A survey analyst remarked: "When covariance analysis is used with survey data, there is a danger that the treatments may be related to the concomitant variable." What is the nature of the problem? Does this same problem exist when the treatments are randomly assigned to the experimental units?