A structural learning model of retail product selection with information spillovers

My research examines how a manager in a chain of grocery stores decides whether to adopt a new product when information about the product's uncertain demand can be shared across stores. When there exists the possibility of learning from others, the manager faces a tradeoff between adopting the new product right away and waiting to see how well the new product sells in other stores. To capture this tradeoff, I specify a dynamic structural model of Bayesian learning in which each manager learns the unobserved quality of new products by observing sales in his own store as well as sales in other stores throughout the store chain. With these signals of the unobserved quality, the manager updates his beliefs and makes a stocking decision to maximize expected lifetime profits. My model differs from most previous empirical learning models in the literature in two ways. First, my model includes both sides of the market: a dynamic firm side and a static consumer side. Most previous research focused on one or the other. By modeling both sides, I can recover signals of the unobserved quality that are structural in nature. Second, my model allows for informational spillovers and learning from others, a topic widely explored in the theoretical literature but on which little empirical work exists. There are no papers that I am aware of that include both of these features.;I estimate the model with data on fruit juices from a grocery store chain. First, I estimate the demand side of the model by GMM. After recovering signals of unobserved quality, I estimate the firm side of the model by maximum likelihood. I recover reasonable estimates for parameters not associated with learning. Estimates of learning parameters also seem reasonable and indicate that a modest amount of learning occurred. Moreover, I find that managers relied on the average signal across stores, indicating that there was information sharing. The learning parameter estimates are not statistically significant however. I conducted two policy experiments. I find that decreasing the uncertainty over unobserved quality or the amount of information sharing has little effect on adoption rates.