As managers strive to keep up with the ongoing “Data Revolution,” data-driven decision-making (DDD) is the new best-practice, rapidly diffusing across businesses. Yet reliable estimates of the impact of DDD on business outcomes are few, and available only for large publicly traded firms. We derive the first experiment-based estimates of the impact of data-driven decision-making on small businesses, analyzing the randomization-based gradual introduction of eBay’s Seller Hub (SH) tool, a data rich seller dashboard. We demonstrate that SH adoption is associated with increased DDD and find that access to SH increases e-retailers’ sales by 3% on average. The benefits of SH access increase with sellers’ intensity of performance monitoring, with 5.1% sales increases for adopters with monitoring levels in the top 50%, and 7.3% for adopters with top-quartile monitoring levels. In addition, an aptitude for analytics leads to larger increases in DDD for SH adopters. Policies to support small businesses’ transition to the data era should therefore address DDD gaps by both ensuring access to tools and improving managerial practices and analytical skills.

Firms often decide whether to adopt an innovation of uncertain value in markets where the outcomes of earlier adopters are observed. This paper introduces a flexible Bayesian model suitable for the analysis of social learning, competition, and diffusion in such environments. Agents in the model have (potentially misspecified) theories of how others’ profits relate to their own, and use these to make their adoption decisions. When adopting, agents steal business from and inform others. I estimate the model exploiting a unique reform in Illinois that legalized slot machines, and empirically study how information and adoption diffuse through a network. This setting is well-suited for such analysis, since gambling data are publicly available, adoption is a discrete action, and the set of potential adopters (liquor license holders) is defined by law. I find that establishments that observe more adoption or higher neighbors’ profits are more likely to adopt themselves, yet learning could improve since they do not use all the relevant information. Establishments have diffuse priors and they learn from more neighbors than they compete with. The direction and extent to which learning affects adoption are ex-ante ambiguous. In two counterfactual exercises I show that increasing information availability or learning substantially increases both adoption and total profits in the market. 

Agents learn about a changing state using private signals and past actions of neighbors in a network. When can they learn efficiently about recent changes? We find two conditions are sufficient: (i) each individual's neighbors have sufficiently diverse types of private information; (ii) agents are sophisticated enough to use this diversity to filter out outdated information and identify recent developments. If either condition fails, learning can be bounded far from efficient levels—even in networks where, with a fixed state, learning is guaranteed to be efficient without (i) or (ii). We thus identify learning externalities that are distinctive to a dynamic environment, and argue that they can be quite severe. The model we develop to make our argument provides a Bayesian foundation for DeGroot learning in networks, permitting new counterfactual and welfare analyses for that commonly-used behavioral model.

We develop a structural model of dynamic network formation and apply the model to data on academic collaborations. In the model, new connections form randomly with probabilities depending on the existing network, individual characteristics, and geographic and topical distances. Using our model, we estimate parameters determining the importance of each of these factors on link formation using data on formal and informal collaboration in Financial Economics. Our theoretical analysis shows that with high probability, the network will be `close' to a steady state in the sense that researcher's centralities are close to their steady state values. Understanding these steady states lets us determine which researchers will be most central, and thus most productive, in the networks that would form under counterfactual policies (e.g. cutting all costs of collaborating at a distance in half).

The size distribution of firms is an important indicator of market concentration. Studying the distribution of sales for eBay’s commercial sellers, this paper provides the first analysis of a firm size distribution (FSD) at an online market, further examining its evolution, and the underlying growth rate patterns. The evolution of eBay’s FSD is characterized by increasing mass in its right tail, yet it remains better-fitted by a lognormal than by a power law distribution throughout our ten-year period of analysis. This is in line with a possible convergence towards a power law in the long run, while currently market concentration is lower compared to power law FSDs found in many traditional industries. Smaller firms are the fastest growing group of businesses, followed by large firms, while mid-sized firms exhibit the slowest growth rates. Sellers’ growth rate patterns obey Gibrat’s Law only among the subset of top selling firms.