Parallel trends in doubt? The Triply Robust Panel (TROP) estimator relaxes it and now has official Python and Stata packages. Python (pip install trop): github.com/zhaonanq/TROP-Est… Stata: github.com/justinwaddy/TROP Built by Zhaonan Qu, Xuelin Yang, Justin Waddy & Damian Clarke Why "triply robust": TROP combines unit weights, time weights and a regression adjustment based on a low-rank factor model. Its bias is bounded by the product of unit imbalance, time imbalance and misspecification of the regression adjustment, so it is asymptotically unbiased if any one of the three removes the underlying bias. In 20 of 21 simulation designs calibrated to widely used datasets (CPS, Penn World Table, and case studies such as the Basque Country and German reunification), TROP had the lowest RMSE of six estimators, including DID, synthetic control and SDID. With few treated units and periods, other estimators become more competitive. Both packages choose tuning parameters by cross-validation and report bootstrap standard errors and confidence intervals. Paper by @Susan_Athey (our Faculty Director), @guido_imbens, Zhaonan Qu and @VivianoDavide , now in the Journal of Applied Econometrics: doi.org/10.1002/jae.70061 (arXiv version: arxiv.org/abs/2508.21536)

An exciting week at the Golub Capital Social Impact Lab continues: Congratulations to our Faculty Director @Susan_Athey and to Hal Varian @halvarian named Clarivate Citation Laureates 2026 in Economics. Also named in economics this year: Michael Woodford for the theory and practice of monetary policy, and Sidney Winter @sidwindc for evolutionary approaches to organizational behavior.

Development progress has dipped to its weakest pace in 75 years. Can AI change that for low- and middle-income economies, and what has to be true for it to? Our Faculty Director @Susan_Athey is Co-Academic Lead of the World Development Report 2026: The Promise of Artificial Intelligence. Join us for an all-day event at Stanford GSB on Sep 24 as the writing team and experts present the findings. Speakers include @Susan_Athey and @ufukakcigit , @Gaurav__Nayyar, @AndrewYNg, @IndermitGill, @jblumenstock, Christopher Tonetti, Anja Sautmann, Elwyn Davies, Somik Lall. Program: gsb.stanford.edu/events/worl… Reserve your spot: eventbrite.com/e/world-devel… Livestream: youtube.com/watch?v=opvq4e0o… The Word Development Report: worldbank.org/en/publication… The program will present perspectives from the WDR 2026, the first comprehensive assessment of what AI means for the 6.8 billion people in low- and middle-income countries. Some themes: 1. The promise and the perils are both large. Adapted to local needs, AI could help do in a decade what might otherwise take a century. But what does it take to achieve the promise? 2. AI depends on complements: electricity, connectivity, literacy. Many countries are behind. In rural Sub-Saharan Africa, 32% of schools lack regular electricity and 68% lack reliable internet; 89% of 10-year-olds cannot read a simple text. 3. Yet AI can be adapted to reach people where they are. In Ghana, Rori delivers math tutoring by text message on basic phones and weak networks, producing nearly a year of learning for about $5 per student. Basic phones far outnumber smartphones in low-income countries. 4. AI can improve capabilities by supplementing scarce expertise. Sub-Saharan Africa has 2.9 physicians per 10,000 people; the global average is 16.7. Low-income countries have ~40 primary students per teacher vs 15 in rich countries. Farmers lack real-time information; most workers are generalists in tiny firms. 5. It is already happening. AI screening raised the number of patients checked for diabetic eye disease in Bangladesh by nearly 40% a day. Kenya's judiciary used AI to assign 10,000+ cases a year to 1,500+ mediators. AI weather forecasts in Telangana saved farmers up to $560 each. 6. AI is a general purpose technology, like steam, electricity and the internet. But AI is diffusing faster: poorer countries lagged 80 years on steam, 40 on electricity, 20 on the internet. Middle-income countries were half of ChatGPT traffic in 6 months. 7. But fast diffusion is not deep use. By April 2025, about 25% of internet users in high-income countries had adopted ChatGPT, vs 5.8% in upper-middle, 4.7% in lower-middle and 0.7% in low-income countries. Benefits could stay concentrated in richer countries at first. 8. Policy needs a theory of the pace of change. Change is a process with bottlenecks. Electricity paid off only after factories were redesigned; computers, only after workflows changed. Finding where and why change is slow is what lets you predict when impacts arrive. 9. Pockets of benefit can come fast when a specific application meets a specific need. Fitting AI into complex institutions like schools and health systems takes years: adoption runs from information to selection, piloting, evaluation and rollout, each needing scarce time and money. 10. For many developing countries, job disruption will likely be gradual. About 4.5% of jobs in low- and middle-income countries are amenable to automation by generative AI, vs 14.2% in high-income ones. Nearly 90% of workers in low and lower-middle income countries are in firms under 10, where augmentation with AI can improve decision-making and assist with administrative tasks. 11. Some effects are fast. Multinationals can pull back outsourced white-collar work quickly: one large freelance platform reported jobs outsourced to developing countries fell 39% in 2025 (the evidence is mixed). Fraud and cyberattacks face few bottlenecks for bad actors. 12. Concentration matters. Frontier models, chips and data centers are controlled by a few firms in a few countries, mostly the US and China. The largest US AI firms are expected to spend over $750 billion on AI infrastructure in 2026, more than the GDP of many developing countries. 13. Developing countries can build on existing models rather than train their own. GPT-3.5-level capability fell from ~$20 to $0.07 per million tokens in two years, and open-weight models (Llama, Qwen, Gemma, Mistral, DeepSeek) can be adapted locally and optimized for local languages or low-resource environments. 14. Dependence is the risk. The report argues against full AI sovereignty (very costly, and a local data center still runs on foreign chips) and for interoperability: source from many countries and keep layers swappable. India's UPI and Brazil's Pix show how shared rails do this. 15. The World Bank Enterprise Surveys on AI (India, Jordan, Kenya, Mexico, Nigeria, Thailand, plus the US) find that about 1 in 5 small firms in the six developing countries already use AI chatbots, vs 1 in 4 in the US. Use is shallow so far: search, writing, translation. 16. What holds firms back? Firms that do not use AI cite lack of knowledge first, then cost. Adopters are firms with better management, prior innovation and better market access. Roughly 90% of small firms already use messaging apps, so the entry point exists. 17. Governments shape AI in three roles: as enablers (building the complements), as users (buyers for locally adapted AI), and as regulators. 18. The report's country strategy framework: adapt AI to local languages, laws, infrastructure, and data, which makes adoption pay off; adopt off-the-shelf tools where they work; for a few countries, advance the frontier. 19. New survey of 57 governments on AI use: front-line barriers are skills and connectivity; the back-office barrier is data, often on paper or scattered across agencies. Many lack clear information on which tools to choose and how to buy and implement them. 20. An analysis of 6 million procurement contracts: government AI in low- and lower-middle-income countries is overwhelmingly predictive (decision support, geospatial, early warning), not generative. 21. Buying AI is hard for governments and funders because AI products improve through iteration, not one-time delivery. The report suggests buying against performance benchmarks (e.g. WHO's TB screening standard), procurement built for ongoing improvement, and independent evaluation. 22. Procurement can also pull local solutions into existence. The report proposes advance commitments to buy AI that meets a performance bar, modeled on vaccine purchase commitments, especially in health, education and justice, where poorly adapted AI can do real harm. 23. The "too many pilots" problem: AI coding tools make prototypes cheap, but evidence on what works is thin. Of 9,762 studies of AI health tools, under 1% were in low-resource settings and only 17 give causal evidence on outcomes. Funders can't choose; developers can't scale. 24. Preparation is uneven. 80+ countries have national AI strategies, but only 1 of 25 low-income countries (Rwanda). The countries with the most to gain look least prepared. The report's message: tune out both the hype and the hysteria, and start now.

In our theory we show how the estimator’s bias depends on product of imbalance across units, time and bias of the regression adjustment, so that effectively only one of these three needs to be small. This gives flexibility in the choice of unit, time weights and outcome model.

In contrast, TROP embeds all these estimators as special cases, while it learns which components of such estimators are most relevant to accurately predict the counterfactual. We evaluate TROP through a large set of simulation studies calibrated to match real-world applications.

Why this matters: TROP learns the combination of time unit weights and regression adjustments that most accurately predict the counterfactual. Common estimators of causal effects in panel data (DID/TWFE, SC, MC, SDID) rest on different assumptions but all can fail in applications

New paper: Triply Robust Panel Estimators (TROP) by @Susan_Athey @guido_imbens Zhanonan Qu @VivianoDavide. TROP estimates causal effects in panel models by combining a flexible outcome model (regularized low-rank factors + FE), unit weights, time weights arxiv.org/pdf/2508.21536

Listen to Raj Chetty talk about how surrogate indices make it possible to make decisions more quickly using multiple short-term outcomes to predict long-term effects with @nberpubs. nber.org/research/videos/202…

Watch @Susan_Athey’s talk on the implications of AI on the organisation of industry and work at #G20SouthAfrica youtube.com/watch?v=okDGsZak…

AI & digitisation are rapidly reshaping the way we work. Policymakers need to understand how, and what to do about it. Watch @Susan_Athey speak to G20 leaders about these issues tomorrow 16 July @ 13:30 CET. #G20SouthAfrica bit.ly/3GyMFgm or bit.ly/44PTXFP

Analyzing representations tells us where history explains the gap. Ex: there are two kinds of managers: those who used to be engineers and those who didn’t. The first group gets paid more and has more males than the second. Models that don’t use history omit this distinction.

We use these methods to estimate wage gaps adjusted for full job history, following the literature on gender wage gaps. History explains a substantial fraction of the remaining wage gap when compared to simpler methods. But there’s still a lot that history can’t account for.

This result motivates new fine-tuning strategies. We consider 3 strategies similar to methods from the causal estimation literature. E.g. optimize representations to predict the *difference* in male-female wages instead of individual wages. All perform well on synthetic data.