|
Frank Harrell
@
f2harrell
Nashville, TN
|
|
Prof, Founding Chair of Biostatistics, Vanderbilt U. Expert Statistical Advisor, Office of Biostatistics, FDA CDER. Blog: Statistical Thinking -fharrell.com
|
|
|
7,143
ಟ್ವೀಟ್ಗಳು
|
78
ಹಿಂಬಾಲಿಸುತ್ತಿರುವವರು
|
9,043
ಹಿಂಬಾಲಕರು
|
| ಟ್ವೀಟ್ಗಳು |
|
Frank Harrell
@f2harrell
|
9 ಗಂ. |
|
What makes a calculation that uses a single arbitrary effect size valuable?
|
||
|
|
||
|
Frank Harrell
@f2harrell
|
9 ಗಂ. |
|
Don't see that logic, sorry.
|
||
|
|
||
|
Frank Harrell
@f2harrell
|
9 ಗಂ. |
|
Don't see this for cost-effectiveness because cost-effectiveness varies drastically by patient. On other point, what are the top examples of blanket public policy decisions?
|
||
|
|
||
| Frank Harrell ಮರುಟ್ವೀಟಿಸಿದ್ದಾರೆ | ||
|
https://toot.cafe/@chartier
@chartier
|
19 ಗಂ. |
|
1. Shut #Facebook down
2. Jail key decision makers and possibly some of the engineers involved
3. Give the money back to the people it abused (its users)
4. Build ethical replacements for its useful bits
nytimes.com/2018/12/18/tec…
|
||
|
|
||
|
Frank Harrell
@f2harrell
|
14 ಗಂ. |
|
What are the problems with patient-specific absolute risk estimates?
|
||
|
|
||
|
Frank Harrell
@f2harrell
|
14 ಗಂ. |
|
Can't think of why NNT would be good in that context.
|
||
|
|
||
|
Frank Harrell
@f2harrell
|
16 ಗಂ. |
|
Still trying to find an example where population effect is clearly the best target. Since it's usually not, non-representative RCTs are generally better than representative observational studies.
|
||
|
|
||
|
Frank Harrell
@f2harrell
|
16 ಗಂ. |
|
Glad that @venkmurthy has restarted this hot discussion on datamethods. I just posted a response that tries to generalize the question. twitter.com/venkmurthy/sta…
|
||
|
|
||
|
Frank Harrell
@f2harrell
|
16 ಗಂ. |
|
I think that @ken_rothman did something related to this. Perhaps he can comment.
|
||
|
|
||
| Frank Harrell ಮರುಟ್ವೀಟಿಸಿದ್ದಾರೆ | ||
|
🔥 Kareem Carr 🔥
@kareem_carr
|
ಡಿಸೆಂ 19 |
|
Statistically, machine learning has poor modeling practices which should limit generalizability. It doesn't. Why? For tasks like image recognition, all humans on the ML team understand ground truth; in healthcare/science tasks, it won't be the same. We can expect lots of failure.
|
||
|
|
||
|
Frank Harrell
@f2harrell
|
17 ಗಂ. |
|
(1) Use splits other than the median and shows it matters, (2) Show that quantiles are divorced from biology, (3) Explain why a pre-specified statististical analysis plan was needed and since it wasn't written this is all a game.
|
||
|
|
||
| Frank Harrell ಮರುಟ್ವೀಟಿಸಿದ್ದಾರೆ | ||
|
Paul Bürkner
@paulbuerkner
|
23 ಗಂ. |
|
My Christmas present for the #rstats community: #brms 2.7 is now on CRAN extending support for Gaussian processes and fixing a substantial performance bug in some post-processing methods among other new features and improvements. github.com/paul-buerkner/…
|
||
|
|
||
|
Frank Harrell
@f2harrell
|
17 ಗಂ. |
|
I haven't commented on this because I don't like twitter for this. Too disjointed. See if you want to reopen this discussion on datamethods.org.
|
||
|
|
||
|
Frank Harrell
@f2harrell
|
17 ಗಂ. |
|
I wouldn't equate HTE with subgroups but rather interaction with treatment on an appropriate scale (a scale for which such interaction may be absent).
|
||
|
|
||
|
Frank Harrell
@f2harrell
|
ಡಿಸೆಂ 19 |
|
What a wonderful graph. Love it.
|
||
|
|
||
| Frank Harrell ಮರುಟ್ವೀಟಿಸಿದ್ದಾರೆ | ||
|
Roger J. Lewis, MD, PhD
@RogerJLewis
|
ಡಿಸೆಂ 17 |
|
Please read—well worth your time. Challenges of Non–Intention-to-Treat Analyses | Research, Methods, Statistics | JAMA | JAMA Network jamanetwork.com/journals/jama/…
|
||
|
|
||
| Frank Harrell ಮರುಟ್ವೀಟಿಸಿದ್ದಾರೆ | ||
|
Colin Begg
@ColinBBegg
|
ಡಿಸೆಂ 17 |
|
It's a big stretch to put such a positive spin on the wastage of resources that occurs when the scientific pathway is led astray building upon results that can't be replicated. @ReplicateEU twitter.com/dirnagl/status…
|
||
|
|
||
|
Frank Harrell
@f2harrell
|
ಡಿಸೆಂ 17 |
|
I'm not seeing that Dan. Think about linking continuous height and continuous weight into BMI (granted BMI is an imperfect index of obesity). The first thing is to understand risk. Up-front categorization of lab measurements doesn't accomplish that.
|
||
|
|
||
|
Frank Harrell
@f2harrell
|
ಡಿಸೆಂ 17 |
|
Heartily disagree. It's a fundamental building block I use in many general functions.
|
||
|
|
||
|
Frank Harrell
@f2harrell
|
ಡಿಸೆಂ 17 |
|
Glad to see questionable research practices elevated with respect to severity of harm to science twitter.com/RetractionWatc…
|
||
|
|
||