Sample Distribution of the Estimators
What Repeated Samples Say About an OLS Estimate
An estimator is random because the sample is random. Its sampling distribution describes the values it would take across repeated samples generated under a specified design and model.
For ordinary least squares, unbiasedness means the distribution is centered at the target parameter under the relevant exogeneity conditions; it does not promise a correct estimate in any single sample.
In simple regression with fixed explanatory values and independent errors with conditional mean zero and common variance sigma2, Var(beta-hat1 | x) = sigma2/Sxx, where Sxx = sum (xi - x-bar)2. Thus a larger Sxx, not merely a wider endpoint range, means a smaller slope variance when error variance is held fixed.
Sample size can affect precision through Sxx, but it is not the sole determinant.
Independent normal errors give exact normal sampling distributions for OLS estimators conditional on the design. Without that assumption, exact finite-sample normality is not guaranteed; approximate normality can arise in large samples under additional regularity conditions. Bias, variance, distributional shape, and causal interpretation are separate claims that should not be merged.
In each simple linear regression, treat the observed x values as fixed. Conditional on those values, errors are independent, have mean zero, and share variance sigma2. Both studies use the same number of observations, and their designs satisfy Sxx,A > Sxx,B > 0, where Sxx = sum (xi - x-bar)2. Which statement is best?
Conditional on fixed explanatory values, which additional assumption gives an exact normal sampling distribution for the OLS slope estimator in the usual simple linear model?