Nakul Sathaye

While working on a personal project, I stumbled upon an interesting case. It starts with the dataset GHG Emissions of Chicago's buildings and the task is simple: predict the GHG emissions. But this soon turns out to be a very interesting case about how understanding the Data Generation Process (DGP) can help solve the problem of structural missing data caused by selection bias. This is what I regard as compliance-gated data: the data is generated from a selection process where a unit is selected only when its observed features exceed chosen thresholds. In this case, only those buildings larger than 50,000 sqft are required to report their emission statistics to the authority. The catch is easy to miss without knowledge of this selection process: the observed data isn't a representative sample of all buildings, but rather a biased sample of only those that complied. This characteristic isn't a result of the sampling process but of the DGP itself — so a better model or a better pipeline won't correct the bias unless it integrates the DGP. This is an interesting case to analyze using my familiarity with structural causal models, but it isn't the point — it's just a demonstration. The real idea: a lot of our data isn't a neutral snapshot, it's produced by a documented process — a regulation, an accounting standard, an eligibility rule. That document doesn't just describe the data, it constitutes it. So knowledge of the DGP can be extracted from these documents and integrated into the analytical pipeline. This idea is developed further in my working paper, and I also built an app to demonstrate the process in action. 📄 Paper: https://lnkd.in/e4anvdZZ 💻 App: https://lnkd.in/eu3-32V6 --- A note: this is still a developing idea, and I'm sharing it precisely to think it through in the open. I'd genuinely appreciate any feedback, critique, or comments on the paper — whether on the framing, the method, or cases where you've seen data shaped this way. Thanks for reading. #DataScience #CausalInference #MachineLearning #BayesianStatistics

12

As experimenters, we’re increasingly asked to measure what cannot be randomized. Partnerships. Long-term effects. Post-launch policy evaluation. In this piece, Ross Chu and I share how we validated and productionized AIPW (doubly robust estimation) inside Lyft’s quasi-experimentation platform — and what we learned when we compared observational estimates to experimental IV ground truth. A few key insights: - ⚖️ Doubly robust ≠ automatically unbiased. - 🧠 Confounder rigor is non-negotiable. - 📏 Hidden confounding must be quantified, not ignored. If you're thinking about scaling causal inference beyond A/B testing, this might resonate. 🔗 https://lnkd.in/gF27Qs7M

164