I find very hard to believe gives you a total filter-length of 61.) Since the matlab-code is available to inspect you can simply go to the code to see what's done, R is an open source project so that source-code should also be available. Perhaps R uses median absolute deviation an matlab uses mean absolute deviation? I cant recall what the default of mad in matlab is at the moment.

HTH

[Claudio Viecelli]

Thank you very much for your answer. I corrected the windows length but still experiencing differences :(

[Bjorn Gustavsson]

Have you compared the source-codes?

[Claudio Viecelli]

Examining the differences between the two implementations, I found out that R leaves the window-length at the beginning and ending of the time series untouched while matlab is modifying it. How exactly, I havent found out yet.

[Bjorn Gustavsson]

My guess for the matlab-hampel's handling is that it simply cuts out all indices that fall outside your array-limits, something like this:

idxWindow = [max(1,idxCentre-K):min(array_length,idxCentre+K)]

and then calculates the mad on that array. You'll have to investigate using a suitably small array where you can change the values at the ends and compare the output with what you'd calculate "by hand", using sensibly small K (3, 4, 5?)...

How R can use the full window-length when it runs outside the array-limits, is a question for someone else, perhaps simply repicating the last element?