An Alternative to Data Imputation in Analgesic Clinical Trials David Petullo, Thomas Permutt, Feng Li Division of Biometrics II, Office of Biostatistics Office of Translational Sciences, Center for Drug Evaluation and Research U.S. Food & Drug Administration
2 Chronic Pain Trial - Not Missing Data If patient discontinues study drug, the pain score at planned endpoint becomes irrelevant Completers with bad outcomes All considered equally bad outcomes You have an outcome just not a numerical score
3 Continuous Responder Curve
4 Continuous Responder Curve
5 Continuous Responder Curve
6 Why Half? Doesn’t have to be half Can be adaptive – Trim only dropouts for the group with more dropouts – Dropouts plus others for the other group to make fractions equal Why isn’t this completers analysis?
7 Trim Same Fraction If equal fractions – It is difference in completer means If active has more completers – Bonus: compare your best completers to all placebo completers – Because more completers is good drug effect If placebo has more completers – Test drug not “penalized” or disqualified – But you picked your best – So it’s fair to compare to placebo best – This is an insoluble problem with MAR methods
8 Permutation Test Calculate difference in trimmed means Rerandomize and recalculate, many times Use rerandomization distribution as reference See how far real value is in tail of reference distribution Exact, randomization-based test – That’s easy, even completers analysis can do that – But based on interpretable statistic Especially in the case of more dropouts in the active group
9 Case Study 1
10 Lyrica- Efficacy Indication: management of pain associated with spinal cord injury (SCI) Two randomized, multi-center, placebo-controlled, double-blind trials – Study 1107: 4-week dose escalation, 12-week fixed, 1-week taper – Study 125: 3-week dose escalation, 9-week fixed
11 Study Efficacy Primary efficacy parameter was duration adjusted average change (DAAC) – A significant result on DAAC by itself would not support efficacy – Must use a conservative imputation strategy
12 Disposition
13 Results
14 Difference in Better Half Mean change from baseline pain at Week 16
15 Permutation Test
16 Case 1: Summary Discontinuations – 15% placebo, 17% active BOCF and mBOCF – treatment effect Better Half estimand – treatment effect
17 Case Study 2
18 Drug X ( approved 2010) Drug X Management of pain severe enough to require daily, around the clock, long term, opioid treatment Two studies provided evidence of efficacy Study 1: DB, R, PC, AC study in subjects with moderate to severe chronic low back pain Study 2: DB, R, PC, AC study in subjects with OA of the knee
19 Study 2: Efficacy Change from baseline to the end of the maintenance period in the average pain Pain measured using 11-point NRS scale during the past 12- hours Analysis ANCOVA with treatment, baseline pain score and pooled center Analysis population was all randomized subjects that received at least 1 dose of study drug Missing data imputed using LOCF
20 Disposition Active
21 Results Active
22 Sensitivity Analyses Active
23 Difference in Better Half Average change from baseline pain at Week 12
24 Permutation Test
25 Case 2: Summary Discontinuations – 39% placebo, 43% active LOCF –treatment effect BOCF and mBOCF – no treatment effect Better Half – no treatment effect
26 Conclusion Complete observations – just not numerical Exact test for hypothesis of no drug effect All randomized patients Accounts for – Excess placebo dropouts for lack of efficacy (bonus) – Excess active dropouts for toxicity Guaranteed fair comparison in an adherent subset