Intertek ASG Computational Chemistry Group

Design of Experiments and Statistical Analysis

Design of Experiments (DOE) is an efficient statistical procedure for planning a series of experiments such that the data obtained can be analysed to yield valid and objective conclusions. It can be used when you are looking to screen a set of variables to identify those with most effect, design a series of experiments to investigate the influence and effect of different parameters, optimise an experimental process or retrospectively analyse a set of experimental data. Both the design and analysis steps require the application of techniques for statistical data analysis.

To find out more on these areas see below:

• Design of Experiments
• Statistical Analysis

Design of Experiments

In an experiment, we deliberately change one or more process variables (or factors) in order to observe the effect that the changes have on one or more response variables. The (statistical) design of experiments (DoE) is an efficient procedure for setting up a detailed experimental plan in advance of doing any experiments. Well chosen designs maximise the amount of information that can be obtained for a given amount of experimental effort. The analysis of an experimental design is built on the foundations of Analysis of Variance (ANOVA).

A complete DoE procedure will involve the following steps:

• Factorial designs for screening

Email ASGlab@intertek.com for more information

Statistical Analysis

Robust statistical techniques are often required to reveal the information contained in a multi dimensional data set. They are also used to confirm the statistical significance of observed trends.

Data summarisation

All factors and responses can be summarised to give basic statistical data such as minimum, maximum, standard deviation, confidence interval etc. This data can also be displayed graphically as a box plot.

Understanding variability

It is crucial that you understand the extent and sources of variability within your data as this will impact the level of confidence for any model built upon it. We use statistical tools that allow the nested analysis and quanitification of batch to batch variation. This understanding will allow you to focus more experimental time and resources on understanding and minimising the major sources of your variabiity.

Sampling strategies

In order to make most efficeint use of experiments you should create a sampling strategy based on the experimental variability.

Numerical models and trends (MLR,PLS)

For a given block of data, numerical models can be built and trends analysed. A numerical model will allow you to find the optimum conditions and response. It will also predict results for areas where there is no experimental data. These numerical models are analysed statistically against the known variability to provide confidence and avoid over-fitting. Techniques used include:

• Multiple Linear Regression (MLR)
• Partial Least Squares (PLS)

Diversity and Domain of Applicability

It is important to understand the domain of applicability for any model that is built. Similarly, it is important to take account of the diversity of your data. Outliers do not build good numerical models.

Data clustering (PCA, Hierarchical, K-Means, Self-Organising Maps)

One route to simplifying data analysis is to cluster data. This clustering can also reveal patterns and trends in the data.

Some of these areas are also described on the following web pages:

• Informatics Modelling and QSAR
• Scienfific Data Management

Email ASGlab@intertek.com for more information