BookmarkSubscribeRSS Feed
stan_koprowski

Community Manager

Joined:

Aug 8, 2012

Choose Language Hide Translation Bar

Equivalence Approach in Design of Experiments for Robustness Evaluation: Applications in Vaccines Development With JMP(R) ( 2019-EU-30MP-102 )

Level: Intermediate
Job Function: Analyst / Scientist / Engineer
Industry: Health Care
Bernard Francq, Principal Statistician, GlaxoSmithKline
Dan Lin, Expert Biostatistician, GlaxoSmithKline
Waldemar Miller, Statistician, GlaxoSmithKline
Réjane Rousseau, Senior Statistician, GlaxoSmithKline
Sylvie Scolas, Statistician, GlaxoSmithKline
Walter Hoyer, Senior Manager, Innovation and Automation, CMC Stats, GlaxoSmithKline

Current state-of-the-art vaccines development is based on the Quality by Design paradigm. In design of experiments (DOE), the design space is defined as a subspace of process parameter combinations “that have been demonstrated to provide assurance of quality.” (ICH guidance Q8) The robustness of a process is its property to stay within the specification limits (target ± Δ) after a change in experimental conditions. A “DOE for flatness” extends the classical equivalence test (two one-sided test) to the multidimensional case (continuous or categorical factors). We’ll discuss the use of the multi-t distribution, as the entire experimental domain is compared to a reference level (usually the center point). The design space is then the subspace where the predicted means are equivalent to the reference level (confidence intervals of mean contrasts lie within ± Δ). Our methodology will be illustrated with applications in CMC statistics and vaccines development. We will discuss the limitations of the multiple comparisons and the Simulator in Prediction Profiler in JMP. We will show how to modify the default predicted values and the Contour Profiler to obtain an interactive design space to better communicate with the scientists. A JSL script will be shown to automate this procedure.

Comments
frankderuyck

Hi Bernard, excellent presentation! Maybe one comment: as the model you specified around the reference point has non-significant coefficients can it not be considered as an intercept model?

BeFrancq

Hi Frank, Thanks! Good point! If all parameters are non-significant (process very robust), then the model moves closer to an intercept only model. The contrasts converge to 0 and their CIs collapse, as illustrated at the end of my talk (see the animated heatmap). A practical difference (equivalence test) is, anyway, considered in our context, so a process can be assessed as robust whatever the significance of the parameters. Reducing a model should be done with caution, and my recommendation is to keep the model for which the DoE was designed (as this is not a screening phase).