Base correlation larger than alpha (ME - Spdep package

Hello Roger,

Thanks a lot for your help. I fit a model as you said but with a binomial
response and I did ok! However, in ME model, I had to increase the alpha
value to 0.5.

Is it bad to have such a "high" alpha value?


The results for ME and Spatial Filtering models were a little bit different:
anova(glmerr12,test="Chisq")

Analysis of Deviance Table

Model: binomial, link: logit

Response: OCC_BIN

Terms added sequentially (first to last)


               Df Deviance Resid. Df Resid. Dev Pr(>Chi)
NULL                               81     984.43
DEPTH            1   322.27        80     662.17   <2e-16 ***
DISTCOAST        1   114.42        79     547.74   <2e-16 ***
SLOPE            1     0.00        78    2739.32        1
fitted(errcol2)  1   360.44        77    2378.88   <2e-16 ***


anova(glmerr1,test="Chisq")
Analysis of Deviance Table

Model: gaussian, link: identity

Response: OCC_BIN

Terms added sequentially (first to last)


               Df Deviance Resid. Df Resid. Dev  Pr(>Chi)
NULL                               81      20189
DEPTH            1   1916.3        80      18272 0.0006276 ***
DISTCOAST        1   5093.0        79      13179 2.484e-08 ***
SLOPE            1    557.4        78      12622 0.0651683 .
fitted(errcol1)  1      1.7        77      12620 0.9180173
---
Signif. codes:  0 ?***? 0.001 ?**? 0.01 ?*? 0.05 ?.? 0.1 ? ? 1


Both models showed significant relationship with DEPTH and DISTANCE TO
COAST and not with SLOPE. However, one model was significant for fitted
spatial autocorrelation (ME model) and the other did not.


When I tried to compare the results via anova (as done in spdep manual -
pag.118) for these binomial models, I could not obtain the p level. My code:
##Anova comparing both models
anova(glmerr12,glmerr1,test="Chisq")
Analysis of Deviance Table

Model 1: OCC_BIN ~ DEPTH + DISTCOAST + SLOPE + fitted(errcol2)
Model 2: OCC_BIN ~ DEPTH + DISTCOAST + SLOPE + fitted(errcol1)
 Resid. Df Resid. Dev Df Deviance Pr(>Chi)
1        77     2378.9
2        77    12620.3  0   -10241

##glmerr 12 is the ME model
##glmerr1 is the Spatial Filtering model


Can you help me (or others in this list) one more time?

Once again thanks a lot for everyone who can help me here.

Rodrigo


2013/6/10 Roger Bivand <Roger.Bivand at nhh.no>

On Mon, 10 Jun 2013, Rodrigo Tardin wrote:

 Dear list,

I am new to this list and I am trying to fit a Spatial Eigenvector GLM to
count data.

I am trying to fit a ME model in spdep package to investigate
environmental
drivers on Bryde's whale distribution. I have 3 environmental variables
(depth, distance to coast, slope), my response variable is the occurrence
(counts) of whales in 82 km2 grids.

I used the dnearneigh function to build .nb data using the coordinates
data.

   nb1  <-  dnearneigh(xy,  0,  10)  ----

and: lw1  <-  nb2listw(nb1,  style="W")

x - is latitude coordinates
y - is longitude coordinates


Using the glm and Spatial filtering function (as done in spdep manual) I
had no problems. However, when I used the 'ME' it says: base correlation
larger than alpha. My alpha is 0.05.

Here it is my code:
##This part works well
glmbase1  <-  glm(OCC  ~  DEPTH  +  DISTCOAST+SLOPE,  offset=RT,
data=bryde3)

lagcol1  <-  SpatialFiltering(OCC_BIN  ~  1,  ~  DEPTH  +  DISTCOAST+
SLOPE, data=bryde3,  style="W",  nb=nb1, alpha=0.05,  verbose=TRUE)

glmlag1  <-  glm(OCC ~  DEPTH  +  DISTCOAST+SLOPE +  fitted(lagcol1),
offset=RT,data=bryde3)

These are not the same models - lagcol1 is as you know the process in the
response (here OCC - RT??), not in the error. In ME, the process is in the
error, and it looks as though the covariates mop up much of the spatial
dependence (it is also a Poisson model with a default log link).
SpatialFiltering() does not handle an offset (as you have realised, I
think), so again, this is a possible source of difference. ME is testing
slightly diffferently too. Do:

errcol1  <-  SpatialFiltering(I(OCC-RT)  ~  DEPTH  +  DISTCOAST +

 SLOPE, data=bryde3,  style="W",  nb=nb1, alpha=0.05,  verbose=TRUE)
glmerr1  <-  glm(OCC ~  DEPTH  +  DISTCOAST+SLOPE +  fitted(errcol1),
 offset=RT, data=bryde3)

and

errcol2  <-  ME(OCC ~  DEPTH  +  DISTCOAST+SLOPE,  offset=RT, data=bryde3,
family="gaussian", listw=lw1,  alpha=0.05,  verbose=TRUE)
glmerr12  <-  glm(OCC ~  DEPTH  +  DISTCOAST+SLOPE +  fitted(errcol2),
 offset=RT, data=bryde3)

give similar results?

Hope this helps,

Roger

anova(glmlag1)
anova(glmbase1,  glmlag1)
set.seed(123)

##From here on, it does not
lagcol1  <-  ME(OCC ~  DEPTH  +  DISTCOAST+SLOPE,  offset=RT, data=bryde3,
family="poisson",
listw=lw1,  alpha=0.05,  verbose=TRUE)

Error in ME(OCC ~ DEPTH + DISTCOAST + SLOPE, data = bryde3, family =
"poisson",  :
 base correlation larger than alpha

When I increase alpha to 0.6 (lower values keep saying the same error),
this error message does not appear. However, when I try to rerun GLM with
eigenvectors included as covariates (as done in spdep manual - page 118) I
have the following error message:

Error in glm.fit(x = c(1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,  :
 NA/NaN/Inf in 'x'
warning message: step size truncated due to divergence

Can anyone please help me?
Other question, if I may, is there a strong difference between Spatial
Filtering and ME?

Many thanks in advance,
Rodrigo

--
Roger Bivand
Department of Economics, NHH Norwegian School of Economics,
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