Work
Worked examples
Each example takes one of the three questions from data to decision, on public or fictional data, so you can see the method before you hire it. None of it is client work.
Finished
-
Demonstration · Public data · Not client work
What happens next?
Who Leaves, Who to Save, and What It Costs
A retention playbook from churn analysis to the targeting decision: calibrated models, profit-based contact rules, uplift on a randomized experiment and a fairness audit.
Calibrated classification · Profit-based targeting · Uplift modeling · Fairness audit
Public and fictional datasets; findings checked against operators' public disclosures.
-
Demonstration · Public data · Not client work
Did it work?
Did Medicaid expansion reduce uninsurance among low-income adults?
A staggered difference-in-differences analysis of state Medicaid expansions and the uninsured rate among adults under 138% of the poverty line, with a pre-registered plan, robustness checks and sensitivity bounds.
Difference-in-differences with staggered timing · Callaway–Sant'Anna estimator · HonestDiD sensitivity bounds
Public data: Census American Community Survey microdata and KFF expansion dates.
-
Demonstration · Public data · Not client work
Can you trust the number?
Who files the most severe-injury reports? It depends on how you count employers
A step-by-step audit of OSHA's Severe Injury Reports: duplicates, name normalization and fuzzy entity resolution, with the matcher's precision and recall measured on a labeled sample.
Data-quality profiling · Entity resolution and duplicate detection · Precision and recall from a stratified labeled sample
Public data: OSHA Severe Injury Reports (federal-OSHA jurisdictions only) and company Form 10-K filings.
In progress
Unfinished. Listed so you know what's coming; nothing here is a result yet.
-
Demonstration · Public data · Not client work
Did it work?
When the Network Goes Dark
Gulf Coast cell-site outages and recovery in hurricanes, from FCC, EAGLE-I and NOAA public data.
In progress