The Population You Cannot See.
Claims data records encounters that happened. The women who were dismissed, who lost trust in the clinician, who ran out of money to keep trying, are structurally invisible. Every fiscal model built on the claim record undercounts them.
Billed visits, dispensed prescriptions, completed procedures. Every record produced by a formal, compensated clinical interaction.
The appointment that ended with "it is probably just stress." The patient who did not rebook. The woman who stopped going. The specialist referral that never got written. None of it appears in any claim, because no claim was ever filed.
The measurement problem, stated plainly
A claims database is an encounter log, not a population log. It is an ordered list of billed clinical events. That is the record the hospital generates, the payer reimburses against, the state aggregates for fiscal planning, and the academic epidemiologist eventually analyses. It has one structural limitation that is not a limitation of any specific database but of the category of database: it records what happened, and therefore cannot record what did not.
For most of medicine, the absence is tolerable. An acute event enters the record because the event forced itself into the record. A patient with a myocardial infarction arrives at the emergency department because he cannot not. Ambulatory care, chronic disease, and diagnostic odysseys do not work this way. They are elective. The patient chooses whether to continue engaging. The record only captures the patients who chose to.
The asymmetry of engagement is sex-stratified. Women are statistically more likely to be dismissed at the initial presentation (the 7-to-10-year diagnostic delay in endometriosis, the 4-year average delay across 770 diseases documented by Westergaard et al. 2019). Dismissal in turn reduces the probability of continued engagement. Patients who are told they are anxious do not reliably come back. Patients who come back and are told again, more definitively, that they are anxious reliably stop coming back. The system produces a self-erasing cohort.
Why this collapses every claim-based fiscal model
Every fiscal estimate in the FemTechnology Intelligence Suite, and in every comparable body of work, uses published per-patient cost deltas multiplied by a patient count. The per-patient numbers are peer-reviewed. The patient count is the problem. The Soliman 2018 $10,002 excess cost per endometriosis patient-year is a measurement of patients who were eventually diagnosed. The women who were dismissed out of the pathway at year one, three, or five, and who never returned to receive the diagnosis, do not appear in the Soliman denominator. They still consumed clinical care, generated administrative claims for indeterminate pain, received inappropriate prescriptions, and accrued disability. They were counted against the wrong disease category, if they were counted at all.
The three examples above are not controversial in the clinical literature. Each is documented, in the respective condition's specialty journals, as a large and persistent gap between what populations self-report and what claims databases capture. The consistency of the direction (always a claims undercount) is the evidence that the measurement architecture, not the self-report, is what is wrong.
Funnel values are illustrative of the combined attrition described in the endometriosis diagnostic-delay literature (Soliman 2018; Staal 2016; Ballard 2006). Stage-specific attrition rates are condition-dependent; the shape of the cascade is not.
The three failure modes that produce ghosts
- Dismissal at point of entry. The clinician determines the symptom does not meet threshold, attributes it to anxiety or stress, and closes the visit without referral or diagnostic workup. The encounter appears in the record as an unspecified complaint (R codes). The underlying condition never receives a specific diagnostic code.
- Abandonment after repeated dismissal. The patient, having received the above response two or three times, stops rebooking. She becomes functionally invisible. If she is absorbed into alternative medicine, self-management, or simple endurance, no claim is generated. Her condition is still progressing.
- Affordability attrition. Even in jurisdictions with universal coverage, the real cost of diagnostic persistence (time off work, childcare, transport, specialist co-pays) is high enough that women in lower-income deciles disengage from the pathway at significantly higher rates. The class of patient who drops out is systematically different from the class who does not.
The measurement instrument already exists
The correct instrument is the Patient-Reported Experience Measure and Patient-Reported Outcome Measure, collected from populations directly rather than via clinical encounter. PREMs and PROMs are an established methodology in health-services research. The historical obstacle to scaling them was the cost of administration. That obstacle has collapsed in the last three years.
Conversational AI systems can now conduct structured, empathetic, multi-turn interviews with very large respondent populations at essentially linear marginal cost. The ORI prototype architecture (see pipeline index, Patient Voice module) demonstrates the technical feasibility. The scale of the addressable measurement, a population of tens of millions of women carrying dismissed diagnoses, is exactly the scale AI-mediated interviews are designed to reach.
The policy and payer move
For state health departments, Medicaid managed-care plans, and commercial payers, the immediate action is to commission a paired-cohort PREMs survey against an existing claims extract for one condition, in one jurisdiction, in one year. Three instruments, identical respondents, cross-linked by a de-identified key.
- A claims-based prevalence count.
- A PREMs/PROMs self-report extraction via AI-mediated interviews.
- A reconciled delta that quantifies the ghost population by condition.
The output of that exercise is a corrected true-prevalence denominator for every fiscal model that department runs. It is a one-time methodological investment that permanently improves every future cost model, reform evaluation, and public-health priority-setting exercise the jurisdiction undertakes.
The women the claim record does not contain are not a rhetorical device. They are a measurable population, and until they appear in the measurement, every calculation of the cost of the gender health gap is a lower bound.
Related reading: The Dark Matter of Culture for the cultural substrate, Architecture of the Gender Data Gap for the upstream research-design failure.