iiinova — where your digital self keeps evolving

Dan Lewer @danlewer.bsky.social · 28/09/2026

In OLS you probably get the same point estimate (is that what FWL says??) but worse CI coverage due to ignoring the weather variance. In other types of regression I guess the point estimate might be different because you have a different estimand (effect on "deweathered" pollution)

000

Dan Lewer @danlewer.bsky.social · 28/09/2026

I think we're only interested in the effect of the intervention at the time of the "interruption". The other variables aren't interpretable in this analysis (table 2 fallacy). So the question is just if you get a better effect estimate by "deweathering" the pollution or modelling weather as normal

100

Dan Lewer @danlewer.bsky.social · 25/09/2026

Debating with colleagues about the practice of "deseasoning and deweathering" air pollution data before use in time series analysis. I can't see why you don't use the raw pollution data and model season/weather variables like normal confounders (independent variables). Can anyone defend this?

162

Reposted by Dan Lewer

January Weiner @january3.github.io · 23/09/2026

(I finally re-did it in #typst, though)

2111

Dan Lewer @danlewer.bsky.social · 21/09/2026

Our new Born in Bradford study suggests that teenagers of Pakistani heritage in Bradford have better mental health and wellbeing than teenagers of white British ethnicity bmjpublichealth.bmj.com/content/4/3/... @bihr.bsky.social

000

Dan Lewer @danlewer.bsky.social · 17/09/2026

"I made this with the aid of chatGPT"

000

Reposted by Dan Lewer

Adam Holland @holla-adam.bsky.social · 16/09/2026

Today we’re launching a series of films - Drugs: It’s Complicated - intended to tackle the stigma directed towards people who use criminalised drugs www.youtube.com/watch?v=2NAj...

youtube.com

Reimagining Drug Use (Drugs: It's Complicated - Film 2)

YouTube video by Drugs: It's Complicated

154

Reposted by Dan Lewer

Jack Wilkinson @jdwilko.bsky.social · 14/09/2026

This is from the same journal that brought us ‘add p-vals to Table 1, and write what you would have done had any been significant’

131

Dan Lewer @danlewer.bsky.social · 10/09/2026

- let me summarize your email for you

000

Dan Lewer @danlewer.bsky.social · 10/09/2026

You might have a modest and highly caveated causal claim but it's still a causal claim and might be quite useful

000

Dan Lewer @danlewer.bsky.social · 10/09/2026

A failure in causal inference is the idea that it needs sophisticated methods. This has led researchers to say "it's not causal, we're just doing simple exploratory work". You don't need G computation to be interested in the effects of exposures and interventions. Even common sense is valid evidence

131

Reposted by Dan Lewer

Per Engzell @pengzell.bsky.social · 28/11/2025

24. Foo Fighters: www.rki.de/DE/Aktuelles...

1142

Reposted by Dan Lewer

Per Engzell @pengzell.bsky.social · 28/11/2025

20. Rod Stewart: www.ajconline.org/article/s000...

151

Dan Lewer @danlewer.bsky.social · 22/08/2026

Old coding workflow: (1) Solve a problem New workflow: (1) Immediately ask AI (2) AI suggests 3 obscure solutions, which I half heartedly try implementing while getting AI to help debug (3) I think up a fairly obvious solution and for no reason suggest it to AI (4) AI declares me a genius

011

Dan Lewer @danlewer.bsky.social · 22/08/2026

When doing simple operations in R, always prefer simple functions #rstats

130

Dan Lewer @danlewer.bsky.social · 20/08/2026

Meetingitis (rather than think about something, let's set up a meeting) is now extended to chatbotitis

000

Dan Lewer @danlewer.bsky.social · 14/08/2026

Why do we dichotomise things? We all deeply want to know if we are normal or deviant

000

Dan Lewer @danlewer.bsky.social · 12/08/2026

Had this a couple of times - I like to declare a COI

000

Dan Lewer @danlewer.bsky.social · 12/08/2026

If there is no statistical evidence of an association between Schoenfeld residuals and time, this does not suggest the proportional hazards assumption is met. It doesn't really tell you anything!

000

Dan Lewer @danlewer.bsky.social · 07/08/2026

Can we get a UX engineer involved in the clinical trial governance process please

001

Dan Lewer @danlewer.bsky.social · 06/08/2026

They even ask you to do the copy editing yourself, point by point!

000

Dan Lewer @danlewer.bsky.social · 23/07/2026

As much as a problem, it's a defence against the urge to change things. These venerable institutions can survive endless pointless strategies and change programmes, but occasionally fail when one of them somehow succeeds

000

Dan Lewer @danlewer.bsky.social · 22/07/2026

Yes of course. I just had my plain English hat on when I posted that. I wear a lot of hats

100

Dan Lewer @danlewer.bsky.social · 11/07/2026

I met Ann Widdicombe when I was an undergraduate at LSE maybe in 2002. She gave a lecture, I think on the topic of "Europe", and for some reason only four people turned up. She approached it no differently to a hundred people and just gave more time for questions

000

Dan Lewer @danlewer.bsky.social · 10/07/2026

Completely agree Adam. When hiring people, you now have to think "is this person going to copy stuff from chatGPT or actually do some thinking?" In the former case, you end up doing the work/thinking yourself anyway

120

Dan Lewer @danlewer.bsky.social · 09/07/2026

If you try to describe research in plain English with very short sentences and only the most basic words, you can make it more difficult to understand. The reader is probably an interested and intelligent but not expert, rather than a toddler or an alien who doesn't understand terms like "the NHS"

010

Dan Lewer @danlewer.bsky.social · 01/07/2026

TBF Pilates probably does help a bit tho?

000

Dan Lewer @danlewer.bsky.social · 01/07/2026

You can also reduce your adjusted future risk of heart attack by having a heart attack

110

Dan Lewer @danlewer.bsky.social · 01/07/2026

I've often thought that a good way to improve your chances of survival is to take part in a health survey 10 years in the future

110

Dan Lewer @danlewer.bsky.social · 29/06/2026

Thank you for this great blog! I think "descriptive" means giving average measures or distributions of an outcome in one group. If you compare groups (defined by an exposure) in terms of an outcome it's analytical regardless of the method, adjusted or not. Maybe that's a bit extreme. Unsure!

020

Dan Lewer @danlewer.bsky.social · 29/06/2026

The idea that everyone is a leader is completely mad

010

Dan Lewer @danlewer.bsky.social · 28/06/2026

```N <- 1000 time <- 500 x <- matrix(rbinom(N*time, size = 1, prob = 0.01), ncol = time) zero <- rowSums(x) == 0 x <- max.col(x, ties.method = 'first') x[zero] <- NA x <- table(factor(x, levels = 1:time)) plot(cumsum(x), type = 'l'))```

000

Dan Lewer @danlewer.bsky.social · 28/06/2026

If events happen at random, the time until the first event follows a geometric distribution. This an be understood intuitively by simulating random events with a fixed probability across a population (R code below) #RStats

120

Dan Lewer @danlewer.bsky.social · 17/06/2026

"Our primary outcome was XXX scale. Our secondary outcome/sensitivity analysis was a different version of the same scale (eg. dichotomised if the primary outcome is continuous, or vice versa" <- please don't do this!

010

Reposted by Dan Lewer

Anna Dixon MP for Shipley @annadixonmp.bsky.social · 11/06/2026

Reform’s plan to effectively scrap Bradford’s Clean Air Zone is premature and reckless. Evidence shows the CAZ has improved air quality and cut GP appointments for breathing issues by 600 per month. Scrapping it now would be a public health mistake our children's generation will pay for.

1112

Dan Lewer @danlewer.bsky.social · 08/06/2026

I'm doing a bit of light p-hacking in preparation for an NIHR seminar on natural experiments - www.nihr.ac.uk/news/nihr-an... - I asked chatGPT if it could help me do more thorough p-hacking:

120

Dan Lewer @danlewer.bsky.social · 08/06/2026

I say "The D-P-L-Y-E-R package" to sound old

010

Dan Lewer @danlewer.bsky.social · 04/06/2026

I think the review contains the full analysis!

010