The Biostatistics Team offer a rare combination of statistical and genetic expertise, providing clients with optimal strategies and solutions in the molecular diagnostics field. Our biostatisticians are available from early stage assay design to late clinical assessment, and have excellent knowledge of the current regulatory recommendations and guidelines.
Much more than a traditional group of analysts, our biostatistics team help ensure that the assay meets client’s exact customised needs, that studies are designed to ensure success while minimising cost, and that the right analyses will be chosen to facilitate appropriate characterisation of the assay.
To find out more on each of our biostatistical methodologies, select the appropriate drop-down below:
Design – Early Clinical
Multivariate Statistical Learning
While it is always hoped that the trait of interest (e.g. response to drug) will be Mendelian, complex traits are much more common. We have extensive experience developing complex trait models using genetic and genomic data with techniques such as LASSO, classification approaches, and Bayesian regression (e.g. Bayesian LASSO). We will work with our clients to ensure that the most appropriate model is chosen and appropriately describe its expected performance.
While many diagnostic tests are quantitative in nature, there is often a need to define one or more thresholds and produce a qualitative output. We have experience assigning such thresholds using many approaches such as percentiles, orthogonal tests, and clinical data. We help our clients determine which approach is appropriate for their situation and work through the details, such as incorporating recruitment rate and sensitivity/specificity costs.
Design – Early + Late Clinical
Outlier Detection in qPCR Triplicates
While much statistical work has been devoted to outlier detection, this has primarily focused on “big data” and multi-dimensional problems. When working with qPCR triplicates, we have the exact opposite problem…how does one detect an outlier when only three data points are present? Given that statistical test-based methods are inappropriate for such small sample sizes, we have developed a nearest-neighbour based approach that leverages our experience from running thousands of qPCR chips.
Independent Verification Scripts
There are many activities that will increase the odds of accurate results, and one of the most basic and fruitful is having an independent operator verify the final output. We employ multiple levels of verification to meet the specific needs of our clients and utilize automated scripts to ensure highly accurate, fast, and reproducible results.
Standardisation and Normalisation
Given that most RNA quantification techniques produce estimates of relative expression, standardization (and possibly normalization) is always necessary. We have extensive experience dealing with qPCR, NGS, and nanoString data and will help our clients understand the different options and choose the one most applicable for their situation.
Early & Late Clinical
Regardless of the quality of a diagnostic, a poor study design will do no more than ensure that time and money are wasted. We work with the client to quantify risks and ensure that the design is appropriate for the questions being asked. Then, we incorporate cost (actual processing cost, sample availability, etc.) to design a tailored study that has the best chance of succeeding and meeting the client’s needs.
Statistical Analysis Plans that Meet CLIA and /or Regulatory Requirements
While guidance documents are helpful, they are rarely detailed enough or necessarily pertinent to diagnostic devices, especially ones that use emerging technologies like Next Generation Sequencing. Our statisticians have written dozens of statistical analysis plans that are compliant with CLIA and/or regulatory requirements and we can help our customers make sure the appropriate factors are considered.
Single / Dual Censoring
When dealing with quantitative or semi-quantitative assays, there is always a lower point when quantitation becomes unreliable. Such measurements are considered censored, but they still contain valuable information. We work with our clients to understand the information contained in such values and how to apply it so that sample failure rate is minimized while introducing no bias.
For many analytic validation studies, the goal is to demonstrate that a perturbation (sample input amount, potentially interfering substance, etc.) does not affect the measurement. Classical statistical tests, where the null hypothesis is that no difference exists, are inappropriate for such circumstances. Instead, equivalency tests should be used, and we can help our clients understand their interpretation as well as how to power the analysis and set appropriate acceptance criteria.
Adaptive Sample Size Designs and Bayesian Analyses
For many studies, sufficient information is unavailable to reliably calculate the power for a classical statistical test. Adaptive sample size designs can be helpful in such situations since they allow the user to look at the data and decide if they should stop (since the answer, whether good or bad, is well supported probabilistically) or whether they are in an ambiguous state that requires more information. This, coupled with a Bayesian analysis, produces a very cost effective design that minimizes the number of samples while also lowering the risk of the study failing to due insufficient data.
Censored mixed models
In many analytical validation studies, it is possible to obtain censored values. In such situations, it is inappropriate to “remove” such observations or impute values due to the bias this introduces. We work with our clients to implement censored mixed models to provide unbiased estimates that allow for standard acceptance criteria.
MC simulations for power estimation
Once a typical Phase III trial is ready to begin, there is usually plenty of information about how the diagnostic works. For certain outcomes, conducting power analyses is straightforward, but in other situations, the well-established deterministic methods do not incorporate all of the available information. In such instances, we can help develop Monte Carlo based simulations to provide the most accurate power and sample size estimates.