Are we overestimating harm when high numbers of placebo patients are switched to active therapy as part of an adaptive trial design? A meta-epidemiological study

Tags: Oral
Tarp S1, Dossing A1, Furst DE2, Boers M3, Luta G4, Gluud C5, Bliddal H1, Christensen R1
1Musculoskeletal Statistics Unit, The Parker Institute, Department of Rheumatology, Copenhagen University Hospitals, Bispebjerg and Frederiksberg, Denmark, 2University of California, Los Angeles, David Geffen School of Medicine, Los Angeles, USA, 3Department of Epidemiology and Biostatistics, and Amsterdam Rheumatology and Immunology Center at VU University Medical Center, Amsterdam, The Netherlands, 4Department of Biostatistics, Bioinformatics, and Biomathematics, Georgetown University Medical Center, Washington, DC, USA, 5Copenhagen Trial Unit, Centre for Clinical Intervention Research, Rigshospitalet, Copenhagen University Hospital, Copenhagen, Denmark

Background: Placebo-controlled trials in rheumatoid arthritis (RA) increasingly employ adaptive trial designs that offer early rescue therapy for participants with a poor response at an interim time point. This may compromise inferences drawn from an ITT analysis.

Objective: To examine if adaptive trial design confounds inference from ITT analysis and to explore the quantified impact of adjusting for skewed dropout between trial groups in the evaluation of harm outcomes.

Methods: A systematic search identified RA trials comparing a biologic or tofacitinib vs placebo in approved dose. The main outcome was the ratio of odds ratios (ROR) quantifying the degree of bias associated with adaptive vs non-adaptive trials. An ROR > 1 implies that adaptive trials exaggerate serious adverse event (SAE) estimates. Subsequently, two meta-analyses of trials were performed based on participant number with SAEs (numerator): i) based on the ITT population (denominator) expressed as odds ratio (OR); ii) based on the total person-years of exposure (denominator) expressed as rate ratio (RR). If not reported, exposure was estimated by assuming a linear dropout rate from baseline to follow-up. To compare the pooled OR and RR, the ratio of ratios were calculated as Exp(ln[OR]–ln[RR]), assuming acceptable comparability due to low event rate. A ratio of OR to RR > 1 implies that OR exaggerates harm estimates.

Results: All 10 drugs except infliximab and anakinra were represented in both the 69 non-adaptive and the 31 adaptive trials included in the analysis (32,101 patients, approximately 18,902 person-years). Fig. 1 illustrates pooled OR for adaptive and non-adaptive trials. The estimate was exaggerated in adaptive trials (ROR 1.29, 1.06 to 1.57). The ratio between the OR and the RR found that OR exaggerates harm estimates (ratio 1.11, 1.08 to 1.14).

Conclusion: This study suggests that evaluation of harm outcomes (e.g. SAEs) is affected by adaptive trial designs per se. Applying ITT analysis for odds or risk statistics might overestimate the true effect when analysing harm data from trials with skewed dropout rates. We suggest that an exposure-time adjusted model should be applied.