This week you will practice fitting multivariate state-space models with and without covariates. On Tuesday, Eli will walk through the lab on fitting MARSS models with covariates. You'll need this for next week's homework. On Thursday's lab, we'll work through some more complex examples of fitting multivariate state-space models without covariates.
Starting this week, your homework will be more focused on analyzing datasets and the questions are more open ended. These are some common mistakes that students have made in past years:
* In most cases, you need to de-mean your data, both response variable and the covariates. If you subset your data, you change the mean so you need to de-mean again.
* If you square a covariate c with a mean of 0 (meaning c^2), the mean of c^2 is not zero. So if you use c^2 as your covariate, you need to de-mean that separately.
* You cannot compare AICs across models with different data. So if you subset the data, you cannot compare AICc with a data subset to a model with the full data or different data subset.
* A slowly decaying acf is exactly what you expect for AR-1. A pacf with only a peak at 1 is what you expect for AR-1.
* try method="BFGS" in your MARSS call. That is much faster than the default EM-algorithm used by MARSS
For the homework, do problems 1-5. Optionally, try problems 6 & 7.
Fitting MARSS models.pdf
Introduction to MARSS models without covariates.
Fitting MARSS models with covariates.pdf
Introduction to MARSS models with covariates. This is the material that you will need for homework #4.
Fitting MARSS models with covariates-..nt.R
just the R code for the above