# When the Map Doesn't Predict the Outcome: Phase B of the Youth Mental Health Access Gap

*A long-read working-paper summary from Trellison Institute. May 2026.*

---

Last week we published the Youth Mental Health Access Gap V1 — a state-level need-vs-access analysis of 35 U.S. states and 41.8 million under-18 residents. The headline finding was that 39.4% of high-school students report two weeks or more of persistent sadness or hopelessness in the past 12 months — more than two-and-a-half times the adult rate — and that the state-to-state variation in youth-serving mental-health provider supply is approximately 85-fold. The analysis identified three "positive outliers" (Puerto Rico, North Carolina, New Jersey — supply gap worse than the insurance landscape predicts) and two "negative outliers" (Vermont, Alaska — supply better than predicts).

The framework places a state in a residual class. The natural follow-up question is: does that placement *predict* anything in the real world? Do positive outliers have worse mental-health outcomes? Do the negative-outlier states (well-supplied for their insurance landscape) actually have lower suicide rates?

We tested this in Phase B by joining the gap measures with three categories of state-level outcome data: CDC YRBSS state-level youth suicide ideation/planning/attempt prevalence, NCHS state-level all-age age-adjusted suicide death rates (2017, the latest year available via Socrata), and NCHS provisional drug overdose deaths per 100,000 (2023).

The findings are nuanced enough that we lead with them rather than the answer.

---

## Strong: within-instrument coupling

The strongest signal in the entire analysis is that **state-level distress prevalence is a robust cross-sectional proxy for state-level youth suicide indicators**. The need metric — the share of high-school students reporting sad/hopeless for 2+ weeks in past 12 months — correlates with the YRBSS suicide questions at:

- r = +0.82 with "seriously considered suicide"
- r = +0.80 with "made a plan"
- r = +0.76 with "actually attempted suicide"

This is a high-confidence finding. The within-YRBSS correlation matrix is the strongest empirical result of Phase B and it tells us the Youth V1 paper chose the right need signal. State-level sad/hopeless prevalence is a meaningful proxy for the more severe state-level youth suicide indicators.

For state-level all-age outcomes the signal degrades. The same need metric correlates only +0.34 with the NCHS all-age suicide age-adjusted death rate, and +0.19 with the drug overdose rate. These are real correlations but weaker than the within-instrument relationships. Two factors drive the degradation: the demographic mismatch (YRBSS is grades 9-12; AADR is all-age), and the intrinsically lower signal-to-noise ratio of low-frequency mortality events at state aggregation.

---

## Null: provider supply does not predict reduced state-level mortality

The most policy-consequential finding is what *isn't* there.

The framework's supply-side metric — `access_value`, the count of youth-serving NPI providers per 100,000 under-18 — shows essentially no protective correlation with any outcome:

- YRBSS considered suicide: r = -0.030
- YRBSS made plan: r = +0.000
- YRBSS attempted: r = +0.109
- All-age suicide AADR: r = +0.143
- Drug overdose rate: r = +0.211

The drug-OD coefficient is *positive*. Higher-supply states have higher drug-OD rates.

The Alaska case anchors this. Alaska has the highest youth-serving provider density in our dataset (1,085 per 100K under-18) — by far above the national average. It also has the highest all-age suicide AADR (27.0 per 100K, third-highest in the country) and the highest YRBSS attempted-suicide rate (19.0% — nearly twice the national average). The framework correctly identifies Alaska as well-supplied. The mortality outcome is independently determined by forces that the supply density does not capture: rural isolation, firearm prevalence, intergenerational trauma, and the specific demographic and geographic structure of Alaska's youth population.

Or take New Jersey, the framework's third positive outlier — supply gap worse than the insurance landscape predicts. New Jersey has the lowest YRBSS attempted-suicide rate in our dartboard (5.2% vs 9.8% national average) and the third-lowest all-age suicide AADR (8.3) in the country. The framework correctly says NJ has built less youth-serving workforce capacity than its insurance landscape implies it should have. The mortality outcome shows that capacity gap is not translating to bad mortality — because NJ's population is dense, urban, demographically structured, and policy-anchored in ways that the supply-vs-uninsured residual does not pick up.

This is the central interpretive finding of Phase B: **the framework's residual_z signal captures a policy-relevant capacity gap (given the insurance landscape) — it is not a suicide-risk signal**. The two are different questions, and our published Phase B paper draws the distinction explicitly.

---

## Why the null? Three honest explanations

The paper does not claim to have identified why provider supply is not protective. It proposes three non-mutually-exclusive explanations and tests none formally:

**First, capacity is not count.** Counting NPIs assigned to youth-serving taxonomies is the closest a public researcher can come to measuring state-level supply without licensed access to billing data. It does not measure hours worked, accepting-new-patients status, network adequacy, or actual care delivery. A state with 1,000 NPI-registered child psychologists per 100,000 under-18 may have very few of them accepting Medicaid, very few accepting new patients, very few with appointment slots within four weeks. The supply count would still be 1,000.

**Second, state granularity is too coarse for the care-delivery question.** A state with 1,000 providers per 100K under-18 may have 500 concentrated in one metro area and zero in 60% of its land area. The intra-state heterogeneity — which the adult V1 paper captured at the tract level — is invisible at the state level. The youth analysis was forced to state geography because YRBSS doesn't publish at sub-state resolution. The price of that constraint is the loss of within-state variation.

**Third, geographic confounding.** The rural-state suicide cluster — Montana, Alaska, Wyoming, New Mexico, Idaho — overlaps with the higher-supply rural states in our analysis (Alaska, Vermont, Maine) for *different reasons*. Alaska's high supply is the Indian Health Service plus tribal health organizations; its high suicide is firearm access plus isolation plus intergenerational trauma. Vermont's high supply is the University of Vermont Medical Center plus state Medicaid expansion plus the Designated Agency system; its moderate suicide is firearm access plus rural-state demographic homogeneity. The cross-section confounds.

None of these explanations is novel. The health-policy literature has documented them. The framework's contribution is to demonstrate the orthogonality with a reproducible analytical pipeline applied to publicly-available federal data.

---

## What the framework should be used for

For policymakers and program designers, the framework's outputs are now interpretable:

For **workforce build-out priority**, use `residual_class == positive_outlier`. The framework's residual analysis identifies which states have not built youth-serving capacity proportional to their insurance landscape. Puerto Rico, North Carolina, and New Jersey are the three positive outliers in our published dataset. These are the states where the federal BHWET (Behavioral Health Workforce Education and Training) program with youth-specific tracks, CCBHC certification with explicit youth-serving criteria, and Title V MCH Block Grant youth lines should be deployed.

For **suicide-prevention triage priority**, use NCHS state-level suicide AADR. The highest-AADR states are Montana, Alaska, Wyoming, New Mexico, and Idaho. The drivers in these states are rural isolation, firearm prevalence, occupational stress, and lack of crisis-response infrastructure — and the policy levers are correspondingly different: 988 crisis-line scaling, mobile crisis teams, firearm-access counseling, NHSC loan-repayment incentives for rural crisis-response staff. The federal MHA / SAMHSA / VA / NIH suicide-prevention research portfolio is the right framing for this list.

Confusing the two lists — using the framework's residual to triage suicide prevention, or using AADR to triage workforce build-out — would mismatch the intervention to the problem.

---

## Drug overdose is a parallel question, not the same question

The fundamentally different state-level signature of drug overdose deaths — higher-coverage states have higher drug OD rates — confirms that this outcome operates on a different axis from mental-health-care access. The "deaths of despair" geography (West Virginia at 79.1 per 100K, DC at 95.0, Ohio in the high-double digits) clusters in economically-stressed regions whose insurance coverage profile is heterogeneous. A future Trellison need-vs-access study on substance-use-disorder treatment access would use a different framework binding (NPPES SUD treatment taxonomies, SAMHSA-certified opioid treatment programs, NSDUH OUD treatment-receiving rate) and would not assume YRBSS or PLACES distress measures capture the relevant need.

We treat drug overdose as adjacent context in Phase B, not as a primary outcome of the mental-health access framework. v1.1 of the Trellison need-vs-access series will likely include a dedicated SUD-access study.

---

## What this finding does for the framework's credibility

The framework's value proposition — surface which states have under-built workforce capacity given their insurance landscape — is *preserved* by Phase B. The Youth V1 published list of positive outliers (PR, NC, NJ) and negative outliers (VT, AK) is correct: those are the states where the supply-vs-uninsured residual deviates from zero.

What is *qualified* is the claim that the framework's outputs are directly outcome-predictive. The framework should not be sold as a suicide-prevention triage tool. The framework's outputs are the workforce-build-out priority list, and the published paper now draws that distinction explicitly.

We believe transparent reporting of this orthogonality strengthens, not weakens, the framework's published case. Trellison's published methodology audit standard is: measure honestly, report what the data does and does not show, and refuse to overstate. Phase B is that standard applied to the framework's own outputs. The framework passes a meaningful test (the within-YRBSS validation, r > +0.75) and fails a different meaningful test (state-level mortality prediction, r ≈ 0). Both results matter.

---

## What's next

v1.1 of Phase B will integrate:

- **FBI Crime Data Explorer juvenile arrests** (pending api.data.gov key issuance — a HIT to Rob is open). Crime correlation would test whether the framework's gap measures predict youth justice involvement, an independent and meaningful proxy.
- **CDC WONDER state-level youth-specific (10-19) suicide mortality** via the XML POST API with explicit license-acceptance handling. This is the gold-standard youth-mortality data and we deferred it from v1.0 because the WONDER programmatic interface requires interactive license flow.
- **Multivariate residual analysis** adjusting for rural-share, opioid-exposure index, household-gun-ownership proxy, and state-level Medicaid expansion status. The cross-section confounding is the largest source of error in v1.0; v1.1 begins to address it.

v2.0 will integrate HCUP State Emergency Department visit licensed data and shift to a longitudinal panel using YRBSS biennial trajectories (2019, 2021, 2023) combined with multi-year CDC WONDER mortality.

---

## The published content arsenal

The Phase B working paper, the methodology supplement, the per-state outcome profiles, the executive brief, the press release, the replication package, and the slide deck are all published as a single content arsenal at:

**https://trellison.com/research/youth-mental-health-outcomes-correlation**

The full joined dataset — 35 states × 11 fields, sha256-stamped — is downloadable from the same page. License: CC-BY-4.0.

The Need-vs-Access Framework v1.1 — which produced the Youth V1 study and which the Phase B analysis tests — is registered as `atlas.need_vs_access_framework_v1` v1.1.0 in the DaedArch tool registry. The Phase B correlation analysis is documented in the methodology supplement; v1.1 will register `atlas.outcome_correlation_v1` as a reusable tool for any need-vs-access study to plug in any set of outcome metrics.

---

*Trellison Institute · methodology-rated data journalism · trellison.com*
*Newsletter version (~1,500 words) derived from the v1.0 working paper. The peer-review article is ~3,100 words.*