Bureau of the Census death registration records, as reported in Mortality Statistics, are a primary source for early twentieth-century U.S. homicide statistics. Those data appear to show a massive rise in homicide during the first decade of the century, with a continuing increase through 1933. This increase is quite at variance with the trend away from violence in other industrialized societies. During the first one-third of the century, however, death registration was incomplete; it occurred only in an expanding "registration area" that was composed, in the earlier years, primarily of states with typically low rates of homicide. Further, in the first decade of the century homicides within the registration area often were reported as accidental deaths. As a result, apparent increases in rates of homicide in the United States between 1900 and 1933 may be illusory. I use a two-step process to address these problems. Drawing on internal evidence and commentaries in early volumes of Mortality Statistics, I use GLS regression to estimate the prevalence of undercounts. Then I create a series of GLS models that use registration area data to estimate early twentieth-century national rates. These estimates call into question the extent of homicide change early in the century.