Skip to content

Intersectional Stratification Table

Tool — instantiates Vulnerability Hotspot Mapping and Hardening

Cross-tabulates an outcome across intersecting attributes so the subgroup where several disadvantages coincide appears instead of being washed out by the average.

An Intersectional Stratification Table finds the hotspot in category space rather than physical space: it takes a harm outcome and cross-tabulates it across the joint combinations of several attributes, so the cell where disadvantages intersect — not the marginal totals for each attribute alone — is what you read. Its defining discipline is refusing to average: a group can look fine on every single axis yet be acutely at risk at the intersection of three of them, and only the joint table exposes it. That distinguishes it from the Layered Risk Heatmap, which maps co-location across a spatial grid, and from a vulnerability index, which collapses strata into one score — this tool deliberately keeps them apart so the worst cell stays visible.

Example

A manufacturer's overall recordable-injury rate looks acceptable and steady. A safety analyst builds a stratification table crossing three attributes: shift (day / night), tenure (under one year / over), and employment type (direct / subcontracted). The single-axis views are reassuring — night shift is only slightly worse than day, new hires only slightly worse than veterans. But the joint cell — night-shift, first-year, subcontracted workers — shows an injury rate several times the plant average, on a headcount small enough that it never moved the overall number. Every marginal average had hidden it. The table turns "we're within tolerance" into "there is one specific crew configuration carrying the risk," which is a hardening target a rate alone could never have named.

How it works

The mechanism's distinct step is joint stratification, not the tabulation itself:

  • Fix the outcome. Choose the harm or failure metric and how it's expressed (rate, severity, count) so every cell is comparable.
  • Enumerate the attributes as the sensitivity axes. The columns and rows are the susceptibility inventory — the dimensions along which units differ in vulnerability.
  • Cross to the intersection cells. Populate the joint combinations, not just row and column margins, and read the worst cell — the intersectional hotspot the average dissolves.

The point of the tool is everything the marginal totals throw away.

Tuning parameters

  • Attributes crossed — more axes resolve finer subgroups but split the data into ever-sparser cells; each added dimension trades resolution for stability.
  • Cell-size floor — the minimum count below which a cell is suppressed or merged. Too low and a single event spikes a rate; too high and the smallest, often most vulnerable, groups disappear.
  • Outcome expression — rate vs. absolute count vs. severity-weighted. A rate spotlights small high-risk groups; a count spotlights large ones.
  • Margin vs. joint emphasis — how hard the table steers readers to the intersection cells rather than the comforting marginal totals.

When it helps, and when it misleads

Its strength is catching the subgroup that both a single-variable analysis and the grand average conceal — the classic case where an aggregate trend reverses or vanishes once you disaggregate (Simpson's paradox). Naming the exact intersecting configuration converts a diffuse "some workers are at risk" into a precise hardening target.

Its failure modes cluster at the sparse cells. Slice finely enough and some intersection will look extreme by chance; a table with many thin cells invites reading noise as signal, so a striking cell still needs statistical validation. It is also easy to run backwards — keep adding or regrouping attributes until a cell appears that flatters or damns a predetermined conclusion, the "garden of forking paths." And the attributes are not neutral: a table can reify crude categories as if they were the mechanism of harm. The discipline is to pre-specify the strata and cell-size floor before looking, and to treat the worst cell as a lead to validate, not a verdict.[1]

How it implements the components

Intersectional Stratification Table realizes the outcome-and-susceptibility side of the mapping machinery — the categorical layers, not the spatial or interventional ones:

  • harm_or_failure_metric — defines the outcome and computes it per cell so subgroups are directly comparable.
  • sensitivity_layer_inventory — the crossed attributes are the susceptibility inventory, held disaggregated rather than summed.

It does NOT map co-location spatially (Layered Risk Heatmap) or collapse the strata into a single composite score (Vulnerability Index Construction); it does NOT test the worst cell's significance (Multiple-Testing Holdout Check) or check that the targeting it invites is equitable (Equity Impact Review).

References

[1] The lens of examining overlapping attributes jointly rather than one at a time is intersectionality, a term coined by legal scholar Kimberlé Crenshaw to describe how disadvantages compound at the intersection of categories. Used here as the analytic frame for reading joint cells — not as a claim about any particular dataset. ↩