Master Thesis:
The correct interpretation of logistic regressions.
A summary and application of the research practice by analyzing the determinants of unemployment.
Abstract: Over the last few decades the effect sizes Logits and Odds Ratios from logistic regression models were criticized for being difficult to use and hard to understand. Additionally they seem to be biased by the unobserved heterogeneity and therefore can't be interpreted as stable predictors for binary dependent variables or compared between groups. Instead it was recommended to use linear predictors like the AME or a different regression model like the LPM. But those also showed distortions. As a first step towards providing clarity this master thesis summarized the entire statistical theory and the current research regarding the logistic regression. The second step consisted of applying the correct usage on the empirical example of the determinants of unemployment. The models were theoretically based on the human capital theories and used their data from the ALLBUS survey of 2012. Interpretations of their outcome as well as the comparisions between effects were done in context of the model specification and focused on the model fit measurements. Additionally the AMEs and the LPM were used for comparing the effect sizes between different groups and the results between different models. Odds Ratios and AMEs corresponded to each other. The LPM however showed a worse model fit and partly different results.
