SEM with time series?
On 11/29/2010 11:05 AM, Mudrak, Erika [EEOBS] wrote:
I am helping a colleague with stats analysis, and though it's a seemingly simple setup, it's becoming quite complicated! The system is a deciduous forest with treefall gaps of different carefully chosen sizes. The response variable is amount of NH3 found in the rainwater collected under each gap, sampled once a month during the growing season. Explanatory variables includes gap size (main variable of interest), soil temperature, soil moisture, microbial biomass, etc.... They are all continuous variables, so we would like to do a regression context. We expect the response variable to be autocorrelated over time, so that leads us to want to do a time-series regression. But the other explanatory variables may also be correlated with each other and autocorrelated across time. There are also lots of instances of missing data, for example when no rainfall occurred, there was no opportunity to measure the chemical composition of it. Is there a way to do structural equation modeling (to account for correlation between explanatory variables) with a time series component (to account for autocorrelation of explanatory variables)? Or is there another more appropriate technique? Thank you, Erika Mudrak
My guess (not having done much of this stuff myself) is that a full Bayesian setup (WinBUGS etc.) would be the simplest (!!) way to handle this kind of problem. Of course, there's a lot of conceptual and programming overhead in learning to set it up ... if you want to go this route and you are new to Bayesian stats and WinBUGS I would suggest McCarthy's book for basics and one or more of (1) Clark [comprehensive and oriented toward ecology but dense in places] (2) Gelman and Hill [extremely clear treatment of multi-level modeling in general] or (3) my book [not as specific to Bayes/WinBUGS, but long on general explanation] for tackling your real problem. good luck ... Ben Bolker