Mobile money networks with tax-incentives

In the developing world, mobile money (MM) is viewed as a key tool for financial inclusion, enabling deposits, withdrawals, transfers, and payments via mobile phones. Despite private-sector successes (e.g., M-PESA in Kenya), it is unclear whether government-initiated MM programs can achieve similar adoption and usage. Using the first comprehensive dataset of a government-run MM project (Ecuador, January 2015–December 2017), the study examines user behavior and responses to tax incentives designed to promote non-cash transactions. Ecuador’s Central Bank launched a government-operated, centrally managed e-money system (a form of CBDC) to alleviate liquidity constraints in a dollarized economy and extend financial services to the unbanked (about 60% of the population). During the project, tax incentives—refunds of 1–2% of VAT for non-cash payments—were introduced to encourage adoption. The research questions include: how did agents behave over time in this MM network, how did they respond to tax incentives, and did these incentives foster sustained adoption, increased economic activity, or greater interconnectedness? The study uses temporal network analysis and regression to quantify immediate and longer-term impacts, and compares outcomes to broader economic objectives. It finds that incentives increased activity among continuing users but had modest impact on network interconnectedness, diffusion, and long-term adoption, and came at a high fiscal cost.

The literature on MM highlights impacts on remittances, financial inclusion, and risk sharing (e.g., Jack and Suri 2011, 2014; Donovan 2012). Success factors vary by context, implementation characteristics, and the role of macro-agents (Suri 2017; Lal & Sachdev 2015). Network effects are critical for adoption; Fafchamps et al. (2016) and Murendo et al. (2018) document social network externalities influencing MM usage. Prior work often relies on surveys; few use actual transaction network data. This study addresses that gap by leveraging complete transaction records and network analysis to observe system-level structure and diffusion (contrasting with Rwanda/Kenya cases). The paper also situates Ecuador’s state-run monopoly model within broader debates about implementer credibility, business models, and agent networks (Camner et al. 2009; Heyer & Mas 2011), noting that macro-agent quality and competition can shape demand (Balasubramanian & Drake 2015). Diffusion theories (e.g., complex contagion; Centola & Macy 2007; threshold models; Valente 1996; Granovetter 1977, 1983) suggest adoption requires clustered exposure and wide bridges, which the observed MM transaction network in Ecuador lacked (low clustering, hub-centric ties).

Data and setting: The Central Bank of Ecuador (CBE) provided a complete, de-identified dataset of all MM activities from January 2015 to December 2017, including account activations, balance checks, deposits, ATM withdrawals, P2P transfers, payments, incentives, and accounting movements. Agents are categorized as users (natural persons), companies (legal entities), and macro-agents (authorized distributors with multiple service points, including public/private institutions and financial entities). All transactions between phones used SMS; the CBE covered telecom operating costs. Usage charges were set by regulation. Government tax-incentives: The Organic Law for Equilibrium in Public Finances (OLEPF; 04/29/2016) introduced VAT refunds: 2% for e-money payments and 1% for debit/credit card transactions (VAT was 12%). The Organic Law of Solidarity (OLSRRAZE; 05/20/2016) raised VAT to 14% for one year but maintained the 2% e-money refund. In late 2017, the Organic Law for Reactivation (OLRE) ended the MM project and transferred administration to the private system, with accounts to be zeroed by March 2018. Network construction: The authors build four networks from the transaction logs: (1) Transaction Network (payments/charges for goods/services), (2) Cash-in Exchange Network (loading e-money with cash), (3) Cash-out Exchange Network (withdrawing cash from MM), and (4) Incentives Network (government VAT refunds into accounts). Each is represented as a multi-graph (directed edge per transaction, annotated with amount, date, type) and a simple-graph (edges aggregated per pair, summing values). Temporal analysis: Time is partitioned into 30-day spans. Time span 0 centers on the legislative interventions (04/25/2016–05/24/2016). Spans −1, −2, … are before; 1, 2, … are after. For each span, graphs are constructed and isolated nodes (no edges) are removed to analyze active agents. Structural and activity metrics are computed per span, including actor and transaction counts, mean transactions per actor, mean transaction value (per transaction and per actor), mean number of partners, local clustering coefficient, degree/type assortativity, Louvain modularity and community counts, and exchange-specific metrics (cash-in/out actor and transaction counts, mean values per transaction and per actor). Econometric evaluation: To quantify policy effects, the study shifts to larger windows (90-day and 150-day windows sliding by 30 days) to reduce zeros and increase observations per agent. Agents are labeled Early (any transaction before OLEPF), Late (any after OLSRRAZE), Continuing (intersection of Early and Late), and Total (union). For fixed sets of nodes across time, the following model is estimated with Newey–West robust standard errors to address serial correlation and heteroskedasticity: Y_it = α + β·t + γ·After_t + δ·(After_t × t) + ε_it, where Y_it is one of five outcome metrics per node i in span t: number of transactions, mean transaction value, total value of transactions, number of partners (degree), and local clustering coefficient (transitivity). β captures pre-policy trend; γ captures immediate level shift at policy; δ captures change in post-policy trend. Regressions are run separately for continuing users, companies, and macro-agents (main text reports 90-day windows; 30- and 150-day results are in Supplementary Information). A complementary regression collapses metrics across all continuing agents per span (N=33 spans) to assess aggregate effects.

• Incentives increased activity among continuing users immediately but did not sustain growth over time; effects on interconnectedness were marginal, and there were no significant effects for companies or macro-agents.
• Quantified effects for continuing users (90-day spans): • Number of transactions: +4.67 per user at policy (≈+130%); no significant change in trend thereafter. • Mean transaction value: +$4 at policy (≈+62%); significant negative post-policy trend (−0.336 per 90-day span), dissipating gains by project end. • Total value transacted: +$120 per user per 90 days at policy (>+200%); post-policy trend negative but not statistically significant (near flat or declining). • Number of partners: +1.56 at policy (≈+114%); no significant change in growth over time (partners did not continue to expand). • Local clustering coefficient: +0.013 at policy (≈+34% relative to baseline), with a significant negative post-policy trend—remaining low overall, indicating few user-to-user triangles.
• Network structure and diffusion: • Mean number of partners across agents hovered around ~2.4 before and after incentives; incentives did not expand local connectivity meaningfully. • Mean per-transaction value after incentives ≈$11.3; with 5.2 transactions per active agent per 30 days, mean dollars exchanged per active agent per span ≈$58.6—far below typical monthly consumption ($700 in 2017), indicating MM did not become a primary payment method. • Clustering decreased over time; degree assortativity became negative as low-degree users connected primarily to high-degree hubs (companies, macro-agents), not to each other—limiting peer-to-peer diffusion. • Louvain modularity stabilized high (~0.81) with many small communities (≈852 per month post-policy), reflecting numerous dyads/small clusters rather than cohesive MM user communities.
• Exchange behavior (cash-in/out): • Average per 30-day span after incentives: ~2,657 agents cashing in vs ~13,070 cashing out (peaking at 24,682 cashing out), indicating strong withdrawal orientation. • Mean values: cash-in ≈$64 per transaction with ≈4.3 transactions per cash-in actor (≈$276 per actor per span, growing); cash-out ≈$52 per transaction with ≈2.4 transactions per cash-out actor (≈$128 per actor per span). Conclusion section summarizes cash-in ≈$273 and cash-out ≈$140 per month, consistent in magnitude. • Spikes corresponded to policy enactment and project termination (users cashing out remaining balances).
• Incentives network and fiscal cost: • Up to 164,441 accounts per span received incentives (≈62% of 265,240 agents in the incentives network); mean incentive values rose over time consistent with 1–2% VAT refunds. • Government paid ≈$16 million in refunds to induce ≈$8 million in primary MM transactions after the policy—indicating high cost per dollar of transacted value.
• Adoption/usage penetration: • Maximum of ~22,106 agents conducting real transactions in a span (~0.25% of economically active population of ~8 million in 2016). • By Dec 2017: 402,515 MM accounts existed; only 41,966 (10.43%) used for purchases/payments; 76,105 (18.91%) deposited/withdrew only; 284,444 (70.67%) never used—consistent with incentive-driven account activation without sustained transactional use.
• No significant effects detected for companies or macro-agents across economic or network metrics, consistent with user-targeted incentives and small samples for these agent types.

The findings address whether tax incentives can drive adoption, network expansion, and sustained usage in a government-run MM system. Incentives triggered immediate increases in transaction counts and values among existing (continuing) users but did not catalyze longer-term growth or deeper network integration. The number of partners rose slightly at policy but did not continue to expand, and clustering remained low, indicating that users largely transacted with preexisting hubs (macro-agents, companies) rather than forming peer-to-peer ties. Consequently, diffusion was limited and MM did not become a primary payment instrument for daily consumption. Structurally, the transaction network exhibited heavy-tailed degree distributions centered on hubs with low clustering—conditions favorable for information spread but not necessarily for behavioral adoption, which often requires complex contagion and wide bridges (multiple reinforcing ties). The observed lack of user-to-user triangles and stable low average partners (≈2.4) are inconsistent with strong network effects for adoption. Policy relevance: Incentives that broadly reward non-cash payments (including banked card users) distorted behavior toward capturing refunds and cashing out rather than building MM-based commerce, particularly in a dollarized context facing liquidity constraints. The fiscal cost was high relative to induced MM transaction value. Targeting incentives to actors with high centrality (macro-agents, merchants) or to unbanked segments, aligning with agent network quality, and integrating trusted implementers could better foster diffusion. Contextual factors (implementer credibility, exclusion of mobile operators, macro-agent network management, and public trust in the Central Bank given historical crises) likely contributed to low penetration. The results underscore that the design of incentives should account for heterogeneous agent roles, network evolution, and country-specific institutional features to avoid incentive gaming and ensure durable adoption.

The study contributes the first analysis of a complete national MM transaction dataset for a government-run system, quantifying how tax incentives affected users’ economic activity and network structure over time. Incentives produced immediate increases among continuing users in number of transactions (+4.67 per 90 days), mean value (+$4), and total value (+$120), and modest increases in partners (+1.56) and clustering (+0.013). However, these effects did not translate into sustained growth, broader interconnectedness, or significant network diffusion. Most activity remained hub-centric (users to macro-agents/companies), MM’s role in monthly expenditures was small ($58.6 per active agent per 30 days), and cash-outs dominated. The policy imposed a high fiscal cost (≈$16M refunds versus ≈$8M in MM purchases). No significant effects were found for companies or macro-agents. Implications: Effective MM implementation requires incentive designs that promote network growth and peer interconnections, target key intermediaries (macro-agents/merchants) and unbanked users, and consider trust and institutional credibility. Future research directions include: evaluating alternative, targeted incentive schemes (e.g., merchant or macro-agent subsidies, network-wide referral/cluster incentives), integrating mobile operators and agent network quality metrics, modeling complex contagion thresholds in payment adoption, and leveraging linked datasets to observe broader economic/social networks beyond MM to better identify diffusion pathways.

• Absence of a formal control group and reliance on before/after designs with assumed linear pre-policy trends may limit causal inference.
• Data are confined to MM transactions; no observation of agents’ broader social/economic networks or alternative payment behaviors, precluding direct measurement of complex contagion or substitution effects.
• Limited agent attributes (beyond type) restrict heterogeneity analyses (e.g., income, geography).
• Small sample sizes for companies and macro-agents reduce power to detect effects for these groups.
• Windowing choices (30-, 90-, 150-day) and removal of inactive nodes per span in temporal graphs may affect metric levels (though robustness checks suggest consistent patterns).
• External validity is constrained by Ecuador’s specific institutional context (government-run CBDC-like system, dollarization, public trust dynamics).