“None” means AHI <5
The reference group comprises referred adults with sleep studies and AHI below 5. It is not a population-level healthy control. Severe versus None is the default pairwise contrast.
Research guide
Methods & interpretation
How to read OSA associations across disease endpoints, adjustment models, statistical flags, and staged atlas releases.
Study frame
Cohort, reference, and time direction
Interpret every estimate within the sleep-clinic referral cohort.
Disease already present
L1-regularized logistic models summarize disease presence before index as odds ratios. Estimates are shrunk and intervals are non-classical; they should be labeled accordingly.
Disease arising during follow-up
Ridge cause-specific Cox models follow at-risk participants after index and report hazard ratios. At-risk counts, event counts, and proportional-hazards diagnostics belong with each result.
Odds ratios and hazard ratios answer different questions. They are shown in separate views and should not be compared as though they were the same estimand.
Questionnaire-wide association study
Questionnaire responses anchored to the sleep-study index
QWAS asks how questionnaire responses vary with OSA severity within the referred cohort.
74,061 pre-index questionnaires
The analysis selects the closest questionnaire on or before the sleep-study index for each participant. It is cross-sectional, not a one- or five-year longitudinal window, and no selection weighting was applied for questionnaire availability.
83 binary · 49 continuous
Binary positive-or-worse response states use logistic odds ratios. Ordinal and continuous responses are rank-inverse-normalized before linear modeling and report standardized betas. The 92 unique items include 40 represented in both forms; the two scales never share an axis.
M4 adds BMI
M1 includes the OSA exposure; M2 adds age, sex, race/ethnicity, and study year; M3 adds smoking, income, and observation window; M4 adds BMI. The AHI below 5 reference remains a symptomatic referral group, not a population control.
Models are complete-case for each item. Displayed N is the answered sample supplied to the source formula and can exceed the fitted N after covariate missingness. FDR and Bonferroni flags are calculated within contrast, model, and effect type. STOP1, STOP3, GASP, and the snoring/apnea chief-complaint item are flagged because they can help drive referral, making ascertainment part of their observed association. Numbered instrument labels are conservative concepts rather than verbatim questionnaire text and remain under review.
Explore QWAS →Cross-domain OSA phenotypes
Three gradients combined into eight octants
The taxonomy separates physiologic severity, symptom burden, and coded comorbidity burden.
Complementary latent axes
Model-derived scores summarize sleep-study physiology, symptoms, and comorbidities. The axes are nearly independent and should be described as regions of continuous gradients—not biological subpopulations.
Median split × three axes
Each score is split at its shared-cohort median, classifying all 70,880 people into eight near-balanced groups. The three-square glyph records whether physiology, symptoms, and comorbidity are above the median, in that order.
Cohort-specific research taxonomy
Cut points are not clinical thresholds and do not transfer unchanged. The comorbidity axis also reflects healthcare contact and record completeness; downstream M4 models adjust for age and sex because octants are strongly structured by both.
Measure-specific denominators
The phenotype explorer reports 26 aggregate measures with their own coverage. Follow-up and encounter measures cover the 25,380-person IncWAS spine; age, BMI, and sex cover 53,012 people. The displayed standardized difference is the octant estimate minus the containing cohort estimate, divided by cohort SD—not a two-independent-group SMD.
M4 ridge Cox
Each focal octant is compared with the pooled other seven. M4 adjusts for an age spline, sex, study year, race/ethnicity, smoking, income, baseline observation and encounter measures, and BMI. CPAP is not adjusted because it is treated as a mediator.
Omnibus gate, then one-vs-rest
All eight contrasts are shown only for outcomes whose 7-df M4 omnibus test reached nominal p < 0.05. Within each focal-octant contrast, Bonferroni uses 0.05/15 for gated PheCodes and 0.05/6 for gated systems; it is not 0.05/168.
Octant cumulative-incidence curves are unadjusted tie-aware Aalen–Johansen estimates with death competing; M4 hazard ratios are adjusted estimates from a separate model. Mean follow-up is 1.42 years and only 8.7% of the pooled baseline risk set remains observed at year 3. Eighty of 168 panels have EPV below 10, and proportional-hazards diagnostics were not computed for any octant model.
The phenotype explorer releases documented PAP setup, retained-record, adherence-field, and usage-field coverage across five display windows. These are extract-documentation measures, not true initiation or adherence: absent rows cannot distinguish no treatment from no capture, and fixed-window eligibility uses an existing diagnosis, insurance, and death censor rather than a confirmed PAP-capture boundary. Adherent percentages and usage distributions remain unavailable because those estimates are absent and their implementation or averaging window is not fully confirmed. Race/ethnicity remains an M4 adjustment covariate, but its descriptive octant breakdown is not published because one cell is below the disclosure threshold and release is not approved. Curve coordinates are withheld for the 16 panels whose focal event count is suppressed below 11.
Explore octant phenotypes →Incidence PheDAS survival analysis
Cumulative incidence, with death treated as competing
The curves describe absolute observed incidence; they are not M4-adjusted hazard ratios.
Aalen–Johansen curves
Severity curves report cumulative incidence from index through six years. Landmark CPAP curves report five years from index plus the selected grace period. Both treat death as a competing event and do not use one minus Kaplan–Meier.
Post-selection description
Curves are available for the 40 PheCodes that were FDR-significant in at least one M4 OSA-severity contrast, including one visibly labeled OSA-recoding control. The curves describe those selected outcomes; they are not independent confirmation.
Landmark-based, still descriptive
The 90- and 180-day grace-period landmarks require OSA patients to remain observed and event-free to the landmark, then restart the analysis clock. This design addresses immortal-time bias. It does not remove healthy-adherer confounding, so curves remain associational rather than treatment effects; 180 days is the primary view.
Aggregate cumulative-incidence percentages are released publicly. Exact monthly at-risk and cumulative-event arrays, uncertainty bands, and curve downloads remain outside the browser bundle. Follow-up ends May 31, 2023, and late curve tails may be supported by thinner risk sets.
Explore cumulative incidence curves →PheDAS adjustment ladder
Read disease-model M1 through M4 as a sequence
ModelFramingInterpretationDisplay
M1CrudeUnadjusted association model.Comparison
M2DemographicsAdds demographic adjustment.Comparison
M3Extended · excludes BMIAdds family-specific adjustment while excluding BMI. Incidence PheDAS M3 also includes baseline encounter rate.Adjacent
M4Primary · adds BMIAdds BMI to M3. M3 remains adjacent because BMI may be treated as a confounder or as part of the pathway, depending on the scientific question.Primary
Model labels are family-specific. Other WAS views therefore retain neutral M1-M4 identifiers and do not assume that their covariate sets are interchangeable with PheDAS.
Statistical reading
Multiplicity, diagnostics, and uncertainty
Significance is context for an estimate, not a scientific conclusion.
Keep correction families intact
Current disease outputs expose FDR and Bonferroni significance flags. Exact adjusted q-values are not generally available. The atlas does not infer one global correction across incompatible analyses, models, or contrasts.
Warnings stay visible
Unstable estimates remain labeled rather than silently removed. Incidence results additionally expose event-per-variable warnings and proportional-hazards diagnostic information where available.
Association, not causation
Results support hypothesis generation. They do not establish causal, protective, or independent risk effects, and they do not estimate population OSA prevalence.
Other association families
Keep each clinical signal on its native scale
AnalysisEstimandInterpretation ruleDefault
LabWAS
β / OR / IRRMean and median value betas use rank-inverse-normal standard-deviation units and condition on being tested. Ordering propensity and ordering rate remain separate extensive and intensive signals.M4 · Severe vs NoneMedWAS
ORBinary ever-filled medication classes use zero-preserved GPI-4 codes. Fallback class labels remain visibly marked for review.M4 · Severe vs NoneBehWAS
OR or βBinary and rank-inverse-normal standardized continuous representations remain on separate axes. This small curated scan is preliminary and forest-first.M4 · Severe vs NoneQWAS
OR or βBinary positive-or-worse response states use logistic odds ratios; continuous and ordinal responses use rank-inverse-normal betas. Both are itemwise complete-case and index-anchored.M4 · Severe vs NoneProcWAS
IRRProcedure counts are annualized rates over one- and five-year pre-index windows; sleep-study and PAP-pathway procedures were excluded.M4 · Severe vs NoneUtilWAS
OR / IRRPresence and count among present are a two-part analysis. Physician and allied-health definitions must not be summed while their overlap rule remains under review.M4 · Severe vs NoneOmnibus rows are non-directional Wald statistics. They appear only in Manhattan and table views, never on a signed volcano or forest axis.
Release roadmap
Eight clinical lenses, released in stages
AnalysisScientific focusStatusRelease rule
Prevalence PheDASExisting disease · ORResearch previewDisease scan
Incidence PheDASNew-onset disease · HRResearch previewDisease scan
LabWASValues, ordering propensity, and ordering rateArchived snapshotVisible with snapshot or review warning
MedWASMedication fills · ORArchived snapshotVisible with snapshot or review warning
BehWASEHR-derived behaviors · OR or βPreliminary archivedVisible with snapshot or review warning
QWASIndex-anchored questionnaire responses · OR or βResearch previewDisease scan
ProcWASProcedure rates · IRRArchived snapshotVisible with snapshot or review warning
UtilWASHealthcare-use presence and countsArchived reviewVisible with snapshot or review warning
Ready to explore?
Begin with the primary M4 Severe-versus-None view, then inspect M3 and the alternate contrasts before drawing a scientific conclusion.
Explore other WAS families →