Dear Editor,

Patient selection can affect required sample sizes and outcomes in randomised clinical trials and the optimal selection may depend on how the tested drug differentially affects patients. Bruynesteyn K et al. [1,2] have investigated these effects in rheumatoid arthritis based on an outcome measure of radiological damage progression.

An absolute reduction (AR) drug which reduces the expected number of new erosions per year by, say, 2 in all patients irrespective of their initial erosive risk is conceivable. Similarly, a relative reduction (RR) drug which reduces the expected number of new erosions per year by, say, 20% in all patients is also conceivable, thus by, say, 1 new erosion per year in medium risk patients and by, say, 3 new erosions in high risk patients.

However, the authors’ absolute and relative risk reduction drug models (ARR, RRR, respectively) differ from those described above in that they reduce the risk of the number of new erosions exceeding a pre- determined limit, smallest detectable change, by absolute or relative amounts [1]. Although valid as means of quantifying the effect of drugs [3,4], ARR and RRR are not appropriate statistical models of drug action. In fact, the authors’ ARR model represents a drug which reduces the number of new erosions in ‘high’ and ‘low’ risk patients by similar absolute amounts but is 50% less effective in patients at ‘medium’ risk of disease progression. It is this inappropriate drug model which gives rise to the statement (1): “From a statistical point of view, it would be wise to avoid groups of patients with an average prior risk (of disease progression) …. such patient groups provide less statistical power, as compared to trials with patient groups with a low or a high prior risk” rather than any valid reason to avoid enrolling average risk patients.

In a realistic absolute reduction (AR) drug model, the ‘medium’ risk group is the optimum choice (representing patients with progression critically near the pre-determined on/off event limit, Table 1.2 in [1]). If the drug effect is averaged across all patients, results for an AR drug are unaffected by patient selection. Thus the article inadvertently highlights the caution required if on/off events, as opposed to group averaged results, are used as outcome measures in trials.

Within the limits of such an evaluation, re-examination of the given trial results [1] using corrected AR, RR models, indicate that: methotrexate in combination with cyclosporin A is a pure RR drug, reducing the occurrence of erosions in patients in proportion to their initial risk; whilst the combination therapy in the COBRA trials follows a mixed AR/RR model, reducing the occurrence of erosions more in high risk patients than in low risk patients but not as much as might be expected by their relative risk. Both studies confirm that patients at medium and high risk of radiological disease progression are the optimum choice in clinical trials where radiological damage is the outcome.

R References

(1). Bruynesteyn K, Wanders A, Landewé R, van der Heijde D. How the type of risk reduction influences required sample sizes in randomised clinical trials. Ann Rheum Dis 2004; 63:1368-1371.

(2). Bruynesteyn K, Landewé R, van der Linden S, van der Heijde D. Radiography as primary outcome in rheumatoid arthritis: acceptable sample sizes for trials with 3 months follow-up. Ann Rheum Dis 2004; Published Online First 22/03/2004. doi:10.1136/ard.2003.014043.

(3). Osiri M, Suarez-Almazor ME, Wells GA, Robinson V, Tugwell P. Number needed to treat (NNT): implication in rheumatology clinical practice. Ann Rheum Dis 2003; 62:316-321.

(4). Rothwell PM. Can overall results of clinical trials be applied to all patients?[see comment]. Lancet 1995; 345:1616-1619.