Existing criteria for craze initiation are reviewed, and their limitations are discussed. The most obvious problem is that they are formulated simply in terms of principal stresses, making no provision for the known effects of small inclusions and surface imperfections. To solve this problem, a new criterion is proposed, which is based on linear elastic fracture mechanics. Craze initiation is treated as a frustrated fracture process rather than a yield mechanism. Calculations show that the strain energy release rate, G(I)(nasc), required to generate a typical 20 nm thick nascent craze, is less than 1 J m(-2). This explains why flaws less than I mu m in length are capable of nucleating crazes at stresses of 20-30 MPa. Subsequent craze propagation is dependent upon two flow rates, one relating to fibril drawing at the craze wall and the other to shear yielding at the craze tip. Under biaxial stress, the second principal stress sigma(2) affects craze tip shear yielding but not fibril drawing. This model is used in conjunction with the von Mises yield criterion to derive a new expression for the crazing stress sigma(1)(craze), which provides a good fit to data on visible crazes obtained by Sternstein, Ongchin and Myers in biaxial tests on cast PMMA [Sternstein SS, Ongchin L, Silverman A. Appl Polym Symp 1968;7:175; Sternstein SS, Ongchin L. Polym Prepr Am Chem Soc Div Polym Chem 1969; 10: 1117; Sternstein SS, Myers FA. J Macromol Sci Phys 1973;B8:539]. (c) 2007 Elsevier Ltd. All rights reserved.