I forgot to reply to all: This is what I sent earlier to him: """
I would look at the data. You don't really have any information other than what appears to be the weekly passage. I would look to see if there is any relationship between the current value of passage and previous values of passage. You can look at the ACF or PACF of the time series. If these are significant think about creating an arma model for prediction. Ie an ar(1) model would say that there is some linear relationship between past values of passage. an MA(1) would say that there is a relationship between the current value and past shocks (ie differences from the expected passage value) and the current value. An arma model combines both of these. You can use BIC/AIC to fine to your model. Another thing that may exist in your dataset is seasonality ie more fish pass during the summer months than the winter months for example. Hope that helps, Rob """ Given that you don't have a strong background in statistics you would at the very least want to think about a few things. Do you think that the current level of weekly passage has an impact on the next week's passage? Does this make physical sense for fish? You can test this by regressing last week's passage on this week's passage. This would be the same as an ar(1) model. Seasonality also makes sense, do you find that the average value in certain months of the year or weeks of the year are higher or lower than others? What Ben said is basically take a look at the average value and use that to help you predict. It looks like he is basing this relative to a seasonal idea. AIC/BIC is a measure of the liklihood the model fit the data taking into account how many parameters you used to create the model. AIC uses a linear penalize function in the number of paramters and BIC uses a more strict penalizing function in the number of paramters. The intuition is that you can always increase the liklihood model fits the data by adding more degrees of freedom; however, as you add more degrees of freedom you run the risk of overfitting. AIC/BIC basically measure the trade off in additional explanatory power of a variable with the overfitting cost of adding another variable to the model. -Rob On Tue, May 11, 2010 at 2:11 PM, Ben Bolker <bol...@ufl.edu> wrote: > Felipe Carrillo <mazatlanmexico <at> yahoo.com> writes: > > ## snip > > In the absence of any other information, I would say your > best bet would just be to take the weekly average across the > previous years. There are lots of ways to do this (tapply, > aggregate, etc.), but cast() works: > > fallavg <- cast(fallmelt,value="value",WEEK~.,fun.aggregate=mean, > na.rm=TRUE) > names(fallavg)[2] <- "value" > fallavg$variable <- "predicted" > ggplot(fallmelt,aes(WEEK,value/1000,linetype=variable, > colour=variable,fill=variable)) + > geom_line(size=1)+ > theme_bw() + > scale_x_continuous(breaks=seq(1,52,3), > labels=levels(fall$week)[seq(1,52,3)],) + > opts(title="Fall Cumulative") + > labs(y="Number of fish X 1,000",x="WEEK")+ > geom_line(data=fallavg,size=2) > > ______________________________________________ > R-help@r-project.org mailing list > https://stat.ethz.ch/mailman/listinfo/r-help > PLEASE do read the posting guide > http://www.R-project.org/posting-guide.html > and provide commented, minimal, self-contained, reproducible code. > [[alternative HTML version deleted]] ______________________________________________ R-help@r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-help PLEASE do read the posting guide http://www.R-project.org/posting-guide.html and provide commented, minimal, self-contained, reproducible code.