1 OCTOBER 2010 VOL 330 SCIENCE www.sciencemag.org 38 EDUCATIONFORUM E ducation researchers have struggled for decades with questions such as “why are troubled schools so difficult to improve?” or “why is the achievement gap so hard to close?” We argue here that con- ceptualizing schools and districts as com- plex adaptive systems, composed of many networked parts that give rise to emergent patterns through their interactions ( 1), holds promise for understanding such important problems. Although there has been consid- erable research on the use of complex sys- tems ideas and methods to help students learn science content ( 2), only recently have researchers begun to apply these tools to issues of educational policy. We roughly categorize existing educa- tion research into two categories, “mecha- nism based” and “effects based.” Mecha- nism-based studies include ethnographies, case studies, and laboratory experiments that focus on understanding individuals and their interactions inside schools. Such work has provided insight into the motivation and cog- nition of students, teachers, and school lead- ers, as well as how social phenomena unfold inside schools ( 3, 4). Effects-based research treats factors contributing to academic per- formance of schools as inputs that work together to yield a particular level of student achievement ( 5). By analyzing variation in quantitative observational data on inputs and outcomes ( 6), or through the execution of field experiments ( 7), effects-based research has increased our understanding of the rela- tive importance of factors such as teacher- pupil ratio, family background, and teacher quality and has established effects (or lack thereof) of specific interventions designed to improve achievement. What Works, How It Works Despite of such successes, and as evidenced by the call for more “mixed-methods” designs ( 8), a key challenge facing education research is to integrate insights about “micro- level” processes with evidence about aggre- gate, “macro-level” outcomes that emerge from those processes. For example, suppose we have results of a well-executed experi- ment using a nationally representative sam- ple of schools that indicates students in small classes perform better than those in large ones. Although it is tremendously helpful to know that, on average, students in small classes do better, this alone is not enough to fully understand what changes policy-makers and school leaders should implement. One reason for this is the issue of hetero- geneous treatment effects ( 9). If small class size mattered under certain conditions but not others, school leaders would need to under- stand what happened differently in some small classrooms that led to better student outcomes. Both mechanism- and effects- based research may be helpful here, exam- ining differing contexts and how programs are implemented. But we still face the chal- lenge of aligning micro-level accounts with aggregate data. This is all the more difficult when considering inherent impediments to understanding complex systems: Effects are disproportional to cause; cause and effect are separated in time and space; and prop- erties of the macro-level system may be confused with properties of constit- uent, micro-level elements (e.g., attrib- uting intelligence to individual ants when observing an entire ant colony intelligently gathering food) ( 10). Additionally, we need to consider what are often referred to as “general equilibrium effects,” i.e., the systemic impli- cations of class-size reductions enacted at a large scale ( 11). For example, partly on the basis of results of a randomized field exper- iment in Tennessee, California mandated statewide class-size reductions. However, many school districts had to hire teachers with limited training and credentials because the supply of qualified teachers was too small to handle the sudden increase in demand ( 12). If identified a priori, we can try to account for such effects using econometric models esti- mated from observational data ( 13). Although such models can often help characterize par- ticular equilibrium states of educational sys- tems at a larger scale, we are still interested in an additional, policy-relevant step: how to best move the system from one equilibrium state to another. Regardless of how well we account for heterogeneous treatment and general equilibrium effects, complex systems methods can help bridge these aggregate out- comes to underlying mechanisms at work in the system, as well as discover new and unan- ticipated systemic consequences. Bridging the Gap Applying a complex systems perspective to education research parallels the recent use of complex systems methods to model the spread of epidemics ( 14). Traditionally, one relied on (i) detailed case studies that traced social con- Complex Systems View of Educational Policy Research EDUCATION S. Maroulis, 1,2,3 R. Guimerà, 1,4 H. Petry, 2 M. J. Stringer , 1 L. M. Gomez, 5 L. A. N. Amaral, 1,6 U. Wilensky 1,2 * Agent-based modeling and network analysis can help integrate knowledge on “micro-level” mechanisms and “macro-level” effects. 1 Northwestern Institute on Complex Systems, Northwest- ern University, Evanston, IL 60208, USA. 2 Center for Con- nected Learning and Computer-Based Modeling, North- western University, Evanston, IL 60208, USA. 3 Ford Motor Company Center for Global Citizenship, Kellogg School of Management, Evanston, IL 60208, USA. 4 Institució Cata- lana de Recerca i Estudis Avançats (ICREA) and Chemical Engineering, Universitat Rovira i Virgili, Tarragona 43007, Catalonia. 5 School of Education, University of Pittsburgh, Pittsburgh, PA 15260, USA. 6 Howard Hughes Medical Insti- tute, Northwestern University, Evanston, IL 60208, USA. *Author for correspondence. E-mail: uri@northwestern.edu ABM of school choice in Chicago. Students, represented by small dots, are shown in their census block. Dark red indicates high-poverty locations; dark green, low. Each circle rep- resents a school. Color reflects the expected academic performance of students attending the school, given the experimental conditions of the model. Light blue indicates high mean achievement; dark blue, low. Circle size is pro- portional to expected enrollment. For more information, see the text and (32). Published by AAAS on May 13, 2017 http://science.sciencemag.org/ Downloaded from