37 Model interpretation
We have considered a sequence of increasingly refined constructions for defining a Cox process. Beginning with the most general non-central formulation, we examined its structural properties, identified limitations in interpretability, and progressively imposed restrictions to obtain simpler and more tractable models.
This process has revealed a trade-off between generality, interpretability, and modelling flexibility, which we now summarise and use to guide the selection of a preferred construction.
37.1 Summary of constructions
We have considered three primary constructions:
The non-central model: \[ \Lambda(u) = Z(u)^2, \quad \mathbb{E}[Z(u)] = m(u) \]
The centered model: \[ \Lambda(u) = Z(u)^2, \quad \mathbb{E}[Z(u)] = 0 \]
The shifted model: \[ \Lambda(u) = \mu + Z(u)^2, \quad \mathbb{E}[Z(u)] = 0 \]
Each of these models defines a valid Cox process, but differs in terms of interpretability and flexibility.
37.2 Interpretation of the constructions
Each construction corresponds to a different interpretation of how spatial variation in intensity is generated.
In the non-central model, the intensity comes from a quadratic transformation of a Gaussian field with non-zero mean. This leads to interaction between the mean and covariance structure, resulting in a pair correlation function containing both linear and quadratic terms in the correlation function. While flexible, the interaction complicates interpretation, as it is difficult to untangle the effects of first- and second-order structure.
In the centered model, the intensity is driven purely by stochastic fluctuations of a mean-zero Gaussian field. This leads to a particularly simple form of the pair correlation function, depending only on \(\rho(r)^2\). However, this simplicity comes at a cost: the clustering strength is fixed at \(g(0) = 3\), and cannot be adjusted through model parameters.
In the shifted model, the intensity is decomposed into a deterministic baseline component \(\mu\) and a stochastic component \(Z(u)^2\). This provides a natural interpretation as a signal-plus-noise model, where \(\mu\) represents background intensity and \(Z(u)^2\) introduces spatial clustering.
Looking back on this, one thing I forget to mention is that a potential benefit of the non-central over the shifted, is that the shifted model can only have intensity as low as the value of \(\mu\), whereas the non-central allows intensity to drop all the way to 0.
37.3 Comparison of structural properties
The key differences between the constructions can be summarised as follows:
- Interpretability:
- Non-central: difficult due to mixed terms
- Centered: simple and interpretable
- Shifted: simple and interpretable
- Clustering behaviour:
- Non-central: influenced by both mean and covariance
- Centered: fixed clustering strength (\(g(0) = 3\))
- Shifted: clustering strength is parameter-dependent
- Flexibility:
- Non-central: flexible within a bounded range
- Centered: highly restrictive
- Shifted: flexible within a bounded range
- Analytical tractability:
- Non-central: slightly cumbersome
- Centered: squeaky clean
- Shifted: retains simplicity while improving flexibility
37.4 Selection of the canonical construction
From the above comparison, it is clear that the centered model, while elegant, is too restrictive for practical use, as it does not allow control over clustering strength. The non-central model, although more general, suffers from a lack of interpretability due to the interaction between mean and covariance structure.
The shifted model strikes a balance between these two extremes. It preserves the analytical simplicity of the centered construction, while introducing sufficient flexibility to model a range of clustering behaviours through the parameter \(\mu\).
For this reason, we adopt the shifted construction as the canonical form of the single-component CSCP.
37.5 Final definition
Throughout the remainder of this work, we define the single-component Chi-square Cox process (CSCP) as the Cox process with random intensity
\[ \Lambda(u) = \mu + Z(u)^2, \]
where \(\mu \ge 0\) and \(Z(u)\) is a mean-zero Gaussian random field with covariance function \(C(u,v)\).
So I’m kind of fine with this part of the thesis overall, however there is one thing kind-of bothering me:
We really only properly address 3 possible constructions of the CSCP.
I’m fine with our justification as to why the construction we selected from the three is preferred, however it is not exactly clear to me why the one we have settled on might be the best of ALL possible constructions. For example, what if a scaled and shifted CSCP is better, \(\mu + \alpha Z(u)^2\)?
(I guess this is just the same as shifted).
Or any other possible transformation?
I can understand if instead we are trying to go along with something like: “we have been trying a bunch of things, and out of the stuff we tried, this seemed to work the best”, but it does seem a little loose, no?
What other possible transformations are there, involving \(Z(u)^2\)?
(I understand that we want to preserve the fact that marginal distribution of location is Chi-square, so this limits the number of possibilities I guess).
Suggested?
While other transformations of Gaussian random fields could be considered, the quadratic form represents the simplest construction leading to a chi-square-type intensity. More complex transformations introduce additional parameters or nonlinearities that complicate both interpretation and inference. The shifted quadratic form therefore provides a natural balance between simplicity and modelling flexibility within this class of processes.
Also, I definitely should have mentioned that we decided to primarily use the \((\lambda, \phi)\) parameterization, and justify why.