Mridul K. Thomas

@mridulkthomas

ecology, statistics, experimental design, temperature, multiple drivers, plankton | www.mridulkthomas.com

685 Folgt2.695 Follower349 Posts

Ist das dein Profil?Übernimm es — dann gehört dir mridulkthomas.dirty.social.Profil übernehmen

I've been baffled by this figure. Why is Europe low when the big European countries show more - sometimes record - levels of fire this year? It turns out this pattern is entirely driven by Russia, which I didn't realise was counted as Europe in these datasets. ourworldindata.org/grapher/cumu...

Screenshot of figure showing cumulative area burnt by wildfires in Russia, by year. 2026 is the lowest since the dataset begins in 2012.

Mridul K. Thomas@mridulkthomas· 16d

Odd wildfire fact: 2026 has seen the least European territory burned of any year in the last 15. Caveat: this runs till a week ago and so misses the big ones of the past few days. And I don't know the source dataset myself, I am relying on OWID. ourworldindata.org/grapher/cumu...

A graph of Cumulative area burnt by wildfires by week in Europe, across several years. Source: https://ourworldindata.org/grapher/cumulative-area-burnt-by-wildfires-by-week?country=~OWID_EUR

Mridul K. Thomas@mridulkthomas· 20d

And this: www.outsourceaccelerator.com/articles/ai-...

Screenshot of "List of common AI models" that mentions linear and logistic regression, among other things

Screenshot of "4 types of AI models listing machine learning, supervised learning, unsupervised learning, and deep learning.

Mridul K. Thomas@mridulkthomas· 21d

They have a point - plenty of work is uninformative and a waste of effort. But I would have liked to see them grapple with the trade-offs here. We want surprise as well as rigour, and we want it to be tractable, fast, and cheap. Eventually all the low-hanging fruit gets picked!

"A paper that changes nothing about how people think is uninformative and uninteresting."

Mridul K. Thomas@mridulkthomas· 21d

The authors do seem to belatedly recognise this. Their solution is "independent methods, diverse datasets, distant researchers, and intersecting theoretical mechanisms". But all of this isn't possible in one surprising study!

"A surprising finding supported by a single, underpowered study is hype. But a surprising finding supported by independent methods, diverse datasets, distant researchers, and intersecting theoretical mechanisms is the most valuable kind of scientific contribution."

Mridul K. Thomas@mridulkthomas· 21d

The authors advocate for doing work that surprises. But this is part of what led to the replication crisis. And surprise is selected for in science journalism, which routinely inflates unreliable results. Selecting for surprise is in a sense anti-Bayesian. It is selecting for unreliability.

"... surprising combinations of research contents (what the paper is about) and contexts (which fields it draws from) are highly predictive of a paper landing in the top 10% of citations. The most surprising papers are roughly five times more likely to be massive hits."

Mridul K. Thomas@mridulkthomas· 21d

This is immediately followed by the opposing claim that scientists who shift the community's opinion the most reap the largest rewards. These claims together imply a bimodal reward structure: either be boring or revolutionary, not in-between. Perhaps - but this isn't spelt out, let alone defended.

"Progress requires persuading the scientific community to accept new empirical claims. What is less appreciated is that the farther a scientist can move the predominant position of a community, the more people will acknowledge the finding, and the more importance will be attributed to the work. The predominant view is often supported by considerable evidence, which is why moving it demands extraordinary evidence and why the system offers its largest rewards for successes and largest penalties for failures, the farther a claim sits from that view. A researcher who convinces the scientific community, with robust evidence, that objects can travel faster than the speed of light would become famous, and the researcher who presses such claims without convincing evidence would be ignored or sanctioned."

Mridul K. Thomas@mridulkthomas· 21d

The core argument is that science rewards boring research that does not change our understanding of anything.

"A paper that confirms what readers already believe tells them little they did not already know and perhaps nothing that changes their worldview. Yet, the incentives of modern science reward this type of uninformative communication. Peer reviewers are drawn from one’s own field and publication norms favor deference to established paradigms. Scientists cite sources in their immediate field roughly 500% more than those in distant fields. This is performance art designed for reviewers rather than a search for the most informative surprise designed to advance knowledge. Thus, most papers are written to be unsurprising, and consequently, have minimal impact. "

Mridul K. Thomas@mridulkthomas· 21d

I think there's plenty of scope to improve how we do science and enjoy debating this. But this is one of the more puzzling essays I have read. www.science.org/doi/10.1126/...

Screenshot of essay title "Advancing science by designing for surprise"

Mridul K. Thomas@mridulkthomas· 23d

An interest in cuisine pairs oddly well with one in etymology. (Stumbled on this reading about the origin of phyllo dough)

Screenshot of wikipedia page describing 'placenta cake'. Text below:

Placenta cake is a dish from ancient Greece and Rome consisting of many dough layers interspersed with a mixture of cheese (such as ricotta) and honey and flavored with bay leaves, baked and then covered in honey.[1][2] The dessert is mentioned in classical texts such as the Greek poems of Archestratos and Antiphanes, as well as the De agri cultura of Cato the Elder.[2] It is thought to be related to baklava.[3][1]

Etymology
The Latin word placenta is derived from the Greek plakous (Ancient Greek: πλακοῦς, gen. πλακοῦντος – plakountos, from πλακόεις – plakoeis, "flat") for thin or layered flat breads.[4][5][6]

The placenta of mammalian pregnancy is so named from the perceived resemblance between its shape and that of a placenta cake.[citation needed]

Mridul K. Thomas@mridulkthomas· Jul 13

A Shiny app to calculate and evaluate Bayesian optimal experimental designs, for common nonlinear functions in biology. Mathematical optimisation leads to some unintuitive designs! Play with priors, true values, and measurement error. See how these shape the optimal design and how well it performs.

Screenshot of the app linked at the quoted post. This image shows the function selection, some inputs (sample size, design space, custom design specification) and some outputs (calculated D-optimal and uniform designs)

Screenshot of the app linked at the quoted post. This image shows some additional inputs (true values and priors) and outputs (violin plots and summary of CRPS values across 50 simulations for design evaluation).

Mridul K. Thomas@mridulkthomas· Jul 9

A Shiny app to simulate data and fit a Bayesian nonlinear regression for several common functions in biology. This is intended to get a feel for Bayesian nonlinear regression and how much different experimental designs matter based on your assumptions.

Screenshot of the app linked at the quoted post. This image shows the function selection, parameter entry, model fit visualisation, and MCMC chains.

Screenshot of the app linked at the quoted post. This image shows the posteriors and model evaluation using a 'strictly proper scoring rule', the Continuous Ranked Probability Score.

Mridul K. Thomas@mridulkthomas· Jul 8

A Shiny app to help understand why some treatment levels matter more than others for common nonlinear functions in biology. I found that this helped me develop better intuition about experimental design. bsky.app/profile/mrid...

Screenshot of Shiny app to explore parameter sensitivity for common nonlinear functions in biology. Access the app through the link in the quoted thread.

Mridul K. Thomas@mridulkthomas· Jul 7

A Shiny app to visualise and define Bayesian priors for common nonlinear functions in biology: growth responses to nutrients, light, temperature, and toxins. Lognormal priors are often a good choice but are horribly unintuitive. So I made this to visualise them and the curves that arise from them.

Screenshot of Shiny app linked to in the quoted thread.

Mridul K. Thomas@mridulkthomas· Jul 6

BOED consistently outperformed a standard uniform design for parameter estimation, which is what we optimised for. In the most extreme case, an optimal design performed as well as a standard uniform design using just one-third as many experimental units! (8)

Nutrient (Monod) curves: Visualising curves drawn from the prior, optimal and uniform designs, and comparisons of the performance of the two designs for parameter estimation and prediction. See preprint for details.

Light (Eilers-Peeters) curves: Visualising curves drawn from the prior, optimal and uniform designs, and comparisons of the performance of the two designs for parameter estimation and prediction. See preprint for details.

Temperature (Norberg) curves: Visualising curves drawn from the prior, optimal and uniform designs, and comparisons of the performance of the two designs for parameter estimation and prediction. See preprint for details.

Toxin (log-logistic) curves: Visualising curves drawn from the prior, optimal and uniform designs, and comparisons of the performance of the two designs for parameter estimation and prediction. See preprint for details.

Mridul K. Thomas@mridulkthomas· Jul 6

We can help make your experiments better! 🌊🧪🌎 🧮➕📏 @raviranjan.bsky.social and I have been thinking about how to improve experimental design. We have a preprint with maths, simple rules-of-thumb, and R code and Shiny apps ready to help you.    www.biorxiv.org/content/10.6... (1)

A visualisation of Bayesian optimal designs for Monod (nutrient) curves at four different sample sizes, and a comparison of those designs against a uniform design using 1000 simulations per sample size. Across all 4 sample sizes, the Bayesian optimal design outperformed the uniform design for parameter estimation in >80% of simulations. We quantified this using the Energy Score, a generalisation of the Mean Absolute Error that applies to distributions and multiple parameters. See preprint for details.

Mridul K. Thomas@mridulkthomas· Jun 28

It's either going to be +36 or -12 °C today. (Interesting visualisation problem - wide gradient with fine resolution needed for a part of it)

Map showing expected temperatures today with an odd colour scale and has pink for both for very low and very high temperatures.

Mridul K. Thomas@mridulkthomas· Jun 17

Do better, French and German birds. Czech birds have left sexism in the past, learn from them. besjournals.onlinelibrary.wiley.com/doi/10.1002/...

Birds fly away from women while they are further away than men (Y-axis: Flight Initiation Distance, in metres). The mean FID seems to differ from country to country, and the difference between mean FID for men and women also differs. Lowest difference is in the Czech Republic, highest by ratio is Germany.

Mridul K. Thomas@mridulkthomas· May 21

I traced this as far as the UN World Urbanization Prospects report, which seems to be the ultimate source for this surprising graph (which is also on the World Bank and OWID sites). Based on a cursory check, the 2025 version doesn't seem to show the sharp drop. population.un.org/wup/download...

Japanese rural population till 2024.

Source: UN World Urbanization Prospects 2025

Mridul K. Thomas@mridulkthomas· Apr 24

Claude seems fantastic at writing Shiny apps. You can in principle go from knowing nothing about how to create one to having one online in minutes (obviously, best to spent more time checking it!). Fabulous for sharing ideas quickly and an excellent accompaniment to papers and lectures.

Screenshot of a Shiny app for simulating data from two different experimental designs, fitting them using Bayesian methods, and comparing the posteriors.

Mridul K. Thomas@mridulkthomas· Feb 20

Remarkably, there's even a modest association between the growth niche width and the occurrence niche width. The latter is wider on average, possibly because our growth niche widths miss a lot of intraspecific variation. 5/n

The growth niche width in the lab is positively associated with the occurrence niche width in the field. Occurrence niche width estimates are likely most reliable for unimodal curves, followed by monotonic and finally bimodal curves. The six bimodal curves (hollow squares) were therefore excluded from the regression.

Mridul K. Thomas@mridulkthomas· Feb 20

The 'median growth temperature' in the lab matches the 'median occurrence temperature' in the field reasonably well. There's certainly scatter, especially at intermediate values. But the regression line is pretty close to 1:1. 4/n

Median growth temperature in the lab predicts median occurrence temperature in the field reasonably well across 39 marine phytoplankton species. The mean response (solid regression line) is close to the 1:1 line (dashed line). Median occurrence temperature estimates are likely most reliable for unimodal curves, followed by monotonic and finally bimodal curves. The six bimodal curves (hollow squares) were therefore excluded from the regression.

Mridul K. Thomas@mridulkthomas· Feb 20

We compared thermal performance curves (TPCs) estimated from lab experiments and occurrence probability curves estimated with SDMs and a dataset of global occurrences. We excluded a few occurrence curves that were bimodal, probably reflecting data/SDM limitations. 3/n

The temperature-dependence of population growth rate estimated from lab experiments (top row) and of uncalibrated occurrence probability estimated from SDMs fitted to presence/pseudoabsence data (bottom row), for three marine phytoplankton species. These species are examples of the three shapes of occurrence probability curves that we manually classified curves into - monotonic, unimodal, and bimodal. We consider the few bimodal curves to be biologically unrealistic and exclude these species from our analyses. The vertical red lines show the median growth temperatures and median occurrence temperatures. The grey shaded regions indicated the growth niche widths and occurrence niche widths. For Trichodesmium erythraeum and Thalassiosira gravida, different growth curves (in black) correspond to different published experimental datasets, largely representing different strains. We overlay the envelope of the individual curves in blue and use this to calculate the indicated median growth temperature and growth niche width.

Mridul K. Thomas@mridulkthomas· Feb 17

A paper on temperature effects on fish, written by Bass Dye. Nominative determinism strikes again! (Interesting-looking paper too!)

Screenshot of paper titled "Disentangling the effects of resource level and temperature dependence on the performance of fish in different guilds" with a first author named Bass Dye.

Mridul K. Thomas@mridulkthomas· Sep 15

@raviranjan.bsky.social & I are teaching a free online workshop with on experimental design for environmental scientists on the 23rd. We'll focus on using simulations to evaluate how well different experimental designs help achieve your goals. Please sign up & share! forms.gle/MZTxeQs4UpMr...

Poster for 90 minute workshop on HOW TO DESIGN BETTER EXPERIMENTS, by Mridul Thomas & Ravi Ranjan.

Details: September 23rd 16:00 CEST ; 14:00 UTC

Description:

Experimental designs can make or break an experiment. A good experiment has a clear goal and efficiently uses experimental resources to achieve that goal. In this workshop, we will review what experiments are for, basic and advanced principles of designing experiments, and how to use simulations to evaluate designs before actually doing the
experiment. We’ll do a moderate amount of coding in R and so experience with this would be helpful but is not required. We intend to have small-group discussions to help participants develop their own experiments, and encourage participants to think of a specific question they would like to answer with an experiment.

Mridul K. Thomas@mridulkthomas· Sep 8

Lunar eclipse over the waters

Mridul K. Thomas@mridulkthomas· Aug 4

My first Amanita muscaria! Spotted this weekend by @joelyndelima.bsky.social

Mridul K. Thomas@mridulkthomas· Jul 18

I will never understand how this became so influential. Instead of using either evidence-based teaching or tried-and-tested methods (both defensible), we somehow ended up with fad-based education.

https://en.m.wikipedia.org/wiki/Whole_language

Whole language is a philosophy of reading and a discredited[8] educational method originally developed for teaching literacy in English to young children. The method became a major model for education in the United States, Canada, New Zealand, and the UK in the 1980s and 1990s,[7] despite there being no scientific support for the method's effectiveness.[9] It is based on the premise that learning to read English comes naturally to humans, especially young children, in the same way that learning to speak develops naturally.[10][11][12]

Mridul K. Thomas@mridulkthomas· Apr 7

Rather pleased that my posts on stats.stackexchange.com have been viewed a million times. I've learnt so much from the community there, it's nice to have helped out as well.

Screenshot showing "Impact: ~1 million people reached"

Mridul K. Thomas@mridulkthomas· Apr 3

Hell of a day for a workshop on careers in finance.