Recently, I had just gotten home from PT when I got a call that one of my kids was sick and needed to be picked up. By the time I got to the school, it was clear it was a stomach virus, and I immediately worried about how many of us were going to catch it. The combination of frustration at a lingering (and highly impactful) injury & the prospect of a stomach virus moving through the family led me to draw this:

Drawing silly cartoons has always been a way for me to process things (here’s one example), and this one also prompted me to sit down and re-evaluate my sabbatical goals. Clearly the ‘crank out lots of papers and go for mid-morning runs’ goals have gone out the window. As I thought about it more, I realized a key goal for the remainder of my sabbatical: emerge from it less harried. 

I think harried is a fun word (if not a fun feeling). I think this definition* is apt: rushed, panicked, overly busy or preoccupied. So, my goal for the second half of the semester is to try to feel less like I need to crank things out and more like I should enjoy the chance to take my time working through things.

I also have a few secondary goals:

• take advantage of being less busy to do more reading of work-related books. I love reading fiction and do a lot of that, but rarely read non-fiction, even though there are many books I want to read. Sabbatical seems like a great time to try to carve out some time for reading work-related books.
• be more intentional about connecting and interacting with people. I spent a lot of time during my first sabbatical meeting with people around campus as I worked to set up a task force on graduate student mental health. I really loved that and have missed that on this sabbatical so far. I’m realizing that, for me, being away from my normal teaching, committee work, etc is a mixed blessing – time to focus on research (and injury recovery), but also fewer interactions.

The first of those secondary goals seems in line with typical sabbatical advice. The second is not, and feels a bit like doing sabbatical wrong. However, I’ve never regretted spending my first sabbatical the way I did, and I think that doing things like meeting with undergrads who want to meet, going to our regular seminar series, etc. could be useful to me, even if they’re the sort of thing we’re often told we should step away from on sabbatical.

I also want to spend some time thinking about what goals I want to carry forward. There are things about sabbatical that just won’t be possible once I’m back to teaching, service, etc. – that’s sort of the point of a sabbatical, after all! At the same time, I think this time for experimenting and reflecting could help me identify a few things that are helpful and that could be carried forward. As a friend who is also grateful for sabbatical opportunities put it: sabbatical is a great time to take a step back and reflect on different aspects of our work lives, including to think about what we want to carry forward.

I certainly wish the first half of the semester had gone differently, but hopefully the perspective I’ve gained will be helpful. I’m hoping the second half feels less harried, and also looking forward to spending more time reflecting on what to try to carry forward. 

*From the GNU version of the Collaborative International Dictionary of English, which appeared when I searched ‘harried’ in DuckDuckGo.

The Pokémon Company is hiring PhD-holding ecologists! My best friend from college knows Japanese and lived in Japan for a while; he and I could apply for this job together. I can’t see any downside to this plan.

Maryland tv weather forecaster discovers Peak Analogy. That’s it, it’s over, it’s all downhill from here for analogies.

In honor of the Word Baseball Classic, I’m reupping this. Great post if I do say so myself.

An argument that academic papers should be written to be easily readable by their primary audience: LLMs. We now go live to Stephen Heard.

Sorry, my apologies for the technical glitch, that wasn’t Stephen Heard. I’m told the problem is fixed and now we can go live to Stephen Heard.

Under what circumstances should non-experts defer to the judgement of experts? That is, when should not experts believe something on an expert’s say-so, purely because that person is an expert? In a new book chapter, philosopher Alex Worsnip looks at how the answer to that question changes when expert testimony involves covert value judgements. If you’re looking for an change-of-pace reading for your next lab group meeting, this might be a good pick. Purely anecdotally, my impression is that many science grad students don’t even realize when they are making value judgments.

Jim Olds on the “chronology problem“–we often underestimate the extent to which recent scientific advances depend on, or are minor variants on, much older advances. The various examples discussed are all very interesting. But I didn’t come away from the post convinced that the “chronology problem” is actually all that big a problem, in terms of slowing the progress of current research, or in terms of creating a mistaken impression in the minds of the public or policymakers regarding what’s “settled science” vs. what remains unknown.

LLMs exhibit “emergent misalignment”: if you train them badly in one domain (e.g., train them on lots of bad computer code), they start producing bad outputs, not only in that domain but also in other unrelated domains. Is the same true for people and organizations? Good blog post, good example of thinking out loud. Doesn’t have a well-worked out argument or any firm conclusions, but lots of interesting little thoughts and passing remarks.

Reading Schelling in Tehran. I thought this was very good, though obviously I’m not really qualified to judge.

It’s the end of the world as we know it–just like always.

The interview was conducted by email.

Summarize Song and Levine (2025) for our readers. The paper proposes a new analytical approach (well, new to ecology). What’s that approach and why is it useful?

Ecology has a model hoarding problem. We have way too many models and, at the same time, way too little confidence in any of them. Stefano Allesina once joked in a Quanta Magazine interview that physics textbooks stay roughly the same length over time, because experiments mercilessly execute old theories as fast as new ones are born. Ecology textbooks, on the other hand, just keep getting fatter. We’re fantastic at proposing models, but we’re notoriously terrible at retiring them. Just for predator-prey interactions alone, there are over 40 competing models for how predators eat. There’s a saying I’m fond of: if you keep your mind too open, your brain falls out.

So how do you actually discipline these models? That’s been a central research obsession of mine. We have another project in progress—on what we call label invariance—that’s also about getting better at retiring models. But this paper focuses specifically on what you can squeeze out of time series data.

What we did was borrow a tool from a completely different field. The idea comes from queueing theory—you know, the branch of mathematics that deals with how long you’ll wait in line at Disneyland. Biophysicists, notably Andreas Hilfinger and Johan Paulsson, later borrowed it to study gene expression in cells. The key insight is simple. Every model of population dynamics splits the world into “gain processes”—births, immigration, things making more of you—and “loss processes”—deaths, emigration, things making less of you. It turns out that for any given model, there’s a precise, inescapable mathematical relationship that must hold between how these gain and loss rates covary with the observed population abundance over time. We call this the “covariance criteria.”

This relationship holds no matter what else is going on in the messy real ecosystem you haven’t bothered to model. Indirect interactions? Doesn’t matter. Environmental chaos? Doesn’t matter, as long as things are roughly stationary. You often don’t even need to know the parameter values. So it’s a genuine lie detector test for your model’s structure— not just checking whether one is clever enough to twist knobs until the curve looks right.

We let this loose on three classic ecological headaches: the decades-long functional response debate (is predation prey-dependent or ratio-dependent?), figuring out how to model rapid evolution, and the hunt for higher-order species interactions. When the biophysicists first used this method in their field, it went on a brutal killing spree— falsified almost every published model of gene expression. We fully expected a similar bloodbath in ecology. But much to our surprise, the classic Lotka-Volterra model—the one that textbooks routinely dismiss as too simple and unrealistic—actually survived the gauntlet. So the method is a genuinely useful tool—not just for retiring bad models, but for proving that some of our oldest approximations actually know what they’re doing.

Can you say a bit about how that paper came about? It’s an unusual paper, in that it’s taking a methodological approach originally developed in one field—physics—and applying it to a very different field (ecology).

I did my PhD in an engineering department, so there weren’t many theoretical ecologists to grab coffee with. I ended up just hanging out with the biophysics community in Boston instead. One day, a friend casually mentioned some work coming out of Johan Paulsson’s group at Harvard—they had just built this mathematical method to audit models of stochastic gene expression. Honestly, it was never meant to be a research project. Like any easily distracted grad student, I was just curious to see how the mathematical gears turned under the hood, so I sat down and worked through the equations for fun.

Then I did absolutely nothing with it for years. Classic academia. It wasn’t until my postdoc at Princeton that the idea came back to life. My advisor, Jonathan Levine, gave me complete freedom to chase whatever I found interesting. I pitched this idea to him and we started working on it together. But the translation was far from straightforward. Biophysics deals with discrete counts of individual mRNA and protein molecules inside cells—the math is built on master equations that track every single reaction event. Ecology works with continuous variables like biomass and density. On top of that, ecological time series are typically shorter and noisier than what biophysicists get to play with.

So Jonathan and I spent a lot of time figuring out exactly when and why the math survives the border-crossing into ecology. We wanted to be very careful on that front. And I want to acknowledge that the editor and reviewers at Nature Ecology & Evolution were incredibly rigorous and constructive about it too—you can actually read our lengthy back-and-forth in the published peer review file.

I’ll say something more general here. There’s a seductive illusion in interdisciplinary work that you can just grab a method from field A and slap it onto field B. That rarely works and honestly shouldn’t work. The intellectual challenge—and where the real value lives —is in the translation: understanding precisely when the mathematics transfers, where it breaks down, and what genuinely new insight the translation reveals about your field. Think of “complex networks science”—it was initially pitched as a universal theory that would unify everything from the internet to ecosystems to the brain. Two decades later, the most productive work in network science is deeply discipline-specific.

The approach described in Song & Levine (2025) looks like it could be quite powerful. And it can be applied to observational time series data; ecologists already have a lot of time series data. Do you worry that that makes the approach ripe for abuse? Or maybe a better way to phrase the question is: when you’re proposing a new approach that you hope others will use, how do you find the right balance between “selling” the approach—highlighting its strengths —and making its limitations and weaknesses clear?

That is the eternal tightrope walk of methods development, isn’t it? You want people to use your new tool, but you also definitely don’t want to watch them accidentally hit their thumbs with it.

When it came to “selling” the approach, we were very deliberate about transparency. Instead of burying the limitations in a supplementary file, we put them front and center in the main text as guardrails. We explicitly spelled out that the method is for systems driven by deterministic forces, like Lotka-Volterra predation cycles, not just stochastic noise jittering around an equilibrium position. We also highlighted that the test is better at catching bad models than at guaranteeing good ones: a “pass” should strictly be read as “promising, let’s investigate further,” not “case closed.”

Of course, being candid can be risky—I’ve certainly had my share of rejections. But I think it’s absolutely worth it. In my experience, the ecology community largely rewards honesty about limitations. And if a reviewer does penalize you for admitting what your method can’t do, well, that’s a trade-off I can live with.

The “use and abuse” part of your question is a harder nut to crack. On the one hand, we want to make the method as accessible as possible—we built an R package to standardize the entire statistical workflow so people can apply it correctly out of the box. On the other hand, accessibility inevitably invites misuse. The history of quantitative ecology is littered with sophisticated techniques that were initially met with hype, then applied carelessly, and eventually discredited—not because the math was wrong, but because the field’s enthusiasm outran its caution. I don’t pretend to have a magic bullet for that. To paraphrase the famous quote, I’d add that all methods are abusable, but some are worth the risk. At the end of the day, you just have to be radically honest about the boundaries of your method and hope the community meets you halfway.

Your bachelor’s degree is in mathematics and your PhD is in civil and environmental engineering. And you have numerous papers besides Song & Levine (2025) that take ideas from physics and engineering and apply them to ecology. There’s of course a long and proud history, going all the way back to AJ Lotka and Vito Volterra, of people with training in mathematics, physical sciences, and engineering bringing ideas and techniques from those fields into ecology. There’s also a long history of resistance to those ideas and techniques from some other ecologists. What’s your sense of where things stand today? Is the conversation around the application of physics and engineering ideas to ecology different than it was 20 or 50 or 100 years ago, and if so, how?

I’m no historian and still fairly junior as a researcher, so take this with a generous grain of salt. But my sense is that the conversation has shifted enormously, and mostly for the better.

Robert May, who was trained as a physicist, once described the cultural shock he experienced upon entering ecology in the 1960s. The equations ecologists were using, he said, were “in some important ways different from the more familiar ones of physics.” What struck him wasn’t that ecology was less mathematical—it was that the relationship between the math and the empirical claims was looser, more tentative, less disciplined. I think what’s changed since then isn’t so much that ecology has become more quantitative —it has, obviously—but that the ambition of the quantitative work has shifted. We’ve gone from borrowing individual equations to importing entire intellectual frameworks— and then doing the hard work of figuring out what they reveal about ecology specifically.

Let me give you two concrete examples. The first is what I consider the most successful application of physics to ecology in recent years: the cavity method. It’s a technique from statistical physics, originally developed to understand disordered magnets—spin glasses, if you want the jargon. A biophysicist by training, Guy Bunin, had the key insight that the same mathematics could calculate how many species survive when you assemble a large, random ecological community. Since then, a serious community of mathematicians and theoretical physicists has built on that foundation, extending and refining it. Before the cavity method, our understanding of multi-species coexistence was mostly simulation-based or limited to the assumption that all species must coexist. After it, we could map out entire phase diagrams of ecological communities—and the method has already inspired new experimental designs and uncovered new patterns hiding in old data. That’s a profound upgrade, and it came directly from physics.

The second is very much a work in progress, but it’s something I’m genuinely excited about: working with biophysicist collaborators to bring the concept of irreversibility —a fundamentally non-equilibrium idea—into ecology. And here’s where the history gets deliciously ironic. Almost exactly a century ago, Alfred Lotka himself—the Lotka of Lotka-Volterra, arguably the founding figure of mathematical ecology—tried to do precisely this. He published a paper in Science arguing that irreversibility was central to understanding living systems. If you read the first few chapters of his classic Elements of Physical Biology (1925), there was hardly any math, but philosophical pondering of why irreversibility is the fundamental feature of life. It was prescient, it was profound, and it was almost completely ignored. That Science paper has been cited only a handful of times in a century.

So Lotka planted this seed in the 1920s, the soil wasn’t ready, and now—armed with tools from non-equilibrium statistical physics that didn’t exist in his lifetime—we’re finally in a position to grow it. I personally find it enormously encouraging. It suggests that the real barrier to physics ideas in ecology was never intellectual resistance per se. It was a lack of the right infrastructure—both mathematical and empirical—to make those ideas precise and testable. As that infrastructure develops, ideas that were once dismissed as too abstract or too foreign keep turning out to be exactly what we needed.

Two long-term trends in the field of ecology are that it’s becoming a more applied field, and that it’s becoming a more quantitative field, with an increasing emphasis on quantitative methods development. How do those trends affect the sort of research you do and how you present it to others (in talks, in papers, in grant proposals, etc.)? Because on the one hand, your work is quite quantitative and proposes lots of new quantitative methods, but on the other hand your work often concerns quite abstract and fundamental research questions.

Here’s what I think makes ecology genuinely unusual compared to other branches of biology: the field still has a deep soft spot for theory. If you talk to theorists in molecular or cell biology, they often feel completely decoupled from the empirical work happening around them. But in ecology, an enormous amount of empirical research is still explicitly framed around testing or extending theoretical predictions. That’s a rarity in modern life science, and it’s a big part of why I consider myself incredibly lucky to be doing this right now.

That said, the job market has definitely noticed the trend toward applied and quantitative work. If you scroll through the eco-evo job boards nowadays, finding a posting for a “pure theoretician” is like spotting an ivory-billed woodpecker—it’s practically extinct. The market overwhelmingly wants “quantitative ecologists” or “computational biologists,” titles that carry a strong implication of being closely tethered to data and immediate application. My PhD advisor gave me a friendly warning early on in grad school that getting hired as a professor as a pure theorist was an uphill battle bordering on a cliff climb.

Given that tension, I actually consider myself incredibly lucky to enjoy both sides. On the one hand, I love the abstract, fundamental questions—the “pure” theory—because that’s where you figure out the fundamental rules of the game. On the other hand, I genuinely love working with data. If my work can’t eventually shake hands with messy empirical reality, it’s just a glorified math puzzle.

Do I think we still need space for pure theory? Absolutely—and I want to be clear, this isn’t a knock on quantitative ecology at all. The rise of data-driven, computational approaches has been fantastic for the field and has produced tons of genuinely important work. But a field that only ever solves today’s applied problems eventually runs out of the fundamental insights needed to solve tomorrow’s. I think ecology departments should be actively making room for theoreticians—even the unapologetically pure ones—because the payoff may not be immediate, but it compounds. Of course, I am biased

First, a bit of throat clearing: I think it’s fairly rare for fundamental research in ecology, or any other field of science, to have direct practical applications. I think a lot of fundamental research does end up getting applied somehow, but usually in indirect and diffuse ways that are hard to steer (or even predict, or even trace in retrospect). I also think that people who do fundamental research also are pretty good at fooling others, and often themselves, into thinking that fundamental research has direct applications, even when it doesn’t.* Conversely, I think a lot of the most directly relevant applied ecological research isn’t all that interesting from a fundamental perspective, often because it’s very species- or location-specific. For instance, if you need to parameterize an integral projection model for some species of conservation concern, as a step towards modeling possible management interventions, it’s unlikely that your work is going to speak to any interesting fundamental conceptual issues in population ecology, or lead to the development of novel, broadly-applicable statistical methods. So I think it’s pretty rare for any ecological research to be of both great fundamental interest and direct practical importance. But surely there are a few examples?

The first one that comes to mind is trophic cascades research, especially in lakes, and its applications for managing algal blooms. But I don’t actually know anything about how algal blooms are managed, so maybe I’m wrong about this one?

The second one that comes to mind are simple stochastic population growth models and their applications to managing species at risk. Correct me if I’m wrong, but I vaguely recall that, in some countries, some of the official legal standards for designating species as “at risk” or “endangered”, entitling them to legally-mandated management intervention, are based on “rules of thumb” derived from simple stochastic population growth models?

The third example I can think of off the top of my head isn’t a single example, it’s a type of example: system-specific, applied case studies that became textbook illustrations of broadly applicable, interesting fundamental concepts. Stuff like applied research on various pest and disease organisms (e.g., large budmoth, measles), that ended up providing textbook illustrations of cyclic population dynamics.

Ok, over to you! Looking forward to your comments, as always.

*I’m thinking for instance of Peter Adler’s old guest post, taking fundamental ecologists–including himself–to task for bullshitting about how fundamental research in ecology will improve forecasting in a way that will be useful to land managers and other ecological decision-makers.

A reader of The Ecology of Ecologists just posted a very positive review on Story Graph. Remember: you can get a free signed bookplate if you tell the world–or even just me–what you think of my book.

An argument that, rather than replacing genuine expertise, AI will make it more valuable.

On Stewart Brand. I don’t really know much about Brand or the various broader movements he’s participated in or influenced over the past several decades. FWIW, just based on the bits and pieces I’ve picked up over the years, he seems like an interesting figure. He combines ideas and attributes that usually are seen as opposing. For instance, he was a towering figure in the environmental movement–but he’s also a pro-technology optimist who admires Elon Musk. I’m always interested in people who resist easy categorization.

The editorial board and reviewer pool at the Journal of Applied Ecology have both become more gender-balanced and geographically diverse over the past 20 years. The trends in reviewers reflect changes in both the composition and behavior of editors: the editorial board is not only more gender-balanced and geographically diverse than it used to be, but both men and women editors have been inviting more women reviewers.

Scholars of Old English focus disproportionately on just a few canonical texts in the entire corpus of Old English writing, although to a lesser extent than they used to decades ago. This may sound like a boring blog post, but it was actually really interesting–and it has graphs!

Political economist Chris Blattman shares his Claude-based workflow, which he swears by for everything from managing his inbox, to grant proposal drafting, to trip planning. Includes instructions, including for some bits that should work “straight out of the box” with no customization needed and no need to install Claude Code.

Cosma Shalizi’s unsolicited opinions (basically, ideas for blog posts he’s never going to get around to writing).

Stephen Heard on the strange mix of feelings prompted by correcting the proofs for the third edition (!) of The Scientist’s Guide to Writing. Speaking as a book author whose book would be doing very well if it sold 1/10 as many copies as Stephan’s, I was left wondering: what about the strange feeling of sleeping on a mattress full of cash from book royalties, huh? Tell us what’s that feels like, Stephan!

New Brunswick is looking to slash higher education spending…somehow. Rather than looking to cut spending, they should instead raise revenue by appropriating Stephen Heard’s book royalties. Boom, problem solved.

Artist uses frozen Alberta lake as his canvas.

So here are a bunch of posts I haven’t written yet, some of which I’ll never write. But they’re all mixed up, so you’ll have to guess which is which.

When should you just keep doing what you’re doing? When does that demonstrate clear-eyed focus/an admirable resistance to faddishness/etc., and when does it amount to becoming lazy/self-indulgent/stuck in a rut/deadwood?

Are recent TT hires at R1 unis less likely to have ever been VAPs than TT hires at less research-intensive unis?

Something about Keynes’ line in Economics Possibilities for our Grandchildren about how economists should become like dentists. The thing about dentistry is that it’s super-useful but also super-boring. It’s routine, not creative. It doesn’t have unsolved puzzles or new ideas. Nor does it involve clashing values, mediating conflicts between stakeholders, etc. Could, or should, ecologists become more like dentists?

A post on my failure to replicate a classic protist microcosm experiment by Leo Luckinbill

Is it true that Ivy League unis and elite SLACs tend to hire their own alums? Could compile data on this from department websites, use other R1s and non-elite SLACs as a control group.

Maybe we should stop studying stabilizing vs. equalizing mechanisms?

The modern coexistence theory backlash has begun

What’s the strongest trade-off in ecology? Are there any that always show up in the data, without any need to control for the organism’s health/quality/environment?

How much does meta-analysis improve on unweighted averaging of effect sizes?

Have ecology faculty job seekers’ collective preferences to live in certain places, or for certain types of jobs, gotten stronger over time? Look at predictors of # applied from ecoevojobs.net, broken down by year. Problem: would need to compile a lot of new data on job-level predictors, not just state-level predictors as I did in a previous post.

Best examples of ecological research that’s of great fundamental interest AND has important, direct applications?

Read some old articles on what ecology is and how to do it, and then post about them. Taylor 1936 Ecology. Stiling 1994 ESA Bulletin. Bergman and Adams 1993 ESA Bulletin.

Ecology is full of variation. So why do ecologists’ hypotheses about the drivers of that variation so rarely pan out?

Is it more common these days than it was 10 years ago for TT ecology hires to have MSc degrees? Relatedly, are TT ecology hires these days spending longer from bachelor’s degree to getting hired into their first faculty job, because they’re spending longer in grad school due to more frequently completing an MSc before a PhD?

Should you go into admin? How should you think about the pluses and minuses? How can you maintain your research program while you’re a dept chair, or dean, or whatever? Problem: I don’t know anything about this topic.

What’s the most substantively I’ve ever been cited? Or more broadly, the most substantively I’ve ever influenced anyone else’s research? Has anyone besides my own collaborators ever published a paper that couldn’t have been written without my work, or that would’ve looked very different without my work? Or has my work every prevented someone from publishing what would’ve been a major paper?

I find these sorts of sociological models interesting to think about, but I’m often unsure how seriously to take them. I feel like there are lots of different plausible-ish models of the peer review system, and the behavior of the (many!) people involved in it, that lead to quite different predictions. I say this as someone who was arguing more than 15 years ago for urgent, fundamental changes to the incentives to do peer review, on the grounds that the peer review system was in imminent danger of meltdown. Those changes did not come to pass, and yet the peer review system seems to be chugging along more or less as it was 15 years ago. I was wrong in thinking that the peer review system was in imminent danger of meltdown unless something radical changed. So if you say that the incentives associated with the peer review system will make it melt down eventually, my question is: when?

Related: my old review of The Bet, a book about a famous prediction of imminent ecological and societal meltdown.

Years ago, Meghan asked whether bird papers have the best figures. The answer was “yes,” and it’s still “yes.” Check out Fig. 1-2!

Related to a recent link on how environmental NGOs went off the rails, here’s how NGOs in general went off the rails. Makes for an interesting comparison with that recent link; there’s a lot of overlap even though the previous link was by someone coming from a politically leftwing perspective while today’s link is by people coming from the political center or slightly right-of-center. But there are some differences too. Obviously, I’m not qualified to weigh in on this topic myself, but I found both links interesting.

Commenter Dan Elton sends us to the Metascience Observatory project. They’re trying to compile and synthesize a comprehensive database of all replications in all fields of science, and make it available on the web with an interactive dashboard. I’ve only had a quick glance so far, but what I saw impressed me.

Sticking with metascience, here’s Jessica Hullman on how were on the cusp of living the metascience dream–or nightmare. Thanks to LLMs, it’s about to get a lot easier to do certain kinds of checks for robustness and replicability that many people have been calling for for years. But is this a case of needing to be careful what you wish for because you might get it? Very good post. It’s from a computer science perspective, but most of the points generalize. For instance, this very interesting remark about how it’s not useful to think of “researcher degrees of freedom” as sampling analyses from some statistical population of possible analyses:

Treating model variants as if they form a random sample turns analytic flexibility into an uninterpretable frequency. This is why robustness testing at scale does not guarantee insight: we still have to figure out how to interpret the results, and that is hard.

Still sticking with metascience: are we really so sure that science in the past was broken, or that it was broken in ways that can be fixed by doing more close replications? I have mixed feelings about this. Related: techniques aren’t powerful, scientists are. Also related: transparency vs. trust.

No, AI is not going to lead to mass unemployment, come on. I mean, how is that even supposed to work?

Writing in Science, Philip Ball reviews Maria Popova’s new book Traversal. I wish the review could’ve been longer; sounds like an unusual book in ways both good and bad.

Cory Doctorow on how he uses LLMs (scroll down). Includes some spicy pushback against blanket ethical objections to LLM use. Not that Cory Doctorow (or anyone) is necessarily correct, obviously. But FWIW, he’s someone who’s thought hard about this stuff, and has gone to greater lengths than most people to live by his publicly stated principles regarding technology use. A couple of quotes to give you the flavor:

You know what’s better than refusing to use a technology because you hate its creators? Seizing that technology and making it your own.

AI is like the dotcom bubble, awash in sin and inflicting untold misery, but it will leave something useful behind

Here’s how some physicists are using LLMs: to do real physics research. Apparently, the latest version of ChatGPT appreciably simplified some crucial mathematical expressions, and proved they were correct, within minutes to hours–expressions that human physicists had been trying and failing to simplify and prove for months. Link goes to a Science news article.

The owner of the former campus of Green Mountain College wants to give it away. I link to this purely because it got me wondering: what would you do with a college campus, besides operate a college? The linked article gives one possible answer–the owner wants to give it away to a Catholic mission-based organization. But what are the other possible answers? And whatever you do, you have to either gut or abandon the lab buildings, right?

Recently, I collected some new data on the true, exact number of applicants to tenure-track EEB jobs advertised in North America in the last five years. I mostly just wanted to make the total sample size bigger, but I also figured it couldn’t hurt to check whether there have been any other changes to the geography of EEB faculty job applications in the last few years. Unfortunately, that new survey didn’t yield all that much additional data: just 29 new tenure-track jobs advertised in the last five years, to add to the 86 jobs for which I already had data on the true exact number of applicants.

So here’s an analysis of the geography of EEB faculty job applications in North American, using the combined dataset of 115 tenure-track EEB jobs. What factors determine which tenure-track EEB faculty positions get more applicants?

tl;dr: the answers haven’t changed all that much from my previous analysis of the smaller 86 job dataset.

Overview of the dataset

Let me emphasize up front that this is a small and somewhat biased dataset. It’s big enough and unbiased enough to improve on whatever anecdotes you happen to have heard. But it’s not big or unbiased enough to give you a complete, accurate picture.

The main bias is that jobs that attracted a lot of applicants are overrepresented in this dataset. That’s for at least two reasons. First, the more applicants a job attracts, the greater the number of people who are likely to receive information about exactly how many applicants there were, since some colleges and universities send out rejection letters that state the number of applicants. All else being equal, the more people who know how many applicants a given job attracted, the more likely someone is to report that information in my surveys. Second, this blog’s readership skews towards faculty at big research universities. So when our readers share their first-hand knowledge of how many people applied to EEB searches at their own universities, they tend to report the results of searches at big research universities. As shown below and in previous posts, those searches tend to attract lots of applicants.

Ok, with those caveats out of the way, here’s an overview of those 115 jobs. They came from 88 different colleges and universities (80 in the US, 7 in Canada, one unspecified). The most-represented institutions are Cornell and Georgia (4 jobs each). 48% of the 115 jobs were at R1 universities according to the most recent Carnegie classification or their Canadian equivalents.* 29% of the jobs were at AAU universities (a particularly prestigious and research-intensive subset of R1 universities and their Canadian equivalents). Having said that, the dataset also includes positions at teaching-intensive liberal arts colleges (e.g., Coe College, etc.), more research-intensive liberal arts colleges (e.g., Williams, Middlebury, Reed, etc.), R2 universities (e.g., Villanova, etc.), regional public universities (e.g., Southeast Missouri State, Southern Illinois University – Edwardsville, etc.), and other sorts of institutions (e.g., New Jersey Institute of Technology, etc.).

Those 88 institutions were located in 37 states and provinces. A few states and provinces are heavily represented, but their representation only correlates loosely with state or provincial population size. Of the 115 jobs, 16 were in California, 7 each were in Texas, Georgia, and Ohio, 6 each were in New York, North Carolina, and Michigan, 5 were in Pennsylvania. Conversely, the dataset doesn’t include any jobs in Florida, and only includes one job in New Jersey and two in Ontario. If you’ve been reading carefully, you will infer, correctly, that most states are represented in this dataset by jobs at just one or two institutions. For instance, the only New York jobs in this dataset were at Cornell and Skidmore. So although I think this dataset is big enough to give some insight into broad-scale trends across many institutions in many states and provinces, it’s definitely not big enough to drill down and say anything about any one specific state or province.

The grand mean number of applicants per job was 134 and the median was 100. The distribution was right-skewed, so the median’s probably the best measure of central tendency here. But remember, both the mean and median are probably overestimates for the EEB faculty job market as a whole, due to the biases discussed above.

There was a lot of variation around the central tendency: the range was 9-1000 applicants (!)

Data analysis

In order to explain some of that variation, I compiled data on variables associated with each job, or in some cases variables associated with the state/province in which the job was located:

• Was the job at an R1 university in the most recent Carnegie classification, or the Canadian equivalent
• Was the job at an AAU university
• Was the job part of a cluster hire
• For jobs in US states, the fraction of the state’s votes obtained by Hillary Clinton in the 2016 US Presidential election, as a proxy for state political orientation. (note: I already had this data for the 86 jobs on which I compiled data several years ago, so I just updated it for the 29 new jobs.) I also tried coding each state as politically “blue” (liberal), “red” (conservative), or “purple” (moderate), but that variable gave pretty much the same results as Clinton vote share, so I just went with Clinton vote share.
• Latitude and longitude of the hiring institution
• Population of the city or town in which the hiring institution is located
• Year the job was advertised
• Was it a fisheries/wildlife/natural resources job
• Was the job in the US or Canada

As I hope is obvious, many of these predictors are crude surrogates for all sorts of attributes that (some) EEB faculty job seekers likely take into account when deciding what jobs to apply to. FWIW, I doubt you could get much more refined insight if you used different surrogates (e.g., some other surrogate for state political orientation, or some other surrogate for the “urbanness” of the area in which each hiring institution is located.)

Many of these predictors are correlated with one another to some extent. For instance, R1 status is correlated with city/town population size, because there aren’t any R1 universities in really small towns. As another example, latitude is correlated with Clinton vote share, because on average Clinton did better in northern states than southern states. But I looked at the generalized variance inflation factors, and there’s not nearly enough collinearity to be worth worrying about. The issues with precisely and accurately estimating the marginal effect of each predictor are mostly down to the small size of the dataset, not collinearity among predictors.

No doubt there are some variables that some EEB job seekers care about, for which my dataset includes no surrogates at all. In particular, I tried and failed to think of a good practical measure of how close each job is to the sorts of natural areas that EEB faculty job seekers as a group tend to enjoy. I also tried and failed to think of good practical measures of “quality of life”, either with or without consideration of housing costs or broader cost of living. (If you can think of any, let me know in the comments!)

A bit of exploratory analysis indicated that a couple of the predictors either had no detectable effect whatsoever, or else were redundant with other predictors. So I dropped the USA/Canada predictor, the year the job was advertised, and the AAU-or-not predictor. It also looked like a good idea to log10-transform town/city population size, so I did that.

I then analyzed the data two different ways: with a regression tree, and with a general linear model (type III SS), in both cases ignoring possible non-independence among different jobs at the same hiring institution. I also looked at some graphs.

Read the footnote** if you are inclined to hassle me about data transformations/distributions of residuals/homoscedasticity/bootstrapping/cross-validation/AIC/GAMs/hierarchical nonlinear generalized mixed models/dimension reduction/machine learning/clustering algorithms/Bayesian approaches/etc.

Results

The GLM had an R^2 of 0.63, so it did a pretty decent job of capturing the variation in the data. The effects of cluster hire, log-transformed town/city population size, Clinton vote share, and fish/wildlife/NatRes all came out significant at the 0.05 level, while the latitude effect was marginally significant (p=0.06). Slightly to my surprise, the R1 university effect, though estimated to be positive, was non-significant (p=0.19). The mean number of applicants increased by 468 for cluster hires vs. non-cluster hires, decreased by 74 for fish/wildlife/NatRes jobs vs. other sorts of jobs, increased by 51 for each 10-fold increase in town/city population size, increased by 3.9 for each additional degree of latitude north of the equator, and increased by 3.6 for each additional percentage point of Clinton vote share. R1 jobs got 23 more applicants than non-R1 jobs on average, all else being equal, but as noted earlier the effect wasn’t significant.

You can see most of these effects in bivariate scatterplots, shown below. Bivariate scatterplots also reveal why the GLM doesn’t pick up an effect of longitude: any effect of longitude is non-monotonic because jobs near the coasts tend to draw more applicants.

In a regression tree (shown below), the first break was associated with city/town size, with jobs in cities with >615,000 people (basically: Portland, OR on up) getting many more applications on average than jobs in smaller cities and towns (basically: Milwaukee on down). Then jobs in big cities got subdivided geographically (latitude and longitude), with big city jobs in the Pacific Northwest drawing lots of applicants, then big city jobs in New York and New England, then big city jobs elsewhere. Meanwhile, jobs in smaller cities and towns drew many more applicants if they were R1 jobs than otherwise. Then, the non-R1 jobs in smaller cities and towns got subdivided by Clinton vote share right down the middle of the politically moderate states (NC and GA), with jobs in states towards the more liberal end of the political spectrum drawing more applicants.***

Discussion

The regression tree and the GLM pick up somewhat different effects, and summarize those effects in different ways. And I definitely wouldn’t take the exact numbers from either analysis all that seriously. But putting it all together, including with my previous posts on this topic and related topics, a reasonably clear picture emerges. It makes sense to think in terms of the number of “boxes” that a given job checks for many/most EEB faculty job seekers:

• A tenure-track job that checks all the boxes for many/most EEB faculty job seekers is likely to draw several hundred applicants. Think of a broad cluster hire at an R1 university in the Pacific NW or New England (two of the most politically liberal regions of the US), in a big city that’s close to mountains.
• Conversely, a job that checks no boxes for many/most EEB faculty job seekers–a job at a teaching-intensive undergraduate-only college in a rural area in a politically conservative state far from mountains–is likely to draw, oh, 10-30ish applicants or so. Could be more or fewer, obviously, but that’s my best guess as to the typical range.
• In between those extremes, there’s a lot of variation. If it’s an R1 job in a politically liberal or moderate state, and it’s not a fish/wildlife/NatRes job, I’d say you’re likely looking at 150-300ish applicants. Probably towards the upper end of that range if it’s in the Pacific NW or New England, and towards the lower end of that range if it’s in the Midwest. If it’s an R1 job in a politically conservative state, there’s a ton of variation likely driven by all sorts of factors, it wouldn’t be surprising to see anything from 25ish applicants to 250ish. If it’s a somewhat more teaching-intensive job that still has significant research expectations–say, an R2 or an elite liberal arts college–then you’re probably looking at 50-200ish applicants if it’s in a politically liberal or moderate state, and somewhat lower than that (25-125?) if it’s in a politically conservative state. If it’s a very teaching-intensive job, but ticks at least a couple of other boxes for many/most EEB faculty job seekers, it’s hard to say; there are very few jobs like that in this dataset. Probably fewer than 100ish applicants I guess?

Also worth keeping in mind that number of applicants is a very crude measure of how “competitive” the applicant pool for a given job is. In particular, jobs that check many boxes for many EEB faculty job seekers tend to draw lots of uncompetitive long-shot applications from applicants who aren’t good fits.

In conclusion, I’m not sure what EEB faculty job seekers can do with this information. Obviously, it wouldn’t make sense to apply for a job you wouldn’t take just because few others are likely to apply for it, or conversely to not apply to a job you’d take just because many others are likely to apply. Trying to apply for only the jobs for which (you think) you’re “competitive” is a mug’s game. But I guess if you just wanted confirmation that you’re going to have a lot of competition if you apply for a cluster hire at the University of Washington, well, now you have it.

*As an aside, that’s appreciably higher than the fraction of all TT N. American ecology faculty jobs advertised in a given year that are at R1 unis. Like I said, it’s a biased dataset.

**Come on, quit trying to commit statistical machismo, or get me to do so. This a blog post analyzing a small and very imperfect dataset. Part of the art of doing statistics well is knowing when to keep it simple.

***There were further splits in the regression tree, but they resulted in tiny groupings of jobs and were hard to interpret, so I trimmed them. I doubt they’re meaningful.

Image manipulation is as old as images. Some very fun photos here.

At this point, we’re all well aware that different data analysts trying to answer the same scientific question often will obtain different results even when provided with the same data. Data analysis inevitably involves a sequence of somewhat subjective or arbitrary decisions (what statistician Andrew Gelman calls the ‘garden of forking paths’ and others call ‘analyst degrees of freedom’). One solution to this is a ‘multiverse analysis’: an analysis summarizing the results of many (all?) possible analyses of a given dataset to answer a given question. But what problem does a multiverse analysis solve, exactly? Does it actually accomplish anything besides reminding the reader (and the analyst) about the garden of forking paths?

How to make a (very good) living as an artist. The usual caveats about following the advice of highly successful people apply. Obviously I’m not qualified to judge the linked piece, but I found it interesting and FWIW it all sounded sensible to me. It’s especially interesting that a lot of it isn’t so much advice as…I dunno, therapy? It tries to help readers reconcile themselves to the advice. I often find myself doing the same thing with my posts about the ecology faculty job market, though I admit that I’m not sure if any readers find the therapeutic bits helpful. Plausibly, the people who would benefit most from reconciling themselves to a reality that they deeply dislike or refuse to face up to, but also can’t change, are the same people who will be most resistant to unasked-for help achieving that reconciliation.

I got that last link from here; click through for many more interesting links. Several of the links show that classic results in psychology continue to not replicate, in some cases because they turn out to have been faked. And item #10 in the linked list is a particularly deep and weird rabbit hole.

How many times do you feel like you need to teach (and revamp) a course before you’re happy with it?

How Emory University prof Pedro Sant’Anna uses Claude Code for academic work. Includes a GitHub repo to make it easy for you to try it yourself.

Life after p-hacking. Old preprint from psychology, but new to me, and much of it generalizes to other fields. I link to it mostly because I like the idea of teaching students to guesstimate the sample size required to achieve high statistical power by using familiar everyday cases. For instance, if you wanted reasonably high power to detect a difference in average weight between men and women (a substantial difference in means that everyone is familiar with), you’d need a sample size of roughly 50 men and 50 women. It’s a sobering thought that that’s a much bigger sample size than the typical ecological experiment, even though the typical ecological experiment likely isn’t looking for such a big difference in means.

New preprint (which I have yet to read; just passing news of it along) argues that scientific papers should be machine readable, and uses LLMs to build a workflow making them machine readable. Could be a useful tool for automated “sentiment analysis” of large chunks of the scientific literature. For instance, in many applied fields of science it would be interesting to ask how the tone with which scientists write about their results lines up with the results themselves.