On Thu, 17 Nov 2011 10:39:11 -0500, Susan Durham <sdurham@BIOLOGY.USU.EDU> wrote:

>On Wed, 16 Nov 2011 15:41:47 -0500, Adam Smith <raptorbio@HOTMAIL.COM> wrote: > >>All, >> >>Please consider the following binomial GLMM, which models the the number of >>fruits removed from a plant during a given count period relative to the >>total number of fruits present at the start of the count, as a function of a >>treatment (TRT) and geography (GEOG): >> >>PROC GLIMMIX DATA=abs2009 method=laplace order=data; >>CLASS TRT GEOG PLOT COUNT; >>MODEL REMOVED/TOTAL = GEOG COUNT / S DIST=BIN LINK=CLOGLOG COVB; >>RANDOM INTERCEPT / SUBJECT=PLOT; *16 total plots; >>RANDOM INTERCEPT / SUBJECT=TRT(PLOT); *32 trt within plot combinations; >>RUN; >> >>(Selected) fit statistics, for comparison to a subsequent parameterization: >> >>-2 Log Likelihood 2931.64 >>AICC (smaller is better) 2967.18 >> >>However, the fit statistics for the conditional distribution suggest >>overdispersion (i.e., Pearson Chi-Square / DF = 4.28). >> >>I thought an appropriate way to model the overdispersion would be to create >>an individual-level random effect (i.e., assign each observation a unique >>identifier, OBS, and include that as a random variable): >> >>PROC GLIMMIX DATA=abs2009 method=laplace order=data; >>CLASS OBS TRT GEOG PLOT COUNT; >>MODEL ABS/TOTAL = GEOG COUNT / S DIST=BIN LINK=CLOGLOG COVB; >>RANDOM INTERCEPT / SUBJECT=PLOT; >>RANDOM INTERCEPT / SUBJECT=TRT(PLOT); >>RANDOM INTERCEPT / SUBJECT=OBS; >>RUN; >> >>(Selected) fit statistics suggest an improvement: >> >>-2 Log Likelihood 2150.06 >>AICC (smaller is better) 2187.78 >> >>as do the fit statistics for the conditional distribution (i.e., Pearson >>Chi-Square / DF = 0.29). >> >>The fixed parameters and their standard errors are calculated, and agree >>well with the same model evaluated in other software, but there are no df >>present to evaluate their significance: >> >> Num Den >>Effect DF DF F Value Pr > F >> >>geog 1 0 9.31 . >>count 13 0 9.79 . >> >>Can I evaluate these effects, then, with LRTs or AIC? Is there a more >>appropriate way to model the overdispersion in this case (e.g., R-side >>effects)? Beta-binomial model? >> >>Thanks very much for considering, >> >>Adam Smith >>University of Rhode Island > >Adam, > >Over-dispersion could be due to one or more components of the model that are >incorrectly specified: linear predictor, distribution of the data, link, >and/or random effects and the associated covariance structures. > >The first thing to consider is your linear predictor. I don't know how GEOG >and COUNT fit in your design, but I definitely suspect that TRT should be in >the MODEL statement. An incorrect MODEL statement might also be the cause of >your 0 df problem. > >HTH, >Susan

Susan,

Thanks for the thought. TRT was included initially in the MODEL statement, and it does have df associated with it. I dropped it due to its lack of contribution to the model. TRT's inclusion does not affect the df problem with GEOG and COUNT.

To summarize the design, there where 16 plots, which were then halved and treatment assigned randomly to the halves. Half of the plots where in the North, half in the south (GEOG). COUNT indicates the 14 repeated visits to count the number of fruits remaining.

If I'm not mistaken, the RANDOM INTERCEPT / SUBJECT = PLOT statement is meant to account for the paired treatment design w/in plots, while the RANDOM INTERCEPT / SUBJECT = TRT(PLOT) statement accounts for the repeated counts on a given half of a plot (there is little to no temporal autocorrelation, as removal of fruits is quite stochastic), and the RANDOM INTERCEPT / SUBJECT = OBS statement is meant to model overdispersion...

Cheers, Adam