screen shot of hierarchical data

We’re Asking Our Customers “What If Your Data?”

What if your data could generate predictions that made a material difference to your business: actuarial insights that are actionable, patient behavior, provider trends, predictions of new high-cost claimants? Building predictive machine learning models requires significant infrastructure and capabilities, not to mention patience. As we talk to our customers, we find there are common hurdles that are slowing read more…


Optimizing Individual-Level Models for Group-Level Predictions.
Part 2 — The Median of a Sum of Independent Identically Distributed

Introduction In Part 1 — An Analysis of Bias I explored bias in group-level predictions coming from individual-level models. There I used the following theorem by Peter Hall (“On the Limiting Behaviour of the Mode and Median of a Sum of Independent Variables”¹): But why is that true? I found Hall’s paper to be somewhat difficult to read more…

Photo with Title: Optimizing Individual-Level Models for Group-Level Predictions. Part 1 — An Analysis of Bias

Optimizing Individual-Level Models for Group-Level Predictions
Part 1 — An Analysis of Bias

Introduction The goal of this post is to present some of my thoughts about a very common, yet scantily addressed, problem in machine learning at large (and healthcare in particular): how do you construct group-level predictions given only individual-level data in a way that optimizes a pre-determined loss function (e.g., MAE or MSE)? If you read more…