Factors that need to be considered in the analysis of time trends in disease incidence are age, year of diagnosis, and birth cohort. When these are included in a log-linear model, a nonidentifiability problem arises from the linear dependence among these three time factors so that only specified functions of the parameters can be unambiguously determined. One of these invariant functions is the drift or the sum of the period and cohort trend. Non-Hodgkin's lymphoma incidence rates from Connecticut for the period 1935-1989 were analyzed for males and females. In addition to an age effect, both period and cohort significantly improved the fit of the model. The estimated drift shows that there has been a 10.3% increase in risk every 5 years since 1965 for females and 9.2% for males. It is unlikely that a trend of this magnitude can be attributed entirely to data artifact.