I'm working on my dissertation and my data has missing data. The missing data isn't horrible (6%-12% on a few variables), but I am imputing the data to restore some of the lost variance. After exploring the data and refining the model, I am satisfied with the output for proc MIANALYZE.
My question concerns the next steps in my analysis. I am performing geographically weighted regression (GWR) in ArcMap, so I can't analyze the pooled imputed data sets with a simple command like "By _imputation_". Unfortunately, this problem means I would have to perform X GWR's, where X is equivalent to the number of imputations, and pool the individual results. I could follow Little and Rubin (1987) methodology for manually combining data (i.e., averaging beta's and using matrix algebra to compute covariances). Does anyone have the SAS code, so that I could upload output from each GWR result (e.g. covariance matrix's) and automate this process to obtain variance, df, and p-values?
The real problem is creating a pooled estimate for local and global Moran's I autocorrelation results across data sets. I would have no idea how to do that. Any suggestions?
General note: The data I'm imputing is at the client admission level and not the census tract level - or else I would use another technique (e.g., kriging) to estimate missing data.
Perhaps you can prove with other software, for example ILWIS if you have raster images.
It should be fairly easy in 'R' ( I am assuming that you are familiar to the R environment ). Try the packages 'SimComp', 'gstat' etc. In parallel, have a look into this example : http://www.public.iastate.edu/~maitra/stat501/Rcode/BoxMTest.R. You might find here what you have been looking for.
I also agree, the options in R present far more opportunities -- there is the package VIM (Visualization and Imputation of Missing Data), which has several methods for multiple imputation, some of which are model based -- but it sounds like you want to integrate spatial analysis into a model based approach...there is an excellent book on the subject, Applied Spatial Data Analysis with R by Bivand, Pebesma and Gomez-Rubio. Additionally the user group R-sig-geo is very active in this area...you'll probably want to switch from ArcGIS to QGIS or another open source GIS and possibly also introduce postgreSQL, but in the long run -- it will definitely be worth it
Stay tuned. We'll soon have a UCI Social Science Gateway site for this (Dow-Eff 2SLS)
It's a good question. I may have the same question in the coming months, so I will stay tuned.
Thank you for the responses; I will be integrating spatial analysis into a model-based approach. While I can code in SPSS and SAS, I'm not familiar with nuts and bolts of R. The book looks very interesting. Would you say it's written to introduce you to the language too?
Also, would a package in R help me to understand how to pool global and local Moran's I values? (I am unfamiliar with ILWIS.)
R has a bit of a front end on the learning curve, but with some time and patience you should be fine -- its not really a drop down menu environment, but you could look at an IDE like R Studio to try and ease the front end of the learning curve. It really just takes learning and doing, and I would guess after 2 to 3 months you will be rolling forward. You can search for a multitude of packages, and I am fairly certain there is one which uses Moran values, or you can look at the package as a guideline/model and then go back and recode it as you like. This is what I often do. If you are looking to do something like integrate spatial weighting into a model based approach, then you will probably have to cook something up yourself, but R is so powerful and the community is so big and supportive, the most important ingredients tend to be willpower and determination. Good luck!
oh I almost forgot -- a great resource for getting started with R is the Quick R website
http://www.statmethods.net
HTH!
- Badges
- Science topic
- Similar topics
- Geoscience
- Geoinformatics
- Geographic Information System