[R] AICc vs AIC for model selection

Spencer Graves spencer.graves at pdf.com
Mon Jul 17 17:37:32 CEST 2006

	  I understand that you have only 26 observations.  Model 
identification always requires more observations than estimating a model 
you already think you know.  If it were my problem, I think I'd first 
plot the data over time and make a normal probability plot of the data. 
  Then I'd fit the simplest, most parsimonious model I could think of 
that would include the trend and seasonal.  Then I'd examine the 
residuals and check the p values.  If everything looked sensible, 
wouldn't push it further.  If I had several such series, I'd study 
Pinheiro and Bates (2000) Mixed-Effects Models in S and S-Plus 
(Springer) and use the 'nlme' package to do more.

	  Hope this helps.
	  Spencer Graves

Sachin J wrote:
> Hi Spencer,
> I did go through the previous postings in the mailing list. But couldn't 
> find satisfactory answer to my question. I am dealing with univariate 
> time series. I suspect that my  data may contain some trend and seasonal 
> components. Hence, rather than just fitting just AR(1) model, I am 
> trying to find the right model which fits the data well and then use 
> that model to forecast. In order to achieve this I am using best.arima 
> model. If you have any other thoughts on this please let me know.
> Thanx in advance for your help.
> Regards
> Sachin
> */Spencer Graves <spencer.graves at pdf.com>/* wrote:
>     Regarding AIC.c, have you tried RSiteSearch("AICc") and
>     RSiteSearch("AIC.c")? This produced several comments that looked to me
>     like they might help answer your question. Beyond that, I've never
>     heard of the "forecast" package, and I got zero hits for
>     RSiteSearch("best.arima"), so I can't comment directly on your question.
>     Do you have only one series or multiple? If you have only one, I
>     think it would be hard to justify more than a simple AR(1) model.
>     Almost anything else would likely be overfitting.
>     If you have multiple series, have you considered using 'lme' in the
>     'nlme' package? Are you familiar with Pinheiro and Bates (2000)
>     Mixed-Effects Models in S and S-Plus (Springer)? If not, I encourage
>     you to spend some quality time with this book. My study of it has been
>     amply rewarded, and I believe yours will likely also.
>     Best Wishes,
>     Spencer Graves
>     Sachin J wrote:
>      > Hi,
>      >
>      > I am using 'best.arima' function from forecast package
>     to obtain point forecast for a time series data set. The
>     documentation says it utilizes AIC value to select best ARIMA
>     model. But in my case the sample size very small - 26
>     observations (demand data). Is it the right to use AIC value for
>     model selection in this case. Should I use AICc instead of AIC.
>     If so how can I modify best.arima function to change the selection
>     creteria? Any pointers would be of great help.
>      >
>      > Thanx in advance.
>      >
>      > Sachin
>      >
>      >
>      >
>      >
>      > ---------------------------------
>      >
>      > [[alternative HTML version deleted]]
>      >
>      > ______________________________________________
>      > R-help at stat.math.ethz.ch mailing list
>      > https://stat.ethz.ch/mailman/listinfo/r-help
>      > PLEASE do read the posting guide!
>     http://www.R-project.org/posting-guide.html
> ------------------------------------------------------------------------
> Do you Yahoo!?
> Next-gen email? Have it all with the all-new Yahoo! Mail Beta. 
> <http://us.rd.yahoo.com/evt=42241/*http://advision.webevents.yahoo.com/handraisers> 

More information about the R-help mailing list