In this article, we use two formats of contingent valuation (CV) questions to elicit willingness to pay (WTP) for maintaining biodiversity in forests: the dichotomous choice (DC) or referendum format and the open-ended (OE) question. A large population of French households were surveyed nationwide by phone. The sample of respondents was later divided into two subsamples: people who have recreational activities in forests and those who do not. This dichotomy potentially biases WTP as the decision to have recreational activities in forests is endogenous. We estimate a Probit model with sample selection to correct this bias in the DC question. With the OE question, a second source of selection bias related to non-random censoring is present: some respondents are unwilling to pay. We use an extension of Heckman's approach to the double selection problem. The empirical application shows that ignoring these sample selection problems leads to biased estimates of mean WTP for biodiversity in a national survey for France.