RE: st: Filtering methods with short time series

[kokootchke <kokootchke@hotmail.com>]

Dear Nick,Thanks again for your comments. I am intrigued about your last comment:> Also, I get the impression that you are filtering these series individually, but pooling them as longitudinal data would come to mind first to many readers of this list.Are you suggesting I append the different series I have and then filter the constructed longer series? I don't think that would be appropriate given that my different variables correspond to the amount of time spent by individuals on various activities (eating, sleeping, shopping, personal care, ...), and many of these would have very different behaviors over time (for example, after the financial crisis, people shop less and eat out less, but sleeping patterns don't change much--controlling for employment status). But perhaps I could group similar activities and construct the longitudinal series the way you suggest for various broader groups, e.g., leisure activities, work, etc.? The other thing I thought I could do was to wor!
 k with principal components.

> I think much of the literature is coy about short time series because (1) that's the way the data often arrive and there is still a practical problem of what to say about them and (2) people who think time series analysis with short time series is intensely problematic don't usually write texts or papers on it.
>
> Flippancy aside, from the point of view of seasonality your sample size is 7, not 84, but it's not very clear what you can do about that, but first be cautious.I couldn't agree more. I'm being cautious and this is only an exploratory study to see if a case can be made to keep collecting more time use data at the national level. >
> However, in principle, you could explore the properties of whatever you do by simulating extra datasets. That would ideally require some idea of an appropriate generating process, which I can't supply.

Yes, I had thought of this and I'm working on it.
Thanks a lot.Adrian
----------------------------------------
> From: n.j.cox@durham.ac.uk
> To: statalist@hsphsun2.harvard.edu
> Date: Wed, 2 Feb 2011 11:26:49 +0000
> Subject: RE: st: Filtering methods with short time series
>
> I think much of the literature is coy about short time series because (1) that's the way the data often arrive and there is still a practical problem of what to say about them and (2) people who think time series analysis with short time series is intensely problematic don't usually write texts or papers on it.
>
> Flippancy aside, from the point of view of seasonality your sample size is 7, not 84, but it's not very clear what you can do about that, but first be cautious.
>
> However, in principle, you could explore the properties of whatever you do by simulating extra datasets. That would ideally require some idea of an appropriate generating process, which I can't supply.
>
> Also, I get the impression that you are filtering these series individually, but pooling them as longitudinal data would come to mind first to many readers of this list.
>
> Nick
> n.j.cox@durham.ac.uk
>
> kokootchke
>
> Thanks, Jorge, Kit, and Nick for your answers.
> Regarding -cfitzrw-, I have also obtained the output from the CF filter in R and, although there are some slight differences, they are not substantial (at least in my case). The advantage is that R allows for non-RW, and I have run several tests and this RW assumption also does not seem to affect my results.
> Sorry for the confusion about the I(2) comment. I just meant to say that my series are not integrated and hence the robustness of the test to high degrees of integration, like that of an HP filter, are not something of concern to me...
> I have also used the -couliari- filter. This produces results very similar to -cfitzrw-, perhaps a bit less wavy, but not too different from each other.
> My biggest concern is the short length of the series. I didn't know if there was a rule of thumb regarding the number of periods one should use to obtain "credible" trends/cycles. I also didn't know if there are any caveats to using short time series and the literature doesn't seem to be concerned on this regard.
>
> *
> * For searches and help try:
> * http://www.stata.com/help.cgi?search
> * http://www.stata.com/support/statalist/faq
> * http://www.ats.ucla.edu/stat/stata/
 		 	   		  
*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/