Article type
Year
Abstract
Background: It is unclear whether data on continuous outcomes, such as pain intensity, should be summarised as differences in means, fully respecting their continuous nature, or used for dichotomous categorisation of patients into treatment response and treatment failure. The second approach allows direct calculation of clinically meaningful measures such as numbers needed to treat or harm. Many trials do not report treatment response, however, and it is unclear whether methods to approximate odds ratios (OR) of treatment response from continuous data are accurate.
Objective: To determine the agreement between OR of treatment response reported in published papers and OR approximated from continuous data.
Methods: Meta-epidemiological study of meta-analyses of large-scale osteoarthritis trials comparing active treatment with placebo, sham, or non-intervention. Trials had to include at least 100 patients per group, use pain or global symptoms as an outcome and report both, dichotomised treatment response and mean effects for the outcome. For each trial, we calculated log OR of treatment response from the odds of response in experimental and control groups based on reported response criteria. Then, we calculated standardised mean differences (SMDs) and used a previously described approach towards approximating log OR from SMDs (Chinn S. Stat.Med. 2000; 19:3127). We pooled differences in directly calculated log OR and approximated log OR using standard random-effects models that accounted for variability between trials and meta-analyses.
Results: We included 21 large-scale trials (9802 patients) from 5 meta-analyses. The pooled overall ratio of OR was 0.95 (95%-CI 0.81 to 1.11, see Figure1). There was little evidence for heterogeneity between trials (I2=0%) or meta-analyses (I2=0%).
Conclusions: ORs derived from SMDs are accurate approximations of directly calculated ORs of treatment response if dichotomized data on treatment response are unavailable and could also be used to calculate numbers needed to treat or harm.
Objective: To determine the agreement between OR of treatment response reported in published papers and OR approximated from continuous data.
Methods: Meta-epidemiological study of meta-analyses of large-scale osteoarthritis trials comparing active treatment with placebo, sham, or non-intervention. Trials had to include at least 100 patients per group, use pain or global symptoms as an outcome and report both, dichotomised treatment response and mean effects for the outcome. For each trial, we calculated log OR of treatment response from the odds of response in experimental and control groups based on reported response criteria. Then, we calculated standardised mean differences (SMDs) and used a previously described approach towards approximating log OR from SMDs (Chinn S. Stat.Med. 2000; 19:3127). We pooled differences in directly calculated log OR and approximated log OR using standard random-effects models that accounted for variability between trials and meta-analyses.
Results: We included 21 large-scale trials (9802 patients) from 5 meta-analyses. The pooled overall ratio of OR was 0.95 (95%-CI 0.81 to 1.11, see Figure1). There was little evidence for heterogeneity between trials (I2=0%) or meta-analyses (I2=0%).
Conclusions: ORs derived from SMDs are accurate approximations of directly calculated ORs of treatment response if dichotomized data on treatment response are unavailable and could also be used to calculate numbers needed to treat or harm.
Images
