@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...

Lunar eclipse over the waters

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

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.

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.

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

I had no idea that there were nutria/coypu in Switzerland (or Europe) till I stumbled on this one last week.

Didn't know much about them, but they are bigger than I thought, have large appetites, and can cause big changes to ecosystems.

Livening up my stats lectures with lots of help from GPT #rstats

To mathematically optimise an experiment, we need to have a candidate function that describes how the response changes across interacting drivers. For most driver sets, we don't have this!

So step 1 is to do full factorial designs (~ 5 x 5 or better) to develop these functions.

(11/n)

We show that optimal & sequential designs are better at minimising prediction error.

In these simulations, they are a little better than a 5 x 5 full factorial design. But more importantly, they are basically as good even when using *half* the sample size (not shown).

(10/n)

These designs look different because when dealing with nonlinear functions (and complex interactions), not all points on the surface have the same *leverage*. For the function illustrated on the left, leverage is highest for points at low nutrients and high temperatures.

(9/n)

We can optimise our experiments mathematically if we have a clear objective.

@raviranjan.bsky.social dug deep into the statistical and engineering literature to review the theory of optimal and sequential design.

These designs look very different from the ones we are used to!

(8/n)

Regression designs are superior but even with them, there are many choices. We review them all and recommend when to use which design. Their strengths and weaknesses are summarised below:

(6/n)

Before for even a simple multiple-driver experiment, there are more ways to design it that we often think about in ecology - and there's a lot of scope for improvement.

Each point here represents a treatment, and point size indicates the number of experimental units at that treatment.

(4/n)

"Experiments should be tied to mathematical theory/models" is not a new observation. But where do we go from that?

Here are the kinds of experiments that tend to be most useful in informing models:

(3/n)

Experiments are hard work. If we think carefully about our scientific goals and how experiments relate to them, we can squeeze a lot more out of them. And sometimes with less effort than we presently put in!

(2/n)

Our group in Geneva has a 2-year postdoc position to classify phytoplankton from images and analyse time series. We are specifically looking for experience with machine learning / image classification, not necessarily in ecology/phytoplankton.

Please share!