Re: Fwd: regression coefficient confidence intervals and degrees of freedom

The classic df for autocorrelated time series data with "large"
sample size is

   df = N * (1 - acr) / (1 + acr)

For smaller sample sizes, the equivalent sample size is recommended.

>
> Hi,
>
> I'd like to compute regression coefficient confidence intervals and the
> timeseries I'm examining have significant autocorrelation. The
> confidence intervals depend on the degrees of freedom (df). The page at
> http://www.ncl.ucar.edu/Document/Functions/Built-in/regline.shtml states
> that df should be adjusted as
>
> df = N * (1 - acr) / (1 + acr)
>
> where acr is the lag-1 autocorrelation of the dependent variable,
> assuming that this autocorrelation is statistically significant.
>
> A message on ncl-talk, archived at
> http://www.ncl.ucar.edu/Support/talk_archives/2011/1866.html suggests
> for the related problem of significance testing of the regression
> coefficient to use the ncl function equiv_sample_size to compute the df.
>
> For the timeseries I'm examining, equiv_sample_size yields 84, and the
> formula given with the regline documentation yields 90.5. These values
> are different, though not dramatically so.
>
> Is there a reason to prefer one of these df computations over the other
> in the computation of regression coefficient confidence intervals?
>
> (I've browsed the Zwiers& von Storch paper mentioned in the
> equiv_sample_size documentation, but it is over my head.)
>
> Thanks, Keith
> _______________________________________________
> ncl-talk mailing list
> List instructions, subscriber options, unsubscribe:
> http://mailman.ucar.edu/mailman/listinfo/ncl-talk
>
_______________________________________________
ncl-talk mailing list
List instructions, subscriber options, unsubscribe:
http://mailman.ucar.edu/mailman/listinfo/ncl-talk
Received on Fri Apr 6 13:54:04 2012