Bayesian spline-type smoothing in generalized regression models (Q1966017)

Generalized additive models assume that, given observations \((y_i,x_{i1},\dots,x_{ip})\), the distribution of \(y_i\) belongs to an exponential family with mean \(\mu_i=E(y_i |x_i)\) linked to an additive predictor \(\eta_i\) by an appropriate response function \(h\), i.e. \(\mu_i=h(\eta_i),\;\eta_i=\gamma +f_1(x_{i1})+\cdots+f_p(x_{ip})\). To estimate the regression functions \(f_1,\dots,f_p\), the penalized log-likelihood criterion \[ PL(f_1,\dots,f_p)=\sum_{i=1}^n l_i(y_i |\eta_i)-2^{-1}\sum_{j=1}^p \lambda_j\int\left(f_j^{(m)}(u)\right)^2du\to\max \] with the likelihood contributions \(l_i\) from \(y_i |x_i\) and with separate penalty terms and smoothing parameters is used. Since the \(m\)-th derivative is used in the penalty term the maximizing functions \(\hat{f_1},\dots,\hat{f_p}\) are polynomial splines of order \(2m-1\). The regression functions \(f_1,\dots,f_p\) satisfy the stochastic differential equations \[ d^m f_j(x)=\sigma_j dW_j(x),\quad x>x_{1j}, \] where \(W_j(x)\) are mutually independent standard Wiener processes. In order to compute the Bayesian spline-type smoother based on the Markov chain Monte Carlo simulation from the posterior, these stochastic differential equations are reformulated as stochastic difference equations for the vectors \(f_j(x_{1j}),\dots,f_j(x_{nj})\), \(j=1,\dots,p\). If the data are dense, this technique produces sufficiently smooth curves. For comparison the suggested Bayesian smoothing technique is applied to data sets which are already analyzed by different methods. All algorithms are implemented in C++.