B0626
Title: Confidence intervals for finite mixture regression based on resampling techniques
Authors: Colin Griesbach - Georg-August-University Goettingen (Germany) [presenting]
Tobias Hepp - Friedrich-Alexander-Universitaet Erlangen-Nuernberg (Germany)
Abstract: Mixture regression models are widely used to quantify associations between outcomes and various covariates in scenarios with unobserved heterogeneity. However, meaningful uncertainty estimates are not immediately available as regular statistical inference neglects any variance regarding class assignments yielding biased results. This issue has been addressed for ordinary mixture models by employing resampling techniques like various bootstrapping routines or the jackknife and in the case of mixture regression models, bootstrapping was already used to detect identifiability issues of fitted mixture regression models. A resampling approach is proposed for uncertainty estimates of regression parameters in finite mixture regression models. The method applies empirical bootstrapping and in addition, uses a matching mechanism based on correlations of posterior class probabilities to aggregate estimates across all bootstrapping iterations and prevent label switching. Simulations and real-world applications reveal that applying the proposed resampling approach results in slightly wider confidence intervals which are now capable of holding the type-I error threshold.
