January 2009 Archives

The guest teacher will be Prof. Shigeo Hirano of the Dept. of Political Science. If he has any extra readings for you, I'll let you know!

A McKinsey interview (sorry, it's behind a registration wall, but the registration is worth it if you're interested in business topics or "futurism") with Google's chief economist Hal Varian has an interesting quote:

I keep saying the sexy job in the next ten years will be statisticians. People think I'm joking, but who would've guessed that computer engineers would've been the sexy job of the 1990s? The ability to take data--to be able to understand it, to process it, to extract value from it, to visualize it, to communicate it--that's going to be a hugely important skill in the next decades, not only at the professional level but even at the educational level for elementary school kids, for high school kids, for college kids. Because now we really do have essentially free and ubiquitous data. So the complimentary scarce factor is the ability to understand that data and extract value from it.

I think statisticians are part of it, but it's just a part. You also want to be able to visualize the data, communicate the data, and utilize it effectively. But I do think those skills--of being able to access, understand, and communicate the insights you get from data analysis--are going to be extremely important. Managers need to be able to access and understand the data themselves.

I'm sure everyone reading this blog will feel warmer and fuzzier now. :) But the teaching of introductory statistics really has to convey how to:

• capture data relevant to the problems


• visualize, communicate and effectively utilize the data


• access, understand, and communicate the insights of data analysis

Update 5.28.09: Michael E. Driscoll has a better written description of the above points, along with the guidelines.

Sidney Redner sends along this article by D. Volovik, M. Mobilia, and himself (apparently physicists think it's tacky to include first names in their articles), which begins:

The phone just rang. I picked it up and heard: "Could I speak with the youngest female in the household who is eligible to vote?" My reply: "Sorry, we're busy."

I'm either a traitor to my profession by not participating in a poll, or a contributor by increasing the problem of missing data.

Ed Vul, Christine Harris, Piotr Winkielman, and Harold Pashler wrote an article where:

1. They point out that correlations reported in FMRI medical imaging studies are commonly overstated because researchers tend to report only the highest correlations, or only those correlations that exceed some threshold.

2. They suggest that these statistical problems are leading researchers, and the general public, to overstate the connections between social behaviors and specific brain patterns.

After posting on this article, I received a bunch of comments and questions as well as some responses:

This article by Jabbi, Keysers, Singer, and Stephan argues that, because brain imaging resesarchers adjust their p-values and significance thresholds for multiple comparisons (the thousands of voxels in a brain image), their statistical methods don't have the problems that Vul et al. claimed.

This reply by Vul to the Jabbi et al. article. Here Vul argues that adjustment of significance levels does not stop the selected correlations themselves from being too high. I found Vul's argument here to be convincing. Multiple comparisons methods control the rate of false alarms in a setting where true effects are zero--but I don't see that to be relevant to the imaging setting, where differences are not in fact zero. Lots of things affect blood flow in the brain, and we would never expect the average scans of two different groups of people to be the same.

This article by Lieberman, Berkman, and Wager, who defend social neuroscience and argue the following:

1. They accept Vul et al.'s point 1 above (correlations are overstated) but present some evidence that the correlations aren't as overstated as Vul et al. might fear.

2. They disagree with the implied claim that the overstated correlations have distorted scientists' understanding of social neuroscience research.

3. They object to Vul et al's focusing on social neuroscience, given that the same statistical issues arise in all sorts of brain imaging studies.

4. They point out some specific areas where Vul et al. mischaracterized the data-analytic methods used in this field.

I think Lieberman et al. make some good points, but, as Vul et al. point out, researchers often do use correlations to summarize their results. And, even if said correlations survived a multiple-comparisons analysis, readers might interpret these at face value without understanding the selection issue. So all this shake-out is probably a good thing, especially where correlation estimates are being compared to each other.

My thoughts

First off, I haven't worked seriously in medical imaging for nearly 20 years and have only one published paper in the area, so my comments are mostly informed by my perspective on general statistical issues, as well as my own experience thinking about estimation of effect sizes in studies with low statistical power.

Regarding the singling-out of social neuroscience, I see the point of Lieberman et al. I was thinking that maybe one reason for this is that in social neuroscience it's perhaps more difficult to get external validation in the way that might be more possible in other areas of neuroscience where there is some measurement in the blood or whatever that can be taken. I'm not sure about this, just a conjecture.

It's hard for me to believe that the approach based on separate analyses of voxels and p-values, is really the best way to go. The null hypothesis of zero correlations isn't so interesting. What's really of interest is the pattern of where the differences are in the brain.

Related to this point is that, ultimately, when trying to understand differences in brain processing between different sorts of people (or between people doing different tasks), the maximum correlation among voxels is ultimately not what you're looking for. That is why researchers summarize using regions of interest (as in p.7 of the Lieberman et al. article). Vul et al. were correct to warn about overinterpretation of correlations that have been selected as the maximum: the naive reader can see such correlations (and accompanying scatterplots) to think that certain personality traits are more predictable from brain scans than they actually are.

I think the way forward will be to go beyond correlations and the horrible multiple-comparisons framework, which causes so much confusion. Vul et al. and Lieberman et al. both point out that classical multiple comparisons adjustments do not eliminate the systematic overstatement of correlations. A hierarchical Bayes approach (using some sort of mixture for the population of pixel differences, ideally modeled hierarchically with pixels grouped within regions of interest) would help here..

And now for some amateur psychologizing (unsupported by any statistical analysis, correlational or other)

I suspect that one of the motivations of Vul et al in writing their article was frustration at too-good-to-be-true numbers which they felt led to exaggerated claims of neuro super-science.

Conversely, I suspect one of the frustrations of Lieberman et al. is that they are doing a lot more than correlations and fishing expeditions--they're running experiments to test theories in psychology, they're trying to synthesize results from many different labs. And from that perspective it must be frustrating for them to see a criticism (featured in the popular press) that is so focused on correlation, which is really the least of their concerns.

It also seems that both sides were irritated by what they saw as giddy press coverage: on one side, claims of dramatic breakthroughs in understanding the biological basis of behavior and personality; on the other, claims of a dramatic Emperor-has-no-clothes debunking. As scientists, most of us welcome press coverage--after all, we think this work is important and we'd like others to know about it--but . . . fawning press coverage of something that we think is wrong--that's just annoying.

P.S. Wager is a friend--he teaches in the psychology department here--but I don't think my personal knowledge has hindered my evaluation here.

P.P.S. I ran the above by various people involved and they gave some helpful clarifications. But I've probably left in a couple of sloppy statements here and there.

John Lanchester asks this question about video games. I have a few observations to add:

1. When I was a teenager, my friends and I spent tons of time at the arcade, just throwing quarters into videogames. (I always preferred pinball, but videogames were more available.) I haven't played a videogame in decades and have zero interest in doing so. (This is not something I'm necessarily proud of, just a statement of fact.) And now I can't figure out what about videogames was so appealing to us, back then.

2. Different people read different sorts of books, often with little overlap. Stephen King, John Grisham, Danielle Steele, etc., are the super-sellers, but lots of people would read one of these and not the others--and lots of readers don't read these blockbusters at all. In contrast, everybody who's into movies is aware of the latest major releases, and it's my impression that people who would rarely read a bestseller of the Stephen King sort would still watch a big-budget movie.

To put it another way: My impression is that the default for reading is to pick something in a niche that you're interested in, but the default for movie watching is to start with the blockbusters and then go from there.

3. But with old movies, I think it's different. Back in the old days, everybody might watch whatever old movie was being shown on TV that night, but what with videos people will now make their choices. And if a movie is old, the whole blockbuster thing seems irrelevant.

I don't know exactly how video games fit into all this; I just wanted to point out that the same audiences seem to expect different things from different media.

I got an email from the people at Revolution computing reporting:

REvolution Computing has now made a public version of its commercial grade REvolution R program available for download from its website. REvolution R is REvolution Computing's distribution of the popular R statistical software, optimized for use in commercial environments. A key feature includes the use of powerful optimized libraries capable of boosting performance by a factor of 5 or 10 for commonly used operations.

Here it is.

There's also a payware version, which "features advanced functionality, including ParallelR, which speeds deployment across both multiprocessor workstations and clusters to enable the same codes to be used for prototyping and production. REvolution R Enterprise is functional with 64-bit platforms and Linux enterprise platforms and provides for telephone support and response guarantees."

I don't know anything about this but at least in theory it sounds like a good idea! If anyone has any comments, feel free to share them.

Seth asks the above question. I have a couple of thoughts:

1. It's hard to write a book that's easy to read. As Seth points out, it's a Venn diagram situation: you're looking for the overlap between three groups of people: those who know the material, those who can write well, and those who are willing to put in the effort to write a book.

2. Some subjects are so urgent that they're worth writing--and reading--about even if hard to read. For example, if you're a terrorist and want to build an A-bomb (or, to use another example, if you're a social scientist and you want to use quantitative methods), sure, you'd prefer clear instructions, but something that's hard to read is still much better than nothing.

3. What's hard for you to read might be easy for somebody else. An extreme example is foreign language or the passage of time. But, beyond that, there's familiarity. For example, I find the news section of the newspaper much easier than the comics to read. The comics are much simpler, but I'm not familiar with them, and to read them I have to enter all these different stories. In contrast, news stories follow a predictable pattern: drought in the Great Plains, people blowing each other up in the Middle East, and so forth.

4. Standards differ. Bayesian Data Analysis is considered by many to be well-written but it's much less easy to read than Data Analysis Using Regression and Multilevel/Hierarchical Models, which in turn is, I'm sure, much less easy to read than the collected works of Len Mlodinow. But, then again, Mlodinow is harder to read than Stephen King. If Stephen King wants to write statistics books, that's fine, but I'm afraid that would wreck things for the rest of us. It would set the bar too high.

5. And, of course, some people write books to be hard to read on purpose. James Joyce and so forth. If you, like me, aren't a big James Joyce fan, just pick your favorite writer who you find challenging to read (in a good way).

John Updike just passed away, and coincidentally I noticed a story by him in the most recent New Yorker. Well, actually the story was by someone called "Antonya Nelson" but it was clearly a John Updike story. Not angry enough to be a John Cheever story, not clipped enough to be a Raymond Carver story, not smooth enough to be a Richard Ford story. Based on this evidence, I expect we'll continue to see new Updike stories for awhile.

P.S. I'm a big fan of Updike. Rabbit, Run is my favorite.

When we have a grad school applicant who's taken the GRE or TOEFL multiple times, we typically just look at the highest score. It's my impression that pretty much everybody does this, even though basic statistical principles would suggest taking the average. Eric Rasmusen reminded me of this point in the context of the SAT, which apparently has changed its policy to encourage multiple test-taking even more, by allowing students to report only their highest score. Throwing away information--that doesn't sound like a good idea! But, as Rasumusen points out, it might make more money for the organizations that administer the test.

According to the linked news article, students "will have the option of choosing which scores to send to colleges while hiding those they do not want admissions officials to see." My question is: will their score report state whether they've chosen this option? If so, it should be possible to at least try to correct for the bias.

In any case, all this discussion makes me think we should be more careful about just looking at the maximum when our applicants take the GRE multiple times. And then there's the possibility of cheating. . . . I guess the real lesson is that these admissions decisions aren't going to be perfect, and we should think more about how to incorporate this perspective into our admissions process.

Pippa Norris writes:

I realize that it is clunky but if you could always, always cite the survey source and date below each figure, this would make the book much easier to read and interpret. If I know the source, then it is easier to judge the meaning of the presentation, the exact questions used, and the reliability of the data. I use a lot of figures in presenting my own work, to the despair of my publisher, and I know how difficult it is to both combine elegance and simplicity with technical details in a compact space. But if we don't provide these details, then this is such a bad role model for our students!

She's got a point. I'm a big believer in having the graph and caption be a self-contained entity--as I tell my students, you have to think about people like me who only read the graphs--but I've rarely put the data source right in there too. In our book we have all the data sources listed in the notes at the end, but I agree that putting sources right on the graph would be a good idea. Actually, I think what I want to do is write some R functions to make graphs just the way I like them, and one option on the graph will be to give the data sources in small print near the bottom.

Brad Miner wrtes:

With the Super Bowl coming up this weekend, I [Miner] want to write about sports, which I consider a key to building a larger conservative coalition in America. . . .

If you did a survey of the political philosophies of 75,000 randomly selected Americans you'd expect the usual--if somewhat mystifying--results: "Only about one-in-five Americans currently call themselves liberal (21%), while 38% say they are conservative and 36% describe themselves as moderate." So said the folks at Pew Research, and this was after the November election.

Do that same poll among the fans at Raymond James Stadium in Tampa on Sunday and the results would likely be more like 15% liberal, 30% moderate, and 50% conservative. And a bunch of those liberals would probably be gun owners.

Obviously those numbers are just speculation on my part, but I guarantee that Steelers fans are more conservative than all Pennsylvanians and ditto Cardinals devotees and the rest of Arizona. Which is not to say that these folks cast their ballots in November more for McCain than Obama. That's the problem.

What do the data say?

Yu-Sung and I looked at the "attended sporting event in the past year" item in the General Social Survey. (Unfortunately, the question was only asked once, in the 1993-1996 survey.) 56% of respondents said they attended an amateur or professional sports event" during the past twelve months. How do they differ from the 44% who didn't?

So, at least in the mid-1990s, sports attenders were quite a bit more Republican than other Americans (the categories in the graph above are Strong Democrat, Democrat, lean Democrat, Independent, lean Republican, Republican, strong Republican), but not much different in their liberal-conservative ideology.

So these data do not appear to support Miner's claim. Miner expected sports fans to be label themselves as more conservative but maybe not to be more likely to vote Republican; actually, sports fans were more likely to call themselves Republican but no more likely to describe themselves as conservative.

Some other issues:

1. The sporting event attended could be the Super Bowl or your kid's soccer game. Maybe more dramatic results would be obtained by considering a more restricted group of sports fans.

2. There are lots of surveys of TV watching, so I'm sure there are tons of data that would let you crosstab ideology, voting, and spectator sports watching.

3. More generally, we never want to rely too strongly on just one survey. Still, it's fun to look.

P.S. Sometimes people ask me how much time blogging takes me. This took about an hour: 15 minutes for me to read Miner's article and think about it, 10 minutes for Yu-Sung to get the crosstabs, 20 minutes for me to make the graphs, and 15 minutes for me to write the blog entry.

And, yes, this means I have a lot of real work that I've been putting off. . . .

Burt Monroe writes:

I [Burt] sent an entry for the Chance visualization contest. By the time I'd ferreted out the original in our library and quizzed family friends (my father was a biologist) about bacterial taxonomy, I ended up writing a goofy little paper about the whole thing.

Here's Burt's article, and here's his graph:

This is pretty, although to my eye it looks a little busy. I'd probably favor including this information in multiple plots. That said, I didn't actually look at the data or the problem--all I did was post the announcement--so maybe it's the best thing to do. I like that Burt investigated the context of the problem and didn't just treat this as a "dataset" to be graphed.

Jess sends along this, which isn't a bad idea although I disagree with how it's organized. For one thing, I think just about all graphs are comparisons; for another, I think line graphs are often the way to go, so I'm unhappy to see them in only a few of the pictures here; for another, the scatterplot-plus-regression-line, which I love, isn't anywhere to be found. But I appreciate the thought.

Hal Pashler writes:

I [Pashler] thought you guys would enjoy this charming little 1950 paper by Edward Cureton entitled "Validity, Reliability, and Baloney" (Dirk Vorberg, a German math psych guy, sent it). Long before machine learning, it seems that psychometrics people were confronting this issue--and the concrete form it took was "What should we make of validation measures computed with the same data that were used to select out particular items for inclusion in the test?". Just swap voxels for items, and it's the same problem [as in the Vul, Harris, Winkelman, and Pashler paper on suspiciously high correlations in bran imaging studies].

This reminds me of a longstanding principle in statistics, which is that, whatever you do, somebody in psychometrics already did it long before. I've noticed this a few times. Once, about ten years ago, I was at a conference where computer scientists were talking about some pretty elaborate statistical models, and I realized these were the same as some things I'd seen Iven Van Mechelen and his colleagues working on in the Psychology Department at Leuven. Then, more recently, I wrote this article with David Park on splitting a predictor into three parts, and it turned out that similar work had been done in 1928! by psychometric researcher T. L. Kelley (and, oddly enough, E. Cureton in 1957).

I received the following email from a Ph.D. student who wishes to remain anonymous:

I came into operations research with a masters in mechanical engineering therefore my statistical analysis is very unbalanced. I tool stochastic processes I and II (both PhD level courses) but I have no applied background in statistics, now I am doing a lot of number crunching using R but I believe I still lack broad understanding of statistical tools and I don't know enough about analysis (although I believe I have a solid understanding of the basics like mean, divergence, confidence interval, etc)

Given all my embarrassing situation I would appreciate it if you would please, in a blog post, lay out a learning plan for people like me who like to dive deeper into statistical analysis and do it in daily basis but come from weird backgrounds like mechanical engineering.

My reply: Since you're not at Columbia, I can't simply recommend that you take my course. You could read my books, that might help. More seriously, you gotta think about all the great skills you have that many statistics students don't have: you can make yourself useful in a lot of different sorts of projects . . .

Aleks sends in this link to an interactive statistics web course and writes, "It requires registration, but I highly recommend you go through this. You will probably not like the statistics teaching style that goes back to Fisher's 1930s textbook, but it's hard not to appreciate the effort that went into interactive 3D demonstrations of various concepts."

There is something incredibly old-fashioned about what they're teaching, but I agree that the 3-D graphics are extremely impressive. Probably the wave of the future, once it can be combined with a more up-to-date (that is, ARM-style) set of statistical concepts and methods.

Peter Flom is a statistical consultant I know, who has worked with graduate students and researchers in the social sciences, education, medicine and other fields. He earned his PhD in psychometrics from Fordham University in 1999, and has been first author and co-author on numerous publications. He has also assisted in grant writing and dissertation preparation. He is based in New York City, but also works remotely.

Nate points out that Obama's approval rating at the start of his presidency is higher than anyone since John Kennedy, and he writes, "In comparison with Ronald Reagan, however, Obama's approval is quite a bit more impressive. Indeed, it is hard to mount a credible argument that Reagan began his term with more political capital than Obama."

I recall reading that Reagan's popularity jumped after his attempted assassination two months into his presidency. (See here for some graphs.) So I suspect that the impression of Reagan being initially popular comes from that point, not from January 20.

P.S. Recall that Reagan, like Roosevelt, was a statistician.

Mark Blumenthal links to this article by Nate Silver, who writes, "If Bill Clinton was the first black president, then Barack Obama might be the first urban one."

This reminds me of some of our recent discussions here:

- Who was the last urban president before Obama? (I said Nixon, who lived in New York before his 1968 presidential run; a commenter said Kennedy, who was from Boston.)

- County-level vote swings by population. Democrats have been gaining in urban areas. The gain has been pretty steady over the past three elections, so I don't know how much should be attributed to Obama's urban-ness in particular. Graphs here along with much discussion:

What do we see?
1. The large-county/small-county differential in Obama's gains was particularly strong in the south and did not occur at all in the northeast. For example, Obama won 84% of the two-party vote in Philadelphia--but Kerry got 80% there four years ago. This 4% swing was about the same as Obama's swing nationally. Part of the issue here is that Obama had almost no room for improvement in these places.

2. The pattern of Democrats improving more in large-population counties is not unique to 2008. Gore did (relatively) well in big counties in all regions in 2000.

Christian Robert has some thoughts on my paper with Aleks, Yu-Sung, and Grazia on weakly informative priors for logistic regression. Christian writes:

I [Christian] would have liked to see a comparison of bayesglm. with the generalised g-prior perspective we develop in Bayesian Core . . . starting with a g-like-prior on the parameters and using a non-informative prior on the factor g allows for both a natural data-based scaling and an accounting of the dependence between the covariates. This non-informative prior on g then amounts to a generalised t prior on the parameter, once g is integrated.

This sounds interesting. I agree that it makes sense to use a hierarchical model for the coefficients so that they are scaled relative to each other.

Regarding the pre-scaling that we do: I think something of this sort is necessary in order to be able to incorporate prior information. For example, if you are regressing earnings on height, it makes a difference if height is in inches, feet, meters, kilometers, etc. (Although any scale is ok if you take logs first.) I agree that the pre-scaling can be thought of as an approximation to a more formal hierarchical model of the scaling. Aleks and I discussed this when working on the bayesglm project, but it wasn't clear how to easily implement such scaling. It's possible that the t-family prior can be interpreted as some sort of mixture with a normal prior on the scaling.

In any case, maybe Aleks can try Christian's model on our corpus and see what happens. Christian links to his code, which would be a good place to start.

Hey, I've wanted to do this for awhile! Example code here.

I gotta say, I find the expression() function incredibly difficult to use. Examples are key.

Arthur Brooks writes:

Over the past several years, studies have consistently shown that people on the political right outperform those on the left when it comes to charity. This pattern appears to have held -- increased, even -- in 2008.

In May of last year, the Gallup polling organization asked 1,200 American adults about their giving patterns. People who called themselves "conservative" or "very conservative" made up 42% of the population surveyed, but gave 56% of the total charitable donations. In contrast, "liberal" or "very liberal" respondents were 29% of those polled but gave just 7% of donations.

I had a grant application turned down and wrote the following polite email to the program director:

Dear Dr. ***,

I am sorry to hear this. In particular, I can't understand how the panel could've thought that the methods are "not in themselves new." Clearly we have more work to do in explaining our proposal.

But I will look on the upside, which is that ** must have received some excellent proposals to fund that were even better than ours! So congratulations on that.

Yours
Andrew Gelman

I was surprised that he did not respond, but when I related the story to my colleagues, they explained to me that the director might have thought I was being sarcastic in my email. I was actually sincere. But intonation is notoriously difficult to convey via email.

Justin Phillips, Jeff Lax, and I wrote this article summarizing some of the findings of their recent research on gay rights in the states:

In his address at the Democratic convention, Barack Obama said, "surely we can agree that our gay and lesbian brothers and sisters deserve to visit the person they love in the hospital and to live lives free of discrimination."

What was he thinking, saying this to the nation? California was on the way to a contentious battle over same-sex marriage and the issue has arisen in other states as well. Isn't gay rights a wedge issue that Democrats should try to avoid?

Yes, Americans are conflicted about same-sex marriage, but one thing they mostly agree on is support for antidiscrimination laws.

In surveys, 72% of Americans support laws prohibiting employment discrimination on the basis of sexual orientation. An even greater number answer yes when asked, "Do you think homosexuals should have equal rights in terms of job opportunities?" This consensus is remarkably widespread: in all states a majority support antidiscrimination laws protecting gays and lesbians, and in all but 10 states this support is 70% or higher.

But people do not uniformly support gay rights. When asked whether gays should be allowed to work as elementary school teachers, 48% of Americans say no. We could easily understand a consistent pro-gay or anti-gay position. But what explains this seeming contradiction within public opinion so that gays should be legally protected against discrimination but at the same time not be allowed to be teachers?

If anything, we could imagine people holding an opposite constellation of views, saying that gays should not be forbidden to be public school teachers but still allowing private citizens to discriminate against gays. A libertarian, for example, might take that position, but it does not appear to be popular among Americans.

We understand the contradictory attitude on gay rights in terms of framing.

Elizabeth Suhay sent me this article examining the mechanisms underlying social influence. Here's the abstract:

Citizens often feel pressured to adopt the beliefs held by their peers, conforming to the views of the majority even in the absence of rational argument. However, few scholars have investigated the mechanisms underlying this "mindless" conformity to group pressure. Drawing on recent research in psychology, this manuscript puts forward a new theory of group influence called Social-Emotional Influence Theory which states that subjective group identification and self-conscious emotions (e.g., pride and shame) are critical to understanding political conformity. We feel pride when we conform to, and shame when we deviate from, in-group beliefs and behaviors; these emotional reactions motivate conformity. SEI Theory is tested with an experimental study of group influence among Midwestern American Catholics with respect to social conservatism. The evidence supports SEI theory: Identification with other Catholics mediated group influence over participants' conservative views, and self-conscious emotions appeared to play a key role in explaining that influence.

I like the idea that Suhay presents her theory as complementary rather than competitive with more traditional quasi-rational information-processing models of Diana Mutz and others.

Regarding the social conservatism of American Catholics, I wonder what Suhay would say about Rudy de la Garza's finding that Latino Americans have conservative views on abortion, but very few of them state abortion as an important determinant in their vote. This suggests that there's another choice point, which is how much to consider attitudes on social issues when deciding how to vote.

And here's her picture (sorry, no cool data graphs this time):

My thoughts here.

And, for a different audience, a discussion of Red State, Blue State.

As always, I thank my collaborators for a lot of the analysis that I'm summarizing in these articles and discussions.

I discussed here the gradually decreasing decline in relative vote swings by state:

The next step was to do this calculation by counties. For each presidential election year in the graph below, I computed the interquartile range (that is, the 75th percentile minus the 25th percentile) of the swings in vote proportions for the Republicans for the 3000 or so counties in the United States. I exclude third-party votes. For each year, I computed the interquartile range for all the counties in the U.S. and also just for the counties outside of the South.

I only seem to have the data at hand back to 1968 which is why this graph only goes back to 1972. As with the statewide swings, there has been a steady decline, in this case much more dramatic in the South. The decline is gradual but we're clearly at a lower level of variation now than we were 30 years ago.

A few days ago, I discussed an interesting article that said that it's actually not so unusual for countries to be multilingual. Ubs disagrees:

Although there is more to nationalism than just language, the idea of identifying a political state with a single language is a central idea of nationalism. When you contemplate why it is that today we expect any state to have a single language and think of Canada or Belgium as "weird" (and don't forget Switzerland), what you're really contemplating is why the nation-state has become dominant in the modern world.

Among those who study such things, the standard and mainstream thesis is that there is a fundamental incompatibility between a multinational state and a modern state. Historical evidence is abundant that nationalism tends to occur simultaneously with industrialization, and there are plenty of plausible reasons why this should be so. In a traditional society, where the economy is primarily agricultural and power flows hierarchically, a local noble who does not speak the language of the capital is at no particular disadvantage; in a modern society, where production is aimed for the market, literacy is essential to economic success, and political power flows through a central establishment, he is disenfranchised.

Historians continue to debate the exact nature and significance of the connection between modernization and nationalism, but no one can ignore the question. The two central examples are the Habsburg and Ottoman empires. The 19th century history of both states is completely dominated by their efforts to reconcile the multinationalism with modernization. The Ottoman Empire remained multinational and as a result failed to effectively modernize. The Habsburg empire did modernize but was unable to remain multinational.

The notion that language became a problem in Habsburg lands only after the end of World War I, as your quoted excerpt seems to imply, is ridiculous. The language problem dominated the empire's politics from its founding in the Napoleonic wars until its defeat. Following the links, I see that Kamusella's book is 1,168 pages long, so I'm sure he has plenty to say about this, but if his thesis is merely that multilingualism is extinct in central Europe because the mean British, French, and Americans "delegitimized" multilingualism, then either he is naive or he thinks we are.

I would say -- and I believe this is a pretty mainstream view -- that the multilingual nature of the Habsburg empire (and likewise the Ottoman) put it at a disadvantage vis-à-vis nation-states like France and Britain. As a result it was unable to recover after defeat as Germany did. The victors of World War I did not "choose to delegitimize" Austria-Hungary; they defeated it, and they destroyed it. The bundle of smaller states that filled the void were created not in the pursuit of any unilingual ideal, but simply for the usual geopolitical reasons.

It is true that Wilson paid lip service to the idea of drawing state boundaries to match national identities, but this premise was used only where it was politically convenient. It was easily abandoned in South Tyrol, Sudetenland, and Asia Minor, among other places, and the victors' preservation of bilingual Belgium was the very opposite of delegitimization. The principle behind the carving up of Central Europe after World War I was not any "ideal of ethnolinguistic homogeneity", but rather an attempt to secure all the economically productive property under the political control of the victors.

But I digress. Reading the post again, I think I probably don't disagree with Kamusella nearly as much as I initially thought I did. I think the excerpt and the context in which it was presented rubbed me the wrong way.

The point is that if you had "always thought of monolingual countries as a default rather than a construct", you were absolutely right. At least for any modern state.

Far be it from me to argue with someone who can not only use the expression "vis-à-vis" in a sentence but also knows how to put the accent into an email. But . . . what about China, India, and the former Soviet Union?

P.S. Ubs also points out:

If you say "monolingual" you're mixing languages (not that that hasn't been done before): Bilingual, multilingual, unilingual; polyglot, monoglot (and I suppose, though I've never heard it, "duoglot").

I'll try "unilingual" on for size. (I can't bring myself to say "monoglot," given that "polyglot" sounds weird enough as it is.

This one just came from Life Science Journals :

Based on your research profile, we would like to offer you the following free subscription to Nature Biotechnology.

Click here to sign-up for your free subscription which is available without any obligation to qualified scientists.

What part of my "research profile" are they talking about??

Mark Thoma has an interesting discussion of the challenge that the economics profession, and individual economists, have when they give policy recommendations.

Mark's basic point goes as follows. Consider the following four stages of a model:

(a) assumptions about fundamental principles of how the world works,
(b) normative principles (that is, fundamental goals, views about how the world should be),
(c) conclusions about the likely effects on policy,
(d) recommendations about policies.

In any rigorous economic model, there should be a mapping leading from (a) to (c). Further reasoning (possibly mathematical modeling, as in cost-benefit analysis) will take you from (b) and (c) to (d).

That's all fine. But Mark's point is that the reasoning can go the other way too: start with (b) and (d), and then you can figure out what (c) needs to be, and then you can go back one more step and figure out what model (a) you need to get started! Even if economists are not doing this reasoning-from-conclusions-to-assumptions explicitly, you could well believe it's going on implicitly as well as being induced by various pressures such as the selection of what research results to report and even what problems to work on.

This is inevitable, and I discuss it in the decision analysis chapter (22, I think it is) of Bayesian Data Analysis. We call it the garbage-in-garbage-out problem: If you can come with any decision you'd like by just altering the inputs of your analysis, then what's the point of decision analysis (or, by extension to the above-linked example, economic modeling) at all?

My answer is something that I call "institutional decision analysis," which has two principles:

1. It can be a good idea to provide reasoning to justify your decisions. As an individual person, you might not have to justify your personal decisions to anyone (except to your spouse), but an institution--whether it be a business, a government agency, a nonprofit organization, or some other grouping--often needs some path of bread crumbs connecting assumptions to recommendations. (Here, I carefully say "connecting" rather than "leading from" to be agnostic about the direction of the reasoning.)

2. As Mark noted, an overall decision recommendation on anything important is likely to be so dependent on assumptions to such an extent that it's probably fair to say that the analyst is reasoning from conclusions to assumptions (from (d) to (c) and then to (a), in my above notation). But, even then, formal decision analysis can be useful in making relative recommendations. This is the point that we made in our article about decision making for home radon [link fixed]. In the economics context, this might suggest that economists of different political persuasions could still give useful recommendations about how to spend money or cut taxes, or where in the economy such policies would make more or less sense.

Strange Maps has this cool picture of Polish election results compared to the pre-1914 partition border:

I can't tell what the colors represent, but it's striking nonetheless. In linking to it, Matt Yglesias writes,

History's impact can often be surprisingly long-lasting. It's been a long time since taking midwestern agricultural products via train to Chicago and then by boat across the Great Lakes, across the Eerie Canal, down the Hudson, and to the port at New York was a major element in the American economy. But we still have two giant cities in Chicago and New York . . . I wouldn't be surprised if the German-run bit of Poland was richer in 1918 than the rest of it, and that the differential has persisted since then. By the same token, we can expect the East Germany part of Germany to remain poorer than the West Germany part for a long time.

Here are some graphs that I posted a couple of years ago and that found their way into chapter 5:

More pictures here (for those of you who haven't bought the book yet). For the book, we cleaned up the graphs a bit, but the general point remains, that the states that are rich and poor now are the ones that were rich and poor 80 years ago.

Mark Blumenthal writes about some of the elaborate analyses done by Catalist and other political consulting firms that helped organize the Democrats' get-out-the-vote efforts in 2008. Presumably the Republicans will be doing this next time as well. It brings swing-voter targeting to the next level. It would also be interesting to do some multilevel modeling to put together their county- and individual-level analyses (and, soon, their precinct-level analyses). Also it's worth thinking about how national politics might change as these techniques become more widely used.

In a discussion of her recent Aikenesque historical novel, Jenny links to a reviewer who liked the book but didn't think the fantasy elements fit in so well with the rest of the book. In this case, the fantasy part involved communication between living and dead people ("spiritualism"). Jenny then links to Colleen Mondor, who also liked the book and said that she didn't mind the fantasy element since she (Colleen) has talked to dead people herself.

On this particular point, I have no problem with people talking with dead people. But I'm skeptical about claims that the dead people are talking back.

Here I'm talking about real life, not fiction. I certainly agree that fiction can be "true to life" even while violating recorded history or the laws of physics or just about anything else. Think about Stephen King, for example. I imagine there must even be some stream of science fiction (if you'd call it that) centered around, not new technology or alternate history or fantasy, but violations of logic and continuity. For example, a guy goes out of the house wearing a red shirt and later it's green. Or he gets in his car to go to work, but when he gets to work, he's getting off the bus. That sort of thing is impossible by anyone's standards--of course I'm excluding rational explanations, blackouts, Mission Impossible-style kidnappings and staged sets, etc.--but in some ways it's true to lived experience. (Yes, I realize that some of Philip K. Dick's books are sort of like this--for example, Time out of Joint--but here I'm thinking of even more extreme continuity violations, the sort of thing you'd see in a poorly made low-budget movie where somebody lost the script.)

Anyway, my point in bringing this up is to separate any disagreements about the ability of dead people to talk, from the larger question of getting human insight (or at least a good story) from something that not only didn't happen, but couldn't happen.

P.S. All this blogging is a clear sign that I have lots of work I've been putting off! One thing that happened is we just moved (around the corner). Our new apartment has an airy living room with lots of bookshelves, and sitting here seeing all the books gets me thinking more about literature. . . .

Kenny sent me this article by Bill James endorsing Hal "Bayesian Data Analysis" Stern's dis of the BCS. I'd like to add a statistical point, which is a point that Hal and I have discussed once or twice: There is an inherent tension between two goals of any such rating system:

1. Ranking the teams by inherent ability.

2. Scoring the teams based on their season performance.

Here's an example. Consider two teams that played identical opponents in the season, with team A having a 12-0 record and team B going 9-3. But here's the hitch: in my story, team B actually had a much better better point differential than team A during the season. That is, team A won a bunch of games by scores of 17-16 or whatever, and team B won a bunch of games 21-3 (with three close losses). Also assume that none of the games were true run-up-the-score blowouts.

In that case, I'd expect that team B is actually better than team A. Not just "better" in some abstract sense but also in a predictive sense. If A and B were playing some third team C, I'd guess (in the absence of other information) that B's probability of winning is greater than A's.

But, as a matter of fairness, I think you've gotta give the higher ranking to team A. They won all 12 games--what more can you ask?

OK, you might say you could resolve this particular problem by only using wins/losses, not using score differentials. But this doesn't really solve the general problem, where teams have different schedules, maybe nobody went 12-0, etc.

My real point with this example is not to recommend a particular ranking strategy but to point out the essential tension between inference and reward in this setting. That's why, as Hal notes, it's important to state clearly what are the goals.

P.S. It's been argued that a more appropriate system is to change the rules of football to make it less damaging to the health of the players (see here for a review of some data). I certainly agree that this is a more important issue than the scoring system. In statistics we often use sports examples to illustrate more general principles, but it is always good to be aware of
the reality underlying any example. It also makes sense to me that people who are closer than I am to the reality fo the situation would be less amused by the thoughts of Bill James and others about the intellectual issues in the idealized system.

Gur Huberman writes, regarding the Edlin, Gelman, and Kaplan article in The Economist's Voice:

Can you extend the charity/rationality argument to explain why people in non battleground sates (e.g., NY) vote? Even if charity motivation is a partial explanation for voting, an implication would be that voter turnout is higher in battleground states, other things being equal. However, I am afraid that this prediction is consistent with many other explanations of why people vote.

Another issue that has intrigued me for years: I am under the impression that voter turnout is lower in local elections and in midterm elections. In midterm elections there's less at stake, so your charity story seems to cover that. But, selfishly speaking, it may well be that who my mayor is may have a stronger impact on my life than who my president is. (Quantifying this last statement is challenging.) If so, why am I more likely to vote in a presidential election than in a mayoral one? Your charity theory may help answer the question.

My reply

1. I think there are many reasons for voting, and in NY it's not particularly rational for instrumental reasons.

2, In our article a couple years ago in the journal Rationality and Society, Edlin, Kaplan, and I discuss the coexistence of many different models for voting. For example, there is the "psychological" model that we are more likely to vote in an election that more people are talking about. People are more likely to talk about an election that is close and that is viewed as important. So the psychological and economic/rational explanation coincide in this way. (Similarly, you could consider psychological or economic rationales for purchases. For example, if I buy something on sale, I'm economically motivated to save money and psychologically motivated because of the pleasure in "getting a deal.") These two things reinforce each other; I see them as parallel, not competing, explanations.

3. Your mayor may have more of an impact on _your_ life, but total impact is proportional to total #people affected. And that doesn't even get into foreign policy (not an issue for local politics unless you happen to live in, say, Berkeley, California).

AT points me to this column by Andrew Leonard, who writes:

I [Leonard] asked my readers to explain Felix Salmon's statement that "I'd say p=0.3 right now that Barack Obama's first major act as POTUS will be the nationalization of Citigroup."

What follows are some amusing quotes, including this one from Kobi "diagnostics for multiple imputations" Abayomi:

People pick a cutoff -- arbitrarily, really -- and a p-value lower than the cutoff is pronounced "Significant." Alchemy! . . . p-values are a bit passe, if not completely gauche -- statistically speaking. Modelling, these days, is more particular (if not exact). Macho statisticians are proud of tight (posterior (bayesian)) confidence intervals. Real men keep p-values to themselves.

I think there's some confusion here. Leonard's correspondents were making things too complicated. Salmon was using "p" as jargon to represent "probability," not "p-value." Thus, he was saying that he saw it as a 3 in 10 chance that Obama's first major act would be to nationalize Citigroup. Everybody is so hot under the collar about p-values that they didn't notice the direct interpretation.

As I've discussed before, in 2008 the red/blue map was not redrawn; it was more of a national partisan swing:

I also posted some graphs of previous vote swings that were less uniform.

But maybe it makes sense to study this more systematically. For every pair of consecutive elections since 1952 (that is, 1952/1956, 1956/1960, . . . , 2004/2008), I compute the interquartile range (that is, the 75th percentile minus the 25th percentile) of the swings in statewide vote proportions for the Republicans. I exclude third-party votes and also exclude states that were won by third parties. For each year, I computed the interquartile range for all 50 states (plus D.C. when appropriate) and also just for the non-southern states. Here's what I found:

The past several decades have seen a steady decline in the variation of statewide vote swings. (The big spike in the graph is 1976, when Jimmy Carter did very well in a bunch of southern states that Nixon carried in 1972.)

To put it another way, the red-blue map is much more stable from election to election than it used to be. What's going on? I'm not sure, but I think this is an important stylized fact.

Here's the introduction to Jenny's new book, which is all about the pure nature/nurture distinction, "pure" in the sense of being uncontaminated by the scientific perspective of modern biology. In that sense it reminds me of The Passions and the Interests: Political Arguments for Capitallism before Its Triumph, by the great Albert Hirschman.

Reading this intro reminds me that authors often say that nobody reads the introduction, but in my experience a lot of people do. One way I can tell is that reviews sometimes pick up on items mentioned in the intro, another way is that people pick up on personal info in the acknowledgments.

I was also reminded that the first Jenny you told me about her book-in-progress, I thought she said the title was "Braiding" (it was her mid-Atlantic accent), which oddly enough wouldn't be a bad title for the book.

It's funny how often such malapropisms are possible; for example, I had a friend who once said she just wanted to bleed into the woodwork. Another time she said she wanted to get on the right tract. In the case of Braiding the malapropism came from the listener not the speaker, but I think the principles are the same.,

By Aleks, Grazia, Yu-Sung and myself. Here's the article, and here's the abstract:

We propose a new prior distribution for classical (nonhierarchical) logistic regression models, constructed by first scaling all nonbinary variables to have mean 0 and standard deviation 0.5, and then placing independent Student-t prior distributions on the coefficients. As a default choice, we recommend the Cauchy distribution with center 0 and scale 2.5, which in the simplest setting is a longer-tailed version of the distribution attained by assuming one-half additional success and one-half additional failure in a logistic regression. Cross-validation on a corpus of datasets shows the Cauchy class of prior distributions to outperform existing implementations of Gaussian and Laplace priors.

We recommend this prior distribution as a default choice for routine applied use. It has the advantage of always giving answers, even when there is complete separation in logistic regression (a common problem, even when the sample size is large and the number of predictors is small), and also automatically applying more shrinkage to higher-order interactions. This can be useful in routine data analysis as well as in automated procedures such as chained equations for missing-data imputation.

We implement a procedure to fit generalized linear models in R with the Student-t prior distribution by incorporating an approximate EM algorithm into the usual iteratively weighted least squares. We illustrate with several applications, including a series of logistic regressions predicting voting preferences, a small bioassay experiment, and an imputation model for a public health data set.

I love this stuff, and I'm interested in applying the concept of weakly informative prior distributions for many other models.

This semester I'm teaching my "how to teach" class: The Teaching of Statistics at the University Level. (Stat 6600, or those of you here at Columbia.) I'll post more on that in a bit. Here I want to talk about an idea I had as I was falling asleep last night, of a new course I'd like to teach sometime.

The new course will be called Statistical Communication and it will cover the following topics:

- Graphical presentation (not just of raw data, also visualization of inferences)

- Writing research reports

- Writing computer code that can be used by others

- Working with colleagues (including "consulting" but also research collaborations)

- Email, blogging, hallway conversations, and other informal interactions

I think there was some other aspect of statistical communication that I wanted to include that I can't remember right now. The big idea is that maybe something is to be learned by thinking about all these activities as modes of communicating statistical ideas.

Helen DeWitt links to this interesting exercpt from a book by Tomasz Kamusella about the politics of language in central Europe. The basic idea is that we're all too used to thinking that a country should have a single language, and the exceptions (for example, Canada, Belgium, old-time Austria Hungary) seem weird to us. For example, it's always seen as a big joke in the U.S. that some people in Canada insist on speaking French. China shouldn't be a joke but, hey, they all speak "Chinese," right? And India doesn't really count because they're all supposed to speak English. Anyway, India's not just a country, it's most of a subcontinent, so that's different. And African countries have "tribes" so that doesn't really count either. And, sure, they speak 23 languages in Guatemala, but the official language is Spanish, so that's fine, right? Back when Russia was the U.S.S.R., I certainly had no idea that they spoke Ukranian and all those other languages there. And of course lots of people in the U.S. get upset that people insist on speaking Spanish here.

Kamusella writes,

Although the Western European pedigree of politics of language is at present conveniently forgotten, the phenomenon of language politicization is said to be now most visible in Central Europe. It is so because after World War I, the formerly multilingual Western European powers of France and the United Kingdom with the support of the United States chose to delegitimize the existence of Austria-Hungary on the account of its multilingualism and multiethnicity. By the same token, the victorious powers legitimized various ethnonational (formerly, often marginal) movements, which defined their postulated nations in terms of language. The national principle steeped in the ideal of ethnolinguistic homogeneity allowed these movements to carve up Central Europe into a multitude of ethnolinguistic nation-states. What followed with vengeance was forced ethnolinguistic homogenization pursued to assimilate 'non-national elements' within a nation-state. . . .

The declaration of more than one language per person was not permitted, which by default excluded the phenomenon of bi- and multilingualism from official scrutiny. The logic of this exclusion stemmed from the conviction that a person can belong to one nation only. By the same token, declarations of variously named dialects, already construed as 'belonging to' a national language, were noted as declarations of this national language. . . .

And then some statistics:

Nowadays, in comparison to the majority of extant polities worldwide, most of the nation-states of Central Europe are unnaturally homogenous in their ethnolinguistic composition. Non-Polish-speakers constitute less than 1 percent Poland's population, non-Magyar-speakers amount to 2 percent of Hungary's inhabitants, non-Czech-speakers are less than 3 percent in the Czech Republic's populace, non-Romanian-speakers constitute less than 11 percent of Romania's inhabitants, and non-Slovak-speakers amount to less than 15 percent of Slovakia's populace. . . .

I like to say I speak 1 3/4 languages. I wish I could speak more. But, until reading this, I'd always thought of monolingual countries as a default rather than a construct. Interesting stuff.

Chris Blattman writes,

Several aspiring graduate students have written me [Blattman] about becoming an impact evaluator. . . . I think the best advice is: don't get a PhD to do evaluations. The randomized evaluation is just one tool in the knowledge toolbox. . . . Yes, the randomized evaluation remains the "gold standard" for important (albeit narrow) questions. Social science, however, has a much bigger toolbox for a much broader (and often more interesting) realm of inquiry. . . .

I pretty much agree with Chris on the substance of his remarks, but I think he's missing something when he merges "impact evaluation" and "randomize evaluation" into a single concept. Policy analysis is a big area, and it certainly includes observational studies. We care about the impacts of all sorts of policies that can't be directly studied using experimentation.

P.S. In a different direction, it's interesting to me that policy evaluation is considered part of economics (a little bit) but not really part of political science--but maybe things are changing.

I'll be speaking on Red State, Blue State on Mon 12 Jan at the Medical University of South Carolina, Dept of Biostatistics, Bioinformatics, and Epidemiology. The location is Cannon Place, room 301, and the time is noon.

Nate Silver and Greg Mankiw have an interesting exchange about the use of exogenous instruments to estimate causal effects. Unfortunately, the subject is macroeconomics, a topic on which I know next to nothing beyond what I learned in Mr. Cutlip's econ class in 11th grade. But I think it is, in Greg's phrase, "a teachable moment" on the subject of causal inference.

Greg summarizes the exchange pretty well, although I think he's missing a key point.

Nate noticed a newspaper article where Greg related research by Christina and David Romer on the effects of "exogenous" tax cuts on the economy. Nate writes:

The type of tax cut that Romer and Romer think falls into this category is what they call an "exogenous" tax cut -- one designed not to counter business cycles, but rather a "spontaneous" tax cut under relatively healthy economic circumstances.

This is very much not the type of tax cut that we are contemplating right now. Instead, what is being contemplated is a countercyclical action in an unhealthy economy designed to return the economy to normal growth. Romer and Romer are not all that keen on this type of tax cut; in fact, they argue that such "countercyclical fiscal policy is not achieving its intended purpose" . . .

Greg repiies:

Why did the Romers focus on exogenous policy changes? The reason is that these are the only changes that can be used to reliability identify the effects of tax policy. . . . The Romers focus on exogenous tax changes for the same reason doctors conduct randomized drugs trials--not because they are interested in randomization as a prescriptive tool, but because randomization solves a statistical identification problem.

And now here are my thoughts, again with full recognition that I can really only comment on the statistical issues here, not the economics.

First, Greg is right that it is generally considered desirable or even optimal to estimate treatment effects using randomized experiments or exogenous implementations (but see here for an opposite view from James Heckman), even when the ultimate goal is to understand how the intervention works in the wild, so to speak.

But there is the potential for treatment interactions--that is, a treatment might be more effective in some conditions than in others. There's lots of evidence for treatment interactions in various settings, ranging from education to job training. And this is what Nate is talking about. Again, without attempting to comment on the economics, the treatment effect could vary enough that Nate could be right about the direct relevance of the Romers' study of exogenous tax changes.

To put it another way, Greg is talking about identifiability and Nate is talking about generalizability.

Greg writes, "I usually don't respond to blogosphere commentary on my work because, after all, time is scarce." But since he's had time to respond once, perhaps he'll be able to respond again and clarify this issue. (I think my time is particularly non-scarce since I'm responding to blogosphere commentary on somebody else's work!) In any case, I like the idea of shifting the debate to a discussion of treatment interactions since then it might be more possible to resolve this on a technical level. Perhaps a teachable moment for me as well as for others.

Seth writes:

I'm writing a critique of how epidemiologists analyze their data and one thing they rarely do is provide scatterplots. I suppose their excuse is they have too many points. Do you know of any papers about how to make scatterplots with large numbers of points?

My reply: The book, Graphs of Large Datasets: Visualizing a Million (which I've been planning to review on this blog for literally over two years, I even took notes on it and everything) discusses this issue in details, including tricks such as alpha-blending.

My impression is that if you have millions or even thousands of points, a density plot can do the trick.as in page 149 of Red State, Blue State. Perhaps readers have other suggestions. But if you just want to make the point and give a definitive reference, I'd go with the Graphics of Large Datasets book.

It's all R all the time around here, as Chris Blattman asks:

How can busy economists and political scientists learn R quickly? Is there a good guide? A set of handy ready-made programs? I learned MATLAB and STATA inside out in econ grad school, but now that I'm in a poli sci dept, and because the technology has progressed, I feel like I should learn R. I tried to teach myself a year ago (with no aids) and it was not pretty, even though I am usually pretty good at these things. So I abandoned it. But the article suggests R is easy to use. Did I miss the magic instruction book? If you have time to post on this, it would be a real boon to me and others I am sure.

My suggestion is to start with John Fox's book, An R and S-Plus Companion to Applied Regression and follow up with my book with Jennifer Hill, Data Analysis Using Regression and Multilevel/Hierarchical Models, which is full of R code. Sure, lots of the code is messed up, but we're planning to put together a clean version soon. . . .

If anyone has other suggestions for Chris, feel free to say something in the comments.

In a comment here, Steve Polili of SAS writes, "neither SAS nor Anne [Milley] hates R or open source. We run on Linux and we love Apache. For a little more info on SAS and R, take a look at a followup on Anne's blog [entry]," where Anne writes:

First, SAS and I applaud the innovative contributions and passion of the R community, and users who apply R to solve problems. In a very real sense, we are grateful for R, as it provides a freely available venue for bleeding-edge and experimental data analysis methods, and underscores the increasing importance of advanced analytical and graphical methods in this age of massive data volumes.

Over the past three decades, SAS has made and continues to make many noteworthy contributions to advanced data computing. As a trusted supplier to a large and diverse set of organizations, SAS provides analysis software that has been refined over years of customer application and feedback. SAS software is fully supported and used daily in countless ways. SAS is also scalable to very large data sets (multi-threaded, grid-enabled, etc.). These issues remain very important to the organizations we serve.

This seems reasonable to me. There's certainly room in the world for SAS, R, Stata, and SPSS, as well as Fortran, C, Python, and even Excel. I find it useful to have different software for different purposes. I'd like R to have better data input and merging tools, but until I become aware of such tools, I'll continue to use Stata (indirectly) to do some of this.

So, in that case, how could I go around saying, "I just hate SAS"?

In May 2005, I wrote:

After the 2004 election I had this idea that Bush's victory over Kerry is analogous to Truman's over Dewey in 1948. . . . A key issue in the analogy, I suppose, is the implicit suggestion that Bush in 2005-2008, like Truman in 1949-1952, could see a continuing decline in approval followed by a loss in his party's control of Congress.

Bush's popularity has sometimes been compared to Nixon's, but I think Truman's a better comparison because it is gradual rather than the result of a single scandal.

In response to my comments here, Kevin writes:

On Andrew's point about interpretation: Bob Carpenter is correct in his post that we were trying to convey the idea that our method is (in our opinion) an improvement over previous approaches that tended to be quite imprecise. Dan and I would not claim that our method is arbitrarily precise.

On Andrew's point about the MCMC run length: our choice of 4 million scans is more than necessary. We simply did this because we had the available computer time.
Diagnostics looked fine on a 1 million scan run and continue to look fine on the longer run. Results don't change from the 1 million scan run to the 4 million scan run.

On Andrew's point about parameterization: I think the suggestion is to do something similar to what Andrew and coauthors suggested here (Section 2). I think one could do something like that here although we really tried to keep the model as simple as possible while still being an accurate representation of the data. As a side note, we did a number of things (including posterior predictive checks) to assess model fit-- only some of these things made it into the paper.

I think the appendix of the paper deals with Bob Carpenter's question about data collection. See also the discussion about data collection in this paper that is forthcoming at Stanford Law Review.

Zubon is right to note that what counts as "centrist" depends entirely on what the references points are. We try to be as clear as possible about this in the body of the paper. Note that the 0 position on the scale is arbitrary-- there's no reason to think that 0 is "the center". The data can speak only to relative locations.

BTW, the version of the paper that Bob Carpenter linked to here is the final QJPS version. This should be freely available to everyone.

Bob pointed me to this article by Ashlee Vance following up on the recent newspaper article. Bob writes that "Pfizer's bought into the FUD (fear, uncertainty and doubt) argument that marketers employ to discourage the use of open-source or other free software." From the Vance article:

Pfizer was a prominent R user mentioned in the story. The company relies on R for its nonclinical drug studies and has shied away from using the technology for clinical research that will ultimately be presented to regulators. For such work, Pfizer instead turns to software from SAS Institute, which brings in more than $2 billion a year in revenue from data analytics software and services.

Were Pfizer to use R in clinical studies, it would run the risk of seeing its research questioned or even rejected by regulators doubting the veracity of results based on what they view as an unknown quantity.

"It's very hard to displace the industry standard in those types of cases," said Max Kuhn, associate director of nonclinical statistics at Pfizer.

I'm actually working with Neal Thomas and other people at Pfizer on an improved and more trustworthy OpenBugs implementation that they can use for their research. It's actually worth it for Pfizer to put resources into an open-source project: open-source can mean more beta-testing and more reliability.

At a technical level, Sam Cook and I are working with them on implementing unit testing (the so-called "self-cleaning oven"; see item 5 here) for Bayesian modeling, following our earlier work in this area.

Finally, Vance concludes with a discussion of the size of R's user community. I imagine this is tricky to define--for example, do you count students?

From my forthcoming book co-edited with Jeronimo. . . This part is by Chuck Cameron and myself:

What percentage of votes cast in national elections in Germany's Weimar period were cast in favor of the Nazis or other antidemocratic parties? The answer is: 33% for the Nazi Party and 17% for the Communist Party of Germany in the eighth German Federal Election on November 21, 1932, and 44% for the Nazi Party and 12% for the Communist Party of Germany in the ninth German Federal Election on March 5, 1933. In other words, a large portion of the electorate did not support antidemocratic parties.

This factoid suggests a very different world from one in which an overwhelming majority of the German public voted for the Nazis. Knowing this factoid might lead us to ask a question requiring systematic data: When democracies perish, how much support for the antidemocratic forces is there in the public at large? This question has a real normative punch, because it is another way of asking whether mass electorates can be trusted with democracy. And then: If mass electorates can be trusted and the problem is antidemocratic elites, how can we protect democracy from its elite enemies?

An interesting related question, perhaps falling closer to psychology or history than political science, is why it is such a common belief that the Nazis won in a democratic election. Perhaps this (false) belief is popular because it leads to an interesting story and a "paradox" of democracy - What should be done if an antidemocratic party wins in a democratic election? - or perhaps it simply arose from a misunderstanding of historical writings.

John Shonder points to this article by Carl Bialik discussing this article by Steve Ansolabehere and Charles Stewart discussing the 2008 election. Ansolabehere and Stewart write:

Obama won because of race . . . Obama captured ten million more votes in 2008 than John Kerry did in 2004, resulting in a 4.6 percentage point swing toward the Democrats from 2004 to 2008. This swing did not occur similarly or uniformly among all politically relevant groups, as forecasting models might suggest. Most of the additional Democratic votes were cast by black and Hispanic voters--4.3 million and 2.7 million more, respectively. Democrats also gained among white voters, but the increase was a modest 3 million votes. . . . Obama gained not only by bringing new minority voters into the electorate, but also by converting minority voters who had previously been in the GOP stable.

This is consistent with instant election-night analysis (see item 4 here).

Ansolabehere and Stewart also write, "had Blacks and Hispanics voted Democratic in 2008 at the rates they had in 2004 while whites cast 43 percent of their vote for Obama, McCain would have won." I don't think that's really a reasonable model, though, because that would be assuming that Obama would've outperformed Kerry more among whites than among nonwhites, which hardly seems plausible. To put it another way, Obama's baseline swing among any group is his national swing, not zero. Given the state of the economy in November 2008, zero just doesn't make sense as a baseline.

Similarly, Ansolabehere and Stewart write, "Had Obama relied only on a surge among young voters, holding other groups at the 2004 voting behaviors, he would have fallen short of victory." Again, I think this is slightly misleading: Obama's strategy was not to do better only young voters but rather to improve upon Kerry's performance in general, but piling up a particular margin among the young. Which is what he did.

You can also slice up the vote swing geographically, by counties in different regions of the country, and you find that Obama did close to uniformly better than Kerry nearly everwhere, except for Republican-leaning poor counties in the South (where Obama pretty much stayed even with Kerry). The geographic patterns are striking (see graph at the end of this post).

Race matters, yes, but we're still seeing a national swing.

Finally, I noticed that some of Bialik's commenters focused on Obama's racial appeal. I'd like to remind them that the Democrats gained even more in elections for the House of Representatives (compared to 2004) than Obama gained on Kerry. The House gains just weren't so obvious because they were spread over two elections.

2008 was a Democratic year, Obama was a Democrat, and he won in one of the ways the Democrats could've won. With a different candidate there might have been different demographics but roughly the same national swing, and maybe a slightly different electoral map with a similar electoral vote total.

I think Ansolabehere and Stewart are right on the money when they write, "the results of the 2008 election challenge much of what has been conventionally thought about race and politics in America. Barack Obama has accomplished an astonishing political move [by] disproportionately energizing nonwhite voters and converting erstwhile Republican supporters within the minority community without alienating white voters."

My summary: as Carl said, the election outcome is multidimensional. Because Steve and Charles were writing a short article, they very properly focused on a single feature of the election--race. I'd say that the #1 feature of the election was a bad economy that produced a national swing toward the Democrats in general and Obama and particular. But once you want to break this down by demographics, I agree that ethnicity is the biggest factor.

P.S. Comments here from John Sides. who links to this article by Mark Blumenthal and this by Marc Ambinder. John writes that "Most likely, the economy and race both mattered. Andy sees the economy as more important. I'm inclined to agree, but ultimately time, and more evidence, will tell."

My response: 'd say the economy was more important in determining the ultimate outcome of the election, and that race was more important in describing relative differences between the Obama and Kerry vote.

That is, the economy predicted the uniform partisan swing, and race described much of the discrepancies from uniform partisan swing.

P.P.S. Here's further discussion from Blumenthal.

See here. I pretty much agree with what they're saying, except that I think R occupies a position as much as it serves a function. By this I mean that, if R didn't exist, we'd be doing similar things using something else, whether it be Matlab, Mathematica, or some Python-based confection. When I think about what I actually do in R, it wouldn't actually be so hard to do most of it from scratch. This is not to disparage R, just to say that it's filled a niche.

And I certainly wouldn't characterize R as "a supercharged version of Microsoft's Excel spreadsheet software." Or maybe I should say that I didn't know that R had spreadsheet capabilities. One more thing to learn, I guess. Also another motivation for Jouni to finish Autograph.

And it's good to hear that SAS is in trouble. I just hate SAS. It's not just a matter of it having poor capabilities; SAS also makes people into worse statisticians, I think.

P.S. The reporter contacted me about this story a few weeks ago, but I don't actually remember what I said (or even whether I was ever actually reached). Certainly nothing memorable enough to quote.

John Sides posts this graph:

You will perhaps not be surprised to hear that I have no comments on the substance but I have some thoughts on the presentation. I'd bound the y-axis at 0 and 100% (currently it goes beyond these limits), also I'd put the year labels between the hash marks rather than on them (think about it: on this scale, 1995 is a time period, not a single point), also I'd put percent signs on the y-axis (e.g., "25%" rather than "25") for some useful redundancy. But other than these minor comments, I think the graph is beautiful.

The year-labeling issue is not completely trivial, especially when trying to interpret when the series ends. I've noticed that people often have difficulties representing time on the x-axis of graphs. Other times, you'll see, for example, a graph going from 1950 to 2000 with 50 little hash marks and tiny slanted labels at 1951, 1953, 1955, 1957, etc. Instead of simply labeling every 20 or 25 years.

Daniel Ho and Kevin Quinn write:

We amass a new, large-scale dataset of newspaper editorials that allows us to calculate fine-grained measures of the political positions of newspaper editorial pages. Collecting and classifying over 1500 editorials adopted by 25 major US newspapers on 495 Supreme Court cases from 1994 to 2004, we apply an item response theoretic approach to place newspaper editorial boards on a substantively meaningful--and long validate--scale of political preferences.We validate the measures, show how they can be used to shed light on the permeability of the wall between news and editorial desks, and argue that the general strategy we employ has great potential for more widespread use.

Here's their key graph, which aligns the estimated ideological positions of major newspapers with recent Supreme Court justices:

They used Bayesian ideal point estimation. Their main substantive conclusion:

Our book will be discussed on the Firedoglake book salon this Saturday, 10 Jan, from 5-7pm Eastern time. I don't know exactly how this works, but my impression is that you can email in your questions/comments and I'll be online to answer them. The discussion will be moderated by Matt Yglesias.

I was pointed to Paolo Guardia's excellent Data Protection, Information Privacy, and Security Measures: an Essay on the European and the Italian Legal Frameworks. Here's an excerpt:

Data Protection Principles

Data protection regulations in the EU set the main principles that establish how data processing shall be performed. We can summarize privacy principles as follows:

• Fair and Lawful Processing: the collection and processing of personal data shall neither unreasonably intrude upon the data subjects' privacy nor unreasonably interfere with their autonomy and integrity, and shall be compliant with the overall legal framework.

• Consent: personal data shall be collected and processed only if the data subject has given his explicit consent to their processing.

• Purpose Specification: personal data shall be collected for specified, lawful and legitimate purposes and not processed in ways that are incompatible with the purposes for which data have been collected.

• Minimality: the collection and processing of personal data shall be limited to the minimum necessary for achieving the specific purpose. This includes that personal data shall be retained only for the time necessary to achieve the specific purpose.

• Minimal Disclosure: the disclosure of personal data to third parties shall be restricted and only occur upon certain conditions.

• Information Quality: personal data shall be accurate, relevant, and complete with respect to the purposes for which they are collected and processed.

• Data Subject Control: the data subject shall be able to check and influence the processing of his personal data.

• Sensitivity: the processing of personal data, which are particularly sensitive for the data subject, shall be subject to more stringent protection measures than other personal data.

• Information Security: personal data shall be processed in a way that guarantees a level of security appropriate to the risks presented by the processing and the nature of the data.

Will the pervasive data mining on the web ever become compliant?

John Quiggin sent me this article of his from 1987 that made the same argument as my paper with Edlin and Kaplan on why and how it's rational to vote. In his article, Quiggin wrote:

There is strong evidence that voting behaviour is both ends-directed and rational. That is, electors choose to vote because of the effects their vote will have, and do not vote if these effects are insufficient to outweigh the costs of voting. However, as Downs' paradox shows, rationality and egoism together imply non-voting. The evidence suggests that egoism is the postulate which must be abandoned. . . . voters' interest in political information increases with the importance of political choices. Once again, this is consistent with rationality but not with egoism.

Our article had more math and more focus on U.S. politics but the basic point is the same.

Also let me use this as yet another excuse to plug a wonderful article, The Norm of Self-Interest, by psychologist Dale Miller, in which he argues the following:

A norm exists in Western cultures that specifies self-interest both is and ought to be a powerful determinant of behavior. This norm influences people's actions and opinions as well as the accounts they give for their actions and opinions. In particular, it leads people to act and speak as though they care more about their material self-interest than they do.

NY Times ran an article Risk Mismanagement:

VaR uses this normal distribution curve to plot the riskiness of a portfolio. But it makes certain assumptions. VaR is often measured daily and rarely extends beyond a few weeks, and because it is a very short-term measure, it assumes that tomorrow will be more or less like today. Even what's called "historical VaR" -- a variation of standard VaR that measures potential portfolio risk a year or two out, only uses the previous few years as its benchmark.

Nonsense. VaR is an innocent and useful mathematical construct, completely independent of the distribution or the model used. It can be as simple as subtracting variance from the mean to penalize for the risk of a distribution. Don't throw the baby (VaR) out with the bathwater (dubious VaR practices).

The real failure of risk management was in the bad short-tailed models (that underestimated the probability of a default) that they fit in a bad way (overfitting to a small amount of one-sided historic data, without using priors that would include the possibility of a disaster).

But even once these problems are fixed, economy will still be a feedback system, not something pretty, simple and linear. Let's hope for a marriage of statistical modeling with systems and complexity theory. In the meantime, I'll hope for using more common sense based on substance and less mathematics based on arbitrary metrics. This would help prevent disasters in the first place.

Ted Dunning sent me this graph:

So, how do the polling data compare to the contract prices from Intrade on the day before the election? Below is a graph with a data point for each state, with the horizontal axis representing the polling data and the vertical axis representing the Intrade contract price.

The quick message that I get from here is that Intrade prices are way biased toward 50/50. For example, the price for DC is something like .04, which is ridiculous. (To two decimal places, it should certainly be .00).

John Quiggin writes:

The strong version [of the efficient markets hypothesis], which gained some credence during the financial bubble era says that asset prices represent the best possible estimate taking account of all information, both public and private. It was this claim that lay behind the proposal for 'terrorism futures' put forward, and quickly abandoned a couple of years ago. It seems unlikely that strong-form EMH is going to be taken seriously in the foreseeable future, given the magnitude of asset pricing failures revealed by the crisis.

I have two comments:

1. It was my impression that under classical economic theory, the economy is always at a phase transition (by analogy to ice water): there's said to be enough "water" (i.e., trading) that prices reflect a consensus, but there's enough "ice" (i.e., new information entering the system) that it is rational for prices to move. I don't know any of the theory beyond this, but I imagine that much of the debate must center on how large the fluctuations are in this phase transition. In may mathematical systems (although not ice water, I think), these fluctuations can be large.

2. A point that I always thought was under-emphasized is that the proposed terrorism futures markets were to be run by by convicted criminal John Poindexter--a guy who was actually involved in what were arguably terrorist activities (the project of secretly sending weapons to Iran in the 1980s). The defenders of the terrorism futures markets never seemed interested in this point; see, for example, the comments here. But to me this was a pretty serious issue: really we're talking about an unrepentant criminal here. I'm not saying that terrorism futures are necessarily a bad idea, but I was highly skeptical of that particular implementation.

A couple of months ago, my article on the probability of a single vote being decisive in the presidential election (at most 1 in 10 million, according our calculations) was picked up by the Associated Press, and shortly after I received the following email from Deron Reynolds, a pilot in the U.S. Air Force:

Seth writes:

A friend of mine who works for one of these newspapers said that the end has been coming for a long time. In the early 1990s, if I remember correctly, the audience started to shrink. At the time, and for a long time thereafter, this was ignored. Had the problem been recognized back then it might have been possible, given a lot of time to experiment, to find a solution, a way to survive much longer. But now it is too late.

I don't know about this. I've been reading for at least 10 years about the decline of newspapers. Newspapers and magazines have had a lot of stories on this topic for awhile. In the 1990s, the story was how daily newspapers were reducing the number of reporters because investors were demanding huge profit margins. Everyone understood that there were long-term problems; if nothing else, old newspapers were disappearing faster than new ones were being created.

So, I think lots of people did know about the problem--it was not being ignored. But it wasn't clear what to do about it. "Time to experiment" sounds like a good idea in theory, but in practice there was no solution available.

Some fun things you might have missed . . .

No justice, no foul

Ed Park is a Democrat

Walter de la Mare was a statistician

Just quaid

The effectiveness of peacekeeping

The statistics of art

Objects of the class "Whoopi Goldberg"

The "all else equal" fallacy

Basketball statistics

Oh yeah, also we wrote a book a few months ago. I can't recall if I mentioned it on the blog.