Attribute non-attendance in choice experiments affects willingness-to-pay (WTP) estimates and therefore the validity of the method. A recent strand of literature uses attenuated estimates of marginal utilities of ignored attributes. Following this approach, we propose a generalisation of the mixed logit model, whereby the distribution of marginal utility coefficients of a stated non-attender has a potentially lower mean and lower variance than those of a stated attender. Model comparison shows that our shrinkage approach fits the data better and produces more reliable WTP estimates. We further find that while reliability of stated attribute non-attendance increases in successive choice experiments, it does not increase when respondents report having ignored the same attribute twice.