Optimal Design Points For A Simple Linear Model
Problem statement
You are studying the simple linear model
$$y_i = \alpha + \beta x_i + \varepsilon_i, \qquad i = 1, \dots, 24,$$
where the design points $x_i$ may be chosen anywhere you like in the interval $[-1, 1]$, and the errors $\varepsilon_i$ are uncorrelated with $E[\varepsilon_i] = 0$ and constant variance $\operatorname{Var}(\varepsilon_i) = \sigma^2$.
You have a budget of exactly $24$ experiments: you pick $x_1, \dots, x_{24} \in [-1,1]$, then observe the corresponding $y_1, \dots, y_{24}$.
(a) Given the data, what estimator should you use for $\alpha$ and $\beta$, and why is it "best"?
(b) How should you choose the $24$ design points $x_i$ so that the resulting estimates of $\alpha$ and $\beta$ are as precise as possible? Give the variances your design achieves.
(c) What is the practical drawback of the design you found in (b)?
Show hint
Which formula tells you how the slope's variance depends on where the design points sit?
Relevant roles: Data Scientist, Applied Scientist and Research Scientist.
Related free problems
- Computing OLS When the Design Matrix Is Too Large for MemoryLinear Algebra & OptimizationMedium