Robin Blythe
@rbly.bsky.social
Assistant professor at Duke-NUS medical school. Mostly interested in health economics, biostats and clinical informatics.
Nice thread over on the DataMethods discussion forum (an invaluable resource for me, personally!) on differentiating between conditional and average treatment effects
discourse.datamethods.org/t/propensity...
#statsky #episky
discourse.datamethods.org/t/propensity...
#statsky #episky
Propensity score weights
Question for @ngreifer Considering ATE, ATT, ATO and ATM weights, it is commonly said that a weighted estimator (ATE/ATT/ATO/ATM) targets a marginal estimand — an average over a specified population...
discourse.datamethods.org
November 26, 2025 at 2:38 AM
Nice thread over on the DataMethods discussion forum (an invaluable resource for me, personally!) on differentiating between conditional and average treatment effects
discourse.datamethods.org/t/propensity...
#statsky #episky
discourse.datamethods.org/t/propensity...
#statsky #episky
Reposted by Robin Blythe
IHEA members - the largest ever Board election in history is now open. I’m standing for the role of President Elect. More about why is here healtheconomics.org/paula-lorgelly/ #votepaula
Paula Lorgelly – International Health Economics Association
healtheconomics.org
November 25, 2025 at 5:14 AM
IHEA members - the largest ever Board election in history is now open. I’m standing for the role of President Elect. More about why is here healtheconomics.org/paula-lorgelly/ #votepaula
Just boosting this solution if you want to do some brms model bootstrapping. It's a very simple model but the whole thing comes together in just a couple minutes and it's very easy to code up
#rstats #statsky #bayes
#rstats #statsky #bayes
Basically, fit the model with chains = 0, then create an update() function with recompile = FALSE and the sampling specs you want. I've usually had success with {{furrr}} for the parallelisation part. Can't take full credit for this one as I needed an LLM to show me how the recompile step worked.
November 21, 2025 at 4:45 AM
I'm pretty sure this joke has been made (probably by @zachweinersmith.bsky.social over on SMBC) but I'm starting to wonder whether my inability to prove anything empirically using mathematics and relying on simulation every time is an education issue, a me issue, or both
#statsky #rstats
#statsky #rstats
November 20, 2025 at 4:06 AM
I'm pretty sure this joke has been made (probably by @zachweinersmith.bsky.social over on SMBC) but I'm starting to wonder whether my inability to prove anything empirically using mathematics and relying on simulation every time is an education issue, a me issue, or both
#statsky #rstats
#statsky #rstats
Reposted by Robin Blythe
My sister is having a rough time with her Master's thesis, which was graded by a rude external reviewer. Is it normal for a thesis to go out for a letter grade? This seems particularly cruel and uncommon given a thesis is a 200+ page research document. How can you give it a grade?
#academicsky
#academicsky
November 17, 2025 at 1:35 PM
My sister is having a rough time with her Master's thesis, which was graded by a rude external reviewer. Is it normal for a thesis to go out for a letter grade? This seems particularly cruel and uncommon given a thesis is a 200+ page research document. How can you give it a grade?
#academicsky
#academicsky
This has me wondering whether a perfect replication study can actually be bad - wasting time and resources on something we strongly suspect is effective when we could be using those funds to feel out the edges of that knowledge or doing something novel.
#metascience #statsky #medsky
#metascience #statsky #medsky
Right! And assuming that (i) what we can learn from a study is limited to what's captured by a p-value, and (ii) a close enough replication of that p-value is the only way science can validate existing evidence.
November 16, 2025 at 3:16 AM
This has me wondering whether a perfect replication study can actually be bad - wasting time and resources on something we strongly suspect is effective when we could be using those funds to feel out the edges of that knowledge or doing something novel.
#metascience #statsky #medsky
#metascience #statsky #medsky
New paper: we validated the Clinical Frailty Scale for predicting 90-day mortality in frail older adults. We show that the scale's predictive utility is low, but argue that mortality risk shouldn't be what catalyses end-of-life care discussions anyway
doi.org/10.1136/bmjo...
#statsky #medsky #episky
doi.org/10.1136/bmjo...
#statsky #medsky #episky
Validation of the Clinical Frailty Scale for predicting 90-day mortality in hospitalised older adults screened as at risk of nearing the end of life in Queensland, Australia: a multisite observational...
Background The Clinical Frailty Scale is an ordinal scale from 1 (very fit) to 9 (terminally ill) commonly used to assess frailty in older patients. It is simple for clinicians to apply and can help i...
doi.org
November 14, 2025 at 1:12 AM
New paper: we validated the Clinical Frailty Scale for predicting 90-day mortality in frail older adults. We show that the scale's predictive utility is low, but argue that mortality risk shouldn't be what catalyses end-of-life care discussions anyway
doi.org/10.1136/bmjo...
#statsky #medsky #episky
doi.org/10.1136/bmjo...
#statsky #medsky #episky
It would be nice if this forces people to recognise that the point of a degree is learning how to think deeply/problem solve, rather than get a highly paid job or put some letters after your name. Not everyone needs to go to university!
Relying on LLMs to get your diploma and is similar to forging PhD research for boosting your political career in Germany
You think it creates shortcuts and notability for getting the top job. But as you are only interested in the CV line, you don't learn anything.
And one day, you are exposed.
You think it creates shortcuts and notability for getting the top job. But as you are only interested in the CV line, you don't learn anything.
And one day, you are exposed.
November 10, 2025 at 9:19 AM
It would be nice if this forces people to recognise that the point of a degree is learning how to think deeply/problem solve, rather than get a highly paid job or put some letters after your name. Not everyone needs to go to university!
Got a strange analysis to do: I want to estimate how long prophylaxis can defer surgery. Problem is many Drs will just send straight to surgery, especially if patient is complex. If prophy is done, they'll check back every month or so until prophy fails, or pt ages out of the cohort #statsky #rstats
November 8, 2025 at 10:48 AM
Got a strange analysis to do: I want to estimate how long prophylaxis can defer surgery. Problem is many Drs will just send straight to surgery, especially if patient is complex. If prophy is done, they'll check back every month or so until prophy fails, or pt ages out of the cohort #statsky #rstats
Now up with a gift link to the full text:
authors.elsevier.com/a/1m2nJ38vD3...
We're hopeful that this paper gives payers leverage to insist on best practice guidelines that improve outcomes and lower costs, and gives patients more info about their treatment options.
authors.elsevier.com/a/1m2nJ38vD3...
We're hopeful that this paper gives payers leverage to insist on best practice guidelines that improve outcomes and lower costs, and gives patients more info about their treatment options.
November 6, 2025 at 2:46 AM
Now up with a gift link to the full text:
authors.elsevier.com/a/1m2nJ38vD3...
We're hopeful that this paper gives payers leverage to insist on best practice guidelines that improve outcomes and lower costs, and gives patients more info about their treatment options.
authors.elsevier.com/a/1m2nJ38vD3...
We're hopeful that this paper gives payers leverage to insist on best practice guidelines that improve outcomes and lower costs, and gives patients more info about their treatment options.
Reposted by Robin Blythe
🧠⚙️ Interested in decision theory+cogsci meets AI? Want to create methods for rigorously designing & evaluating human-AI workflows?
I'm recruiting PhDs to work on:
🎯 Stat foundations of multi-agent collaboration
🌫️ Model uncertainty & meta-cognition
🔎 Interpretability
💬 LLMs in behavioral science
I'm recruiting PhDs to work on:
🎯 Stat foundations of multi-agent collaboration
🌫️ Model uncertainty & meta-cognition
🔎 Interpretability
💬 LLMs in behavioral science
November 5, 2025 at 4:40 PM
🧠⚙️ Interested in decision theory+cogsci meets AI? Want to create methods for rigorously designing & evaluating human-AI workflows?
I'm recruiting PhDs to work on:
🎯 Stat foundations of multi-agent collaboration
🌫️ Model uncertainty & meta-cognition
🔎 Interpretability
💬 LLMs in behavioral science
I'm recruiting PhDs to work on:
🎯 Stat foundations of multi-agent collaboration
🌫️ Model uncertainty & meta-cognition
🔎 Interpretability
💬 LLMs in behavioral science
Been really unimpressed with NEJM AI so far. Publishing paid articles (with DOIs!) from CEOs of health AI companies and half-baked puff pieces is not a good look. This journal should stay un-indexed for the foreseeable future, and if they ever get indexed, it shouldn't apply retroactively.
#medsky
#medsky
November 4, 2025 at 4:59 AM
Been really unimpressed with NEJM AI so far. Publishing paid articles (with DOIs!) from CEOs of health AI companies and half-baked puff pieces is not a good look. This journal should stay un-indexed for the foreseeable future, and if they ever get indexed, it shouldn't apply retroactively.
#medsky
#medsky
Glad to share our full paper on why deployed clinical prediction models should not discard predicted risks in favour of thresholds with @rexwp.bsky.social and @aidybarnett.bsky.social now up on @jclinepi.bsky.social. A brief explainer (1/4):
#statsky #rstats
authors.elsevier.com/a/1m1D83BcJQ...
#statsky #rstats
authors.elsevier.com/a/1m1D83BcJQ...
authors.elsevier.com
October 31, 2025 at 10:49 AM
Glad to share our full paper on why deployed clinical prediction models should not discard predicted risks in favour of thresholds with @rexwp.bsky.social and @aidybarnett.bsky.social now up on @jclinepi.bsky.social. A brief explainer (1/4):
#statsky #rstats
authors.elsevier.com/a/1m1D83BcJQ...
#statsky #rstats
authors.elsevier.com/a/1m1D83BcJQ...
Absolutely loving A City on Mars by @zachweinersmith.bsky.social and @weinersmith.bsky.social but feel the need to point out an error in the text: Kelly and Zach appeared to have accidentally reversed the axis on this figure
October 28, 2025 at 9:40 AM
Absolutely loving A City on Mars by @zachweinersmith.bsky.social and @weinersmith.bsky.social but feel the need to point out an error in the text: Kelly and Zach appeared to have accidentally reversed the axis on this figure
In health, we commonly measure changes over time in a control and intervention group. In other fields, I often see "difference-in-difference" analyses with a sort of handwavy explanation and no regression equation published. Am I missing something? Is it just having Tx group and time as X vars?
October 26, 2025 at 3:36 AM
In health, we commonly measure changes over time in a control and intervention group. In other fields, I often see "difference-in-difference" analyses with a sort of handwavy explanation and no regression equation published. Am I missing something? Is it just having Tx group and time as X vars?
Dumb q about reporting posterior density of a hurdle Gamma in a meta-analysis. Do I report posterior interval of the non-zero responses separately, or together with the zeroes? I.e., hurdle param was 0.09. Should I report the median [IQR] of all, or just non-zero results? Both would have their uses.
October 24, 2025 at 9:09 AM
Dumb q about reporting posterior density of a hurdle Gamma in a meta-analysis. Do I report posterior interval of the non-zero responses separately, or together with the zeroes? I.e., hurdle param was 0.09. Should I report the median [IQR] of all, or just non-zero results? Both would have their uses.
Adding to this that PPI often has unexpected benefits. In the case of my doctoral thesis, learning that many clinicians interpret predictive models causally, and want things from them that are often not even possible. This work also informed a future grant application - it's not wasted!
We did this paper as part of a methods-focused grant that required patient involvement and I honestly found it so valuable that I tried to add similar involvement to all my grant submissions. Not that any of those got funded, but…
pubmed.ncbi.nlm.nih.gov/29966732/
pubmed.ncbi.nlm.nih.gov/29966732/
Patients and investigators prefer measures of absolute risk in subgroups for pragmatic randomized trials - PubMed
We made four recommendations for pragmatic trials to improve patient centeredness: (1) focus on superiority in effectiveness or safety, rather than noninferiority; (2) involve patients in specifying a...
pubmed.ncbi.nlm.nih.gov
October 24, 2025 at 7:23 AM
Adding to this that PPI often has unexpected benefits. In the case of my doctoral thesis, learning that many clinicians interpret predictive models causally, and want things from them that are often not even possible. This work also informed a future grant application - it's not wasted!
Maybe just me, but I'm often encouraged to see peer-reviewed papers and blog posts from people in industry. We (academics) are paid to publish, but they're doing it for love of the game. Sure, there's often a big fat COI, but often not for methods papers.
#statsky
#statsky
October 24, 2025 at 6:46 AM
Maybe just me, but I'm often encouraged to see peer-reviewed papers and blog posts from people in industry. We (academics) are paid to publish, but they're doing it for love of the game. Sure, there's often a big fat COI, but often not for methods papers.
#statsky
#statsky
EVERY time I raise a clinical informatics issue - statistical, logistical, theoretical, whatever - there is always a clinician or administrator who will say, "But wouldn't this be fixed by AI?" MFer, unless you are going to offer the exact way a loss function is going to be applied, just shut up.
October 24, 2025 at 2:40 AM
EVERY time I raise a clinical informatics issue - statistical, logistical, theoretical, whatever - there is always a clinician or administrator who will say, "But wouldn't this be fixed by AI?" MFer, unless you are going to offer the exact way a loss function is going to be applied, just shut up.
Got a LinkedIn ad to join ICE.
In unrelated news, did you know LinkedIn has a feature to report posts as coming from a 'Dangerous or extremist organization'?
In unrelated news, did you know LinkedIn has a feature to report posts as coming from a 'Dangerous or extremist organization'?
October 16, 2025 at 10:24 PM
Got a LinkedIn ad to join ICE.
In unrelated news, did you know LinkedIn has a feature to report posts as coming from a 'Dangerous or extremist organization'?
In unrelated news, did you know LinkedIn has a feature to report posts as coming from a 'Dangerous or extremist organization'?