Current Research

TMB and R-INLA

While I was working to develop methods and code to handle very large geostatistical models, we started studying Tempalte Model Builder (TMB) as a viable alternative to Integrated Nested Laplace Approximations (R-INLA).

In addition to being quite fast and resource efficient, TMB can also be applied to a wider class of problems than R-INLA.

I’ve taken the time to really dig into the TMB algorithm and have compared it against the methods used in R-INLA. I also ran many many geostatistical simulations to compare the two in a variety of settings.

You can check out a working version of a paper that digs into this on arXiv.

A bit of code to get you started modeling the same continuous spatial models in both R-INLA and TMB is available in my Docs.

EU Cancer Mortality and Incidence Estimation

Across the European Union, different country members have different systems for recording the cancer mortality and incidence of their populations. Some record only national-level mortality data, while some record ssubnational cancer mortality and incidence occurence via registry systems, and many others have a system that falls in between. To complicate matters even further, most coountries have changed their data collection methods sometime in the last few decades.

Cancer incidence, the occurence of new cases, is generally more useful for planning and resource allocation when compared mortality records, which record deaths. On the other hand, mortality data is more quickly and more frequently available.

In this project I am jointly model cancer incidence and mortality, using the mortality-incidence ratio, across age-space-time. Then, we use more recently available mortality data to predict the recent incidence that is not yet available.

More notes on this project will soon be available in my Docs.

Cross-validation for spatial model assessment

Assessing the quality of random effects (REs) estimates can be tricky because (assuming the correct model) expected qualities of the REs, such as coverage, are true when averaged across the distribution of the predicted REs. As the magnitude of one aprticular true random effect varies around the mean of the REs, the expected coverage for a particular confidence or credible level is known to vary the nominal target level. This begs the question about what may be expected in leave-one-out (or other) cross-validation schemes for model assessment in spatial effects models.

In this project I’m studying the expected behavior of goodness-of-fit statistics from various cross-validation approaches for spatial models.