Factorial effects in SAS/AUTOREG

[GPerry1]

I've been investigating the AUTOREG procedure for a while now and I had a question about the support for discrete class effects. The option exists but how well is this supported? Has anyone had any experience with it?

Specifically, I have a system with numerous (>130k records) across about ten groups of organisms (moving through the same facility over time (anywhere from 3-12 months depending on stage)), reared in several (ranges from 10-20) units, with contiguous observation arrays of two to several months, for a single dependent variable. Is the class option in AUTOREG sufficient to support this kind of near-mixed modeling longitudinal array?

Best,

GP

[sbxkoenk]

Hej,

• What's the measurement scale of your dependent variable?
• Why are you considering AUTOREG procedure? You want a GARCH-Type Model?
• Why are you calling it near mixed-modelling? (it seems plain mixed modelling to me and the units seem a random effect to me)
• Many procedures in SAS support mixed and random effects (not only the procedures that contain the string 'mix')

I would like to hear more before answering this.

Koen

[SteveDenham · Accepted solution]

I would follow one of @sbxkoenk 's suggestion.  This design looks like a mixed model to me. I would assume correlated errors over time and some possible random effects. Fixed effects that jump out at me are organism, time in facility, and their interaction.  Random effects would be unit and observation arrays within units. The latter may have some sort of geospatial correlation.  So, a complex model with a lot of parameters. I think you have adequate data, but you may not have adequate computing power.

SteveDenham

[GPerry1]

Hi Steve - thanks also for this. It's pretty much as you and Koen say: definitely mixed for the arrays and tanks within arrays. I analyzed within the arrays using AUTOREG, but I'll check the other procs and see if they have any longitudinal-mixed options.

[GPerry1]

Thanks for the comment Koen.

• Measurement scale would be a quantitative with an interval of 1, since the dependent variable is mortality. I don't think I should worry about this violating some kind of longitudinal CDF; the absolute range runs from 0 to about 100 or so, so I think it's at least close to continuously distributed.
• I'm looking at AUTOREG (maybe using GARCH where indicated) because it's a longitudinal track of mortalities with a long coincident track for environmental variables.It's the environmental effects on morts that we wanted to figure out.
• I have to agree that the system would definitely be a mixed model with units as random; the longitudinal platforms on SAS don't seem to handle this specifically, so what I did was to take the coward's way out and address each 'lot' of animals independently. It's not optimal, but I understand that the factorial effects in AUTOREG are a bit experimental anyway. Our final conclusions will be a bit 'thumb in the wind' but I don't think I have another option.
• I've looked around at the other procs but I think AUTOREG is kind of the right choice; it's for sure that there's autoregression for our mort counts over short-term day tracks in the system.

Thanks very much - I agree with your positions and I think we've got some reasonable conclusions out of the results.

G