We use Hubbell's neutral theory to predict the impact of habitat fragmentation on Amazonian tree communities. For forest fragments isolated for about two decades, we generate neutral predictions for local species extinction, changes in species composition within fragments, and increases in the probability that any two trees within a fragment are conspecific. We tested these predictions using fragment and intact forest data from the Biological Dynamics of Forest Fragments Project in central Amazonia. To simulate complete demographic isolation, we excluded immigrants--species absent from a fragment or intact forest plot in the initial census but present in its last census--from our tests. The neutral theory correctly predicted the rate of species extinction from different plots as a function of the diversity and mortality rate of trees in each plot. However, the rate of change in species composition was much faster than predicted in fragments, indicating that different tree species respond differently to environmental changes. This violates the key assumption of neutral theory. When immigrants were included in our calculations, they increased the disparity between predicted and observed changes in fragments. Overall, neutral theory accurately predicted the pace of local extinctions in fragments but consistently underestimated changes in species composition.