Equity Analytics

Statistical explanation and normative justification are entirely different operations. Controlling for rank, job, or experience explains variance — it does not endorse the processes that generated those distributions. The gap that "disappears" when you control for job title is not solved; it is a prompt to ask why jobs are gendered in the first place.

The determinants of pay, promotion, and evaluation — performance ratings, tenure, job title — are themselves generated by processes that may carry bias. Whether to adjust for them is a normative and strategic choice, not a purely statistical one. Treating control variables as exogenous systematically understates disparity.

Any algorithm, policy, or rule trained on data where inputs are correlated with identity will produce disparate impact — regardless of whether identity is explicitly included. Excluding a variable does not eliminate its influence; it only hides it. There is no criterion-based approach that avoids demographic consequences when applied to a structurally unequal organization.

Opposite biases across units can cancel in aggregate, producing a misleading zero net effect — a textbook instance of Simpson's Paradox. Enterprise-level parity at Elemental Systems masked significant division-level disparities. Always interrogate the distribution, not just the mean.

Differential treatment requires behavioral interventions (structured evaluation, blind review, bias training). Disparate impact requires process and policy redesign (revised criteria, alternative pathways, job redesign). Matching the wrong solution to the wrong problem wastes resources and can produce new inequities.

To address inequity, analysts must focus on the process that generates disparity over time rather than cross-sectional snapshots. The Data-Generating Process is the unit of analysis. Without career history data — who entered, at what level, how they were promoted, when they left — it is impossible to distinguish equitable sorting from biased sorting.

Observed differences can be produced through both fair and unfair processes. The analytical task is to distinguish which mechanism generates the disparity — not to assume the answer. A system can produce unequal outcomes equitably if the process is fair; it can produce equal outcomes inequitably if offsetting biases cancel.

Inequity can result from treating people unequally, but can also be addressed by doing so. Like staggered starting positions on a 400-meter track, treating people differently can be equitable when underlying structural conditions differ. Understanding the precise DGP allows analysts to design identity-conscious corrections.

In statistics, you know the DGP and predict outcomes. In equity analysis, you observe outcomes and must work backwards to infer the DGP — requiring both analytical rigor and theoretical imagination about plausible data-generating processes. You are an inferential detective, not a predictive modeler.

A program can be found lawful, shown effective, and still collapse if the institution running it cannot absorb the cost of defending itself. The Mansfield Rule was not killed by a court order; it was killed by the cost of legal defense. Program design must account for organizational resilience, not just legal compliance.