The rational addiction model is usually tested by estimating a linear second-order difference Euler equation, which may produce unreliable estimates. We show that a linear first-order difference equation is a better alternative. This empirical specification is appropriate under the reasonable assumption that people are uncertain about the time of their death, it is based on the same structural assumptions used in the literature, and it retains all policy implications of the deterministic rational addiction model. It is also empirically convenient because it is simple, it allows using efficient estimation strategies that do not require instrumental variables, and it is robust to the possible non-stationarity of the data. As an application we estimate the demand for smoking in the US from 1970 to 2016, and we show that it is consistent with the rational addiction model.