# Using auto-correlation to subsample

• Feb 24th 2013, 08:30 AM
cetacean
Using auto-correlation to subsample
Hello,

I would like to use the auto-correlation plots of my data for subsampling. For example, if you look at the photo attached, the data seems to no longer be correlated at lag#9. Since each lag is 28 minutes, this means I should be able to subsample every 4 hours and 12 minutes to get an independent data set. However, the plot shows that the data becomes correlated again at lag#22.

My question is: is it appropriate to subsample every 4 hours and 12 minutes? Or is it only appropriate to subsample according to the last correlated lag (eg. lag#146 = 2 days, 20 hours and 8 minutes)?

Thank you!
• Feb 24th 2013, 05:40 PM
chiro
Re: Using auto-correlation to subsample
Hey cetacean.

Can you outline what you are trying to do in plain non-statistical terms?
• Feb 25th 2013, 04:36 AM
cetacean
Re: Using auto-correlation to subsample
Sure. My data is temporally correlated, and since most statistical analyses require the data to be independent, I want to get rid of that correlation. Therefore, I have ran auto-correlation functions which tell me how much time must be between data points for them to no longer influence each other. For example, if I ran an auto-correlation function that said there was no correlation after 10 minutes, I would subsample my dataset every 10 minutes. However, if the auto-correlation function showed non-significance at 10 minutes, but then a significant correlation again at 20 minutes, then I'm not sure if it's appropriate to use the 10 minute mark or the last point at which it's correlated.

I hope that is more clear.
• Feb 25th 2013, 05:30 PM
chiro
Re: Using auto-correlation to subsample
Can you use some kind Principal Component Analysis for longitudinal data? (Basically you get a set of un-correlated random variables when you do PCA).
• Feb 26th 2013, 02:48 AM
cetacean
Re: Using auto-correlation to subsample
I'm not sure. I've done PCA to determine the main relationships out of a set of variables, but are you thinking time would act as a set of variables?
• Feb 26th 2013, 03:04 PM
chiro
Re: Using auto-correlation to subsample
Basically I'm suggesting you do it with regard to time.