Analysing the SF-36 in population-based research : a comparison of methods of statistical approaches using chronic pain as an example.

Torrance, Nicola and Smith, Blair H. and Lee, Amanda J. and Aucott, Lorna and Cardy, Amanada and Bennett, Michael I. (2009) Analysing the SF-36 in population-based research : a comparison of methods of statistical approaches using chronic pain as an example. Journal of Evaluation in Clinical Practice, 15 (2). pp. 328-334. ISSN 1356-1294

Abstract

Background The Medical Outcomes Study 36 Item Short-Form (SF-36) questionnaire is one of the most widely used measures of health related quality of life in medical research, including studies on pain-related conditions. Although scores in each of its eight domains rarely conform to a normal distribution, it is most widely analysed using simple parametric statistical techniques. Some have suggested a need for more complex or non-parametric analytical approaches, and this quandary faces researchers recurrently when using the SF-36. In this study of chronic pain, we compared results arising from the SF-36 between three study sub-samples, using conventional parametric, non-parametric, bootstrapping and log transforming methods. Methods Respondents to a postal survey conducted in Aberdeen, Leeds and London (n = 3002, response rate 52%) were categorized in three groups according to previously validated questionnaires: those with chronic pain of predominantly neuropathic origin (POPNO, n = 241), those with chronic pain (non-POPNO, n = 1179), and those with no chronic pain (n = 1537). SF-36 scores were compared between these groups, using: ANOVA and t-tests; Kruskall–Wallis and Mann–Whitney U-tests; bootstrapping methods; and log transformation with ANOVA. Results There were highly significant differences between the three groups, with lower scores in all SF-36 domains found those with chronic pain (P < 0.001). Those with chronic POPNO had lower scores in all domains than those with chronic pain (non-POPNO) (P < 0.001). These results were the same after applying each statistical method Conclusions In this study, the choice of statistical approach had no influence on the results. We conclude that the conventional approach, using straightforward parametric tests, is both simplest and the best for allowing comparison with other studies. We are likely to adopt this in future studies.