The number
In the 2020 U.S. Census, 49,874,246 people checked "Some Other Race." A 129% increase from the previous count.
They are mostly Hispanic Americans whose self-understanding does not separate ethnicity from race the way the form does. The Census asks two questions: one about ethnicity (Hispanic or not) and one about race (White, Black, Asian, American Indian, and so on). The form treats these as independent dimensions — you answer one, then the other, as if they are separate facts about you. For 49.9 million people, they are not separate facts. They are the same fact, and the form has no box for it.
The form determines who counts. Change the form, change the count. This is not a metaphor. It is what happened. The 2020 re-coding did not discover 49.9 million people who had been hiding. It created a visible population by changing how the instrument registers identity. The people were always there. The form could not see them until it changed, and even after changing, it saw them only as an anomaly — a spike in a residual category, a box that was never supposed to be this full.
This is not a Census problem. It is a classification problem. Every form has boxes. Every box has edges.
The form
A form is not a mirror. It is a machine with built-in parameters — what it measures, how finely it measures it, what counts as the same thing, where the line falls — and those parameters determine what can be seen. What falls outside them does not register as wrong. It does not register at all.
The Census form separates ethnicity from race: two dimensions, two questions, checkbox format. The assumption built into the instrument is that these are independent aspects of identity. If your experience of identity does not separate them — if "Latino" is not an ethnicity layered on top of a race but a single way of being in the world — the form has no way to record what you are telling it. You check "Some Other Race" and write in what you mean, and the form files you in a category that was designed to be small.
The DSM operates a different form on a different population with the same structure. Major depressive disorder: five of nine symptoms, two weeks, functional impairment. That is the line. Fourteen days of symptoms is diagnosed. Thirteen is not. On one side of that line: a name for what you are experiencing, a treatment protocol, insurance coverage. On the other side: nothing official. Not because you are well. Because the form cannot see you yet.
Different instrument. Same architecture.
The U.S. poverty threshold is a number that varies by household size, but in practice functions as a single cliff. Federal programs peg eligibility to percentages of it — 130% for food assistance, 138% for Medicaid in expansion states. Fall just above the relevant cutoff and you qualify for nothing. The "near-poor," people clustered around these thresholds, experience the same deprivation but exist in no administrative category. Too poor to be comfortable. Too rich to be counted as poor.
HUD's annual Point-in-Time count measures homelessness by observing who is in a place not designed for habitation, or in emergency shelter, on a single night. If you are sleeping on a friend's couch, cycling between relatives, doubled up in an overcrowded apartment — you are not counted. Federal funding follows the count. The resources go where the numbers are. The people the instrument cannot see do not receive help designed around numbers they were never part of.
Every one of these instruments has built-in parameters that determine who appears and who does not. The parameters are not neutral. They are choices — about what to measure, how finely, where to draw the threshold. And the choices determine the result. Not by biasing it. By constructing it.
The fit
Hold a classification in place long enough and the world reshapes around it.
The DSM's diagnostic categories were designed as clinical tools — operational definitions meant to standardize communication among clinicians. But the categories are also the codes that insurance companies require for reimbursement. So clinicians learn to code toward the categories that get covered. Patients learn to describe their experience in the instrument's language — the only language the system can hear. Pharmaceutical companies organize research, marketing, and drug development around the diagnostic labels. The label attracts funding. The funding produces evidence. The evidence validates the label.
The category starts to look real — not because it was always real, but because it has made itself real. The distinction produces the population it appears to describe.
This is not conspiracy. It is not corruption. It is what happens when a measurement instrument is embedded in the thing it measures. The Census categories shape how people understand and report their own identity. Representation, redistricting, and resource allocation all flow through those boxes. People learn to navigate them — checking multiple boxes, choosing strategically, adapting self-description to the form's logic. The form does not passively record identity. It actively shapes it, and then records the shape it has produced.
The poverty line was designed for administrative convenience — a single threshold to determine eligibility. But once food assistance, Medicaid, and housing subsidies are pegged to it, the line is no longer just a measurement. It is a cliff. People cluster around it. Institutions organize programs around it. The line appears to describe a real boundary in the distribution of need because it has produced one.
The category appears to track real joints in the world because it has produced the joints it appears to have found.
This is the deeper problem. The form is not merely inaccurate. It is generative. It makes things. And because it makes things that look like the things it was supposed to find, the fact that it made them is invisible. Institutional success is mistaken for truth.
The revision
So we revise. We update the form, improve the categories, learn from the boundary cases. And the boundary cases move.
DSM-5 removed the bereavement exclusion — previously, you could not be diagnosed with major depression if your symptoms followed the death of a loved one within two months. Removing it resolved one boundary population: grieving people with genuine depression could now be diagnosed. It immediately created a new dispute: is normal grief now being medicalized? Are we pathologizing sadness? Each revision addresses some hard cases and generates others. The hard cases are not fixed. They are redistributed.
The 2020 Census re-coding was itself an attempt to improve — to give respondents more room to describe themselves. It succeeded in one sense: the 49.9 million became visible. But the fundamental mismatch — ethnicity and race treated as separate dimensions — remains. The 2030 Census is set to combine the questions into a single item. This will resolve the "Some Other Race" anomaly. It will produce new boundary cases, new populations that the combined question cannot capture, new people in a different wrong box. The box changes. The structure does not.
California's ABC Test tightened the definition of "employee," reclassifying many gig workers from contractors to employees. Proposition 22 then carved out a new category — "app-based driver" — institutionalizing the gap rather than closing it. Neither resolved the mismatch between how platform work actually functions and how employment law classifies it. They relocated the boundary. The people at the edge are still at an edge. It is just a different edge.
The rock does not respond to being classified as igneous. The person classified as "near-poor" can, and does. Human populations push back. They organize. They articulate why they do not fit. The 49.9 million people who checked "Some Other Race" are not passively miscounted — they are actively telling the Census something the Census cannot hear. And when enough people say the box is wrong, the box eventually has to move. But moving the box does not solve the problem. It draws a new edge, and new people find themselves on it.
Each revision is presented as progress. Sometimes the accuracy improves. But it is never finished. The form always has edges. The edges always leave someone out.
The question
The standard question about a classification is: is it accurate? Does it describe reality?
This is the wrong question. Not because accuracy doesn't matter, but because it assumes there is a form that could, in principle, get it right — that the gap between the category and the world is a defect to be repaired rather than a structural feature of categorization itself. Any form with finite parameters, applied to the continuous grain of human experience, will leave someone out. You can draw the line more carefully. You cannot draw it without cutting.
The better question is: who disappears when you draw the line here?
Every classification could be preceded by a different question: not is this accurate? but what will this exclude? How will the excluded population respond? Where will the pressure build?
There is no correct form. There is the question of what each form leaves out, and whether what it leaves out is treated as information or as noise.
The 49.9 million people who checked "Some Other Race" are not confused. They are not edge cases. They are not anomalies to be resolved by the next revision. They are telling the Census something about how identity works, and the Census cannot hear it, because hearing it would require a form the Census does not have.
