Abstract Background Selective non-participation at baseline (due to non-response and non-consent) and loss to follow-up are important concerns for longitudinal observational research. We investigated these matters in the context of baseline recruitment and retention at 18 months of participants for a prospective observational cohort study of knee pain and knee osteoarthritis in the general population. Methods Participants were recruited to the Knee Clinical Assessment Study – CAS(K) – by a multi-stage process involving response to two postal questionnaires, consent to further contact and medical record review (optional), and attendance at a research clinic. Follow-up at 18-months was by postal questionnaire. The characteristics of responders/consenters were described for each stage in the recruitment process to identify patterns of selective non-participation and loss to follow-up. The external validity of findings from the clinic attenders was tested by comparing the distribution of WOMAC scores and the association between physical function and obesity with the same parameters measured directly in the target population as whole. Results 3106 adults aged 50 years and over reporting knee pain in the previous 12 months were identified from the first baseline questionnaire. Of these, 819 consented to further contact, responded to the second questionnaire, and attended the research clinics. 776 were successfully followed up at 18 months. There was evidence of selective non-participation during recruitment (aged 80 years and over, lower socioeconomic group, currently in employment, experiencing anxiety or depression, brief episode of knee pain within the previous year). This did not cause significant bias in either the distribution of WOMAC scores or the association between physical function and obesity. Conclusion Despite recruiting a minority of the target population to the research clinics and some evidence of selective non-participation, this appears not to have resulted in significant bias of cross-sectional estimates. The main effect of non-participation in the current cohort is likely to be a loss of precision in stratum-specific estimates e.g. in those aged 80 years and over. The subgroup of individuals who attended the research clinics and who make up the CAS(K) cohort can be used to accurately estimate parameters in the reference population as a whole. The potential for selection bias, however, remains an important consideration in each subsequent analysis.