Test against non-Poisson (in-)homogeneous models
Functions to automate testing of 'univariate' and 'bivariate' point pattern hypothesis against non-Poisson (in-)homogeneous models.
Kci(mod1, mod2, correction="trans", nsim=99, ngrid=200, nrep=1e+05, r=NULL, simu="both", spctype=1) Ki(mod1, correction="trans", nsim=99, ngrid=200, nrep=1e+05, r=NULL, spctype=1) ## S3 ploth method for objects of class 'ecespa.kci': ## S3 method for class 'ecespa.kci' plot(x, type=1, q=0.025, kmean=TRUE, add=FALSE, maine=NULL, xlabe=NULL, ylabe=NULL, xlime=NULL, ylime=NULL, lty=c(1,2,3), col=c(1,2,3), lwd=c(1,1,1), ...)
mod1: Fitted model. An object of class ppm or ecespa.minconfit.
mod2: Fitted model. An object of class ppm or ecespa.minconfit.
correction: A character item selecting any of the options "border", "bord.modif", or "translate". It specifies the edge correction to be applied when computing K-functions.
nsim: Number of simulated point patterns to be generated when computing the envelopes.
ngrid: Dimensions (ngrid by ngrid) of a rectangular grid of locations where predict.ppm
would evaluate the spatial trend of the fitted models.
nrep: Total number of steps (proposals) of Metropolis-Hastings algorithm that should be run by rmh
to simulate models of class ppm.
r: Numeric vector. The values of the argument at which the functions should be evaluated.
simu: A character item indicating if both models will be simulated for the computation of the envelopes (simu = "both") or just the second model (simu != "both").
spctype: Type of 'pre-thinning' method employed by rIPCP in the simulation of ecespa.minconfit models.
x: An object of class 'ecespa.kci'. The result of runing Kci or Ki.
type: What to plot. One of 1 (K1), 2 (K2), 12 (K12), 21 (K21), 112 (K1-K12) or 221 (K2-K21).
q: Quantile for selecting the simulation envelopes.
kmean: Logical. Should the mean of the simulated envelopes be ploted?
add: Logical. Should the kci.plot be added to a previous plot?
maine: Title to add to the plot.
xlabe: Text or expression to label the x-axis.
ylabe: Text or expression to label the y-axis.
xlime: Max and min coordinates for the x-axis.
ylime: Max and min coordinates for the y-axis.
lty: Vector with the line type for the estimated Kmm function, the simulated envelopes and the mean of the simulated envelopes.
col: Vector with the color for the estimated K-function, the simulated envelopes and the mean of the simulated envelopes.
lwd: Vector with the line width for the estimated K-function, the simulated envelopes and the mean of the simulated envelopes.
...: Additional graphical parameters passed to plot.
These functions are designed to automate the testing of 'univariate' and(/or) 'bivariate' point pattern hypotheses (based on K-functions) against non-Poisson (in-)homogeneous models. These non-Poisson (in-)homogeneous models should have been fitted with pseudolikelihood tools (spatstat ppm function) or with minimum contrast methods (ecespa.minconfit).
Function Ki is designed to test 'univariate' hypotheses. It will compute the (in-)homogeneous K-function (using spatstat Kinhom function) of the point pattern to which the ppm or ecespa.minconfit model has beeen fitted and will compute 'envelopes' simulating from the fitted model. The computed envelopes can be considered as a pointwise test of the point pattern been a realisation of the fitted model.
Function Kci is designed to test 'bivariate' hypotheses. It will compute the (in-)homogeneous cross K-function (using spatstat Kcross.inhom function) and will compute envelopes simulating from the fitted models. As, when dealing with inhomogeneos patterns != , Kci will compute both functions. If simu = "both" (default option), Kci will simulate from mod2 to test K12 and from mod1 to test K21. If simu != "both", only mod2 will be simulated. This option may be useful when only K12 is of interest. Function Kci will also compute univariate (in-) homogeneous K-functions and envelopes for each individual point pattern.
The S3 ploth method will plot the selected K-function and envelopes (actually, it will plot the most usual L-function = ). The appropriate K function can be selected with the argument type. If type = 1 (default option), it will plot the univariate K function (of the analized model in Ki or of the first model [mod1] in Kci). If type = 2, it will plot the univariate K function of the second model (mod2 in Kci). When type = 12 or type = 21, it will plot respectively K12 or K21. Options type = 112 or type = 221 will graph a kind of 'segregation test' (see K1K2), and will represent de differences K1-K12, K2-K21 and their envelopes.
Both Kci and Ki return an object of class 'ecespa.kci', basically a list with the following items:
r: Numeric vector. The values of the argument at which the functions have been evaluated.
kia: Numeric vector. Observed (in-)homogeneous K function for mod1 point pattern.
kib: Numeric vector. Observed (in-)homogeneous K function for mod2 point pattern.
kci.ab.o: Numeric vector. Observed (in-) homogeneous cross K-function (K12) for mod1 and mod2 point patterns.
kci.ba.o: Numeric vector. Observed (in-) homogeneous cross K-function (K21) for mod2 and mod1 point patterns.
kci.ab.s: Matrix of simulated (in-) homogeneous cross K-function (K12) for mod1 and mod2 point patterns.
kci.ba.s: Matrix of simulated (in-) homogeneous cross K-function (K21) for mod2 and mod1 point patterns.
kib.s: Matrix of simulated (in-)homogeneous K function for mod2 point pattern.
kia.s: Matrix of simulated (in-)homogeneous K function for mod1 point pattern.
datanamea: Name of mod1 point pattern.
datanameb: Name of mod2 point pattern.
modnamea: Name of model mod1.
modnameb: Name of model mod2.
type: Type of analysis. "Kci" or "Ki".
De la Cruz, M. and Escudero, A. 2008. Null models and tools for multivariate heterogeneous point patterns. Submitted.
De Soto, L., Olano, J.M., Rozas, V. and De la Cruz, M. 2009. Release of Juniperus thurifera woodlands from herbivore-mediated arrested succession in Spain. Applied Vegetation Science, 13: 1525. tools:::Rd_expr_doi("10.1111/j.1654-109X.2009.01045.x")
Marcelino de la Cruz Rot
As this implementation involves the use of images as the means of evaluation of the (inhomogeneous) spatial trend, and a mask based on those images will be used as the point pattern window, the "Ripley's" or "isotropic" edge correction can not be employed.
It is usual that during the simulation process some warnings are produced. They are related to some simulated points being rejected as lying outside the specified window.
Even when one of the two point patterns is assumed to be homogeneous Poisson (and, apparently not worth of fitting any model), an homogeneous Poisson model can be easily fitted and passed to Kci with ppm. See the examples.
data(urkiola) # set the number of simulations (nsim=199 or larger for real analyses) nsim<- 5 ##################### ## univariate case # get univariate pp I.ppp <- split.ppp(urkiola)$birch # estimate inhomogeneous intensity function I.lam <- predict (ppm(I.ppp, ~polynom(x,y,2)), type="trend", ngrid=200) # Compute and plot envelopes to Kinhom, simulating from an Inhomogeneous # Poisson Process: I2.env <- envelope( I.ppp,Kinhom, lambda=I.lam, correction="trans", nsim=nsim, simulate=expression(rpoispp(I.lam))) plot(I2.env, sqrt(./pi)-r~r, xlab="r (metres)", ylab= "L (r)", col=c(1,3,2,2),legend=FALSE) # It seems that there is short scale clustering; let's fit an Inhomogeneous # Poisson Cluster Process: I.ki <- ipc.estK(mippp=I.ppp, lambda=I.lam, correction="trans") # Compute and plot envelopes to Kinhom, simulating from the fitted IPCP: Ipc.env <- Ki(I.ki, correction="trans", nsim=nsim, ngrid=200) plot (Ipc.env, xlab="r (metres)", ylab= "L (r)") ## Not run: ##################### ## bivariate case: test independence between birch and quercus in Urkiola J.ppp <- split.ppp(urkiola)$oak # We want to simulate oak from a homogeneous Poisson model: J.ppm <- ppm(J.ppp, trend=~1, interaction=Poisson() ) IJ.env <- Kci (mod1=I.ki, mod2=J.ppm, nsim=nsim) plot(IJ.env, type=12) plot(IJ.env, type=21) ## End(Not run)
Useful links