Numbers-needed-to-treat analyses--do timing, dropouts, and outcome matter? Pooled analysis of two randomized, placebo-controlled chronic low back pain trials

Abstract

Numbers-needed-to-treat (NNT) are useful for presenting treatment response, conveying the clinical relevance of results. NNTs are typically calculated at a landmark endpoint (end of trial), but often using the last observation carried forward (LOCF), which ignores patient discontinuations. We compared NNTs in chronic low back pain (CLBP) using three separate imputation methods, using data from two identical 12-week trials comparing etoricoxib 60 mg (N=210), 90 mg (N=212), and placebo (N=217). We calculated the number of patients with improvements in pain intensity from baseline of ≥15%, ≥30%, ≥50%, and ≥70% at 2, 4, 8, and 12 weeks of treatment. For longitudinal response over time, patient discontinuations were assigned a 0% improvement from dropout forward. Landmark response at week 12 was assessed using LOCF and completer approaches, using only observed (non-missing) data. The longitudinal approach was most conservative; after 12 weeks 65% of patients taking etoricoxib had ≥15% improvement, 60% had ≥30% improvement, 45% had ≥50%, improvement, and 30% had ≥70% improvement, with placebo rates approximately 55%, 45%, 30%, and 15%, respectively. Response rates were higher with landmark analyses. Landmark NNTs at week 12 were generally similar or slightly lower (better) than those from a longitudinal approach, but results were inconsistent. Landmark analyses provide no information on response variability, as is obtained with longitudinal analysis. Outcome, imputation method, and reporting method are intimately connected and need to be considered alongside trial quality and validity to make sensible comparisons between treatments.