This dissertation studies the impact of data revisions on macroeconomic expectations. In the first chapter I show theoretically how revised data bias estimates of information frictions taken from surveys of expectations. The bias is due to the dynamic property of revised data which introduces additional past dependency into forecasts, breaking the informational symmetry between forecast errors and forecast updates. I propose a new estimation specification which corrects the bias and apply my method to data from the Survey of Professional Forecasters. I find the degree of information frictions is 33% different from existing estimates on average. Using my new estimates, I find equilibrium models of incomplete information more closely match the data. In the second chapter, I study socially optimal data provision when the economy resembles a beauty-contest. The budget-constrained statistical office (S.O.) faces a trade-off: provide imprecise yet timely data about the present, or wait to provide more precise yet delayed data about the past. I find the S.O.’s welfare-maximizing choice of current versus revised signal precisions depends on the accuracy of private information and strength of strategic incentives. In some cases the S.O. offers a mixture of both signals, while in others it provides only one or neither signal in order to reduce the crowding-out of private information caused by the coordination incentive. The S.O. treats the revised signal as an inferior good, reducing its precision in favor of the current signal when the budget limit rises. In the third chapter, I evaluate the interaction between revised data and higher-order dynamics when estimating information frictions. I prove analytically that the existing estimation method is valid when the fundamental possesses additional past dependency relative to the case in chapter one. I find additional bias terms in estimates of information frictions in the presence of revised data and higher-order dynamics. I propose a correction which eliminates the bias and does not depend on the specific time series properties of the data. I use simulation to study the nature of this new bias, finding it to be quantitatively small relative to the bias caused solely by revised data.