When I started my career in biology (almost forty years ago), only the luckiest of us had access to advanced statistical software. Licences were very expensive, and it was not easy to convince the boss that they were really necessary: “Why do you need to spend so much money just to perform an ANOVA?” In fact, simple one- or two-way ANOVAs were not difficult to carry out, and one of my colleagues had already written the appropriate routines for several experimental designs using GW-BASIC. But I wanted more!
My agricultural experiments often involved complex designs. Split-plot, strip-plot, subsampling, and repeated-measures experiments were the rule rather than the exception. I therefore started writing a series of QuickBASIC programs to analyse those designs on my PC. In the early 1990s, nonlinear response models became increasingly popular, and I had to write my first Gauss-Newton optimiser, again in QuickBASIC. Generalised linear models were not yet widely used by biologists, so I mainly relied on variance-stabilising transformations whenever the assumptions of linear or nonlinear models were not met.
That rather humble statistical life gave me an undeniable advantage: it forced me to study and understand the principles behind every new method and algorithm before I could even think of programming it. I had to become familiar with all those strange mathematical objects—matrices, eigenvalues, determinants—which are not usually part of a biologist’s mathematical background (at least in Italy). And once my BASIC programs were finished, they could do only what they had been programmed to do: solve one specific problem, in one specific way. No additional options, no hidden defaults, no unexpected behaviour. In short, no general-purpose solutions.
Today I have R. It is free, incredibly powerful, and almost anything is possible. With just a few lines of code, I can fit almost any model that comes to mind in a matter of minutes. I can compare several alternatives: Which one is the best? Which one tells the story I would like my data to tell? Biometry textbooks have changed as well. They have become much more algorithm-oriented, while the mathematics is often confined to boxes that can easily be skipped. I have to admit that I often skip them myself: a few code snippets are usually enough to get the job done, and if they are not, I can find thousands more on the Internet. So why should I bother studying an abstract subject like statistics when I have R?
Of course, this is an exaggeration. Nevertheless, I sometimes wonder whether the widespread availability of such a powerful free tool has a downside. Biologists—especially students—may mistake learning R for learning statistics. I am often surprised by how many sophisticated models are fitted nowadays with little biological or statistical justification, and with very little attention paid to the assumptions on which those models rely.
A few days ago, a PhD student in my department proudly showed me the results of a reduced-rank regression fitted to a biological dataset. He was particularly pleased that he had managed to write the correct R code after a long search on the Internet. He had even succeeded in getting rid of a “rather strange” warning message. Unfortunately, he had misunderstood what that warning meant, and the analysis had therefore been incorrect from the very beginning.
So here is a friendly reminder to all biologists—including myself. R is a wonderful piece of software, but it does not automatically lead to sound statistical analyses.
Let’s use R—but let’s study stats first!
Prof. Andrea Onofri
Department of Agricultural, Food and Environmental Sciences
University of Perugia (Italy)
Send comments to: andrea.onofri@unipg.it
