Thursday October 30, 2008
I’ve been in Australia this week, polishing up and publicizing the results of a study I ran for the United States Studies Centre here at the University of Sydney: we surveyed 800 respondents by phone and 3,000 respondents via Internet in Australia, concentrating on their views about the United States ahead of the presidential election.
Key finding: Australians love Barack Obama and would vote for him at margins approaching 70-20.
The instrumentation largely parallels my CCAP study with Lynn Vavreck fielded in the United States, and the Australian studies were fielded contemporaneously with the September wave of CCAP; so we’ve got a great basis for comparisons of public opinion across the two countries. Kewl.
I’ve been doing a TON of media here in connection with the survey, the USSC has a dynamite PR operation. Lots of press yesterday and radio today, including Eleanor Hall on The World Today, Derryn Hinch in Melbourne on 3AW — he saw all kinds of comparisons with 1960 (Obama as Kennedy, McCain as Nixon). Nice insight, I thought.
Here’s just one of many graphs from the study: average left-right placements (1 to 5 scale) of Australian and American politicians and political parties, from our Australian respondents (horizontal bars are 95% CIs for the mean placements).
Tuesday October 21, 2008
I just pushed an update of pscl to CRAN. Version 1.02 contains some more data sets that will appear in my book, Bayesian Analysis for the Social Sciences (Wiley, due out in 2009), along with an option in ideal() to employ marginal data augmentation.
I’ll write all this up somewhere (or you can buy my book), but here is the story. We’re using a Gibbs sampler plus data augmentation (DA) to explore the posterior density of the parameters of an IRT model fit to rollcall data. The DA here is a la Albert and Chib: i.e., to run a probit, sample a latent variable from truncated normals, run regressions of the latent variable on regressors, repeat, and repeat… Marginal data augmentation (a la Meng and van Dyk) in this case introduces a scale parameter into the DA stage. This parameter isn’t identified (i.e., probit is only identified up to a scale factor), but this is actually a virtue, since now the sampled latent variables skate around their parameter space a little more freely, which in turn means that the Markov chain you get in the space of identified parameters (bill parameters and legislative ideal points, in the rollcall/IRT context) displays better mixing. The improvements are quite dramatic for centrists; less so for the troublesome extremist legislators.
Anyway, mda is now an option in ideal, as of pscl 1.02, and the default is to use mda. mda=FALSE gets you back to the usual Gibbs+DA sampler.
Monday October 20, 2008
Some sad news from Berkeley. I heard through the grapevine that David Freedman was not doing well, but last Friday morning we got this email relayed to us from the Berkeley Statistics Dept:
I am very sorry to tell you that David Freedman died this morning. He was one of the most courageous, clear-minded, and honest persons I have ever met, and he remained that way until the end.
David has been a member of the department since 1961 and has had a profound influence on generations of students and faculty. For all of us, it has been a privilege to be his colleague. We will miss him deeply.
Plans for a memorial service will be announced later.
I last saw David when he gave a talk for us in the MAPSS series down here at Stanford, a paper which would appear in Statistical Science. I knew of Freedman’s “shoe leather” critiques of statistical practice in the social sciences before I was aware of his papers with Persi Diaconis on some truly foundational issues in Bayesian inference (e.g., consistency of Bayes estimates, exchangeability, etc). Late in his career he produced Statistical Models (2005), which is a very nice and compact treatment of a lot of what we teach in our graduate methods sequence.
David wasn’t always a fan of what passed for statistics in the social sciences. But his influence on statistics in the social sciences was profound, and for the good, causing a lot of “methodologists” in the social sciences to lift their game. And so we’ll miss him.
Saturday October 18, 2008
Well, kinda, I suppose. See story #10 here. LOL excerpt:
I’m not a scholar, and Mr. Bartels is, and, to me, the first part — I had to read this a bunch of times. It reads like gobbledygook, and what it is is an analysis of a whole bunch of studies worldwide over many, many decades of the electorates, the electorates in democracies.
Also mentioned is the (award-winning) UCLA study on decay functions for advertising effects.
Friday October 17, 2008
We caught the W movie on opening night. Great political timing by Oliver Stone and friends. The debates are over, but the election is still some 19 days away and the result seems clear, but this movie could well help keep the fires stoked for the Obama crowd. “Maintain your rage” and all that, to evoke Gough Whitlam.
Wednesday October 15, 2008
My undergraduates and I watched the 3rd debate tonight. I thought it was perhaps the most substantive of the debates this time around, with only a few detours into “my ads, your ads”. McCain landed a few good jabs that have (as designed) found their way into the sounds bites that make up the late news debate packages.
One of my students said there was not a lot of the buzz about this debate in the dining halls, or in the dorms: i.e., the polling would indicate that this thing has blown out, that the country has made up its mind. Another student used the NASCAR analogy: people only watch in case there is a horror smash. Interesting. 60 million watched the Palin/Biden debate expecting just that, I suppose.
Anyway, I broke away from the debate for an interview with The World Today.
Sunday October 12, 2008
I’ve flown United almost exclusively for 14 years now and I can recall just two previous occasions when I’ve flown UA 744s on a domestic route: Chicago – SFO on a Sunday morning coming home from a conference a few years back, and even Denver – SFO many moons ago.
This morning I flew SFO to Dulles (Washington, IAD) on a 747. As recently as 24 hours before departure it was listed as 777, but I knew something was up when I couldn’t do the on-line check-in. Yesterday afternoon I could get in, and the equipment was listed as 747 (and somehow I managed to mess up my upgrade request too). A bit of digging suggests that it is not totally rare to find UA 744s running domestically, with SFO-IAD as a route on which you find them from time to time.
The aircraft itself sported one of UA’s refurb’d 747 interiors. With my upgrade request lost in the ether, I didn’t get to experience the new C-class seats etc (some of them face aft, which would be interesting I suppose), but even Economy Plus has had a little bit of a make over. Yet, still no in-seat entertainment. But a new, better video screen on the bulkhead. And now the cheese and fruit plate costs $9. No hot food in C-class anymore, at least not domestically: you get a free version of the $6 snack box from Economy. The cabin crew announced that we should feel free to direct any protests to Glenn Tilton.
I’m on the upgrade list for what is showing up as a 747 for IAD-SFO tomorrow night. We’ll see…
Wednesday October 8, 2008
I flew UAL 1166 SFO to LAX tonight on a 757… Light fuel load, low pax load too. Tower advised of some wind shear on the runway (28L), a 20kt loss. Not a problem, since I think the crew dialed it up a notch or two on the takeoff roll and we were airborne well before the intersection with the crossing runways. Nice westerly wind at our end of the field helped a lot too, I suppose.
We continued a very steep climb and I heard them say we were above 8000ft just as we got offshore (still on the runway heading).
I’ve long heard that these lightly loaded 757s can really perform. I believe it. It was probably the most impressive climbing performance I’ve encountered since flying in very lightly loaded Ansett 747 BNE-SYD.
Tuesday October 7, 2008
I got the following inquiry about my/our “pooling the polls” model (names and places removed for anonymity):
Recently I presented a paper at … and received a question from X, which shows his understanding of your method used in the presidential approval is different from mine. I think you used a weekly aggregated polling series for Bayesian Kalman filtering. Xs understanding is that you used individual polls seperately for the Kalman filtering That is if there are 7 polls in a week, all the seven polls are directly used for the Kalman filtering. My understanding is that you average the seven polls and use it for Kalman filtering while you give different weights to the house effect variables according to the sample sizes.
Will you tell me how you did in the paper?
See “Presidential Approval: the case of George W. Bush” (with Neal Beck and Howard Rosenthal, here) and also the earlier version of this research that looks at polls ahead of the 2004 Australian Federal election (replication materials here).
you don’t want to aggregate the polls if you want to be able to estimate house effects. so no, I didn’t do it that way. each poll goes into the analysis separately.
X’s recollection is right!
also, the preferred term is a DLM (dynamic linear model). A Bayesian Kalman filter is an unusual way of referring to the model, since (a) the Kalman filter is just a filter, but there is more to the model than that; (b) the Kalman filter is Bayesian when you unpack it, and so there is some redundancy to the phrase “Bayesian Kalman filter”.