Statistical Analyses and Modeling
BSSI has a strong team with diverse statistical backgrounds and extensive research experience in genomics and clinical analysis. For each project, we carefully select statistical analysis methodologies to ensure that proper statistical approaches are used for different types of data. Parametric statistics are mostly used in the analysis when the distribution assumption is met; meanwhile, non-parametric methodologies such as Wilcoxon Rank-sum test, Wilcoxon Signed-rank test, Mann–Whitney U test, etc., are also widely employed. In addition, resampling techniques such as bootstrap, jackknife and permutation tests are commonly utilized for various statistical scenarios and purposes.
Power and sample size calculation is essential for research proposal and statistical analysis plan (SAP) development. Given its nature of flexible modeling, simulation is widely used among BSSI statisticians and statistical analysts to estimate power and sample size for a given study. Power analysis has been used to support experimental design and statistical analysis strategies, including but not limited to selection of multiplicity control (Bonferroni, Holm’s, FDR) procedures in the absence or presence of correlated tests, finding the optimal ratio to split datasets into discovery and validation sets in Genome-Wide Associations Studies (GWAS), identifying statistical methods with optimal power for analyses with small sample size for rare variants, adaptive design, and combining different trials to gain power.
We have extensive statistical analyses experience with large scale and complex data from several different data platforms. Based on association testing of high-throughput genetic and non-genetic biomarker data, a novel subgroup identification algorithm (MMMS) was developed at BSSI to identify a biomarker-driven subgroup with enhanced treatment effect for binary, continuous or time-to-event outcomes. A parametric/semi-parametric bootstrap method is proposed to evaluate the significance of treatment-by-subgroup interaction effects. Demonstrated by Monte Carlo simulations, type I error rate is controlled at design level.
In addition, BSSI has experience in the analyses of complex data as it relates to several different study designs and/or disease outcomes. This includes, parametric, non-parametric, variance component and linear models, adaptive design, survival, logistic and linear regression, power and sample size, equivalence, non-inferiority and superiority, Bayesian approaches, permutation, simulation, just to mention a few. We approach these as part of our statistical tools. Just as a set of surgical instruments are life-saving in the hands of a skilled surgeon, so are statistical tools only valuable in the hands of a skilled statistician who knows which tools to use when and under what circumstances for each individual client. In short, we offer the tools to help our clients succeed.
Our internal research and method development efforts evaluate new statistical methodologies to identify statistical approaches that better fit the data and/or address a specific client’s question more accurately.