Screening for depression in primary careBMJ 2003; 326 doi: https://doi.org/10.1136/bmj.326.7396.982 (Published 03 May 2003) Cite this as: BMJ 2003;326:982
Scientific and statistical errors should have been picked up in peer review
- William P Plummer (email@example.com), consultant psychiatrist
- East Kent Community NHS Trust, St Martin's Hospital, Canterbury, Kent CT1 1AZ
- Section of Social and Epidemiological Psychiatry, Department of Psychiatry, University of Leicester, Leicester LE5 4PW
- Tayside Centre for General Practice, University of Dundee, Dundee DD2 4AD
- Department of Epidemiology and Public Health, University of Dundee, Ninewells Hospital and Medical School, Dundee DD1 9SY
- Department of Psychiatry, Ludwig-Maximilians-University Munich, Nußbaumstrasse 7, D-80336 Munich, Germany
EDITOR—The paper by Henkel et al contains several scientific and statistical errors, which should have been identified and eliminated by the BMJ's peer review process before publication.1 The study is described as a cohort study but is a cross sectional study.
The general health questionnaire is inappropriate as a screening test for depression and should not be compared with a gold standard test that is specific for depression. It is a screening test for the “common mental disorders” that often occur in primary care,2 with a good sensitivity but poor specificity for depression.
Henkel et al choose sensitivity and negative predictive value as measures of how well their tests perform. Sensitivity should be paired with specificity. Sensitivity and negative predictive value are not independent.
They do not make clear their primary hypothesis or explain what questions they are seeking to examine in their analysis. With so many different measures that they could compare, a plan of analysis is essential. They could, for example, start with an omnibus test for sensitivity, such as Cochran's Q.3
Henkel et al use one sided statistical tests but should have used two sided tests. Exact P values or confidence intervals should have been given. Fleiss gives the procedure for confidence intervals in matched pairs.3
They have carried out multiple statistical tests without making a correction. To maintain the overall probability of obtaining a type I error at 0.05, a Bonferroni or other correction is usually made to the significance level of the test.
Henkel et …