Removing the blindfold: visualising statistical models

Posted on by

Hadley Wickham’s talk for Monday 13 Sept at noon in the statistics dept:

As the volume of data increases, so to does the complexity of our models. Visualisation is a powerful tool for both understanding how models work, and what they say about a particularly dataset. There are very many well-known techniques for visualising data, but far fewer for visualising models. In this talk I [Wichkam] will discuss three broad strategies for model visualisation: display the model in the data space; look all members of a collection; and explore the process of model fitting, not just the end result. I will demonstrate these techniques with two examples: neural networks, and ensembles of linear
models.

Hey–this is one of my favorite topics!

9 thoughts on “Removing the blindfold: visualising statistical models”

  2. Or posted on his web site http://blog.had.co.nz/

    Already some very interesting material there.

    I used the split-apply-combine strategy here
    http://www.stat.columbia.edu/~cook/movabletype/ar…

    One question I have wanted to ask is if anyone has any sense of how often log priors, likelihoods and posteriors are plotted together?

    My guess is rarely – perhaps sometimes log priors and posteriors – but I don't really know.

    If anyone knows of published examples, it would be nice to hear about them.

    K?

  3. Keith:

    I prefer plotting probability densities, not log densities. We have an example of the plots you suggest on page 396 of ARM. It's for the radon example, and what really makes it work, for me, si that we show four different sets of prior/likelihood/posterior, one for each of four different counties. This emphasizes the hierarchical structure of Bayesian analysis and moves the discussion away from the sterile "What prior should I choose?" perspective.

  7. Thanks Andrew, I'll have a look.

    Scale of plot perhaps should meet the purpose, logs for showing addition of evidence and original scale for discerening where the probablities are, before and after.

    Of course what _someone_ _should_ do is identify a few relevant journals in an application area and search for Bayesian analyses in 2009 and count the number of times certain plots are supplied.

    Posted slides were different from what I was guessing they would be.

    As for showing the coefficients for all possible models, I once suggested that for an observational study of a treatment effect and was initially _overlooked_ as a co-author (the other statisticians boosted they could find the one best model).

    Fortunately the journal reviewers were agast at the overconfidence of the claims in the submitted paper and I was re-asked to be a co-author and we got a summary of all the treatment effect estimates in the paper.

    K?

  8. On the slides, Mr. Wickam decsribes neural networks as one way networks. The way it is pictured is that the lower levels communicate up to the highere levels. However, don't the higher level neurons communicate with the lower level neurons? The neural network isn't one way communication, but he describes it as such.

Comments are closed.