JSM (the Joint Statistical Meetings) are coming up soon, and Jiqiang’s giving a talk on Stan. Here’s the advice I gave him:
in 20 minutes, something like this:
– What is Stan?
– Where does Stan work well?
– Current and future Stan research.For JSM audience it could be good to spend some time on our exciting future research ideas. The goal is not to teach people Stan, it’s to get them excited about it.
You can also look in our JEBS paper for material, as we do some comparison of Stan with other Bayesian model-fitting options.
Regarding surveys, you can say that you personally are not currently working on survey data but the past and current development of Stan has been motivated by various applications of mine, including survey analysis, and we are currently being supported by the polling company YouGov.
Stan has the potential to revolutionize survey inference, as follows: More and more, surveys are not reprsentative of the popualation. Problems with non-response, self-selection, etc. So we want to weight or adjust for as many variables as possible so as to match sample to population. But it’s well known that if you try to weight on a lot of varables, and their interactions, the weights will be super-noisy. Better and more stable to do MRP, i.e. hierarchical Bayes. Stan allows us to build big fast and make models with lots of predcitors and (it will be necessary) informative priors.
The key piece of advice (the secret to giving a good talk) is in bold above.
P.S. And if the computer for the presentation is linked to the sound system, he can start off with the Stan trailer.
Could someone give an example that contrasts the weighting versus MRP approaches applied to the same non-representative survey?
It’s my sense that the bolded statement is largely true for classroom teaching, too.
Any secrets for successfully attending a conference such as JSM with 1,000+ sessions? There are probably no must-see presentations for all 6,000 expected to be there, but perhaps there are especially interesting presentations for people with certain interests.
For a slightly different view on how to captivate an audience see http://deevybee.blogspot.co.uk/2011/03/one-hour-lecture-how-to-captivate-your.html.
The trick is to get the audience excited but not too excited. i.e. No hype. I’ve attended talks where the speaker set the expectations so high about a technology, software, algorithm etc. that people rush to dig in deeper & then are put off because of the mismatch between what was promised vs delivered.
I’m attending JSM myself for the first time this year — not presenting, just scoping out the scene — and I wonder, do you have any secrets to share about successful conference participation? I have this insane list of talks I might want to go listen to based on nothing but abstract keywords and a few familiar names, and I wonder whether there’s a better approach. I’m no stranger to academic conferences, but these huge ones still confound me.
Erin,
I have not been to JSM, but at other large conferences, I have found it beneficial to stay in one general area. If you spend less time moving around, you can spend some time making human contact. The conference feels less large that way.
in helping people prep for conference presentations, my top-level rule-set was:
4I
Information … the easiest part
Insight … about why you did what you did, alternatives considered and rejected, what has been learned,
(things wished to have done differently, in retrospective)*
Inspiration … people get excited, audience is engaged** … this is often the hardest, biggest win if achieved
Impression … mechanics: readable visuals, have practice talk a few times for timing, video/audio recording for yourself to eliminate uhhs and umsms (horrific expeerience first time you do it); be able to give talk without notes or reading script, still be smooth
* people love that … especially since it is really rare to hear