General Health Questionnaire-12 (GHQ-12)

Structural validity Highcontestedevidence form: canonical

direct - large representative general-population and survey samples (England, Spain, Germany) plus meta-analytic aggregation; the definitive evidence is not workplace-specific but the structural conclusion is not population-dependent.

The internal structure is the central and long-running controversy for this instrument, and it is driven by scoring method rather than by any true multidimensionality. Although Goldberg designed the GHQ-12 as a unidimensional measure, factor-analytic studies have reported one-, two- and three-factor solutions, the best known being Graetz's three factors (anxiety/depression, social dysfunction, loss of confidence) and various two-factor solutions splitting positively and negatively worded items. Hankins 2008 showed in Health Survey for England data (n=3705) that the best fitting model is one-dimensional with a response-bias factor on the negatively phrased items, concluding that earlier multifactor structures were artefacts of the analysis method and of the mixing of positively and negatively worded items across scoring schemes (Likert, GHQ 0-0-1-1 and C-GHQ). Rey et al. 2014 replicated this in a large Spanish sample (n=27,674), finding that spurious multidimensionality appears only under corrected and Likert scoring because of ambiguous response categories in the negative items, and recommending standard GHQ scoring with a single global score. Romppel et al. 2013 reached the same conclusion in a German population sample (N=2041), with subscale correlations against external criteria (BDI, PHQ-2, SF-36) not differing substantially from one another. The strongest evidence comes from Gnambs and Staufenbiel 2018, two meta-analyses (summary data from 38 studies, total N=76,473; and individual responses from 84 samples, N=410,640): confirmatory and bifactor modelling showed that although two wording-based factors are recoverable, almost all common variance loads on a general factor, so the GHQ-12 is essentially unidimensional and subscale scores should not be interpreted. Hystad and Johnsen 2020 corroborated the bifactor structure (general factor plus positive- and negative-wording method factors) in military samples. The modern consensus is therefore essential unidimensionality with method effects from item wording; the historical multifactor literature is treated here as contested and largely superseded.

Convergent and discriminant validity Moderatewell-establishedevidence form: canonical

indirect - convergent coefficients come from non-UK, non-workplace samples (India, Australia) and clinical cancer cohorts; no UK working-adult convergent study was located in the latest review pass.

Convergent evidence is consistent though less systematically studied than structure or criterion validity. The GHQ-12 total correlates moderately with other distress and wellbeing measures: r=0.58 with the Subjective Well-being Inventory in older Indian adults (Qin et al. 2018). In head-to-head comparisons against diagnostic interviews the GHQ-12 discriminates common mental disorders comparably to the SRQ, K10, K6 and PHQ (Patel et al. 2007; Gill et al. 2007), although Gill et al. 2007 found the SF-12 mental component and the K6/K10 outperformed the GHQ-12 for detecting diagnosed depression. In UK cancer patients the GHQ-12 tracked the Distress Thermometer and HADS over time (Gessler et al. 2008). Discriminant separation from distinct constructs is less directly documented in the sources retrieved.

Criterion validity: reference standard Highwell-establishedevidence form: canonical

indirect - reference-standard validation is extensive but drawn from general medical, primary-care, community and international samples, and (for Baksheev) adolescents; no criterion validation against a diagnostic standard in a UK working-adult sample was located in the latest review pass.

Validity against a diagnostic reference standard is the GHQ-12's best-evidenced property. In the WHO multi-centre study of mental illness in general health care (5438 patients, 15 centres, CIDI primary-care interview), the GHQ-12 achieved a mean area under the ROC curve of 0.88 (range 0.83 to 0.95), performing as well as the longer GHQ-28, with complex scoring offering no advantage and no significant effect of gender, age or education on validity (Goldberg et al. 1997). Against the CIS-R in Indian primary care the GHQ-12 was among the best of five screeners, though positive predictive value at optimal cut-offs was modest across all five screeners compared, ranging from 51% to 77% (Patel et al. 2007). Against the CIDI in an Australian general-population survey (N=10,504) the GHQ-12 discriminated depression and anxiety, albeit less well than the K6/K10 for depression (Gill et al. 2007). In adolescents against a SCID DSM-IV-TR interview the AUC was 0.781, with sex-specific optimal thresholds (Baksheev et al. 2011). Case-detection performance is thus robust and reproducible across languages and settings, but is anchored in primary-care, community and clinical samples rather than workplaces.

Criterion validity: organisational Absent (a finding about the literature)untestedevidence form: canonical

direct - the absence applies squarely to the UK working-adult deployment question; no evidence located to grade.

No study validating the GHQ-12 against organisational outcomes (sickness absence, staff turnover, job performance, or diagnosed conditions recorded in a work context) as a criterion was located in the latest review pass. The GHQ-12 is very widely used within occupational and workforce samples, for example among South African healthcare workers where its reliability and dimensionality were examined (Kufe et al. 2024) and in UK primary-care detection studies (Plummer et al. 2000), but in those studies the GHQ is deployed as the distress measure itself, not validated against a separate work-outcome standard. The absence of organisational criterion evidence is reported here as the finding it is: the instrument's predictive relationship to absence, turnover or performance in the workplace was not established in the retrieved literature.

Internal consistency Highwell-establishedevidence form: canonical

indirect - alpha estimates come from England (general population), India and international student/community samples rather than UK working adults specifically, though the property is stable across populations.

Internal consistency is consistently high. Cronbach's alpha of about 0.90 has been reported under Likert and GHQ scoring in Health Survey for England data, falling to 0.75 under C-GHQ scoring (Hankins 2008); alpha of 0.9 in older Indian adults (Qin et al. 2018); and a lower KR-20 of 0.70 under dichotomous scoring in Colombian students (Simancas-Pallares et al. 2017). Values across the wider literature typically sit in the 0.80 to 0.90 range. An important caveat is that Hankins 2008 showed Cronbach's alpha overestimates reliability once the negative-item response bias is modelled, so the true measurement precision is somewhat lower than the headline alpha values suggest.

Test-retest reliability Lowthin

indirect - the single located estimate is from Japanese young adults over a two-week interval, not UK working adults; it is also a two-way random-effects ICC for agreement rather than a workplace test-retest.

Genuine test-retest (stability) evidence for the GHQ-12 is thin. Only one clean estimate was located in the latest review pass: an intraclass correlation exceeding 0.70 over a two-week interval in 137 young adults, reported alongside a standard error of measurement of 1.47 (bimodal scoring) and 2.44 (Likert scoring) and corresponding smallest detectable change values (Ohno et al. 2017). Because the GHQ measures a current, changeable state rather than a trait, high test-retest coefficients are not necessarily expected, and the scarcity of stability data is itself a graded finding. Reported alpha values from other studies must not be mistaken for retest reliability.

Measurement invariance Lowthinevidence form: mixed

indirect - invariance evidence comes from military, international and cross-language samples; occupation-level and UK working-adult invariance were not established in the latest review pass.

Invariance evidence is partial. Goldberg et al. 1997 found no significant effect of gender, age or educational level on the validity of the GHQ across the WHO multi-centre sample, and validity coefficients held across ten translated languages, supporting broad cross-population comparability of case detection. Hystad and Johnsen 2020 demonstrated that a bifactor structure was invariant across two independent military samples and, in a multi-group analysis, across time. The meta-analytic confirmation of a common structure across 84 samples (Gnambs and Staufenbiel 2018) is consistent with structural stability across populations. However, formal configural/metric/scalar invariance testing across sex, age band and, in particular, occupation is not comprehensively established in the retrieved literature, and translated versions are noted by the steward as not all validated by the publisher.

Sub-grades (evidence differs by subgroup):

• {"subgroup": "sex / age / education", "grade": "Moderate", "note": "no effect on validity across the WHO multi-centre study (Goldberg et al. 1997); reasonably supported."}
• {"subgroup": "across time / repeated samples", "grade": "Low", "note": "bifactor structure invariant across two military samples and over time (Hystad and Johnsen 2020); limited to one military cohort."}
• {"subgroup": "occupation", "grade": "Absent", "note": "no formal invariance test across occupational groups located in the latest review pass."}

Responsiveness and MIC Very lowthinevidence form: canonical

indirect - the change evidence comes from a UK cancer cohort and Japanese young adults, not UK working adults; no anchor-based minimal important change exists.

Formal responsiveness and a minimal important change value are not established. Gessler et al. 2008 found that GHQ-12 scores changed over four and eight weeks in the same direction as the HADS and Distress Thermometer in UK cancer outpatients, providing indirect support for sensitivity to change. Ohno et al. 2017 derived smallest detectable change values from the standard error of measurement (for example a smallest detectable change of about 4.06 points at the individual level under bimodal scoring), which bounds interpretable change but is a distribution-based statistic, not an anchor-based minimal important change. No validated minimal important change for the GHQ-12 was located in the latest review pass.