[R-sig-Geo] Anova and confidence intervals for kppm models in spatstat

[Rolf Turner]

On 13/01/15 08:41, ASANTOS wrote:
> Dear Members,
>
>               I have two questions, first if there are approaches like
> anova.ppm for comparing a clustered point process models kppm in
> spatstat package (example above)? And second the same question for
> estimate the confidence intervals for kppm model parameters, like
> confint() function?
>
> data(bei)
> ## Model with elevation
> fit1 <- kppm(bei, ~elev, "Thomas", covariates = bei.extra)
> fit1
> ##Null model
> fit.null <- kppm(bei, ~1, "Thomas", covariates = bei.extra)
> fit.null
> #Comparing with ANOVA
> anova.ppm(fit1,fit.null, test="Chi")
> #but don't have anova.kppm?
> ##Confidence intervals
> confint(fit1)

The short answer is "no".  A somewhat longer answer:  If the method of 
minimum contrast is used, there is no analogue of the log likelihood 
ratio available and consequently no form of "anova" based on the 
likelihood ratio test could possibly be done.  If either of the other 
two methods ("clik2" or "palm") were used then at least conceivably some 
form of anova *could* be done, but this possibility is not yet implemented.

Further theoretical development is needed. Adjustment of the analogue of 
the likelihood ratio is required in order for this statistic to have the 
required chi-squared null distribution, and the theory for this 
adjustment needs to be worked out.  This is on our "to-do list" but 
probably will not happen soon.

The necessary theory underlying the use of anova.ppm() for Gibbs models 
has only recently become available.  See the references given in the 
help for anova.ppm.  Previously anova could only be applied to *Poisson* 
models fitted to point patterns.

Until further development is carried out you will have to get by on 
Monte Carlo methods.

cheers,

Rolf Turner

-- 
Rolf Turner
Technical Editor ANZJS
Department of Statistics
University of Auckland
Phone: +64-9-373-7599 ext. 88276
Home phone: +64-9-480-4619