Comparing the Accuracy of Univariate, Bivariate, and Multivariate Estimates across Probability and Nonprobability Surveys with Population Benchmarks

Previous studies have shown many instances where nonprobability surveys were not as accurate as probability surveys. However, because of their cost advantages, nonprobability surveys are widely used, and there is much debate over the appropriate settings for their use. To contribute to this debate, we evaluate the accuracy of nonprobability surveys by investigating the common claim that estimates of relationships are more robust to sample bias than means or proportions. We compare demographic, attitudinal, and behavioral variables across eight German probability and nonprobability surveys with demographic and political benchmarks from the microcensus and a high-quality, face-to-face survey. In the analyses, we compare three types of statistical inference: univariate estimates, bivariate Pearson’s r coefficients, and 24 different multiple regression models. The results indicate that in univariate comparisons, nonprobability surveys were clearly less accurate than probability surveys when compared with the population benchmarks. These differences in accuracy were smaller in the bivariate and the multivariate comparisons across surveys. In addition, the outcome of those comparisons largely depended on the variables included in the estimation. The observed sample differences are remarkable when considering that three nonprobability surveys were drawn from the same online panel. Adjusting the nonprobability surveys somewhat improved their accuracy.