Principle Foundations Home Page
|
|||||
Test of Significance (Statistical Inference) |
|||||
In order to
test for the statistical significance of the parameter estimates of the
regression, we have firstly to know a few basic things about
. The most notable, are the following: Statistical inference Testing the
significance of β1 Example
1: Suppose we have the population relationship of the type Yi = α + βXi + εi for all i,
where the dependent variable is linearly dependent on the
explanatory X variable, but is also influenced by the disturbance ε. Thus, by considering this
population relationship, if β=0 in this equation then
X does not influence Y, which then has an expected value of α, from
which it can only be disturbed by a non-zero ε. If
, then this means that X does influence Y. We can therefore test whether X
influences Y by setting up the null hypothesis H0: β=0 and
testing it against an alternative hypothesis HA:
. We can obtain a test statistic for this purpose. Thus, we apply the
formula
, which has a Student's t distribution with n-2 d. f (degrees of freedom). We can
therefore use a
(1.7) as a test
statistic and reject the null hypothesis β=0 (X
does not influence Y) if the absolute value of this test statistic exceeds
the relevant critical value taken from Student's t tables. Effectively
what we do is to consider whether the OLS estimate
is
sufficiently different from zero for us to reject the null hypothesis that
the true β is non-zero. Dividing
by its estimated standard
error enables us to use Student's t tables to decide what is meant by
"sufficiently different".
Example Again taking the example of the household consumption function H0: implies that a household's income does not influence its consumption. Since if β is non-zero we expect β>0 in this case, we employ an upper tail test. That is, the alternative hypothesis is HA: β>0. Taking the 0.05 level of significance, the critical t value is t0.05= 1.714 (d.f = n-2). We suppose we have n=25 (hence 23 d.f), and , so that the test statistic (1.7) takes a value of 7.19. This clearly exceeds the critical value , so we can reject the null hypothesis β=0 at the 0.05 level of significance and conclude that household income influences consumption. It is possible to test a null hypothesis H0: α=0 in a similar manner. Under
this null hypothesis we have that
has a Student's t
distribution with n-2 d.f. This
may therefore be used as a test statistic and we reject H0 if
its absolute value exceeds the relevant critical t value. Once the
estimated standard errors
and
have been computed, it
clearly is a simple matter to test the null hypothesis α=0 and β=0. The test statistics used,
and
are referred to as t ratios. It is customary to present sample regression
equations with these ratios placed in parenthesis underneath the estimates
and
. That is, we
present
=
+
Example 3:
There are
occasions when we may wish to test hypothesis other than β=0 and α=0. Suppose we wished to test the
hypothesis that the parameter β takes some non-zero value β*. Under H0:β=β* and we set the required
test statistic which is
. Under H0 a Student's
t distribution with the usual n-2 degrees of freedom. For example, in a
four-person household consumption function, we might wish to test whether
the MPC b of households was less than unity. We can set up a null
hypothesis H0: β=1 and test it using the test statistic
. Since the
alternative has to be HA: β<1, we use a lower tail
test. (In the alternative hypothesis we set β<1 because in this
case b represents MPC, and as we know its value varies from 0 to 1,
something which implies that b cannot take a value greater than 1). Under
H0, the test statistic has a Student's t distribution, so we expect its
value to be close to zero. Since we will reject H0 if the OLS
estimate
is sufficiently different
from unity, we reject if the test statistic is sufficiently different from
0. Given an one-tail test and taking the 0.05 level of significance, with
n-2 d.f, the relevant critical value is t= 1.714. We reject H0
if the absolute value (or calculated value) of the test statistic exceeds
this value. We suppose that for our sample we know that
and
. Hence, the test statistic takes a value
.Since the absolute value of the
test statistic does not in this case exceed the critical t value, we
cannot reject the null hypothesis β=1. So we must conclude that on
the bases of the present sample, there is not enough evidence to say that
the MPC of households is less than unity.
Copyright
© 2002
Back
to top Evgenia Vogiatzi <<Previous Next>> |