Not a research scientist, but I get to play one on TV
Design of experiments (DOE) is potentially one of the most strategic weapons in your analytic arsenal. DOE is core to learning faster from data and can be applied in many areas — not just traditional areas like manufacturing, but also in marketing, HR and a whole host of areas. As Bradley Jones, Principal Research Fellow in the JMP division of SAS, says, “designed experimentation is the fastest and least expensive approach to successful innovation.” Why then aren’t people falling all over themselves setting up experiments to learn faster from data?
These and other questions were explored in recent conversations I had with Brad and Peter Goos, full professor at the University of Antwerp. They co-authored the recent book, Optimal Design of Experiments: A Case Study Approach. This is a great book because it leapfrogs you to the new advances in DOE and does so in a way that makes it conceptually very understandable — even to the novice. But it goes further than that.
There has been a slow revolution in DOE. No longer are we limited to using textbook designs. Thanks to more automated design searches and the ability to flexibly account for the many constraints we face in the real world, due to the work of people like Peter and Brad, we are now able to custom-design our experiments and pick the designs that fit our problem rather than force our problems to fit the designs. Through the case studies in their book, the authors show that you can tailor-make experiments that address a variety of problems to provide truly useful results. You no longer have to let the tail wag the dog by limiting your choice of experiments to tabled designs, which all too often don’t fit the problem at hand.
We took an example from their book, Chapter 7 – a response surface design in blocks — and performed it as a skit in the new Analytically Speaking webcast, which premiered today. I got to play the role of the research scientist who seeks help from experts on a design problem. Brad and Peter play the role of expert — in the book and in real life — sharing their expertise to solve the problem. The case study approach is very effective in illustrating the value (and in many cases ease) of custom or optimal designs.
Another of my favorite (paraphrased) quotes from Peter: The most important recent development in DOE is the availability of fast and powerful algorithms for computing good experimental designs for basically any possible experimental scenario — not just for completely randomized experiments, but also for experiments involving correlated observations such as blocked experiments or split-plot experiments. Moreover, we can now also compute diagnostics such as the fraction of design space plots, which visualize the performance of a design in terms of the variance of prediction.
To ensure you are equipped with the latest thinking in DOE and creative problem-solving, check out more from Peter and Brad in the webcast and their book. Peter and Brad have written a number of journal articles in Biometrika, Journal of Quality Technology, Journal of Choice Modeling, Journal of Business and Economic Statistics, and Applied Stochastic Models in Business and Industry.