Characterizing the “Outlier Profiles” of Diagnostic Products
*Krista Michelle Birch, Abbott Labs 

Keywords: outliers, diagnostics

What is an “Outlier Profile?” These days, the word “Profile” often makes us think of social media, right? Facebook, Linked-In and other personal profiles describe us. It is my experience that diagnostic products have their own Outlier Profiles.

Statisticians know that outliers are values that are located a relatively far distance from the rest of the values in a data set. From Webster’s, “An observed value that appears to be discordant from the other observations in a sample.” These definitions make it sound as if outliers are unexpected events. I maintain that outliers are in fact expected events and so we should be prepared, not only to deal with them in our data sets, but to get to know them! Where and when do they occur? What direction and what magnitude do they have? Is their frequency increasing, decreasing or fairly consistent? Most importantly, is their rate of occurrence “bad” or acceptable?

Traditionally, outliers are said to be due to “special cause variation,” while the rest of the variability is said to be due to “common cause variation.” The question often arises, “How can we tell if a data point is truly an outlier and not the result of shifted common-cause variability?” The characteristics of a data set can dramatically affect our ability to identify when an outlier has truly occurred. Outliers themselves can distort the interpretation of data and have undue influence on summary statistics, especially with small data sets. We will see how pooling data from multiple studies, using graphical pictures and only basic statistics can improve our ability to identify these true special cause events. Once we can identify outliers using these methods, we can then characterize their behavior and understand their “Profile.” When enough data is gathered by pooling what would otherwise be unrelated data sets, we often find that the outliers due to special cause variation actually have their own common cause! This can then open the door for us to minimize their frequency and magnitude by implementing product design improvements. Finally, we will see how the characterization of the Outlier Profile allows us to simulate and predict the impact of outliers to the ultimate success of the product.