Mridul K. Thomas

@mridulkthomas.bsky.social

2697 Followers 688 Following

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

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

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

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

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

"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."

"... 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."

"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."

"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. "

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

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]

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 the function selection, parameter entry, model fit visualisation, and MCMC chains.

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

Screenshot of Shiny app linked to in the quoted thread.

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.

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.

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

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.

Japanese rural population till 2024.

Source: UN World Urbanization Prospects 2025

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

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.

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.

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.

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.

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.

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]

Screenshot showing "Impact: ~1 million people reached"

Ad for a workshop on Finance & Banking Careers for Scientific Profiles

Animation showing the numerical optimisation process in a nonlinear regression fit. The example shows simulated Michaelis-Menten kinetics.

Why understanding functional forms is important for understanding and prediction

The out-of-sample prediction error from 100 simulated experiments using each of the designs. Lower RMSE (root mean squared error) is better.

(a) An example response surface for our temperature–nutrient interaction model and (b) the leverage of points across this surface, which indicates their relative influence on a final model fit. The leverage plot in panel b shows the natural logarithm of the tangent plane leverage, which is calculated using a linear approximation of the nonlinear model plotted in panel a. High temperatures and low nutrient treatments have high leverage, meaning that experimental points placed in those regions will have a large impact on the fitted surface.

A brief introduction to the maths of optimal design

A qualitative assessment of different experimental designs’ ability to meet our needs

Illustration of 8 different experimental designs for a 2-driver experiment.

1. Experiments that characterize the shape of the relationship between drivers and a response (saturating, exponential, etc.) to identify how best to describe this relationship mathematically...
2. Experiments that provide reliable parameter estimates with associated uncertainties...

$Abstract: For decades, multiple-driver/stressor research has examined interactions among drivers that will undergo large changes in the future: temperature, pH, nutrients, oxygen, pathogens, and more. However, the most commonly used experimental designs — present-versus-future and ANOVA — fail to contribute to general understanding or predictive power. Linking experimental design to process-based mathematical models would help us predict how ecosystems will behave in novel environmental conditions. We review a range of experimental designs and assess the best experimental path toward a predictive ecology. Full factorial response surface, fractional factorial, quadratic response surface, custom, space-filling, and especially optimal and sequential/adaptive designs can help us achieve more valuable scientific goals. Experiments using these designs are [...] our most promising path toward linking experiments and theory in multiple-driver research and making accurate, useful predictions.$