Construct validity is essentially the degree to which our scales, metrics and instruments actually measure the properties they are supposed to measure. Although construct validity is widely considered an important quality criterion for most empirical research, many software engineering studies simply assume that proposed measures are valid and make no attempt to assess construct validity. Researchers may ignore construct validity because evaluating it is intrinsically difficult, or due to lack of specific guidance for addressing it. In any case, some research inevitably produces erroneous conclusions, because due to invalid measures. This article therefore attempts to address these problems by explaining the theoretical basis of construct validity, presenting a framework for understanding it, and developing specific guidelines for assessing it. The paper draws on a detailed example involving 15 software metrics, which ostensibly measure the size, coupling and cohesion of Java classes.