Educational decentralization, learning, and inequality: A critical review of causal evidence.

Is decentralizing education effective? The question is deceptively underspecified because fiscal devolution to municipalities, transfer of school administration to provinces, school-level budget autonomy, parental participation, principal training, and locally hired teachers constitute different interventions with distinct causal estimands. This critical evidence synthesis appraises three influential research reviews and a purposively selected set of experimental and quasi-experimental studies, paying explicit attention to identification, treatment definition, statistical dependence, precision, implementation fidelity, institutional capacity, distributional effects, and external validity. A 2016 Campbell review reports a pooled improvement of $0.08$ standard deviations (95\% confidence interval $[0.02,0.13]$) in mathematics, with substantial heterogeneity ($I^2=69\%$); its randomized-study subgroup is smaller and statistically inconclusive ($0.05$, $[-0.03,0.14]$). A separate 2023 synthesis of principal-management interventions estimates $0.033$ standard deviations; that finding does not identify the effects of territorial decentralization. Randomized evidence from Indonesia, India, Niger, The Gambia, Kenya, and Madagascar suggests that management authority, information, grants, and community participation are neither interchangeable nor independently sufficient. Quasi-experimental results from Argentina and Chile highlight unequal gains and costs of jurisdictional fragmentation. For Bolivia, historical studies most directly support a change in the geographic and functional targeting of public investment after the 1994 reform; they do not establish a corresponding causal effect on standardized learning. We derive a multi-level theory of change, a structured study-by-study risk-of-bias audit, a transparent evidence-to-policy framework, and a feasible prospective evaluation design featuring pre-specified outcomes, equity constraints, and adversarial validity checks. The review is based on a purposive evidence corpus and does not claim exhaustive coverage or a newly estimated pooled effect.