Disentangling Age, Time, and Cohort Effects in Lifecycle Income: A Proxy Machine Learning Approach

Two key facts economists have emphasized about the lifecycle income process are rapid growth of inequality and the average level of income over the lifecycle. However, empirical estimates of these facts are plagued by the difficulty of distinguishing between age, time, and cohort trends. Foundational work on the lifecycle profile of income inequality assumes that “time effects” are zero (Deaton and Paxson, 1994). But this strong assumption is rejected by the data. We propose a new proxy variable machine learning approach to the age-time-cohort problem. For the variance of income over the lifecycle, our method yields both a significantly smaller slope of the age profile for the U.S., and much less erratic age profiles for 11 other countries. For the average level of income over the life cycle, our method yields a larger slope for the U.S., and confirms earlier estimates showing that the slope of the income profile increases with economic development.