Abhijeet Singh
@singhabhi.bsky.social
Development Economist in Stockholm. Past: Oxford Econ, UCL and Delhi Univ. Mostly posts about Econ, India, education, miscellanea. #econsky
Website: www.abhijeetsingh.dev
Website: www.abhijeetsingh.dev
Big picture: `science of scaling’ matters enormously for welfare.
Unlike medicine, social programs need iterative adaptation & testing beyond finding "what works" in small pilots.
Without this, as
@johnlist.bsky.social
notes, “we are performing efficacy tests on steroids”. 16/16"
Unlike medicine, social programs need iterative adaptation & testing beyond finding "what works" in small pilots.
Without this, as
@johnlist.bsky.social
notes, “we are performing efficacy tests on steroids”. 16/16"
September 5, 2025 at 12:03 PM
Big picture: `science of scaling’ matters enormously for welfare.
Unlike medicine, social programs need iterative adaptation & testing beyond finding "what works" in small pilots.
Without this, as
@johnlist.bsky.social
notes, “we are performing efficacy tests on steroids”. 16/16"
Unlike medicine, social programs need iterative adaptation & testing beyond finding "what works" in small pilots.
Without this, as
@johnlist.bsky.social
notes, “we are performing efficacy tests on steroids”. 16/16"
These results also matter for rich countries.
High-dosage tutoring has been found to be very effective but has been hard to scale since good tutors are scarce, and costly.
PAL could make personalized tutoring feasible at scale. 15/16
High-dosage tutoring has been found to be very effective but has been hard to scale since good tutors are scarce, and costly.
PAL could make personalized tutoring feasible at scale. 15/16
September 5, 2025 at 12:03 PM
These results also matter for rich countries.
High-dosage tutoring has been found to be very effective but has been hard to scale since good tutors are scarce, and costly.
PAL could make personalized tutoring feasible at scale. 15/16
High-dosage tutoring has been found to be very effective but has been hard to scale since good tutors are scarce, and costly.
PAL could make personalized tutoring feasible at scale. 15/16
These results matter beyond India.
Most LMICs face a “learning crisis”, and have large numbers of students below grade-level standards.
Middle-school years are especially challenging with few evidence-backed scalable ideas.
Our results offer a promising way forward. 14/16
Most LMICs face a “learning crisis”, and have large numbers of students below grade-level standards.
Middle-school years are especially challenging with few evidence-backed scalable ideas.
Our results offer a promising way forward. 14/16
September 5, 2025 at 12:03 PM
These results matter beyond India.
Most LMICs face a “learning crisis”, and have large numbers of students below grade-level standards.
Middle-school years are especially challenging with few evidence-backed scalable ideas.
Our results offer a promising way forward. 14/16
Most LMICs face a “learning crisis”, and have large numbers of students below grade-level standards.
Middle-school years are especially challenging with few evidence-backed scalable ideas.
Our results offer a promising way forward. 14/16
Why is this a big deal?
Governments spend billions on hardware, but evidence shows hardware alone ≠ learning.
What matters is thoughtful integration of PAL software into classrooms.
Our study shows how to make EdTech actually work in public schools. 12/16
Governments spend billions on hardware, but evidence shows hardware alone ≠ learning.
What matters is thoughtful integration of PAL software into classrooms.
Our study shows how to make EdTech actually work in public schools. 12/16
September 5, 2025 at 12:03 PM
Why is this a big deal?
Governments spend billions on hardware, but evidence shows hardware alone ≠ learning.
What matters is thoughtful integration of PAL software into classrooms.
Our study shows how to make EdTech actually work in public schools. 12/16
Governments spend billions on hardware, but evidence shows hardware alone ≠ learning.
What matters is thoughtful integration of PAL software into classrooms.
Our study shows how to make EdTech actually work in public schools. 12/16
One disappointment: we find no gains on school exams.
Students were so far behind that even large gains didn’t show up on grade-level tests.
Highlights trade-off between teaching “at the right level” vs “at the curricular level,” and the value of early remediation. 10/16
Students were so far behind that even large gains didn’t show up on grade-level tests.
Highlights trade-off between teaching “at the right level” vs “at the curricular level,” and the value of early remediation. 10/16
September 5, 2025 at 12:03 PM
One disappointment: we find no gains on school exams.
Students were so far behind that even large gains didn’t show up on grade-level tests.
Highlights trade-off between teaching “at the right level” vs “at the curricular level,” and the value of early remediation. 10/16
Students were so far behind that even large gains didn’t show up on grade-level tests.
Highlights trade-off between teaching “at the right level” vs “at the curricular level,” and the value of early remediation. 10/16
A key new result, enabled by multi-year RCT:
PAL significantly reduced mismatch between students and curriculum over time as seen in the pivot over time in the Figure below: 8/16
PAL significantly reduced mismatch between students and curriculum over time as seen in the pivot over time in the Figure below: 8/16
September 5, 2025 at 12:03 PM
A key new result, enabled by multi-year RCT:
PAL significantly reduced mismatch between students and curriculum over time as seen in the pivot over time in the Figure below: 8/16
PAL significantly reduced mismatch between students and curriculum over time as seen in the pivot over time in the Figure below: 8/16
All students gained comparably, consistent with personalization
BUT, since weaker students learn less in control, the same *absolute* effect is a much larger *relative* improvement over business-as-usual gains for weaker students. 7/16
BUT, since weaker students learn less in control, the same *absolute* effect is a much larger *relative* improvement over business-as-usual gains for weaker students. 7/16
September 5, 2025 at 12:03 PM
All students gained comparably, consistent with personalization
BUT, since weaker students learn less in control, the same *absolute* effect is a much larger *relative* improvement over business-as-usual gains for weaker students. 7/16
BUT, since weaker students learn less in control, the same *absolute* effect is a much larger *relative* improvement over business-as-usual gains for weaker students. 7/16
Our main results:
After 18 months, treated students scored 0.2-0.22SD higher in both Math/Hindi.
That’s 50–66% more learning per year than control schools.
Effect size is in ~90th percentile of effects from large education RCTs (N>5000).
@daveevansphd.bsky.social
6/16
After 18 months, treated students scored 0.2-0.22SD higher in both Math/Hindi.
That’s 50–66% more learning per year than control schools.
Effect size is in ~90th percentile of effects from large education RCTs (N>5000).
@daveevansphd.bsky.social
6/16
September 5, 2025 at 12:03 PM
Our main results:
After 18 months, treated students scored 0.2-0.22SD higher in both Math/Hindi.
That’s 50–66% more learning per year than control schools.
Effect size is in ~90th percentile of effects from large education RCTs (N>5000).
@daveevansphd.bsky.social
6/16
After 18 months, treated students scored 0.2-0.22SD higher in both Math/Hindi.
That’s 50–66% more learning per year than control schools.
Effect size is in ~90th percentile of effects from large education RCTs (N>5000).
@daveevansphd.bsky.social
6/16
Our study design:
RCT with 80 public schools (40 treatment and 40 control) across 4 districts in Rajasthan.
All students in Grades 1-8 in treated schools were included (~6500/year).
We focus on independent measures of student learning after 2-years of treatment 5/16
RCT with 80 public schools (40 treatment and 40 control) across 4 districts in Rajasthan.
All students in Grades 1-8 in treated schools were included (~6500/year).
We focus on independent measures of student learning after 2-years of treatment 5/16
September 5, 2025 at 12:03 PM
Our study design:
RCT with 80 public schools (40 treatment and 40 control) across 4 districts in Rajasthan.
All students in Grades 1-8 in treated schools were included (~6500/year).
We focus on independent measures of student learning after 2-years of treatment 5/16
RCT with 80 public schools (40 treatment and 40 control) across 4 districts in Rajasthan.
All students in Grades 1-8 in treated schools were included (~6500/year).
We focus on independent measures of student learning after 2-years of treatment 5/16
Why might PAL work here?
Because the learning crisis is severe:
(i) In grade 8, average math skill is at Grade 4 level
(ii) within class variation in learning spans many grade levels.
Even the best teachers struggle to handle this problem, which PAL can address 4/16
Because the learning crisis is severe:
(i) In grade 8, average math skill is at Grade 4 level
(ii) within class variation in learning spans many grade levels.
Even the best teachers struggle to handle this problem, which PAL can address 4/16
September 5, 2025 at 12:03 PM
Why might PAL work here?
Because the learning crisis is severe:
(i) In grade 8, average math skill is at Grade 4 level
(ii) within class variation in learning spans many grade levels.
Even the best teachers struggle to handle this problem, which PAL can address 4/16
Because the learning crisis is severe:
(i) In grade 8, average math skill is at Grade 4 level
(ii) within class variation in learning spans many grade levels.
Even the best teachers struggle to handle this problem, which PAL can address 4/16
This study:
Working with
Educational Initiatives,
we developed an in-school model where students spent 1 period/day in a Mindspark lab for Math/Hindi.
This displaced ~25–50% of regular class time
Impacts not obvious since tech-induced classroom disruptions can often REDUCE learning. 3/16
Working with
Educational Initiatives,
we developed an in-school model where students spent 1 period/day in a Mindspark lab for Math/Hindi.
This displaced ~25–50% of regular class time
Impacts not obvious since tech-induced classroom disruptions can often REDUCE learning. 3/16
September 5, 2025 at 12:03 PM
This study:
Working with
Educational Initiatives,
we developed an in-school model where students spent 1 period/day in a Mindspark lab for Math/Hindi.
This displaced ~25–50% of regular class time
Impacts not obvious since tech-induced classroom disruptions can often REDUCE learning. 3/16
Working with
Educational Initiatives,
we developed an in-school model where students spent 1 period/day in a Mindspark lab for Math/Hindi.
This displaced ~25–50% of regular class time
Impacts not obvious since tech-induced classroom disruptions can often REDUCE learning. 3/16
n our 2019 AER paper, we found that PAL (Mindspark) had huge positive effects in Delhi.
But that RCT treated ~300 students in an after-school setting with extra instructional time on PAL.
The big question: Can it work in public schools, at scale, within the school day? 2/16
But that RCT treated ~300 students in an after-school setting with extra instructional time on PAL.
The big question: Can it work in public schools, at scale, within the school day? 2/16
September 5, 2025 at 12:03 PM
n our 2019 AER paper, we found that PAL (Mindspark) had huge positive effects in Delhi.
But that RCT treated ~300 students in an after-school setting with extra instructional time on PAL.
The big question: Can it work in public schools, at scale, within the school day? 2/16
But that RCT treated ~300 students in an after-school setting with extra instructional time on PAL.
The big question: Can it work in public schools, at scale, within the school day? 2/16
Many interventions “work” in small trials but fail at scale
Also, EdTech often promises much, but delivers little
In a new paper (bit.ly/3JKLgVn)
@karthik-econ.bsky.social
& I show how personalized adaptive learning (PAL) software can sharply improve learning outcomes at scale🧵1/16
Also, EdTech often promises much, but delivers little
In a new paper (bit.ly/3JKLgVn)
@karthik-econ.bsky.social
& I show how personalized adaptive learning (PAL) software can sharply improve learning outcomes at scale🧵1/16
September 5, 2025 at 12:03 PM
Many interventions “work” in small trials but fail at scale
Also, EdTech often promises much, but delivers little
In a new paper (bit.ly/3JKLgVn)
@karthik-econ.bsky.social
& I show how personalized adaptive learning (PAL) software can sharply improve learning outcomes at scale🧵1/16
Also, EdTech often promises much, but delivers little
In a new paper (bit.ly/3JKLgVn)
@karthik-econ.bsky.social
& I show how personalized adaptive learning (PAL) software can sharply improve learning outcomes at scale🧵1/16
This take away also seems to be basically the heuristic that Stuart Buck used to question the paper (convincingly, early on).
May 17, 2025 at 9:14 AM
This take away also seems to be basically the heuristic that Stuart Buck used to question the paper (convincingly, early on).
I'll be presenting tomorrow at the joint Harvard-MIT Development seminar (2:30 p.m.- 4 p.m.).
Friends working on econ or edu in Boston/Cambridge, come over to say hi if you've got nothing else on!
Friends working on econ or edu in Boston/Cambridge, come over to say hi if you've got nothing else on!
April 22, 2025 at 11:04 PM
I'll be presenting tomorrow at the joint Harvard-MIT Development seminar (2:30 p.m.- 4 p.m.).
Friends working on econ or edu in Boston/Cambridge, come over to say hi if you've got nothing else on!
Friends working on econ or edu in Boston/Cambridge, come over to say hi if you've got nothing else on!
Gotta love some paper titles!
Go on, Chuck, tell us what you really think about multiple imputation 😂
www.journals.uchicago.edu/doi/abs/10.1...
Go on, Chuck, tell us what you really think about multiple imputation 😂
www.journals.uchicago.edu/doi/abs/10.1...
April 12, 2025 at 8:48 PM
Gotta love some paper titles!
Go on, Chuck, tell us what you really think about multiple imputation 😂
www.journals.uchicago.edu/doi/abs/10.1...
Go on, Chuck, tell us what you really think about multiple imputation 😂
www.journals.uchicago.edu/doi/abs/10.1...
Like Klaus Adam + @lugaricano.bsky.social I think this one-off opportunity *could* be transformative for Europe (and cheap for the price!).
But, even w/ funding, making it happen is hard. Top researcher salaries would need to go up 3-4x to compete with the US...
But, even w/ funding, making it happen is hard. Top researcher salaries would need to go up 3-4x to compete with the US...
February 22, 2025 at 11:12 PM
Like Klaus Adam + @lugaricano.bsky.social I think this one-off opportunity *could* be transformative for Europe (and cheap for the price!).
But, even w/ funding, making it happen is hard. Top researcher salaries would need to go up 3-4x to compete with the US...
But, even w/ funding, making it happen is hard. Top researcher salaries would need to go up 3-4x to compete with the US...
@yann-lecun.bsky.social on what the EU needs to attract top scientists from the US.
AFAIK only Swiss unis plus very few others solve all 7 criteria. Eg Sweden IMO solves the rest (=some great clusters!), but not #3.
That retains some intra-EU talent but it's hard to hire from R1 US unis
AFAIK only Swiss unis plus very few others solve all 7 criteria. Eg Sweden IMO solves the rest (=some great clusters!), but not #3.
That retains some intra-EU talent but it's hard to hire from R1 US unis
February 22, 2025 at 11:12 PM
@yann-lecun.bsky.social on what the EU needs to attract top scientists from the US.
AFAIK only Swiss unis plus very few others solve all 7 criteria. Eg Sweden IMO solves the rest (=some great clusters!), but not #3.
That retains some intra-EU talent but it's hard to hire from R1 US unis
AFAIK only Swiss unis plus very few others solve all 7 criteria. Eg Sweden IMO solves the rest (=some great clusters!), but not #3.
That retains some intra-EU talent but it's hard to hire from R1 US unis
I could get very old and I'd still find frozen lakes with trapped boats gorgeously exotic. Stockholm was looking lovely today.
(There were also people jumping into ice water -- that I'll never find natural or a reasonable thing to do!)
(There were also people jumping into ice water -- that I'll never find natural or a reasonable thing to do!)
February 16, 2025 at 5:32 PM
I could get very old and I'd still find frozen lakes with trapped boats gorgeously exotic. Stockholm was looking lovely today.
(There were also people jumping into ice water -- that I'll never find natural or a reasonable thing to do!)
(There were also people jumping into ice water -- that I'll never find natural or a reasonable thing to do!)
I'm assuming this was an online poll?
Otherwise, I'm scratching my head that (i) only 10% of Indian survey respondents say "don't know" re whether Trump's rise is good for India, and (ii) a staggering majority (86%!) think it's good.
Heck, even a majority of Chinese respondents seem positive?
Otherwise, I'm scratching my head that (i) only 10% of Indian survey respondents say "don't know" re whether Trump's rise is good for India, and (ii) a staggering majority (86%!) think it's good.
Heck, even a majority of Chinese respondents seem positive?
January 22, 2025 at 2:15 AM
I'm assuming this was an online poll?
Otherwise, I'm scratching my head that (i) only 10% of Indian survey respondents say "don't know" re whether Trump's rise is good for India, and (ii) a staggering majority (86%!) think it's good.
Heck, even a majority of Chinese respondents seem positive?
Otherwise, I'm scratching my head that (i) only 10% of Indian survey respondents say "don't know" re whether Trump's rise is good for India, and (ii) a staggering majority (86%!) think it's good.
Heck, even a majority of Chinese respondents seem positive?
Here is one example I use in slides (from a paper w/ Mauricio @marome1.bsky.social)
We needed to do this because the treatment effect is entirely driven by the control mean. Works well in talks.
The paper, of course, has the full tables.
We needed to do this because the treatment effect is entirely driven by the control mean. Works well in talks.
The paper, of course, has the full tables.
September 18, 2024 at 1:33 PM
Here is one example I use in slides (from a paper w/ Mauricio @marome1.bsky.social)
We needed to do this because the treatment effect is entirely driven by the control mean. Works well in talks.
The paper, of course, has the full tables.
We needed to do this because the treatment effect is entirely driven by the control mean. Works well in talks.
The paper, of course, has the full tables.
TFW you really miss @instrumenthull.bsky.social on the other app. This could've been such a spicy exchange!
November 14, 2023 at 9:42 PM
TFW you really miss @instrumenthull.bsky.social on the other app. This could've been such a spicy exchange!
Apparently, I'm closest to Marianne Bertrand.
If only I had even a fraction of her productivity!!
If only I had even a fraction of her productivity!!
October 12, 2023 at 9:25 PM
Apparently, I'm closest to Marianne Bertrand.
If only I had even a fraction of her productivity!!
If only I had even a fraction of her productivity!!
Today's Nobel for Claudia Goldin is inspiring in so many ways.
One of them is recognizing important *descriptive* work built painstakingly over decades. This is rare.
Two big examples are Fogel (C.Goldin's advisor) and Kuznets (Fogel's advisor). Few other examples (Stone, Deaton, Ostrom?).
One of them is recognizing important *descriptive* work built painstakingly over decades. This is rare.
Two big examples are Fogel (C.Goldin's advisor) and Kuznets (Fogel's advisor). Few other examples (Stone, Deaton, Ostrom?).
October 9, 2023 at 8:10 PM
Today's Nobel for Claudia Goldin is inspiring in so many ways.
One of them is recognizing important *descriptive* work built painstakingly over decades. This is rare.
Two big examples are Fogel (C.Goldin's advisor) and Kuznets (Fogel's advisor). Few other examples (Stone, Deaton, Ostrom?).
One of them is recognizing important *descriptive* work built painstakingly over decades. This is rare.
Two big examples are Fogel (C.Goldin's advisor) and Kuznets (Fogel's advisor). Few other examples (Stone, Deaton, Ostrom?).
Dunno the cross posting etiquette from the other place.
But @marome1.bsky.social and I won a prize for our paper and we're just thrilled that folks studying Latin America find our India work interesting!!
(And yes, I know it's a bit cringe to post about prizes but honestly this made our day)
But @marome1.bsky.social and I won a prize for our paper and we're just thrilled that folks studying Latin America find our India work interesting!!
(And yes, I know it's a bit cringe to post about prizes but honestly this made our day)
September 25, 2023 at 9:30 PM
Dunno the cross posting etiquette from the other place.
But @marome1.bsky.social and I won a prize for our paper and we're just thrilled that folks studying Latin America find our India work interesting!!
(And yes, I know it's a bit cringe to post about prizes but honestly this made our day)
But @marome1.bsky.social and I won a prize for our paper and we're just thrilled that folks studying Latin America find our India work interesting!!
(And yes, I know it's a bit cringe to post about prizes but honestly this made our day)