Probability-based web surveys are increasingly challenged by decreasing response rates and high costs. A cheap and convenient solution is to use ‘opt-in’ online panels, which
are based on non-probability samples. However, the quality of the data such panels produce is subject to debate.
To improve our understanding in this regard, especially in the Swiss context, we compare conditional distributions of sociodemographic variables and voting behaviour of two probability-based web surveys and three opt-in panels. Indeed, point estimates in opt-in panels are well studied, but bivariate relationships between variables, arguably more important for researchers in political science research, have received less attention. Our analysis has the advantage of most variables of interest being included in the sampling frame and thus the true values are known for each conditional distribution. Our results show a lack of consistency and reproducibility in the results from opt-in panels, which leads us to recommend care when using this type of data.