James Bland

v0.0.6 is here with some somewhat major adjustments to what goes on behind the scenes

1. A better measure of how much uncertainty you've resolved
2. Grid search for the best lottery pair (this is more robust)

#TeachEcon

jamesblandecon.shinyapps.io/RDUAdaptive/

Data Visualization A Practical Introduction

Here’s a full draft of the upcoming second edition of my “Data Visualization: A Practical Introduction”: socviz.co

Data Visualization A Practical Introduction

@kjhealy.co has a new version of his data visualization book coming out and 1) you’d be a fool not to get it especially if you do R stuff 2) it’s gonna be even more beautiful than the first one, which is truly lovely book 3) he put the ENTIRE content on his website for free, you lucky so-and-so

Makes sense. Interested to see how it turns out.

Out of curiosity, what are you teaching? I could see this working for principles or a theory-heavy class, but I can't imagine teaching my econometrics class without devices. We code too much.

As in, who is more likely to agree to answer the survey?

My first guess is that there might be a greater fraction of male respondents in these countries.

On #WorldBookDay check out our last set of recommendations for economics books to read based around behavioural economics. As discussed in the podcast. #EconSky podcasts.apple.com/gb/podcast/n...

Another shiny app. This one helps you see the implications of selected hyper-priors in a hierarchical model.

It probably has an audience of about four people, one of whom is me, but it is something I worry about a lot.

jamesblandecon.shinyapps.io/CalibrateMyH...

#TeachEcon #EconSky

That's mostly coming from the prior. Come back when you've made 50 decisions!

😜

It’s fun! 🤩

My CRRA coefficient is 0.34. My probability weighting function evaluated at 50% is 57.6%. I made 15 decisions. Estimate your rank-dependent utility preferences here: jamesblandecon.shinyapps.io/RDUAdaptive/

I'm not above tooting my own horn

jamesblandecon.shinyapps.io/RDUAdaptive/

v0.0.5 is now live!

This update adds in a risk premium posterior plot, becuase interpreting RDU parameters on their own is hard.

jamesblandecon.shinyapps.io/RDUAdaptive/

(2) speeds things up a bit

v0.0.4 is up and running!

This is a major change to what goes on behind the scenes.

1. My dodgy MH sampler replaced with R's adaptMCMC functions
2. Fewer samples used to evaluate the objective function

jamesblandecon.shinyapps.io/RDUAdaptive/

#TeachEcon

My CRRA coefficient is 0.10. My probability weighting function evaluated at 50% is 66.2%. I made 50 decisions. Estimate your rank-dependent utility preferences here: jamesblandecon.shinyapps.io/RDUAdaptive/

#TeachEcon #EconSky

So. Much. This.

And why I want to keep my book free to access.

my course notes on a bayesian workflow for (single agent) cognitive modeling are now fully revised and online: fusaroli.github.io/AdvancedCogn...

Predictive checks, updating checks, sensitivity analyses and simulation based calibration in @mc-stan.org

Feedback is very welcome!

This is all based on my working paper about optimizing economic experiments for structural estimation, but the paper just focuses on static designs.

papers.ssrn.com/sol3/papers....

But if you stop after making T decisions, then at least the Tth decision was optimized.

How optimal is this process? I'm not too sure. The algorithm isn't forward-looking like (say) DOSE is. That is, it does not consider the value of being able to ask you the next question.

jnchapman.com/assets/pdf/d...

This is equivalent to (approximately) having a squared loss function between your parameter estimates and their true value.

But there isn't just one information matrix. There's a distribution of them, because the app has beliefs over your parameters. So I optimize over the expected value of the sum of these elements.

If I also included the diagonal element corresponding to lambda, this would be an A-optimal design.

But the model has a parameter I don't really care about: lambda, which measures choice precision. So I just add up the diagonal elements of the information matrix that correspond to the parameters that I do care about.

The app then uses this distribution to calculate the most informative lottery pair to present to you next.

A good place to start here is an A-optimal or D-optimal design, which seek to maximize a function of the information matrix.

At the start of the app, and after every time you submit an answer, the app simulates the posterior distribution of your parameters. This becomes the prior for the next decision.

A bit about my app that elicits your rank-dependent utility preferences using an adaptive algorithm

#EconSky #TeachEcon

jamesblandecon.shinyapps.io/RDUAdaptive/

No worries. And here are your labelled axes and bigger font sizes

The model is of course wrong in the sense that *all* models are wrong (but some are useful). But its parameters and transformations of them tell you something about your risk preferences. Again, if you answer truthfully.