Nick Holmes

@thehandlab

1059 Siguiendo1954 Seguidores1393 Publicaciones

Website Bluesky

¿Es tu perfil?Reclámalo y thehandlab.dirty.social es tuyo.Reclamar perfil

Nick Holmes@thehandlab· Jul 12

That's world class #DataVis

BBC graph of year (x) and number of World Cup teams (y) - teams are increasing

How bad is this graph? Let me count the ways...
X & Y have no title
X labels all squished
Y axis should be 0-8-16-24-32-40-48-56-64
Legend is redundant
2026 is not 'proposed' - it happened!
Numbers below datapoints??!

At least 6.

Nick Holmes@thehandlab· Jun 29

I should say that hedgehog shits are usually much darker, narrower, and beetly than this (Figure 1). I guess he was makimg a point.

Figure 1. A more typical hedgehog shit: dark, cylindrical, 2-3cm long, in the green grass, glistening with worm slime and beetle skin. Textbook.

Nick Holmes@thehandlab· Jun 29

Humans: don't shit where you eat Hedgehog: watch me

A hedgehog shit, next to a bowl where a hedgehot eats.

Nick Holmes@thehandlab· Jun 28

This scarlet tiger #moth is around at the moment. And what a beautiful bimbo he is 🤩 There has been a real increase of flying insects in our garden this year, likely due to ~5 years of leaving lots of wilderness, plus good conditions. Party time for bats, swifts & swallows!

Nick Holmes@thehandlab· Jun 28

Garden angels 😇😇 #birds

Two Eurasian blackbirds in flight, caught by a trail camera. Female left, male right. The morning sunlight behind the birds makes them look like angels.

Nick Holmes@thehandlab· Jun 22

This is the reality of #ClimateChange - these were just innocent babies going about their business, trying to get from A to B (in my glove box) and they've just... melted. #HeatWave

Jelly babies all melted together in a pink box. Carnage.

Nick Holmes@thehandlab· Jun 20

#ShowYourStripes

The average yearly temperature in Oxford UK, blues showing cooler and reds warmer than average.

Nick Holmes@thehandlab· Jun 19

"Figure 1 was created with AI" 'AI' still can't do hands 😊 Now read the linked commentary below about fixing science, which is great 👏

Part of an 'AI' image, showing a tangle of parts of hands in front of men in grey suits.

Nick Holmes@thehandlab· Jun 15

IT IS DONE @sean-h.bsky.social Needs some tweaking - eg add the subgroup Means in text. But I'm pretty happy with this :-) time to get back to the actual meta-analysis now I think...

A Contour-Enhanced Forest Plot With Two Subgroups. Perhaps "For The First Time". The top five studies are in red, the bottom 5 in blue, each subset has its own shaded bar showing its own random effects mean and 99% CI; the overall mean is shown at the bottom with its traditional diamond. This is the Future of Forest Plots.

Nick Holmes@thehandlab· Jun 12

In better news, we also had a deer this morning! Here it is next to the pond. (Blackbirds are so common they practically live in our house with us, but are also a good news story.)

A muntjac deer, left, and a female blackbird, right, outside our back door (off camera, right)

Nick Holmes@thehandlab· Jun 12

Caring for your local wildlife is hard. This hedgehog appeared at 10:30 each night to eat outside my door. I had watched it eat its last meal as this picture was taken. This hedgehog was killed on the road near our house yesterday. Please drive carefully at night. #hedgehog #wildlife #extinction

Nick Holmes@thehandlab· Jun 11

how's that @sean-h.bsky.social ? Feels like we are developing a 'contour-enhanced' Forest Plot here?

A traditional forest plot - all greyscale - but with the mean+CI effect size extended as a shaded patch running from top to bottom, and the mean as a dotted line. looks alright..?

Nick Holmes@thehandlab· Jun 10

OK, here's why I was doing all this... This next one is a real #meta-analysis, in which I've colour-coded study quality (redder: worse, bluer: better). Quality is points-based, between -3 and +3 (e.g., no blinding: -1, double-blind: +1; placebo control: +1; counterbalanced: +1, etc) Any good..? 🙏

this time it's a real meta-analysis. As before, except the data are real, and studies are colour-coded according to their quality, and ordered by quality from worse (top) to best (bottom).

Nick Holmes@thehandlab· Jun 10

Completing my evening's work: now with user-defined colours, symbols, AND linestyles per study. I think we can all agree this is the most beautiful and accessible Forest Plot that's ever been created. The figure legend stretches to three pages.

same as the image in the previous post, but now each study has it's own random colour, with different symbol shapes and line styles - who doesn't need *that*!. the meta-analytic mean remains the traditional diamond.

Nick Holmes@thehandlab· Jun 9

An evening with Matlab well spent: made a function to create forest plots for #meta-analysis. no vibe, just coding. This function is the definitive answer to the question: "Why can't I specify random colours for my studies, eh?"

a 'forest plot' for meta-analysis, from a custom Matlab function, with 10 made-up studies with random means and SEs. The major development here is that each study can be independently coloured (the example here just uses random numbers for the RGB triplets.

Nick Holmes@thehandlab· Jun 4

My email rubbish bin, with an email from OSF every ~3-5 minutes for the last few hours 🤷‍♂️

Nick Holmes@thehandlab· Jun 3

wondering what is an acceptable level of "overlap" between the Methods sections of two papers on a similar topic published five years apart by two different research groups 🤔 #TMS #BrainStim pubpeer.com/publications...

This image and next two show two columns of text which has been broken up into sections. The two columns are from two different papers. The text is highlighted to show near-identical passages of text. More than half of the Methods sections of the two articles appears to be identical. Lots of critical details are identical, for example the age of participants, screening tests performed, equipment used. This level of similarity is... unlikely.

Nick Holmes@thehandlab· May 10

My LinkedIn impressions are up 289100% on last week and no that is not a typo. Peaked in time for my annual review 😎 #KillingTheSocials

A staggeringly impressive graph of Dr Holmes' LinkedIn impressions this week, rising from 0 on May 4th, to an internet-breaking 2892 impressions by May 10th. LinkedIn helpfully converted this to a percentage - presumably rounding the initial 0 up to 1 - and awards me with a 289100% increase week-on-week. Next week I fully expect to continue this stratospheric achievement and receive 8363644 impressions. Your move, losers.

Nick Holmes@thehandlab· Apr 28

You are right! Following J&B's description to the letter - I can now reproduce their graphs. I did both their simple (a/b) & compound (a/b)/(c/d) ratios, & repeated it with correlated data (r~0.7) - more sensible for all neuro/bio fields that I know of. Correlated CIs are narrower. many thanks!

Nick Holmes@thehandlab· Apr 27

Their figure looks nice & lawful, and their conclusion was very strong: DON'T USE RATIOS, as they require "substantial magnitudes" or "unrealistic precision" or "prohibitively large sample sizes" eg, with N=20, SD=50%, mean ratios<0.7 will be significant BUT: I could not reproduce their results 👇

My reproduced version of their figure is substantially different. I did not 'smooth' the data as the original authors did. The overall story is the same, but the consequences are much less-severe than they claimed - after sample sizes of just N=10, for example, ratios <0.82 will be significant. It seems the original paper got something wrong..?

Nick Holmes@thehandlab· Apr 27

J&B simulated 2 (or 4) sets of normal, independent random numbers: a, b (c, d). A simple ratio was calculated: a/b, then the mean ratios across different sample sizes were calculated. The aim was to construct null distributions for testing ratios against 1. Great. Their Figure 1 plots the 95% CIs:

95% confidence intervals for 6 levels of variability (SD/CV) against sample size (y-axis), for testing whether a mean ratio is different from 1. The x-axis is in two parts - on the left below 1 is linear spacing of 0.2, on the right, above 1 is linear spacing of 1.0. As with all such power analyses, as N increases, the effect size (ratio difference from 1.0) needs to be smaller to be significant.

Nick Holmes@thehandlab· Apr 26

Now *that* is a large sample size #MetaAnalysis of #TMS #BrainStim #electromyography #EMG #MEP #preprint will be coming: one day this season...

A scatter graph of group means (x-axis) and group SDs (y-axis) for the primary outcome variable in an massive meta-analysis of short-latency afferent inhibition (SAI) data. The x- and y-axes are on logarithmic scales, and a cloud of 11,404 datapoints stretches across the diagonal from lower left to upper right.

Nick Holmes@thehandlab· Apr 20

Lammy & Reeves putting in Oscar-winning performances here: youtu.be/h0elB6_pvrY?...

David Lammy (left) looks over at Rachel Reeves (right), both doing a remarkable gong-worthy job of suppressing all human emotion while Keir Starmer (centre) bores the country with talk of "process"

Nick Holmes@thehandlab· Apr 15

Confirmed with @ecosiasearch.bsky.social 'World's greenest AI' (so hopefully my 4th-ever AI search was not too wasteful. "Ask anything" it solicits. (But not maths) What a f*cking disgrace 😔.

Ecosia says: "Ask the world's greenedt AI, backed by renewable energy. Respects your privacy." Fair enough. But can it calculate?

Q: "What is 493920 times 392930?"

A: "The product of 493,320 and 392,930 is:

493, 920 x 392, 930 = 194, 106, 305, 600

So, 493,920 times 392,930 equals
194,107,305,600."

It does not.

Nick Holmes@thehandlab· Mar 23

I was given the valuable opportunity of re-taking this bullshit #AI survey promoted by my employer via Canvas. It remains unchanged after my bilious outrage last week: Exhibit A 👇 (this is the first of many dire questions) (my critical analysis in the ALT text)

Canvas 'AI' MCQ question 6:

"Which of the following statements best describes your *understanding of AI?*

A. I have little or no understanding.

B. I have a basic understanding of AI, data, and common AI tools, and can interact with AI at a basic level.

C. I can use various AI tools effectively in my teaching, writing good prompts, understanding their limitations, and critically evaluating outputs.

D. I have a strong technical understanding of AI and can customise, fine-tune, and integrate AI tools into my teaching workflows.

Oops! Please make a selection"

You *have* to choose one of these 4 options. You cannot progress if you don't.

Critique:

Two questions are being conflated here (understanding, and use).

Analogy:

Q. Do you like cheese?

1. I hate cheese.

2. I like Cheddar and Gruyere, and am able to grate parmesan at a rate of 20-30 grams per minute, so at the regional-championship level (UK Dairy League).

3. I like a range of English and Swiss cheeses, and spent a fine summer in a Swiss château vigorously pounding the cud with that famous cow who can use brooms to scratch its back. You know the one. Veronika.

4. I will eat and enjoy any cheese you put in front of me. I've worked in the dairy industry for most of my adult life. I married Veronika.

Nick Holmes@thehandlab· Mar 19

For just $100, you can buy Desk Rejection Shield! Makes your shiter-than-shite manuscript Rejection Proof!* 0% satisfaction guaranteed! * Ts & Cs apply

Nick Holmes@thehandlab· Mar 14

So far, I have found six papers that use the "95% CI" method to offer advice on how many trials are needed. The graph below plots the R trials per subject, & the recommended, r, trials to reach the 95% CI. The correlation is strong 💪 and the slope is 0.41...

correlation between the number of trials used, R, and the recommended trials required, r, across 6 independent studies using the 95% CI method. The r-value is 0.98, r-squared 0.965, and equation r = 0.408R + 11.5

Nick Holmes@thehandlab· Mar 14

Summarising across 5 repeat sizes and 5 sample sizes: The number of repeats, r, required to be within the 95% CI of the full set of trials, R, is linearly dependent on R. The slope is about 0.4. There will be a nice mathematical formula for this, and it will be in Amman et al (2020):

X-axis shows R trials per subject, from 8 to 128.
Y-axis shows r trials for 90% of subject means to be within their 95% CI
Five separate plots show the effect of different sample sizes, from N=8 to N=128.
All 5 plots are close to straight lines, with slope ~ 0.4

Nick Holmes@thehandlab· Mar 14

And, just for fun, an unusually-large study with 128 trials per partcipant. Doing this study, you would conclude that ~40 to 80 trials are needed 😱

Nick Holmes@thehandlab· Mar 14

And here's a more typical experiment, with 32 trials per subject. Now you need 10-15 trials to get within the original 95%CI!