|
Abstract:
|
Managers often face the problem of limited data at the individual customer level. A common practice is to augment the limited available customer level data with averages for the group to which the customer belongs. We demonstrate using a target selection problem that this standard practice of using group (zip code) averages as a proxy for individual information leads to biased inference and erroneous managerial decisions. We therefore propose that firms use "conditional averages," i.e., rather than use the raw averages for the group, use averages conditional on the available individual information in the firm's internal databases. However, this is hard to implement in practice because group level joint distributions are unavailable. We develop a flexible and scalable approach to obtain group level joint distributions.
|
