This paper tests the validity of Present Value (PV) models of stock prices by employing a two-step strategy for testing the null hypothesis of no cointegration against alternatives which are fractionally cointegrated. Monte Carlo simulations are conducted to evaluate the power and size properties of this test, which is shown to outperform existing ones, and to compute appropriate critical values for finite samples. It is found that stock prices and dividends are both I(1) nonstationary series, but they are fractionally cointegrated. This implies that, although there exists a long-run relationship which is consistent with PV models, the equilibrium errors exhibit slow mean reversion. As the error correction term possesses long memory, deviations from equilibrium are highly persistent.