Skip navigation

Responder analysis and NNTs for osteoathritis


Clinical bottom line

Responder analysis is possible for osteoarthritis trials. Underlying distributions are U-shaped, and average results cannot be trusted. NNTs calculated at 12 weeks for active treatments against placebo are reasonable, and offer to discriminate between them.


RA Moore et al. Responder analysis for pain relief and numbers needed to treat in a meta-analysis of etoricoxib osteoarthritis trials: bridging a gap between clinical trials and clinical practice. Annals of the Rheumatic Diseases 2009 doi:10.1136/ard.2009.107805


The study use information on 3,554 patients, two thirds women, with an average age of 62 years. Six trials involved patients with osteoarthritis of knee or hip, and one of knee only. Osteoarthritis was established clinically and radiographically. Initial pain had to be a minimum of 40/100 mm at inclusion, plus at least 15 mm increase and worsening in investigator global assessment since baseline.

Treatments used in the study included:

The study calculated the number of patients in each treatment group in each trial achieving various Initiative on Methods, Measurement, and Pain Assessment in Clinical Trials (IMMPACT) thresholds of pain relief over baseline of at least 15% (minimal benefit), at least 30% (moderate), at least 50% (substantial), and at least 70%, which we defined as extensive improvement. All but one trial lasted 12 weeks.

For the purposes of determining response at various levels, Bandolier has calculated the following at 12 weeks:

Criteria in defining responders included:


The proportion of patients with various levels of pain relief is shown in Figure 1. For each treatment, the proportion achieving the outcome fell as the degree of benefit increased.

Figure 1: Levels of response at 12 weeks with different active treatments. Light and dark red indicate at least moderate and at least substantial response respectively

Figure 2 shows that for the 2,676 patients with an active treatment for 12 weeks, there was a U-shaped response curve. Almost 60% of patients had a moderate or substantial decrease in pain intensity, 10% had minimal relief, while 35% had no meaning pain relief, were worse, or could not take the tablets.

Figure 2: For all active treatments at 12 weeks of treatment, there was a U-shaped response curve

The responder analysis could be used to calculate NNTs for active treatments compared with (their own) placebo. Figure 3 shows that NNTs were low (good) at between 4 and 5 for etoricoxib, celecoxib, and naproxen, but that ibuprofen 2400 mg had a much higher (worse) NNT of about 12.

Figure 3: NNTs compared for at least moderate pain relief at 12 weeks for active treatments compared with placebo.


What this shows is that responder analysis provides important insights into results from clinical trials. Differences between active and placebo on average may look small, and are reported here as between 10 mm and 15 mm on a 100 mm scale. But this overlooks the fact that some 60% of the patients getting active treatment get a useful level of pain relief, and 40% do very well; 30% have major reductions in pain.

With these responder analyses we can also calculate NNTs, which offers the possibility of meaningful comparison between treatments as more responder analyses are published.

And finally, we must learn to be cautious about using average results when no-one is average. It may well be that meta-analyses based on pooling average results from U-shaped curves are misleading, and have to be eliminated.