Quantifying the Unobserved: Gathering and Organizing Difficult Data
March 1, 2019
Some people tend to be more agreeable than others. Some suffer more depression symptoms than others. Yet personality traits such as agreeableness, and psychopathological syndromes such as depression, are difficult to measure statistically because they cannot be observed directly.
As a Quantitative psychologist, Sonya Sterba looks for better ways to measure and model such unobserved constructs. Her research both informs and corrects the way researchers interpret results from their own models, and it impacts the way models are used in both clinical practice and in theoretical research.
The first of Dr. Sterba’s three research focus areas concerns “item parceling,” which refers to the common practice of combining subsets of measurements of the same underlying condition, such as depression, to facilitate statistical analysis.
Dr. Sterba explains that there are typically dozens of items serving as measurements of the same “construct,” or condition. If there are, for example, 30 items measuring depression, five parcels might be created from the original 30 items by averaging subsets of six items each. In that case the five parcels would then be used instead of the original 30 items.
However there are thousands of ways to create five six-item parcels, she explains. “For instance, we could average the first six items together to create the first parcel, the next six items to create the next parcel, and so on. Or we could parcel together items 1, 10, 12, 13, 17, and 20 to create the first parcel, and items 5, 8, 15, 19, 25, and 29 to create the second parcel, and so on.”
Dr. Sterba’s research showed that the way the parcels are created affects the statistical conclusions drawn from the original model, and she developed ways to account for this variability in results.
A second area of research concerns models used to represent different unobserved categories or “latent classes” within a given population–such as diseased or not diseased.
“We cannot directly observe who belongs to each class, and we don’t know how many classes exist,” Dr. Sterba says.
To help determine the probability that certain people belong in a specific class, Dr. Sterba uses finite mixture models. One model might classify people according to the kinds of psychopathological symptoms they exhibit, or might distinguish persons with different patterns of change, or treatment response, over time. With competing mixture models she can determine which number of classes best fits the data. Dr. Sterba has evaluated the performance of mixture models and adapted them to handle different kinds of data.
A final focus of Dr. Sterba’s research concerns partially nested data, which frequently arise in clinical and educational settings. An example would be data from a randomized controlled trial looking at patient depression scores where patients were randomly assigned either to a wait-list control or group therapy. Depression scores are expected to be more similar among those “nested” in the same therapy group because group members all work with the same therapist. As such, partially nested data arise when individuals in one study arm are nested within groups but individuals in another study arm are not. Dr. Sterba explains that ignoring partial nesting can lead to bias in inferences about the effect of treatment. She introduced and extended models to accommodate partially nested data for a variety of designs. Her work on this topic increased modeling options for partially nested data.
Dr. Sterba says she hopes going forward to bridge her various research programs by, for instance, “developing parallel methods of model assessment that are applicable to both finite mixture models and multilevel models.”
Dr. Sterba is an associate professor and director of the Quantitative Methods Program within the Department of Psychology and Human Development at Vanderbilt University. She received her PhD from the University of North Carolina at Chapel Hill. She is the recipient of the 2018 Anne Anastasi Distinguished Early Career Contribution Award (APA), 2015 Rising Star Award (APS), and 2015 Cattell Early Career Research Award (SMEP).
Sonya Sterba is a recipient of the 2018 Federation of Associations in Behavioral & Brain Sciences (FABBS) Early Career Impact Award and was nominated by the Society of Multivariate Experimental Psychology.