Data sets of sea ice and snow thickness, acquired as drilling profiles and shipboard observations during three Antarctic expeditions in the Ross, Amundsen and Bellingshausen Seas are studied with respect to their statistical robustness towards weed-out, sectional, and random subsampling. For individual profiles of the drilling data the resulting subsampling errors, associated with the means and standard deviations of ice and snow thicknesses, and freeboard were generally smallest for the weed-out method, where the spacing between drilling points was increased, while the profile length was unchanged. On the cruise level, floe drilling data and shipboard observations (consisting of individual sets of 25 single observations) were subsampled by picking every nth floe (or observation set). Up to n=3, subsampling errors were negligible against the inherent errors of the full data sets. The effects of subsampling depend on the measured probability density functions (PDFs) of the full data sets whose statistical properties are studied in the section 2. It was found that lognormal or gamma distributions are better statistical representations for both drilling and shipboard data than normal distributions. A simple statistical ice thickness model is presented, which explains how different ice growth and development processes influence the shape of the ice thickness PDF. Appropriate choices of the parameters, consistent with cruise dates and observational evidence of ice deformation, give good representations of the measured ice thickness PDFs for the three cruises. ¿ 1999 American Geophysical Union