[R-sig-Geo] Error in Moran eigenvector GLM filtering ("base correlation higher than alpha")

[Roger Bivand]

On Wed, 18 Mar 2009, Jamie Mihoko Doyle wrote:

> Hi everyone,
>
> I could really use your help. I'm doing a study of homicide mortality among 
> Foreign Born and US Born Hispanics in the U.S. and using Moran eigenvector 
> filtering for the Poisson models. The data are from US counties having 
> populations greater than 100,000 people since data from smaller counties is 
> not publicly available.
>
> I seem to get the following error when I'm running the model for Foreign Born 
> Hispanics only: "base correlation higher than alpha".

You see this message when the observed autocorrelation is so weak that no 
eigenvectors are chosen for inclusion for the value of alpha given (here 
0.2). You could increase alpha, or, better, just fit a regular glm(), 
because autocorrelation does not seem to be a problem in this case. I say 
"does not seem to be", because you would need to examine the residuals of 
the glm() fit carefully. You may also be running into negative 
autocorrelation (ME() assumes positive), but I haven't looked at your data 
(yet).

Hope this helps,

Roger

>
> Is there a more appropriate method I should use?
>
> Please find my code below. The data that I am using can be found here: 
> http://www.jamiemihoko.com/Rquestions.htm
>
> I would greatly appreciate any help.
>
> Thank you in advance for your time,
>
> Jamie
>
> ######
> ##Data
> ######
> # Read in flat data
> d <- read.csv("D:/Vital Statistics Mortality/R Version of 
> Dataset/homic_mortality_v14_R.csv",header=TRUE)
>
> #Read in shape file
> homic<-readOGR("D:/Vital Statistics Mortality/ArcGIS Shapefiles", 
> layer="mortality_data2_edited_Disso3")
>
> # Store Coordinates
> coords<-coordinates(homic)
>
> ##################################
> # Create Nearest Neighbor Weights
> ##################################
> homic.nn.cont<- poly2nb(homic)
> homic.nn.cont.W<-nb2listw(homic.nn.cont, zero.policy=TRUE)       # Need to 
> use "zero.polcy=TRUE" because some areas do not have neighbors
>
> ##############################################
> # Need to round numbers (5-year average of homicide deaths)
> ##############################################
> d_HispFore.r <- round(homic$d_HispFore)
>
> ###################
> # Spatial filter ###################
>         sf.HF <- ME(d_HispF5.r  ~ mindistCA2 + median_age + pct_single + 
> pct_male2 +  pct_ling_i +  pct_povert +  pop_densit,
>                                    data = homic, family = 'poisson', listw = 
> homic.nn.cont.W, alpha = 0.2, verbose = TRUE)
>
> # Data Dictionary- All county-level:
>  # d_HispF5.r :     Foreign Born Hispanic Deaths
>  # mindistCA2:     Minimum distance from county, population-weighted 
> centroids to the U.S.-Mexico border
>  # median_age:     Median age of county residents
>  # pct_male2:       % male residents
>  # pct_ling_i:        % linguistically isolated
>  # pct_povert:      % poverty
>  # pop_densit:      Population density
>
>

-- 
Roger Bivand
Economic Geography Section, Department of Economics, Norwegian School of
Economics and Business Administration, Helleveien 30, N-5045 Bergen,
Norway. voice: +47 55 95 93 55; fax +47 55 95 95 43
e-mail: Roger.Bivand at nhh.no