Successfully reported this slideshow.
We use your LinkedIn profile and activity data to personalize ads and to show you more relevant ads. You can change your ad preferences anytime.
A Novel Phase 3 Design IncorporatingHistorical Information for the Development ofAntibacterial AgentsJeff Wetherington, GS...
Complicated Urinary TractInfections (1)• Occurs in men and women– structural or functional abnormalities of the urinarytra...
Complicated Urinary TractInfections (2)Yet unmet treatment needs for patients with cUTIcontinue to exist given the emergen...
Traditional Phase 3• Randomized, parallel group• Active controlled, non inferiority• Narrow non inferiority margins• Infea...
Historical StudiesBayesian Augmented Control for Antibacterial Agents 50.00.20.40.60.81.0Naber 2009 Redman 2010Doripenem M...
Fixed Design• “Standard” 1:1 trial requires 750 subjects total– 90% power and one sided α=0.025– power for p=0.83 with non...
Goals• Reduce sample size!• Reduce sample size!• Maintain– controlled type I error (most complex…won’t beable to get “comp...
Preview• Proposed design incorporates historicalborrowing on the control arm using ahierarchical model.• 20% reduction in ...
Experimental parameters• Dichotomous endpoint, p=Pr(ME)• Control = Doripenum versus Treatment• 20% dropout rate• Goal is 9...
Notation and Model• γ0 = logit(true current control rate)• γ1 = logit(true Naber rate)• γ2 = logit(true Peninsula rate)• γ...
Intuition• The three Doripenem arms are connectedthrough the “across studies” distributionN(µ,τ).– similar to random effec...
Intuition• τ has a prior, and is estimated as part of themodel fitting.• Datasets with high across study variation– produc...
Intuition• Only 3 Doripenem arms available– so only three γ used to estimate τ• Enough that borrowing is dynamic, but prio...
Proposed Design A• N=600 with 2:1 randomization andborrowing– 20% savings on N compared to fixed design– 400 subjects on t...
Bayesian Augmented Control for Antibacterial Agents 15Data
Bayesian Augmented Control for Antibacterial Agents 16horizontal dashedline is historical rate
Bayesian Augmented Control for Antibacterial Agents 17control arm CIsunder none,fulland hierarchicalborrowing
Bayesian Augmented Control for Antibacterial Agents 18Treatment CI(same for all borrowingas prior is noninformative)
Bayesian Augmented Control for Antibacterial Agents 19EffectiveborrowingN=mean(p)*(1-mean(p)) / var(p)numborrowed=N-233
Bayesian Augmented Control for Antibacterial Agents 20probability of trialsuccess under eachkind of borrowing
Dynamic borrowing• The effective amount of borrowing for193/233 on control (almost identical tohistory) is 225 of the 530 ...
Bayesian Augmented Control for Antibacterial Agents 22
Bayesian Augmented Control for Antibacterial Agents 23
Bayesian Augmented Control for Antibacterial Agents 24Observed control proportion (x) versuseffective number of borrowed o...
Summary• The hierarchical model produces dynamicborrowing (through estimation of τ)– Many historical subjects are borrowed...
Operating Characteristics• For all designs, evaluated Pr(trial success) for– control rates 0.780, 0.805, 0.830, 0.855, 0.8...
Operating CharacteristicsTrt -10% Trt -5% Trt equal Trt +5%Control 0.780 0.024 / 0.006 0.309 / 0.119 0.842 / 0.702 0.996 /...
Operating CharacteristicsTrt -10% Trt -5% Trt equal Trt +5%Control 0.780 0.024 / 0.006 0.309 / 0.119 0.842 / 0.702 0.996 /...
Operating CharacteristicsTrt -10% Trt -5% Trt equal Trt +5%Control 0.780 0.024 / 0.006 0.309 / 0.119 0.842 / 0.702 0.996 /...
Operating CharacteristicsTrt -10% Trt -5% Trt equal Trt +5%Control 0.780 0.024 / 0.006 0.309 / 0.119 0.842 / 0.702 0.996 /...
Summary for Design A• Pros– 20% reduction in sample size– more patients on treatment– increased or equivalent power for tr...
Design B• As with design A, hierarchical borrowing and2:1 randomization, with maximum N=600.• Incorporates futility stoppi...
Operating CharacteristicsTrt -10% Trt -5% Trt equal Trt +5%Control 0.780 0.006 / 0.005456.00.119 / 0.133572.50.702 / 0.692...
Operating CharacteristicsTrt -10% Trt -5% Trt equal Trt +5%Control 0.780 0.006 / 0.005456.00.119 / 0.133572.50.702 / 0.692...
Summary for Design B• Similar Pros/Cons relative to fixed design• Pros relative to design A– can save 60-140 subjects on a...
Statistical Summary• Hierarchical Modeling allows for competitivealternatives to fixed designs with significantsample size...
Conclusions• Hierarchical Modeling allows forcompetitive alternatives to fixed designs withsignificant reduction in cycle ...
Upcoming SlideShare
Loading in …5
×

Tessella webinar 6 18-2013 v3

394 views

Published on

Published in: Business, Technology
  • Be the first to comment

  • Be the first to like this

Tessella webinar 6 18-2013 v3

  1. 1. A Novel Phase 3 Design IncorporatingHistorical Information for the Development ofAntibacterial AgentsJeff Wetherington, GSKKert Viele, Berry ConsultantsTessella Series WebinarJune 18th, 2013
  2. 2. Complicated Urinary TractInfections (1)• Occurs in men and women– structural or functional abnormalities of the urinarytract– hospitalized patients with significant medical orsurgical co-morbidities• Major cause of hospital admission, extendedhospitalizations morbidity, mortality, and excesshealthcare costs• Prescribing physicians have several options forempiric and pathogen-specific treatmentBayesian Augmented Control for Antibacterial Agents 2
  3. 3. Complicated Urinary TractInfections (2)Yet unmet treatment needs for patients with cUTIcontinue to exist given the emergence andprevalence of multi-drug resistance in uropathogensBayesian Augmented Control for Antibacterial Agents 3
  4. 4. Traditional Phase 3• Randomized, parallel group• Active controlled, non inferiority• Narrow non inferiority margins• Infeasible in the face of increasing unmet need– >1500 patients enrolled into 2 independent trials– >5 year clinical development programs• Urgent need for novel clinical designs forantibiotic drug developmentBayesian Augmented Control for Antibacterial Agents 4
  5. 5. Historical StudiesBayesian Augmented Control for Antibacterial Agents 50.00.20.40.60.81.0Naber 2009 Redman 2010Doripenem Microbiological Eradication Rate
  6. 6. Fixed Design• “Standard” 1:1 trial requires 750 subjects total– 90% power and one sided α=0.025– power for p=0.83 with noninferiority δ=0.10– 20% dropout assumed– note 375 patients on treatment• Can we do better using the historicalinformation?Bayesian Augmented Control for Antibacterial Agents 6
  7. 7. Goals• Reduce sample size!• Reduce sample size!• Maintain– controlled type I error (most complex…won’t beable to get “complete” control)– comparable power around (and slightly below perexpectation) p=0.83– similar numbers of subjects on treatment (forsecondary analyses)Bayesian Augmented Control for Antibacterial Agents 7
  8. 8. Preview• Proposed design incorporates historicalborrowing on the control arm using ahierarchical model.• 20% reduction in sample size• Similar power near or slight below p=0.83• Slightly MORE subjects on treatment• Type I error control– in a region near 0.83– based on E[type I error] for a range of perceivedlikely amounts of “drift” in the true control rateBayesian Augmented Control for Antibacterial Agents 8
  9. 9. Experimental parameters• Dichotomous endpoint, p=Pr(ME)• Control = Doripenum versus Treatment• 20% dropout rate• Goal is 90% for p=0.83 (NI δ=0.10) withone sided α=0.025• If possible, would like to leverage twohistorical studies from control arm.– Naber, 230 successes in 280 subjects (82.1%)– Peninsula, 209 successes in 250 subjects (83.6%)Bayesian Augmented Control for Antibacterial Agents 9
  10. 10. Notation and Model• γ0 = logit(true current control rate)• γ1 = logit(true Naber rate)• γ2 = logit(true Peninsula rate)• γ0,γ1,γ2 ~ N(μ,τ)• π(μ)=N(1,1)• π(τ2)=IGamma(α=0.001,β=0.001)• Treatment effect θ ~ N(μθ=0,σθ=100)• γ0+θ = logit(true treatment rate)Bayesian Augmented Control for Antibacterial Agents 10
  11. 11. Intuition• The three Doripenem arms are connectedthrough the “across studies” distributionN(µ,τ).– similar to random effects model on studies– common model in meta-analysis• τ is the most important parameter (acrossstudy variance)– τ≈0 corresponds to γ0≈γ1≈γ2– τ large corresponds to no borrowingBayesian Augmented Control for Antibacterial Agents 11
  12. 12. Intuition• τ has a prior, and is estimated as part of themodel fitting.• Datasets with high across study variation– produce higher estimates of τ– the N(µ,τ) across distribution exerts less influence oneach group (acts as less informative prior)– less borrowing• Datasets with low across study variation– produce lower estimates of τ– the N(µ,τ) across study distribution can act as quiteinformative prior– more borrowingBayesian Augmented Control for Antibacterial Agents 12
  13. 13. Intuition• Only 3 Doripenem arms available– so only three γ used to estimate τ• Enough that borrowing is dynamic, but priorwill not fully wash out.• Important to consider operatingcharacteristics (as always)• Big goal– borrow robustly when current data near p=0.83– borrow less as current data diverges from p=0.83Bayesian Augmented Control for Antibacterial Agents 13
  14. 14. Proposed Design A• N=600 with 2:1 randomization andborrowing– 20% savings on N compared to fixed design– 400 subjects on treatment, more than fixeddesign• Trial declared success if– Pr(trt rate > ctrl rate – 10%)>0.975Bayesian Augmented Control for Antibacterial Agents 14
  15. 15. Bayesian Augmented Control for Antibacterial Agents 15Data
  16. 16. Bayesian Augmented Control for Antibacterial Agents 16horizontal dashedline is historical rate
  17. 17. Bayesian Augmented Control for Antibacterial Agents 17control arm CIsunder none,fulland hierarchicalborrowing
  18. 18. Bayesian Augmented Control for Antibacterial Agents 18Treatment CI(same for all borrowingas prior is noninformative)
  19. 19. Bayesian Augmented Control for Antibacterial Agents 19EffectiveborrowingN=mean(p)*(1-mean(p)) / var(p)numborrowed=N-233
  20. 20. Bayesian Augmented Control for Antibacterial Agents 20probability of trialsuccess under eachkind of borrowing
  21. 21. Dynamic borrowing• The effective amount of borrowing for193/233 on control (almost identical tohistory) is 225 of the 530 historical subjects• Hierarchical modeling produces dynamicborrowing.– as the current control varies away from 82.9%(history), the amount of borrowing decreasesBayesian Augmented Control for Antibacterial Agents 21
  22. 22. Bayesian Augmented Control for Antibacterial Agents 22
  23. 23. Bayesian Augmented Control for Antibacterial Agents 23
  24. 24. Bayesian Augmented Control for Antibacterial Agents 24Observed control proportion (x) versuseffective number of borrowed observations (y)
  25. 25. Summary• The hierarchical model produces dynamicborrowing (through estimation of τ)– Many historical subjects are borrowed when thecurrent data is consistent with history– Few historical subjects are borrowing when thecurrent data is inconsistent• Most effective (gains the most information) ifthe historical data is “on point”.• If you happen to get inconsistent data, lessborrowing occurs mitigating the costs.Bayesian Augmented Control for Antibacterial Agents 25
  26. 26. Operating Characteristics• For all designs, evaluated Pr(trial success) for– control rates 0.780, 0.805, 0.830, 0.855, 0.880– treatment rates 10% lower, 5% lower, equal, or5% greater than control• Treatment rates 10% lower were used tocompute type I error• Equal treatment rates used to computepower.Bayesian Augmented Control for Antibacterial Agents 26
  27. 27. Operating CharacteristicsTrt -10% Trt -5% Trt equal Trt +5%Control 0.780 0.024 / 0.006 0.309 / 0.119 0.842 / 0.702 0.996 / 0.990Control 0.805 0.026 / 0.007 0.321 / 0.229 0.874 / 0.861 0.996 / 0.997Control 0.830 0.026 / 0.017 0.342 / 0.376 0.910 / 0.942 0.998 / 0.999Control 0.855 0.028 / 0.045 0.372 / 0.564 0.929 / 0.966 1.000 / 1.000Control 0.880 0.030 / 0.100 0.421 / 0.642 0.954 / 0.976 1.000 / 1.000Bayesian Augmented Control for Antibacterial Agents 27Table shows Pr(trial success)Two entries per cell in the form FIXED / DESIGN ADesign aims to produce greater power AND lower typeI error for consistent control data, with 20% savings on N.Here type I error reduced from 0.026 to 0.017 ANDpower increased from 91% to 94.2%.Type I error Power
  28. 28. Operating CharacteristicsTrt -10% Trt -5% Trt equal Trt +5%Control 0.780 0.024 / 0.006 0.309 / 0.119 0.842 / 0.702 0.996 / 0.990Control 0.805 0.026 / 0.007 0.321 / 0.229 0.874 / 0.861 0.996 / 0.997Control 0.830 0.026 / 0.017 0.342 / 0.376 0.910 / 0.942 0.998 / 0.999Control 0.855 0.028 / 0.045 0.372 / 0.564 0.929 / 0.966 1.000 / 1.000Control 0.880 0.030 / 0.100 0.421 / 0.642 0.954 / 0.976 1.000 / 1.000Bayesian Augmented Control for Antibacterial Agents 28Table shows Pr(trial success)Two entries per cell in the form FIXED / DESIGN AWith slight reduction in true control rate, design stillobtains nearly equivalent power.Type I error Power
  29. 29. Operating CharacteristicsTrt -10% Trt -5% Trt equal Trt +5%Control 0.780 0.024 / 0.006 0.309 / 0.119 0.842 / 0.702 0.996 / 0.990Control 0.805 0.026 / 0.007 0.321 / 0.229 0.874 / 0.861 0.996 / 0.997Control 0.830 0.026 / 0.017 0.342 / 0.376 0.910 / 0.942 0.998 / 0.999Control 0.855 0.028 / 0.045 0.372 / 0.564 0.929 / 0.966 1.000 / 1.000Control 0.880 0.030 / 0.100 0.421 / 0.642 0.954 / 0.976 1.000 / 1.000Bayesian Augmented Control for Antibacterial Agents 29Table shows Pr(trial success)Two entries per cell in the form FIXED / DESIGN AIf true current control rate is higher than observedhistorical rate, power is increased, but we also observeinflated type I error.Type I error Power
  30. 30. Operating CharacteristicsTrt -10% Trt -5% Trt equal Trt +5%Control 0.780 0.024 / 0.006 0.309 / 0.119 0.842 / 0.702 0.996 / 0.990Control 0.805 0.026 / 0.007 0.321 / 0.229 0.874 / 0.861 0.996 / 0.997Control 0.830 0.026 / 0.017 0.342 / 0.376 0.910 / 0.942 0.998 / 0.999Control 0.855 0.028 / 0.045 0.372 / 0.564 0.929 / 0.966 1.000 / 1.000Control 0.880 0.030 / 0.100 0.421 / 0.642 0.954 / 0.976 1.000 / 1.000Bayesian Augmented Control for Antibacterial Agents 30Table shows Pr(trial success)Two entries per cell in the form FIXED / DESIGN AIf true current control rate is much lower than the observedhistorical rate, then power is reduced (type I error rateis negligible).Type I error Power
  31. 31. Summary for Design A• Pros– 20% reduction in sample size– more patients on treatment– increased or equivalent power for true control ratesnear observed historical data.– essentially, there is a “sweet spot” where Design Adominates the fixed trial.• Cons– inflated type I error for true control rates muchabove observed historical control rates.– decreased power for true control rates much belowobserved historical dataBayesian Augmented Control for Antibacterial Agents 31
  32. 32. Design B• As with design A, hierarchical borrowing and2:1 randomization, with maximum N=600.• Incorporates futility stopping– interim analyses N=300, 400, 500, 600.– trial stopped for futility ifPr(non-inferiority)<0.15• Evaluated operating characteristics as before,including expected sample size.Bayesian Augmented Control for Antibacterial Agents 32
  33. 33. Operating CharacteristicsTrt -10% Trt -5% Trt equal Trt +5%Control 0.780 0.006 / 0.005456.00.119 / 0.133572.50.702 / 0.692598.50.990 / 0.994599.7Control 0.805 0.007 / 0.010499.70.229 / 0.237589.90.861 / 0.851599.70.997 / 0.998599.7Control 0.830 0.017 / 0.019537.90.376 / 0.376595.70.942 / 0.938599.70.999 / 0.999599.7Control 0.855 0.045 / 0.048569.30.564 / 0.567598.20.966 / 0.968599.71.000 / 0.999599.7Control 0.880 0.100 / 0.096580.20.642 / 0.638598.30.976 / 0.970599.71.000 / 0.999599.7Bayesian Augmented Control for Antibacterial Agents 33Three entries per cell. Top two are Pr(trial success)in the form DESIGN A / DESIGN B. Lower numberis expected N (max 600)Can save another60-140 subjectsunder null hypothesisdepending on truecontrol rate.
  34. 34. Operating CharacteristicsTrt -10% Trt -5% Trt equal Trt +5%Control 0.780 0.006 / 0.005456.00.119 / 0.133572.50.702 / 0.692598.50.990 / 0.994599.7Control 0.805 0.007 / 0.010499.70.229 / 0.237589.90.861 / 0.851599.70.997 / 0.998599.7Control 0.830 0.017 / 0.019537.90.376 / 0.376595.70.942 / 0.938599.70.999 / 0.999599.7Control 0.855 0.045 / 0.048569.30.564 / 0.567598.20.966 / 0.968599.71.000 / 0.999599.7Control 0.880 0.100 / 0.096580.20.642 / 0.638598.30.976 / 0.970599.71.000 / 0.999599.7Bayesian Augmented Control for Antibacterial Agents 34Three entries per cell. Top two are Pr(trial success)in the form DESIGN A / DESIGN B. Lower numberis expected N (max 600)Very slightchanges topower
  35. 35. Summary for Design B• Similar Pros/Cons relative to fixed design• Pros relative to design A– can save 60-140 subjects on average when nullhypothesis is true– when drug is noninferior, design very rarely stopsfor futility.– successful trials retain 400 subjects on treatment.• Cons relative to design A– very slight power loss (simulations all have 1% orless power loss)Bayesian Augmented Control for Antibacterial Agents 35
  36. 36. Statistical Summary• Hierarchical Modeling allows for competitivealternatives to fixed designs with significantsample size savings• Possible risks are associated with drift in thetrue control rate from observed historicalrate.• Futility stopping can result in additionalsample size savings under null hypothesiswithout significant cost to power.Bayesian Augmented Control for Antibacterial Agents 36
  37. 37. Conclusions• Hierarchical Modeling allows forcompetitive alternatives to fixed designs withsignificant reduction in cycle time• Possible risks are associated with drift in thetrue control rate from observed historical rate• Futility stopping can result in minimizingexposure to ineffective therapy under nullhypothesis without significant cost to powerBayesian Augmented Control for Antibacterial Agents 37

×