Browsing by Subject "Bayesian statistics"

The focus of this Master’s degree thesis is on the role of priors in a small sample Bayesian serial mediation. In particular, it examines how changing the informativeness of priors influences posterior parameter estimates and probability distributions. The empirical example in the study is taken from the field of applied linguistics. The research hypotheses address how the linguistic composition of two bilingual municipalities in Finland impact on the frequency of interpersonal contact young Finnish speakers have with Swedish speakers, which, in turn, increases young Finnish speakers’ communicative competence in Swedish enhancing the use of Swedish in interpersonal communication. The datasets used in the study are exchangeable; they were collected in the same population, in the Finnish secondary schools in Kokkola/Karleby and Pietarsaari/Jakobstad, with identical measures. The first dataset was collected in 2014 (N = 202) and it is used as prior sample in the analyses. The second dataset was collected in 2014 (N = 55) and it is used as the focal sample in the analyses. The results demonstrate that the grade of informativeness of priors fundamentally influences the outcomes of the mediational analyses. When using noninformative priors, some parameters do not differ from zero and the model suggests no mediated effect. However, when using informative priors, both parameter estimates and the mediated effect appear to be different from zero. Controversies around using priors in Bayesian analysis and their significance in empirical applications are discussed.