Social scientists regularly rely on population estimates when studying change in small areas over time. Census tract data in the United States are a prime example, since there are substantial shifts in tract boundaries from decade to decade. This study compares alternative estimates of the 2000 population living within 2010 tract boundaries to the Census Bureau's own re-tabulation. All methods of estimation are subject to error; this is the first study to directly quantify the error in alternative interpolation methods for U.S. census tracts. A simple areal weighting method closely approximates the estimates provided by one standard source (the Neighborhood Change Data Base or NCDB), with some improvement provided by considering only area not covered by water. More information is used by the Longitudinal Tract Data Base (LTDB), which relies on a combination of areal and population interpolation as well as ancillary data about water-covered areas. Another set of estimates provided by NHGIS uses data about land cover in 2001 and the current road network and distribution of population and housing units at the block level. Areal weighting alone results in a large error in a substantial share of tracts that were divided in complex ways. The LTDB and NHGIS perform much better in all situations, but are subject to some error when boundaries of both tracts and their component blocks are redrawn. Users of harmonized tract data should be watchful for potential problems in either of these data sources.