> -----Original Message-----
> From: [EMAIL PROTECTED] [mailto:[EMAIL PROTECTED]]On
> Behalf Of Jacek Gomoluch
> Sent: Wednesday, October 02, 2002 7:21 AM
> To: [EMAIL PROTECTED]
> Subject: Calculation of the confidence interval
>
>
> Hi,
>
> I have a question concerning the calculation of the confidence interval of
> the mean of a measured variable.
>
> I am running several (m>30) simulation runs with same parameters but
> different random seeds.
----------------------------------------------------------------------------
This doesn't make sense, unless you are testing the random number generator
(RNG). The characteristics of a good RNG do not change with seed. Even poor
ones show little dependence on seed. Patterns in linear congruential
generators (LCG) show up quickly. The lattice structure of LCGs does not
change with seed.
----------------------------------------------------------------
> In each of these simulation runs, n
> samples (usually
> several thousand) of a measured variable are taken. In my simulations, the
> samples within each simulation run can not be assumed to be independent.
-----------------------------------------------------------------
The covariance/correlation structure of typical RNGs is small. You can
assume each RNG value as being independent of others. Only in very critical
simulations would you use true random numbers (such as from G. Masaglia's
CDROM), which have statistically no correlations.
DAH
--------------------------------------------------------------------------
> However, according to the literature I read so far, the sample
> means of the
> different simulation runs are normally distributed (due to the
> central limit
> theorem).
In the literature, it is assumed that the number of
> samples (n) is
> the same for each simulation run.
>
> With this assumption the confidence interval of the mean is calculated by:
>
> overall_sample_mean +/- z_value* ( standard_dev_of_the_sample means)
>
> However, in my experiments the number of samples (n) is not the same for
> each simulation run, but slightly different (e.g. 3105, 2934, 3050,...).
>
> My question: does this matter? Or can I still use the above formula?
>
> Thanks in advance!
>
> Jacek Gomoluch
-----------------------------------------------------------------------
Theoretically the confidence intervals are based on a given sample size. For
each of your sample sizes, you would have different confidence intervals. It
would be simpler to just compute one for the smallest sample. If all are
within the one for the smallest sample, you are ok. For each outside,
compute the exact interval and determine which exceed their individually
calculated interval.
DAH
-----------------------------------------------------------------------
>
.
.
=================================================================
Instructions for joining and leaving this list, remarks about the
problem of INAPPROPRIATE MESSAGES, and archives are available at:
. http://jse.stat.ncsu.edu/ .
=================================================================
