Kernel Density Estimation

References

  1. Duong et al. "Closed-form density-based framework for automatic detection of cellular morphology changes" http://m.pnas.org/content/109/22/8382.full
  2. Kelsall, Diggle, "Non-parametric estimation of spatial variation in relative risk", Statistics in Medicine 14 (1995) 2335--2342

Similar to the use of likelihood, we might instead first use the actual events to form a density surface, using some KDE method, and then comparing this risk surface against the predicted one.

This is suggested in [1], in a medical imaging context (cellular morphology). The proposed way to compare two kernels is via integrated square error $$ T = \int \big|f(x) - p(x)\big|^2 $$ where $f$ is the kernel estimated from the actual events, and $p$ is the prediction. When both $f$ and $p$ are estimated from a fixed bandwidth KDE, it is possible to expand out the integral defining $T$ and proceed at least partly analytically, which is how [1] proceeds. In our case, $p$ will often not be given by an explicit KDE method. Furthermore, we also wish to consider the "edge correction" methods as suggested by [2].

Advantages

  • Clear interpretation

Disadvantages

  • Computationally expensive.
  • Very hard to choose the correct bandwidth for $f$, compounded by the small expected number of actual events.

In [ ]: