09-01-2015
tdlang
Calcite | Level 5
Member since
12-10-2013
- 4 Posts
- 0 Likes Given
- 0 Solutions
- 0 Likes Received
-
Latest posts by tdlang
Subject Views Posted 2996 12-16-2013 05:28 PM 2996 12-12-2013 12:25 PM 2996 12-11-2013 12:57 PM 3468 12-10-2013 06:30 PM -
Activity Feed for tdlang
- Posted Re: GLIMMIX spatial autocorrelation, syntax on SAS Procedures. 12-16-2013 05:28 PM
- Posted Re: GLIMMIX spatial autocorrelation, syntax on SAS Procedures. 12-12-2013 12:25 PM
- Posted Re: GLIMMIX spatial autocorrelation, syntax on SAS Procedures. 12-11-2013 12:57 PM
- Posted GLIMMIX spatial autocorrelation, syntax on SAS Procedures. 12-10-2013 06:30 PM
12-16-2013
05:28 PM
Not much progress here. Even after including the group=layer option, the sp(exp) term for each group is still 1. I double checked that there are no duplicated coordinates within the Name*Layer combinations. As for data richness in N, each Name*Layer combination has at least 1 observation (usually many, many more, dataset is 37,000 observations). I can see how those groups with a single observation would cause a problem, but I would think that since I have tried to "partition" the calculations, the problem would be confined to only those groups with a single observation. I may be mistaken, as I don't know exactly how SAS is computing this giant covariance matrix. Even when I restrict the analysis to those Name*Layer combinations with at least 3 observations, I have the same problem. Besides this, I cannot think of anything else that might be causing the problem. I am open to suggestions and, as always, appreciative of your comments.
... View more
12-12-2013
12:25 PM
Thanks again Steve. I have given this a run as well, looks to increase computing time dramatically so I am still waiting on output (11hrs and counting!) but I will report back once it finishes. From reading about the options again, this seems to make sense. I appreciate your help putting these things in a context I can understand.
... View more
12-11-2013
12:57 PM
Thank you very much Steve. I gave that a shot and it does run, however in the resulting "Covariance Parameter Estimates," the spatial term comes back with an estimate of 1.000 and no standard error. Cov Parm Subject Estimate Standard Error Variance Name 0.1279 0.04371 SP(EXP) Name 1.0000 . Residual 8.4454 0.06207 This yields a model that is no different from one that was run without the type= option included. Could this be due to the easting/northing duplication you mentioned? It is possible that there is duplication of easting/northing pairs: Within a Survey (since grid cells are in multiple layers, stacked cells would receive the same easting/northing) Can the type=sp option handle three dimensions? I think that would solve this problem. Between surveys (since surveys were conducted in the same locations multiple times) I appreciate the response very much, one step closer!
... View more
12-10-2013
06:30 PM
Hello all, First question I have posted, so please let me know if addition information is needed beyond that provided below: I am looking at a dataset of fish density values, recorded over 3 years at 3 different locations. These density values are obtained during surveys that are composed of multiple transect lines at a single site Transect "curtains" are further divided into grids (distance x depth), of which each cell is assigned a density and coordinate Additional parameters (Bottom habitat, salinity, temperature, etc) are assigned to the cells The density values have a very very skewed distribution (many small, few very large), hence the use of GLIMMIX to attempt to address this. I am interested in the effects of these "additional parameters" on density, though I cannot ignore the spatial nature of the data (fish group together, nearby cells will tend to have more similar densities). I am assuming that spatial correlations would only exist within a single survey. My model currently looks like this: proc glimmix data=Data.BankPresent pconv =1e-5 maxopt=30; class Site Habitat Season Survey; model Density= Relief DShallow MeanSalinity MeanTemp Site|Habitat|Season / dist=gamma link=log; random Survey / subject=intercept type=sp(exp) (Easting Northing); run; I am unsure of where to put the spatial autocorrelation component [type=sp(exp) (Easting Northing)], or if this model is doing what I have attempted to describe above. If needed/of interest, the variables included are: Relief: continuous, vertical relief of the seafloor below cell i DShallow: continuous, distance from cell i to the shallowest point at the site MeanSalinity: continuous, salinity of cell i MeanTemp: continuous, temperature of cell i Site: categorical, site at which the survey was conducted (1,2,3) Habitat: categorical, habitat classification of seafloor below cell i (1,2,3,4,5) Season: categorical, season in which the survey was conducted (1,2,3,4) To reiterate, I want to run the above model with a consideration for spatial autocorrelation on a survey-by-survey basis. Thank you for any information you may be able to provide, applicable references, or questions of clarification.
... View more