I'm analyzing data from a designed agricultural experiment using linear mixed models.  At one point, I collected data for what I thought would serve a covariate (weed rating) for explaining crop yield, but further analysis revealed overlap between this variable and one of my experimental treatments.  A major hint that these variables might be collinear was that, when both variables were included in the model as predictors for crop yield, the treatment variable was no longer significant (it previously had been consistently significant).  Note that both variables are categorical, and the treatment variable has two levels while the covariate (weed rating) is an ordered variable with 5 levels (1 to 5).    I would like to determine a standard way to demonstrate redundancy between these variables, so that I can justify not including them simultaneously in models.   A chi-squared test (code below) seems to be appropriate.  Is this approach adequate for establishing predictor redundancy?  I'm not using the term "collinearity" here, because I'm under the impression that collinearity applies more to regression and not ANOVA.  I don't think I need to determine VIF here. title "Chi-squared test for predictor redundancy"; proc freq data=df; tables treatment * other_variable / chisq; run; Ultimately, my goal is to establish a causal link between the response (crop yield), the treatment, and weed rating.  I've accomplished part of this by treating the covariate (weed rating) variable as the response variable and modeling that with my experimental treatments.  That confirmed a high degree of correlation (not sure if "correlation" is technically correct to use here, but you know what I mean) between the treatment in question and the weed rating.   Now, I believe that I just need to establish the link between weed rating and yield, and justify not including the redundant treatment in the model.  Thank you for reading. ... View more

I've configured two repeated measures models, one using PROC MIXED and the other using PROX GLIMMIX, as consistently as possible to the best of my understanding.  The experiment is a factorial, RCBD. Trt_Amend_App and Trt_CC are experimental treatments with 8 and 2 levels, respectively. ID_S are sites or locations (2) Block has 4 levels Year_Time is the time component with 7 levels. ID_S_P is the plot ID or subject     proc mixed data=df plots=studentpanel; class Trt_Amend_App Trt_CC Year_Time ID_S Block ID_S_P; model Soil_PAN_LN = Trt_Amend_App | Trt_CC | ID_S | Year_Time; repeated Year_Time / type=un sub=ID_S_P; random Block(ID_S); run; proc glimmix data=df plots=studentpanel method=rspl; class Trt_Amend_App Trt_CC Year_Time ID_S Block ID_S_P; model Soil_PAN_LN = Trt_Amend_App | Trt_CC | ID_S | Year_Time; random Year_Time / residual type=un subject=ID_S_P; random Block(ID_S); /*nloptions maxiter=500; /* Set this to avoid "did not yield a valid objective function error"*/ /*nloptions technique=nrridg;*/ run; The within-subject correlations are complex, and unstructured yields the best (lowest) information criteria results (AIC, BIC).  The MIXED model doesn't execute when using compound symmetry, giving the "WARNING: Stopped because of infinite likelihood" error.    I'm curious as to why I cannot get the GLIMMIX model to execute, and why I get the error "The initial estimates did not yield a valid objective function." every time that I attempt to run it.  To address this I've attempted increasing the iterations (via INITITER and MAXITER, the latter of which is shown commented out).  I've also tried changing the optimization technique to NRRIDG, which I've read is the default for PROC MIXED but not PROC GLIMMIX.  None of that has worked, and I get the valid objective function error each time.  I considered passing PARMS based on results from MIXED, but that seems unpractical given the amount of covariance parameters with unstructured correlation (28).  Why might GLIMMIX be failing where MIXED works? Also, it occurred to me that within-subject measurements may be so irregular that attempting to model within-subject correlation may be pointless.  How might one verify whether there is any reason to use repeated measures period?  When executing the MIXED model with ar(1) structure, the within-subject correlation is less than the estimate for residual, but not very low (see picture).  I could use some practical advice in this regard.   Thanks for reading! ... View more

I have data on the severity of weed pressure for an agricultural study with a factorial, RCBD.  The weed pressure variable is the area in square footage within a plot where weeds were measured to be above a certain height.  It is an atypical metric for weediness, and I have not been able to find anything in the literature like it, so I'm unsure of how to analyze this data.   The raw, untransformed variable "Area_Weed", in units of square feet, exhibits right-skewness.  A few of the observations are zero.  The histogram is shown below:   One approach that I've tried is to convert the response variable "Area_Weed", in square feet, to a proportion of weed area divided by total plot area, the apply an arcsine transformation.  The residuals plot exhibits non-constant variance.  The code and residuals are shown below: proc glimmix data=df plots=studentpanel method=rspl; class Trt_Amend_App Trt_CC Block; model Area_Weed_PROP_ANG = Trt_Amend_App | Trt_CC / ddfm=kr2; random Block; run;   Considering an alternative distribution for the untransformed response variable "Area_Weed", which is positive and right-skewed, the gamma distribution seems promising, aside from the fact that it can't accommodate zeros in the data.  If I apply a transformation of adding a small constant (+1) then I can get the model to run using PROC GENMOD, with the caveat that I can't include the variable "Block" as a random effect. proc genmod data=df plots=all; class Trt_Amend_App Trt_CC Block; model Area_Weed_PLUS = Trt_Amend_App | Trt_CC / dist=gamma link=log; ods output ParameterEstimates=pe; output out=outmean pred=mu; run; My questions amount to the following: Based on what I've presented here, is the gamma distribution appropriate for my response variable, which is positive, right-skewed, and includes a few zeros? If I use PROC GENMOD and a gamma distribution to model, how might I go about evaluating the residual plots?  What might be a good option for the parameter plots aside from "all"? If I attempt to incorporate Block as a random effect using PROC GLIMMIX, the model does not converge.  How can I address this issue? Thank you for reading.  Please let me know if I can provide any other information. ... View more