Abstract
We study optimal income taxation when workers’ productivity is stochastic and evolves endogenously because of learning by doing. Learning by doing calls for higher wedges and alters the relation between wedges and tax rates. In a calibrated model, we find that reforming the US tax code brings significant welfare gains and that a simple tax code invariant to past incomes is approximately optimal. We isolate the role of learning by doing by comparing the aforementioned tax code to its counterpart in an economy that is identical to the calibrated one except for the exogeneity of the productivity process. Ignoring learning by doing calls for fundamentally different proposals.