What the discipline has produced.
After teaching mathematics, I’ve spent 24 years in modeling, simulation, and analysis (MS&A), primarily for national security. The independent studies below test how Analysis Driven Modeling — a methodology developed through that work — translates to other consequential domains.
← The Mission Professional background →By Michael Key · ORCID
The method, tested on real problems.
These independent 2026 projects were completed outside my professional roles. They test where general MS&A skills and Analysis Driven Modeling transfer, where they fail, and where domain knowledge or higher fidelity changes the answer. They are not client work or claims of long professional experience in each subject area. The studies closest to the mission come first.
Human-cohort & health studies
The method on real human-health data — the nearest neighbors to the Alzheimer’s and ALS work. New studies aimed closer to cognition and aging will join this group as they’re finished.
What predicts how long we live?
Thirteen investigations across NHANES, CHNS, and HRS, with question-specific samples ranging from 1,907 to 39,839 people. Richer learned models help on some questions; simpler or domain-led models are enough on others. The right method depends on the question — the same judgment the neurodegeneration work turns on.
Longevity study →California freight & ozone
54 current investigations and three withdrawn records retained for audit, including a 705,610-person health-effects cohort. In the modeled T2 ozone-health channel across the LA, San Joaquin, and Sacramento basins, only 0.05% of 24,576 uncertainty draws produced net benefit; the rest produced net disbenefit. That is narrower — and more useful — than claiming every NOx cut raises ozone.
RFAQ study →The method on other domains — eleven studies · energy, water, environment, emergency response, education
Can the grid absorb the data-center surge?
A current capacity-auction ledger, nine historical demand years, and bounded Elliott and Dominion stress checks. The answer is conditional: new load needs accredited, deliverable supply and a tested flexibility fallback.
PJM study →Military base groundwater
The screening model cannot choose a remedy. MODFLOW makes pump-and-treat plausible; a finite Monte Carlo ensemble tests that choice across the encoded uncertainty and asks whether more characterization is worth buying.
PFAS study →Evacuation & insurance
Four real wildfires, NASA FIRMS satellite data, USGS elevation, NIFC perimeters. Right model shifts evacuation triggers 1–7 hours earlier among the 9 of 16 communities with a positive modeled gain. Of the remaining seven, six had no deterministic trigger and Centerville’s two triggers coincided.
Wildfire study →Can the river keep its promises?
Legal over-allocation vs. physical supply. Twelve investigations. Global sensitivity shows the delivery obligation drives just 7.3% of breach risk — inflow drives 92.9%.
Colorado River study →Offshore wind & right whales
Pile-driving noise propagation vs. NARW critical habitat. Eight investigations, three sites, real SSP profiles. Coordinated scheduling cuts total whale takes 3.3× and creates 9 quiet hours/day — at no added cost.
Ocean acoustics →Texas grid resilience
Five questions about the grid that failed in Winter Storm Uri. Physics-based dispatch, investment tradeoffs, location sensitivity.
Texas energy →Where should a data center go?
Where should a large data center go when transmission is the binding constraint? Spatial optimization over grid physics and siting cost.
DC Siting →Where do the charging gaps fall?
Where do the coverage gaps fall, and how big should each station be? Network planning under grid constraints, with battery-storage tradeoffs.
EV Charging →How much fidelity does the decision need?
3 fires, 4 fidelity levels, 200 Monte Carlo draws. Three viability gates, two interactive tools. How fidelity affects the insurance decision.
Wildfire insurance →College closure risk
The current public output covers 1,516 real-data institutions; a separate 4,344-case synthetic cohort tests the models under known outcomes. College Scorecard and IPEDS inputs, three model tiers, and three interactive tools.
College closure →Is your degree worth it?
65,935 college programs analyzed with real IRS earnings data. 34% have negative lifetime ROI. Data analysis, not modeling — but the same question-first discipline applied to a hard, consequential question.
College ROI →