Despite the great importance of plant–insect interactions to the functioning of terrestrial ecosystems, many temporal gaps exist in our knowledge of insect herbivory in deep time. Subsampling of fossil leaves, and subsequent extrapolation of results to the entire flora from which they came, is practiced inconsistently and according to inconsistent, often arbitrary criteria. Here we compare herbivory data from three exhaustively sampled fossil floras to establish guidelines for subsampling in future studies. The impact of various subsampling routines is evaluated for three of the most common metrics of insect herbivory: damage type diversity, nonmetric multidimensional scaling, and the herbivory index. The findings presented here suggest that a minimum fragment size threshold of 1 cm2 always yields accurate results and that a higher threshold of 2 cm2 should yield accurate results for plant hosts that are not polyphyletic form taxa. Due to the structural variability of the plant hosts examined here, no other a priori subsampling strategy yields consistently accurate results. The best approach may be a sequential sampling routine in which sampling continues until the 100 most recently sampled leaves have caused no change to the mean value or confidence interval for damage type diversity and have caused minimal or no change to the herbivory index. For nonmetric multidimensional scaling, at least 1000 cm2 of leaf surface area should be examined and prediction intervals should be generated to verify the relative positions of all points. Future studies should evaluate the impact of subsampling routines on floras that are collected based on different criteria, such as angiosperm floras for which the only specimens collected are those that are at least 50 % complete.
