Testing the Information Matrix Equality with Robust Estimators

Figure 4: Power curves: Student’s t

Let S be the score test statistic, defined in the usual way. Under the
null hypothesis (p = ∞), S has a limiting χ2 distribution. Under Hn, as we
show in Appendix C.1,

S →^d χ1²(δ),

with non-centrality parameter

In conclusion, the score test and the IM tests have the same non-centrality
parameter in their limiting distribution. The power curves (as a function
of e) of 5%-level tests are given in Figure 4. The differences in power are
entirely due to differences in degrees of freedom: 1 for the score test, 2 for
the IM test with the ML estimator, and 3 for the IM test using any other
M-estimator. The difference in power between the IM tests is small.

4.3 Skewed normal alternative

Let Z — N(0, 1) and denote the distribution of ZI(Z ≤ 0) + (1 + γ) ZI(Z >
0) as FYk (Fernandez and Steel, 1998). Under the sequence of local alterna-
tives

<<   <   14   15   16   17   18   19   20   >   >>