New empirical models of consumer demand that incorporate social preferences, observational learning, word-of-mouth or network effects have the feature that the adoption of others in the reference group - the installed-base - has a causal effect on current adoption behavior. Estimation of such causal installed-base effects is challenging due to the potential for spurious correlation between the adoption of agents, arising from endogenous assortive matching into social groups (or homophily) and from the existence of unobservables across agents that are correlated. In the absence of experimental variation, the preferred solution is to control for these using a rich specification of fixed-effects, which is feasible with panel data. We show that fixed-effects estimators of this sort are inconsistent in the presence of installed-base effects; in our simulations, random-effects specifications perform even worse. Our analysis reveals the tension faced by the applied empiricist in this area: a rich control for unobservables increases the credibility of the reported causal effects, but the incorporation of these controls introduces biases of a new kind in this class of models. We present two solutions: an instrumental variable approach, and a new bias-correction approach, both of which deliver consistent estimates of causal installed-base effects. The bias-correction approach is tractable in this context because we are able to exploit the structure of the problem to solve analytically for the asymptotic bias of the installed-base estimator, and to incorporate it into the estimation routine. Our approach has implications for the measurement of social effects using non-experimental data, and for measuring marketing-mix effects in the presence of state-dependence in demand, more generally. Our empirical application to the adoption of the Toyota Prius Hybrid in California reveals evidence for social influence in diffusion, and demonstrates the importance of incorporating proper controls for the biases we identify.