Data collection for economic valuation by using Internet surveys and pre-recruited Internet panels can be associated with severe disadvantages. Problems concerning sample coverage and sample representativeness can be expected. Representation errors may occur since people can choose whether to be part of an Internet panel and subsequently whether they wish to participate in the survey, thereby introducing two elements of potential self-selection. These elements may be correlated with preferences, thus making the respondents a non-random and non-representative sample, ultimately biasing results. This paper analyses a sample used for an Internet contingent valuation method survey eliciting preferences for improvements in water quality of a river. We find that some variables that affect the survey participation decision also affect willingness-to-pay, consequently biasing our welfare estimates. We show how adjusting willingness-to-pay for this bias can be accomplished by using a grouped data model incorporating a correlation parameter to account for selection.