Association for Behavior Analysis International

Behavioral data science is an emerging interdisciplinary field at the interface of behavioral science and data science. Behavioral science aims to understand why people emit specific behaviors in specific contexts. Data science aims to generate insight from large data sets using mathematical and computational analyses. Behavioral data science aims to gain behaviorally-grounded insights from large-scale data sets to answer questions of basic or applied interest. This symposium provides the attendee with a broad understanding of what behavioral data science is by describing the skills and methods behavioral data scientists use and the types of questions they ask. This is accomplished via example wherein researchers across four presentations demonstrate how: (1) time-series and geographical analyses forecast BACB certificant demand; (2) network analyses identify trends and gaps in published behavior analytic science; (3) computational techniques efficiently compare multiple behavioral models of choice in natural contexts; and (4) machine learning allows us to predict the next response made in dynamic contexts. Behavior analysts who learn the skills of data science can likely ask questions novel to the science of behavior analysis and develop novel applied behavior analytic interventions.

The audience should have a general understanding of operant contingencies and issues of relevance to the field. However, every presentation is aimed at explaining what behavioral data science is and how it can be used. The goal is to be an introduction to this topic so interested audience members can follow-up afterward to learn more.

The generalized matching equation (GME) predicts behavior allocation based on the relative amount of reinforcement contacted by each behavior. Dynamic state variable (DSV) models predict behavior allocation based on variables that change dynamically over time. To use these models in nonlaboratory settings, researchers must identify the temporal window over which to aggregate response and reinforcer rates. This study demonstrates how computational techniques can identify the optimal temporal bin for fitting the GME and DSV models to data collected in the nonlaboratory context of basketball games. For both models, the dependent variable was the logged ratio of three-point and two-point shots taken. For the GME, the independent variable was the logged ratio of three-point and two-point shots made. For DSV models, the independent variable was the difference in points scored between opponents during the previous temporal bin. For each model we: calculated prediction accuracy over temporal bins ranging from 30 s intervals to 2880 s (entire game); identified the optimal temporal window; and determined the conditions under which each model generated the highest predictive accuracy. Overall, the methods used here demonstrate how computational analyses can be used to efficiently describe and predict nonlaboratory human behavior.

Forecasting methods for time series data have been used for quite some time in various applications of behavior analysis, and yet are rarely used in our literature. For example, a popular single-subject experimental design textbook (Barlow, Nock, & Hersen, 2009) has a chapter on statistical methods—including forecasting methods—for behavioral data. In this presentation I will review an application of forecasting to BACB certificants in the state of Texas and show how those data are changing both in the aggregate and geographically by region. The field of behavior analysis, as measured by certificant numbers, has been growing exponentially. While these are not behavioral data, the methods discussed are relevant to behavioral interventions. Understanding certificant trends can help behavior analytic organizations plan for the provision of behavior analytic services. In the context of this application, both the advantages and disadvantages of forecasting methods will be discussed. Further resources for learning about these methods will be provided.

Communities survive if the behavioral repertoires of individual members within the population vary enough to withstand selective pressures. For scientists, one way to measure the total population repertoire and the evolutionary dynamics of ideas might be through analysis of peer-reviewed publications. Natural Language Processing (NLP) is one set of tools that allow researchers to analyze textual data at scale. Here, we used NLP to describe the evolution of behavior analysis by identifying the key characteristics of publications over time. To do this, we gathered 1500+ peer-reviewed publications from the Journal of Applied Behavior Analysis and the Journal of Experimental Analysis of Behavior. For each article, we collected data on publication year, title, authors, abstract, keywords, manuscript, and references. Once obtained, we analyzed the differences and similarities in research topics between the two journals and used network analyses to identify citation patterns within the research literature. Future research aimed at understanding the variation, selection, and evolution of topics studied by behavior analysts might be important for three reasons. First, it gives data to conversations about how the field allocates resources to promote understudied topics, variation in studied topics, or high impact topics and that are low-hanging fruits. Second, it may help junior researchers identify gaps and niches upon which to build a career. Lastly, it could highlight gaps in the research literature that, if filled, would benefit applied practitioners. The methods of behavioral data science make these benefits easier to obtain and more robust in their methodology and findings.

Molecular analyses predict and control behavior through discrete responses strengthened by contiguous reinforcers. Molar analyses predict and control behavior through response-reinforcer relationships aggregated across a temporal window. Unified analyses aim to leverage molecular and molar analyses to describe, predict, and control behavior. Here, we sought to take a unified analytic approach wherein quantitative analyses of behavior and machine learning combined to predict the next response a human made. To do this, we obtained data on every pitch thrown by a pitcher during the 2016-2019 Major League Baseball seasons. The dataset contained information about the game context, the pitch type and characteristics, and the consequences that followed every pitch. Molecular information was included through a weighted decay function placing greater weight on more recent reinforcers and by making response-by-response predictions. Molar information was included through dynamically updating covariance relations between game context, pitch type, and pitch consequence via the generalized matching equation. Machine learning combined raw data, molecular information, and molar information to predict the next pitch. The dynamic unified model of behavior led to higher response-by-response prediction accuracy than the molecular and molar approaches alone. This experiment demonstrates how behavioral data science can describe and predict dynamic human behavior.