Fitting Bayesian models has gotten even better with the R package INLA. The fact that it’s integrated into R has led to more simplistic scripting, and the approach means much reduced computational time. There is a lot of literature out there about “why and how” the integrated nested Laplace approximation, which is why my treatment of it in this blog will likely be more practical. “The” paper about INLA for Bayesian stats starts out talking about the class of structured additive regression models. Fortunately, for non-statisticians like me, they throw me a bone and give me a list of examples from which I can pick out a few that I recognize.
- (generalized) linear models
- (generalized) additive models
- smoothing spline models
- state space models
- semiparametric regression
- spatial/temporal models
- log-Gaussian Cox processes
- geostatistical/additive models
The focus is on latent Gaussian models (about which there will be an entire conference this fall; the postdoc I work with and I have joked about going). As the name would imply, the latent field is Gaussian, “controlled by a few hyperparameters and with non-Gaussian response variable” (Rue et al. 2009).
Another cool feature is that you can run it remotely within your R script. With that general question of why my models crash or how the fit is going, I found a reference to error checking that might help me diagnose.