Parallel strand lumber (PSL), a type of structural composite lumber, is composed of wood strands that have cross section dimensions on the order of 5mm by 25 mm and length of order tens of centimeters. These strands are formed, using pressure, adhesive, and microwave curing, into structural members that have cross section dimensions up to approximately 40cm and the material properties of such members have a high degree of heterogeneity. In this paper we describe measurements of the void structure of PSL using a serial sectioning approach and the stochastic characterization of the ensemble of voids present in a typical PSL member. We have developed a suite of probabilistic models for the void structure, each of which is emphasizes certain characteristics of the void structure such as volume fraction, void aspect ratio, or void size distribution. Using the measured mesostructures and synthetic mesostructures generated from the probabilistic we use nonlinear finite element analysis to evaluate whether the probabilistic models can generate void structures that not only match the statistics of the measured void structures but also mimic the stress-strain and strength response of the material. Although the models do require a certain degree of calibration through the fitting of free parameters, they are able to reproduce the response of PSL to mechanical loads with good accuracy.