Improving Digital Competence of Prospective Mathematics Teachers through Project-Based Learning in Co-Teaching with Practitioners: A Structural Equation Modeling Analysis

Background Twenty-first-century competency demands on mathematics teachers require robust digital technology mastery. A needs analysis of prospective mathematics teacher students at Universitas PGRI Mpu Sindok found 70.31% lacked confidence designing technology-based instructional media and 67% were hesitant selecting appropriate technology. This study examined whether the Project-Based Learning in Co-Teaching with Practitioners (PjBLCTP) model improves digital competence relative to conventional Project-Based Learning (PjBL). Methods A quasi-experimental nonequivalent control group design was used with 64 fifth-semester mathematics education students (32 experimental, implementing PjBLCTP; 32 control, implementing PjBL). The same lecturer-practitioner team taught both classes, so treatments differed only in instructional syntax, isolating syntax rather than practitioner presence as the active ingredient. Digital competence was measured with a 33-item self-performance instrument based on the six-dimension DigComp 2.2 framework. Data were analyzed using covariance-based Structural Equation Modeling (maximum likelihood, semopy in Python) in three stages: item-level confirmatory factor analysis (CFA), a composite measurement model, and a structural model testing learning group and pretest digital competence as predictors of latent digital competence. Results Item-level CFA showed variable item validity; three items (mainly the problem-solving dimension) showed weak, non-significant loadings. The composite measurement model showed adequate convergent validity for latent Digital Competence in the pooled sample (Composite Reliability = .891; Average Variance Extracted = .576), although an exploratory group-split check did not confirm this factor structure as invariant within each condition separately. The structural model showed learning group significantly predicted latent digital competence (β = .999; p < .001), while pretest digital competence did not (p = .716); model fit was good (χ 2 (22) = 18.05; p = .703; CFI = 1.012; TLI = 1.017; RMSEA = .000). Conclusions PjBLCTP significantly improves prospective mathematics teachers’ digital competence relative to conventional PjBL, with instructional syntax rather than practitioner presence isolated as the active ingredient; however, this near-total effect coincided with complete separation of total posttest scores between groups and should be replicated in larger, more heterogeneous samples before being considered a stable estimate. The digital competence instrument, particularly its Problem Solving dimension, requires substantial refinement before further use.