Linear Regression for Experimental Research

Simple linear model

The population model is

\[Y_i=\beta_0+\beta_1X_i+\varepsilon_i.\]

Ordinary least squares chooses the coefficients that minimize the sum of squared residuals:

\[\min_{\beta_0,\beta_1} \sum_i(Y_i-\beta_0-\beta_1X_i)^2.\]

The estimated slope and intercept are

\[\hat\beta_1= \frac{\sum_i(X_i-\bar X)(Y_i-\bar Y)} {\sum_i(X_i-\bar X)^2},\] \[\hat\beta_0=\bar Y-\hat\beta_1\bar X.\]

The fitted value and residual are

\[\hat Y_i=\hat\beta_0+\hat\beta_1X_i, \qquad e_i=Y_i-\hat Y_i.\]

With an intercept and one slope, the residual variance estimate is

\[\hat\sigma^2=\frac{\sum_i e_i^2}{n-2}.\]

Coefficient inference

To test $H_0:\beta_j=\beta_{j,0}$,

\[t=\frac{\hat\beta_j-\beta_{j,0}} {SE(\hat\beta_j)}.\]

Under the classical linear model, use the residual degrees of freedom. Robust or design-based standard errors may be appropriate when the classical variance assumptions do not hold.

Coefficient of determination

With

\[SSE=\sum_i(Y_i-\hat Y_i)^2, \qquad SST=\sum_i(Y_i-\bar Y)^2,\]

the coefficient of determination is

\[R^2=1-\frac{SSE}{SST}.\]

$R^2$ measures in-sample fit. A high $R^2$ does not establish causality, and a randomized treatment effect can be credible even when $R^2$ is low.

Multiple regression in matrix form

The classical linear regression model is

\[Y=X\beta+\varepsilon.\]

When $X^\top X$ is invertible, the OLS estimator is

\[\boxed{\hat\beta=(X^\top X)^{-1}X^\top Y}.\]

Under homoskedastic uncorrelated errors with variance $\sigma^2$,

\[\operatorname{Var}(\hat\beta\mid X) = \sigma^2(X^\top X)^{-1}.\]

For the common null $H_0:\beta_j=0$,

\[t_j=\frac{\hat\beta_j}{SE(\hat\beta_j)}.\]

The matrix variance formula is model-based. Randomized experiments can also justify inference from the assignment mechanism; see Regression Adjustment in Experiments.

See