Blog 🎤

The green whistle test

I have some lived-experience to report and a brief reflection on an RCT. Content warning: it involves a brief description of an injury, but I lived happily ever after!

I recently dislocated my shoulder after slipping on a traverse climbing wall – ten minutes before the end of an introductory bouldering session. Thankfully, I wasn’t far off the ground. As far as these things go, it was a “good” dislocation: I feel fine and am starting physio imminently.

It wasn’t my first. Over a decade ago, I had a dislocation and was given nitrous oxide (gas and air) at A&E while they reattached my arm. This time, I was handed Penthrox – also known as the green whistle – a vape-sized handheld device that delivers methoxyflurane.

Subjectively, Penthrox felt much better than gas and air. During the procedure, I’d describe my mental state as “relaxed, pain-free, and a bit messed up but in a good way.” Once my shoulder was back in place, I would say I was “elated and very chatty” (emphasis on the “very”: apologies to everyone else waiting for an x-ray with me).

In contrast, gas and air didn’t do as much for the pain, though it did have the “messed up” component, which was helpfully distracting – at least as I remember it. There were plenty of confounding variables: injury severity (previously I needed an ambulance, this time not) and memory effects among them. So I wondered what the evidence says about the average differences between the two.

I came across a systematic review (Porter et al., 2017). Both gas and air and Penthrox outperformed placebo; however, no statistically significant difference was found between them in direct comparisons. But hang on – someone ran a study comparing Penthrox with placebo? I’m not sure I’d have met the inclusion criteria, but imagining being invited to take part in a placebo-controlled trial while hanging onto my arm, I suspect my response would have involved expressive language and a firm decline.

The justification for the placebo-control was: “Use of an active comparator, although preferred, would have posed considerable challenges to keep the study blind because of the unique mode of delivery and smell of methoxyflurane” (Coffey et al., 2014, p. 614).

An interesting case example to discuss.

References

Coffey, F., Wright, J., Hartshorn, S., Hunt, P., Locker, T., Mirza, K., & Dissmann, P. (2014). STOP!: A randomised, double-blind, placebo-controlled study of the efficacy and safety of methoxyflurane for the treatment of acute pain. Emergency Medicine Journal, 31(8), 613–618.

Porter, K. M., Siddiqui, M. K., Sharma, I., Dickerson, S., & Eberhardt, A. (2017). Management of trauma pain in the emergency setting: Low-dose methoxyflurane or nitrous oxide? A systematic review and indirect treatment comparison. Journal of Pain Research, 11, 11–21.

LLM refactors code into infinite loop

Here’s a lovely example of an LLM going awry in a deployed software product, causing a service to slow to a crawl.

A section of code was “refactored” by an LLM – meaning it was restructured without intending to change its behaviour. The refactoring involved moving a block of code between files, but in the process the LLM made a subtle edit that did dramatically alter the code’s behaviour.

Originally, the code included a loop that checked for errors on each iteration. If an error occurred, the loop exited. In the refactored version, the loop instead continued after encountering an error, which meant that errors became infinite loops. This tiny change – in code, the word “break” becoming “continue” – was missed during human review, buried among other edits. But if a human developer had done the refactor, they probably would have copy-pasted the entire block and so preserved the original logic.

Here’s the software company’s blog post explaining what happened.

DAGs are not necessarily associated with causation

The bad old days: an association was found between X and Y; however, correlation does not imply causation. Check out this long list of covariates and how they were adjusted in a big ol’ regression model.

These enlightened times: X causes Y; here’s a causal DAG, which you critically appraise by determining which important confounding variables are absent; whether something that’s currently a confounder should, e.g., really be a mediator or vice versa; and whether the regression, weighting, or matching model reflects what you think is the true causal structure, having critically appraised the authors’ DAG.

The main benefit of causal DAGs is not in identifying confounders but rather not mistakenly conditioning on the wrong things. The hard work of non-experimental causal inference is in the topic theory. X causes Y in the DAG does not imply causation in the world without a solid tested theory.

Subverting randomisation

One of the reasons that people test covariate balance in trials is because they have a vague sense that they should check whether something has gone wrong with random assignment – but often without a clear idea of what that something is. One thing that could go wrong is that people involved in participant recruitment, such as clinicians, could be deliberately trying to subvert assignment so people receive the treatment clinicians think they need.

Brown et al. (2005) asked 25 clinicians and research nurses whether they tried to predict treatment allocation and, if they did, how, and why.

Reassuringly, most (21 out of 25) said they did not, citing that they were “aware that this is wrong”, wanted to “remain impartial to the treatments”, were “unable to, allocation is out of our hands”, or thought it would be too complicated to do so.

The 4 out of 25 who said that they did try to predict treatment allocation reported keeping a log of previous allocations:

“A few investigators stated that they hoped some patients would be randomized to specific treatments, giving such reasons as that patients would be able to travel more easily or that some would have a perceived benefit from a particular treatment.”

References

Brown, S., Thorpe, H., Hawkins, K., & Brown, J. (2005). Minimization—Reducing predictability for multi‐centre trials whilst retaining balance within centre. Statistics in Medicine, 24(24), 3715–3727.

Fei Wan (2025) on propensity score matching

Quasi-experimentalists will be familiar with King and Nielsen’s (2019) landmark paper, Why Propensity Scores Should Not Be Used for Matching. This specifically critiqued the use of propensity scores for matching; other common uses of propensity scores, such as inverse probability weighting, were not affected. An interesting paper has appeared by Fei Wan (2025) critiquing King and Nielsen’s findings.

Fei Wan’s targets include:

  1. The inappropriateness of using an average pairwise covariate distance between treated and closest comparison match to evaluate PSM. Distances are always nonnegative so can’t take account of positive and negative differences averaging out. It’s already known that two observations with the same propensity score are likely to have different covariate values, but it’s ok if they are random.
  2. King and Nielsen used a cherry-picking approach for model selection, trying 512 different specifications and selecting the one that yields the largest average treatment estimate. This is just poor analysis practice.

Fei Wan recommends machine learning approaches to estimate propensity scores, rather than the common use of logistic regression.

Lots more to digest in this…

References

King, G., & Nielsen, R. (2019). Why Propensity Scores Should Not Be Used for Matching. Political Analysis, 27(4), 435–454.

Wan, F. (2025). Propensity Score Matching: Should we use it in designing observational studies? BMC Medical Research Methodology, 25(1), 25.

The magic words are squeamish ossifrage

Interesting post on RSA, a type of encryption that allows the key used to encrypt messages to be shared publicly. It works thanks to the difficulty of factoring the product of large prime numbers. We rely on public key encryption approaches like RSA everytime we use a secure internet connection, for instance to share credit card details or check our GP record.

Martin Gardner posed a factorisation problem in 1977, which was solved in 1993 with the help of ~600 people and ~1600 computers. The solution was used to decrypt a message, which turned out to be the title of this post. In 2025, the problem was solved again in 47 minutes using 8 CPUs on Google Cloud.

Bigger numbers lead to more secure encryption; however, quantum computers threaten to destroy the whole approach, hence the development of post-quantum cryptography.

Translating folk psychology into neuroscience using an LLM

Something LLMs are good at. The prompt I used was: “Pretend that you are a famous clinical psychology researcher, such as [redacted], who uses neural models to describe psychological states and processes. Write in the style of an Oxford psychology professor. Translate the following folk psychology into the appropriate neuroscience description, making a table with the folk psychology on the left and translation on the right.”

Folk psychologyNeuroscience/neurobollocks
I had an ice cream and it was enjoyable.Consumption of the ice cream likely activated the mesolimbic dopaminergic pathway—particularly projections from the ventral tegmental area to the nucleus accumbens—facilitating reward anticipation and hedonic experience. Gustatory input would have engaged the anterior insula and orbitofrontal cortex, supporting sensory integration and affective appraisal.
I had an argument with someone and said something stupid because I was tired.The interpersonal stressor likely elicited heightened amygdala reactivity, increasing emotional salience and diminishing prefrontal inhibitory control. Fatigue is associated with reduced dorsolateral prefrontal cortex activity, impairing executive functions such as verbal regulation and impulse suppression. This neurofunctional compromise may have attenuated mentalisation capacity, resulting in dysregulated speech and diminished perspective-taking.
I snogged my partner and felt tingles down my spine.Intimate contact likely activated oxytocinergic and dopaminergic systems, enhancing affiliative bonding and reward sensitivity. Somatosensory stimulation along the spinal dermatomes may have engaged the posterior insula and periaqueductal grey, contributing to embodied affective resonance. Concurrent activation of the ventromedial prefrontal cortex and anterior cingulate cortex would support interoceptive awareness and emotional attunement.
I had a first date and found myself falling in love when we held eye contact and I realised how much we had in common.Prolonged mutual gaze likely engaged the social salience network, including the superior temporal sulcus and medial prefrontal cortex, facilitating affective attunement and mentalisation. Recognition of shared values or interests may have activated the default mode network, supporting autobiographical memory and self–other integration. Oxytocin release and ventral striatal activation would contribute to the emergent experience of romantic attachment.
I tried to write a report; however, found myself spelling things phonetically because I was tired.Fatigue-induced attenuation of prefrontal cortical activity—particularly within the left inferior frontal gyrus and dorsolateral prefrontal cortex—may have impaired orthographic retrieval and phonological encoding. This reduction in executive functioning and linguistic precision likely led to reliance on sublexical phonetic strategies, reflecting a shift from automatic to effortful processing under cognitive load.
I cycled at speed down a hill and felt so happy.Rapid cycling likely engaged vestibular and proprioceptive systems, stimulating the cerebellum and somatosensory cortex. The associated thrill may have activated the locus coeruleus–noradrenergic system, enhancing arousal, while dopaminergic signalling within the nucleus accumbens supported the experience of euphoria. Simultaneous suppression of the default mode network may have facilitated a state of embodied flow and affective elevation.
I was dancing in a club to uplifting house music and loved the experience despite being completely sober.Rhythmic entrainment and auditory stimulation likely activated the basal ganglia and auditory cortex, while motor synchrony engaged the supplementary motor area and cerebellum. The affective uplift, in the absence of pharmacological modulation, may reflect endogenous dopaminergic activity and increased connectivity between the limbic system and prefrontal regions. Group synchrony and embodied movement may have enhanced oxytocin release, supporting social cohesion and emotional resonance.
A complex R function I wrote ran first time and yielded the result I expected.Successful execution of a cognitively demanding task likely activated the frontoparietal control network, particularly the dorsolateral prefrontal cortex and intraparietal sulcus, supporting working memory and abstract reasoning. The match between predicted and actual outcomes may have engaged the anterior cingulate cortex and ventral striatum, reinforcing prediction error minimisation and eliciting reward-related dopaminergic signalling.