Biologics for rheumatoid arthritis: an overview of Cochrane reviews

The six Cochrane reviews used a variety of outcome measures to compare the drugs with a placebo, but the overview focused on two which were used in all the individual reviews and in most of the randomised trials within this. For benefit, the overview used the ACR50. This is a composite measure that counts the number of patients who achieve a 50% improvement in the number of painful, swollen joints and who experience a 50% improvement in at least 3 out of 5 other assessments: these are

• Patient global assessment,
• Physician global assessments,
• Pain scores,
• Health Assessment Questionnaire (HAQ) disability score,
• Acute phase reactants (Erythrocyte Sedimentation Rate (ESR) or C-Reactive Protein (CRP)

For harm, it used the proportion of people who withdrew from the randomised trials because of adverse effects.

Five of the six biologic DMARDs were statistically significantly more beneficial than placebo, whether used in conjunction with traditional DMARDs such as methotrexate or not. Less than 10% dropped out due to adverse events in most of the trials from the individual reviews, but etanercept, abatacept and infliximab seemed to be responsible for the fewest withdrawals, with little difference between the numbers dropping out in the treatment and the placebo groups in the trials of those drugs.

The overview allowed a common pair of analyses (ACR and withdrawal due to side effects) for each drug by using the results of the trials in each of the six Cochrane reviews and also an indirect comparison of the different drugs. This used a single treatment group (the recommended standard dose of the relevant biologic drug) and a single comparator group (the placebo) from each of the trials that contained relevant outcome data. The results were analysed using a Restricted Maximum Likelihood (REML) linear mixed model in SAS, adjusting for important patient characteristics. The hierarchical model was used as it incorporates the most important study-level characteristic (i.e. the biologic drug) using another empirical Bayes statistical approach. The linear mixed (REML) model applied, with drug as a fixed factor and study as a random factor, in order to reduce the between-study variance. The observed 'Control Event Rate' (i.e. the result for the placebo group) and the trial duration were expected to be important effect modifiers and were included in the model as a (fixed-effect) interaction term.

In order to compare the benefits (ACR50) and harms (withdrawal due to adverse effects) of each drug, a hierarchical model was used. Although handled via one big model, the summary statistics were subsequently converted into the number needed to treat to benefit (NNTB) and number needed to treat to harm (NNTH) for clinical interpretation. This was done with Visual Rx, after pooling the placebo group data across all trials to obtain the assumed result for patients not receiving biologics. This allowed the calculation of a separate result for the effects of each drug compared to placebo and, also, the indirect comparisons between pairs of drugs that are not currently available from direct, head-to-head randomised trials.

Robin Christensen, from the Parker Institute: Musculoskeletal Statistics Unit, at Frederiksberg Hospital, in Denmark, who was one of the statisticians on the overview explains more in the podcast.