Our proposed model provides a behaviorally plausible approach to examine the impact of return policies on consumers' purchase and return behavior.

In this dissertation, we study referral programs and preference estimation in two essays. In the first essay, we propose that a firm can enhance the effectiveness of its referral program by promoting better matching between referred customers and the firm. We develop three treatments aimed at promoting better matching, including (1) offering current customers a gift before inviting them to refer friends, (2) notifying current customers about the value that they have received from the firm before inviting them to refer friends, and (3) rewarding referring customers based on the value of their referred customers. We test these three treatments by conducting two field experiments in collaboration with a Chinese online financial services firm. We find that all three treatments substantially enhanced the effectiveness of the focal referral program, measured for each current customer as the total value of his referred customers. We also find that the enhancement was primarily driven by the acquisition of higher-value new customers rather than the acquisition of more new customers. In addition, we investigate customer heterogeneity in treatment effects and explore the mechanisms through which these treatments impacted customer referrals. In the second essay, we develop a new model for effective modeling of consumer heterogeneity in choice-based conjoint estimation. Assuming that most variations in consumers' partworth vectors are along a small number of orthogonal directions, we propose that shrinking the individual-level partworth vectors toward a low-dimensional affine subspace that is also inferred from data can be an effective approach to pooling information across consumers and modeling consumer heterogeneity. We develop a low-dimension learning model to implement this information pooling mechanism that builds on recent advances in rank minimization and machine learning. We evaluate the empirical performance of the low-dimension learning model using both simulation experiments and field choice-based conjoint data sets. We find that the low-dimension learning model overall outperforms multiple benchmark models in terms of both parameter recovery and predictive accuracy. While addressing two different marketing topics, both essays share a common theme--careful modeling of consumer heterogeneity plays a key role in understanding consumer behavior and developing effective marketing strategies.

A realistic specification for consumer heterogeneity should allow for both multiple market segments and within-segment heterogeneity. Such a flexible heterogeneity specification, which we term as multimodal consumer heterogeneity (MCH), raises considerable modeling challenge. We propose an innovative sparse learning approach for modeling MCH and apply it to conjoint analysis, where an adequate modeling of consumer heterogeneity is critical. The unique perspective of our approach is to characterize MCH via a special form of structured sparsity defined on conjoint partworths that can be recovered efficiently using recently developed optimization techniques. The proposed approach is intuitive and easy to implement in practice. We use extensive simulation experiments and two empirical conjoint data sets to demonstrate the performance of our sparse learning approach in modeling MCH and recovering both accurate individual-level partworths estimates and managerially relevant market segmentations.

The influence of metal interconnect resistance on the performance of vertical and lateral power MOSFETs is studied. Vertical MOSFETs in a D2PAK and DirectFET package, and lateral MOSFETs in power IC and flip chip are investigated as the case studies. The impact of various layout patterns and material properties on R[subscript DS(on)] will provide useful guidelines for practical vertical and lateral power MOSFETs design.

We consider a single-installation, single-product inventory management problem with positive fixed ordering costs. We allow the unfulfilled demand to be either backlogged or lost and the lead time to be either zero or positive, and assume that the stochastic demand is correlated and information about the demand distribution is limited. A simple implication of non-negligible fixed ordering costs is that any viable inventory control policy implicitly divides the planning horizon into a sequence of { em inventory cycles}, where an order is only placed at the beginning of the cycles. Building on this observation, we propose a new policy that manages inventory by directly selecting the order quantity and the cycle length for each cycle using robust optimization. Our cycle-based policy offers a simple and unified approach to solving the inventory problem, and we show that the policy decisions can be computed extremely efficiently by solving a small number of linear programs for all model instances (i.e., backlogging versus lost sales, zero lead time versus positive lead time). To the best of our knowledge, our policy is the first robust policy that can accommodate lost sales, and also the first linear programming-based robust policy that can handle positive fixed ordering costs. Results of extensive numerical experiments show that our policy outperforms several strong benchmark policies.