Intersectionality — What It Claims Empirically

Part 5 covered intersectionality as a theoretical position (see 5.6.2). This lesson does the other half: what it claims as an empirical proposition, how that is tested, and what the tests have found.

The results are more mixed than either its advocates or its critics usually report — which is the most useful thing about them, and exactly what this course's standards require reporting.

The case in which the harm fell between two categories.

In 1976 five Black women sued a car manufacturer in Missouri. Their claim was that its seniority system discriminated against them specifically as Black women.

The facts made the claim precise.

The company had employed women — white women — in office roles since before 1964. It had employed Black men in the plant. It had not hired Black women until 1970.

Then came a recession, and layoffs proceeded on a last-hired-first-fired basis. Because no Black woman had been hired before 1970, every Black woman was laid off. The women hired before 1964 kept their jobs; so did the men.

The plaintiffs argued the harm was specific to the combination.

The court disagreed, and its reasoning is why the case matters. It held that they could bring a claim as women, or a claim as Black people, but not a claim as Black women — writing that the plaintiffs should not be permitted to combine two statutory remedies to create what it called a new "super-remedy" unavailable to others.

And then each single-axis claim failed on the facts.

The sex discrimination claim failed because the company had employed women — the white women in the offices demonstrated that it did not discriminate on sex.

The race claim was consolidated with a separate action and did not proceed on the terms the plaintiffs had brought it.

So the women lost, and they lost precisely because the harm they suffered fell into the space between two legal categories, each of which the employer could show it had not violated.

That is not a metaphor about identity. It is a documented procedural outcome, and it is the concrete problem the concept was constructed to name: an analytical framework that examines one axis at a time can produce a finding of no discrimination where discrimination has plainly occurred.

Three claims, of increasing strength, and they need separating before anything can be tested.

The weak claim: single-axis analysis can miss things. Examining race while holding gender aside, or gender while holding race aside, can fail to detect a harm falling on a specific combination. The case above is a demonstration, and the claim is not seriously contested by anyone who has looked at it.

The medium claim: the effects are not additive. Being disadvantaged on two axes produces something other than the sum of the two disadvantages — usually more, on the "double jeopardy" reading. This is a statistical claim about interaction, and it is testable.

The strong claim: the categories are mutually constituted. Race and gender are not separable variables that combine; each is partly constituted by the other, so that "the effect of gender, holding race constant" is not a coherent quantity in the first place. On this view the variable-based approach is not merely incomplete but misconceived.

These require different responses. The weak claim requires disaggregation. The medium claim requires an interaction test and adequate power to run it. The strong claim is a challenge to the framework in which the other two are stated, and it cannot be tested within it — which is a genuine philosophical position and a real methodological difficulty (see 7.1.3).

Most arguments about intersectionality are conducted with participants holding different ones of these three, and it is why they go nowhere.

How the medium claim is tested, and why the tests are often bad

An interaction term, and a power problem that invalidates a great many null findings.

The additive model estimates a penalty for one characteristic and a penalty for another, and predicts that someone with both experiences the sum.

The interaction model adds a term allowing the combination to differ from the sum. If the interaction is positive, the disadvantages compound; if negative, they are less than the sum; if zero, they add.

And here is the technical fact that most of this literature ignores (see 7.6.3).

Detecting an interaction requires far more statistical power than detecting a main effect. As a rule of thumb, testing an interaction of the same magnitude as a main effect needs roughly four times the sample; testing an interaction that is half the size of the main effects needs something like sixteen times.

Which means that a great many studies reporting "no significant interaction" were not capable of detecting one. They are not evidence of additivity. They are evidence of an underpowered test — the absence-of-evidence error from 7.2.3, in a literature where it has real consequences.

And the problem is compounded by cell sizes. Testing many intersections requires groups defined by combinations, and combinations get small quickly. A national survey with ten thousand respondents may contain a few dozen people at some intersections, which is not enough for anything.

The best current method, and what it found.

The most promising recent approach handles the small-cell problem directly. Rather than estimating a separate effect for each intersectional group, it treats the groups as units within a multilevel model, allowing information to be shared across them so that estimates for small groups are stabilised — and, critically, producing a principled estimate of how much of the total variation in an outcome lies between intersectional strata at all.

It has now been applied to a range of outcomes across several countries — health, mental health, obesity, mortality, education, income.

Three findings recur, and they are what the honest reading of this literature has to accommodate.

Most of the variation in outcomes is within intersectional strata, not between them. People who share a race, a gender, a class and an age still differ enormously from each other. This is a finding about how much any categorical scheme explains, and the answer is: some, and much less than the discourse implies.

Of the variation that is between strata, the great majority is accounted for by the additive main effects. Knowing someone's position on each axis separately predicts most of what knowing their combination predicts.

And interactions exist, are generally modest, and are outcome-specific. They are real, they are not nothing, and they are not the dominant term.

How to read this proportionately. "Multiple jeopardy" as a general law — that disadvantages always multiply — is not supported. Nor is the claim that intersections are analytically unnecessary. The pattern is additive-dominant with real, modest, outcome-specific interactions, and the size and even the direction of those interactions differs by outcome.

And the interactions that have been found run in both directions, which is the part most often omitted.

Some are compounding, as the theory predicts. Audit studies varying more than one signal at once find combinations that are penalised more than either signal alone.

Some are sub-additive, and this is genuinely instructive. In several analyses of earnings, the gender penalty is smaller among some minority groups than among the majority — not because those women are advantaged, but because the men in the comparison have themselves been subject to a substantial racial penalty. The gap is smaller because the reference point is lower.

This is a real finding and it is easy to misuse in both directions. It does not show that those women are doing well; their absolute position is worse. And it does show that a framework predicting uniform multiplicativity gets the pattern wrong, and that comparisons must be specified carefully: a "gender gap" is always a gap relative to a particular group of men.

And a third pattern is documented that neither addition nor multiplication captures: invisibility.

Experimental and observational work finds that people at intersections are systematically less noticed — less likely to be remembered as having spoken, less likely to be treated as prototypical of either group they belong to, and less likely to be the subject of advocacy by organisations representing either.

This is the case at the start of this lesson, generalised. The harm is not that the disadvantages multiply. It is that the group falls out of the frame — of the data, of the category, of the remedy — which is a distinct mechanism and arguably the one the original argument was about.

How the framework is used in research

Three research strategies, and each answers a different question.

The intercategorical strategy takes existing categories as provisionally valid, constructs groups from their combinations, and compares outcomes across them. This is what the statistical work above does. It requires large samples, and it accepts the categories in order to document the relationships between them.

The intracategorical strategy studies a single intersection in depth — usually qualitatively — to show what a group's situation actually consists of, and how it is not captured by either single-axis account. This is where ethnography and interview research earn their place (see 7.5.1, 7.5.2), and it is how mechanisms get identified rather than measured.

The anticategorical strategy treats the categories themselves as the problem, and works to show their instability and their construction (see 8.4.1). This is the strong claim in practice, and its output is a critique of the classification rather than an estimate.

All three are legitimate and they are not substitutes. A great deal of confusion comes from evaluating one by another's standards — demanding effect sizes from an intracategorical study, or demanding thick description from an intercategorical one (see 7.7.2 on exactly this error).

Four criticisms, three of which have force.

One — operationalisability. The strong claim resists measurement by design: if the categories are mutually constituted and cannot be separated, then no analysis holding one constant is coherent. This is a genuine difficulty rather than an evasion, and the honest response is that the strong claim is a philosophical position about ontology (see 7.1.3) rather than a hypothesis, and should be argued as such rather than presented as an empirical finding.

Two — proliferation. If every combination of characteristics constitutes a distinct position requiring its own analysis, the number of positions grows exponentially and generalisation becomes impossible — and with it any basis for policy that applies to more than a handful of people. The standard response is that not every combination is socially consequential and the analysis should focus on those that are — which is correct and requires a criterion for which those are, which the framework does not itself supply.

Three — the empirical one, which is the strongest. The medium claim's central prediction is that disadvantages compound. The best-designed tests find additive effects dominating, with modest and outcome-specific interactions that sometimes run the other way. A framework whose distinctive quantitative prediction is not generally confirmed has to be more modest about that prediction — and the version that survives is the invisibility mechanism and the single-axis inadequacy claim, both of which are well supported.

Four — a criticism that does not have force, and it should be named. The objection that intersectionality amounts to ranking oppressions or determining who may speak is an objection to a use, not to the analysis. Nothing in the framework as stated is about standing to speak, and the original argument was about legal remedy and analytical adequacy. Criticising a body of work for its worst popularisation is the error 7.9.1 warns about, and it is committed frequently in both directions.

The legal record, which is the framework's home ground and where it has done least well.

The problem the case identified has been recognised in some jurisdictions and not others.

Some legal systems permit compound or combined discrimination claims, allowing a claimant to plead the combination directly.

Others do not, requiring each protected characteristic to be pleaded separately — which reproduces exactly the gap the original case fell into.

And one instructive case sits between. A major equality statute in the United Kingdom included a provision explicitly permitting claims of combined discrimination on two characteristics — drafted, debated and passed into the Act. It was never brought into force, and was subsequently abandoned, on grounds including complexity and business burden.

So the legal problem the concept was constructed to solve remains substantially unsolved in several major jurisdictions, four decades after it was identified — which is a fact about how analytical arguments translate into institutional change, and it belongs in this lesson more than any amount of theoretical elaboration.

Because the proportionate reading is more useful than either the enthusiastic or the dismissive one, and it is specific.

What is well supported.

Single-axis analysis can miss real harms, and the founding case is a documented instance. This alone justifies routine disaggregation, and it is cheap to do.

People at intersections are systematically less visible — in data, in categories, in advocacy and in remedy. This is the mechanism with the best evidence, and it is not a claim about effect sizes at all.

And interactions exist and are outcome-specific. Where they are found, they are informative about mechanism.

What is not well supported.

That disadvantages generally multiply. The best current methods find additive effects dominating, with modest interactions that sometimes run the other way — and much of the earlier evidence on both sides was too underpowered to say.

Four practical rules.

Disaggregate as a default, because it is cheap and the invisibility finding says it matters.

When you see a null interaction, check the power. Most such tests could not have detected the effect they report finding absent.

Specify the comparison. A "gender gap" is a gap relative to a specific group of men, and which group changes the number and sometimes the sign.

And separate the three claims before agreeing or disagreeing with anybody. Most of the public argument is people holding different ones, talking past each other, and both sides citing the same case.

One closing observation. The five women in Missouri lost. The concept exists because a court said their harm could not be named, and the analytical apparatus that followed is, at bottom, an argument that categories used for remedy should match the categories in which harm occurs. That is a modest and defensible proposition, it is well evidenced, and it has still not been implemented in much of the law it was addressed to.

The founding case: Black women laid off under a seniority system because none had been hired before 1970 were told they could sue as women or as Black people but not as Black women — the court declining to permit a combined claim it called a "super-remedy" — and each single-axis claim then failed. The harm fell in the space between two categories.

Three claims of increasing strength. Weak: single-axis analysis can miss harms — demonstrated, and uncontested. Medium: effects are not additive — a testable statistical claim about interaction. Strong: the categories are mutually constituted, so holding one constant is incoherent — a position about ontology that cannot be tested within the framework it rejects.

Testing the medium claim has a power problem that invalidates much of the literature. Detecting an interaction of the same size as a main effect needs roughly four times the sample; half the size, around sixteen times. Most reported null interactions are underpowered tests, not evidence of additivity — and combination cells get small fast.

The best current method treats intersectional strata within a multilevel model, stabilising small-group estimates and quantifying how much variation lies between strata at all. Three recurring findings: most variation is within strata; most of the between-strata variation is accounted for by additive main effects; and interactions are real, modest and outcome-specific.

Interactions run in both directions. Some compound. Some are sub-additive — a gender penalty smaller within a minority group because the men in the comparison have themselves been penalised, which means the smaller gap reflects a lower reference point rather than an advantage. And a third mechanism is neither: systematic invisibility at intersections, in data, categories, advocacy and remedy — which is the founding case generalised, and arguably what the argument was always about.

Three research strategies: intercategorical (comparison across combinations), intracategorical (depth on one intersection), anticategorical (critique of the categories). They answer different questions and are not substitutes.

Four criticisms. Operationalisability — genuine, and the strong claim should be argued as ontology rather than presented as finding. Proliferation — met by focusing on socially consequential combinations, which requires a criterion the framework does not supply. The empirical one is strongest: the distinctive quantitative prediction is not generally confirmed. And the objection about ranking oppressions is an objection to a popularisation, not to the analysis.

And the legal problem remains substantially unsolved — including in one jurisdiction where a combined-discrimination provision was drafted, debated, passed, and never brought into force.

Single-axis framework — analysis examining one protected characteristic at a time; the target of the original argument.

Compound / combined discrimination — a claim brought on the basis of two characteristics together.

Additive model — the prediction that combined disadvantage equals the sum of separate disadvantages.

Interaction — a departure from additivity; positive if compounding, negative if sub-additive.

Double or multiple jeopardy — the claim that disadvantages compound.

Sub-additivity — a smaller gap arising from a lower reference group rather than from advantage.

Intersectional invisibility — systematic under-recognition of people at intersections in data, categories, advocacy and remedy.

Intercategorical / intracategorical / anticategorical — comparison across combinations; depth within one; critique of the categories themselves.

Multilevel intersectional analysis — modelling strata as units to stabilise small-cell estimates and quantify between-strata variation.

Power for interactions — the substantially larger sample required, which invalidates most reported null interaction findings.

One — read the case. Find the 1976 judgment and read the paragraph about the "super-remedy". Note that the issue was procedural, not philosophical.

Two — check a null. Find a study reporting no significant interaction between two characteristics. Estimate whether its sample could have detected an interaction half the size of its main effects.

Three — specify a comparison. Take any published gender gap figure for a minority group and identify which men it is measured against. Then compute what it would be against a different group of men.

Four — look for invisibility. Take a policy area you know and ask which combination of characteristics is least represented in its data, its advocacy organisations and its remedies. That group is the one the framework was built to find.

Five — separate the claims. Take a public argument about intersectionality and identify which of the three claims each participant is making. Usually they are not making the same one, and the disagreement dissolves once that is visible.