Dear all,
I have maybe for the "time series" experts a silly question:
-I have a dataset of European rivers =80
-In 50% of the rivers I have more than 1 project; in the other 50% is 1 river = 1 project
-In 50 % of the projects I have data collected only for 1 year; in the other 50% of the projects data were collected over years (from 2 untill 20 years, depending on the project)
->I want to assess the Fish diversity depending on the altitude, latitude, catchment size.
After exploring data for the model assumption of normality, variance heterogeneity etc..I though to run this model:
mod
Alessandro,
I am not familiar with ecology, but hopefully the following answers help. The software you're using looks like the nlme package in R, correct?
1/ Your model looks decent; I am sure you have theoretical reasons for using the logarithms of your covariates. I'd recommend testing a baseline model using gls, then use the lme commands. Add the covariates one at a time in a heirarchical fashion (i.e, log (altitude), then log(latitude), then log (catchment)). Also, add the time variable. You may add the time variable before or after the log covariates, but that would depend on the theory of what you do. So you would know best. The responses that follow are based on your question only.
UPDATE: I forgot to mention to randomize your slopes after adding the time variable to allow for varying growth over time for each river. This is in addition to the other models recommended above.
3/To set up the data, yes. Organize the dataset chronologically using "long form". This means that you make a variable for the year with 9 levels (e.g, 2000-2008 in Project 1) and input the data for all the years as such. Do not use the "wide form" used in repeated measures designs in SPSS where each year has its own column.
2/ Assuming you keep the syntax presented in your question, you can then model the covariance structure using something such as below: modAR
Briefly, I would indeed expect to observe time-related correlations in the model residual. Theoretically, the observations measured are not iid, as required by OLS and related models, because consecutive observations are expected to be related. Intuitively that should make sense - one could reasonably expect that measurements that are closer in time to have higher covariance. Time series analysis developed a series of models beyond the basic AR(X) to deal with this. I'd suggest you take a look at vector error correction (VEC) models and variable co-integration. I'm not going to try to cover the models/theory in this post, but the basic idea is to separate the relationship between variables from their mutual relationship with time.
That being said, your model, perhaps with an AR(x) correction as suggested by Keston, seems like reasonable form a mixed effect model i saw used for similar data in ecotoxicology.
Dear Keston and Fjdor,
thank you very much for your suggestions. These were really usefull. I will check both options (including VEC models).
Thank you very much
- Badges
- Science topic