This thesis introduces innovative models and estimation techniques within the generalized linear model framework. Chapter 1 develops the GAMBoost procedure for estimating generalized additive models, utilizing boosting and gradient descent in function space. This approach generalizes repeated fitting of residuals for exponential family responses, allowing for a flexible number of updates per covariate. Consequently, the selection of smoothing parameters is simplified to choosing the number of boosting steps, determined by approximate effective degrees of freedom. The procedure demonstrates strong performance across various examples with binary and Poisson response data, particularly excelling with many covariates and limited information. Real data applications are included. Chapter 2 presents a flexible model for discrete time survival data, accommodating non-linear covariate effects that change over time. An iterative two-step estimation procedure based on Fisher scoring is provided, alongside a simulation study that evaluates model performance under varying data complexities. This study highlights the effectiveness of using effective degrees of freedom for selecting smoothing parameters and models. Chapter 3 focuses on classification with binary response data through logistic regression, employing local models and predictor selection to reduce complexity and enhance numerical stability. Simulated data examples assess the alg
