1111 stories

What readings should be included in a seminar on the philosophy of statistics, the replication crisis, causation, etc.?

1 Share

André Ariew writes:

I’m a philosopher of science at the University of Missouri. I’m interested in leading a seminar on a variety of current topics with philosophical value, including problems with significance tests, the replication crisis, causation, correlation, randomized trials, etc. I’m hoping that you can point me in a good direction for accessible readings for the syllabus. Can you? While the course is at the graduate level, I don’t assume that my students are expert in the philosophy of science and likely don’t know what a p-value is (that’s the trouble—need to get people to understand these things). When I teach a course on inductive reasoning I typically assign Ian Hacking’s An Introduction to Probability and Inductive Logic. I’m familiar with the book and he’s a great historian and philosopher of science.

He’d like to do more:

Anything you might suggest would be greatly appreciated. I’ve always thought that issues like these are much more important to the philosophy of science than much of what passes as the standard corpus.

My response:

I’d start with the classic and very readable 2011 article by Simmons, Nelson, and Simonsohn, False-Positive Psychology: Undisclosed Flexibility in Data Collection and Analysis Allows Presenting Anything as Significant.

And follow up with my (subjective) historical overview from 2016, What has happened down here is the winds have changed.

You’ll want to assign at least one paper by Paul Meehl; here’s a link to a 1985 paper, and here’s a pointer to a paper from 1967, along with the question, “What happened? Why did it take us nearly 50 years to what Meehl was saying all along? This is what I want the intellectual history to help me understand,” and 137 comments in the discussion thread.

And I’ll also recommend my own three articles on the philosophy of statistics:

The last of these is the shortest so it might be a good place to start—or the only one, since it would be overkill to ask people to read all three.

Regarding p-values etc., the following article could be helpful (sorry, it’s another one of mine!):

And, for causation, I recommend these two articles, both of which should be readable for students without technical backgrounds:

OK, that’ll get you started. Perhaps the commenters have further suggestions?

P.S. I’d love to lead a seminar on the philosophy of statistics, unfortunately I suspect that here at Columbia this would attract approximately 0 students. I do cover some of these issues in my class on Communicating Data and Statistics, though.

The post What readings should be included in a seminar on the philosophy of statistics, the replication crisis, causation, etc.? appeared first on Statistical Modeling, Causal Inference, and Social Science.

Read the whole story
10 days ago
Share this story

The Difference Between Amateurs and Professionals

1 Share

Why is it that some people seem to be hugely successful and do so much, while the vast majority of us struggle to tread water?

The answer is complicated and likely multifaceted.

One aspect is mindset—specifically, the difference between amateurs and professionals.

Most of us are just amateurs.

What’s the difference? Actually, there are many differences:

  • Amateurs stop when they achieve something. Professionals understand that the initial achievement is just the beginning.
  • Amateurs have a goal. Professionals have a process.
  • Amateurs think they are good at everything. Professionals understand their circles of competence.
  • Amateurs see feedback and coaching as someone criticizing them as a person. Professionals know they have weak spots and seek out thoughtful criticism.
  • Amateurs value isolated performance. Think about the receiver who catches the ball once on a difficult throw. Professionals value consistency. Can I catch the ball in the same situation 9 times out of 10?
  • Amateurs give up at the first sign of trouble and assume they’re failures. Professionals see failure as part of the path to growth and mastery.
  • Amateurs don’t have any idea what improves the odds of achieving good outcomes. Professionals do.
  • Amateurs show up to practice to have fun. Professionals realize that what happens in practice happens in games.
  • Amateurs focus on identifying their weaknesses and improving them. Professionals focus on their strengths and on finding people who are strong where they are weak.
  • Amateurs think knowledge is power. Professionals pass on wisdom and advice.
  • Amateurs focus on being right. Professionals focus on getting the best outcome.
  • Amateurs focus on first-level thinking. Professionals focus on second-level thinking.
  • Amateurs think good outcomes are the result of their brilliance. Professionals understand when outcomes are the result of luck.
  • Amateurs focus on the short term. Professionals focus on the long term.
  • Amateurs focus on tearing other people down. Professionals focus on making everyone better.
  • Amateurs make decisions in committees so there is no one person responsible if things go wrong. Professionals make decisions as individuals and accept responsibility.
  • Amateurs blame others. Professionals accept responsibility.
  • Amateurs show up inconsistently. Professionals show up every day.

There are a host of other differences, but they can effectively be boiled down to two things: fear and reality.

Amateurs believe that the world should work the way they want it to. Professionals realize that they have to work with the world as they find it. Amateurs are scared — scared to be vulnerable and honest with themselves. Professionals feel like they are capable of handling almost anything.

Luck aside, which approach do you think is going to yield better results?

Food for Thought:

  • In what circumstances do you find yourself behaving like an amateur instead of as a professional?
  • What’s holding you back? Are you hanging around people who are amateurs when you should be hanging around professionals?

Sponsored by: Royce & Associates – Small Cap Specialists with Unparalleled Knowledge and Experience..

Read the whole story
10 days ago
Share this story

The Lost Cause Rides Again

2 Comments and 11 Shares

HBO’s prospective series Confederate will offer an alternative history of post-Civil War America. It will ask the question, according to co-creator David Benioff,  “What would the world have looked like … if the South had won?” A swirl of virtual protests and op-eds have greeted this proposed premise. In response, HBO has expressed “great respect” for its critics but also said it hopes that they will “reserve judgment until there is something to see.”

This request sounds sensible at first pass. Should one not “reserve judgment” of a thing until after it has been seen? But HBO does not actually want the public to reserve judgment so much as it wants the public to make a positive judgment. A major entertainment company does not announce a big new show in hopes of garnering dispassionate nods of acknowledgement. HBO executives themselves judged Confederate before they’d seen it—they had to, as no television script actually exists. HBO hoped to communicate that approval to its audience through the announcement. And had that communication been successful, had Confederate been greeted with rapturous anticipation, it is hard to imagine the network asking its audience to tamp down and wait.

HBO’s motives aside, the plea to wait supposes that a problem of conception can be fixed in execution. We do not need to wait to observe that this supposition is, at best, dicey. For over a century, Hollywood has churned out well-executed, slickly produced epics which advanced the Lost Cause myth of the Civil War. These are true “alternative histories,” built on “alternative facts,” assembled to depict the Confederacy as a wonderland of virtuous damsels and gallant knights, instead of the sprawling kleptocratic police state it actually was. From last century’s The Birth of a Nation to this century’s Gods and Generals, Hollywood has likely done more than any other American institution to obstruct a truthful apprehension of the Civil War, and thus modern America’s very origins. So one need not wait to observe that any foray by HBO into the Civil War must be met with a spirit of pointed inquiry and a withholding of all benefit of the doubt.

Skepticism must be the order of the day. So that when Benioff asks “what would the world have looked like … if the South had won,” we should not hesitate to ask what Benioff means by “the South.” He obviously does not mean the minority of  white Southern unionists, who did win. And he does not mean those four million enslaved blacks, whom the Civil War ultimately emancipated, yet whose victory was tainted. Comprising 40 percent of the Confederacy’s population, this was the South’s indispensable laboring class, its chief resource, its chief source of wealth, and the sole reason why a Confederacy existed in the first place. But they are not the subject of Benioff’s inquiry, because he is not so much asking about “the South” winning, so much as he is asking about “the white South” winning.

The distinction matters. For while the Confederacy, as a political entity, was certainly defeated, and chattel slavery outlawed, the racist hierarchy which Lee and Davis sought to erect, lives on. It had to. The terms of the white South’s defeat were gentle. Having inaugurated a war which killed more Americans than all other American wars combined, the Confederacy’s leaders were back in the country’s political leadership within a decade. Within two, they had effectively retaken control of the South.

Knowing this, we do not have to wait to point out that comparisons between Confederate and The Man in the High Castle are fatuous. Nazi Germany was also defeated. But while its surviving leadership was put on trial before the world, not one author of the Confederacy was convicted of treason. Nazi Foreign Minister Joachim von Ribbentrop was hanged at Nuremberg. Confederate General John B. Gordon became a senator. Germany has spent the decades since World War II in national penance for Nazi crimes. America spent the decades after the Civil War transforming Confederate crimes into virtues. It is illegal to fly the Nazi flag in Germany. The Confederate flag is enmeshed in the state flag of Mississippi.

The symbols point to something Confederate’s creators don’t seem to understand—the war is over for them, not for us. At this very hour, black people all across the South are still fighting the battle which they joined during Reconstruction—securing equal access to the ballot—and resisting a president whose resemblance to Andrew Johnson is uncanny. Confederate is the kind of provocative thought experiment that can be engaged in when someone else’s lived reality really is fantasy to you, when your grandmother is not in danger of losing her vote, when the terrorist attack on Charleston evokes honest sympathy, but inspires no direct fear. And so we need not wait to note that Confederate’s interest in Civil War history is biased, that it is premised on a simplistic view of white Southern defeat, instead of the more complicated morass we have all around us.

And one need not wait to ask if Benioff and D.B. Weiss are, at any rate, the candidates to help lead us out of that morass or deepen it. A body of work exists in the form of their hit show Game of Thrones. We do not have to wait to note the persistent criticism of that show is its depiction of rape. Rape—generational rape, mass rape—is central to the story of enslavement. For 250 years the bodies of enslaved black women were regarded as property, to be put to whatever use—carnal and otherwise—that their enslavers saw fit. Why HBO believes that this duo, given their past work, is the best team to revisit that experience is a question one should not wait to ask.

And all this must be added to a basic artistic critique—Confederate is a shockingly unoriginal idea, especially for the allegedly avant garde HBO. “What if the white South had won?” may well be the most trod-upon terrain in the field of American alternative history. There are novels about it, comic books about it, games about it, and a mockumentary about it. It’s been barely a year since Ben Winters published Underground Airlines.

Storytellers have the right to answer any question they choose. But we do not need to wait to examine all the questions that are not being chosen: What if John Brown had succeeded? What if the Haitian Revolution had spread to the rest of the Americas? What if black soldiers had been enlisted at the onset of the Civil War? What if Native Americans had halted the advance of whites at the Mississippi? And we need not wait to note that more interesting than asking what the world would be like if the white South had won is asking why so many white people are enthralled with a world where the dreams of Harriet Tubman were destroyed by the ambitions of Robert E. Lee.

The problem of Confederate can’t be redeemed by production values, crisp writing, or even complicated characters. That is not because its conceivers are personally racist, or seek to create a show that endorses slavery. Far from it, I suspect. Indeed, the creators have said that their hope is to use science fiction to “show us how this history is still with us in a way no strictly realistic drama ever could.” And that really is the problem. African Americans do not need science-fiction, or really any fiction, to tell them that that “history is still with us.” It’s right outside our door. It’s in our politics. It’s on our networks. And Confederate is not immune. The show’s very operating premise, the fact that it roots itself in a long white tradition of imagining away emancipation, leaves one wondering how “lost” the Lost Cause really was.

It’s good that the show-runners have brought on two noted and talented black writers—Nichelle Tramble Spellman and Malcolm Spellman. But one wonders: If black writers, in general, were to have HBO’s resources and support to create an alternative world, would they choose the world dreamed up by the progenitors of the Ku Klux Klan? Or would they address themselves to other less trod areas of Civil War history in the desire to say something new, in the desire to not, yet again, produce a richly imagined and visually beguiling lie?

We have been living with the lie for so long. And we cannot fix the lie by asking “What if the white South won?” and waiting for an answer, because the lie is not in the answer, but in the question itself.

Read the whole story
10 days ago
10 days ago
Share this story
2 public comments
10 days ago
Ta-Nehisi Coates is a national treasure. Also that Andrew Johnson = Trump link is pure gold.
Somerville, MA
11 days ago
Coates brings up a point that I have never seen quite so well articulated before. The South did win the Civil War. When the War began, 40% of the South was enslaved. When the war ended 0% of the South was enslaved. That's a win by any decent ethical standards.
Washington, District of Columbia

Saturday Morning Breakfast Cereal - Drugs


Click here to go see the bonus panel!

Me: I wonder if there'll ever be bars that just serve uppers. Also me: GOD I'd kill for a triple espresso right now.

New comic!
Today's News:
Read the whole story
11 days ago
Share this story

Seemingly intuitive and low math intros to Bayes never seem to deliver as hoped: Why?

1 Share

This post was prompted by recent nicely done videos by Rasmus Baath that provide an intuitive and low math introduction to Bayesian material. Now, I do not know that these have delivered less than he hoped for. Nor I have asked him. However, given similar material I and others have tried out in the past that did not deliver what was hoped for, I am anticipating that and speculating why here. I have real doubts about such material actually enabling others to meaningfully interpret Bayesian analyses let alone implement them themselves. For instance, in a conversation last year with David Spiegelhalter, his take was that some material I had could easily be followed by many, but the concepts that material was trying to get across were very subtle and few would have the background to connect to them. On the other hand, maybe I am hoping to be convinced otherwise here.

For those too impatient to watch the three roughly 30 minute videos, I will quickly describe my material David commented on (which I think is fairly similar to Rasmus’). I am more familiar with it and doing that avoids any risk of misinterpreting anything Rasmus did. It is based on a speculative description of what what Francis Galton did in 1885 which was discussed more thoroughly by Stephen Stigler. It also involves a continuous (like) example which I highly prefer starting with. I think continuity is of overriding importance so one should start with it unless they absolutely can not.

Galton constructed a two stage quincunx (diagram for a 1990 patent application!) with the first stage representing his understanding of the variation of the targeted plant height in a randomly chosen plant seed of a given variety. The pellet haphazardly falls through the pins and lands at the bottom of the first level as the target height of the seed. His understanding I think is a better choice of wording than belief, information or even probability (which it can be taken to be given the haphazardness). Also it is much much better than prior! Continuing on from the first level, the pellet falls down a second set of pins landing at the very bottom as the height the plant actually grew to. This second level represents Galton’s understanding of how a seed of a given targeted height varies in the height it actually grows. Admittedly this physical representation is actually discrete but the level of discreteness can be lessened without preset limits (other than practicality).

Possibly, he would have assessed the ability of his machine to adequately represent his understanding, by running it over and over again and and comparing the set of heights plants represented on the bottom level with knowledge of past, if not current heights this variety of seed usually did grow to. He should have. Another way to put Galton’s work would be that of building (and checking) a two stage simulation to adequately emulate one’s understanding of targeted plant heights and actual plant heights that have been observed to grow. Having assessed his machine as adequate (by surviving a fake data simulation check) he might have then thought about how to learn about a particular given seeds targeted height (possibly already growing or grown) given he would only get to see the actual height grown. The targeted height remains unknown and actual height becomes know. It is clear that Galton decided that in trying to assess the targeted height from an actual height one should not look downward from a given targeted height but rather upward from the actual height grown.

Now by doing multiple drops of pellets, one at a time, and recording only where the seed was at the bottom of the first level if and only if it lands at a particular location on the bottom level matching an actual grown height, he would doing a two two stage simulation with rejection. This clearly provides an exact (smallish) sample from the posterior given the exact joint probability model (physically) specified/simulated by the quincunx. It is exactly the same as the conceptual way to understand Bayes suggested by Don Rubin in 1982. As such, it would have been an early fully Bayesian analysis, even if not actually perceived as such at the time (though Stigler argues that it likely was).

This awkward to carry out, but arguably less challenging way to grasp Bayesian analysis can be worked up to address numerous concepts in statistics (both implementing calculations and motivating formulas ) that are again less challenging to grasp (or so its hoped). This is what I perceive, Rasmus, Richard McElreath and and others are essentially doing. Authors do differ in their choices of which concepts to focus on. My initial recognition of these possibilities lead to this overly exuberant but very poorly thought through post back in 2010 (some links are broken).

To more fully discuss this below (which may be of interest only to those very interested), I will extend the quincuz to multiple samples (n > 1) and multiple parameters, clarify the connection to approximate Bayesian computation (ABC) and point out something much more sensible when there is a formula for the second level of the quincunz (the evil likelihood function) . The likelihood might provide a smoother transition to (MCMC) sampling from the typical set instead of the entirety of parameter space. I will also say some nice things about Rasmus’s videos and of course make a few criticisms.

Here I am switching back to standard statistical terminology, even though I agree with Richard McElreath that these should not be used in introductory material.

As for n > 1, had Galton conceived  of cloning the germinated seeds to allow multiple actual heights with the same targeted height, he could have emulated sample sizes of n > 1 in a direct if not very awkward way using multiple quincunzes. In the first quincunz, the first level would represent the prior and the subset of the pellets that ended up on the bottom of the second level matching the height of the first plant actually grown, would represent the posterior given the first plant height. The pellets representing the posterior in the first quincunx (the subset at the bottom of the first level) would then need to be transferred to the bottom of the first level of second machine (as the new prior). They then would be let fall down to the bottom of its second level to represent the posterior given both the first and second plant height.  And so on for each and every sample height grown from the cloned seed.

As for multiple parameters, Rasmus moved on to two parameter in his videos by simply by running two quincunzes in parallel. At some point the benefit/cost of this physical analogue (or metaphor) quickly approaches zero and should be discarded. Perhaps at this point just move on to two stage rejection sampling with multiple parameters and sample sizes > 1.

The history of approximate Bayesian computation is interesting and perhaps especially so for me as I  thought I had invented it for a class of graduate Epidemiology students in 2005. I needed a way of convincing them Bayes was not a synonym for MCMC and thought of using two stage rejection sampling to do this. Though, in a later discussion with Don Rubin it is likely I got it from his paper linked above and just forgot about that. But two stage rejection sampling is and is not ABC.

The motivation for ABC was from not having a tractable likelihood but still wanting to do a Bayesian analysis. It was the same motivation for my DPhil thesis which was not having a tractable likelihood for many published summaries (with no access to individual values) but still wanting to do a likelihood based confidence interval analysis (I was not yet favouring a Bayesian approach). In fact, the group that is generally credited with first doing ABC (with that Bayes motivation) included my internal thesis examiner RC Griffiths (Oxford).  Now, I first heard about ABC in David Cox’s recorded JSM talk in Florida. Afterwards, whenever I exchanged emails with some in Griffiths’ group and others doing ABC, there arose a lot of initial confusion.

That was because in my thesis work, I did have the likelihood in closed form for individual observations but only had summaries which usually did not have a tractable likelihood. The ABC group did not have a tractable likelihood for individual observations ever (or it was to expensive to compute). Because of this, when I used ABC to get posteriors from summarised data, because that was all that I had observed, it would be actually approximating the exact posterior (given one had only observed the summaries). So to some of them I was not actually doing ABC but some weird other thing. (I am not aware if anyone has published such an ABC like analysis, for instance a meta-analysis of published summaries).

So now into some of the technicalities of real ABC and not quite real ABC. Lets take a very simply example of one continuous observation that was recorded only to two decimation places and one unknown parameter. In general, with any prior and data generating model, two stage rejection sampling matched to two decimation places provides a sample from the exact posterior. So not ABC just full Bayes done inefficiently. On the other hand, if it was recorded to 5 or 10 or more decimal places, matching all the decimal places may not be feasible and choosing to match to just two decimal places would be real ABC. Now think of 20, 30 or more samples recorded to two decimal places. Matching all, even to 2 decimal places is not feasible but deciding to match the sample mean to all decimal places of the sample mean recorded will be feasible and is ABC having used just the summary. Well, unless one assumes the data generating model is Normal as then by sufficiency its not ABC – its just full Bayes. These distinctions are somewhat annoying – but the degree of approximation does need to be recognised and admittedly ABC is the wrong label when there is no approximate rather than exact posterior.

Now some criticism of Rasmus’s videos. I really did not like the part where the number of positive responses to a mail out of 16 offers was analysed using a uniform prior – primarily motivated as non-informative – and the resulting (formal) posterior probabilities discussed as if they were relevant or useful. This is not the kind of certainty about uncertainty to be obtained through statistical alchemy that we want to be inadvertently instilling in people. Now, informative priors were later pitched as being better by Rasmus, but I think the damage has already been done.

Unfortunately, the issue is not that well addressed in the statistical literature and someone even once wrote that most Bayesians would be very clear about the posterior not being the posterior but rather dependent on the particular prior. At least in any published Bayesian analysis. Now, if I was interested in how someone in particular or some group in particular would react, their prior and hence their implied posterior probabilities given the data they have observed would be relevant, useful and could even be taken literally. But if I was interested in how the world would react, the prior would need to be credibly related to the world for me to take posterior probabilities as relevant and in any remote sense literal.

Now, if calibrated, posterior probabilities could provide useful uncertainty intervals. That’s a different topic. For an instance of priors being unconnected to the world, Andrew’s multiple comp post provided an example of a uniform prior on the line that is horribly not credibly related to the effects sizes in the research area one is working in.  Additionally, the studies being way too week to in any sense get over that horribly not relatedness. In introductory material, just don’t use flat priors at all. Do mention them but point out that they can be very dangerous in general (i.e. consult an experienced expert before using) but don’t use them in introductory material.

I really did like the side by side plots of the sample space and parameter space for the linear regression example. The sample space plot showing the fitted line (and the individual x and y values) and the parameter space plot initially having a dot at the intercept and slope values that give the maximum probability of observing the individual x and y values actually observed. Later dots where added and scaled to show intercept and slope values that give less probability and then posterior probabilities where printed over these.  Now, I do think it would be better if there was something in the parameter space plot that roughly represented the individual x and y values observed.

Here one could take the value of the intercept (that jointly with the slope gave the maximum probability) as fixed or known and then with just one unknown left, find the maximum slope using each individual x and y value and plot those. Then do the same for the intercept by taking the slope value (that gave the maximum probability) as known. The complication here comes from the intercept and slope parameters being tangled up together. Much more can be done here, but admittedly I have able to convince very few that this sort of thing would be worth the trouble. (Well one journal editor, but they found that the technical innovations involved were not worthy of getting a published paper in their journal). What about the standard deviation parameter? It was taken as know from the start? Actually that does not matter as much as that parameter is much less tangled up with the intercept and slope parameters.

When one does have a closed form for the likelihood (and it is not numerically intensive), two stage rejection sampling is sort of silly. If you think about two stage rejection sampling, in the first stage you get a sample of the proportions certain values have in the prior (i.e. estimates of the prior probabilities of those values). In the second stage you keep the proportion those certain values that generated simulated values that matched (or closely approximated) observed values. The proportions kept in the second stage are estimates of probabilities of generating the observed values given the parameter values generated in the first stage. Hence they are estimates of the likelihood – P(X|parameter) – but you have that in closed form. So take the parameter values generated in the first stage and simply weight them by the likelihood (i.e. importance sampling) to get a better sample from the posterior much more efficiently. Doing this, one can easily implement more realistic examples such as multiple regression with half a dozen or more covariates. Some are tempted to avoid any estimation at all by using a systematic approximation of the joint distribution on grids of points leading to  P(discretized(X),discretized(parameter)) and P(parameter) * P(X|parameter) for some level of discreteness. I think this is a mistake as it severely breaks continuity, does not scale to realistic examples and  suggests a return to unthinkingly plugging and chugging mindlessly through weird formulas that one just needs to get used to.

These more realistic higher dimension examples may help bridge people to the need for  sampling from the typical set instead of the entirety of parameter space. I did work it through for bridging to sequential importance sampling by _walking_ to prior to the posterior in smaller, safer steps. But bridging to the typical set likely would be better.

In closing this long post, I feel I should acknowledge the quality of Rasmus videos which he is sharing with everyone. I am sure it took a lot of time and work. It took me more time than I want to admit to put this post together, perhaps since its been a few years since I actually worked on such material. Seeing other make some progress, prompted me to try again by at least thinking about the challenges.



The post Seemingly intuitive and low math intros to Bayes never seem to deliver as hoped: Why? appeared first on Statistical Modeling, Causal Inference, and Social Science.

Read the whole story
11 days ago
Share this story

`Profit Was a By-Product of the Pleasure'

1 Share
“President [Theodore] Roosevelt was an omnivorous reader. The legend holds that he read a book a day, and while that might be an exaggeration, it seems unlikely that any president in the century since he held the office read as much as he, or in such a wide variety of fields, including poetry.”

This comes not from a biography of the twenty-sixth president of the United States but from Scott Donaldson’s Edwin Arlington Robinson: A Poet’s Life (2007). Donaldson is retelling the familiar story of how, in 1904, Kermit Roosevelt, the president’s son, brought Robinson’s second poetry collection, The Children of the Night (1897), to his father’s attention. TR persuaded Charles Scribner’s Sons to republish the volume, and reviewed it himself in Outlook magazine. Roosevelt got Robinson’s name wrong ("Edwin," not “Edward”), but he rightly detected “an undoubted touch of genius” in the poems. Roosevelt arranged for Robinson to receive a sinecure at the New York Customs House, with a $2,000 annual stipend. In 1910, Robinson repaid the debt by dedicating his next collection of poems, The Town Down the River, to the former president.

Roosevelt wasn’t being strictly altruistic. He was the only American president who, for significant periods, lived as a professional writer, earning much of his living with his pen. He understood the teetering balancing act the writing life might pose for a poet like Robinson. But Roosevelt also had good taste in literature (though he did favor the almost unreadable Jack London), and he had the interests of his country at heart. We might think of him as a literary patriot with an open mind (he loved Gibbon). The president writes about Robinson in a 1905 letter to James Hulme Canfield (Theodore Roosevelt: Letters and Speeches, Library of America, 2004): “. . . –I hunted him up, found he was having a very hard time, and put him in the Treasury Department. I think he will do his work all right, but I am free to say that he was put in less with a view to the good of the government service than with a view to helping American letters.”

In 1903, Roosevelt made his literary tastes explicit in a letter to the president of Columbia University, Nicholas Murray Butler. He fills three pages with a list of the books he has read over the previous two years, including Herodotus, Plutarch, Dante, Shakespeare, Keats, Browning and Carlyle. At the end of his list, Roosevelt writes:

“There! that is the catalogue; about as interesting as Homer’s Catalogue of the Ships, and with about as much method in it as there seems in a superficial glance to be in an Irish stew. The great comfort, old man, is that you need not read it and that you need not answer this!”

Dedicated, unpretentious, pleasure-driven reading remains a theme across Roosevelt’s life, most memorably articulated in Theodore Roosevelt, an Autobiography (1913):  

“Books are almost as individual as friends. There is no earthly use in laying down general laws about them. Some meet the needs of one person, and some of another; and each person should beware of the booklover's besetting sin, of what Mr. Edgar Allan Poe calls `the mad pride of intellectuality,’ taking the shape of arrogant pity for the man who does not like the same kind of books.”

That characterizes nine-tenths of book chat in general and an even greater proportion of the blogosphere’s bookish precincts. Later in the same paragraph Roosevelt writes: “Personally, the books by which I have profited infinitely more than by any others have been those in which profit was a by-product of the pleasure; that is, I read them because I enjoyed them, because I liked reading them, and the profit came in as part of the enjoyment.”

Here he echoes Dr. Johnson, who is quoted by Boswell as saying: “. . . what we read with inclination makes a much stronger impression. If we read without inclination, half the mind is employed in fixing the attention; so there is but one half to be employed on what we read.” And Roosevelt, bless him, found Dickens a confounding irritant. In a letter to his son Kermit in 1908 he writes:

“. . . he had himself a thick streak of maudlin sentimentality of the kind that, as somebody phrased it, `made him wallow naked in the pathetic.’ It always interests me about Dickens to think how much first-class work he did and how almost all of it was mixed up with every kind of cheap, second-rate matter. I am very fond of him. There are innumerable characters that he has created which symbolize vices, virtues, follies, and the like almost as well as the characters in Bunyan; and therefore I think the wise thing to do is simply to skip the bosh and twaddle and vulgarity and untruth, and get the benefit out of the rest.”

[Hereis a marvelous portrait of Roosevelt the reader and dog lover, taken in Colorado in 1905.]
Read the whole story
14 days ago
Share this story
Next Page of Stories