- Draw a Paik-Agresti diagram
- Find the conditional and marginal effects of X
- Find the least-squares regression coefficients (without solving for
them with \((X'X)^{-1}X'Y\))
- Why is the word ‘effects’ misleading?
Suppose you have data on continuous variable, \(Y\), and two variables \(X\) and \(Z\) that are dichotomous with values 0 and
1.
The following table shows the values of \(\bar{Y}\) for all the combinations of
values of \(X\) and \(Z\):
The number of observations in each cell is:
Z = 0 |
300 |
10 |
Z = 1 |
100 |
40 |
Questions:
- Draw the Paik diagram for this data.
- What are the ‘conditional effects of X’?
- What is the ‘marginal effect of X’?
- If you were to fit the model ’Y ~ X*Z’ in R, i.e. the model \[\hat{Y} = \hat{\beta}_0 + X \hat{\beta}_1 + Z
\hat{\beta}_2 + X Z\hat{\beta}_3\] what would the values of the
estimated regression coefficients be? (Hint: It’s a saturated model and
it gives an exact fit to \(\bar{Y}\))
- How would you get the ‘conditional effects of X’ from the regression
coefficients? Express your answer in the form of a ‘hypothesis matrix’
multiplying the vector of fitted values, i.e. a matrix \({\mathrm L}\) so that your answer would
have the form \({\mathrm L}
\hat{{\boldsymbol\beta}}\).
- What is the interpretation of the interaction coefficient, \(\hat{\beta}_3\)?
- Why is the use of the word ‘effects’ problematic in this context?
Can you think of a better word?
- “Challenging – or tedious – so not on the quiz?” How would you get
the marginal effect of X from the regression coefficients? What is the
connection with the chain rule in multivariate calculus?