September 2008 Archives

Marshal Zeringue asked, and I replied:

Phil Klinkner writes:

History doesn’t repeat itself, the saying goes, but it does rhyme.

When I was about 9 years old, I read just about every book of fairy tales in the library. 398.2 in the Dewey decimal system, I remember it well.

Brendan Nyhan offers this amusing example of a newspaper hyping poll noise. From the LA Times:

Registered voters who watched the debate preferred Obama, 49% to 44%, according to the poll taken over three days after the showdown in Oxford, Miss.

That is a small gain from a week ago, when a survey of the same voters showed the Democratic candidate with a 48% to 45% edge.

A small gain, indeed.

From Mark Blumenthal.

Henry Farrell referred here to his blog as a "place." Which seemed funny to me because I think of a blog as a "thing." Henry replied:

That's the way that I [Henry] think about blogs (or at least group blogs and blogs with comments) - places where people meet up, chat, form communities, drift away from each other etc.

My analogy was blog-as-newspaper, the self-publishing idea, and I'm not used to thinking of a newspaper, or even a listserv, as a place. I think there is an aspect of the analogy that I'm still missing.

P.S. See Mark Liberman's thoughts in his blog here.

See here for my failed attempt to construct a political conspiracy theory around Lehman Brothers and the financial crisis.

My blog discussion with Eyal Shahar (see comments #3 and onward here) reminded me of a persistent challenge I face when talking with outsiders about Bayesian statistics.

Laura Wattenberg writes, "in baby naming as in so many parts of life, style, not values, is the guiding light."

Dan Lakeland writes:

I am working with some biologists on a model for time-to-response for animals under certain conditions. The model(s) ultimately are defined in terms of a differential equation that relates a (hidden) concentration of a metabolic product to the (cumulative) probability that an animal will respond within a given time by changing its behavior.

Here.

Chris Zorn pointed me to this graph and asked for my thoughts. I replied that I'd seen worse, but the use of two dimensions doesn't help, and the comparison to the GDP of Kazakhastan is just weird. I mean, who has any idea what is the GDP of Kazakhstan??

Chris replied,

I'm teaching first-year Ph.D. methods on PoliSci this term, and we have a feature called "Graph of the Day," where -- for five minutes or so at the beginning of every class -- the students all look at and comment on some graph from a paper, the press, etc. I used this one yesterday, and the response (from people with a grand total of three weeks of graduate education) was identical: "What's up with Kazakhstan?", and "Isn't a reference point supposed to be *non-obscure*?"

Jamie points out this interesting article by Douglas Oxley et al. that appeared in Science last month. Here's the abstract:

Although political views have been thought to arise largely from individuals? experiences, recent research suggests that they may have a biological basis. We present evidence that variations in political attitudes correlate with physiological traits. In a group of 46 adult participants with strong political beliefs, individuals with measurably lower physical sensitivities to sudden noises and threatening visual images were more likely to support foreign aid, liberal immigration policies, pacifism, and gun control, whereas individuals displaying measurably higher physiological reactions to those same stimuli were more likely to favor defense spending, capital punishment, patriotism, and the Iraq War. Thus, the degree to which individuals are physiologically responsive to threat appears to indicate the degree to which they advocate policies that protect the existing social structure from both external (outgroup) and internal (norm-violator) threats.

I myself am extremely sensitive to sudden noises, so make of that what you will . . . Seriously, though, this seems related to John Jost's work on personality profiles and political affiliation.

Here. See here for my thoughts.

When the sign says the train will be 0:05 late, it won't be 0:05 late. If it were going to be 0:05 late, they wouldn't say anything at all. In reality it will be 0:30 late. But they won't say it will be 0:30 late, because that would mean the train will be 1:30 late.

I got off a good line when I got on the train. I stepped in, saw a retirement-age couple already seated, and asked, Philadelphia? They said, yeah, that's where they're going, but they're not sure either, they hope they're in the right place. I said, yeah, I think this is right. (Pause) And, if there's one thing last week's news has taught us, it's that you can trust a guy in a suit.

That got a laff.

Ads in the Newark train station

A big picture of a hot-dog guy holding a mustard-slathered beauty, next to the words: You Want Cancer With That? Medical research shows hot dogs increase your risk of cancer... [An ad for some law firm that's suing food manufacturers.]

The Retreat at Princeton
Inpatient Alcohol and Drug Treatment for Executives & Professionals
www.RetreatAtPrinceton.com

The following is my discussion of the article, "Approximate Bayesian inference for latent Gaussian models by using integrated nested Laplace approximations" by H. Rue, S. Martino and N. Chopin, for the Journal of the Royal Statistical Society:

Statisticians often discuss the virtues of simple models and procedures for extracting a simple signal from messy noise. But in my own applied research I constantly find myself in the opposite situation: fitting models that are simpler than I would like—models that clearly miss important features of the data and, more importantly, important features of the underlying system I am modeling—because of computational limitations.

I'll be talking about Red State, Blue State on the Kathleen Dunn show on Wisconsin Public Radio tomorrow (Tues 23 Sept), from 10-11 Central Time (that's 11-12 Eastern Time). For the second half of the show, you can call in with questions!

Here.

I'll be speaking on the Red State, Blue State book this Monday (22 Sept) at 4:30pm at the University of Pennsylvania. It'll be at the Annenberg School for Communication, Room 109. The address is 3620 Walnut Street, Philadelphia, PA. This is your chance to ask questions and also to meet some interesting people: the talk is cosponsored by the departments of Statistics, Biostatistics, and Political Science as well as the Annenberg School.

Larry Bartels writes about how "the contemporary electoral landscape, which is less volatile and more partisan than it has been at any time in the past half-century or more." Larry's presentation is clean and well illustrated by graphs, adding nicely to earlier discussion of this topic by John Sides.

Larry also has some comments about the problems that can occur when a historian is "moonlighting as a political scientist." Which reminds me of my own rants:

Every four years, some hardworking and enterprising journalists do some digging around in the political science literature, talk with some people who sound like they know what they're talking about, and then resurface to tell the world about the counterintuitive finding that the Electoral College actually benefits voters in large states.

Well, as I like to say to my social science students: Just 'cos it's counterintuitive, that don't make it true.

The Electoral College benefits voters in swing states, and it slightly benefits voters in small states (on average). Large states are not benefited (except when they happen to be swing states such as Ohio or Florida, but we knew that already).

See here for the fuller discussion.

I just wanted to put this out here to get out in front of the discussion. So that if any of you do see this argument floating around, youall can shoot it down before it fully takes off...

A conference celebrating the 100th birthday W.S. "Student" Gosset's "The Probable Error of a Mean" and three other classic papers:

Here's a pretty picture (from Charles Franklin, link from John Sides):

What a great graph! I won't be picky, but if I were, I'd make the following suggestions:
- Bigger numbers on the axes--as is, they're hard to read.
- Add percentage signs on the y-axis.
- Label age every 20 years rather than every 10.
- Put the "80-84" age group at 82 (rather than 80), and put the "85 and up" group at 88 (rather than 85).
- Pick colors other than red and blue.

Dipankar Bandyopadhyay writes:

I am currently running an autologistic regression model where I have some fixed effects and also spatial (autlogistic) terms. Is there any recommendation from you on the appropriate choice of prior on the variance when I put a normal prior on the regression coefficients? I mean, do you recommend a folded-t, or a half-cauchy, or a uniform over the traditional inverse gamma, and in such a case, where can I get the WinBUGS codes to put folded-t, or half cauchy/half-normal priors?

I've blogged about this before but it's worth mentioning again as a good teaching example. The site How Many of Me purports to estimate how many people in the U.S. have any particular name. But it can give wrong results; as "the other Craig Newmark" noted, it said there was only one of him, and there are actually at least two.

What the site actually does is to plug in esitmates of the frequency of the first name and the frequency of the last name and assume independence. The results can be wrong.

This could be a great example for teaching probability. Three questions: first, how can you check that the site really is assuming independence; second, how many people does the site assume are in the U.S.; third, how could you do better?

1. How can you check that the site really is assuming independence? We'll check four names and see how many it says there are of each:

Rebecca Schwartz: 171
Rebecca Smith: 6600
Mary Schwartz: 1047
Mary Smith: 40941

Calculate the ratios: 6600/171=39, 40941/147=39. Check.

Actually, to one more digit, the ratios are 38.6 and 39.1. Why the difference? Shouldn't they be exactly the same? Playing around with the last digits reveals that it can't be simple rounding error. Maybe some internal rounding error in the calculations? (Perhaps another good lesson for the class?) Hmm, let me go back and check. Number of Mary Schwartzes: 1047. Check. Number of Mary Smiths? 40491. Uh oh, I'd transposed the digits when copying the number. Now the ratios agree (to within rounding error)

The website is definitely assuming independence. I have no doubt that there are some Mary Schwartzes out there but no way that the frequencies of Marys among Smiths and Schwartzes is exactly identical.

2. How many people does the site assume are in the U.S.? The site says there are 4,024,977 people in the U.S. with the first name Mary, 3,069,846 people in the U.S. with the last name Smith, and 40,491 Mary Smiths. 4024977*3069846/40491 = 305 million. So that's what they're assuming.

3. How could you do better? Phone books are an obvious start. They don't have everybody and there are other sampling difficulties involved (for example, a telephone that's under the name of only one person in the family, leaving the others unlisted) but it would give you some clear information about how large are the discrepancies from indepdence.

And, a bonus:

4. A bad idea (which might be tried by a naive instructor who doesn't get the point): Using this to teach the chi-squared test for statistical independence. This is a bad idea for two reasons: first, the data in HowManyofMe.com are not a sample under statistical independence; they are exactly statistically independent (a/b=c/d) and so a chi-squared test is beside the point. Second, for real data the point is not whether they could be explained by statistical independence--they can't--but how large the discrepancy is. This can be expressed using probabilities or odds ratios or whatever but not by the magnitude or the p-value of a chi-squared test. (If you want to use this example to illustrate chi-squared, this is the point you'd have to make.)

P.S. I've never met the other Andrew Gelman, but I did once meet someone who lives down the street from him (in New Jersey).

A student writes:

One of my most favorite subjects is model-selection. I have read some papers in this field and know that it is so widely used in almost every sub-field in statistics. I have studied some basic and traditional criterion such as AIC, BIC and CP. The idea is to set a consistent optimal criterion, usually it's not easy when the dimensionality is high, but my question is, what is the biggest problem and why it is so hard?

Jim Manzi says yes, and he has some data. He says that in 46 out of 48 states, there's a positive correlation between a county's neighborhood-level inequality and its vote for Kerry.

P.S. Also see interesting thoughts in the comments section below.

P.P.S. This paper by Mark Frank also seems relevant to the discussion. Frank writes:

For many states, the share of income held by the top decile experienced a prolonged period of stability after World War II, followed by a substantial increase in inequality during the 1980s and 1990s. This paper also presents an examination of the long-run relationship between income inequality and economic growth. Our findings indicate that the long-run relationship between inequality and growth is positive in nature and driven principally by the concentration of income in the upper end of the income distribution.

P.P.P.S. See also the graphs here (from chapter 5 of the Red State, Blue State book).

Bob Carpenter writes the following regarding Extreme Programming, focusing specifically on some of my struggles in statistical computing:

Well, it's a bit extreme. I think you'll find better overall advice in Hunt and Thomas's Pragmatic Programmer without all the silver bullet rhetoric. I wouldn't bother with the Agile development stuff, but that's the currently trendy descendant of this whole line of thinking.

Having said that, there are several good take-away messages from the Extreme Programming (XP) process, but I don't believe, as its proponents do, that you need to do everything their way.

I love pair programming -- it's not only a great way to learn for a novice/expert or expert/expert pair, it's a great way to keep quality high by keeping each other honest and it's a great way to catch bugs early on. But if you follow the XP advice to the letter, that's the only way you'd program, which is impractical in most groups.

You should start using unit testing, which I believed your group referred to as a "self-cleaning oven" (though don't keep the tests in the same file as the code).

Research coding is both good and bad for XP. The specifications tend to move around even more than in the usual XP project, because you don't even know if what you're trying to do is possible before you start much of the time.

And you definitely need version control, which Masanao and Yu-Sung set up for you through RForge.

With some R and BUGS under my belt, I really wish it was easier to stick to the don't repeat yourself (DRY) principles. All of the R code I've seen could use much more modularity.

Finally, you should get into refactoring; it's really what you're trying to do with BayesGLM, though it may be easier to refactor GLM first if you're going to start from scratch.

I'll be speaking on Red State, Blue State this Wed, 17 Sept, 12-1:30, in the Government Dept at Harvard. It's at 1737 Cambridge St., Room K-354. If you live in the Boston area, this is your chance to come and ask your questions and give your suggestions.

Aleks points me to this paper by Jerry Friedman on non-Bayesian regularization methods. I'd also recommend our Bayesian approach (see this Annals of Applied Statistics paper). Once you're going to assume a probability model for the data (a likelihood), it's a pretty small step to include prior information as well. But read Friedman's paper in any case. He focuses more on computational issues than we do. There are really two parallel literatures.

Phil pointed me to this. Very nice. I have to get someone to program my emailer to do something similar...

Rachel writes that she gave our students (it's a grad class in applied statistics, based on the Gelman and Hill book) what she thought of as a "Taxonomy of Confusion"... types of things they might be confused about and what they should do before asking the T.A.:

1. statistics-related questions that are prerequisite to the course--get an Intro to Stats book, don't ask the T.A. unless you really must.

2. statistics-related questions that are part of the course--read the book, ask a friend, then ask the T.A.

3. you know what you want statistically but you don't know the name of the function--google "R standard deviation" or write the function yourself... if you can't find it, ask a friend then ask the T.A.

4. you know the function's name but you can't figure out how it works: type help(sd), then ask a friend then ask the T.A.

5. you wrote code but you get error messages: DEBUG using tips like, print things out, break into smaller steps (we should do more on this later).

6.you wrote code and it doesn't do what you think it should do but there are no errors: DEBUG (more on this later).

This just seemed hilarious to me. Maybe it was the deadpan tone.

Bruce McCullough writes:

Don't know if you're aware of this, but if you need more evidence for the primacy of interaction effects, data mining is a great place to look. My degree is in economics. I was taught to use interaction effects as a test for nonlinearity, and that was about it.

My data mining experience of the past few years has taught me that interaction effects can be neglected at my own peril. A wonderful paper that illustrates this is "Variable selection in data mining: Building a predictive model for bankruptcy," by Dean P. Foster and Robert A. Stine in the Journal of the American Statistical Association (2000). The usual linear regression doesn't work. The model with lots of interactions works very well.

In response to something Robin Hanson wrote on his blog (sorry I can't find the exact link, I think it was at the end of July, 2008), I wrote:

If you're in D.C., you should stop by. . . . I'm speaking in the statistics department at George Washington University on the topic of interactions. Here's the powerpoint and here's the abstract:

As statisticians and practitioners, we all know about interactions but we tend to think of them as an afterthought. We argue here that interactions are fundamental to statistical models. We first consider treatment interactions in before-after studies, then more general interactions in regressions and multilevel models. Using several examples from our own applied research, we demonstrate the effectiveness of routinely including interactions in regression models. We also discuss some of the challenges and open problems involved in setting up models for interactions.

The talk will be today, Wed 10 Sept, at 3pm at 1957 E Street, Room 212. If you don't know where that is, you can call the department (202-994-6356) and they should be able to give you directions.

Tomorrow (Thurs) I'll be speaking with Boris at noon at the Cato Institute on Red State, Blue State. It's not too late to sign up for that.

Rick Shenkman reminds us that voters are "grossly ignorant" about many issues. Now that the Cold War is over, we don't have to worry about voters not knowing about "throw weights" and such, but I think it's still probably a bad thing that "six in ten young people (aged 18 to 24) could not find Iraq on the map," that people overestimate by a factor of 50 the percentage of the federal budget that is spent on foreign aid, and so forth.

John Payne writes:

I am writing a Java program to do ecosystem modeling and we wish to use Bayesian MCMC methods for parameter estimation. I am interested in finding flexible, customizable Bayesian MCMC code that can be called in Java. I looked at documentation for JAGS, BUGS, and also Gregory Warnes's Hydra program (which is in Java). I have been unable to get a reply from Dr. Warnes but it seems Hydra is no longer being supported. As far as I can tell, BUGS is written in Component Pascal, which I am ignorant about. I have never tried JAGS; would you have any advice as to which avenue would be the most fruitful to pursue?

My reply: Right now, I think Jags is probably the best way to go. But others might have other suggestions here.

David Frum responded at his blog to my graph-laden comments on his New York Times article.

Frum emphasizes the difference between looking at county-level inequality as compared to state-level inequality. He also makes the point that inequality (at the state and county level) is often associated with big cities. Interesting stuff.

Frum also mentions Missouri, which is one of the states where richer counties favor the Democrats. Richer counties also lean Democratic in Nebraska, and most of the western and northeastern states (see pages 68-70 of the book), but in Indiana, South Dakota, Wisconsin, New Jersey, and most of the South, it goes the other way, with richer counties being more Republican. (I showed this in the map of Texas in my previous blog entry.) The patterns really do look different in different parts of the country, and Missouri is not like Texas in this respect. In any case, I haven't crunched the numbers on county-level inequality, and I agree with Frum that the patterns within a state can differ from those between states. Individually, richer Americans still lean Republican, but location matters a lot also.

David Frum, author of “Comeback: Conservatism That Can Win Again,” wrote an op-ed in the New York Times yesterday that has some interesting insights and but also suffers from some of the usual confusions about rich and poor, Democrats and Republicans. Overall I think Frum has some interesting things to say but I want to point out a couple of places where I think he may have been misled by focusing too strongly on the D.C. metropolitan area.

I read Richard Cook's biography of Alfred Kazin recently. It was surprisingly interesting--I say "surprisingly" because Kazin didn't live a particularly eventful life. I wanted to read the book in the first place because I like a lot of Kazin's writing and I wanted to understand how the pieces fit together. One thing I learned is that his sister married Daniel Bell. Not that the book featured any interesting anecdotes about Bell; still, it was satisfying to see the map filled in. I was struck by how financially precarious Kazin's life was. After the late 1930s, he was never poor, but it was a long time before he had a permanent job. There definitely seems to be a conceptual divide between those of us with steady jobs (the sort that pay us even if we're not really working) and people who start each year from baseline of zero income and have to earn every penny. (Well, I guess Kazin had book royalties, but I don't suppose that was enough to pay the rent.)

My favorite writing of Kazin's are his book reviews, especially of post-1950 literature, which is what I'm most likely to have read and to be able to relate to. (I just can't get into that Henry James stuff.) I'd love to read more of that. I have a collection of his reviews that came out around 1962, and it's excellent. (Not perfect; he sometimes irritates me with a smug all-knowing attitude of condescension, but most of the time it's interesting. For example, I got a lot out of his essay on John P. Marquand, even though Kazin is less of a fan of Marquand than I am. My take on this: Marquand made it look so easy that his skills were hard to appreciate until decades later, when nobody has come along to replace him.)

Cook takes a lot from the memoir, "What I Saw at the Fair," that Kazin's third wife, Ann Birstein, published a few years after Kazin's death. I went to the library and picked it up and gave it a quick read. She was still mighty angry at Kazin, even to the end, when she found out that he'd sold a collection of letters, including many from her, to the New York Public Library. It's gotta be a weird feeling to go to the library and come across your own decades-old letters.

"What I Saw at the Fair" is readable and interesting, but running through it is a funny idea--I'd call it pre-modern--that people's true essences are reflected in their physical appearance. Character after character is introduced as ugly or beautiful, and almost always this is an indicator to the inner being. This strategy works for Dickens, and in addition I'm willing to believe that there's some correlation between inner and outer beauty (especially given that both are in the eye of the beholder). But I know enough people to know that any such pattern is far from universally true. In reading Birstein's memoir, I was continually wondering whether she really believed that beautiful people are nicer, that ugly people compensated by being nasty, that Hannah Arendt was really "a Nazi," Along the same lines, she disparages Norman Mailer's machismo because his penis was small.

But what really struck me about Birstein's memoir is that she strongly identifies herself as a writer--she's published several novels--and she knew lots of writers and intellectuals, including Sylvia Plath, Bernard Malamud, Saul Bellow, and the aforementioned Daniel Bell--but she expresses no interest in any of their writings. Birstein's anecdotes about these people are interesting, but I'm surprised to see no discussion of their literature or their ideas. Perhaps this is her revenge on them for ignoring her writing all these years. In any case, I think she missed an opportunity. It would be like writing a book that takes place in the Giants' locker room and not talking about football. Birstein identifies being a writer with fiction writing and thinks it's funny that Kazin called himself a writer when he was only a critic. "A Walker in the City" has some beautiful phrases and images, but to me it doesn't read as smoothly as a good novel or even as smoothly as good criticism.

Kazin told Birstein that he couldn't love her if she weren't a writer (or something like that; I don't recall the exact wording), but he didn't show much respect for her actual writing. But maybe that has to do with Kazin's career as a critic of classic writing. If he was comparing to Mark Twain, Edith Wharton, etc., then it's no surprise that Birstein came off second best.

This brings me to a more general point. Birstein appeared to evaluate people based on their looks (or, perhaps, retroactively evaluated people's looks based on how much she liked them). Kazin perhaps evaluated Birstein unfairly because, as a writer, she was no Saul Bellow. Do we all do this sometimes? I evaluate statisticians based on their ability--not necessarily technical ability (although that's part of it) but more on whether they "get it" and can solve problems. And the statisticians who are really good at this? I like them as people, almost without exception. Conversely, I get irritated by statisticians who can't do it--especially those who seem to pump up bad ideas or disparage good ideas--I tend to think of them as lesser on a personal level. Some of this is legitimate, I think--part of being a good person is to recognize how one can be most helpful to others--but I probably lean too far in this direction. Even people who are nearly universally disliked, if they're good statisticians, I'll give them the benefit of the doubt. But if I don't like their ideas, it's hard to avoid disliking them.

For a sillier example, I remember reading in Susan Cheever's memoir that John Cheever rated people based on how strong were the drinks they served. Higher alcohol content = better person. And in playing pickup frisbee, I think that, on average, you'll be more liked as a person if you're a better frisbee player. (Although maybe in basketball it goes the other way...)

P.S. What happened to Kazin's first wife? After they broke up, she didn't want to get back together with him--a reasonable enough decision, especially considering how his life proceeded in the years after--but then I was mildly curious what happened with her after that, and the biography didn't say.

Nick Firoozye writes:

Mike McLaughlin writes:

I was wondering about MCMC burn-in and whether the oft-cited emphasis on this in the literature might not be a bit overstated.

My thought was that the chain is Markovian. In a Metropolis (or Metropolis-Hastings) context, once you establish the scale of the proposal distribution(s), successful burn-in gets you only a starting location inside the posterior -- nothing else is remembered, by definition! However, there is nothing really special about this particular starting point; it would have been just as valid had it been your initial guess and the burn-in would then have been superfluous. Moreover, the sampling phase will eventually reach the far outskirts of the posterior, often a lot more extreme than the sampling starting location, yet it will still (collectively) describe the posterior correctly. This implies that *any* valid starting point is just as good as any other, burn-in or no burn-in.

The only circumstance that I can think of in which a burn-in would be essential is in the case in which prior support regions for the parameters are not all jointly valid (inside the joint posterior), if that is even possible given the min/max limits set for the priors. Am I missing something?

My response: What you're missing is that any inference from a finite number of simulations is an approximation.

Consider an extreme example in which your sample takes independent draws from a N(mu,sigma^2) distribution, but you pick a starting value of X. The average of n simulations will then have the value, in expectation, of (1/n)X+ ((n-1)/n)mu (instead of the correct value of mu). If, for example, X=100 and n=100, you're in trouble! But a burn-in of 1 will solve all your problems in this example. (And in this example, n=100 would work just fine for most purposes.) True, if you draw a few gazillion simulations, the initial values will be forgotten, but why run a few zillion simulations if you don't have to? That will just slow down your important work.

More generally, your starting values will persist for awhile, basically as long as it takes for your chains to mix. If your starting values persist for a time T, then these will pollute your inferences for some time of order T, by which time you can already have stopped the simulations if you'd discarded some early steps.

P.S. See here for a different perspective, from Charlie Geyer. For the reasons stated above, I don't agree with what he writes, but you can read for yourself.

P.P.S. In my example above, you might say that it would be ok if you were to just start at the center of the distribution. One difficulty, though, is that you don't know where the center of the distribution is before you've done your simulations. More realistically, we start from estimates +/- uncertainty as estimated from some simpler model that was easier to fit.

Boris and I will be speaking at the Cato Institute in Washington, D.C., next Thurs (11 Sept) at noon on our Red State, Blue State book (also written with David Park, Joe Bafumi, and Jeronimo Cortina). The event will be moderated by Will Wilkinson; see the description here of the event on his blog.

All are welcome to come, but you should register online for the event. We'll be having a panel discussion with Michael McDonald (Brookings Institution and George Mason University) and Brink Lindsey of the Cato Institute. I'm curious what they have to say about our work, especially some of the stuff at the end of chapter 9 about the connections between public opinion and policy.

Jim points me to this article by Don Berry, which argues that studies of doping in sports often don't correctly perform probability calculations.

Tyler Cowen discusses the possibility of economics prodigies. I refer him and his commenters to Dick De Veaux's saying, "Math is like music, statistics is like literature." You can decide yourself where economics is or should stand in this spectrum. I will say, though, that it can take decades to develop a good idea, just because you can be busy doing other things.

Mohammed Mohammed points me to this article by John Nichols, which begins:

Observational studies comparing groups or populations to evaluate services or interventions usually require case-mix adjustment to account for imbalances between the groups being compared. Simulation studies have, however, shown that case-mix adjustment can make any bias worse. One reason this can happen is if the risk factors used in the adjustment are related to the risk in different ways in the groups or populations being compared, and ignoring this commits the ‘‘constant risk fallacy’’. Case-mix adjustment is particularly prone to this problem when the adjustment uses factors that are proxies for the real risk factors. Interactions between risk factors and groups should always be examined before case-mix adjustment in observational studies.

This is interesting, and it connects to my struggles with survey weighting. Survey weighting is similar to adjustment of control and treatment groups in an observational study. (The survey analogy is respondents and nonrespondents.) Nichols's article points out difficulties with adjustment if you ignore interactions, which is a problem we've found in survey adjustment as well. The solution is to include all interactions that are potentially important, but then a model becomes large, and we have to go beyond least squares and exchangeable models . . .

We discuss in chapters 9 and 10 of ARM the general problem of adjusting for differences between treatment and control groups, but we don't specifically focus on the importance of interactions.

Jamie pointed me to this graph in the NYT:

Nice! Especially the margin of error, the subtle colorings and the use of the gray background. The y-axis labels are a little weird (why not simply 60, 80, 100, and 120), and I'm not sure how to think about the x-axis. (Given the scale on the y-axis, should we really care about a change of 1/2 mile per hour in wind speed.) Also, I don't really understand what it means to measure changes in wind speed, if the storms are themselves categorized by wind speed! But, as a graph, it has many excellent features.

Marcus Brubaker writes:

I am currently working on a problem in computational biology using Bayesian inference and I've come to a question for which I hope you have an answer. In this problem there are a large number of noisy 2D images of a molecule, from which we wish to infer the 3D structure. Much of the modeling is straightforward but I have hit a roadblock. Specifically, the noise parameters for these images.

Following up on my Lauryn Hill question (and thanks much, Scott Cunningham, for the detailed answer there): what's the deal with Matthew Klam. The stories in "Sam the Cat" (2001) are hilarious, and they also give the impression that Klam could just whip them out. Since then, though . . . almost nothing. Klam has a bizarrely professional-looking webpage but not the output I'd expect. What's the deal? Maybe author/blogger J. Robert Lennon knows?

See here for discussion.

David Remnick, writing in The New Yorker about the Democratic convention:

Michelle Obama tore up the wing-nut caricatures of herself as a closet radical by revealing, without exploiting, the irresistible charms of her children and delivering a warm, genuine, and impassioned introduction to her husband.

Huh? Where do I start on this?
- Can't a "closet radical" have irresistibly charming children?
- Can't a "closet radical" deliver a good speech?
And, the biggest thing of all: if you're really a "closet" radical, then of course you'll try to act like a normal person when you're on national TV.

I mean, sure, you can say she gave a good speech or that you agree with what she had to say or even that she seems likable (although that seems to be stretching it; after all, it's just a prepared speech). But a public speech has gotta be the last place to look if you're trying to evaluate whether someone has a hidden agenda!

P.S. Just to be clear: I'm not saying anything at all about Michelle Obama here. I'm just stunned at the gap in logical reasoning here. Isn't The New Yorker famous for its fact-checkers???

That header got your attention, huh?? John Hull writes:

Reading an article on "non-Aristotelean" logic, where P(A) is my confidence of A being true, I found (on page 10) the equation P(B=>C)=P(B[AND]C)/P(B). Since I work in municipal government, an obvious interpretation of this is the following:

Josh Tucker sent me this paper by Alexander Herzog and himself on attitudes toward the European Union in different European countries. Here's the abstract:

In this paper, we [Herzog and Tucker] document a hitherto unrecognized “micro-macro paradox” of EU accession in post-communist countries: on the micro-level, economic prosperity increases the likelihood of supporting EU membership; while on the macro-level, economic prosperity decreases aggregate levels of support for EU membership.

Dan Lakeland writes:

I recently enrolled as a PhD student in a civil engineering program. My interest could be described as the application of data and risk analysis to engineering modelling, design methods, and decision making.

The field is pretty ripe, and infrastructure risk analysis is a common topic these days, but the simulations and statistical approaches taken so far have been a bit unsatisfactory. For example people studying the impact of bridge failures during earthquakes on the local economy might assume a constant cost per person-hour of delay throughout the rebuild period, or people might build statistical models of probability of building collapse, but I would call them pretty much prior distributions, not really based on much data, or based on a finite element computer model of the physics of a single model building.

John Skilling wrote this response to my discussion and rejoinder on objections to Bayesian statistics.

John's claim is that Bayesian inference is not only a good idea but also is necessary. He justifies Bayesian inference using the logical reasoning of Richard Cox (1946), which I believe is equivalent to the von Neumann and Morgenstern (1948) derivation that those of us in the social sciences are familiar with.

I have no objection to the Cox/Neumann/Morgenstern argument (which is also associated with Keynes and others). However, all our models have flaws, so the big step that all of us have to make when recommending Bayesian methods is to believe that they work well even with the imperfect models that we use in practice. Models that in social science are surely far more imperfect than those used in physics (which may be one reason that physicist Skilling is so accepting of Bayes). Skilling cites Jaynes, whose writings have been a major influence in my own conception of Bayesian statistics. (The first scientific conference I ever attended, I think, was Skilling's Jaynes-inspired maximum entropy conference in 1988.) In particular, I got from Jaynes the idea that one should take any model, even a simple model, seriously enough to try to understand where it doesn't fit reality and how it should be improved.

In conclusion: I think Skilling's normative argument is powerful, but not definitive on its own, because we want statistical methods to work well even when models have serious flaws. That's why I take a more pluralistic approach. (But, as I explain in my rejoinder (linked above), I am not at all convinced by arguments about classical unbiasedness or coverage.)

To put it another way, the classical ideas of sufficiency, ancillarity, confidence coverage, hypothesis testing, etc etc: I'm happy to trash all these. But there is another set of principles out there, based on external validation (sometimes approximated by cross-validation), that seems valid to me, and does not necessarily rely on Bayes.

P.S. See also Larry Wasserman's further comments.

In Red State, Blue State, I attributed the distorted maps to "computer scientists Michael Gastner, Cosma Shalizi, and Mark Newman." Cosma writes that "none of us are or were 'computer scientists', and in fact we were all trained as physicists, and working as physicists at the time." My bad.

Ernest Sergenti writes with a question regarding how to determine or quantify how many simulation iterations are "enough" to calculate quantities of interest, after burn-in has been achieved:

A few months ago I noticed on my friend Seth's website that he was an "emeritus professor." I called him up, first thinking it was a mistake--he's well under sixty years old--but, no, he really is retired. He taught at Berkeley for 30 years. We had the following exchange on the phone:

Me: Why retire? As a professor, they pay you even if you don't work.

Seth: They pay you for not working if you're retired, too.

He's got a point.

Gabriel Katz writes:

Josh Tenenbaum sent me a link to a paper, The discovery of structural form, by C. Kemp and himself. Also commentary by Keith Holyoak and some supporting information. Code and datasets are here.

For my own thoughts on this work, see here. Josh's talk at Columbia made me realize that all these years I'd been thinking of life as part of a "great chain of being" without realizing it.

Here's the announcement: