Over the past two decades, statistical methods have increasingly become part of the industrial R&D toolkit – from the pharmaceutical industry to high-tech manufacturing and beyond. The promise of such methods as statistical design and analysis of experiments was initially to slim the material and time inputs needed to arrive at an optimal result; that is, to learn as much as possible from as small a data set as possible. But now, in the age of automated labs, machine learning, AI, and statistical methods can offer a new advantage: a way to sift through gluts of data and extract meaningful relationships quickly.

Dr. Phil Kay, Director of Global Presales at JMP, spent more than a decade in laboratory science before entering the field of data science. He explains that “ten years ago, modern experimentation meant optimal DOE and customizing experiments to give you a bespoke experiment for your particular situation.” Nowadays, though, he explains that “the state of the art is how we use lab automation and these bigger experiments that are now possible.” This seemingly dramatic shift has the same end goal, Kay says: “having a data-driven approach to learning faster and having more information-rich data.”

Whether in the context of DOE or lab automation, the crucial piece to integrate data-forward thinking across an enterprise is education. Importantly, this education doesn’t need to focus solely on the statistics; in fact, quite the opposite is true. Kay says that previous pedagogies would mandate “spending two days learning about regression and ANOVA before anyone would allow you to learn about DOE,” but that “those ideas have been debunked now.”

Instead of getting down into the mathematical weeds of statistically designed experimentation, Kay argues for developing intuition about what’s really going on, for instance, when performing multivariate regression. If a scientist feels comfortable fitting a line through a scatter plot, then extending that rationale to multiple dimensions is not a formidable leap from simply fitting Y by X. The task is bridging the gap between expectation and reality in terms of statistical education. Scientists need to know they don’t have to “understand all the T-tests and standard errors and all these statistical terms that might be intimidating [to a non-statistician],” Kay says, and instead the focus should be on building an intuitive sense of what these statistical methods really do.

Fortunately, there is a plethora of content available to learn statistical methods, matching each scientist’s preferred learning style. When just a few researchers become faithful users of statistical methods, they help get their colleagues on board, and a statistically minded researcher base is created. Then, a synergistic relationship with company leadership can emerge. Kay explains “leadership should be brought into [the conversation] to really empower those enthusiastic champions and allow them to dedicate some of their time to helping others.” Once early adopters have documented the value delivered by initial use cases, momentum to adopt statistical methods invariably grows. Scientists seeking out training can point to colleagues’ successes to secure buy-in from management to invest time – or resources – in honing their statistical skills and thinking critically about where methods like DOE can augment their science. “To spread statistical fluency across an organization,” Kay says, “you need to make it fun, and you need to make it accessible.”

Once these pockets of data advocacy grow, they act as catalysts for a burgeoning shift in mindset with the potential to reach across the enterprise. Kay explains that an end goal when doing statistical experimentation is that it challenges you to be more rigorous. “It’s really going to uncover where previously you were not making solid progress,” he continues, explaining that seeing trends contradictory to previous paradigms should be expected. In his view, DOE shouldn’t just be a tool for finding new relationships but also for uncovering entirely new classes of insight. Kay says “the biggest misconception is that DOE is only for special occasions, but DOE should be the default. Any time you are experimenting to understand a multifactor situation, DOE is the way to do it.”

These principles of intuitive education will only become more important with the implementation of automated labs. Kay explains “as we move increasingly to higher-throughput experimentation, where you can do hundreds or even thousands of runs within a day, we’re moving from the world of statistics to the world of machine learning.” He continues that “if you’ve got an organization that’s bought into using DOE, then you’ve got the potential for even more powerful data analytics, and you can start to think on a more holistic scale.” In this way, the adoption of more traditional statistical techniques like DOE enables the most effective exploitation of novel innovations such as automated laboratories and high-throughput experimentation.

Ineffective education – that is, a paradigm centered heavily on statistical theory – may discourage scientists to even try new methods, with the downstream effect of stymying statistical adoption across an enterprise. Conversely, if scientists feel they can seamlessly integrate techniques like DOE into their existing workflow, they are more likely to truly champion them. Then, adoption in teams and departments follows naturally, aided by mentorship. Data advocates should mentor new teams that want to learn statistical methods and walk them through the process to a point where they can see some success, Kay says, and it will grow from there. Then, teams can recognize they don’t need a Ph.D. in statistics to be successful; they just need the right education to realize that statistical methods are a bolster, not a headwind, to doing their best bench science.