In the world of training research, many of the relationships we consider rest on a variety of statistical assumptions. First, we assume that we’re using comparable units of measurement that are at similar levels of specificity—in other words, not only are we comparing apples to apples, but we’re using the same ruler to measure how big the seeds are. But perhaps more importantly, we often assume linearity between the variables under study. That is, for a unit of A to go up, a unit of B goes up or down accordingly, often with little or no attention given to where we are on the continuum of A units. We double A, we’ve just doubled something about B. We decrease A, B will sensibly follow suit.

However, in the real world, we know that all sorts of funky bell curves and inverted U-shaped relationships abound. An easy example is the effect of acetaminophen on a headache. A little bit won’t touch your headache, too much of it will shut down your liver, so somewhere in the middle lies a sweet spot of impact. Is L&D immune to this sort of phenomena? Sorry to disappoint, but not at all. As much as we’d like to be able to say something like, “If we can increase the number of employees in this training, it will result in better ROI!” the truth is never that simple or straightforward. I’m not carelessly using the word “never,” either. While we can demonstrate a linear relationship for many things, the common fault lies in believing that these sorts of relationships will hold across all applications, or even within a single application across time. The world of work is constantly churning, and we all wrangle with a barrage of tasks and inputs and outputs, within which training often gets shoehorned into the mix. When things look linear within a narrow focus, because of the constricted nature of data that’s available, or because of the way the information is being analyzed, there could be a hidden curve that’s being obscured by incomplete information. Said another way, thinking the trees are all pointing straight up because the sky is visible through the leaves might distract you from the fact that the entire forest slopes up a mountainside.

So why do we often start by assuming things are related in a linear fashion? Quite simply, it’s the easiest to wrap our heads around. But that doesn’t necessarily mean an analysis is accurately doing anything that we’d like it to. Can we describe a nonlinear relationship? Of course, it’s a matter of finding a model that fits the data and is able to describe it with some degree of certainty—and describe it better than relying on a straight line. The trouble comes when we want to begin making predictions about phenomena that are described by a nonlinear model. The reason for this is that in linear relationships, we can parse out the effects of other things. If we want to control for the influence of training class size, or the use of different modalities, on the relationship between course content and learning outcomes, that’s not incredibly hard to do if things are comfortably linear. But when they’re not, class size may sometimes make a difference, modality may interact with class size in some instances but not always, and our ability to predict the relationship between course content and learning outcomes can get significantly more complex and riddled with caveats and dependencies that need to be accounted for.

So are our models of how training works basically masquerading as best practices? No, of course not. They all have a sweet spot. However, we need to be aware that they have limited utility in the sense that one size does not necessarily fit all, no matter how much we’d like it to.