Sign in

Lachlan Cribb

@lachcribb.bsky.social
47 followers 160 following 24 posts

PhD student working on causal inference methods & dementia at Monash University. lachlancribb.netlify.app

PostsRepliesMedia
Lachlan Cribb @lachcribb.bsky.social · 20/08/2026
nix (e.g. via rix) is the answer. Easier, faster, and more reliable than renv + rig or whatever else. I will never go back
000
Lachlan Cribb @lachcribb.bsky.social · 11/08/2026
Great stuff! Just FYI, there are new FDA approved treatments for Alzheimer's disease (though they are risky and not very effective). E.g. www.nejm.org/doi/full/10.... Your argument about incentives is convincing. But why should the task of drug development not simply be absorbed into the state?
nejm.org
Lecanemab in Early Alzheimer’s Disease | NEJM
The accumulation of soluble and insoluble aggregated amyloid-beta (Aβ) may initiate or potentiate pathologic processes in Alzheimer’s disease. Lecanemab, a humanized IgG1 monoclonal antibody that b...
021
Lachlan Cribb @lachcribb.bsky.social · 02/07/2026
Besides the last statement, none of this seems unreasonable? A richer dataset will make it easier to adjust for confounding, all else equal. And most PH research probably does suffer more from underadjustment than from overfitting.
100
Lachlan Cribb @lachcribb.bsky.social · 21/06/2026
What kind of loss function do you have in mind?
100
Lachlan Cribb @lachcribb.bsky.social · 17/06/2026
"Patients and investigators prefer measures of absolute risk in subgroups for pragmatic randomized trials" Very few papers have I recommended more.
140
Lachlan Cribb @lachcribb.bsky.social · 11/03/2026
Can anyone recommend tools for collaborating on latex files?? Looking for something friendly to ppl with little tech background (no git) and with capacity for adding comments. There's overleaf but it's expensive and limits the number of collaborators
000
Lachlan Cribb @lachcribb.bsky.social · 04/03/2026
Great stuff! Are you finding that julia works well with nix? There were issues I remember related to some julia libraries (e.g Plots) on NixOS. I can imagine that using T with a pipeline based on julia and R nodes would be super useful
100
Lachlan Cribb @lachcribb.bsky.social · 01/03/2026
In fact not just IMO* pubmed.ncbi.nlm.nih.gov/29966732/
100
Lachlan Cribb @lachcribb.bsky.social · 01/03/2026
The PP effect is the most relevant estimand for those patients who do expect to adhere, IMO. Plus, the ITT effect in the trial population (with unusually high adherence) will differ from the ITT effect in routine practice, or in other pops with different adherence patterns, unlike the PP effect
110
Lachlan Cribb @lachcribb.bsky.social · 26/02/2026
Certainly for *naive* per protocol analyses, but well conducted per protocol analyses are still useful. Can overcome some issues with ITT effects (e.g. relevance, adherence dependence, generalisability). Identification does require assumptions, but same goes for ITT effects in realistic trials!
130
Lachlan Cribb @lachcribb.bsky.social · 26/02/2026
It's also great having all your important configs managed declaratively in one place and committed to git. Any changes you make are easy to track and undo. Plus, because your whole set up is managed by a config file(s), you can reproduce it exactly across any number of machines
110
Lachlan Cribb @lachcribb.bsky.social · 25/02/2026
Big appeal is that it solves all the issues you describe above! Because the whole system is managed declaratively, if everything gets borked, just wind back to a previous build. And if you want latest software, easy, set nixpkgs to unstable. Or use unstable for neovim and leave others at stable
110
Lachlan Cribb @lachcribb.bsky.social · 25/02/2026
Time for NixOS??
100
Lachlan Cribb @lachcribb.bsky.social · 28/01/2026
You would rather someone share code without a targets pipeline rather than with one?? Why?
110
Lachlan Cribb @lachcribb.bsky.social · 28/01/2026
I don't use rix but use nix directly for project environments. I've got to say it's incredibly convenient, even besides the reproducibility benefit. The ability to drop into any project with any set of dependencies and it just works, no matter how old. Incredible. Worth some initial pain imo!
050
Lachlan Cribb @lachcribb.bsky.social · 06/01/2026
Really? I can think of some common use cases. For example, if you want to use biomarkers to predict disease but are constrained in what you can measure in practice, you may want to know which subset of biomarkers is the most "important" (e.g. in terms of R^2, PPV, etc)
000
Lachlan Cribb @lachcribb.bsky.social · 23/12/2025
Also tar_map_rep for simulation studies/bootstrapping. The parallelization with crew is so fast and clean. Will Landau is making such good software
000
Lachlan Cribb @lachcribb.bsky.social · 16/12/2025
Other end of the spectrum: www.nytimes.com/2025/12/16/o...
000
Lachlan Cribb @lachcribb.bsky.social · 05/12/2025
I was wondering the same. This looks promising cran.r-project.org/web/packages...
cran.r-project.org
GPUmatrix: Basic Linear Algebra with GPU
GPUs are great resources for data analysis, especially in statistics and linear algebra. Unfortunately, very few packages connect R to the GPU, and none of them are transparent enough to run the compu...
010
Reposted by Lachlan Cribb
ViCBiostat @vicbiostat.bsky.social · 12/11/2025
Mark your diaries! The ViCBiostat Summer School returns from 13-20 Feb 2026, in Melbourne and online. Courses include causal inference, cluster randomised trials, meta-analysis and the estimand framework. Further details TBA shortly - sign up to our mailing list at www.vicbiostat.org.au #statistics
01510
Lachlan Cribb @lachcribb.bsky.social · 13/08/2025
Science people!! What software are you using to draft papers (besides MS Word 🤮)? Looking for something that 1. Allows for easy collab (with non-tech ppl, so latex is out) 2. In format accepted by journals (so no typst?) 3. Has reference manager integration Is it just googledocs/libreoffice?
010
Lachlan Cribb @lachcribb.bsky.social · 24/04/2025
Just to note, shared folders on google drive etc (anything that syncs changes) can be a real nightmare if you do end up collaborating via git & github
120
Lachlan Cribb @lachcribb.bsky.social · 09/03/2025
Frequentist alternative to Bayesian model averaging: Incorporate all 6 specifications into a superlearner stack, letting the data dictate how much weight each model receives in the ensemble. (Might need a doubly robust estimator/bootstrapping for valid inference). cran.r-project.org/web/packages...
cran.r-project.org
Guide to SuperLearner
091
Lachlan Cribb @lachcribb.bsky.social · 04/03/2025
That's great! With MSM's, is it straightforward to incorporate the estimation of the weights and the weighting of the regression model (Y on A*) with those estimated weights?
120
Lachlan Cribb @lachcribb.bsky.social · 02/02/2025
New paper!! We find that prevalent exposure designs are very common in dementia research, explain why they are ill-suited for causal inference, and show how the target trial framework helps us move beyond them. Read here👇 www.thelancet.com/journals/lan...
thelancet.com
Moving beyond the prevalent exposure design for causal inference in dementia research
As randomised trials are not always feasible or practical, observational studies remain crucial for addressing many causal questions in the dementia prevention field. Through a systematic search, we f...
063