Ehud Karavani
@ehudk.bsky.social
Research Staff Member at IBM Research.
Causal Inference 🔴→🟠←🟡.
Machine Learning 🤖🎓.
Data Communication 📈.
Healthcare ⚕️.
Creator of 𝙲𝚊𝚞𝚜𝚊𝚕𝚕𝚒𝚋: https://github.com/IBM/causallib
Website: https://ehud.co
Causal Inference 🔴→🟠←🟡.
Machine Learning 🤖🎓.
Data Communication 📈.
Healthcare ⚕️.
Creator of 𝙲𝚊𝚞𝚜𝚊𝚕𝚕𝚒𝚋: https://github.com/IBM/causallib
Website: https://ehud.co
The March of Folly by Barbara Tuchman can make a good read while the US is writing its next chapter
November 30, 2025 at 8:54 PM
The March of Folly by Barbara Tuchman can make a good read while the US is writing its next chapter
kinda depends if it's a methods of applications paper, but I generally have the same strategy, though I'll try to be more gracious and skim all figures/tables
November 27, 2025 at 9:27 AM
kinda depends if it's a methods of applications paper, but I generally have the same strategy, though I'll try to be more gracious and skim all figures/tables
I mean not all of her points are substantive, but you've got to give it to her -- there's no sleuthing like future-mom's sleuthing 🧑🍳💋
November 27, 2025 at 7:01 AM
I mean not all of her points are substantive, but you've got to give it to her -- there's no sleuthing like future-mom's sleuthing 🧑🍳💋
huh, nice, i should read this. i had a similar tree-based approach for point interventions a while back
arxiv.org/abs/1907.08127
arxiv.org/abs/1907.08127
A discriminative approach for finding and characterizing positivity violations using decision trees
The assumption of positivity in causal inference (also known as common support and co-variate overlap) is necessary to obtain valid causal estimates. Therefore, confirming it holds in a given dataset ...
arxiv.org
November 19, 2025 at 4:05 PM
huh, nice, i should read this. i had a similar tree-based approach for point interventions a while back
arxiv.org/abs/1907.08127
arxiv.org/abs/1907.08127
and you probably know this, but i've yet to encounter a continuous treatment problem where overlap has been well-behaved (not to mention in time-varying settings). to compensate, i suggest having a good outcome model companion, flexible enough to trust its extrapolations in these overlap gaps.
November 19, 2025 at 3:57 PM
and you probably know this, but i've yet to encounter a continuous treatment problem where overlap has been well-behaved (not to mention in time-varying settings). to compensate, i suggest having a good outcome model companion, flexible enough to trust its extrapolations in these overlap gaps.
ugh, gps, i'm so sorry for you, but yes, although i'm personally more of a ridgeline plot [1] kind of guy myself
[1] medium.com/data-science...
[1] medium.com/data-science...
November 19, 2025 at 3:49 PM
ugh, gps, i'm so sorry for you, but yes, although i'm personally more of a ridgeline plot [1] kind of guy myself
[1] medium.com/data-science...
[1] medium.com/data-science...
well, it of course depends on your positivity metric, dag, and estimator, but if you can calculate the propensity scores at time 𝘵, for example:
𝚐𝚕𝚖(𝚊𝟹 ~ 𝚇𝚋𝚊𝚜𝚎 + 𝚇𝟷 + 𝚊𝟷 + 𝚇𝟸 + 𝚊𝟸 + 𝚇𝟹, 𝚍𝚊𝚝𝚊=𝚍𝚊𝚝𝚊.𝚠𝚒𝚍𝚎, 𝚏𝚊𝚖𝚒𝚕𝚢="𝚋𝚒𝚗𝚘𝚖𝚒𝚊𝚕")
then you can, for example, plot them colored by treatment assignment (a3).
𝚐𝚕𝚖(𝚊𝟹 ~ 𝚇𝚋𝚊𝚜𝚎 + 𝚇𝟷 + 𝚊𝟷 + 𝚇𝟸 + 𝚊𝟸 + 𝚇𝟹, 𝚍𝚊𝚝𝚊=𝚍𝚊𝚝𝚊.𝚠𝚒𝚍𝚎, 𝚏𝚊𝚖𝚒𝚕𝚢="𝚋𝚒𝚗𝚘𝚖𝚒𝚊𝚕")
then you can, for example, plot them colored by treatment assignment (a3).
November 19, 2025 at 1:04 PM
well, it of course depends on your positivity metric, dag, and estimator, but if you can calculate the propensity scores at time 𝘵, for example:
𝚐𝚕𝚖(𝚊𝟹 ~ 𝚇𝚋𝚊𝚜𝚎 + 𝚇𝟷 + 𝚊𝟷 + 𝚇𝟸 + 𝚊𝟸 + 𝚇𝟹, 𝚍𝚊𝚝𝚊=𝚍𝚊𝚝𝚊.𝚠𝚒𝚍𝚎, 𝚏𝚊𝚖𝚒𝚕𝚢="𝚋𝚒𝚗𝚘𝚖𝚒𝚊𝚕")
then you can, for example, plot them colored by treatment assignment (a3).
𝚐𝚕𝚖(𝚊𝟹 ~ 𝚇𝚋𝚊𝚜𝚎 + 𝚇𝟷 + 𝚊𝟷 + 𝚇𝟸 + 𝚊𝟸 + 𝚇𝟹, 𝚍𝚊𝚝𝚊=𝚍𝚊𝚝𝚊.𝚠𝚒𝚍𝚎, 𝚏𝚊𝚖𝚒𝚕𝚢="𝚋𝚒𝚗𝚘𝚖𝚒𝚊𝚕")
then you can, for example, plot them colored by treatment assignment (a3).
I always read sequential positivity as "do whatever you do for point interventions but stratified for each time point"
November 18, 2025 at 10:14 AM
I always read sequential positivity as "do whatever you do for point interventions but stratified for each time point"
not dml per se, but seems that simpler propensity-feature regressions may offer some protection from regularization-induced confounding. See section 4.3 in projecteuclid.org/journals/bay...
so my hunch is that dml is likely to be relatively better (tho not necessarily absolutely good).
so my hunch is that dml is likely to be relatively better (tho not necessarily absolutely good).
projecteuclid.org
November 15, 2025 at 7:48 PM
not dml per se, but seems that simpler propensity-feature regressions may offer some protection from regularization-induced confounding. See section 4.3 in projecteuclid.org/journals/bay...
so my hunch is that dml is likely to be relatively better (tho not necessarily absolutely good).
so my hunch is that dml is likely to be relatively better (tho not necessarily absolutely good).
I tried to do that in the past and failed. I'll be interested if you could share some code. I think it could be really beneficial to tailor a seaborn object around that. the new interface is pretty customizable and I'd love some ggdist in my python
November 15, 2025 at 5:54 AM
I tried to do that in the past and failed. I'll be interested if you could share some code. I think it could be really beneficial to tailor a seaborn object around that. the new interface is pretty customizable and I'd love some ggdist in my python
I don't mind complaining about how Python lacks ggdist or brms, but
𝚍𝚏.𝚐𝚛𝚘𝚞𝚙𝚋𝚢(
["𝚜𝚙𝚎𝚌𝚒𝚎𝚜", "𝚒𝚜𝚕𝚊𝚗𝚍"],
)["𝚋𝚘𝚍𝚢_𝚖𝚊𝚜𝚜_𝚐"].𝚊𝚐𝚐(
["𝚖𝚎𝚊𝚗", "𝚜𝚝𝚍"]
)
is not that bad to be honest
𝚍𝚏.𝚐𝚛𝚘𝚞𝚙𝚋𝚢(
["𝚜𝚙𝚎𝚌𝚒𝚎𝚜", "𝚒𝚜𝚕𝚊𝚗𝚍"],
)["𝚋𝚘𝚍𝚢_𝚖𝚊𝚜𝚜_𝚐"].𝚊𝚐𝚐(
["𝚖𝚎𝚊𝚗", "𝚜𝚝𝚍"]
)
is not that bad to be honest
November 14, 2025 at 6:08 PM
I don't mind complaining about how Python lacks ggdist or brms, but
𝚍𝚏.𝚐𝚛𝚘𝚞𝚙𝚋𝚢(
["𝚜𝚙𝚎𝚌𝚒𝚎𝚜", "𝚒𝚜𝚕𝚊𝚗𝚍"],
)["𝚋𝚘𝚍𝚢_𝚖𝚊𝚜𝚜_𝚐"].𝚊𝚐𝚐(
["𝚖𝚎𝚊𝚗", "𝚜𝚝𝚍"]
)
is not that bad to be honest
𝚍𝚏.𝚐𝚛𝚘𝚞𝚙𝚋𝚢(
["𝚜𝚙𝚎𝚌𝚒𝚎𝚜", "𝚒𝚜𝚕𝚊𝚗𝚍"],
)["𝚋𝚘𝚍𝚢_𝚖𝚊𝚜𝚜_𝚐"].𝚊𝚐𝚐(
["𝚖𝚎𝚊𝚗", "𝚜𝚝𝚍"]
)
is not that bad to be honest
Coming from a computer science background this was one of my biggest confusions at first -- statisticians reserving "nonlinearity" to link functions and transformed outcomes rather than to transformed covariates
November 14, 2025 at 12:04 PM
Coming from a computer science background this was one of my biggest confusions at first -- statisticians reserving "nonlinearity" to link functions and transformed outcomes rather than to transformed covariates
nice. could easily wait a month for the bmj christmas issue
November 14, 2025 at 6:46 AM
nice. could easily wait a month for the bmj christmas issue
Reposted by Ehud Karavani
Rubin's gonna Rubin, but Robins Rubins harder.
November 8, 2025 at 2:08 PM
Rubin's gonna Rubin, but Robins Rubins harder.
it ain't nuthin' but a g thang
November 8, 2025 at 5:07 PM
it ain't nuthin' but a g thang
Also a refreshing example for a successful English song from east of the pond deliberately not sang in an American-adjacent accent
November 4, 2025 at 10:28 AM
Also a refreshing example for a successful English song from east of the pond deliberately not sang in an American-adjacent accent
Well, this quote is now 16 years old, and judging by the current state and the trends of the past few years (GenAI for scientific processes, the pipeline-ization of data science), I think his prediction was spot on. Statistics did have significant spotlight on it for 10 years, but now it diminishes.
November 4, 2025 at 10:12 AM
Well, this quote is now 16 years old, and judging by the current state and the trends of the past few years (GenAI for scientific processes, the pipeline-ization of data science), I think his prediction was spot on. Statistics did have significant spotlight on it for 10 years, but now it diminishes.
Figure 4 completely missed the joke about Normal distributions and paranormal distributions
October 30, 2025 at 3:54 AM
Figure 4 completely missed the joke about Normal distributions and paranormal distributions
I just spent like 30 minutes to find that obligatory bmj segment about how surgeons are carefully trained or whatever while we keep providing everyone with unregulated access to spss
October 28, 2025 at 2:02 PM
I just spent like 30 minutes to find that obligatory bmj segment about how surgeons are carefully trained or whatever while we keep providing everyone with unregulated access to spss
which creates an ironic situation where actual monarchies are nowadays more democratic then the OG monarchy overthrowers
October 26, 2025 at 6:21 AM
which creates an ironic situation where actual monarchies are nowadays more democratic then the OG monarchy overthrowers
oh man, i love it, but we already have enough options for what a "music program" can mean.
October 24, 2025 at 9:07 AM
oh man, i love it, but we already have enough options for what a "music program" can mean.
An excellent opportunity to sneak in a DAG into a flowchart
October 23, 2025 at 2:05 PM
An excellent opportunity to sneak in a DAG into a flowchart