Publications

Individuals update their beliefs and respond to new information in idiosyncratic ways. I show that an individual’s idiosyncrasies such as under-reaction, over-reaction, or frustration can have spillover effects and adversely affect the long run beliefs of society. I derive sufficient conditions for convergence of beliefs for all possible networks of agents with heterogeneous idiosyncrasies. Beliefs converge if the magnitude of over-reaction and frustration in any agent's network neighbourhood is below a threshold determined by how much they trust their own private signals. I find that the absence of disproportionately influential agents is not sufficient to ensure the accuracy of long-run beliefs if learning idiosyncrasies also grow with the network. Finally, I show that agent under-reaction can partition the network, create bottlenecks, and delay convergence. Simulations on artificial and Indian village networks validate the results.

Example of convergence and divergence of individual beliefs in a randomly generated network. Agents  are randomly endowed with learning idiosyncrasies from low/high distributions and update beliefs after observing each others' choices. The colors indicate heterogeneity in choices as beliefs are updated.

Working Papers

2023 Econ Job Market Best Paper Award by the UniCredit Foundation and European Economic Association.

Coverage: VoxDev,  Ideas of India Podcast at the Mercatus Centre. Spotify Link. 

Inaccurate beliefs about social norms can reduce useful social interactions and adversely affect an individual’s ability to deal with negative shocks. We implement a randomized controlled trial with low-income workers in urban India who lack access to formal financial and healthcare support. We find that the majority of individuals underestimate their community's willingness to engage in dialogue around financial and mental health concerns. Belief correction leads to a large increase in the demand for network-based assistance. We show that the effects are driven by a reduction in the perceived costs of violating social norms arising due to concerns around reputation and insensitivity. We structurally estimate a network diffusion model and predict that our belief correction intervention will not lead to a shift in equilibrium engagement. Consistent with this, 2 years later, we find that the average beliefs of those exposed to the intervention are significantly more optimistic but still lower than the information delivered in the experiment. We compute the strength of counterfactual interventions needed to generate a sustained effect and find that belief correction can be used to generate both the demand and funding for such policies.

Presentations: (2024) Workshop on Networks and Development, Naples; CSAE Conference, Oxford; PacDev, Stanford. (2023) European Winter Meeting of the Econometric Society, Manchester; NEUDC, Harvard; CEPR-TCD TIME Conference on Economic Development, Trinity College Dublin; OxDev, Oxford ; CSAE Research Workshop, Oxford; Informal Presentation at the King Center of Global Development, Stanford; Presentation to the Communities Project Team, Stanford; Interface of Theory and Experiments at the Economics Research Jamboree, Oxford. 


Funded by JPAL Post Primary Education Grant (see JPAL Project Page), and RISE.  

Coverage: VoxDev

We study the effect of communicating student-specific teacher expectations on academic performance. We randomize whether students (a) receive high-performance expectations, (b) are additionally paired with a classmate for encouragement, (c) receive information about past performance, or (d) receive no message. Expectations increase math scores by 0.19 standard deviations, with especially large effects among students who randomly received ambitious expectations and were predicted to perform poorly. Information provision has comparably large effects (0.16 standard deviations), particularly in schools with low parental literacy. However, pairing students only improves scores when peers have similar characteristics. Our findings highlight low-cost, sustainable ways of leveraging teachers to improve performance. 


The spatial configuration of domestic supply chains plays a crucial role in the transmission of shocks. This paper investigates the representativeness of formal firm-to-firm trade data in capturing overall domestic trade patterns in Kenya — a context with a high prevalence of informal economic activity. We first document a series of stylized facts and show that informal economic activity is not randomly distributed across space and sectors, with a higher incidence of informality in downstream sectors and smaller regional markets. We then link granular transaction-level data on formal firms with data on informal economic activity to estimate a structural model and predict a counterfactual network that accounts for informal firms. We find that formal sector data overstates the spatial concentration of aggregate trade flows and under accounts for trade within regions and across regions with stronger social ties. Additionally, the higher the informality in a sector and region, the more we underestimate its vulnerability to domestic output shocks and overestimate its vulnerability to import shocks.


How can social networks be strategically leveraged to improve economic outcomes? We implement a randomized controlled trial in rural Nepal in which we vary whether women attend an entrepreneurship training program alone or with a randomly chosen peer. While the training significantly improves short-term outcomes, pairing matters only when the individual is paired with a socially close peer who is more connected than them. One year later, we find that while treated individuals have taken steps to open a business, those treated in pairs instead expand their agricultural activities, with higher effects concentrated among those paired with a socially close peer who is less connected than them. We show that these differential peer effects over time can arise due to motivation from central peers in the short term and ease of collaboration with less central peers in the longer term. Finally, we leverage data from the two survey waves to identify peer effects and implement counterfactual exercises to compare the effect of strategic versus random interactions. We find that strategic pairing can increase average outcomes by 0.8-0.9 standard deviations in the short and long run respectively.

Work in Progress