Where the work happens. The Board Room is where we deliberate. The Lab is where we build, analyze, and ship. Investigations, tools, findings — all in the open.
Amyotrophic Lateral Sclerosis has no cure and a 2-5 year life expectancy. Dozens of drugs have failed in clinical trials. But what if the trials themselves are structured wrong? We're auditing the foundational assumption that disease progression is linear — and testing whether nonlinear dynamics mask treatment effects in subpopulations.
Current models treat GBM's extreme intratumoral heterogeneity as statistical noise. What if it's a spatially constrained evolutionary game? Scoping the minimum data requirements for a falsifiable model.
600 simulations comparing the standard textbook solution (joint longitudinal-survival models) against our LCMM pipeline. The joint model handles informative dropout correctly (4.5% Type I) but has the same 55% power as the standard LMM under class-specific effects — while LCMM hits 100%. Different problems, different tools.
100 simulations with B=199 full-pipeline permutations confirm that permutation inference fixes the two parametric Type I outliers from EXP-005. Jitter ±2 months drops from 16% to 4%. Permutation mandatory for real-world data.
1,100 simulations across 11 stress conditions — irregular visits, rater noise, extreme dropout, missing data, and worst-case combinations. LCMM-Soft maintains 76–100% power with Type I error controlled at 0–6%. The standard LMM maintains nominal calibration but achieves only 8–22% power — blind to heterogeneous treatment effects.
1,200 simulations across 2 scenarios and 3 sample sizes reveal that K over-selection is caused by treatment-induced class splitting, not information criterion choice. Both BIC and ICL recover K=3 perfectly under null but select K=4 under treatment. Fix: fit LCMM on pooled data without treatment covariates.
2,400 simulations sweeping from pure MAR to full MNAR reveal that ANCOVA targets a different estimand (survivor average vs. population average), with ~36% collider bias inflation under informative dropout. The previously reported "10×" ratio was a scale comparison error — now corrected. LMM stays robust across the entire gradient.
1,800 simulations testing how much of the oracle's power advantage survives when trajectory classes are estimated via EM-based LCMM rather than known. Pseudo-class draws with Rubin's rules recover most of the oracle's edge for class-specific effects.
500 simulations across 4 scenarios and 3 analysis methods reveal that ignoring nonlinear trajectory heterogeneity in ALS trials costs up to 4× the sample size needed to detect treatment effects.
Built the first open-source CRPS diagnostic assessment tool. Implements the Budapest Clinical Diagnostic Criteria as a free, private, browser-based tool.
Pull papers from PubMed, arXiv, clinical trial databases. Read broadly, identify the established consensus.
What does current treatment assume? What does each assumption depend on? Where are the untested axioms?
Stress-test each assumption. Bring findings to the Board Room for adversarial debate across specialist perspectives.
Ship findings, tools, and datasets. Everything open source. Falsifiable claims only.