Taxing the invisible: graph-based information structures for redistributive design under uncertainty

What the paper says

Purpose This study examines how equitable taxation can be designed under conditions of severe informational scarcity, where standard assumptions regarding income observability and the reliable estimation of behavioral elasticities do not hold. It proposes a structure-based approach to redistribution, in which observable economic networks provide indirect but policy-relevant information for fiscal design. Design/methodology/approach This paper integrates graph neural networks with distributionally robust optimization to infer income-relevant structural embeddings from observable non-monetary characteristics such as education, occupation, region and formality status. Individuals are represented as nodes within an economic network, where links capture economic similarity and opportunity structure. Tax rules are optimized over Wasserstein ambiguity sets to explicitly account for income uncertainty and limited observability. The framework is evaluated through simulations of two stylized economies – a cohesive network and a fragmented network – using empirically calibrated synthetic data grounded in household surveys and administrative aggregates. Findings Tax rules based on GNN-derived structural embeddings consistently outperform benchmark scoring rules that rely solely on observable proxies. The proposed approach achieves larger reductions in post-tax inequality and lower regressivity while exhibiting greater robustness to income ambiguity and network perturbations. Redistributive performance improves with higher structural connectivity, underscoring the role of informational topology as a determinant of fiscal capacity. Research limitations/implications The analysis abstracts from strategic tax evasion, labor supply responses and general equilibrium effects. Instead, it focuses on the informational foundations of redistributive capacity under partial observability. Practical implications The framework provides tax administrations with a method to transform existing non-monetary registries and fragmented administrative data into actionable fiscal information without relying on intrusive income monitoring. Social implications By enhancing structural observability, the proposed approach has the potential to strengthen fairness, transparency and trust in fiscal governance, particularly in economies characterized by high informality and limited administrative capacity. Originality/value This study contributes to the literature on optimal taxation by demonstrating that progressive redistribution remains feasible under severe informational constraints when the economic structure serves as a substitute for direct income observability. By combining network-based representation learning with distributionally robust fiscal optimization, it introduces a transparent and policy-relevant framework for tax design in data-poor environments.