improve.kppm(object, type=c("quasi", "wclik1", "clik1"), rmax = NULL, eps.rmax = 0.01, dimyx = 50, maxIter = 100, tolerance = 1e-06, fast = TRUE, vcov = FALSE, fast.vcov = FALSE, verbose = FALSE, save.internals = FALSE)"kppm").
"clik1", "wclik1" and "quasi"
for, respectively, first order composite (Poisson) likelihood,
weighted first order composite likelihood and quasi-likelihood.
rmax
from the tail behaviour of the pair correlation function. Namely
rmax is the smallest value of $r$
at which $(g(r)-1)/(g(0)-1)$
falls below eps.rmax.
Ignored if rmax is provided.
"kppm".
"kppm"
object. If type="clik1" estimates are based on the first order
composite (Poisson) likelihood, which ignores dependence between the
points. Note that type="clik1" is mainly included for testing
purposes and is not recommended for the typical user;
instead the more efficient kppm
with improve.type="none" should be used.
When type="quasi" or type="wclik1" the dependence
structure between the points is incorporated in the estimation
procedure by using the estimated pair correlation function in the
estimating equation. In all cases the estimating equation is based on dividing the
observation window into small subregions and count the number of points
in each subregion. To do this the observation window is first
converted into a digital mask by as.mask where the
resolution is controlled by the argument dimyx. The
computational time grows with the cube of the number of subregions, so fine
grids may take very long to compute (or even run out of memory).
ppm,
kppm,
improve.kppm
# fit a Thomas process using minimum contrast estimation method
# to model interaction between points of the pattern
fit0 <- kppm(bei ~ elev + grad, data = bei.extra)
# fit the log-linear intensity model with quasi-likelihood method
fit1 <- improve.kppm(fit0, type="quasi")
# compare
coef(fit0)
coef(fit1)
Run the code above in your browser using DataLab