No R packages are permitted for use in this assignment.
- In the previous assignment you were asked to provide pseudocode for how to generate synthetic data from the statistical model posited for your project. The motivation was described as: synthetic data will be helpful for evaluating an estimation procedure and algorithm. Describe how you will design a simulation study based on your synthetic data to evaluate the estimation procedure and algorithm that you proposed a few weeks ago.
- In the simple linear regression model \(y_i = \beta_0 + x_i\beta_1 + u_i\) for \(i \in \{1, \ldots, n\}\), show that \(\beta_0\) is estimable by finding a vector \(a\) and scalar \(c\) such that \(E(c + a'y) = \beta_0\).
- Prove that if \(\lambda^{(1)'}\beta, \ldots, \lambda^{(k)'}\beta\) are estimable, then so is
\[ \sum_{j=1}^{k} d_j \lambda^{(j)'}\beta, \]for any scalar constants \(d_1, \ldots, d_k\).
- Show that if the least squares estimator \(\lambda'\hat{\beta}\) is the same for all solutions \(\hat{\beta}\) to the normal equations, then \(\lambda'\beta\) is estimable.
- Consider the model \(Y_{ij} = \mu + \alpha_i + \beta_i x_{ij} + U_{ij}\), for \(i \in \{1, \ldots, n\}\) and \(j \in \{1 \ldots, m\}\). Further, assume that \(\sum_{j=1}^{m}(x_{ij} - \bar{x}_{i\cdot})^2 > 0\) for all \(i \in \{1, \ldots, n\}\). Derive the necessary and sufficient conditions for an estimable function \(\lambda'\gamma\), where \(\gamma := (\mu, \alpha_1, \ldots, \alpha_n, \beta_1, \ldots, \beta_n)'\).
- The problem of least squares regression can be understood as a special case of the more general problem of ridge regression. For an \(n\)-dimensional column vector \(y\) and an \(n \times p\) design matrix \(X\), the problem of ridge regression is to solve for the parameter vector \(b\) that minimizes
\[ a\|b\|_2^2 + \|y - Xb\|_2^2, \]where \(a \geq 0\) is fixed.
- Derive a closed-form expression of the ridge regression solution.
- Assume that \(X\) has full column rank, and suppose that \(y\) is an observed instance of the random vector \(Y = X\beta + U\), where \(\beta \in \mathbb{R}^p\) is fixed and \(U\) satisfies the Gauss-Markov assumptions. Under what condition(s) is the ridge regression solution the best linear unbiased estimator (BLUE) for any \(\beta\)?