Previous history of back pain has been highlighted

as a prognostic Docetaxel indicator in other studies (Mallen et al., 2007), but this was not supported here, probably due to the very high proportion of the sample with prior back pain (87%). Although a wide range of prognostic indicators were included here, other factors have been identified elsewhere (e.g. Mallen et al., 2007 and Foster et al., 2010) and it would be useful to examine these. Replication in other samples, perhaps with recent onset back pain, would be useful, as the current sample was mixed, and contained many people with long duration of pain. A strength of this study is presentation of the contribution of prognostic factors to poor outcome through the use of adjusted PAF calculations. Whilst adjustment for confounding is considered essential for models of outcome prediction, adjustment of PAFs is rare. Table 3 demonstrates that proportions can change substantially following adjustment, and presentation of unadjusted proportions would considerably overestimate the contribution of several factors. Although there was loss to follow-up in the study, the

sample is representative of baseline responders. Attrition biases are unlikely to substantially influence the RRs reported here, as comparisons are Selleckchem Apoptosis Compound Library within the sample. However, as the proportions corresponding to each factor are based on associations and risk factor prevalence, these may be affected. In this analysis, 47% of the sample had high pain intensity at baseline, compared to 46% in the total baseline sample; loss to follow-up is therefore unlikely Bay 11-7085 to have affected the proportions reported. However, initial response to the survey was 65%, and it is likely that non-responders were different to baseline responders. The impact of this is difficult to assess due to lack of information, but it is likely

that the prevalence of prognostic indicators would be lower among non-responders. However, even a 10% change in the prevalence of the prognostic indicator would only make a difference in the proportion of poor outcome associated with pain intensity of around 4% (e.g. reducing high pain intensity prevalence from 47% to 37% would lead to an unadjusted proportion of 77% compared with 81% in Table 3), indicating that our results are likely to be broadly generalisable. Comparisons are also difficult to make with other samples due to the different measures used, lack of information about prevalence of prognostic indicators, and the inability to produce adjusted figures without the original data. As proportions differ according to the prevalence of exposure and strength of association, estimates of the potential contributions of prognostic indicators should be made for individual settings.