The Impact of a Hausman Pretest on the Size of a

Hypothesis Test: the Panel Data Case

Patrik Guggenberger
+
Department of Economics
UCLA
May 24, 2009
Abstract: The size properties of a twostage test in a panel data model are investigated
where in the rst stage a Hausman (1978) specication test is used as a pretest of the random
eects specication and in the second stage, a simple hypothesis about a component of the
parameter vector is tested, using a tstatistic that is based on either the random eects or the
xed eects estimator depending on the outcome of the Hausman pretest. It is shown that the
asymptotic size of the twostage test depends on the degree of time variation in the regressors
and on the variance of the error term relative to the variance of the individual specic eect and
equals 1 for empirically relevant specications of the parameter space. Monte carlo simulations
document that the size distortion is well reected in nite samples. The size distortion is
caused mainly by the poor power properties of the pretest that lead to frequent unjustied
inference based on the random eects estimator in the second stage. However, it is also shown
that the conditional size of the test, conditional on the Hausman pretest rejecting the pretest
null hypothesis, exceeds the nominal level of the test. Given the results in the paper, the
recommendation then is to use a tstatistic based on the xed eects estimator instead of using
the twostage procedure.
Keywords: Fixed eects estimation, Hausman specication test, panel data, pretest, ran-
dom eects estimation, size distortion.
JEL classication: C12, C23, C52.

Patrik Guggenberger, Department of Economics, UCLA, Bunche Hall 8385, Box 951477, Los Angeles,
CA 90095-1477. Tel.: (310) 825-0849. Email: guggenbe@econ.ucla.edu. I would like to thank a co-editor, an
associate editor, and two referees for excellent comments, the NSF for support under grant number SES-0748922,
the Alfred P. Sloan Foundation for a 2009 fellowship, and seminar and conference participants at various places.
1
1 Introduction
When deciding between inference based on the random eects or the xed eects estimator in
a panel data model, it is quite standard in applied work to rst implement a Hausman (1978)
pretest. If the Hausman pretest rejects the pretest null hypothesis that the random eects
specication is correct, inference based on the xed eects estimator is used in the second
stage, otherwise inference based on the random eects estimator is used which has favorable
power properties. For example, Blonigen (1997) justies the use of random eects inference
based on a Hausman pretest while Hastings (2004) uses xed eects inference as a result of the
Hausman pretest rejecting the random eects specication. The Hausman pretest is a common
tool, used in hundreds of applied papers and discussed in most textbooks in Econometrics, see
e.g. Wooldridge (2002, chapter 10.7.3).
It is shown in this paper that the asymptotic size of the resulting twostage test equals
1 for empirically relevant specications of the parameter space. An explicit formula for the
asymptotic size of the twostage test is derived. It shows that the asymptotic size depends
on the degree of time variation in the regressors and also on the relative magnitude of the
error variance to the variance of the individual specic eect. Our results explain how these
two quantities impact the size of the twostage test. The result that the twostage test is size
distorted is related to the ndings in Guggenberger (2010). In that paper it is shown that the
corresponding twostage test in the linear instrumental variables (IV) model has size 1, where
the Hausman pretest is used as a test of exogeneity of a regressor. As outlined in more detail
below, the analysis of the panel data example is more complicated than the analysis of the
IV example, because in the former case the asymptotic size depends on a higher dimensional
nuisance parameter vector than in the latter case.
Based on the general theory developed in Andrews and Guggenberger (2010a, AG(2010a)
from now on), we characterize sequences of nuisance parameters that lead to the highest null
rejection probabilities of the twostage test asymptotically. It is shown that under certain
local deviations from the random eects specication, the Hausman pretest statistic converges
to a noncentral chisquare distribution. The noncentrality parameter is small when the error
variance is large relative to the variance of the individual specic eect or when the regressors
are positively correlated over time. In this situation, the Hausman pretest has low power
against local deviations of the pretest null hypothesis and consequently, with high probability,
inference based on the random eects estimator is performed in the second stage which leads
to size distortion. However, it is also shown that the conditional size of the twostage test,
2
conditional on the Hausman pretest rejecting the pretest null hypothesis, exceeds the nominal
level of the test.
Given the results in the paper, if controlling the size of a testing procedure is an objective,
the use of the twostage procedure cannot be recommended. Its asymptotic size is severely
distorted and the size distortion is well reected in nite sample simulations. On the other
hand, use of a tstatistic based on the xed eects estimator has correct asymptotic size and
performs well in nite samples. If the random eects specication is correct, inference based
on the random eects estimator has correct size and has favorable power properties, but of
course leads to size distortion otherwise. Given the results in the paper, the random eects
specication should not be tested using a Hausman pretest.
It has been long known that pretests have an impact on the risk properties of estimators
and the size properties of tests, see Judge and Bock (1978) for an early reference and Guggen-
berger (2010) for additional references. As documented further below, the specication tests
proposed in Hausman (1978) are routinely used as pretests in applied work. However, besides
Guggenberger (2010) where the case of the linear IV model is studied, no results are stated
anywhere in the literature regarding the negative impact of the Hausman pretest on the size
properties of a twostage test.
The remainder of the paper is organized as follows. Section 2 describes the model, the
objective, and denes the test statistics. In Section 3, nite sample evidence is provided.
Subsection 4.1 is based on AG(2010a). It provides theoretical background on how to calculate
the asymptotic size of a test in situations where the test statistic has a limiting distribution
that is discontinuous in nuisance parameters. In subsection 4.2 this theory is then applied to
the situation of a twostage test where in the rst stage a Hausman pretest is implemented.
All technical details are given in the Appendix.
The following notation is used in the remainder of the paper. We denote by 1
T
and 1
T
a
1vector of ones and the 1dimensional identity matrix, respectively. For a matrix with 1
rows, let `

= 1
T
1

. where 1

= (
0
)
1

0
is the projection onto the column space of
. By
2
1,o
and .
o
we denote the ,quantile of a chisquare distribution with one degree of
freedom and of a standard normal distribution, respectively. By 1() we denote the indicator
function that equals 1 if the argument is true and 0 otherwise. Denote by [[ [[ the Euclidean
norm. Finally, let 1
+
= r 1 : r _ 0. 1
1
= 1 ' . and 1
+,1
= 1
+
' +.
3
2 Model and Objective
Consider the simple panel data model
(1)
it
= ` +r
it
o +j
i
+
it
.
for i = 1. .... `. t = 1. .... 1 with scalar parameters ` and o and individual specic eect j
i
. see
e.g. Baltagi (2005, eqn. (2.1)).
1
Denote the regressor vector by n
0
it
= (1. r
it
). By
i
. r
i
. n
i
. and

i
we denote the matrices (or vectors) with 1 rows given by
it
. r
it
. n
0
it
. and
it
. respectively. The
observed data are (
i
. r
i
) 1
T2
. i = 1. .... `. The data (r
i
. j
i
.
i
). i = 1. .... ` are assumed to
be i.i.d. with distribution 1 and
it
. t = 1. .... 1 are i.i.d. Assume 1
1
r
it
= 1
1
j
i
= 1
1

it
= 0.
1
1
j
i

it
= 0. and dene o
2

= 1
1

2
it
and o
2
j
= 1
1
j
2
i
. To simplify notation, we do not index o
2

and o
2
j
by 1. Our asymptotic framework has ` . but 1 xed.
The object of interest is to test the null hypothesis
(2) H
0
: o = o
0
against a one or twosided alternative. One possibility to test (2) is to use a tstatistic 1
11
(o
0
)
based on the random eects estimator

o
11
of o. To dene these quantities, let

1
TT
be an
estimator of the variancecovariance matrix
(3) = 1
1
(
i
+ 1
T
j
i
)(
i
+ 1
T
j
i
)
0
= o
2

1
T
+o
2
j
1
T
1
0
T
1
TT
that replaces o
2

and o
2
j
in by estimated counterparts o
2

and o
2
j
. Possible choices for o
2

. o
2
j
.
and for the estimator o
2

introduced below are discussed in subsection 5.1 of the Appendix.

Then
2
(

`
11
.

o
11
)
0
=
_
.

i=1
n
0
i
(

)
1
n
i
_
1
.

i=1
n
0
i
(

)
1

i
.

\
11
= (`
1
.

i=1
r
0
i
(

)
1
r
i
)
1
. and
1
11
(o
0
) =
`
12
(

o
11
o
0
)
(

\
11
)
12
. (4)
1
Additional regressors n
it
could be included into the model at the expense of more complicated notation.
The asymptotic results of the paper are identical if the intercept ` is not included into the model.
2
The denition of T
RE
(0
0
) in (4) could also be altered by replacing
1

N
i=1
r
0
i
(

)
1
r
i
by

1
[
N

i=1
r
0
i
(

)
1
r
i

N

i=1
r
0
i
(

)
1
1
T
(
N

i=1
1
0
T
(

)
1
1
T
)
1
N

i=1
1
0
T
(

)
1
r
i
].
As veried in (40) in the Appendix, this modication makes no dierence asymptotically for the results in
the paper.
4
Alternatively, the test of (2) can be based on the xed eects estimator

o
11
of o
(5)

o
11
= (
.

i=1
r
0
i
`
1
T
r
i
)
1
.

i=1
r
0
i
`
1
T

i
by using the tstatistic
1
11
(o
0
) =
`
12
(

o
11
o
0
)
(

\
11
)
12
. where

\
11
= (`
1
.

i=1
r
0
i
`
1
T
r
i
, o
2

)
1
(6)
for an estimator o
2

of o
2

that may dier from o

2

.
Let r
i
= 1
1

T
t=1
r
it
be the time average of the regressor. Inference based on the tstatistic
1
11
is justied if r
i
and j
i
are uncorrelated but size distorted otherwise. On the other hand
if r
i
and j
i
are uncorrelated, inference based on 1
11
provides power advantages over inference
based on 1
11
. Because of this tradeo between robustness and power, oftentimes in applied
work, before testing (2), a Hausman (1978) pretest is undertaken. The pretest tests whether
the pretest null hypothesis
(7) H
1,0
: Co::
1
(j
i
. r
i
) = 0
is true. If the pretest rejects the pretest hypothesis, then, in the second stage, H
0
: o = o
0
is
tested based on 1
11
(o
0
). the robust testing procedure when H
1,0
is false. If the pretest does not
reject (7), then in the second stage (2) is tested based on 1
11
(o
0
). the more powerful testing
procedure when (7) is true. Thus, denoting by
(8) H
.
=
`(

o
11

o
11
)
2

\
11


\
11
the Hausman statistic and by , the nominal size of the pretest, the resulting twostage test
statistic 1
.
(o
0
) is given by
(9) 1

.
(o
0
) = 1
11
(o
0
)1(H
.
_
2
1,1o
) +1
11
(o
0
)1(H
.

2
1,1o
)
for an upper onesided test and by 1

.
(o
0
) or [1

.
(o
0
)[ for a lower onesided or a symmetric
twosided test, respectively. The nominal size c test rejects H
0
if
(10) 1
.
(o
0
) c
1
(1 c).
where c
1
(1 c) = .
1c
. .
1c
. and .
1c2
for the upper onesided, lower onesided, and
symmetric twosided test, respectively.
5
The goal of this paper is to illustrate the impact of the pretest on the size properties of the
twostage test. We show that the asymptotic size :o.(o
0
) of the test diers substantially
from its nominal size c and determine the parameters that impact :o.(o
0
). Note that if
Co::
1
(j
i
. r
i
) is nonzero and kept xed as ` goes o to innity, then the twostage procedure
has asymptotic null rejection probability equal to the nominal size of the test because in this
case the Hausman pretest statistic diverges to innity, and in the second stage 1
11
is used
with probability approaching 1. However, this is only a pointwise justication of the twostage
procedure and it does not hold uniformly.
3 Finite Sample Evidence
For illustration of the overrejection problem of the twostage test dened in (10), nite sample
simulations are conducted in this section. The theoretical results below show that the only
parameters that impact the null rejection probability asymptotically are
1
= Co::
1
(j
i
. r
i
).

21
= (1o
2
a
i
,1
1
[[r
i
[[
2
)
12
. and
22
= (1o
2
j
,o
2

)
12
. where
(11) o
2
a
i
= 1
1
r
2
i
.
The results below also prove that asymptotically, it does not matter whether or not the intercept
` is included in the model (1) and, for simplicity, we therefore consider a model without the
intercept. We choose sample size ` = 100 and 1 = 2 and generate 1 = 30. 000 i.i.d. draws
from
(12)
_
_
_
_
r
i1
r
i2
j
i
_
_
_
_
~ `(0.
_
_
_
_
1 j
j 1
1
_
_
_
_
) and
_

i1

i2
_
~ `(0.
_
o
2

0
0 o
2

_
).
Note that dierent values for (j. . o
2

) translate into values for

1
.
21
. and
22
through the
relation
1
= ,(.5+.5j)
12
.
21
= (.5+.5j)
12
. and
22
= (2,o
2

)
12
. Finally, we choose nominal
sizes c = , = .05.
3
Small choices for the pretest nominal size , such as 5% are common in
applied work when applying Hausman pretests, both in panel data and linear IV applications.
E.g. Gaynor, Seider, and Vogt (2005, p.245) state A HausmanWu test does not reject the
null hypothesis of exogeneity (j value = 0.06). ... Given these results ... we treat volume as
exogenous hereafter. and Bedard and Deschnes (2006, p.189) state ... the Jerry R. Hausman
3
When calculating the estimators o
2
v
, o
2

, and o
2
v
, we use the degrees of freedom adjustment 1 = 1, see
Subsection 5.1 below.
6
(1978) test, testing the null hypothesis that the dierence between TSLS and OLS coecients is
due only to sampling error, is rejected at the 5percent level. However, many times the pretest
nominal size , is not even reported in applied papers. E.g. Blonigen (1997, p.453) states A
Hausman test indicated that the random eects model estimates are consistent for these data,
and thus I report only the more ecient random eects model estimates. and Banerjee and
Iyer (2005, p.1205) state A Hausman test does not reject the null hypothesis that the OLS
and IV coecients are equal. As further indication, that small values of , are common in
applied work, consider Bradford (2003, p.1757) that states that the Hausman statistic which
is distributed as a chisquare with two degrees of freedom under the null is calculated at 1.46.
This fails to reject the null at any reasonable level of signicance. Consequently, these two
variables are treated as exogenous regressors hereafter. Note that the jvalue in this case is
.48. Thus, choices of , of that magnitude are considered unreasonable.
In Table 1a) we list null rejection probabilities of the symmetric twostage test and in Table
1b) rejection probabilities of the Hausman pretest. Results for the following 30 parameter
combinations, based on 1 = 30. 000 simulation repetitions, are reported:
(13) (j. . o
2

) .3. .6. .9 0. .3. .4. .5. .6 1. 5.

The possible choices for j and translate into a grid of values for
1
in the interval [0,.74] and
the two choices for o
2

translate into the values .63 and 1.41 for

22
.
Insert Table 1 here
Table 1a) reveals that the twostage test overrejects severely. At the nominal size of 5%, the
null rejection rates reach values higher than 80% for the parameter combinations considered
here. When = 0, i.e. when j
i
and the regressors r
i1
and r
i2
are independent, the null
rejection probabilities are relatively close to the nominal size of the test and fall into the interval
[6.2%,8.6%] over the parameter combinations reported here. In this case, inference based on
both 1
11
and 1
11
is justied. However, when is nonzero and thus j
i
and the regressors
r
i1
and r
i2
are correlated, only inference based on 1
11
is justied but inference based on 1
11
is sizedistorted. The twostage test overrejects because the Hausman pretest does not reject
frequently enough (as documented in Table 1b) in these cases but the resulting inference based
on 1
11
in the second stage leads to very frequent rejections. For example, when j = = .6
and o
2

= 5 the Hausman pretest only rejects in 45.7% of the cases. The resulting frequent use
of 1
11
in the second stage leads to the 51.6% rejection rate of the twostage test.
The simulations reveal that overrejection increases in j and, most of the times, also in o
2

.
Picking j close to 1 and/or large enough values for o
2

. the null rejection probability of the two

7
stage test can be made arbitrarily close to 100%. This is consistent with the theoretical results
reported in the next section. Such large values of j are probably not empirically relevant, but
the important message from the simulations reported here is that severe size distortion of the
twostage test also occurs for empirically relevant choices of the parameters.
In contrast to the sizedistorted twostage procedure, the simple onestage test that always
uses the test statistic 1
11
has very reliable null rejection probabilities. Over all the parameter
combinations in (13), the null rejection probabilities of the onestage test fall into the inter-
val [5.0%,5.3%]. Note that the corresponding interval of the twostage test is [5.3%,80.8%].
Therefore, if controlling the size is an objective, use of the twostage procedure can not be
recommended.
4 Asymptotic Size of a Test
In subsection 4.1, the theoretical background is discussed of how to determine the asymptotic
size of a xed critical value test in a situation where the test statistic that the testing procedure
is based on, has an asymptotic distribution that may be discontinuous in certain nuisance
parameters. This theory is taken from AG(2010a) and illustrated in a motivational example,
namely inference in a simple version of the linear instrumental variables model. Then, in
subsection 4.2, this theory is applied to the twostage test with a Hausman pretest in the rst
stage.
4.1 General Theory and Motivation
Consider a general testing problem of nominal size c with test statistic 1
.
(o
0
) and nonrandom
critical value c
1
(1 c). Assume the model contains a possibly innitedimensional nuisance
parameter vector . Then, by denition, the asymptotic size of the test of H
0
: o = o
0
equals
(14) :o.(o
0
) = limsup
.!1
sup
2
1
0
0
,
(1
.
(o
0
) c
1
(1 c)).
where 1
0,
() denotes probability when the true parameters are (o. ). Uniformity over
which is built into the denition of :o.(o
0
) is crucial for the asymptotic size to give a good
approximation for the nite sample size.
Following AG(2010a), the parameter is decomposed into three components: = (
1
.
2
.
3
).
The points of discontinuity of the asymptotic distribution of the test statistic of interest are
8
determined by the rst component,
1
. The parameter space of
1
is
1
. The second component,

2
. of also aects the limit distribution of the test statistic, but does not aect the distance
of the parameter to the point of discontinuity. The parameter space of
2
is
2
. The third
component,
3
. of does not aect the limit distribution of the test statistic. The parameter
space for
3
is
3
(
1
.
2
). which generally may depend on
1
and
2
. The parameter space
for satises
Assumption A. (i)
(15) = (
1
.
2
.
3
) :
1

1
.
2

2
1
q
.
3

3
(
1
.
2
)
and (ii)
1
=

1
.
&
1
| for some _

1
<
&
1
_ that satisfy

1
_ 0 _
&
1
. where denotes
the left endpoint of an interval that may be open or closed at the left end and | is dened
analogously for the right endpoint.
Dene the index set for the dierent asymptotic null distributions of the test statistic 1
.
(o
0
)
of interest. Let
H = / = (/
1
. /
2
) 1
1+q
1
:
.
= (
.,1
.
.,2
.
.,3
) : ` _ 1
such that `
v

.,1
/
1
and
.,2
/
2
. (16)
Denition of
.,I
: ` _ 1: Given : 0 and / = (/
1
. /
2
) H. let
.,I
= (
.,I,1
.
.,I,2
.

.,I,3
) : ` _ 1 denote a sequence of parameters in for which `
v

.,I,1
/
1
and
.,I,2
/
2
.
The sequence
.,I
: ` _ 1 is dened such that under
.,I
: ` _ 1. the asymptotic
distribution of 1
.
(o
0
) depends on / and only /. This is formalized in the following assumption.
Assumption B. For some : 0. all / H. all sequences
.,I
: ` _ 1. and some distribu-
tions J
I
. 1
.
(o
0
)
o
J
I
under
.,I
: ` _ 1.
The next theorem, a special case of Theorem 1(a) in AG(2010a), provides a formula for
:o.. In contrast to the formula of :o. in (14), the formula in the theorem can be used for
explicit calculation. It shows that the worst case sequence of nuisance parameters, a sequence
that yields the highest asymptotic null rejection probability, is of the type
.,I
: ` _ 1.
Theorem 1 (AG(2010a)) Suppose Assumptions A and B hold where J
I
: 1 [0. 1] is a
continuous function. Then, :o.(o
0
) = sup
I21
[1 J
I
(c
1
(1 c))].
4.2 Asymptotic Size After Hausman Pretest
In this subsection, we return to the panel data application and the twostage test with a
Hausman pretest in the rst stage introduced in (10). To apply Theorem1, we have to determine
9
the decomposition of the nuisance parameter vector into discontinuous and continuous
elements, the rate of convergence :. and verify Assumptions A and B. Finally, one needs to
derive the limiting distribution J
I
of the test statistic 1
.
(o
0
) under
.,I
.
Let = (
1
.
2
.
3
) with
2
= (
21
.
22
).
(17)
1
= Co::
1
(j
i
. r
i
).
21
= (
1o
2
a
i
1
1
[[r
i
[[
2
)
12
.
22
= (1o
2
j
,o
2

)
12
.
and
3
= (1. `). The component
21
measures the expected time variation in the regressor
while the component
22
is a function of the ratio of the variances of the individual specic
eect and the error term
it
. In all the examples studied in AG(2009a,b, 2010a-c),
2
is one
dimensional and this is the only example where the
2
component is twodimensional. In
particular, Guggenberger (2010) studies the asymptotic size of the twostage test in the linear
IV context where the Hausman pretest is used as a test of exogeneity of a regressor. The
2

component there is scalar and is a function of the concentration parameter. Here, the situation
is more complex and two separate parameters impact the asymptotic size of the twostage test
through
2
.
The parameter space of is dened as in (15) with = 2.
2
=
21

22
and
(18)
1
= [1. 1].
21
= [i
1
. i
1
]. and
22
= [i
2
. i
2
]
for some 0 < i
1
< i
1
< 1 and 0 < i
2
< i
2
< . Let

3
(
1
.
2
) = (1. `) : ` 1; 1
1
r
it
= 1
1
j
i
= 1
1

it
= 0.
1
1

2
it
= o
2

. 1
1
j
2
i
= o
2
j
for some nite o
2

. o
2
j
0.
Co::
1
(j
i
. r
i
) =
1
. (1o
2
a
i
,1
1
[[r
i
[[
2
)
12
=
21
. (1o
2
j
,o
2

)
12
=
22
.
1
1
r
it

ic
= 1
1
r
it
r
ic

i
j
i
= 1
1
j
i

it
= 0. 1
1
(r
2
i
j
2
i
) = o
2
a
i
o
2
j
+ 2(1
1
r
i
j
i
)
2
.
1
1
r
it
r
ic

i&
= 1
1
r
it
r
ic
1
1

i&
.
_
_
1
1
_
[r
i
,o
a
i
[
2+c
. ([[r
i
[[
2
,1
1
[[r
i
[[
2
)
1+c
.
[r
0
i

i
,(o
a
i
o

)[
2+c
. [r
i

i
,(o
a
i
o

)[
2+c
. [r
i
j
i
,(o
a
i
o
j
)[
2+c
. (
2
it
,o
2

)
1+c
)
_
_
_ ` (19)
for some constants o 0. ` < . and subindices t. :. . n = 1. .... 1. The condition 1
1
(r
2
i
j
2
i
) =
1
1
r
2
i
1
1
j
2
i
+2(1
1
r
i
j
i
)
2
holds, for example, if r
i
and j
i
are jointly normal. The remaining condi-
tions are moment restrictions, that imply that Liapounovtype central limit theorems (CLT) or
weak law of large numbers (WLLN) for independent 1
1+c
bounded random variables hold, and
the conditional homoskedasticitytype assumption 1
1
r
it
r
ic

i&
= 1
1
r
it
r
ic
1
1

i&
. With
the above denitions, Assumption A clearly holds.
10
For H dened as in (16) it follows that
(20) H = H
1
H
2
= 1
1
[i
1
. i
1
] [i
2
. i
2
].
For every / = (/
1
. /
2
) H. denote by
.,I
a sequence of parameters with components

.,I,1
.
.,I,2
= (
.,I,21
.
.,I,22
). and
.,I,3
.
.,I
= (
.,I,1
.
.,I,2
.
.,I,3
). where

.,I,1
= Co::
1
N
(j
i
. r
i
).
.,I,21
= (
11
1
N
r
2
i
1
1
N
[[r
i
[[
2
)
12
.
.,I,22
= (11
1
N
j
2
i
,1
1
N

2
it
)
12
s.t.
`
12

.,I,1
/
1
.
.,I,2
/
2
. and
.,I,3
= (1
.
. `
.
)
3
(
.,I,1
.
.,I,2
).
(21)
In the Appendix, for every / H we derive the limit distribution J
I
of the test statistic
1
.
(o
0
) under the sequence
.,I
. see (49). This veries Assumption B with : = 1,2.
Then, applying Theorem 1 the asymptotic size of the twostage test equals
(22) :o.(o
0
) = sup
I21
[1 J
I
(c
1
(1 c))]
with H dened in (20) and J
I
dened in (49). The formula applies to upper, lower one
sided, and symmetric twosided versions of the test with c
1
(1 c) = .
1c
. .
1c
. and .
1c2
.
respectively. Note that :o.(o
0
) depends on c. ,. and on the boundaries in the denition of

2
. For notational simplicity, this dependence is suppressed.
Figure 2 plots the asymptotic maximal rejection probability of the symmetric twostage
test, where the maximum is taken over /
1
H
1
. as a function of /
21
and /
22
, i.e. the gure
plots the function
(23) ,(/
21
. /
22
) = sup
I
1
21
1
[1 J
(I
1
,I
21
,I
22
)
(c
1
(1 c))]
for c = , = .05.
4
For small values of /
21
. ,(/
21
. /
22
) is close to the nominal size 5% of the test.
For /
21
_ .4. ,(/
21
. /
22
) < .1. The size distortion increases as /
21
increases and the asymptotic
maximal null rejection probability gets arbitrarily close to 1 as /
21
1 (as documented in
additional simulations). For xed /
21
. the function ,(/
21
. /
22
) decreases as /
22
increases and
as /
22
it decreases to the nominal size of the test. However, the slope of the function
4
For each /, the results are based on 1 = 30, 000 random draws from the distribution of J
h
. We consider /
1
values in [-2000,2000] using a grid with stepsize .01 on [0,.1], stepsize .1 on [.1,1] stepsize 1 on [1,10], stepsize 10
on [10,100], and stepsize 50 on [100,2000] and the analogous grid for negative /
1
values.
11
,(/
21
. ) is rather small and it takes rather large values of /
22
to make ,(/
21
. /
22
) small when
/
21
is close to 1. For example, ,(.95. /
22
) equals 63.7, 40.0, 25.9, 16.9, 13.0, 10.8, and 9.4% for
/
22
= 1. 2. .... 7.
Insert Figure 2 here
What is the reason for the size distortion? It is shown in (45) that H
.

o

2
1
(/
2
1
/(/+1)
1
)
for / = /
2
22
(1 /
2
21
) under
.,I
. where
2
1
() denotes a noncentral chisquare distribution with
one degree of freedom with noncentrality parameter given by the expression in brackets. If
/
1
/ ,= 0. the Hausman pretest has nonzero local power. However, the noncentrality parameter
/
2
1
/(/+1)
1
of the limiting distribution in (45) is small when / is small which is the case if /
21
is close to 1 or if /
22
is close to 0. In these cases, the pretest has poor power properties and the
twostage test frequently uses inference based on 1
11
in the second stage. But the test based
on 1
11
tends to reject frequently under moderate failures of the pretest hypothesis (7) which
leads to size distortion of the twostage test. The parameter /
21
is close to 1when there is little
time variation in the regressor, i.e. in the extreme case where r
it
= r
ic
for all :. t = 1. .... 1.
/
21
= 1. In the case where r
it
. t = 1. .... 1 is i.i.d., /
21
= 1
12
; for example, /
21
= .71 and
/
21
= .58 when 1 = 2 or 3, respectively. Note that if i
1
= .71. the simulations for Figure 2
show that :o.(o
0
) is about 30% if i
2
is small. So, even in the case where the regressor r
it
is uncorrelated for dierent time indices t. the twostage test is extremely size distorted.
5
The
parameter /
22
is small when the ratio of the variances of the individual specic eect and of
the error term is small.
Insert Table 3 here
Table 3 reports conditional rejection probabilities of the symmetric twostage test, conditional
on the Hausman pretest rejecting the pretest null hypothesis, 1 C 1 = 1(1
.
(o
0
)
c
1
(1c)[H
.

2
1,1o
). and conditional on the pretest not rejecting the pretest null hypothesis,
1C `1 = 1(1
.
(o
0
) c
1
(1 c)[H
.
<
2
1,1o
). when c = , = .05 and /
1
= 15 for a grid
of /
21
and /
22
values. Table 3 also reports rejection probabilities of the Hausman pretest. The
results in Table 3 are based on 1 = 3. 000. 000 simulation repetitions. Even though /
1
= 15 is
quite large, these latter rejection probabilities can be quite small, especially when /
21
is close
to one and/or /
22
is close to 0. This is consistent with the local power result of the Hausman
pretest described in the previous paragraph because when /
21
is close to one and/or /
22
is close
to 0 then the noncentrality parameter /
2
1
/(/ + 1)
1
is close to 0. For example, when /
21
= .75
and /
22
= .1. then the Hausman pretest rejects in 16.9% of the cases. However, in cases where
5
Values of /
21
close to zero are possible if r
it
is negatively correlated, a case which is probably of lesser
importance in applied work.
12
the Hausman pretest does not reject despite the fact that /
1
is 15 the rejection probability
in the second stage can be very high. This is because then in the second stage inference is
based on 1
11
which takes on relatively large values when /
1
is nonzero. For example, in the
case /
21
= .75 and /
22
= .2. conditional on the Hausman pretest not rejecting (which happens
in 49.5% of the cases), the test rejects with probability 59.1% in the second stage. Perhaps
more surprisingly, size distortion of the twostage test is also caused by the twostage test
rejecting at high frequency conditional on the Hausman pretest rejecting in the rst stage.
This is despite the fact, that then in the second stage inference is based on the statistic 1
11
and the unconditional size of the onestage test based on 1
11
is c = .05. For example, in the
case /
21
= .75 and /
22
= .05. conditional on the Hausman pretest rejecting (which happens in
8.0% of the cases), the test rejects with probability 26.0% in the second stage. The reason for
this overrejection is that the Hausman statistic and the tstatistic 1
11
are correlated and if
the former statistic takes on large values (and therefore the pretest hypothesis is rejected and
the twostage test is based on 1
11
in the second stage) the latter statistic is likely to take on a
large value too. This correlation increases as /
21
approaches one and/or /
22
approaches zero.
6
5 Appendix
The Appendix provides possible choices for the variance estimators o
2

. o
2
j
. and o
2

and contains
the derivation of the asymptotic distribution of the twostage test statistic under sequences

.,I
.
5.1 Possible choices for o
2

. o
2
j
. and o
2

Following Wooldridge (2002, p.260 and 271), one alternative to dene the variance estimators
is as follows. Let
(24) (

`
O1S
.

o
O1S
)
0
=
_
.

i=1
n
0
i
n
i
_
1
.

i=1
n
0
i

i
be the pooled OLS estimator of (. o)
0
and
(25) n
it
=
it
n
0
it
(

`
O1S
.

o
O1S
)
0
6
Note that if we evaluate the limiting distribution
H;h
of H
n
in (43) at /
21
= 1 and /
22
= 0 the result is

2
FE;h
which is the squared limiting distribution of T
FE
, derived in (38).
13
the residuals from pooled OLS regression estimating n
it
= j
i
+
it
. Then let
o
2
&
= (`1 1)
1
.

i=1
T

t=1
n
2
it
.
o
2
j
= (`1(1 1),2 1)
1
.

i=1
T1

t=1
T

c=t+1
n
it
n
ic
.
o
2

= o
2
&
o
2
j
(26)
for 1 = 0 or 2 depending on whether a degreesoffreedom correction is desired. We can also
estimate o
2

based on the xed eects estimator

o
11
in (5). Let
i
= 1
1

T
t=1

it
and let
(27)
it
= (
it

i
) (r
it
r
i
)

o
11
be the xed eects residuals estimating
it
and dene the estimator
(28) o
2

= (`(1 1) 1)
1
.

i=1
T

t=1

2
it
.
For the asymptotic results below, the specic choice of variance estimators o
2

or o
2

does not
matter as long as they are consistent in the sense that o
2

,o
2

j
1 and o
2

,o
2

j
1 under
sequences
.,I
with nite /
1
.
5.2 Derivation of J
/
In this subsection, the limit distribution of the test statistic 1
.
(o
0
) is derived under sequences

.,I
. Two cases are dealt with separately. Case I has [/
1
[ < while Case II has [/
1
[ = .
In Case I,
1
0. Recall that if 1
.
is the true distribution, then o
2
j
= 1
1
N
j
2
i
. o
2

= 1
1
N

2
it
.
and o
2
a
i
= 1
1
N
r
2
i
. Under any sequence
.,I
for which
.,I,1
= Co::
1
N
(j
i
. r
i
)
1
.
(29)
_
_
_
_
_
_
_
_
(o
2
a
i
o
2
j
)
12
`
12
.

i=1
(r
i
j
i
1
1
N
r
i
j
i
)
(o
2
a
i
o
2

,1)
12
`
12
.

i=1
r
i

i
(1
1
N
[[r
i
[[
2
o
2

)
12
`
12
.

i=1
r
0
i

i
_
_
_
_
_
_
_
_

o
_
_
_
_

j,
1

,I
21

,I
21
_
_
_
_
~ `(0.
_
_
_
_
1 +
2
1
0 0
0 1 /
21
0 /
21
1
_
_
_
_
).
The result holds by the Liapounov CLT for independent, mean zero, 1
2+c
bounded ran-
dom variables using the moment restrictions in (19) noting that 1
1
N
r
2
i

2
i
= o
2
a
i
o
2

,1 and
1
1
N
(r
0
i

i
)
2
= 1
1
N
[[r
i
[[
2
o
2

. In particular, the condition 1

1
N
(r
2
i
j
2
i
) = o
2
a
i
o
2
j
+2(1
1
N
r
i
j
i
)
2
from
(19) yields
(30) 1
1
N
(r
i
j
i
1
1
N
r
i
j
i
)
2
= 1
1
N
(r
i
j
i
)
2
(1
1
N
r
i
j
i
)
2
= o
2
a
i
o
2
j
+ (1
1
N
r
i
j
i
)
2
14
and thus (o
2
a
i
o
2
j
)
1
1
1
N
(r
i
j
i
1
1
N
r
i
j
i
)
2
= 1 +Co::
1
N
(j
i
. r
i
)
2
. The limiting distribution
j,
1
is independent of
,I
21
.
,I
21
because of the conditions 1
1
N
r
it

ic
= 1
1
N
r
it
r
ic

i
j
i
= 0 for
t. :. = 1. .... 1. Finally, the covariance between
,I
21
and
,I
21
is /
21
. This holds because
1
1
N
r
i

i
r
0
i

i
= 1
1
N

2
i
1
1
N
r
2
i
which holds because 1
1
N
r
it
r
ic

i&
= 1
1
N
r
it
r
ic
1
1
N

i&
for
t. :. . n = 1. .... 1.
Next, the joint limiting distribution of the tstatistics and the Hausman statistic are de-
rived. We rst assume that and o
2

are known and replace

and o
2

by and o
2

in the
test statistics. We then show that this modication does not matter asymptotically. The rel-
evant ingredients in the tstatistics and Hausman statistic are the expressions

.
i=1
r
0
i

1
n
i
.

.
i=1
r
0
i
`
1
T

i
.

.
i=1
r
0
i

1
r
i
. and

.
i=1
r
0
i
`
1
T
r
i
. where
(31) n
i
=
i
+ 1
T
j
i
.
We rst derive the appropriate normalizations of these expressions and their limit expressions.
We rst consider Case I in which case
2
1
= 0. Note that
(32)
1
= o
2

1
T
o1
T
1
0
T
. for o = o
2
j
o
2

(o
2

+o
2
j
1)
1
.
Then,

.
i=1
r
0
i

1
n
i
=

.
i=1
r
0
i

i
+

.
i=1
(r
0
i

1
1
T
j
i
1
1
N
r
0
i

1
1
T
j
i
+1
1
N
r
0
i

1
1
T
j
i
)
= o
2

.
i=1
r
0
i

i
1
2
o

.
i=1
r
i

i
+ (o
2

+o
2
j
1)
1
1

.
i=1
(r
i
j
i
1
1
N
r
i
j
i
+1
1
N
r
i
j
i
) (33)
and thus by (29)
(34) 1
1
o
1
a
i
o
1
j
(o
2

+o
2
j
1)`
12

.
i=1
r
0
i

1
n
i

o
(/
22
+/
1
22
)/
1
21

,I
21
/
22

,I
21
+
j,0
+/
1
.
Furthermore, because

.
i=1
r
0
i
`
1
T

i
=

.
i=1
(r
0
i

i
1r
i

i
).
(35) (1
1
N
[[r
i
[[
2
o
2

)
12
`
12

.
i=1
r
0
i
`
1
T

i

o

,I
21
/
21

,I
21
.
Also, because

.
i=1
r
0
i

1
r
i
=

.
i=1
(o
2

[[r
i
[[
2
o1
2
r
2
i
) it follows by a WLLN (using the last
two lines in (19)) and straightforward calculations that
(36) o
1
o
2
a
i
1
2
`
1

.
i=1
r
0
i

1
r
i

j
(/
2
22
+ 1)/
2
21
1.
Finally because

.
i=1
r
0
i
`
1
T
r
i
=

.
i=1
([[r
i
[[
2
1r
2
i
) it follows that
(37) (1
1
N
[[r
i
[[
2
)
1
`
1

.
i=1
r
0
i
`
1
T
r
i

j
1 /
2
21
.
15
Results (35), (37) and the continuous mapping theorem immediately imply that
(38) 1
11
(o
0
) = (o
2

`
1
.

i=1
r
0
i
`
1
T
r
i
)
12
`
12
.

i=1
r
0
i
`
1
T

i

o

11,I
~

,I
21
/
21

,I
21
(1 /
2
21
)
12
.
Note that the distribution
11,I
does not depend on /
1
and we extend the denition of
11,I
in (38) to the case [/
1
[ = .
To derive the limit distribution of 1
11
(o
0
). rst note that by partitioned regression, it follows
that
(39) `
12
(

o
11
o) =
`
12
[
.

i=1
r
0
i

1
n
i

i=1
r
0
i

1
1
T
(
.

i=1
1
0
T

1
1
T
)
1
.

i=1
1
0
T

1
n
i
]
`
1
[
.

i=1
r
0
i

1
r
i

i=1
r
0
i

1
1
T
(
.

i=1
1
0
T

1
1
T
)
1
.

i=1
1
0
T

1
r
i
].
By (34) and (36) the normalizations for the numerator and denominator in (39) are 1
1
o
1
a
i
o
1
j
(o
2

+
o
2
j
1) and o
1
o
2
a
i
1
2
, respectively. However, by straightforward calculations
(40)
o
1
o
2
a
i
1
2
`
1
[
.

i=1
r
0
i

1
1
T
(
.

i=1
1
0
T

1
1
T
)
1
.

i=1
1
0
T

1
r
i
] =
o
2

1o
2
j
(`
1
.

i=1
r
i
,o
a
i
)
2
= o
j
(1).
where the last equality holds because 1
1
o
2
j
o
2

/
2
22
and `
1

.
i=1
r
i
,o
a
i
= o
j
(1) by a
WLLNfor independent 1
1+c
bounded randomvariables and (19). Using an analogous argument
for the numerator in (39), it follows that
(41) 1o
a
i
o
j
o
2

`
12
(

o
11
o) = 1o
a
i
o
j
o
2

_
`
1
.

i=1
r
0
i

1
r
i
_
1
`
12
.

i=1
r
0
i

1
n
i
+o
j
(1).
Therefore, using (41), (34), and (36)
1
11
(o
0
) =
_
`
1
.

i=1
r
0
i

1
r
i
_
12
`
12
.

i=1
r
0
i

1
n
i
+o
j
(1)

11,I
~
(/
22
+/
1
22
)/
1
21

,I
21
/
22

,I
21
+
j,0
+/
1
((/
2
22
+ 1)/
2
21
1)
12
(1 +/
2
22
)
12
. (42)
It is easy to verify that 1
11
(o
0
)
o
`(0. 1) when /
1
= 0. Finally, using (34)(37) and straight-
forward calculations, the limiting distribution of the Hausman statistic under
.,I
with
16
[/
1
[ < is given by
H
.
=
`((
.

i=1
r
0
i
`
1
T
r
i
)
1
.

i=1
r
0
i
`
1
T

i
(
.

i=1
r
0
i

1
r
i
)
1
.

i=1
r
0
i

1
n
i
)
2
(`
1
.

i=1
r
0
i
`
1
T
r
i
,o
2

)
1
(`
1
.

i=1
r
0
i

1
r
i
)
1
+o
j
(1)
= (1`
1
.

i=1
r
2
i
)
1
o
2

+o
2
j
1
o
2

_
1
11
(o
2

`
1
.

i=1
r
0
i

1
r
i
)
12
1
11
(`
1
.

i=1
r
0
i
`
1
T
r
i
)
12
_
2
+o
j
(1)

1,I
~ (1 +/
2
22
)[
11,I
(/
2
21
/
2
22
(/
2
22
+ 1)
1
)
12

11,I
(/
2
21
1)
12
]
2
.
(43)
where the last step holds because
(1`
1
.

i=1
r
2
i
)
1
o
2

`
1
.

i=1
r
0
i

1
r
i
= (1`
1
.

i=1
r
2
i
)
1
`
1
.

i=1
r
0
i
r
i
1o
2
j
(o
2

+o
2
j
1)
1
.
(1`
1
.

i=1
r
2
i
)
1
`
1
.

i=1
r
0
i
`
1
T
r
i
= (1`
1
.

i=1
r
2
i
)
1
`
1
.

i=1
r
0
i
r
i
1. (44)
and because (1`
1
.

i=1
r
2
i
)
1
`
1
.

i=1
r
0
i
r
i

j
/
2
21
and 1o
2
j
(o
2

+ o
2
j
1)
1
/
2
22
(/
2
22
+ 1)
1
. It
follows from (43) that the limit distribution of H
.
is
(45)
2
1
(/
2
1
/(/ + 1)
1
) for / = /
2
22
(1 /
2
21
)
and thus H
.

o

2
1
if /
1
= 0.
In Case II, under sequences
.,I
for which [/
1
[ = the following limits hold jointly
(46)
_
1
11
(o
0
)
H
.
_

o
_

11,I

_
and thus in this case, with probability approaching 1, xed eect inference is conducted in
stage two and since
11,I
~ `(0. 1) it follows that the asymptotic rejection probability of the
twostage test equals c in this case.
In summary,
(47) 1

.
(o
0
)
o
J

I
.
under
.,I
. where J

I
is the distribution of the random variable
(48)

I
=
11,I
1(
1,I
_
2
1,1o
) +
11,I
1(
1,I

2
1,1o
)
and
11,I
.
11,I
. and
1,I
have been dened in (38), (42), and (43), respectively.
17
Dene J

I
. and [J

I
[ as the distribution of the random variable

I
and [

I
[. respectively.
For an upper onesided, lower onesided, and symmetric twosided test, dene
(49) J
I
= J

I
. J

I
. and [J

I
[.
respectively. The distribution J
I
depends on , but for notational simplicity, this dependence
is suppressed. The derivations above imply that Assumption B holds with : = 1,2.
To conclude the derivation of the asymptotic distribution of 1
.
(o
0
), it has to be veried
that replacing

and o
2

by and o
2

does not matter asymptotically. We only do so in Case I,

Case II can be dealt with analogously. For Case I, it is clear that it is sucient to show that
(`
1
.

i=1
r
0
i
(

)
1
r
i
)
1
`
1
.

i=1
r
0
i

1
r
i
= 1 +o
j
(1).
(`
12
.

i=1
r
0
i

1
n
i
)
1
`
12
.

i=1
r
0
i
(

)
1
n
i
= 1 +o
j
(1). and
o
2

,o
2

= 1 +o
j
(1). (50)
We verify (50) for the estimators o
2

. o
2
j
. and o
2

dened in Subsection 5.1. We only verify the

rst of the three conditions, the other conditions are veried analogously. To do so, note that
(51) (`
1
.

i=1
r
0
i
(

)
1
r
i
)
1
`
1
.

i=1
r
0
i

1
r
i
=
o
2

o
2

1
o
2

T
o
2
v
+o
2

T
1`
1
.

i=1
r
2
i
,(`
1
.

i=1
r
0
i
r
i
)
1
b o
2

T
b o
2
v
+b o
2

T
1`
1
.

i=1
r
2
i
,(`
1
.

i=1
r
0
i
r
i
)
and thus it is enough to show that
(52)
o
2

o
2

= 1 +o
j
(1) and that
o
2
j
o
2
j
= 1 +o
j
(1).
because using /
2
[i
2
. i
2
] for 0 < i
2
< i
2
< . the second condition in (52) implies o
2

o
2
j
1
/
2
22
= o
j
(1) and thus (52) implies ( o
2

+ o
2
j
1)
1
o
2
j
1 (o
2

+o
2
j
1)
1
o
2
j
1 = o
j
(1). We only show
the rst condition o
2

o
2

= 1 + o
j
(1). the second condition in (52) can be veried analogously.
For notational convenience, assume 1 = 0 in (26). By denition
(53) o
2

= o
2
&
o
2
j
= (`1)
1
.

i=1
T

t=1
n
2
it
(`1(1 1),2)
1
.

i=1
T1

t=1
T

c=t+1
n
it
n
ic
.
where
(54) n
it
=
it
n
0
it
(

`
O1S
.

o
O1S
)
0
= n
0
it
((`. o)
0
(

`
O1S
.

o
O1S
)
0
) +j
i
+
it
.
18
Multiplying out in (53), it follows that all the contributions with a j
i
factor cancel out. For the
contributions with only
it
factors we have (`1)
1

.
i=1

T
t=1

2
it
,o
2

= 1 + o
j
(1) by a WLLN
for independent 1
1+c
bounded random variables and (19) and
(55) (`1(1 1),2)
1

.
i=1

T1
t=1

T
c=t+1

it

ic
,o
2

= o
j
(1)
also by the WLLNbecause 1
1
N

it

ic
= 0. Finally, the terms involving n
0
it
((`. o)
0
(

`
O1S
.

o
O1S
)
0
)
components are negligible. For example, consider the cross term
_
`

`
O1S
o

o
O1S
_
0
_
2(`1)
1
.

i=1
T

t=1
n
it

it
(`1(1 1),2)
1
.

i=1
T1

t=1
T

c=t+1
(n
it

ic
+n
ic

it
)
_
,o
2

.
(56)
It is o
j
(1) using a WLLN for the mean zero vectors n
it

it
and n
it

ic
and by consistency of the
pooled OLS estimators. This concludes the proof of showing that replacing

and o
2

by and
o
2

does not aect the limiting distribution J

I
.
19
References
Andrews, D. W. K. & P. Guggenberger (2009a) Hybrid and sizecorrected subsampling meth-
ods. Econometrica 77, 721-762.
(2009b) Validity of subsampling and plugin asymptotic inference for parameters
dened by moment inequalities. Econometric Theory 25, 669-709.
(2010a) Asymptotic size and a problem with subsampling and with the m out of n
bootstrap. Forthcoming in Econometric Theory.
(2010b) Applications of subsampling, hybrid and sizecorrection methods. Forthcoming
in Journal of Econometrics.
(2010c) Incorrect asymptotic size of subsampling procedures based on postconsistent
model selection estimators. Forthcoming in Journal of Econometrics.
Baltagi, B. (2005) Econometric Analysis of Panel Data. John Wiley & Sons, Ltd, 3rd edition.
Banerjee, A. & L. Iyer (2005) History, institutions, and economic performance: The legacy of
colonial land tenure systems in India. American Economic Review 95, 11901213.
Bedard, K. & O. Deschnes (2006) The longterm impact of military service on health: ev-
idence from world war II and Korean war veterans. American Economic Review 96,
176194.
Blonigen, B.A. (1997) Firmspecic assets and the link between exchange rates and foreign
direct investment. American Economic Review 87, 447465.
Bradford, W.D. (2003) Pregnancy and the demand for cigarettes. American Economic Review
93, 17521763.
Gaynor, M., H. Seider, & W.B. Vogt (2005) The volumeoutcome eect, scale economies, and
learningbydoing. American Economic Review 95, 243247.
Guggenberger, P. (2010) The impact of a Hausman pretest on the size of a hypothesis test.
Forthcoming in Econometric Theory.
Hastings, J.S. (2004) Vertical relationships and competition in retail gasoline markets: Empir-
ical evidence from contract changes in southern California, American Economic Review
94, 317328.
20
Hausman J.A. (1978) Specication tests in Econometrics. Econometrica 46, 12511271.
Judge, G.G. & M.E. Bock (1978) The Statistical Implications of PreTest and Stein Rule
Estimators in Econometrics. NorthHolland, Amsterdam.
Wooldridge, J.M. (2002) Econometric Analysis of Cross Section and Panel Data. MIT Press,
Cambridge and London.
21
Tables and Figures
Table 1
Finite Sample (Null) Rejection Probabilities (in %) of a) Symmetric Twostage
Test and b) Hausman Pretest for ` = 100. 1 = 2. c = , = .05
a) o
2

= 5 o
2

= 1
j 0 .3 .4 .5 .6 0 .3 .4 .5 .6
.9 8.6 37.3 56.0 71.6 80.8 8.2 59.2 71.5 66.6 55.2
.6 7.7 32.7 45.9 53.3 51.6 7.0 35.0 28.3 15.2 7.2
.3 6.9 27.7 34.6 33.1 24.2 6.2 18.6 10.2 5.5 5.3
b)
.9 5.1 7.2 8.9 11.4 14.1 5.2 15.8 23.9 35.1 47.2
.6 5.2 15.3 23.7 33.9 45.7 5.4 44.9 69.6 87.4 96.8
.3 5.5 26.3 42.5 59.7 76.3 5.9 70.3 91.9 99.0 99.9
Table 3
Asymptotic Rejection Probabilities (in %) of Symmetric Twostage
Test Conditional on Pretest (Not) Rejecting c = , = .05. /
1
= 15
1 C 1 1 C `1 1(H
.

2
1,1o
)
/
21
/
22
.05 .1 .2 .3 .05 .1 .2 .3 .05 .1 .2 .3
.75 26.0 13.7 5.0 3.5 8.8 20.1 59.1 88.7 8.0 16.9 50.5 83.2
.8 31.0 16.5 5.8 3.5 9.3 22.3 64.8 92.5 7.4 14.8 43.3 75.9
.85 37.5 20.7 7.3 3.9 9.8 24.6 70.2 95.2 6.9 12.5 35.1 65.1
.9 46.6 27.7 10.0 5.1 10.4 27.0 75.2 97.1 6.3 10.1 25.8 49.6
.95 60.8 41.7 17.5 8.9 11.0 29.5 79.8 98.3 5.7 7.6 15.6 28.9
22