This paper introduces the least absolute deviations estimation of the contingent valuation model, which corresponds to the semi-parametric estimation of discrete choice models by Manski (1975, 1985) and Lee (1992). The least absolute deviations estimation is more robust to mis-specified distributional assumptions in the estimation of the contingent valuation model, compared to the maximum likelihood estimation. The full identification and strong consistency of the estimation are proved and its application to different formats of contingent valuation survey data is discussed. Simulation studies are designed to evaluate its operational characteristics including computational strategies, small sample properties and the efficiency gain of a follow-up question. The bias and efficiency of least absolute deviations and maximum likelihood estimation are compared in the presence of heteroskedasticity.