12-06-2014 10:12 PM
Hello all. I'm new to Proc ARIMA and could use some advice. My question is not about how to set up an ARIMA test, and it's not even really about how to decide what autocorrelation model best fits my data. My question is, how do I include the autocorrelation model in future statistical analyses that involve this data?
Here's the background story:
I am working with tree sap flow data. Basically I wanted to know how much water certain tree species in my study area were using, so I stuck probes in some trees one summer and got a water flow rate from them. The sap flow data was measured continuously and recorded as a half-hourly average. Later the sap flow rate was converted to a daily average. Because my individual datapoints are not independent, my committee member told me to run a Proc ARIMA test to look for autocorrelation.
I've attached an example of the ARIMA test I ran using data from a single tree. You should find both my code and the results output.
I've been told to...
Look at the ACF plot:
- If ACF shows exponential decline, it is an AR model.
- If ACF shows just sharp drop after lag zero, it is not an AR model.
- Clear spike at lag zero and Weak spikes after that indicate weak MA model structure.
Look at the IACF plot:
- Supports 'no AR' if there are no clear spikes after lag zero. Do not over interpret small
spikes above the significance level.
Look at the white noise test:
The test for autocorrelations white noise (i.e., for a null hypothesis of no autocorrelation at
the specified time lag) is a chi-square test. If autocorrelation tests for the time lags are significant, reject the null that there is
just white noise and accept the alternate that the residuals are autocorrelated at each of these lags.
In this particular example, I can conclude from my results that I do not just have white noise. Based on my ACF and IACF plots, it looks like I have a weak MA model structure. My committee member says I should run future statistical tests both with and without the MA model, then compare the AIC and SBC. If the AIC and SBC are not significantly different,
given the weak MA structure, he says I do not have to include an autocorrelation error structure.
I don't have a lot of depth of knowledge here, so this is probably the first of a dumb series of questions, but where exactly in my output is the MA model I'm supposed to be using? And how do I apply it? There were two sets of ACF / IACF plots, which I don't really understand either. Was SAS trying out a linear model and then a quadratic model, as one might do in a regression analysis?
Thanks in advance for your help.