Research

Abstract:  This paper studies identification and estimation of spillovers in binary choice games with egocentric network data. Outcomes and covariates are observed for all agents, but links are observed only for dyads incident to surveyed egos. A tetrad argument identifies the dyadic component of link surplus, and I introduce an anchored-triad method to identify agent-specific degree heterogeneity. Together, these results identify conditional link probabilities. For estimation, I impose a parametric dyadic-logit specification, construct an estimated hybrid network, and use it in a nested pseudo-likelihood estimator of the game. I establish consistency and show that, in dense networks, network-estimation error is asymptotically negligible for root-n inference when the egocentric sample grows faster than root-n. Monte Carlo experiments and a microfinance application show that hybrid-network estimates closely track realized-network benchmarks.

Presentations: Network Science in Economics Conference 2026 (Poster), 2025 World Congress of the Econometric Society, Chinese Economists Society (CES) North America Annual Conference 2025, American Economic Association Annual Meeting 2025, Global GLO-JOPE Conference 2024 (Job Market Session), The North East Universities Development Consortium (NEUDC) 2024 Conference, Asian Meeting of the Econometric Society 2024,  North American Summer Meeting of the Econometric Society 2024 [Certificate], TEXPOP Conference 2024 by the Population Research Center of the University of Texas at Austin, Texas Camp Econometrics XXVII 2024, Southern Economic Association 93rd Annual Meeting 2023, Causal Data Science Meeting 2023, 18th Annual Economics Graduate Student Conference of Washington University in St. Louis,  33rd Annual Midwest Econometrics Group Conference,  SMU Brown Bag Seminar, 2nd Annual SMU Research Computing Day 2023 (Poster)

Abstract:  This paper studies identification and inference in dyadic network formation models with unobserved degree heterogeneity. In standard approaches, agent effects are treated as incidental parameters and are eliminated or profiled out when estimating homophily. I show that, in a dyadic logit model, selected triadic comparisons provide identifying information about the realized agent effects themselves. The key restriction compares two complementary three-node configurations: its log-odds ratio removes the effects of two nodes and leaves a single-node component together with observed dyadic covariates. Under support and smoothness conditions that permit diagonal comparisons in covariate space, this relation identifies realized node effects; conditional on these effects, dyadic link probabilities identify the homophily component on the relevant support. The result yields an estimator in which a triad-based preliminary step is combined with dyadic likelihood refinement and bias corrections for common parameters and node effects. Large-network asymptotic theory provides inference under dense-network regularity conditions. Monte Carlo simulations show that the estimator is competitive with tetrad and joint fixed-effect methods for homophily and improves numerical feasibility in designs with limited usable conditioning variation. Two applications illustrate the empirical content of recovering heterogeneity. In the 1996 network of economic integration agreements, geographic proximity and pair remoteness remain important dyadic determinants after controlling for economy-specific link propensities, while the recovered effects reveal substantial regional heterogeneity. In the Nyakatoke risk-sharing network, kinship, religion, and physical distance remain robust predictors of links, whereas the association with wealth similarity weakens once household-level heterogeneity is incorporated.

Presentation: 36th Annual Midwest Econometrics Group Conference 2026 (scheduled)

Abstract:  This paper studies how to divide a fixed budget between randomized egos and outcomes for their nominated contacts. A larger name quota measures each ego contrast more precisely, while a smaller quota funds more independent randomizations. The local solution is a square root rule that accounts for residual correlation and assignment induced variation. Network interference makes quota choice inseparable from ego placement because another name changes nomination overlap, exposure, covariance, and cost. I propose an exact directional constraint that allocates an isolated effect bias tolerance across the weak sources actually selected. Uniform distance and pairwise radius rules are conservative special cases. The causal target determines which restriction matters. Bernoulli assignment identifies the effect under the assignment policy, while an isolated effect requires a strong interaction adjustment and a weak interaction allowance. A fixed population claim also requires positive ego inclusion probabilities. Under primitive network weak dependence, a feasible variance estimator yields asymptotically valid policy inference and an honest isolated interval. A field procedure deducts network and identity costs, generates candidate quota and ego samples, and ranks them by full planning mean squared error. Simulations and two network applications show that ego placement can be as important as quota choice, while network data pay for themselves only when better placement compensates for the outcomes their collection displaces.

Abstract: We study the identification and estimation of social interactions in large endogenous networks. Our analysis focuses on a binary-action game of incomplete information in which each agent’s expected payoff relies on her observable and unobservable characteristics, the average of her beliefs about peers’ actions, and some preference shocks. We do not restrict the game to have a unique equilibrium, and hence multiple vectors of equilibrium beliefs can rationalize the binary actions. Although these interdependent beliefs cannot be solved, we achieve the identification of them and the payoff functions under mild conditions, since the average of peers’ actions reflects the average of beliefs of peers’ actions asymptotically. Endogeneity in networks results from agents’ unobservable characteristics that affect both network formation and expected payoffs. To address the endogeneity issue, we express the unobservable characteristics as some unknown function of observable variables and apply the control function approach. We employ the strategy to develop a semiparametric estimator and present the finite-sample performance through Monte Carlo simulations.

Presentations: Asian Meeting of the Econometric Society 2026 (presented by coauthor), Texas Camp Econometrics XXVIII, Southern Economic Association 94th Annual Meeting 2024, 34th Annual Midwest Econometrics Group Conference 2024

Abstract:  This paper studies risk improvement for multivariate estimators obtained from cross-fitted orthogonal scores. Such estimators are widely used to estimate event-study paths, subgroup treatment effects, and other vector-valued empirical objects. While Neyman orthogonality delivers first-order robustness to nuisance estimation error, it does not imply optimality under global quadratic loss. We show that, under local nuisance misspecification, the estimator admits a uniform asymptotic linear representation with a local drift term and is asymptotically equivalent to a Gaussian shift experiment. This representation implies that the usual estimator is generally second-order inefficient when the target dimension is at least three. We propose a feasible covariance-aware positive-part shrinkage estimator that shrinks the baseline estimate toward an economically motivated target. The estimator uses only the original estimate and its covariance matrix and uniformly improves asymptotic quadratic risk over bounded local parameter classes. The results justify a simple post-estimation shrinkage adjustment for noisy coefficient vectors without changing the estimand or the identifying assumptions.

Presentations: Southern Economic Association 96th Annual Meeting 2026 (scheduled)

Mobile money is well established across Sub-Saharan Africa, yet cash continues to dominate retail transactions. We hypothesize that this persistence reflects a two-sided coordination failure: merchants hesitate to adopt digital payments because consumers pay in cash, while consumers continue to pay in cash because merchants rarely accept digital payments. We test this hypothesis using a market-level randomized experiment across 79 markets in Tanzania. The intervention substantially shifted same- and cross-side coordination beliefs (0.22–0.40 SD) and increased upstream merchant use of digital payments (0.32 SD on a composite index of adoption behaviors). Consumers, however, updated their beliefs without changing their payment behavior, even when directly exposed to activated merchants. We argue that this asymmetry reflects a fundamental difference between merchant and consumer adoption. Whereas merchant adoption is primarily a coordination decision, consumers’ payment decisions depend not only on merchant acceptance but also on payment inventories and their relative costs. Addressing coordination failures is therefore necessary but not sufficient for the transition from cash to digital payments; it also requires changing the economics of consumer payment choices.

Presentations: IPA–GPRL Annual Researcher Gathering 2026, Northwestern University (presented by coauthor)

In many emerging economies, cash remains the primary medium of exchange despite widespread access to mobile phones and mobile money. Cash dependence is costly, leaving households and firms vulnerable to theft and loss, and limiting access to formal financial services and credit. One fundamental barrier to the transition from cash to digital payments is a coordination problem, in which merchants’ and consumers’ willingness to switch depends on expectations about adoption by the other side. In a field experiment across 79 markets in Tanzania, we study whether locally targeted, two-sided incentives and cross-sided information help correct misperceptions about adoption and lead to sustained increases in digital payment use. Using detailed merchant network data, we estimate within-market spillover effects and decompose the channels through which digital payment use diffuses across agents. We further explore the role of information and network structure in sustaining digital payment adoption.

Presentations: IPA–GPRL Annual Researcher Gathering 2026, Northwestern University

Lead role in randomization design, referral game design, and social networks data collection. Evaluating AVA (Automated Virtual Assistant) adoption, referral networks, and social diffusion of new technologies in economic empowerment and health contexts.