nested factor for which i would like a parameter estimate
Thanks very much Mike. It clears things up immensely now that i'm not thinking about nesting fixed effects inside random effects, and the model and dfs make sense. doug
On 13 Mar 2011, at 21:04, Dunbar, Michael J. wrote:
Hi Doug Your first model can't be right, as AM/PM is clearly a fixed effect: it has just two levels and you are only interested in making inferences about those two levels. So you can't have AMPM on the right of the random part of the model. Try not to think of nesting fixed effects in random effects. Think about specifying the structure of the nesting with the random effects, then the fixed effects are mapped onto the correct level automatically, lme is clever like that. You can always check the degrees of freedom also. The second model sounds more sensible although you probably need to tell it explicitly that AM/PM corresponds to each measurement in your data, something like: Mod.lme3 <- lme(bird_entropy ~ sound_entropy * landuse + AMPM, random = ~ 1 | plot/subplot, data=Bird) This treats subplot as a random effect nested within plot, and each subplot has two measurements, AM and PM. This should handle the correlation between AM/PM, and you can plot residuals to see if there is any further residual correlation. If so, then you can perhaps consider one of the residual correlation structure models. You could go further and fit a full bivariate model between AM and PM, have a look at the paper by Doran and Lockwood 2006 for how to do this with lme. regards Mike
________________________________________ From: r-sig-ecology-bounces at r-project.org [r-sig-ecology-bounces at r-project.org] On Behalf Of dougwyu [dougwyu at gmail.com] Sent: 13 March 2011 07:29 To: r-sig-ecology at r-project.org Subject: [R-sig-eco] nested factor for which i would like a parameter estimate Hello all, I am trying to test whether "sound diversity" predicts "bird diversity." The question i have concerns how to deal with a factor that could be treated either as a random or fixed factor. Response: bird_entropy (continuous) Fixed factors: landtype (factor), sound_entropy (continuous), AM/PM (factor) Random factor: Plot The problematic factor is AM/PM. We have 29 total sampling plots, each of which is measured for sound_diversity at five subplots, once in the AM (dawn) and again in the PM (dusk), for a total 10 measurements per plot. (Each plot has only one, overall measure of bird_diversity). On the one hand, AMPM is nested within plot, but on the other hand, we would like to estimate a parameter for AMPM, since we expect different suites of sound producers (not just birds) at different times of the day. However, it's reasonable to expect temporal correlation between AM and PM. The first model seems uncontroversial to me (these are post-model-selection). Mod.lme1 <- lme(bird_entropy ~ sound_entropy * landuse, random = ~ 1 | plot / AMPM, data=Bird) But i'm curious to know if this second model is reasonable, and if so, how would I code the plot variable? Mod.lme2 <- lme(bird_entropy ~ sound_entropy * landuse + AMPM, random = ~ 1 | plot, data=Bird) Thanks all, Doug _______________________________________________ R-sig-ecology mailing list R-sig-ecology at r-project.org https://stat.ethz.ch/mailman/listinfo/r-sig-ecology -- This message (and any attachments) is for the recipien...{{dropped:6}}