RSS ⧉

Pioneering Spirit (Blyg)

PAtwater

Some quick biographical context: 38, male, So Cal born, bred and still belong here, stereotypically spent a couple stints in the Bay / NYC / DC (also Philly!), once drove from Cincinnati to LA with a new employee at a crazy indie publishing venture, helped launch the CaDC, married, work as a data sophist of sorts, father of two girls (4 years and 8 months).

Mirror worlds

Reading Chetty's newest paper, I am struck by how that approach to causal inference through N=All administrative data plus calibration via RCTs provides a path to realize a big chunk of David Gelenters cyber vision.

[David Gelernter's Mirror Worlds (1991) imagines software models of real-world institutions, such as a city, a school, or a hospital, that mirror their ongoing activity in detail and in real time. Fed continuously by the data those institutions generate, these "mirror worlds" would sit on your desktop and let ordinary people watch, explore, and make sense of complex systems from the inside, rather than relying on expert summaries. The idea anticipates today's digital twins and the dream of seeing a whole system's behavior at once, which is what makes administrative data at N=All feel like a step toward his vision.]

I am excited to try that approach in water use efficiency. I still need to think through and mull the following AI plan.

[The California Data Collaborative's wue-conditions-explorer repo estimates water savings from conservation programs using a matched difference-in-differences design with six controls per participant. As the issue describes it, the repo currently checks the method by confirming that it recovers effects planted in synthetic data. That is necessary but not sufficient, since the planted effects are our own assumptions. The 2018 turf rebate paper by Schmitt, Tull and Atwater had the same gap: no experimental or quasi-experimental benchmark.

Issue 2 proposes closing that gap by borrowing Opportunity Insights' definition of a "digital twin": an observational estimator whose RMSE is comparable to or smaller than that of randomized trials, validated against experiments and valid only within the domain where it was tested. The plan has five steps:

  1. Fix a narrow domain and estimand, such as single-family turf replacement measured in gallons per square foot per year.
  2. Assemble a full-population panel of billing, rebate application and inspection dates, parcel and irrigable area, ET0 and weather, and ACS covariates.
  3. Keep the six-control matched DiD as the baseline and add a comparator, either staggered-adoption DiD or double/debiased machine learning with cross-fitting.
  4. Build a benchmark from quasi-experimental variation, such as rebate budget cutoffs, waitlists, rate changes across agencies or dates, or a randomized rollout order. Add placebo tests with fake rebate dates.
  5. Report RMSE against the benchmark, the validated domain, and the assumption that would let results generalize to other agencies.

The issue names two main threats. Drought-era restrictions may have affected adopters and controls differently, and savings per square foot vary locally. The main practical bottleneck is access to retailer data.]

I also want to reread the Chetty paper more slowly and reflect. I have a strong intution this approach will be very nontrivial.

Created: Nov 13, 2026