Notes
Course introduction and motivation
for the INLA course in Iceland 2016
Haakon Chris. Bakka
Norwegian University of Science and Technology
Why are we here? What is going to happen?
Wide time estimate: 1h 20min presentation, 1h exercises.
I’m Haakon from Trondheim ...
Notes
And I want to talk about INLA
Notes
R-INLA and Bayesian hierarchical modeling
I want to show you this amazing framework!
I Easier to communicate model and results
I Compute anything from your model
I Fast inference
I Powerful R-package
What is INLA?
Notes
The short answer:
INLA is a fast method to do Bayesian inference with latent
Gaussian models and INLA is an R-package that implements this
method with a flexible and simple interface.
The first paper:
Rue, Martino, and Chopin (2009) “Approximate Bayesian
inference for latent Gaussian models by using integrated
nested Laplace approximations.” Journal of the Royal
Statistical Society: Series B. 319–392.
What is INLA?
Notes
INLA is a community
INLA is a community of statistical researchers and applied
researchers, where you can run many applied models, get high
quality modeling advice, and very good technical help.
”The most useful thing about R-INLA is the discussion
group”
That was just a teaser
Notes
Who are you?
1. Found INLA useful; really interested in learning more
2. Heard about it; really interested in learning more
3. Know quite a lot about it; want to learn more / see another
viewpoint
4. Heard about it and giving it a shot
5. Someone said I needed it for my research
6. Someone told me to be here
I now want to sell you R-INLA
Notes
What kind of problems have other people used INLA for?
Notes
See inla review papers ...
Oh, wow, that’s a lot of references
For example, my favourite review paper
[Link]
Malaria Nature paper
Notes
Air quality report by WHO
Notes
[Link]
Notes
What kind of problems have I used INLA for?
Studied a rainforest
Notes
Analysing the reproductive habitats of Smelt in the Finnish
Notes
Archipelago
Created a new set of models
Notes
[Link]/barrier-model
Buzzwords
Notes
We do...
Machine learning!
Robustness!
Data mining!
Big Data!
Optimal decisions!
Space-time models!
Regularisation!
Complex models!
Sensitivity analysis!
Example use of R-INLA
Notes
> demo("Epil")
Notes
And now: The three coolest ideas!
Bayesian hierarchical modeling
Notes
Is super useful!
I One-Slide models
I A simple framework for inference
I Unifies model fitting and prediction and uncertainty
I Adapts to the data
I Also useful with very little data
I Very reasonable multivariate uncertainty
I Very reasonable joint predictions
I Priors and pooling take care of overfitting
But it is not magical!
Sparse matrices (GMRF)
Notes
Gaussian Markov Random Fields
Are necessary, not optional!
I Precision matrix Q is Σ−1
I E.g. for 104 observations of a time series, from N = 108 to
N = 104 .
I Sparse precision matrices are not ”approximations”, but
natural models
I In numerics you use either ”fourier-transform-tools” or sparse
matrices
I Cholesky: Typically ∝ N 3/2 where N is the number of nonzero
entries
Working with them is a bit technical.
Laplace approximation
Notes
The posterior contribution from any observation likelihood is almost Gaussian!
E.g. if the likelihood is poisson
I An observation likelihood represents ”for each value that can
be observed, what is the observation variance around that
value”
I Conditional on the data values, this is near-Gaussian
> N = 1E4; llam = runif(N, min=-4, max=4);
> y = rpois(N, exp(llam)); llam.p = llam[(y==1)];
> plot(density(llam.p))
I Especially if your underlying model is Gaussian
> N = 1E6; llam = rnorm(N, sd=1);
I This approximation is less influential than most other aspects
of modeling
Milestone
Notes
Previous Why are we here?
Next How this course are going to be
How we talk about what we talk about
Notes
Specific to generic
l Messy real world applications
l Solving practical issues
l Specific
Algorithms and tools
Inference procedures
Ω Abstraction
Ω Mathematical properties
Ω Rules of science
The program for this course
Notes
Part 1 Practical INLA introduction
Practical coding, simulation, model setup, prediction,
Part 2 Theoretical INLA ideas and understanding
How does it work? Why is it nice? For what? Numerics & cool ideas!
Part 3 A framework for good modeling
Interpretation, what models to fit, more interpretation
Part 4 Advanced features
Content depending on time and interests
Problem set 1 Many models to run
Practical coding; parallell with the course
Problem set 2 Many questions to answer
Explain or discuss; parallell with the course
Presenter’s goals
Notes
1. You know why I am talking about [what I am talking about]
2. You feel that you learn many useful things
Practical
Notes
I Link to course files and gdocs
I Other things? Anyone? Questions?
I After the course, use bakka@[Link] or the R-INLA
community at [Link]
Attribution
Notes
Special thanks goes to
SPDE paper Making me embark on a PhD in statistics
Håvard Rue The R-INLA code
Andrea Riebler The slides on practical INLA
Håvard Rue The slides on INLA theory and GMRFs
R-INLA community Too much to say
Summary: Why are we here?
Notes
Last slide of the introduction
1. We want to do Bayesian (hierarchical) modeling
2. Rue et. al. developed a fast algorithm INLA
3. I am an expert in INLA
4. Birgir asked nicely
5. You want to learn INLA, I hope!
Exercise
Notes
Exercise 1: Problem 1: All points.
Also: Experimentation