6.8 Summary

  • Regression asks how one variable is related to another rather than whether two groups differ. Splitting attendance at the median gave \(t = 11.8\); using the whole variable gave \(t = 17.6\).
  • Individuals with the same \(x\) obtain very different \(y\). Forty-nine students attended exactly 25 classes and scored between 1.58 and 9.08, so no line can pass through the individual points.
  • The population regression function is the conditional mean \(E(Y \mid X = x)\). It is not a fitted line, and it exists whether or not we ever estimate it.
  • Since each conditional mean is a function of \(x\), the PRF is also the conditional expectation function. It says how the average of \(y\) moves with \(x\), not what any individual obtains.
  • Taking the CEF to be linear is a modelling choice, made after the target is defined rather than before it.
  • Writing \(y_i = E(Y \mid X = x_i) + u_i\) and substituting the linear CEF gives \(y_i = \beta_0 + \beta_1 x_i + u_i\) — the PRF with a name attached to the part it does not describe, not a new assumption.
  • The error term is not a mistake and not measurement noise. It is the rest of the world: effort, ability, teaching, health, luck.
  • \(E(u) = 0\) costs nothing, since a non-zero average can be absorbed into the intercept. \(E(u \mid x) = 0\) costs everything, cannot be tested, and is what separates a description from an effect.
  • \(\beta_0\) and \(\beta_1\) are population parameters; \(\hat{\beta}_0\) and \(\hat{\beta}_1\) are sample estimates defining the sample regression function. Two sections of thirty students gave slopes of 0.300 and 0.162.
  • The error \(u\) is the deviation from the true line and is unobservable. The residual \(\hat{u}\) is the deviation from the fitted line and can be printed.
  • Minimising \(\sum \hat{u}_i\) fails because positive and negative misses cancel. Ordinary least squares minimises \(\sum \hat{u}_i^2\).