Susan,
Sorry for the confusion about study description to help with the questions that I'm asking! My study was conducted on a 1,861 hectare property and was divided into 3 treatments of approximately equal size and vegetative composition. The three treats are control, no hunters; low density, 1 hunter/101 ha; and high density, 1 hunter/30 ha. I decided to evaluate exposure to hunting pressure (initial, days 1-3; prolonged, days 4-13) to assess differences in habitat selection and by treatment levels. These are the two categorical variables in the model. The other categorical variable is habitat value (1 = mixed, 2 = forest, and 3 = field). To conduct a logistic regression analysis using PROC GLIMMIX, I allocated random GPS locations on the study area and paired them with actual animal locations. I've been working with two biometricians and we're all stumped on the issues that I'm having in this analysis. We decided to take a more simplistic approach to get away from the three-way interactions using a "BY" statement for EXPOSURE, TREAT, AND DN to use more of an estimation approach since with a large daatset we're bound to find something statistically significant. See my new model below:
PROC GLIMMIX DATA=HABITAT; BY EXPOSURE TREAT DN;
CLASS ID TREAT YEAR HABVALUE;
MODEL VALUE (EVENT = '1') = HABVALUE ROAD_LOG ELEVATION ELEVATION_QUAD SLOPE SLOPE_QUAD / DIST=BINARY LINK=LOGIT SOLUTION;
RANDOM ID(TREAT) TREAT YEAR / TYPE=VC;
RANDOM ELEVATION SLOPE ROAD HABVALUE / TYPE=VC SUBJECT=ID;
RUN;
I have two years of data (YEAR), TREAT(control, low, and high), and ID(TREAT) selection of resources made by an animal are more similar or correlated within each treatment. Lastly, RANDOM ELEVATION SLOPE ROAD HABVALUE are added to models the correlation of resource selection within individuals, which means that selection of (e.g., elevation) between intervals is correlated within an individual.
I ran this more simplistic model this morning and here's the errors I received back:
NOTE: The GLIMMIX procedure is modeling the probability that Value='1'.
WARNING: Pseudo-likelihood update fails in outer iteration 5.
NOTE: Did not converge.
NOTE: The above message was for the following BY group:
EXPOSURE=Initial TREAT=Control DN=Diurnal
NOTE: The GLIMMIX procedure is modeling the probability that Value='1'.
WARNING: Pseudo-likelihood update fails in outer iteration 3.
NOTE: Did not converge.
NOTE: The above message was for the following BY group:
EXPOSURE=Initial TREAT=Control DN=Nocturnal
NOTE: The GLIMMIX procedure is modeling the probability that Value='1'.
WARNING: Pseudo-likelihood update fails in outer iteration 3.
NOTE: Did not converge.
NOTE: The above message was for the following BY group:
EXPOSURE=Initial TREAT=High DN=Diurnal
NOTE: The GLIMMIX procedure is modeling the probability that Value='1'.
WARNING: Pseudo-likelihood update fails in outer iteration 4.
NOTE: Did not converge.
NOTE: The above message was for the following BY group:
EXPOSURE=Initial TREAT=High DN=Nocturnal
NOTE: The GLIMMIX procedure is modeling the probability that Value='1'.
WARNING: Pseudo-likelihood update fails in outer iteration 3.
NOTE: Did not converge.
NOTE: The above message was for the following BY group:
EXPOSURE=Initial TREAT=Low DN=Diurnal
NOTE: The GLIMMIX procedure is modeling the probability that Value='1'.
WARNING: Pseudo-likelihood update fails in outer iteration 7.
NOTE: Did not converge.
NOTE: The above message was for the following BY group:
EXPOSURE=Initial TREAT=Low DN=Nocturnal
NOTE: The GLIMMIX procedure is modeling the probability that Value='1'.
WARNING: Pseudo-likelihood update fails in outer iteration 2.
NOTE: Did not converge.
NOTE: The above message was for the following BY group:
EXPOSURE=Prolong TREAT=Control DN=Diurnal
NOTE: The GLIMMIX procedure is modeling the probability that Value='1'.
WARNING: Pseudo-likelihood update fails in outer iteration 4.
NOTE: Did not converge.
NOTE: The above message was for the following BY group:
EXPOSURE=Prolong TREAT=Control DN=Nocturnal
NOTE: The GLIMMIX procedure is modeling the probability that Value='1'.
WARNING: Pseudo-likelihood update fails in outer iteration 2.
NOTE: Did not converge.
NOTE: The above message was for the following BY group:
EXPOSURE=Prolong TREAT=High DN=Diurnal
NOTE: The GLIMMIX procedure is modeling the probability that Value='1'.
WARNING: Pseudo-likelihood update fails in outer iteration 2.
NOTE: Did not converge.
NOTE: The above message was for the following BY group:
EXPOSURE=Prolong TREAT=High DN=Nocturnal
NOTE: The GLIMMIX procedure is modeling the probability that Value='1'.
WARNING: Pseudo-likelihood update fails in outer iteration 3.
NOTE: Did not converge.
NOTE: The above message was for the following BY group:
EXPOSURE=Prolong TREAT=Low DN=Diurnal
NOTE: The GLIMMIX procedure is modeling the probability that Value='1'.
NOTE: Did not converge.
NOTE: The above message was for the following BY group:
EXPOSURE=Prolong TREAT=Low DN=Nocturnal
NOTE: PROCEDURE GLIMMIX used (Total process time):
real time 49.51 seconds
cpu time 40.78 seconds
So I decided to remove the last RANDOM statment and used the following model:
PROC GLIMMIX DATA=HABITAT; BY EXPOSURE TREAT DN;
CLASS ID YEAR TREAT HABVALUE;
MODEL VALUE (EVENT = '1') = HABVALUE ROAD_LOG ELEVATION ELEVATION_QUAD SLOPE SLOPE_QUAD / DIST=BINARY LINK=LOGIT SOLUTION;
RANDOM ID ID(TREAT) TREAT YEAR / TYPE=VC;
RUN;
I received the message: "Estimated G Matrx is not positive definite" again and you mentioned that there are some ways to tweek the model to eliminate this error. Additionally, I only had one Warning: Pseudo-likelihood update fails in outer iteration 3. NOTE: Did not converge. See rest of LOG Output below:
NOTE: The GLIMMIX procedure is modeling the probability that Value='1'.
NOTE: Convergence criterion (PCONV=1.11022E-8) satisfied.
NOTE: The above message was for the following BY group:
EXPOSURE=Initial TREAT=Control DN=Diurnal
NOTE: The GLIMMIX procedure is modeling the probability that Value='1'.
NOTE: Convergence criterion (PCONV=1.11022E-8) satisfied.
NOTE: Estimated G matrix is not positive definite.
NOTE: The above message was for the following BY group:
EXPOSURE=Initial TREAT=Control DN=Nocturnal
NOTE: The GLIMMIX procedure is modeling the probability that Value='1'.
WARNING: Pseudo-likelihood update fails in outer iteration 3.
NOTE: Did not converge.
NOTE: The above message was for the following BY group:
EXPOSURE=Initial TREAT=High DN=Diurnal
NOTE: The GLIMMIX procedure is modeling the probability that Value='1'.
NOTE: Convergence criterion (PCONV=1.11022E-8) satisfied.
NOTE: Estimated G matrix is not positive definite.
NOTE: The above message was for the following BY group:
EXPOSURE=Initial TREAT=High DN=Nocturnal
NOTE: The GLIMMIX procedure is modeling the probability that Value='1'.
NOTE: Convergence criterion (PCONV=1.11022E-8) satisfied.
NOTE: The above message was for the following BY group:
EXPOSURE=Initial TREAT=Low DN=Diurnal
NOTE: The GLIMMIX procedure is modeling the probability that Value='1'.
NOTE: Convergence criterion (PCONV=1.11022E-8) satisfied.
NOTE: Estimated G matrix is not positive definite.
NOTE: The above message was for the following BY group:
EXPOSURE=Initial TREAT=Low DN=Nocturnal
NOTE: The GLIMMIX procedure is modeling the probability that Value='1'.
NOTE: Convergence criterion (PCONV=1.11022E-8) satisfied.
NOTE: Estimated G matrix is not positive definite.
NOTE: The above message was for the following BY group:
EXPOSURE=Prolong TREAT=Control DN=Diurnal
NOTE: The GLIMMIX procedure is modeling the probability that Value='1'.
NOTE: Convergence criterion (PCONV=1.11022E-8) satisfied.
NOTE: Estimated G matrix is not positive definite.
NOTE: The above message was for the following BY group:
EXPOSURE=Prolong TREAT=Control DN=Nocturnal
NOTE: The GLIMMIX procedure is modeling the probability that Value='1'.
NOTE: Convergence criterion (PCONV=1.11022E-8) satisfied.
NOTE: Estimated G matrix is not positive definite.
NOTE: The above message was for the following BY group:
EXPOSURE=Prolong TREAT=High DN=Diurnal
NOTE: The GLIMMIX procedure is modeling the probability that Value='1'.
NOTE: Convergence criterion (PCONV=1.11022E-8) satisfied.
NOTE: Estimated G matrix is not positive definite.
NOTE: The above message was for the following BY group:
EXPOSURE=Prolong TREAT=High DN=Nocturnal
NOTE: The GLIMMIX procedure is modeling the probability that Value='1'.
NOTE: Convergence criterion (PCONV=1.11022E-8) satisfied.
NOTE: Estimated G matrix is not positive definite.
NOTE: The above message was for the following BY group:
EXPOSURE=Prolong TREAT=Low DN=Diurnal
NOTE: The GLIMMIX procedure is modeling the probability that Value='1'.
NOTE: Convergence criterion (PCONV=1.11022E-8) satisfied.
NOTE: Estimated G matrix is not positive definite.
NOTE: The above message was for the following BY group:
EXPOSURE=Prolong TREAT=Low DN=Nocturnal
NOTE: PROCEDURE GLIMMIX used (Total process time):
real time 14.18 seconds
cpu time 10.98 seconds
Based on the new SAS code using a more simplistic approach to get away from 3-way interactions, I think I'm getting closer. Overall, I want to estimate the probability of animals selecting different resources within each treatment. Additionally, I'm not missing any data in the entire dataset. Do you have any additional suggestions based on where I'm at now? Thank you very much for all of your input!
... View more