OBJECTIVES--There are no agreed criteria for osteoarthritis (OA) of the knee in population studies. The radiographic scoring system of Kellgren and Lawrence has been the system most used in the past and although other methods have been developed, comparisons have not been performed. Therefore these grading systems were compared in radiographs from a general population sample. METHODS--Anteroposterior weightbearing radiographs of 1954 knees from 977 women aged 45-64 years from the Chingford population study were read by a variety of methods, including quantitative measures of minimum joint space, qualitative measures of osteophytes and of joint space, and a qualitative Kellgren and Lawrence global score. All qualitative methods used standardised atlases. Intra-observer and interobserver reproducibility was tested on a subgroup of 100 films using three observers and two readings. Variables were dichotomised at the tenth and second centiles to define OA. Odds ratios were calculated for each method for the association of OA with knee pain, obesity, and with each of the other methods. RESULTS--Most methods had high intraobserver and interobserver reproducibility, except for measurements of lateral joint space. The best predictors of knee pain were the presence of osteophytes and the Kellgren and Lawrence grade. Methods measuring narrowing performed less well, with measurements of lateral joint space being particularly poor. Similar results were achieved in the comparison with obesity and in the comparisons between methods. CONCLUSIONS--These data suggest that the presence or absence of a definite osteophyte read by a single observer with an atlas is the best method of defining OA of the knee for epidemiological studies in women. Assessment of narrowing may be better used in evaluating severity.