STEP combines systems thinking, simulation modelling, and empirical analysis to study education systems and support policy decision-making.
STEP’s methodological approach
STEP studies education as a complex system in which outcomes emerge from interactions, feedback, and accumulation over time. Our work combines systems thinking, simulation modelling, and empirical analysis to understand learner progression, institutional performance, and the effects of policy interventions across South African education contexts.
The methodology is deliberately mixed. No single technique is sufficient to capture the dynamics of education systems.
Problem framing and system conceptualisation
All projects begin with clear problem framing. We define system boundaries, key actors, time horizons, and outcomes before modelling. Conceptual models use causal loop diagrams to identify feedback, delay, and interaction effects, ensuring that simulations reflect plausible system structure rather than convenient assumptions.
Statistical mapping and structure
Where suitable data are available, we use multivariate statistical methods to quantify relationships within education systems and to reduce complex survey data into interpretable constructs. Methods include factor analysis, multilevel regression, and structural equation modelling. These analyses identify key factors and their relationships across individual, household, school, and institutional levels, and provide an empirical foundation for subsequent simulation modelling.
Computer simulation modelling
We use simulation to study how education systems change over time. System dynamics models capture aggregate behaviour, feedback, and delay, while agent-based models represent individual pathways and interactions. Both approaches are used to test policy scenarios, explore unintended effects, and understand trade-offs between interventions.
Machine learning for prediction and validation
Machine learning is used selectively for prediction and model validation, not as a standalone decision tool. Models trained on survey and administrative data identify key predictors of progression, support simulation design, and test robustness across cohorts, with interpretation and transparency prioritised over predictive accuracy alone.
