The McGraw-Hill Series Economics essentials of economics brue, McConnell, and Flynn Essentials of Economics

OLS
het
GLS
0.5
0.164
0.134
0.110
0.285
0.277
0.243
1.0
0.142
0.101
0.048
0.246
0.247
0.173
2.0
0.116
0.074
0.0073
0.200
0.220
0.109
3.0
0.100
0.064
0.0013
0.173
0.206
0.056
4.0
0.089
0.059
0.0003
0.154
0.195
0.017
Note:
OLS
het
means OLS allowing for heteroscedasticity.
The most striking feature of these results is that OLS, with or without correction for het-
eroscedasticity, consistently overestimates the true standard error obtained by the (correct)
GLS procedure, especially for large values of
α
,
thus establishing the superiority of GLS.
These results also show that if we do not use GLS and rely on OLS—allowing for or not
allowing for heteroscedasticity—the picture is mixed. The usual OLS standard errors are
either too large (for the intercept) or
generally
too small (for the slope coefficient) in relation
to those obtained by OLS allowing for heteroscedasticity. The message is clear: In the pres-
ence of heteroscedasticity, use GLS. However, for reasons explained later in the chapter, in
practice it is not always easy to apply GLS. Also, as we discuss later, unless heteroscedastic-
ity is very severe, one may not abandon OLS in favor of GLS or WLS.
From the preceding discussion it is clear that heteroscedasticity is potentially a serious
problem and the researcher needs to know whether it is present in a given situation. If its
6
From Eq. (5.3.6) we know that the 100(1
−
α
)% confidence interval for 
β
2
is [
ˆ
β
2
±
t
α/
2
se (
ˆ
β
2
)]. But
if se (
ˆ
β
2
) cannot be estimated unbiasedly, what trust can we put in the conventionally computed 
confidence interval?
7
Russell Davidson and James G. MacKinnon, 
Estimation and Inference in Econometrics
, Oxford 
University Press, New York, 1993, pp. 549–550.
guj75772_ch11.qxd 12/08/2008 07:04 PM Page 375

---

376