As I stated in last week’s post, most beginning experimenters try to combine everything into one experiment. The thinking is if they can include everything, they can do one experiment and be done. But it never works that way. Instead it is an iterative approach, where you learn as you go, getting a better understanding with each experiment conducted.
So where and how do you get started? Keep in mind that the primary considerations that distinguish designs are the number of factors or variables they include and the complexity of the model they provide. With that said, there could be many experimental designs to choose from for a specified number of variables.
Screening designs are used to study a large number of factors for the purpose of identifying the most important ones, the most common of these designs being two-level full and fractional factorial designs and Plackett-Burman designs. These designs are useful for fitting first-order models, which detect linear effects, and can provide the existence of curvature if center points are included.
Plackett-Burman (PB) designs can handle a large number of factors and are very efficient if main effects are of interest but are risky because they use only two levels of each factor and cannot resolve interactions between pairs of factors. Below is an example of a PB design where in a twenty run experiment nineteen factors are evaluated. The horizontal heading at the top shows each factor being evaluated and the + indicates the factor high setting and the – indicates the factor low setting. The vertical numbers down the side indicate each experimental run. For instance, the first run requires all factors to be set at the high setting. It is important to randomize the runs in an experiment and not run them in sequential order. By randomizing, the effects of any unobserved changes in the process unrelated to the experimental factors are uniformly and randomly distributed over all the levels of the experimental factors.
So there you have it, a way to determine the vital few from the trivial many when you’re faced with a situation where there is little prior knowledge or experience. The best strategy is to employ a series of smaller experiments instead of committing all time, effort and resources to one experiment. The first experiment you should consider is a screening experiment to determine the most influential variables from the many variables that could affect the process. By doing this you’ll usually identify two or three significant variables that dominate the process. The next step is to build a more complex design using the key variables identified in the screening experiment. This will be a topic for a future blog article.