I came across a bunch of online essays—“posts”—by Matthew Salesses, a professor of writing at Columbia:
Also 22 Revision Prompts, which are so great that I’ll devote a separate post to them.
As a writer and a teacher of creative work (yes, statistical analysis is creative work!), I’m very interested in the above topics, and Salesses has a lot of interesting things to say.
I should warn you that he has a strong political take, and the political perspective is central to his thinking—but I think his advice should be valuable even to readers who don’t share his views on cultural politics. I’d draw the analogy to Tom Wolfe, whose cultural conservatism informs his views on art, views that should be of interest even to people to disagree with him on politics. It’s possible to be a big fan of Tom Wolfe while at the same time thinking it was pitiful for him to take his cultural politics so far as to deny evolution. Anyway, you can think what you want about Salesses’s political views and still appreciate his thoughts on writing and teaching, in the same way as you can still enjoy The Painted Word and From Bauhaus to Our House without having to subscribe to Wolfe’s political views. Pushing a position to its extreme can yield interesting results.
And more!
It seems that Salesses was doing this Pleiades Magazine blog for some period in 2015. Some googling turned up this fun list. Here’s Salesses:
I have been thinking for a while about how our attempts to define craft terms influence our students’ (and our own) aesthetics, and I have wanted to try other definitions. How to define “tone,” for example, seemed especially difficult. Here are some alternate definitions, for now:
Tone: an orientation toward the world
Plot: acceptance or rejection of consequences
Conflict: what gives or takes away the illusion of free will
Character arc: how a character changes or fails to change
Story arc: how the world in which the character lives is changed or fails to be changed
Characterization: what makes the character different from everyone else
Relatability: is it clear how the implied author is presenting the characters
Believability: the differences and similarities between various characters’ expectations
Vulnerability: the real author’s stakes in the implied author
Setting: awareness of the world
Pacing: modulation of breath
Structure: the organization of meaning
I guess that he wrote some other cool posts but I don’t know how many, and I can’t find any link that lets me scroll through them.
A few years after writing those posts, Salesses published a book, Craft in the Real World, which . . . is on sale at the local bookstore. I think I’ll buy it!
According to the publisher’s website, Craft in the Real World is an NPR Best Book of the Year, an Esquire Best Nonfiction Book of the Year, an Electric Literature Best Book of the Year. But I hadn’t heard of it until this recent google, following up on Salesses’s posts. It’s a big world out there, when a book written by a Columbia colleague, on a topic that interests me, which received multiple awards, was unfamiliar to me. The funny thing is, when I read the above-linked posts, I thought, This guy should write a book! And it turns out he did.
I recommend reading all this along with the advice of writing coach Thomas Basbøll.
Interesting. Is the craft he refers to the craft of writing fiction or the craft of teaching fiction writing? Some time ago I saw an interview with Kurt Vonnegut, who spent two years at the Writers’ Workshop in Iowa. He was asked how we could get better writers. He responded that we have plenty of good writers; we need better readers. I’m trying to be a good reader. That’s my contribution to the world of literature.
Oncodoc:
I’d say the book is more about teaching writing than about writing. As a teacher, I found the book interesting. But I’d guess it would be interesting to writers too, in that it discusses how writers can offer feedback to each other.
Thanks for this post. Since I like teaching dataviz as workshops (although I was asked not to use that word), his experiments and comments on workshopping resonate with me.
Kaiser – Which word were you not asked to use. Dataviz? I hadn’t heard of any controversy on that one.
I can’t speak for Kaiser, but for myself, I’ve always hated “dataviz”. Unless we agree to spell it “vizualization”.
IMO “dataviz” and “data visualization” apply to the NYT and dataviz websites, where the “data visualizations” are created by graphic designers for pleasing visuals, not by scientists trying to understand the world we live in. Instead, scientists plot measurements on charts and graphs to observe relationships or to observe the distribution of various features of the natural or social world.
A “data vizualization” is analogous to the Flying V on the wall of the guitar store – it looks cool but doesn’t play or sound much different than any other guitar even though it costs a lot more, and it’s often used as a prop by people with more money and less talent. Plots and charts, on the other hand, are analgous to the teles, strats, les pauls and SGs: they don’t look as cool but they sound just as good if not better and they do almost all of the work.
Chipmunk:
Along these lines, I recommend my paper with Antony Unwin, Infovis and Statistical Graphics: Different Goals, Different Looks.
I just had the big insight that what we largely do in empirical science is write a kind of fiction—we have imagined models (imagined worlds) and we build up stories around them. I don’t mean this sarcastically; it’s just a fact. What we conclude from data is not a universal fact but a fictional construct that could be real, but there are no guarantees.
I will read his work.
I’d say if we’re doing good science, we imagine the way the world might be (usually inspired by data of some type), we formalize that into mathematical relationships between measurements, and we try to discover if those implied relationships actually obtain in the world in a broad enough context to be worthwhile.
At its fundamental level this is why I rail against Frequentist inference, because it fundamentally says the relationships between measurements are a type of random number generator. That’s a very poor ontology imho.
Hi Daniel,
do you mean that mechanistic modeling is possible in the Social Sciences? Or am I reading too much into your comment? If you do believe this, can you provide a successful example?
If it is possible indeed is the crucial question for quantitative social science, I believe. If it is not possible, we can improve methods and statistics further and further, it won’t lead anywhere. Look at all these useless (in my view) laboratory experiments in sociology or economics. Researchers claim to have found some causal effect. And maybe they have. But this causal effect is conditional on the experiment, the sample, the specific measurements etc. There is no way to generalize to the “world in a broad enough context” as you say, because there is no theory aka mechanistic understanding.
Mechanistic modelling was starting to take off in psychology before NHST quashed all progress. Look at eg Thurstone’s learning curve from the 1930s: https://psycnet.apa.org/record/1935-02633-001
If you want to get more reductionist, check Roman y Cajal and his “laws” of neuronal architecture (from the 1910s).
There have been some successful mechanistic models in social sciences. One example is
https://en.wikipedia.org/wiki/Schelling%27s_model_of_segregation
Another example is models of wealth distribution like this: https://www.scientificamerican.com/article/is-inequality-inevitable/
For the most part models will be successful when they model structural issues rather than relying on accurate models of individual behavior.
So, for example, under some structures of interaction it doesn’t matter much what the individual behavior is, you’ll wind up with a certain type of wealth distribution… Then you can plug in simplistic behaviors and still get the same wealth distribution as reality, because it’s not sensitive to particular exact behavior.
That kind of thing is modelable. But if the behavior of the system is chaotic and sensitively dependent on what each person does, then it’s not going to be modelable in detail (think of the weather). Still, we can use weather models to calculate averages or trends in climate, so in some sense even if the weather model doesn’t tell you accurately what will happen in 3 weeks, it may still tell you what the range of things that might happen 10 years out looks like.
Similarly, you probably can’t predict precise quantities for say wheat production under some, say, tariff scheme, but you may be able to predict that there will be some oscillations or changes in prices of other goods and shifts to different consumption patterns which may all take some approximate amounts of time and cause shortages in some other industries etc.
Mechanistic models don’t have to be precisely correct in the details to give useful systemic predictions.
Hi Daniel,
you mention Shelling’s model. For me, this model is not successful at all. Yes, it recreates patterns of segregation, but it does not explain them.
Assume your neighborhood is, over time, segregated more and more. You ask yourself: “What is going on? Is everyone racist?” (You conclude people are racist if 90% of their neighbors have to be from their own group, otherwise they move) By studying Shelling’s model you find that mild in-group preferences may lead to the observed amount of segregation (People are happy if 30% are of their own group). Can you conclude that everyone is in fact not racist? No you can not. Shelling’s model is not a realistic model. It does not include costs of moving. Assume in your neighborhood people have to pay $1000 more for a new flat. For many, this is unaffordable, they do not move at all, even though they are racist. Some move, but move now only if the share of their own group is below 10%. Some are rich and move if their share falls below 70%. The point is you can not infer anything about people’s preferences without explicitly modeling their movement patterns as a function of both: Preferences and costs.
I am surprised that you consider such a simple and unrealistic model as mechanistic because usually you argue for more complex and realistic models (rightly so).
Huan,
Thanks for your questions. I agree with many of your concerns. In many ways I would love to be able to raise the bar up like you are. We should demand social sciences do much much more than they are. But we have to meet the world where it is, which is learning to jump over a 1 inch thick rope on the ground, not high jumping a 6ft bar, yet.
I think we have to ask “successful at doing what?” (we also have to ask why is it that there are relatively few mechanistic social science models for me to point to? but that’s a different question I’ll touch on at the end.)
What Schelling’s model tells you is that segregation is potentially explainable from relatively mild preference to be near people who are similar to you (and you rightly point out it would be good to try out different models with things like moving cost involved).
Schelling’s model doesn’t explain why segregation is happening in any given instance, it basically just tells you that segregation is potentially not an unusual feature, most populations will experience it to some extent because even mild preferences potentially cause it. In fact, what it does is it moves the baseline you want to compare to. If you see segregation you now need to ask, is this somehow “more” than what you’d expect from a model like Schellings? Instead of having the baseline be “equal mixing of all cultures like uniform sized sand grains on the beach” which I think was the prior assumption. “eliminate racism and you eliminate segregation” is just wrong. It’s not “racist” to simply want to be around people who share traits and experiences that you do.
So, it’s successful at moving the theory away from a concept of “segregation is caused by severe individual racism” to something along the lines of “segregation is potentially not giving us much information about racism without additional sources of information”.
To me, that’s a successful model, one which informs our theories and leads us to look for the right kinds of theories. I think for you it’s not successful because it didn’t explain racism in some specific place. But instead it asks a new question: which instances of observed geographic segregation are the result of racism? Prior to that I think the assumption was probably “all of them are obviously racist”.
For example, we might want a theory of structural racism, redlining for example. We could look at different geography where redlining occurred more or less intensively and see what the long term effects are, separately from those predicted using Schelling’s model. Or we might look at places that had redlining change at different times. Or we might look at places where individual measures of racism are high. So then we can see what differences those produce *relative to a more realistic baseline*
Sure, you could say this isn’t a crazy successful model. But then I think we both have to look at how much social sciences have adopted structural or mechanistic models. They just mostly haven’t. So we don’t see a ton of successful sophisticated ones because they’re apparently well beyond the abilities of the typical social science researcher. My impression is until recently (say 2010) even using R to read CSV files and run regressions in lmer was considered “advanced computing” for social sciences, and that economists and demographers for example have historically used more “canned” stuff from SPSS or Stata or SAS. So coding agent based models and running them in massive parallel runs on clusters to get statistics on tens of thousands of runs over hundreds of parameter combinations is kind of well outside what you’d expect a recent undergrad from econ to be able to do.
Which is sad, because this kind of thing is 100% teachable to high school students. I’ve introduced agent based models to a local outreach group at Caltech (where a friend works) and they’ve had success at getting excitement and a willingness to experiment with them from **4th and 5th graders**
Hopefully that helps explain more.
Schelling’s model was successful as a preliminary model, but we’re *still* at the very beginning of being able to use these models in social science. They need some champions, and some funding.
Daniel,
Your point that Shelling’s model provides a more realistic baseline is a fruitful and interesting way to look at it. I have not thought of this before!
However, we seem to have different ideas what modeling is and should do. For me, modeling has one of three purposes: Description (model as data summary), prediction (predicting new observations), explanation (discovering causes and estimating their effect). Shelling’s model achieves neither. For you this is not a problem because it provides an insight: Segregation is normal and is to be expected. The model informs our theories. It appears that your goal is very different from my goals! Are there any relations between your goal and my goal?
Also, I am unsure if Shelling’s model is really such a useful new baseline. Shelling’s model implies that segregation is normal, if you assume zero cost of moving. But what happens if you relax this very restrictive assumption? What if you add costs and budget to the model? Maybe Shelling’s model then implies that segregation is not normal and, under this more realistic model, is to be expected only if people have very strong in-group preferences. Sounds like a wholly different baseline.
My point is that even as a tool to sharpen our theories a model needs to be somewhat realistic and it needs to include the most important variables. I think we agree on this, it just seems we view Shelling’s model differently.
Btw one more thing I do not understand (even more OT :D). Let’s assume we make Shelling’s model more realistic. Individuals have in-group preferences. They derive some utility from their neighborhood that depends on the share of similar persons. Further they try to maximize their utility by moving. A move has associated costs and the individuals have budget constraints, financially but also stuff like attachment to the neighborhood, friends etc. How do we model this? Ummm, wait, …,individuals maximize their utility under constraints… ummm, wait, there is already a theory that models the social world like this. Rational choice theory, the underpinnings of your favorite academic discipline, economics. I find it ironic that you view Shelling’s model positively but economics not. You argue that macro economic models need to include banks, because the standard neoclassical baseline is too unrealistic. But ignoring costs in Shelling’s model is fine? To me it appears that a natural extension of Shelling’s model leads us directly to a rational choice model. I do not understand how you can view the former as mechanistic but the latter not.
huan,
Perhaps I have overstated or you have over interpreted my statement that Schelling’s model is successful. Is Schelling’s model the be-all end-all? no. Does it show us that it’s possible to think mechanistically and then simulate the mechanism and see what happens and gain insight? Yes. It’s a proof of concept which showed us that we could proceed in a certain direction and get fruitful insight.
I agree with you that it shouldn’t have been the end of the story, it should have been an invitation to expand on the whole genre of models, to include questions about moving cost, strategic thinking, etc but that hasn’t happened to the extent that either of us would like, though I do think there is some stuff in the literature that I’m probably not fully aware of.
So Schelling’s model was successful because it showed *geography matters* for choice, and *dynamics matters* because the system changes in time. And maybe there’s some conditions under which in fact we never get to an equilibrium, we have a constant flux of people moving. And maybe there’s a distribution of “desires for homogeneity” and maybe only some people wind up moving more, like people on the edges of neighborhoods, etc.
Note that what an individual agents rational choice looks like depends on the physical layout of the current neighborhoods! Without some accounting for physical layout and the behavior of other agents we can’t just do “rational choice” in some vacuum. So any model of rational choice of neighborhood that doesn’t include geography is automatically broken.
Let’s forget about racism and soforth and think about something more typical for economics. Why does the rental market in San Francisco (or Los Angeles, or NYC) look like what it does? Given where the jobs are, the existence of rent control, certain neighborhood housing stock quality, transportation time, and soforth, what should the distribution of rents, occupancy, and empty housing look like? Why did it change dramatically between 2008 and 2018?
Rational choice with bounded information is potentially fine as a model for individual agents, but that doesn’t give us a solution without running the simulation. And there will be randomness (a person searches for housing but only finds ads for a few in a given week, and applies to several, and the owner gets multiple applicants, and chooses one of them at fairly random etc) So outcomes come from a stochastic dynamic process and we need to run the simulation many times to determine a distribution of outcomes, like time-to-housing and price paid and fraction of housing left empty, and tenancy duration etc.
If you’ve seen anything like that in Economics please let me know. I haven’t but again I’m not looking for it daily.
Hi Daniel,
thanks for your clarification. I understand now better where you are coming from. Apparently we were talking about two different parts of the modeling of an aggregate phenomenon. I talked about modeling of an individual agent. You talked about aggregating these individual behaviors. In Shelling’s model it mattered to me that the model of individual agents is too simple. To you it mattered that the aggregation was done properly.
So essentially much of social science breaks down for you not because of unrealistic assumptions on individual behavior, but because the aggregation of these behaviors is broken. You say social systems are dynamic and these dynamics have to be accounted for. This you want to achieve via agent based modeling. Am I getting this right?
This is of interest to me because much of the critique of rational choice focuses on unrealistic assumptions made about individual behavior. Now that I think about your suggestions, the lacking empirical success of rational choice could be due to a broken aggregation rather a broken model of individual behavior.
Huan,
Yes, that seems like a good explanation.
I believe the structure of systems is important. Given a certain structure, outcomes of a certain type will appear almost independent of individual choices (at least, for a wide variety of individual behaviors). This is so called emergent phenomenon.
Of course individual behavior does matter crucially in some contexts, but in others it’s the aggregate opportunities or structure more than anything else.
Let’s take an example most people on the blog are familiar with. Create a system where continued employment and thus eating, housing, and survival in the career, are conditional on writing and publishing papers in journals that are gatekept by existing groups and the same groups make grant funding decisions.
Now, certainly some people will do studies and write good papers that are informative. But doing so requires real work, and takes time, and the timescales between grants are fixed, so there isn’t a lot of leeway.
Certainly in that structural system, some people will hit on the strategy of taking various shortcuts, idealizing certain methodologies that are easier and faster to get a paper published. As long as there is a plurality of research styles, and such a structure, we should expect the growth of the population to be concentrated among those who succeed in getting the quick papers.
So the largest populations after a while, should be the people doing the quick methodologies and/or those studying the “hottest” topics using the most commonly approved methods.
We don’t need to posit that the individuals are corrupted through time, only that after some time they can be forced out by failure to meet some standards, so that some groups persist and others don’t.
So aggregation and systemic issues can lead to, for example, people doing idealized and standard and accepted forms of economic research and on topics that get approval from the group that must approve them, and not doing things that take a lot of time and effort and skills that are not commonly taught or appreciated by the publication and funding gatekeepers.
Hence, maybe people go running around looking for “natural experiments” or “difference in difference” or “synthetic control” situations and then quickly write some papers after doing some standardized regressions in off the shelf software that gives standardized output tables that the entire population of people have been taught to interpret.
And examples of things that would not persist would be massive parallel computing of agent based models across many economic conditions in custom built software frameworks with ambiguous but suggestive results that embrace uncertainty and seek to explain structural features of economic systems but that give no nicely packaged easily interpretable results.
I would love to see a lot more agent based modeling, both in econ, and in areas of engineering. Examples might include power generation and transmission, water resources management, human migration in climate affected regions, healthcare systems and the effect of mental illness on utilization and costs, housing construction, taxation, and generational wealth transfer… Etc etc
I personally believe that many economic systems have dynamics that extend for decades, and while economists thing of equilibrium as arising rapidly, on the order of microseconds to a week perhaps, they are completely blind to situations such as when a policy is put in place in perhaps 1974 and then it trends through time towards a build up of structural problems of catastrophic proportions by say 2024 and the entire build-up is predictable dynamical consequences of variation in actions, some structural rules, and physical or legal limitations on what is possible.
50 years of dynamics vs a view that equilibrium occurs in a couple weeks…
That sounds like a perceptive insight. Always striving, never arriving. Reminiscent of Freeman Dyson’s scentific theology, maybe (https://johnhorgan.org/cross-check/freeman-dysons-disturbing-scientific-theology )