Suppose I have a log Gaussian Cox process (LGCP) $X$ with log intensity function $\lambda(x)=S(x)$ where $S$ follows a Gaussian process. Since LGCP still falls under the umbrella of inhomogeneous Poisson processes, the likelihood of an observed point process $(x_1,...,x_n)$ should be proportional to the following density

$$ p(X|S) \propto \prod_{i=1}^n \exp[S(x_i)] \exp[-\exp[S(x_i)]] $$

However I've seen an alternative formulation which replaces the double exponent term dependent on $i$ with a constant equal to the average number of events in the study region.

$$ p(X|S) \propto \prod_{i=1}^n \exp[S(x_i)] \exp\left( -\int_\mathcal{D} \exp[S(x)]dx \right) $$

Which one is correct?


1 Answer 1


Found the answer in 4.168 of "Statistics for Spatio-temporal data" by Cressie and Wikle. The information in $X$ represents both $n$ the number of points observed AND the exact locations of each of those points. So the likelihood is technically that of $[Z(D_x), \{x_1,...,x_{Z(D_x)}\}]$.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.