Often, one needs to check if a given linear function of β say, p'β, in the linear model set up is estimable. There is a theorem (Searle, 1972) which says that it is estimable if and only if X'X(X'X)−p = p where (X'X)− is any g-inverse of (X'X). Also, it so happens that checking the above requirement for a given g-inverse is as good as checking it for all g-inverses and hence with just one choice of a g-inverse, we can determine, using above, the estimability of a linear function.
Consider the one way classification model, described above. Suppose we want to know (a) if μ+t2 is estimable and ( b) if t1+t2 is estimable. The corresponding choices for p' are p'1 = (1 0 1 0) and p'2 = (0 1 1 0). With X as obtained earlier, for p1 and p2, the left sides of the estimability condition can be calculated using the SAS statements
left_p1 = (x_with)`*x_with*(ginv( (x_with)`*x_with ) )*p1; left_p2 = (x_with)`*x_with*(ginv( (x_with)`*x_with ) )*p2;
resulting in
and
Since LEFT_P1 is identically equal to p1, estimability of μ+t2 is established. However, since LEFT_P2 is not equal to p2, t1+t2 is not estimable.
13.58.121.131