Objective To construct a classification of patients with shoulder complaints based on their physical examination. To investigate (1) the interobserver reliability, (2) to what extent the setting in which the patients were recruited, and demographic and clinical characteristics are related to the classification. Study design and setting Data from 132 patients with shoulder complaints recruited in various health care settings in The Netherlands were examined. Two observers independently performed a physical examination of the cervical spine and shoulder joint. A nonmetric multidimensional scaling procedure was performed for each observer separately. The interobserver reliability of both observers was computed. Differences between setting, demographic and clinical characteristics, and the resulting dimensions were investigated. Results For both observers two dimensions (severity of complaints of the shoulder joint, and severity of problems of the cervical spine) were sufficient to classify all patients. Agreement between the two observers was good (r=0.84) to moderate (r=0.69). Patients with neck pain in history taking showed higher scores on both dimensions. Conclusion Despite moderate interobserver agreement for each variable from physical examination found in previous studies, observers agree on the scores of the patients on the relevant dimensions. Given the limited number of effective treatments available to the general practitioner, a more sophisticated classification system seems unnecessary.