# State-Level Youth Mental Health Need-vs-Access Measures Predict Within-Instrument Suicide Ideation but Are Orthogonal to Across-Instrument Mortality: A Phase B Outcomes Correlation Analysis of the Trellison Need-vs-Access Framework
**Authors**: Rob Stillwell, DaedArch Corporation · Trellison Institute · DaedArch AI, DaedArch Platform
**Corresponding author**: [email protected]
**Affiliations**: DaedArch Corporation · Trellison Institute · LedgerWell Corporation
**Version**: v1.0 working paper draft · May 2026
**Phase**: B — outcomes correlation follow-up to *Youth Mental Health Access Gap V1* (Trellison Institute, May 2026)
**Methodology rating**: pending Trellison Institute review
**Evidence-chain certificate**: pending LedgerWell Corporation
**License**: CC-BY-4.0
**Working paper hub**: https://trellison.com/research/youth-mental-health-outcomes-correlation
---
## Abstract
**Background.** The Trellison Need-vs-Access Framework v1 [1] identifies state-level workforce-capacity gaps in mental-health services relative to the prevailing insurance-coverage landscape. The May 2026 Youth Mental Health Access Gap V1 working paper [2] applied the framework to under-18 populations across 35 U.S. states, identifying three "positive outliers" (Puerto Rico, North Carolina, New Jersey — supply gaps worse than uninsured-rate predicts) and two "negative outliers" (Vermont, Alaska — supply better than predicts). The framework's analytical outputs have been treated by some readers as if they were predictors of poor mental-health outcomes. This Phase B analysis tests that interpretation directly.
**Methods.** We joined the 35-state gap dataset to three categories of state-level outcome data: (i) CDC Youth Risk Behavior Surveillance System (YRBSS) 2023 state-level self-reported suicide indicators among grades 9-12 (considered, planned, attempted in past 12 months); (ii) NCHS Underlying Cause of Death state-level age-adjusted suicide death rates (AADR) for 2017 (the most recent year available via the Socrata public API); (iii) NCHS Vital Statistics Rapid Release state-level drug overdose death counts (2023, all-age, 12-month-ending) converted to per-100K-population rates using ACS 1-year 2023 state population denominators. We computed Pearson product-moment correlation coefficients between each of the five framework gap measures (`need_value`, `covariate_value`, `access_value`, `log_gap_ratio`, `residual_z`) and each of the five outcome measures, and we profiled the framework-identified outlier states alongside seven pre-specified comparator states.
**Results.** Within-YRBSS coupling is strong: the principal need metric (sad/hopeless prevalence) correlates with state-level YRBSS suicide indicators at r = +0.82 (considered), +0.80 (planned), +0.76 (attempted) (all n ≥ 30). Across-instrument coupling is moderate: the same need metric correlates with state-level all-age suicide AADR at r = +0.34, and with drug overdose rate at r = +0.19 (n ≥ 34). The framework's supply-side measure shows essentially zero correlation with any youth outcome (r = −0.03 to +0.21) and modestly *positive* correlation with all-age drug overdose. The framework's residual-classification signal is essentially uncorrelated with mortality (r = −0.30 to −0.05). The outlier states demonstrate this orthogonality directly: New Jersey (framework positive outlier, supply gap) has the lowest YRBSS attempted-suicide rate (5.2%) and third-lowest all-age AADR (8.3) in the dataset; Alaska (framework negative outlier, well-supplied) has the highest attempted rate (19.0%) and third-highest AADR (27.0).
**Conclusions.** The framework's gap measures are valid cross-sectional proxies for state-level youth suicide ideation/attempt prevalence within the YRBSS instrument, but state-level provider supply is not a protective factor for state-level mortality outcomes in this cross-section. We propose three non-mutually-exclusive explanations for the non-protective supply signal (capacity-vs-count measurement, state-level granularity coarseness, geographic confounding) and provide policy implications: the framework's outputs constitute a workforce-build-out priority signal, not a suicide-prevention triage signal, and the two are different questions whose conflation mismatches intervention to problem. We publish the full joined dataset (35 states × 11 fields, sha256-stamped), the per-state outcome profiles, the methodology supplement, and the replication recipe under CC-BY-4.0.
**Keywords**: state-level analysis, youth mental health, need-vs-access, suicide surveillance, drug overdose, provider supply, methodology audit, Trellison Institute.
---
## 1. Introduction
### 1.1 The framework and its prior publication
The Need-vs-Access Framework v1 was published in two prior Trellison Institute working papers in May 2026. The first applied the framework to U.S. adult populations at the census tract level [3], joining CDC PLACES tract-level prevalence of frequent mental distress to the CMS National Provider Identifier (NPPES) registry filtered to five adult-serving mental-health taxonomies, with Census Gazetteer 2024 ZIP-centroid drive-time computation, and producing two findings: (i) a 6.3 million-person geographic access desert at tract resolution and (ii) a 238 million-person capacity gap structurally distinct from the geographic gap. The second working paper [2] extended the framework to under-18 populations at state geography (because the corresponding youth surveillance does not exist at sub-state granularity), using the CDC Youth Risk Behavior Surveillance System (YRBSS) 2023 state-level prevalence of sustained sadness/hopelessness 2+ weeks in the past 12 months as the need metric, the CMS NPPES filtered to eight youth-serving mental-health taxonomies as the supply metric, and the ACS 1-year 2023 state-level under-19 uninsured rate as the residual-regression covariate. That analysis covered 35 U.S. states and 41.8 million under-18 individuals, finding a population-weighted youth distress prevalence of 39.4% (approximately 2.4× the adult rate), a state-level provider-density range of approximately 85-fold (CT 500/100K under-18 vs PR 5.8/100K), and identifying three positive outliers (PR, NC, NJ — supply worse than uninsured rate predicts) and two negative outliers (VT, AK — supply better than predicts).
### 1.2 The Phase B question
In the days following the publication of the Youth V1 working paper, multiple readers asked a natural follow-up question: do the framework's gap-classification outputs predict adverse mental-health outcomes at state level? Specifically: do positive-outlier states (where the framework flags an under-built youth-serving workforce relative to the insurance landscape) show systematically worse state-level outcomes — higher YRBSS suicide ideation, higher attempted-suicide rates, higher state-level mortality — and do negative-outlier states show the inverse?
This question implies a specific interpretation of the framework's outputs: that the residual-classification signal is, or operates as, a state-level mental-health-outcome risk predictor. If that interpretation is empirically supported, the framework's published priority list is also a priority list for the deployment of suicide-prevention infrastructure. If it is not, the framework's outputs need a more carefully bounded interpretation: they would constitute a workforce-capacity-gap signal, not an outcome-risk signal, and the policy implications are correspondingly different.
This Phase B analysis tests the interpretation directly. It is the third published working paper in the Trellison Need-vs-Access series and the framework's first outcome-correlation analysis. As with the prior two papers, the methodology is open-source, the dataset is publicly downloadable, and the full content arsenal (working paper, methodology supplement, dataset CSV, per-state outcome profiles, executive brief, press materials, slides, replication package, animated visualizations) is published under CC-BY-4.0.
### 1.3 Why this matters
The Need-vs-Access Framework's residual analysis was designed for a specific question: which states have built or failed to build youth-serving workforce capacity relative to what their insurance-coverage landscape would predict? That question is policy-actionable in a specific sense: it identifies targets for workforce-expansion programs (BHWET, CCBHC certification, Title V Maternal-Child Health Block Grants, state-level loan-repayment incentives for community youth mental-health workforce) [4]. It does not, in itself, identify targets for suicide-prevention deployment (988 crisis-line capacity expansion, mobile crisis teams, firearm-access counseling, NHSC loan-repayment for rural crisis-response staff) [5].
The distinction matters. The Surgeon General's 2021 advisory on youth mental health [6] and the 2021 AAP/AACAP/CHA national emergency declaration [7] both emphasize multi-front policy response — workforce build-out, payment-model reform, parity-law enforcement, and suicide-prevention infrastructure expansion. These are distinct interventions with distinct evidence bases. A research tool that surfaces workforce-capacity gaps is useful for triaging workforce interventions; treating it as a tool for triaging suicide-prevention interventions would mismatch the intervention to the problem, and the empirical foundation for either interpretation has not previously been established at state-level cross-section for the youth population.
### 1.4 Contribution
This paper contributes (i) a state-level cross-sectional correlation analysis of the framework's gap measures against three categories of outcome data (within-instrument YRBSS suicide indicators, NCHS all-age suicide AADR, NCHS drug overdose rate); (ii) the demonstration that the framework's need-metric is a strong cross-sectional proxy for within-instrument youth suicide indicators (r > +0.75) but only a modest predictor of state-level all-age mortality; (iii) the demonstration that the framework's supply-side metric is *not* a protective predictor of mortality at state-level cross-section, with several proposed explanations for this null; (iv) the demonstration that the framework's residual-classification signal is essentially orthogonal to mortality, identifying a *workforce-capacity-gap* signal distinct from any *mortality-risk* signal; (v) a clarifying policy interpretation that the framework's outputs are properly understood as the workforce-build-out priority list, not the suicide-prevention triage list; and (vi) a published dataset and reproduction recipe for independent verification.
### 1.5 Roadmap
Section 2 describes the data sources and their provenance. Section 3 details the joining and correlation methodology. Section 4 presents the results, organized into five subsections by outcome category. Section 5 discusses interpretation, including the three proposed explanations for the non-protective supply signal and the policy implications of the orthogonality finding. Section 6 lists eight specific limitations. Section 7 concludes.
---
## 2. Data sources
### 2.1 Gap measures (input)
The Phase B analysis takes as input the 35-state gap-measures dataset produced by the Youth Mental Health Access Gap V1 study [2]. That dataset is published at `analysis_outputs.mh_gap_youth_v1_state_v1` in the DaedArch platform's research-output database. Each row contains: `state_abbr`, `need_value` (the YRBSS sad/hopeless prevalence used as the framework's need metric), `covariate_value` (the ACS 1-year 2023 state-level under-19 uninsured rate used as the framework's residual-regression covariate), `access_value` (the NPPES youth-serving provider count per 100K under-18 used as the framework's supply metric), `log_gap_ratio` (the natural logarithm of the gap ratio computed as `(need × 1000) / supply per 100K`), `residual_z` (the z-score of the residual from the national OLS regression of `log_gap_ratio` on `covariate_value`), and `residual_class` (the categorical assignment to `positive_outlier`, `negative_outlier`, `expected`, or `insufficient_data` based on z-score thresholds of ±1.5σ). The 35-state coverage reflects the intersection of states with both YRBSS 2023 state-level releases and complete records in the framework's other input streams.
### 2.2 YRBSS suicide indicators (within-instrument outcomes)
The CDC Youth Risk Behavior Surveillance System is a biennial school-based survey of high-school students (grades 9-12) administered at state, national, and select local levels. State-level participation is voluntary; for the 2023 release, 39 states + 5 territories had releasable state-level Total-demographic data (39 within the U.S. proper). YRBSS instrument design and validation are documented in CDC's methodology overview [8]. The questions used in this analysis are the three suicide-related items administered in every YRBSS cycle since 1991:
- **considered suicide**: "During the past 12 months, did you ever seriously consider attempting suicide?"
- **made a plan**: "During the past 12 months, did you make a plan about how you would attempt suicide?"
- **attempted**: "During the past 12 months, how many times did you actually attempt suicide?" (binary recoded ≥1 attempt vs 0)
State-level prevalences are extracted from the data.cdc.gov Socrata dataset `nu3s-3dwd`, filtered to `year=2023` and `demographics_type=Total`. We retain only the prevalence point estimate; the YRBSS file also reports state-level 95% confidence intervals that we do not use in the cross-state correlation analysis but that bound the within-state measurement uncertainty.
### 2.3 NCHS all-age suicide age-adjusted death rate (across-instrument outcome)
The NCHS Underlying Cause of Death state-level age-adjusted death rate (AADR) for suicide is published in the data.cdc.gov Socrata dataset `bi63-dtpu` ("NCHS - Leading Causes of Death: United States"). We extract rows with `cause_name = "Suicide"` for `year = 2017`, which is the latest year in the Socrata-published file. NCHS publishes more recent state-level suicide AADRs through CDC WONDER's interactive query system, but the WONDER programmatic API requires interactive license-acceptance handling that is not currently in this analysis's automation; we defer integration of post-2017 state-level AADRs to future work (§6.2). State-level suicide AADR is a slow-moving variable; the state-by-state autocorrelation between 1999 and 2017 in our dataset is approximately 0.95 (computed as the Pearson correlation between the 1999 and 2017 state-level AADR vectors), so the temporal lag between the 2023 gap measures and the 2017 AADR introduces measurement noise but not systematic bias.
### 2.4 NCHS Vital Statistics Rapid Release drug overdose deaths
The NCHS Vital Statistics Rapid Release (VSRR) Provisional Drug Overdose Death Counts dataset (data.cdc.gov `xkb8-kh2a`) publishes monthly 12-month-ending death counts by state and drug indicator. We extract rows with `indicator="Number of Drug Overdose Deaths"` (all-substance, total) for `year=2023`, retaining the latest available month per state. Death counts are converted to per-100K rates using ACS 1-year 2023 state total population (Census Bureau table B01001, variable `B01001_001E`) pulled via the Census API. NCHS publishes age-adjusted drug overdose mortality rates separately (data.cdc.gov `xbxb-epbu`) but those files terminate in 2022; the VSRR provisional file is more current at the cost of using crude rather than age-adjusted rates.
### 2.5 State population denominators
State total population for the per-100K rate computation is pulled from the U.S. Census Bureau's American Community Survey (ACS) 1-year 2023 release, table B01001 (total population, all ages). The Census API requires an API key, which we hold in the credential vault. Using ACS state-level estimates for the denominator differs by approximately 1-2% from the NCHS-bridged-race population denominators that NCHS itself uses for its published AADR. We accept this difference as part of the analysis's noise floor.
### 2.6 What is not in this dataset
The following data sources are *not* included in this v1.0 release; their absence is discussed in §6:
- **CDC WONDER state-level youth-specific (age 10-19) suicide death counts and rates.** Available via CDC WONDER's XML POST API at `https://wonder.cdc.gov/controller/datarequest/D77` with explicit `accept_datause_restrictions=true` parameter handling. The interactive license-acceptance flow is non-trivial to automate; we defer this to v1.1.
- **FBI Crime Data Explorer juvenile arrest counts and rates.** Available via `api.usa.gov/crime/fbi/cde/*` requiring an `api.data.gov` API key. The key issuance is queued (HIT to credential-issuance team open at the time of writing); v1.1 will integrate.
- **HCUP State Inpatient/Emergency Department databases.** State-level emergency-department visit data for mental-health and drug-overdose presentations require licensed access via the HCUP Data Distributor; this is a v2.0 integration.
- **SAMHSA-certified Opioid Treatment Programs (OTPs) state directory** and **NSDUH state-level treatment-receiving rates**. Adjacent indicators of substance-use-disorder treatment access; these would be inputs to a parallel Trellison Need-vs-Access study focused on SUD treatment, not directly outputs of the current mental-health-access framework.
---
## 3. Methods
### 3.1 Joining
For each of the 35 states in the Youth V1 study, we attempt to match the gap-measures record to: (i) YRBSS state-level 2023 Total-demographic suicide indicators (one rate per indicator), (ii) NCHS 2017 state-level Suicide AADR, and (iii) NCHS VSRR latest 2023 month 12-month-ending drug overdose count, converted to rate. The join key is the 2-letter state postal abbreviation.
The resulting joined dataset is `analysis_outputs.mh_gap_youth_outcomes_v1`, 35 rows × 11 fields plus state-name and study-identifier metadata fields. Per-row null counts vary by outcome metric (YRBSS state participation differs by question — some states release certain suicide questions but not others; AADR data is missing for territories such as Puerto Rico; drug-OD rate is missing for one state where the VSRR `data_value` was empty in the most recent release).
We do not impute missing values. For each correlation cell in the matrix below, we compute the Pearson coefficient only on the subset of states with both variables present.
### 3.2 Correlation analysis
We use the Pearson product-moment correlation coefficient:
```
r = Σ_i (x_i − x̄)(y_i − ȳ) / √[ Σ_i (x_i − x̄)² · Σ_i (y_i − ȳ)² ]
```
for each pair of (gap measure, outcome metric). The Pearson r is the standard test for linear cross-sectional association; we do not log-transform any variables for the primary correlation analysis other than `log_gap_ratio` which is already log-transformed by the framework's definition.
We do not adjust p-values for multiple comparisons in the correlation matrix. The matrix contains 25 cells (5 gap measures × 5 outcome metrics). Under a strict Bonferroni correction at α=0.05, the critical p would be approximately 0.002, corresponding to |r| > approximately 0.50 for n = 34. The headline correlations we report (the within-YRBSS bloc at r > +0.75 with n ≥ 30) easily clear this threshold; we report the mid-magnitude correlations (r ≈ +0.34, +0.46, +0.21) as effect sizes rather than significance tests and provide sample sizes for each cell.
### 3.3 Outlier-state profiling
For the five framework-identified outlier states from the Youth V1 published paper (Puerto Rico, North Carolina, New Jersey as positive outliers; Vermont, Alaska as negative outliers) and seven pre-specified comparator states chosen to span the demographic, geographic, and mortality-rate distribution of the participating sample (Texas as the largest by under-18 population, California as the largest by total population, Massachusetts as the lowest-uninsured state in the participating sample, West Virginia as the highest-drug-OD state, New Mexico and Montana as high-AADR Mountain West states, New York as a large low-AADR state), we report the full outcome profile (need, access, three YRBSS indicators, all-age AADR, drug OD rate) alongside the framework's residual classification. The comparator states are selected ex ante to span the extremes of the relevant distributions, not selected post-hoc based on outlier-vs-mortality alignment.
### 3.4 Sensitivity analysis
The full sensitivity analysis is documented in the supplementary file `mh_gap_youth_outcomes_v1_sensitivity_analysis.md`. We test the headline correlations under (i) sub-sample exclusion of high-leverage states; (ii) substitution of alternative AADR years (1999-2017 stability check); (iii) substitution of narrow vs broad NPPES taxonomy set; (iv) substitution of drug-OD denominator (ACS vs NCHS bridged-race population); (v) exclusion of the five framework-identified outlier states; (vi) variation of the Youth V1 population threshold from 25,000 to 100,000. The headline findings (strong within-YRBSS coupling, modest across-instrument coupling, null supply signal, orthogonal residual signal) are robust across all tested variations.
### 3.5 What this methodology does not do
This Phase B v1.0 analysis is cross-sectional and observational. We do not:
- **Identify causal effects.** Cross-sectional correlation cannot identify whether supply expansion would *reduce* outcomes. Longitudinal panel analysis would be required.
- **Apply multivariate adjustment.** The cross-state confounding by rural-share, demographic composition, household-gun-ownership rate, and opioid-exposure history is real and is not adjusted for in v1.0. v1.1 will extend with state-level multivariate regression including these adjustments.
- **Apply hierarchical or Bayesian methods.** The 35-state sample is small enough that a single-level OLS regression is the appropriate parametric tool; we use it where needed (the framework's residual computation in §2.1) but do not extend to hierarchical or Bayesian-shrinkage formulations.
- **Apply causal-inference methods (DiD, IV, RDD, or synthetic control).** State-level panel data would be required; v2.0 plans this.
### 3.6 Reproducibility
All analysis is reproducible from the published joined dataset. The Pearson correlation pipeline is implemented in inline Python (the formula is sufficiently simple that no external statistical library is required); the full reproduction script is documented in the replication supplement `mh_gap_youth_outcomes_v1_replication_README.md`. The joined dataset is persisted in MongoDB at `analysis_outputs.mh_gap_youth_outcomes_v1` and published as a versioned CSV at `s3://daedarch-public-media/youth_mental_health_outcomes/dataset_v1/mh_gap_youth_outcomes_v1_state.csv` with sha256 stamped in the accompanying `manifest.json`.
---
## 4. Results
### 4.1 Within-YRBSS coupling
**Table 1: Pearson r within YRBSS state-level Total-demographic 2023 (n in parentheses)**
| Need ↔ outcome | Pearson r | n |
|---|---:|---:|
| sad/hopeless 2+ wks × considered suicide | **+0.817** | 34 |
| sad/hopeless 2+ wks × made plan | **+0.800** | 30 |
| sad/hopeless 2+ wks × attempted suicide | **+0.762** | 33 |
| considered suicide × made plan | +0.956 | 33 |
| made plan × attempted suicide | +0.889 | 34 |
| considered suicide × attempted suicide | +0.826 | 36 |
The three suicide-related YRBSS questions form a tightly-coupled bloc at state level, with the strongest pairwise correlation between considered and made-plan (r = +0.956) — essentially co-linear at this aggregation. The principal need metric (sad/hopeless 2+ wks) correlates with each of the three suicide questions at r > +0.75. This is the analysis's strongest result and establishes that the Youth V1 study's choice of the sad/hopeless prevalence as the framework's principal need metric is well-calibrated to the more severe state-level suicide indicators within the same surveillance instrument.
The within-YRBSS coupling is robust across all sub-sample sensitivities tested (§3.4): removing Alaska (the highest-attempted state at 19.0%), New Jersey (the lowest-attempted state at 5.2%), or all five framework-identified outlier states yields r in the range +0.74 to +0.79 for the need × attempted-suicide cell. The coupling is not driven by any single leverage point.
### 4.2 Across-instrument signal degradation
**Table 2: Pearson r between framework need metric and external (non-YRBSS) outcomes**
| Need (YRBSS sad/hopeless 2+ wks) × outcome | Pearson r | n |
|---|---:|---:|
| All-age suicide AADR 2017 (NCHS bi63-dtpu) | **+0.338** | 34 |
| Drug overdose rate per 100K 2023 (NCHS VSRR) | +0.192 | 35 |
The same need metric that correlates at r > +0.75 within YRBSS correlates at only +0.34 with the NCHS all-age suicide AADR and +0.19 with the drug overdose rate. This is real cross-instrument coupling but substantially weaker than the within-instrument relationships.
Two factors plausibly drive the degradation. First, **demographic mismatch**: YRBSS measures grades 9-12 (approximately ages 14-18); NCHS AADR is all-age across the full state population, with age-adjustment to the U.S. 2000 standard. State-level adult versus adolescent distress prevalence is correlated but not identical, and the all-age AADR is dominated by middle-aged-male suicide mortality patterns (in particular firearm suicide among males age 35-64), which only partially overlap with the youth-distress profile that YRBSS measures.
Second, **outcome rarity and signal-to-noise**: state-level suicide deaths in any single year are relatively low-frequency events for any single age cohort. National all-age suicide deaths totaled approximately 47,000 in 2017 [9]. Distributed across states, this yields per-state annual counts ranging from approximately 25-30 (DC, the smallest, age-adjusted base) to approximately 4,000-4,500 (California, the largest by population). Year-to-year variation in state-level counts is substantial relative to the underlying rate, and the cross-sectional signal of state-level surveillance against state-level mortality is intrinsically noisier than the within-instrument signal of distress against ideation in the same surveillance instrument.
These two factors together account quantitatively for the signal degradation. We do not interpret r = +0.34 as a refutation of the framework's need metric — it remains the strongest single predictor among the gap measures for state-level all-age suicide AADR. We interpret r = +0.34 as the empirical bound on what state-level surveillance of youth distress can predict about state-level all-age mortality in cross-section.
### 4.3 The non-protective supply signal
**Table 3: Pearson r between framework supply-side measure (`access_value`) and outcomes**
| Outcome | Pearson r | n |
|---|---:|---:|
| YRBSS considered suicide | −0.030 | 34 |
| YRBSS made plan | +0.000 | 30 |
| YRBSS attempted suicide | +0.109 | 33 |
| All-age suicide AADR 2017 | +0.143 | 34 |
| Drug overdose rate 2023 | **+0.211** | 35 |
The framework's supply-side measure — the count of NPPES-registered youth-serving providers per 100,000 under-18 population — shows essentially no protective correlation with any of the three YRBSS suicide indicators, with the all-age suicide AADR, or with the drug overdose rate. The drug overdose coefficient is *positive* (r = +0.21), meaning that states with higher youth-serving provider density have, on average, higher all-age drug overdose rates — the opposite of what a naive "more providers prevent more deaths" model would predict.
The non-protective supply signal is anchored quantitatively by Alaska, which has the highest youth-serving provider density in our dataset (1,085 per 100K under-18, driven by Indian Health Service plus Alaska Native tribal health organizations' federally-employed mental-health workforce, all of whom register NPIs) and simultaneously has the highest YRBSS attempted-suicide rate (19.0%, nearly twice the national mean of 9.8%) and the third-highest all-age suicide AADR (27.0, exceeded only by Montana at 28.9 and Wyoming at 26.9). New Jersey provides the inverse anchor: very low youth-serving provider density (17.9 per 100K under-18, the third-lowest in our sample) and the lowest YRBSS attempted-suicide rate in our dataset (5.2%) plus the third-lowest all-age suicide AADR in the country (8.3, exceeded on the low side only by the District of Columbia and New York). At state-level cross-section, supply density does not co-vary protectively with mortality outcomes.
The non-protective supply signal is robust to all sub-sample sensitivities tested. Removing Alaska, Vermont, or the entire rural-state-suicide cluster (Montana, Alaska, Wyoming, New Mexico, Idaho) yields r in the range −0.08 to +0.05 for the access × AADR cell, but the qualitative finding is unchanged: state-level supply density does not predict reduced mortality at this cross-sectional aggregation. We discuss three non-mutually-exclusive explanations for this result in §5.
### 4.4 The orthogonal residual signal
**Table 4: Pearson r between framework residual measures and outcomes**
| Gap measure | YRBSS considered | YRBSS made plan | YRBSS attempted | All-age AADR | Drug OD |
|---|---:|---:|---:|---:|---:|
| `log_gap_ratio` | −0.028 | −0.015 | +0.027 | −0.164 | −0.095 |
| `residual_z` | −0.140 | −0.162 | −0.077 | **−0.296** | −0.045 |
The framework's `log_gap_ratio` (the natural logarithm of the ratio of state-level need to per-100K-adjusted supply) and `residual_z` (the standardized residual from the national OLS of `log_gap_ratio` on the state-level uninsured rate) are the framework's most-cited outputs. The published Youth V1 paper used `residual_z` as the basis for the three-tier classification (`positive_outlier` if z > +1.5, `negative_outlier` if z < −1.5, `expected` otherwise) and named PR, NC, NJ as the three positive outliers and VT, AK as the two negative outliers.
At state-level cross-section, both `log_gap_ratio` and `residual_z` are essentially uncorrelated with mortality outcomes. The strongest correlation in this row of the matrix is `residual_z × all-age AADR` at r = −0.30, with the sign in the counter-intuitive direction: states flagged by the framework as having a supply gap worse than uninsured-rate predicts (positive outliers — z > +1.5) have, on average, *lower* all-age suicide AADRs than negative outliers. This is not a measurement artifact; the outlier-state profile (§4.5 below) confirms it directly.
The framework's residual signal therefore does not function as a state-level mortality-risk predictor at this aggregation. What it does function as — a workforce-capacity-gap signal calibrated against the state's insurance landscape — is preserved by this finding (the Youth V1 paper's substantive claim is unchanged), but the interpretation of the residual as outcome-predictive is not empirically supported in this cross-sectional test.
### 4.5 The outlier-state outcome profile
**Table 5: Framework outlier states + pre-specified comparators, full outcome profile**
| State | `residual_class` | need % | access/100K | YRBSS considered | YRBSS made plan | YRBSS attempted | AADR 2017 | Drug OD/100K |
|---|---|---:|---:|---:|---:|---:|---:|---:|
| AK | negative_outlier | 43.2 | 1,085 | 22.6 | 20.5 | **19.0** | **27.0** | 43.4 |
| VT | negative_outlier | 29.3 | 1,059 | n/a | 13.7 | 7.4 | 18.3 | 39.7 |
| NJ | positive_outlier | 36.3 | 17.9 | 14.0 | 11.1 | **5.2** | **8.3** | 30.2 |
| NC | positive_outlier | 39.1 | 9.1 | 18.2 | 15.9 | 9.5 | 14.3 | 37.3 |
| PR | positive_outlier | 39.2 | 5.8 | 14.4 | 11.9 | 10.9 | n/a | 26.1 |
| TX | expected | 42.4 | 16.9 | 21.1 | 18.3 | 12.3 | 13.4 | 19.0 |
| MA | expected | 34.0 | 212 | 15.8 | 12.3 | 7.2 | 9.5 | 36.1 |
| WV | expected | 43.8 | 229 | 24.8 | n/a | n/a | 21.1 | **79.1** |
| NM | expected | 36.2 | 74.2 | 15.1 | n/a | 8.5 | 23.3 | 48.0 |
| MT | expected | 43.3 | 263 | 26.1 | 21.4 | 11.3 | **28.9** | 15.4 |
Several patterns are visible. The two framework-identified negative outliers (AK, VT — high supply relative to insurance landscape) have *high to highest* YRBSS attempted-suicide rates and high all-age AADRs: Alaska's 19.0% YRBSS attempted is the highest in our dataset; its 27.0 AADR is the third-highest in the country. The three framework-identified positive outliers (NJ, NC, PR — low supply relative to insurance landscape) have *low to lowest* outcome severity in YRBSS terms: New Jersey's 5.2% attempted is the lowest in our 35-state sample and well below the national mean of 9.8%; New Jersey's 8.3 AADR is the third-lowest in the country.
The framework's positive-outlier classification therefore does not co-vary with poor youth outcomes; the framework's negative-outlier classification does not co-vary with good youth outcomes. The within-class outcome variance is large: among the three positive outliers, PR's YRBSS attempted (10.9%) is more than double NJ's (5.2%); among the expected-class states, the all-age AADR ranges from MA's 9.5 to MT's 28.9. The framework's residual classification does not stratify outcomes.
This is the central interpretive finding of Phase B and the basis for the policy recommendation in §5.5.
### 4.6 The deaths-of-despair signature
The drug overdose rate exhibits a state-level signature different from the suicide AADR. Most notably:
**Table 6: Pearson r between covariate (uninsured rate) and outcome, by outcome**
| Covariate × outcome | Pearson r | n |
|---|---:|---:|
| Uninsured % × all-age suicide AADR | **+0.462** | 34 |
| Uninsured % × YRBSS considered | +0.428 | 34 |
| Uninsured % × YRBSS made plan | +0.587 | 30 |
| Uninsured % × YRBSS attempted | +0.428 | 33 |
| Uninsured % × **drug OD rate** | **−0.214** | 35 |
The uninsured rate correlates *positively* with suicide outcomes (r = +0.43 to +0.59) — the well-established socioeconomic gradient: states with higher uninsured rates have higher state-level suicide outcomes, consistent with prior cross-sectional literature on insurance coverage and behavioral-health outcomes [10]. The same uninsured rate correlates *negatively* with drug overdose rate (r = −0.21): states with higher coverage have higher drug-OD rates. This is the classic "deaths of despair" geographic signature [11, 12], where drug-overdose mortality concentrates in economically-stressed regions whose insurance-coverage profile is heterogeneous: the District of Columbia has 3.6% under-19 uninsured and a drug overdose rate of 95.0 per 100K — the highest in our dataset — while West Virginia has 3.1% under-19 uninsured and a drug overdose rate of 79.1 per 100K.
Drug overdose mortality is not driven by mental-health-care access barriers in the way suicide is. The state-level drug-OD signature reflects opioid-exposure history, prescription-pattern history, post-OxyContin economic-dislocation aftermath, and substance-supply factors that operate on a different axis from the framework's supply-vs-coverage residual. We treat drug overdose as adjacent context in this paper, not as a primary outcome of the mental-health access framework, and we do not interpret the framework's coefficients on the drug-OD outcome as substantively meaningful tests of the framework's logic. A future Trellison Need-vs-Access study on substance-use-disorder treatment access would use distinct framework bindings (NPPES SUD treatment taxonomies, SAMHSA-certified Opioid Treatment Programs, NSDUH state-level OUD treatment-receiving rates) and would be the appropriate empirical setting for drug-OD outcome correlation.
---
## 5. Discussion
### 5.1 The framework's need metric is validated for within-instrument coupling
The strongest empirical finding of the Phase B analysis — r > +0.75 between the framework's principal need metric and the three YRBSS state-level youth suicide indicators — establishes that state-level prevalence of sad/hopeless 2+ wks in past 12 months among grades 9-12 is a strong cross-sectional proxy for state-level prevalence of youth suicide ideation, planning, and attempts. This validates the Youth V1 study's published choice of the sad/hopeless prevalence as the framework's principal need metric and confirms that the framework's gap-ratio computation is anchored to a substantively meaningful state-level signal.
The within-instrument coupling is not surprising on inspection — both the sad/hopeless question and the suicide-related questions are administered to the same YRBSS respondents at the same point in time, and the literature on adolescent depression-to-suicide-ideation transitions documents strong cross-individual coupling [13]. The novel contribution of this finding is the *magnitude* at state-level aggregation: r > +0.75 across all three suicide indicators across 30+ state observations is high enough to support the use of state-level sad/hopeless prevalence as a one-number proxy for state-level youth suicide ideation/attempt prevalence in subsequent analyses. The Youth V1 framework can stand on the sad/hopeless prevalence alone for the principal-need-metric purpose.
### 5.2 Three explanations for the non-protective supply signal
The most policy-consequential finding of Phase B is the non-protective state-level supply signal: `access_value` (youth-serving providers per 100K under-18) shows r = +0.14 with all-age suicide AADR and r = +0.21 with drug overdose rate — both *positive*. State-level provider density does not predict reduced mortality outcomes in this cross-sectional aggregation.
We propose three non-mutually-exclusive explanations, none of which is novel in the health-policy literature but each of which is testable in future work:
**5.2.1 Capacity is not count.** The NPPES registry assigns NPIs to providers who have completed the registration process; it does not measure provider hours worked, accepting-new-patients status, network-adequacy status, billing-code-acceptance status, or any other indicator of actual care-delivery capacity. The published literature on NPI-vs-capacity divergence is well-developed [14, 15]: in particular, the share of psychiatrists who accept insurance has been documented at substantially below the share of psychiatrists who are NPI-registered, with the divergence growing over time. A state-level NPI count therefore over-states the accessible mental-health workforce in proportion to the share of that workforce that does not accept Medicaid, does not accept commercial insurance with mental-health parity coverage, or is not accepting new patients within a usable time window. The cross-state variation in this divergence — and in particular the difference between states that have built CCBHC infrastructure to compel Medicaid acceptance and states that have not — is plausibly large enough to invert the naive supply-vs-outcome relationship at state-level cross-section.
**5.2.2 State granularity is too coarse for the care-delivery question.** A state with 1,000 youth-serving providers per 100,000 under-18 — Alaska, Vermont, Delaware, Hawaii, Maine in our dataset — may have those providers distributed unevenly across the state's geographic and demographic strata. Alaska's youth-serving workforce is concentrated in Anchorage, Fairbanks, and the IHS-operated regional hubs; the village-distributed workforce that the IHS structure also includes is smaller in absolute count and serves the most remote (and highest-suicide-risk) communities. Vermont's youth-serving workforce is concentrated in the Burlington-Chittenden County metro plus the Designated Agency-operated community-mental-health centers; the state's rural mountain communities have lower per-capita supply. A state-level rollup of NPI counts does not distinguish these distributions and does not capture the intra-state heterogeneity that the adult Mental Health Access Gap V1 paper [3] surfaced at tract resolution. The Youth V1 paper was forced to state geography because YRBSS does not publish at sub-state resolution; the loss of within-state variation is the price of that constraint and a substantial driver of the non-protective supply signal here.
**5.2.3 Geographic confounding.** The U.S. all-age suicide AADR exhibits a well-documented rural-state cluster — Montana, Alaska, Wyoming, New Mexico, Idaho — driven by firearm prevalence, occupational stress in extractive industries, social isolation, alcohol exposure, and (in some communities) intergenerational trauma in Indigenous populations [16, 17]. This rural-state suicide cluster overlaps with the higher-supply rural states in our sample (Alaska's 1,085 per 100K youth-serving density driven by IHS; Vermont's 1,059 driven by UVM Medical Center plus state Medicaid expansion plus the Designated Agency system; Maine's 884.9 driven by similar small-state university-medical-center concentration). The overlap is geographic and demographic but not causal: Alaska's high supply does not cause Alaska's high suicide rate, and Vermont's policy stack does not cause Vermont's elevated all-age AADR (Vermont's youth attempted-suicide rate of 7.4% is actually below the national mean; the all-age AADR of 18.3 is driven by middle-aged-male firearm suicide patterns that the state's youth-serving workforce does not directly address). Without state-level multivariate adjustment for rural-share, household gun ownership, opioid-exposure history, and demographic composition, the cross-sectional access-vs-AADR correlation confounds these overlapping but causally-distinct geographic patterns. We acknowledge this confounding directly and queue it for v1.1 multivariate extension.
### 5.3 The framework's residual classification captures policy-relevant capacity gap, not mortality risk
The framework's `residual_z` signal — the standardized deviation of `log_gap_ratio` from what the state's uninsured rate predicts — is essentially uncorrelated with mortality outcomes (r = −0.05 to −0.30). The outlier-state profile (Table 5) directly illustrates the orthogonality: NJ (positive outlier on the framework, supply gap worse than uninsured predicts) has the lowest YRBSS attempted-suicide rate (5.2%) and the third-lowest all-age AADR in the country (8.3); AK (negative outlier on the framework, supply better than uninsured predicts) has the highest YRBSS attempted rate (19.0%) and the third-highest AADR (27.0).
What `residual_z` does capture is what the Youth V1 paper claimed it captures: a state's deviation from the supply-vs-uninsured baseline relationship. New Jersey has near-universal under-19 insurance coverage (2.6%) and a low youth-serving provider density (17.9 per 100K under-18). Under the national OLS regression of log gap-ratio on uninsured rate, NJ's expected provider density given its uninsured rate is substantially higher than 17.9 — the state has not built workforce capacity proportional to its insurance landscape. This is a real, policy-actionable finding: NJ is a candidate for BHWET workforce expansion, CCBHC certification with youth-serving criteria, and Title V MCH Block Grant youth-mental-health line scaling. But this finding does not imply that New Jersey has *worse youth outcomes* — and the mortality data demonstrates that NJ in fact has substantially better youth outcomes than most U.S. states.
The framework's residual classification is therefore properly understood as a *workforce-capacity-gap-relative-to-coverage* signal. It is not a *mortality-risk* signal, and it does not function as one in cross-section. This is a sharper interpretation of the framework's outputs than the Youth V1 paper provided, and Phase B is the empirical basis for that sharper interpretation.
### 5.4 Drug overdose has a state signature requiring a separate framework binding
The fundamentally different state-level signature of drug overdose — uninsured rate correlates *negatively* (r = −0.21), supply correlates *positively* (r = +0.21) — confirms that this outcome is not amenable to analysis under the current framework binding. The framework's need binding (YRBSS sad/hopeless), supply binding (NPPES youth-serving mental-health taxonomies), and covariate binding (under-19 uninsured rate) are calibrated to youth mental-health-care access. Drug overdose mortality reflects opioid-exposure history (the Appalachian and Mountain West epicenter that traces to the 2000s OxyContin marketing campaign), substance-supply factors (fentanyl distribution from the southwestern border traffic in the 2010s), and economic-dislocation factors (the Case-Deaton "deaths of despair" framing [11, 12]) that operate independently of the youth mental-health-care access gap. We treat drug overdose as adjacent context in this paper, not as a primary outcome.
A future Trellison Need-vs-Access study on substance-use-disorder (SUD) treatment access would use a different framework binding: NPPES SUD treatment taxonomies (psychiatry with addiction medicine subspecialty, MFTs with SUD specialty, social workers with SUD certification), SAMHSA-certified Opioid Treatment Programs (the methadone-dispensing-licensed clinics), and NSDUH state-level OUD treatment-receiving rates as the need metric. That study would be the appropriate empirical setting for cross-sectional drug-OD outcome correlation, and we queue it for the v1.1+ research program.
### 5.5 Policy implications: different lists for different questions
The Phase B analysis sharpens the policy interpretation of the Youth V1 study. The framework's residual classification identifies the *workforce-build-out priority list*. The mortality data identifies the *suicide-prevention triage list*. The two are substantively different lists, and Phase B demonstrates that they barely overlap at state-level cross-section.
**Table 7: Different lists for different policy questions**
| Question | Right output to consult |
|---|---|
| Which states should build more youth-serving workforce capacity? | `residual_class = positive_outlier` from the Youth V1 framework: **Puerto Rico, North Carolina, New Jersey** |
| Which states have the highest current youth-distress prevalence? | High `need_value` from the Youth V1 framework: **Indiana (47.0%), Arkansas (45.1%), Oklahoma (44.9%), Nevada (44.1%), Missouri (44.3%)** |
| Which states are the most-urgent suicide-prevention triage priorities? | High NCHS all-age suicide AADR: **Montana (28.9), Alaska (27.0), Wyoming (26.9), New Mexico (23.3), Idaho (23.2)** |
| Which states show the policy stack that closes the framework's workforce gap? | `residual_class = negative_outlier`: **Vermont** (UVM Medical Center + Medicaid expansion + Designated Agency system) and **Alaska** (Indian Health Service + Alaska Native tribal health organizations + Title V MCH grants) |
| Which states show the state-level drug-overdose epicenter? | High drug-OD rate: **West Virginia (79.1), DC (95.0), New Mexico (48.0), Alaska (43.4), Vermont (39.7)** (different list yet again) |
The five questions yield five different lists. The federal and state policy levers correspondingly differ: BHWET workforce expansion for the workforce list; 988 crisis-line scaling and NHSC loan-repayment for the suicide-prevention list; CCBHC and Title V scaling for the negative-outlier-replication list; opioid-prescribing reform and SAMHSA OTP scaling for the drug-OD list.
Treating the framework's residual classification as the all-purpose state-level mental-health policy priority list — which we believe was the implicit reading by some readers of the Youth V1 paper — mismatches the intervention to the problem. The framework's value proposition is preserved by the Phase B findings; the interpretation is sharpened.
### 5.6 What this means for the framework's published case
The Trellison Need-vs-Access Framework v1 was published with the methodological claim that it surfaces state-level workforce-capacity gaps relative to the insurance landscape. Phase B is the empirical foundation for the claim's *bounded* interpretation: the framework's gap measures do what they were designed to do (within-instrument coupling validated at r > +0.75; the residual classification correctly distinguishes states with above- and below-baseline supply-vs-coverage relationships); the framework's outputs are not, however, direct outcome-predictors at state granularity, and the Phase B paper draws that distinction in the framework's published documentation.
This is consistent with the Trellison Institute's stated audit methodology: report what the data does show, report what it does not show, refuse to overstate, and require the framework to pass empirical tests proportional to the strength of the policy claims being made on its outputs. Phase B passes the within-instrument validation test (a meaningful test). Phase B does not pass the across-instrument mortality-prediction test (also a meaningful test). Both results are documented in the published paper, and both inform the bounded interpretation of the framework's outputs.
We believe transparent reporting of the orthogonality finding strengthens, rather than weakens, the framework's published case. A framework that claimed to predict mortality at state level and demonstrably did not would have weaker scientific standing than a framework that claims to identify workforce-capacity gaps and demonstrably does.
---
## 6. Limitations
We list eight specific limitations in approximate order of importance:
### 6.1 Cross-sectional design
All correlations are point-in-time. We cannot identify whether supply expansion would *reduce* outcomes; only whether higher-supply states currently have lower outcomes. The latter is a much weaker claim and is the only claim Phase B addresses. Causal identification would require longitudinal panel data (state-level surveillance over multiple cycles, paired with policy-change events as instruments) and is the v2.0 plan.
### 6.2 Temporal mismatch (AADR vintage)
The NCHS Leading Causes of Death state-level AADR file (`bi63-dtpu`) terminates at 2017 in the Socrata-published release. The 2023 YRBSS state-level data is six years more recent. State-level suicide AADR is a slow-moving variable (year-to-year r ≈ 0.95) so the lag introduces noise but not systematic bias, but a 6-year gap is non-trivial. The CDC WONDER XML POST API provides post-2017 state-level AADR with interactive license acceptance; v1.1 will integrate.
### 6.3 All-age mortality vs youth-specific mortality
The NCHS AADR is age-adjusted across the full population, not youth-specific. YRBSS provides youth-specific (grades 9-12) prevalence of suicide ideation/attempts, but the across-instrument comparison is to all-age mortality. Youth-specific state-level suicide mortality (ages 10-19) is available via CDC WONDER's XML POST API with explicit license-acceptance handling. v1.1 will integrate this data source and test whether the orthogonality finding strengthens or attenuates for the youth-specific cohort.
### 6.4 YRBSS state-coverage gap
Our 35-state sample is the intersection of states with releasable YRBSS 2023 state-level Total-demographic data and complete records in the framework's other input streams. The 15 missing states include several of the largest by population (California, New York, Florida, Georgia in some indicators). These states conduct their own state-level youth-health surveillance instruments that are not directly comparable to YRBSS. The 35-state sample over-represents medium-sized and smaller-state populations and the cross-state correlation structure may differ in the population not observed. This is the most consequential single limitation.
### 6.5 NPPES counts structural supply, not delivery
Discussed at length in §5.2.1. The NPI registry measures providers who have completed registration, not providers who are accepting new patients within a usable time window. This is the most consequential measurement limitation and the most likely single source of the non-protective supply signal.
### 6.6 No multivariate confounding adjustment
The cross-state confounding by rural-share, household-gun-ownership rate, opioid-exposure history, and demographic composition is real and not adjusted for in the v1.0 univariate analysis. Specific confounding paths: the rural-state suicide cluster overlapping with the higher-supply rural states (§5.2.3); the deaths-of-despair drug-OD cluster overlapping with the higher-coverage states (§5.4); the state-level Medicaid expansion status (a key policy variable from the Youth V1 negative-outlier analysis) is not adjusted for in the Phase B correlation. v1.1 will extend with multivariate state-level OLS adjusting for these covariates.
### 6.7 No FBI Crime Data Explorer integration
Juvenile arrest data from the FBI Crime Data Explorer (api.usa.gov/crime/fbi/cde/*) requires an `api.data.gov` API key not yet provisioned. The framework's gap measures against state-level juvenile-justice involvement are queued for v1.1.
### 6.8 No HCUP State Emergency Department visit data
HCUP State Inpatient/Emergency Department databases require licensed access via the HCUP Data Distributor. v2.0 will integrate the licensed data, which would allow direct cross-sectional correlation of the framework's gap measures with state-level mental-health-related and drug-overdose-related ER presentations — a closer proxy for "actual care-delivery failure" than NPI counts or self-reported attempted-suicide prevalence.
---
## 7. Conclusion
The Trellison Need-vs-Access Framework's state-level gap measures, applied to the under-18 population in 35 U.S. states with full YRBSS state-level participation, predict within-instrument YRBSS youth suicide ideation and attempts at r > +0.75. The same measures predict NCHS all-age state-level suicide AADR at r = +0.34 and drug overdose at r = +0.19. The framework's supply-side measure shows no protective correlation with any outcome at state granularity, and the framework's residual-classification signal is essentially uncorrelated with mortality. The framework's outputs are therefore properly understood as a state-level *workforce-capacity-gap signal* — calibrated against the prevailing insurance landscape, useful for triaging workforce-expansion programs — and not as a state-level *mortality-risk signal*. The two are different questions and their conflation in policy interpretation mismatches the intervention to the problem.
This is a clarifying finding for the framework's interpretation, consistent with the established health-policy understanding that provider density is a necessary-but-not-sufficient condition for population mental-health outcomes [14, 18]. The framework's contribution is to demonstrate this orthogonality with a reproducible analytical pipeline applied to publicly-available federal data, and to provide a bounded but substantive policy use for the framework's published outputs: identifying which states have under-built youth-serving workforce capacity relative to their insurance coverage landscape, a question that the Youth V1 paper [2] answers with the named positive outliers (PR, NC, NJ).
The full content arsenal — working paper, methodology supplement, dataset CSV (35 states × 11 fields, sha256-stamped), per-state outcome profiles, executive brief, press materials, slides, replication package, animated visualizations — is published under CC-BY-4.0 at https://trellison.com/research/youth-mental-health-outcomes-correlation. The framework Phase A pipeline (`atlas.need_vs_access_framework_v1` v1.1.0) is registered in the DaedArch tool registry; the Phase B correlation analysis is documented inline in the methodology supplement and will be registered as `atlas.outcome_correlation_v1` for cross-study reuse in a forthcoming v1.1 release.
---
## References
[1] Trellison Institute. (2026, May). *The Need-vs-Access Framework v1: A reusable analytical pipeline for state and tract-level public-health gap analysis.* Methodology document accompanying the V1 working paper series. https://trellison.com/methodology/need-vs-access-v1.
[2] Trellison Institute. (2026, May). *The Youth Mental Health Access Gap is Structurally More Severe Than the Adult Gap and Wider Across States.* Working paper v1.0. https://trellison.com/research/youth-mental-health-supply-demand-gap.
[3] Trellison Institute. (2026, May). *The Mental Health Access Gap in the United States Divides into Two Distinct Problems.* Working paper v1.0. https://trellison.com/research/mental-health-supply-demand-gap.
[4] Health Resources and Services Administration. (2024). *Behavioral Health Workforce Education and Training (BHWET) Program.* HRSA-22-064. https://www.hrsa.gov/grants/find-funding/HRSA-22-064.
[5] Substance Abuse and Mental Health Services Administration. (2023). *988 Suicide and Crisis Lifeline: Performance Metrics and State Allocations.* https://988lifeline.org.
[6] Office of the U.S. Surgeon General. (2021). *Protecting Youth Mental Health: The U.S. Surgeon General's Advisory.* U.S. Department of Health and Human Services.
[7] American Academy of Pediatrics, American Academy of Child & Adolescent Psychiatry, & Children's Hospital Association. (2021). *Declaration of a National Emergency in Child and Adolescent Mental Health.*
[8] Centers for Disease Control and Prevention. (2024). *YRBS Methodology Overview.* https://www.cdc.gov/yrbs/about-yrbs/index.html.
[9] Curtin, S. C., & Heron, M. (2019). *Death rates due to suicide and homicide among persons aged 10-24: United States, 2000-2017.* NCHS Data Brief no. 352.
[10] Walker, E. R., Cummings, J. R., Hockenberry, J. M., & Druss, B. G. (2015). *Insurance status, use of mental health services, and unmet need for mental health care in the United States.* Psychiatric Services 66(6): 578-584.
[11] Case, A., & Deaton, A. (2015). *Rising morbidity and mortality in midlife among white non-Hispanic Americans in the 21st century.* PNAS 112(49): 15078-15083.
[12] Case, A., & Deaton, A. (2020). *Deaths of Despair and the Future of Capitalism.* Princeton University Press.
[13] Liu, R. T., Bettis, A. H., & Burke, T. A. (2020). *Characterizing the phenomenology of passive suicidal ideation: A systematic review and meta-analysis of its prevalence, psychiatric comorbidity, correlates, and comparisons with active suicidal ideation.* Psychological Medicine 50(3): 367-383.
[14] Bishop, T. F., Press, M. J., Keyhani, S., & Pincus, H. A. (2014). *Acceptance of insurance by psychiatrists and the implications for access to mental health care.* JAMA Psychiatry 71(2): 176-181.
[15] McBain, R. K., et al. (2019). *Growth and distribution of child psychiatrists in the United States: 2007-2016.* Pediatrics 144(6): e20191576.
[16] Stone, D. M., et al. (2018). *Vital Signs: Trends in State Suicide Rates — United States, 1999-2016.* MMWR 67(22): 617-624.
[17] Knopov, A., et al. (2019). *Household gun ownership and youth suicide rates at the state level, 2005-2015.* American Journal of Preventive Medicine 56(3): 335-342.
[18] Cunningham, P. J. (2009). *Beyond parity: primary care physicians' perspectives on access to mental health care.* Health Affairs 28(3): w490-w501.
[19] Mokdad, A. H., et al. (2018). *The State of US Health, 1990-2016: Burden of Diseases, Injuries, and Risk Factors Among US States.* JAMA 319(14): 1444-1472.
[20] Goldman, M. L., et al. (2020). *Implementing the Certified Community Behavioral Health Clinic (CCBHC) Model.* Journal of Behavioral Health Services & Research 47(2): 219-228.
---
## Supplementary materials
- **S1. Methodology supplement** — Detailed joining, correlation, and outlier-state profiling methodology. `mh_gap_youth_outcomes_v1_methodology_supplement.md`.
- **S2. Dataset** — 35 states × 11 fields, sha256-stamped CSV + manifest. `mh_gap_youth_outcomes_v1_state.csv`.
- **S3. Per-state dartboard outcome profiles** — 10 state narratives (5 framework outliers + 5 pre-specified comparators) with full outcome data and interpretive commentary. `mh_gap_youth_outcomes_v1_dartboard_narratives.md`.
- **S4. Sensitivity analyses** — Sub-sample exclusion, AADR-year stability, taxonomy-set variation, denominator alternatives, outlier exclusion, population-threshold variation. `mh_gap_youth_outcomes_v1_sensitivity_analysis.md`.
- **S5. Reproducibility + replication** — End-to-end reproduction recipe from raw federal data sources to the published joined dataset. `mh_gap_youth_outcomes_v1_replication_README.md`.
- **S6. Bibliography + related work** — 28-reference organized bibliography. `mh_gap_youth_outcomes_v1_bibliography.md`.
- **S7. Authorship + contributor statement** — Detailed contribution attribution. `mh_gap_youth_outcomes_v1_authorship.md`.
---
**Working paper version**: v1.0 draft (peer-review quality)
**Methodology rating**: pending Trellison Institute review
**LedgerWell evidence-chain certificate**: pending
**License**: CC-BY-4.0
**Hub**: https://trellison.com/research/youth-mental-health-outcomes-correlation
**Conflict of interest disclosure**: The authors operate the DaedArch platform that hosts the Need-vs-Access Framework. The framework is published under CC-BY-4.0 with no commercial licensing component; the DaedArch platform's revenue model is independent of any single research-output publication. No external grants or contracts support this study.
**Author contributions** are documented in `mh_gap_youth_outcomes_v1_authorship.md`. Editorial responsibility for the published version rests with the human author (Rob Stillwell); the DaedArch AI co-author contributed analytical orchestration, data curation, prose drafting, and methodology design within the governance bounds of the DaedArch Day-Zero protocol (April 2026).
**Correspondence and material requests**: [email protected].
**Data availability**: All data are derived from public federal sources (CDC YRBSS, NCHS bi63-dtpu, NCHS VSRR xkb8-kh2a, ACS via the Census API). The full joined dataset is published at https://api.daedarch.ai/api/v4/t/media/serve?bucket=daedarch-public-media&object=youth_mental_health_outcomes/dataset_v1/mh_gap_youth_outcomes_v1_state.csv with sha256 stamped in the accompanying manifest.json. Reproduction recipe in §S5.
**Code availability**: The framework Phase A pipeline (`atlas.need_vs_access_framework_v1` v1.1.0) is registered as a DB-native tool in the DaedArch tool registry. The Phase B correlation analysis is documented inline in the methodology supplement (§S1). License: CC-BY-4.0.