While it is natural for research disciplines to split into more specialized areas, the resulting divide complicates efforts to share the benefits of related research. Constructs are important underpinnings of a research method used in many behavioral and social sciences, and virtually identical constructs have been developed to support different theories, frequently under different names. As a result, useful connections are missed, constructs and theories are reinvented, and little knowledge exists about the construct origins and flow among disciplines.
Intellectual Merit: This project addresses the problem by investigating the following research question: is it possible to identify closely related constructs, including those in different disciplines and use this information to measure the extent to which existing constructs are utilized effectively? The question is examined in the domain of 'latent construct research using scales,' which is typified by the studies using questionnaire scales to assess latent constructs, thus spanning multiple disciplines in the behavioral and social sciences. For this domain, the notion of a Closely Related Construct (CRC), is formulated and operationalized using a method that integrates automated text analysis, citation analysis, and meta-analysis. Using CRC, a science metric termed Construct Utilization Ratio, is formulated that measures the extent to which closely related constructs are recognized between two units, where the unit may be chosen at different collective levels, such as between theories, journals, or research areas.
Broader Impact: The project examines how this metric can be used to quantitatively and qualitatively assess the extent to which CRCs for a given construct are recognized and utilized within and across different research areas. It can also reveal where opportunities and redundancies reside.
The science of behavior and behavior change is still underdeveloped. Epidemiological evidence from the USA and UK suggests that half of morbidity and premature mortality is caused by behavior patterns that could in principle be changed. Interventions to change some of these behavior patterns have been successful and found to be highly cost effective (e.g. behavioral support for smoking cessation). Unfortunately, because research on behavior and behavior change often uses unobservable constructs, there is extensive renaming and recreation of research that has already been done. This project created the first major tools for detection of construct similarity across studies. First, we created the Construct Identity Detection (CID) algorithm and tested it against four other algorithms that had never been applied to the problem of construct synonymy, showing that CID works best for the problem at hand. Second, we since started work on automating the collection of constructs from academic papers and created the Automatic Nomological Network Extraction (ANNE) algorithm which detects hypotheses in articles, detects construct names in those hypotheses and connects the constructs across hypotheses to create a network for each article. Our experiments shows that extracting construct information from articles holds the potential to improve expert finding of synonymous constructs more than six fold. We are now working with teams from the U.S., Europe, and Australia to further develop the science of construct detection and integration for the purpose of creating a theoretical backbone for the behavioral sciences.
