Patient Health Questionnaire-9 (PHQ-9)

Structural validity Highcontestedevidence form: canonical

general nationally representative data; also university students, traumatic brain injury, stroke and coronary heart disease samples (flag basis: Korean nationally representative data; Brazilian university students; traumatic brain injury patients; stroke sample; Italian coronary heart disease sample)

state: assessed · rubric v1.6 · literature as of 2026-09-01 · grade confirmed 2026-09-03 · under rubric v1.3: precondition evidence 5 entries → 3 entries (re-read before first publication, correction C-0007) · under rubric v1.1: indirectness direct → general (re-read before first publication, correction C-0005)

The dimensionality of the PHQ-9 is genuinely contested, but the practical reading across studies is that it behaves as an essentially unidimensional depression severity scale even where a two-factor model fits marginally better. The original development treated the total score as a single severity dimension (Kroenke 2001). Within the individual participant data (IPD) programme, a comparison of unidimensional, two-dimensional (cognitive/affective versus somatic) and bifactor latent models found that all fitted reasonably and that scoring by the more complex latent models improved sensitivity only marginally (by about 0.04 to 0.05) while reducing specificity, relative to the sum score at a cut-off of 10 or above (Fischer et al 2021). Several population studies reach the same conclusion from different angles: in Korean nationally representative data a single-factor model fitted well (CFI 0.944) with satisfactory internal consistency (Lee et al 2023); in Brazilian university students both one- and two-dimensional models fitted, with the authors favouring the unidimensional solution (Rufino et al 2024); and in traumatic brain injury patients a general factor explained about 85% of the variance despite a slight bifactor improvement (Teymoori et al 2020). Where a two-factor (somatic versus cognitive-affective) structure is reported, the factors tend to be very highly correlated: in a stroke sample the two-factor model fitted slightly better than the one-factor model (CFI 0.984 versus 0.974) but the inter-factor correlation of 0.866 pointed back to unidimensionality (Blake et al 2024), and an Italian coronary heart disease sample reported a bi-dimensional somatic/cognitive solution (Di Matteo et al 2025). The disagreement is therefore about whether the somatic items form a distinguishable subfactor, not about whether a single severity score is defensible; the weight of evidence supports scoring and interpreting a single total. Sample sizes were 6000 primary care and obstetrics-gynaecology patients in the development study (Kroenke 2001), 24 calibration studies with 4378 participants and 17 validation studies with 4252 participants in the IPD comparison, where sensitivity gains were 0.036, 0.050 and 0.049 and specificity losses 0.017, 0.026 and 0.019 for the unidimensional, two-dimensional and bifactor models (Fischer et al 2021), 5347 Korean adults, with SRMR 0.040, RMSEA 0.076 and omega 0.812 for the one-factor model (Lee et al 2023), 3163 Brazilian university students (Rufino et al 2024), 2137 patients at six months after traumatic brain injury (Teymoori et al 2020), 787 stroke and 12,016 non-stroke participants with a propensity-matched comparison subsample of 1574 (Blake et al 2024) and 427 Italian coronary heart disease patients, with model-based reliability 0.80 (Di Matteo et al 2025).

High precondition (rubric 1.6): the cited studies that meet the High precondition, each with a sample size and a statistic of this property.

Confidence note (legacy, first pass; not the basis of the grade): High: many good-quality factor-analytic studies across large and varied samples, converging on essentially unidimensional use despite a recurring, well-characterised somatic/cognitive two-factor debate.

Convergent and discriminant validity Moderatecontestedevidence form: canonical

general general population; also students, traumatic brain injury and coronary heart disease samples (flag basis: Chinese general population; students; traumatic brain injury patients; coronary heart disease patients)

state: assessed · rubric v1.6 · literature as of 2026-09-01 · grade confirmed 2026-09-03 · under rubric v1.1: indirectness direct → general (re-read before first publication, correction C-0005)

Convergent evidence is consistent in direction but uneven in magnitude, and discriminant separation from anxiety is imperfect. In development, higher PHQ-9 scores tracked substantial decrements across all six SF-20 functional-status subscales and greater symptom-related difficulty, supporting construct validity (Kroenke 2001). In the Chinese general population the PHQ-9 correlated negatively with SF-36 subscales (r from -0.11 to -0.47) as expected, but its correlation with the Zung Self-Rating Depression Scale was unexpectedly weak (r = 0.29), an inconsistency worth noting against the usual assumption of strong convergence with other depression measures (Wang et al 2014). Convergent relationships with sleep quality, alcohol use and physical activity have also been reported in students (Rufino et al 2024). On discriminant validity, the PHQ-9 and the GAD-7 anxiety scale share a large common factor: in traumatic brain injury patients a general distress factor dominated both scales (about 85% of variance) even though the instruments related differently to SF-36 subscales (Teymoori et al 2020), and in coronary heart disease patients PHQ-9 and GAD-7 scores were significantly positively correlated (Di Matteo et al 2025). Depression and anxiety symptoms measured this way are statistically distinguishable but strongly overlapping.

Confidence note (legacy, first pass; not the basis of the grade): Moderate: multiple studies establish expected convergent and functional-status correlations, but magnitudes vary (including one weak convergent correlation) and discriminant separation from anxiety is only partial.

Criterion validity: reference standard Moderatewell-establishedevidence form: canonical

indirect clinic patients and diagnostic meta-analysis samples; no working-adult or general-population sample (flag basis: primary-care and obstetric-gynaecology clinic patients; major depression cases)

state: assessed · rubric v1.6 · literature as of 2026-09-01 · grade confirmed 2026-09-03 · under rubric v1.2: grade High → Moderate; precondition evidence 5 entries → none (re-read before first publication, correction C-0006)

Criterion validity against a diagnostic interview for major depression is the PHQ-9's best-evidenced property, and it is strong. In the original validation in primary-care and obstetric-gynaecology clinic patients a cut-off of 10 or above gave 88% sensitivity and 88% specificity against an independent mental-health-professional interview (Kroenke 2001). The individual participant data meta-analysis programme is the anchor here. The 2019 IPD meta-analysis (58 studies, n = 17,357, 2,312 major depression cases) found combined sensitivity and specificity maximised at a cut-off of 10 or above against semistructured interviews (sensitivity 0.88, 95% CI 0.83 to 0.92; specificity 0.85, 0.82 to 0.88) (Levis et al 2019). The 2021 update (100 studies, n = 44,503) reproduced this at the same cut-off (sensitivity 0.85, 0.79 to 0.89; specificity 0.85, 0.82 to 0.87) (Negeri et al 2021). A crucial nuance is reference-standard dependence: sensitivity against semistructured clinician interviews ran higher than against fully structured lay interviews (5 to 22 percentage points) or the MINI (2 to 15 points) in the 2019 synthesis (Levis et al 2019), with median differences of around 21% and 11% in the 2021 update, while specificity was similar across standards (Negeri et al 2021). The diagnostic algorithm approach performed worse than the cut-off (algorithm sensitivity around 0.57 to 0.61 versus 0.88 for the cut-off against semistructured interviews) (He et al 2019), and the PHQ-2 with cut-off 3 or above (sensitivity 0.72, specificity 0.85) is a reasonable first-stage screen followed by the PHQ-9 (Levis et al 2020). The original criterion sample was 580 patients drawn from the 6,000 who completed the PHQ-9 across 8 primary care and 7 obstetrics-gynaecology clinics (Kroenke 2001); the cut-off 10 estimate in the 2019 synthesis rests on 29 semistructured-interview studies with 6,725 participants (Levis et al 2019); the algorithm meta-analysis pooled 54 studies (n=16,688, 2,091 cases) and found algorithm specificity of 0.95 against 0.86 for the cut-off (He et al 2019); and the PHQ-2 analysis drew on 100 studies (44,318 participants, 4,572 with major depression), with an area under the curve of 0.88 for semistructured interviews and a PHQ-2 (2 or above) then PHQ-9 (10 or above) sequence giving sensitivity 0.82 and specificity 0.87 (Levis et al 2020).

Confidence note (legacy, first pass; not the basis of the grade): High for diagnostic criterion validity against clinical interviews (large, consistent IPD evidence); Absent for criterion validity against organisational/workplace outcomes (no such studies retrieved). The overall grade is split by criterion.

Criterion validity: organisational Absent (searched; none found in the sweep to date)untestedevidence form: canonical

absence type: population-general searched; none found in the sweep to date

state: assessed_absent · rubric v1.6 · literature as of 2026-09-01 · grade confirmed 2026-09-03

No study retrieved in this pass tested PHQ-9 scores against measured sickness absence, staff turnover, productivity loss or other workplace outcomes as criteria. The closest health-related association is in the development study, where self-reported sick days and health-care utilisation rose with PHQ-9 severity (Kroenke 2001). Criterion validity against organisational endpoints is therefore unestablished.

Internal consistency Highwell-establishedevidence form: canonical

general general population and a nationally representative sample; the pooled studies name no population (flag basis: Chinese general population; Korean nationally representative sample)

state: assessed · rubric v1.6 · literature as of 2026-09-01 · grade confirmed 2026-09-03 · under rubric v1.1: indirectness direct → general (re-read before first publication, correction C-0005)

Internal consistency is consistently high across populations. A reliability generalisation meta-analysis of 60 studies (232,147 participants) estimated a pooled Cronbach's alpha of 0.86 (95% CI 0.85 to 0.87), with self-administered formats slightly higher (alpha 0.87) than face-to-face interview administration (alpha 0.80), though between-study heterogeneity was very large (I-squared 99.3%) (Ajele 2025). Single-study estimates sit in the same range: alpha 0.86 in the Chinese general population (Wang et al 2014) and omega 0.812 in a Korean nationally representative sample (Lee et al 2023). A systematic review of the wider PHQ family likewise reported good internal consistency for the PHQ-9 (Kroenke 2010), and a meta-analysis of the Spanish-language versions evaluated internal consistency alongside accuracy (Martinez et al 2023). Reported coefficients of omega tend to align with alpha where both are given, but omega is reported far less often than alpha. Single-study sample sizes were n = 1045 Shanghai community residents (Wang et al 2014) and n = 5347 Korean adults over 19 (Lee et al 2023); the Spanish-language meta-analysis pooled 10 studies with 5164 adults and reported PHQ-9 Cronbach's alpha of 0.78 to 0.90 and McDonald's coefficient of 0.79 to 0.90 (Martinez et al 2023), and the PHQ family review drew on four multisite studies of 9740 patients but its abstract gives no internal consistency coefficient (Kroenke 2010).

High precondition (rubric 1.6): the cited studies that meet the High precondition, each with a sample size and a statistic of this property.

Confidence note (legacy, first pass; not the basis of the grade): High: a large reliability-generalisation meta-analysis plus multiple primary studies converge on alpha around 0.86, albeit with substantial heterogeneity and sparse omega reporting.

Test-retest reliability Moderatethinevidence form: canonical

general general population; also a late-life depression treatment sample and primary care patients (flag basis: Chinese general population; late-life depression treatment sample; Spanish primary care patients)

state: assessed · rubric v1.6 · literature as of 2026-09-01 · grade confirmed 2026-09-03 · under rubric v1.1: indirectness direct → general (re-read before first publication, correction C-0005)

Test-retest reliability exists but is markedly under-studied relative to the vast diagnostic-accuracy literature, and this asymmetry is itself the finding. The reliability generalisation meta-analysis could pool a test-retest estimate from only 8 of its 60 studies, yielding 0.82 (95% CI 0.74 to 0.90) (Ajele 2025). Primary estimates are favourable where they exist: a two-week retest correlation of 0.86 in the Chinese general population (Wang et al 2014), test-retest described as excellent over a seven-day interval in a late-life depression treatment sample (Lowe et al 2004), and intraclass agreement assessed for self- versus telephone-administration in Spanish primary care (Pinto-Meza et al 2005). No test-retest study in an occupational or workplace sample was located in this pass. The property is present and reassuring in the settings studied, but the evidence base is thin and skewed toward clinical and general-population samples over short intervals.

Confidence note (legacy, first pass; not the basis of the grade): Moderate: several favourable estimates (roughly 0.82 to 0.86) exist and a meta-analytic pooled value is available, but from few studies (n = 8 pooled), short intervals, and no workplace data. Not absent, but comparatively neglected.

Measurement invariance Lowthinevidence form: canonical

general nationally representative sample; also coronary heart disease and implantable-defibrillator cohorts (flag basis: Korean nationally representative sample; Italian coronary heart disease cohort; Danish implantable-defibrillator cohort; clinical or general-population samples)

state: assessed · rubric v1.6 · literature as of 2026-09-01 · grade confirmed 2026-09-03 · under rubric v1.1: indirectness direct → general (re-read before first publication, correction C-0005)

Invariance has been tested piecemeal and mostly reaches metric to scalar level within the groups examined, but coverage of occupational populations is absent. In a Korean nationally representative sample the one-factor PHQ-9 showed equivalent structure, factor loadings and item intercepts across age groups, i.e. up to scalar invariance across ages (Lee et al 2023). Invariance across sex and across age (65 and over versus under 65) was tested in an Italian coronary heart disease cohort (Di Matteo et al 2025). Item response theory analysis in a large Danish implantable-defibrillator cohort found no differential item functioning across educational level, age, clinical indication or heart-failure severity, with only a single item showing DIF by gender (Pedersen et al 2016). This points to broadly stable measurement across sex and age in clinical and general populations. However, the studies are confined to clinical or general-population samples; invariance across occupations, across employed versus unemployed status, and over repeated workplace administrations was not established in retrieved evidence. Longitudinal (over-time) invariance is likewise weakly evidenced for the PHQ-9 specifically.

Confidence note (legacy, first pass; not the basis of the grade): Low to Moderate: scalar invariance is demonstrated across age (one strong study) and DIF is minimal in another, but evidence is scattered across clinical populations, occupational invariance is untested, and over-time invariance is weak.

Responsiveness and MIC Moderatethinevidence form: canonical

indirect late-life depression trial participants; no working-adult or general-population sample (flag basis: IMPACT late-life depression trial)

state: assessed · rubric v1.6 · literature as of 2026-09-01 · grade confirmed 2026-09-03

Responsiveness to change is well supported, and a minimal important change has been proposed, though the MIC rests largely on one study. In the IMPACT late-life depression trial (n = 434 intervention participants) the PHQ-9 was responsive to treatment, with an effect size (about -1.3 at three months) exceeding the SCL-20 depression scale at three months and matching it at six months; change scores discriminated persistent depression, partial remission and full remission against structured diagnostic interviews (Lowe et al 2004). That study estimated a minimal clinically important difference for individual change of about 5 points on the 0 to 27 scale, derived as two standard errors of measurement (Lowe et al 2004). A systematic review of the PHQ family concluded that sensitivity to change is well established for the PHQ-9 (Kroenke 2010). Routine-outcome use in stepped-care services also relies implicitly on responsiveness, with large pre-post effect sizes reported for depression (Richards 2009). The MIC of 5 points is widely cited but should be understood as a single-derivation estimate rather than a triangulated consensus value.

Confidence note (legacy, first pass; not the basis of the grade): Moderate: responsiveness is demonstrated in a good treatment study and endorsed by a review, but the 5-point MIC derives essentially from one study and one method (2 SEM), and no MIC has been established in a workplace context.