Peter,

Confidence intervals are not additive but variances are. Therefore you can 
calculate the variances for all your “parts” and then add them up and calculate 
a confidence interval based upon that. There is one complicating matter and 
that is whether the parts are independent or not. If two variables (parts) are 
independent then you can simply add the variances as:

Var(ΣXi) = Σ Var(Xi)

This extends easily from two to n variables. The complication arises if the 
parts are not independent, as they almost certainly are in your case as you 
have temporal autocorrelation. For two non-independent variables:

Var(ΣXi) = Σ Var(Xi) + 2 Σ Cov(Xi, Xj)

You can extend this to n variables by calculating all of the pairwise 
correlations for i < j, adding them up and then multiplying by two.

The final variance estimate can then be used to calculate a confidence interval 
assuming you have good estimates for covariance.

Temporal autocorrelation (autocovariance) generally decreases, as the time 
difference between the variables increases, so not every one of your samples 
will show a significant correlation with every other sample. This can cause the 
final variance to be negative due to overestimating large-lag covariances and 
so they may need to be downweighted. The following link shows that problem and 
how to potentially correct it using R.

http://stats.stackexchange.com/questions/79216/summing-variance-of-autocorrelated-timeseries

You will need a statistical program that allows you to calculate autocovariance 
or autocorrelation.
><(((º>   ><(((º>   ><(((º>   ><(((º>   ><(((º>   ><(((º>
Jim Novak
Interim Associate Dean and MSNS Coordinator
College of Sciences
2118 Old Main
Eastern Illinois University
Charleston, IL  61920
(217) 581-3328
(217) 581-7486
[email protected]<mailto:[email protected]>
><(((º>   ><(((º>   <º)))><   ><(((º>   ><(((º>   ><(((º>

Confidentiality Notice: This e-mail message, including
any attachments, is for the sole use of the intended
recipient(s) and may contain confidential and privileged
information.  Any unauthorized review, use, disclosure
or distribution is prohibited.  If you are not the intended
recipient, please contact the sender by reply e-mail and
destroy all copies of the original message.

On Apr 13, 2015, at 7:54 PM, Peter Novak 
<[email protected]<mailto:[email protected]>> wrote:

Hello,
I am a post grad student undertaking research on the ecology migratory
freshwater shrimps in northern Australia. These shrimp undergo an mass
migration from the estuary to the river at the end of the wet season and a
goal of my research is to estimate the biomass of shrimp moving upstream.
I have sampled the migration, to determine hourly migration rates, using
fyke nets at periodic interval throughout the night over a period of 6
weeks. I couldn't sample every night and I couldn't leave nets out all
night as there were significant crocodile risks. So I have half hourly
migration rate data for four nights every week, for 6 weeks. I have
completed multiple regression analysis to find the best model (using a
range environmental predictor variables) to predict migration rate, and
discharge came up as the strongest predictor. So I want to use the
equation from the discharge and biomass regression to estimate migration
biomass over the entire 6 week period. The end result will be an estimate
of the total biomass moving upstream over 6 weeks. Because the total
biomass estimate involves summing many predictions (of hourly biomass over
each night) I am not sure how to calculate the confidence interval around
the final prediction. Can I calculate the confidence interval for each
prediction and then sum it to get the total confidence interval estimate?
Is there an easy way to do this with a standard stats package such as
Statisitica, SPSS or Excel? I would appreciate any feed back.

Thank you
Peter

Reply via email to