This paper develops a new approach to testing hypotheses about the causes of discrimination in housing sales. We follow previous research by using data from fair housing audits, a matched-pair technique for comparing the traetment of equally qualified black and white home buyers. Our contribution is to shift the focus from differences in the treatment of teammates during an audit to agent decisions concerning an individual housing unit. Our sample consists of all units seen by either a black of a white auditor in the 1989 national Housing Discrimination Study. We estimate a multinomial logit model to explain a real estate agent's joint decisions concerning whether to show each unit to a white auditor and to a black auditor. We find evidence that real estate agents make and act upon inferences about a customer's preferences on the basis of the customer's initial inequity and that agents practice redlining, defined as the withholding of units in integrated neighborhoods. We find little evidence to support the conclusion that agents discriminate because of their own prejudice, but some evidence that they discriminate because of the prejudice of their white customes. More importantly, we find strong evidence of statistical discrimination; agents withhold houses from blacks when the probability of a successful transaction is perceived to be low.