Background: The twelve-item General Health Questionnaire (GHQ-12) was developed to screen for non-specific psychiatric morbidity. It has been widely validated and found to be reliable. These validation studies have assumed that the GHQ-12 is one-dimensional and free of response bias, but recent evidence suggests that neither of these assumptions may be correct, threatening its utility as a screening instrument. Further uncertainty arises because of the multiplicity of scoring methods of the GHQ-12. This study set out to establish the best fitting model for the GHQ-12 for three scoring methods (Likert, GHQ and C-GHQ) and to calculate the degree of measurement error under these more realistic assumptions.
Methods: GHQ-12 data were obtained from the Health Survey for England 2004 cohort (n = 3705). Structural equation modelling was used to assess the fit of  the one-dimensional model  the current 'best fit' three-dimensional model and  a one-dimensional model with response bias. Three different scoring methods were assessed for each model. The best fitting model was assessed for reliability, standard error of measurement and discrimination.
Results: The best fitting model was one-dimensional with response bias on the negatively phrased items, suggesting that previous GHQ-12 factor structures were artifacts of the analysis method. The reliability of this model was over-estimated by Cronbach's Alpha for all scoring methods: 0.90 (Likert method), 0.90 (GHQ method) and 0.75 (C-GHQ). More realistic estimates of reliability were 0.73, 0.87 and 0.53 (C-GHQ), respectively. Discrimination (Delta) also varied according to scoring method: 0.94 (Likert method), 0.63 (GHQ method) and 0.97 (C-GHQ method).
Conclusion: Conventional psychometric assessments using factor analysis and reliability estimates have obscured substantial measurement error in the GHQ-12 due to response bias on the negative items, which limits its utility as a screening instrument for psychiatric morbidity.