When you are talking about clinical trials, missingness tends to be informative.
Clinical trials are essential for evaluating new medical treatments. Interpreting the results, however, is not necessarily straightforward. Statistician Abby Flynt is working on new analytical methods to help researchers get the most out of their data.
Most clinical trials are longitudinal studies. The same subjects are measured repeatedly over weeks, months or even years to track their progress. Over that time, the subjects might respond in any number of different ways; some might improve quickly, some more slowly, while still others might not improve or even get worse. Being able to cluster subjects according to the trajectories they follow could create a powerful predictive tool for future patients.
"Being able to say after a few weeks, 'With a high probability, you are on this curve and you are going to see improvement in a couple of weeks,' — that's great," Flynt says. "Or the news might be, 'You are following this curve with such a high probability that we don't think you'll get better on this treatment, so let's try something different instead.'" Either way, if Flynt's work on how to cluster subjects enables doctors to predict patient responses early in treatment, that could save time, money, and even lives.
Longitudinal datasets pose an additional challenge. Such repeated measurements lend themselves to missing data. "People just don't come, they drop out, things happen," Flynt says.
The most common way to deal with missing data is to throw them out. That, Flynt says, can cause a complete misinterpretation of the results. "When you are talking about clinical trials, missingness tends to be informative," she says. For example, in a study of people being treated for depression, Flynt says subjects may be more likely to miss appointments when their depression is worse. By figuring out how to tap the information in missing data, Flynt's work could lead to more accurate conclusions being drawn from longitudinal studies.
Clinical studies are only one type of longitudinal data. Flynt, whose undergraduate degree is in secondary mathematics education, also plans to apply her work to studies of educational strategies that test student learning over time.
As an educator herself, Flynt has seen a positive shift in student attitudes toward statistics, driven by the promise of good jobs. Companies such as Google, Facebook and Twitter are hiring statisticians to analyze the enormous datasets they are gathering. "It's becoming a cool field to get into," she says.
Posted October 2012