Mauricio Romero
@marome1.bsky.social
Colombian in Mexico. Mountaineer, cyclist, climber. Associate Professor of Economics at ITAM. https://mauricio-romero.com/
Indeed, in 2019, the Ministry rolled out the guide nationwide; cohort differences vanished thereafter, consistent with this.
October 8, 2025 at 11:07 PM
Indeed, in 2019, the Ministry rolled out the guide nationwide; cohort differences vanished thereafter, consistent with this.
ENTRE was designed for scale: led by Guatemala’s Ministry of Education using existing staff, costing only USD 2–3 per student. This gov’t ownership reduced the risk of ‘fade out’ often seen when NGO pilots expand.
(See "experiment at scale" by @karthik-econ.bsky.social Paul Niehaus)
(See "experiment at scale" by @karthik-econ.bsky.social Paul Niehaus)
October 8, 2025 at 11:07 PM
ENTRE was designed for scale: led by Guatemala’s Ministry of Education using existing staff, costing only USD 2–3 per student. This gov’t ownership reduced the risk of ‘fade out’ often seen when NGO pilots expand.
(See "experiment at scale" by @karthik-econ.bsky.social Paul Niehaus)
(See "experiment at scale" by @karthik-econ.bsky.social Paul Niehaus)
Admin data let us see longer-term: ENTRE boosted 6th→7th grade transition in 2019, but by 2020–22 gains vanished. Many later dropped out, likely due to a lack of follow-up support in secondary (and other structural factors)
October 8, 2025 at 11:07 PM
Admin data let us see longer-term: ENTRE boosted 6th→7th grade transition in 2019, but by 2020–22 gains vanished. Many later dropped out, likely due to a lack of follow-up support in secondary (and other structural factors)
We use admin data to track student enrollment for several years. What do we find? Dropout decreases by ~1.2 pp from a 34% base (≈−3.3%). Effects similar across arms; lists and nudges add ~0 beyond training+guide.
October 8, 2025 at 11:07 PM
We use admin data to track student enrollment for several years. What do we find? Dropout decreases by ~1.2 pp from a 34% base (≈−3.3%). Effects similar across arms; lists and nudges add ~0 beyond training+guide.
These 4,000 schools were randomly split:
• 1,000 got training + guide
• 1,000 got training + guide + risk list
• 1,000 got training + guide + risk list + nudges
• 1,000 were controls.
Schools were spread nationwide, making this one of the largest dropout-prevention RCTs.
• 1,000 got training + guide
• 1,000 got training + guide + risk list
• 1,000 got training + guide + risk list + nudges
• 1,000 were controls.
Schools were spread nationwide, making this one of the largest dropout-prevention RCTs.
October 8, 2025 at 11:07 PM
These 4,000 schools were randomly split:
• 1,000 got training + guide
• 1,000 got training + guide + risk list
• 1,000 got training + guide + risk list + nudges
• 1,000 were controls.
Schools were spread nationwide, making this one of the largest dropout-prevention RCTs.
• 1,000 got training + guide
• 1,000 got training + guide + risk list
• 1,000 got training + guide + risk list + nudges
• 1,000 were controls.
Schools were spread nationwide, making this one of the largest dropout-prevention RCTs.
We evaluated ENTRE in 4,000 Guatemalan schools (from 6,080 eligible, ~44% of all 6th graders). Eligibility required enough secondary school supply, excluded the smallest primaries, and only included schools with some predicted at-risk students.
October 8, 2025 at 11:07 PM
We evaluated ENTRE in 4,000 Guatemalan schools (from 6,080 eligible, ~44% of all 6th graders). Eligibility required enough secondary school supply, excluded the smallest primaries, and only included schools with some predicted at-risk students.
Finally, some principals received 5 monthly nudges (via the ministry’s portal) to keep dropout top of mind.
October 8, 2025 at 11:07 PM
Finally, some principals received 5 monthly nudges (via the ministry’s portal) to keep dropout top of mind.
Some schools also received a list of 6th-graders at highest risk of dropping out, predicted using sex, age, GPA, grade repetition & school history. The model identified 82% of future dropouts (See doi.org/10.1080/0964... for more details on the predictive model)
October 8, 2025 at 11:07 PM
Some schools also received a list of 6th-graders at highest risk of dropping out, predicted using sex, age, GPA, grade repetition & school history. The model identified 82% of future dropouts (See doi.org/10.1080/0964... for more details on the predictive model)
The core of ENTRE is a half-day training + a practical guide for principals & 6th-grade teachers. The guide offered simple, low-cost strategies: motivate students, help with scholarship information, provide remedial support, engage families, and ease enrollment logistics.
October 8, 2025 at 11:07 PM
The core of ENTRE is a half-day training + a practical guide for principals & 6th-grade teachers. The guide offered simple, low-cost strategies: motivate students, help with scholarship information, provide remedial support, engage families, and ease enrollment logistics.
Our themes relate to other great work:
RCTs: See work by Ganimian @karthik-econ.bsky.social and Walters on improving AWCs; by @joshtdean.bsky.social @seema.bsky.social on vouchers for private preschool
Facts: See ASER Centre Early Years and main ASER Reports
RCTs: See work by Ganimian @karthik-econ.bsky.social and Walters on improving AWCs; by @joshtdean.bsky.social @seema.bsky.social on vouchers for private preschool
Facts: See ASER Centre Early Years and main ASER Reports
September 29, 2025 at 2:44 PM
Our themes relate to other great work:
RCTs: See work by Ganimian @karthik-econ.bsky.social and Walters on improving AWCs; by @joshtdean.bsky.social @seema.bsky.social on vouchers for private preschool
Facts: See ASER Centre Early Years and main ASER Reports
RCTs: See work by Ganimian @karthik-econ.bsky.social and Walters on improving AWCs; by @joshtdean.bsky.social @seema.bsky.social on vouchers for private preschool
Facts: See ASER Centre Early Years and main ASER Reports
Villages with better public pre/primary schools also have better private sectors ➡️ unequal access to quality education across villages.
Why?
Maybe public sector quality induces better performance from the private sector (as Tahir, Bau,Das @nkarachiwalla.bsky.social and Khwaja find in Pakistan)?
Why?
Maybe public sector quality induces better performance from the private sector (as Tahir, Bau,Das @nkarachiwalla.bsky.social and Khwaja find in Pakistan)?
September 29, 2025 at 2:44 PM
Villages with better public pre/primary schools also have better private sectors ➡️ unequal access to quality education across villages.
Why?
Maybe public sector quality induces better performance from the private sector (as Tahir, Bau,Das @nkarachiwalla.bsky.social and Khwaja find in Pakistan)?
Why?
Maybe public sector quality induces better performance from the private sector (as Tahir, Bau,Das @nkarachiwalla.bsky.social and Khwaja find in Pakistan)?
Result 3: Private preschools outperform public ones in nearly ALL villages.
In primary schooling, this is more variable across geographies.
(Doing this correctly needs Bayesian shrinkage; details in paper)
In primary schooling, this is more variable across geographies.
(Doing this correctly needs Bayesian shrinkage; details in paper)
September 29, 2025 at 2:44 PM
Result 3: Private preschools outperform public ones in nearly ALL villages.
In primary schooling, this is more variable across geographies.
(Doing this correctly needs Bayesian shrinkage; details in paper)
In primary schooling, this is more variable across geographies.
(Doing this correctly needs Bayesian shrinkage; details in paper)
Result 1: Average value-added in private preschools larger than public AWCs by 0.74σ in math & 0.59σ in Tamil.
By contrast, private primary schools have NO positive value-added over government schools in these subjects (we did not test English)
By contrast, private primary schools have NO positive value-added over government schools in these subjects (we did not test English)
September 29, 2025 at 2:44 PM
Result 1: Average value-added in private preschools larger than public AWCs by 0.74σ in math & 0.59σ in Tamil.
By contrast, private primary schools have NO positive value-added over government schools in these subjects (we did not test English)
By contrast, private primary schools have NO positive value-added over government schools in these subjects (we did not test English)
Sample: Household-based panel data on ~19k students (age 3-10) across 215 villages. All children tested one-to-one. Age-appropriate tests linked on a common scale using IRT (using overlaps across ages). Items align with national goals for foundational skills.
September 29, 2025 at 2:44 PM
Sample: Household-based panel data on ~19k students (age 3-10) across 215 villages. All children tested one-to-one. Age-appropriate tests linked on a common scale using IRT (using overlaps across ages). Items align with national goals for foundational skills.
Context: Tamil Nadu, nearly all kids 4-5y enroll in *some* preschool
Public: Mostly anganwadi centres. Free to attend, only 1 worker + 1 helper. In TN, ~38 mins/day on cognitive tasks.
Private: Nursery/KG, often linked to primary sch. Fee-charging, focus on early learning.
Public: Mostly anganwadi centres. Free to attend, only 1 worker + 1 helper. In TN, ~38 mins/day on cognitive tasks.
Private: Nursery/KG, often linked to primary sch. Fee-charging, focus on early learning.
September 29, 2025 at 2:44 PM
Context: Tamil Nadu, nearly all kids 4-5y enroll in *some* preschool
Public: Mostly anganwadi centres. Free to attend, only 1 worker + 1 helper. In TN, ~38 mins/day on cognitive tasks.
Private: Nursery/KG, often linked to primary sch. Fee-charging, focus on early learning.
Public: Mostly anganwadi centres. Free to attend, only 1 worker + 1 helper. In TN, ~38 mins/day on cognitive tasks.
Private: Nursery/KG, often linked to primary sch. Fee-charging, focus on early learning.
Early Childhood Education is central to India’s NEP and global edu goals. Our(@petterberg.bsky.social @singhabhi.bsky.social) new paper at EJ (@resmedia.bsky.social ky.social, bit.ly/4gOtoVV), shows private ECE outperform public options, explaining 60% of the SES gap. In primary, NO private premium.
September 29, 2025 at 2:44 PM
Early Childhood Education is central to India’s NEP and global edu goals. Our(@petterberg.bsky.social @singhabhi.bsky.social) new paper at EJ (@resmedia.bsky.social ky.social, bit.ly/4gOtoVV), shows private ECE outperform public options, explaining 60% of the SES gap. In primary, NO private premium.
Practical guidance for new experiments
1) Only care about main effects? Drop the interaction cells and boost N in treatment/control
2) Want interactions? Power the study to detect them
3) Still running the short model? Specify that the effect you’ll report is a weighted average
1) Only care about main effects? Drop the interaction cells and boost N in treatment/control
2) Want interactions? Power the study to detect them
3) Still running the short model? Specify that the effect you’ll report is a weighted average
May 28, 2025 at 1:12 PM
Practical guidance for new experiments
1) Only care about main effects? Drop the interaction cells and boost N in treatment/control
2) Want interactions? Power the study to detect them
3) Still running the short model? Specify that the effect you’ll report is a weighted average
1) Only care about main effects? Drop the interaction cells and boost N in treatment/control
2) Want interactions? Power the study to detect them
3) Still running the short model? Specify that the effect you’ll report is a weighted average
The solutions we evaluate are:
Nearly-optimal tests (onlinelibrary.wiley.com/doi/abs/10.3...) that target power near β₁₂ = 0 while guarding size.
Bias-aware CIs (arxiv.org/abs/2012.14823) or bounds (tinyurl.com/5n86u6a6, tinyurl.com/56mcxe2a) that incorporate credible limits on |β₁₂|
Nearly-optimal tests (onlinelibrary.wiley.com/doi/abs/10.3...) that target power near β₁₂ = 0 while guarding size.
Bias-aware CIs (arxiv.org/abs/2012.14823) or bounds (tinyurl.com/5n86u6a6, tinyurl.com/56mcxe2a) that incorporate credible limits on |β₁₂|
May 28, 2025 at 1:12 PM
The solutions we evaluate are:
Nearly-optimal tests (onlinelibrary.wiley.com/doi/abs/10.3...) that target power near β₁₂ = 0 while guarding size.
Bias-aware CIs (arxiv.org/abs/2012.14823) or bounds (tinyurl.com/5n86u6a6, tinyurl.com/56mcxe2a) that incorporate credible limits on |β₁₂|
Nearly-optimal tests (onlinelibrary.wiley.com/doi/abs/10.3...) that target power near β₁₂ = 0 while guarding size.
Bias-aware CIs (arxiv.org/abs/2012.14823) or bounds (tinyurl.com/5n86u6a6, tinyurl.com/56mcxe2a) that incorporate credible limits on |β₁₂|
What if we just test for interactions first? Beware: a data-dependent “model selection” (pre-test for interactions, then drop them if p>0.05) does not solve the problem. In fact, this two-step approach can itself generate invalid inference.
May 28, 2025 at 1:12 PM
What if we just test for interactions first? Beware: a data-dependent “model selection” (pre-test for interactions, then drop them if p>0.05) does not solve the problem. In fact, this two-step approach can itself generate invalid inference.
This has big implications for statistical significance. More than half of the main effects that were significant using the short model become insignificant once interactions are included.
May 28, 2025 at 1:12 PM
This has big implications for statistical significance. More than half of the main effects that were significant using the short model become insignificant once interactions are included.
What happens to the main results when we account for these interactions? In many cases, the findings change. The average absolute change in treatment effects was 0.05σ. The median percentage change was 96% and 26% of estimates flipped sign when estimated with interactions.
May 28, 2025 at 1:12 PM
What happens to the main results when we account for these interactions? In many cases, the findings change. The average absolute change in treatment effects was 0.05σ. The median percentage change was 96% and 26% of estimates flipped sign when estimated with interactions.
Were interactions truly negligible? Not really. We reanalyzed data from 15 of these papers, where data were available. While the median interaction was 0.00 (in standard deviations), the median absolute interaction was 0.07σ: about 37% of the size of the main effect.
May 28, 2025 at 1:12 PM
Were interactions truly negligible? Not really. We reanalyzed data from 15 of these papers, where data were available. While the median interaction was 0.00 (in standard deviations), the median absolute interaction was 0.07σ: about 37% of the size of the main effect.
How big is this problem in practice? Of the 27 factorial RCTs published in top-5 journals, 8 (30%) reported the fully interacted model. The other 19 (≈70%) relied on short models that ignored interactions, implicitly assuming any interaction effects were zero or negligible.
May 28, 2025 at 1:12 PM
How big is this problem in practice? Of the 27 factorial RCTs published in top-5 journals, 8 (30%) reported the fully interacted model. The other 19 (≈70%) relied on short models that ignored interactions, implicitly assuming any interaction effects were zero or negligible.
Ignoring interactions can indeed raise power when interactions are zero. But if interaction exists, the short model’s estimator is biased, and its test fails to control type I error. The model with interactions remains unbiased and correctly sized, at the cost of some power.
May 28, 2025 at 1:12 PM
Ignoring interactions can indeed raise power when interactions are zero. But if interaction exists, the short model’s estimator is biased, and its test fails to control type I error. The model with interactions remains unbiased and correctly sized, at the cost of some power.
First, we document that factorial designs are common in field experiments: 27 out of 124 RCTs published in top-five economics journals between 2007 and 2017 have a cross-cutting design.
May 28, 2025 at 1:12 PM
First, we document that factorial designs are common in field experiments: 27 out of 124 RCTs published in top-five economics journals between 2007 and 2017 have a cross-cutting design.