Applied Econometrics

ECON2046 & ECON2005

Week 6 Lecture 10
Inference IV: Testing hypotheses about multiple coefficients
Prof. Abigail Barr

[email protected]
Room B32 SCGB
Last week
• We looked at how to draw inferences about single coefficients by
formulating hypotheses and conducting tests

Earlier this week

• We looked at confidence intervals and learnt that they are closely
related to tests
• Then, we shifted our focus away from drawing inferences about single
coefficients, towards drawing inferences about groups of coefficients
ଶ
• Specifically, we looked at the coefficient of determination, , which is
a measure of the fit of a regression model to the data

Today
• We are going to look at F tests, which are used to test hypotheses
about groups of coefficients
The test
When we want to test a hypothesis about a single parameter or coefficient
• e.g. ଴: ଵ ; ଵ: ଵ
• we use the test

When we want to test a hypothesis relating to 2 or more parameters ( ଵ, ଶ,

ଷ…) at the same time
• we use the test, which is a multiple or joint hypothesis test

Most often, we want to test Null hypotheses of the form

଴: ଵ ଶ ଷ … or ଴: ଵ ଶ ଷ

• i.e., ଴: ଵ and ଶ and ଷ …

• i.e., there is no relationship between ଵ and , ଶ and , ଷ and …
The test
When we want to test a hypothesis about a single parameter or coefficient
• e.g. ଴: ଵ ; ଵ: ଵ
• we use the test

When we want to test a hypothesis relating to 2 or more parameters ( ଵ, ଶ,

ଷ…) at the same time
• we use the test, which is a multiple or joint hypothesis test

Most often, we want to test Null hypotheses of the form

଴: ଵ ଶ ଷ … or ଴: ଵ ଶ ଷ

• i.e., ଴: ଵ and ଶ and ଷ …

• i.e., there is no relationship between ଵ and , ଶ and , ଷ and …

A Null hypothesis of this form A Null hypothesis of this form

is often referred to as an is often referred to as a
Exclusion Restriction Set of Exclusion Restrictions
The test
When the Null hypothesis is
଴: ଵ ଶ ଷ or ଴: ଵ ଶ ଷ

• i.e., ଴: ଵ and ଶ and ଷ ,

the Alternative hypothesis is
ଵ: ଵ ଶ ଷ

• i.e., ଵ: ଵ and/or ଶ and/or ଷ

• i.e., there is a relationship between at least one of the variables ( ଵ ଶ,

and ଷ) and

We test ଴ by comparing two models

• one that embodies the exclusion restrictions - the restricted model
• one that does not - the unrestricted model
The test
Given the model,
଴ ଵ ଵ ଶ ଶ ଷ ଷ ௞ ௞

suppose we want to test whether the last regressors in the model have no
effect on
଴: ௞ି௤ାଵ ௞ି௤ାଶ ௞

the alternative is
ଵ: ௞ି௤ାଵ ௞ି௤ାଶ ௞

•i.e., ଵ: ௞ି௤ାଵ and/or ௞ି௤ାଶ and/or...and/or ௞

•i.e., there is a relationship between at least one of the variables ( ௞ି௤ାଵ

௞ି௤ାଶ, … and ௞ ) and

If ଴ is true, we can restrict the model (without affecting its trueness) to

଴ ଵ ଵ ଶ ଶ ଷ ଷ ௞ି௤ ௞ି௤

which has fewer parameters. To get this restricted model we apply the
exclusion restrictions listed in ଴
The test This is the unrestricted model
Given the model,
଴ ଵ ଵ ଶ ଶ ଷ ଷ ௞ ௞

suppose we want to test whether the last regressors in the model have no
effect on
଴: ௞ି௤ାଵ ௞ି௤ାଶ ௞

the alternative is
ଵ: ௞ି௤ାଵ ௞ି௤ାଶ ௞

•i.e., ଵ: ௞ି௤ାଵ and/or ௞ି௤ାଶ and/or...and/or ௞

•i.e., there is a relationship between at least one of the variables ( ௞ି௤ାଵ

௞ି௤ାଶ, … and ௞ ) and
This is the restricted model
If ଴ is true, we can restrict the model (without affecting its trueness) to

଴ ଵ ଵ ଶ ଶ ଷ ଷ ௞ି௤ ௞ି௤

which has fewer parameters. To get this restricted model we apply the
exclusion restrictions listed in ଴
The test
We test ଴: ௞ି௤ାଵ ௞ by comparing the fit of the restricted model
to the fit of the unrestricted model
More precisely, we compare how much of the variation in each model leaves
unexplained, i.e., each model’s sum of squared residuals,
To do this, we calculate an statistic (or ratio)
௥ ௨
௨
where
• ௥ sum of the squared residuals for the restricted model
• ௨ sum of the squared residuals for the unrestricted model
Essentially, the statistic is measuring the increase in when moving from
the unrestricted to restricted model
Since ௥ ௨, the statistic is always .
The statistic measures the increase in the when the restrictions set out
in ଴ are applied, with adjustments made for the number of restrictions and
degrees of freedom
The test
We test ଴: ௞ି௤ାଵ ௞ by comparing the fit of the restricted model
to the fit of the unrestricted model
More precisely, we compare how much of the variation in each model leaves
unexplained, i.e., each model’s sum of squared residuals,
To do this, we calculate an statistic (or ratio)
௥ ௨
௨

• If ଴ is true, ௥ ௨ and
• The larger the difference between ௥ and ௨ , the larger is
• A large indicates that the explanatory power of the restricted model is
poor relative to that of the unrestricted model, implying ଴ is false and
should be rejected
• has two sets of degrees of freedom
• Numerator number of restrictions imposed under ଴
• Denominator df in the unrestricted model
The test
To test ଴: ௞ି௤ାଵ ௞ we
• calculate an statistic (or ratio)
௥ ௨
௨

• If ଴ is true, this statistic has an distribution with and

degrees of freedom
• We reject ଴ when is larger that the critical value implied by our pre-
selected significance level,
• If we pre-selected 5% then the critical value, , that we use in our
decision rule is the 95th percentile of the ௤,௡ି௞ିଵ distribution
• The critical value , , depends on:
• significance level (5, 10 or 1% are usual)
• number of restrictions ( )
• degrees of freedom when estimating the unrestricted model ( )
• We reject ଴ if
The distribution

probability
• if we select 5% then the critical
value, , that we use in our decision
rule is the 95th percentile of the
௤,௡ି௞ିଵ distribution

• We reject ଴ if
The distribution

probability
• if we select 5% then the critical
value, , that we use in our decision
rule is the 95th percentile of the
௤,௡ି௞ିଵ distribution

• We reject ଴ if
• We fail to reject ଴ if
The distribution

probability
• if we select 10% then the critical
value, , that we use in our decision
rule is the 90th percentile of the
௤ǡ௡ି௞ିଵ distribution
 test for the significance of the regression
In a model with regressors, we may want to test
• ଴ ଵ ଶ ௞ do not help explain
• ଵ At least one of ଵ ଶ ௞ does help explain

This is equivalent to testing

• ଶ
଴
• ଶ
ଵ

Either way, this ଴ is a special set of exclusion restrictions

• ଴: ଵ ଶ ௞
• ଵ At least one of ଵ ଶ ௞

So, we can test the hypothesis by calculating the statistic which

• is referred to as the F statistic for the significance of the regression
• under the CLM assumptions, will have a ௞,௡ି௞ିଵ distribution
 test for the significance of the regression
଴: ଵ ଶ ௞ is very ‘pessimistic’
However, when ଶ is small, it is useful to have a formal test of whether the
model is explaining any of the variation in
The statistic for the significance of the regression and the corresponding
p value are routinely generated by econometric packages
In this case
௥ ௨
௨

But note that if we divide the top and bottom by ,

௥ ௨
௨

ଶ
and then use we get
ଶ ଶ ଶ
௥ ௨
ଶ ଶ
௨
 test for the significance of the regression
Number of restrictions

௥ ௨
௨

ଶ ଶ ଶ
௥ ௨
ଶ ଶ
௨

test for a set of exclusion restrictions,

Number of restrictions

௥ ௨
௨

ଶ ଶ
௥ ௨
ଶ
௨
The step-by-step guide to conducting tests
• Formulate the hypothesis (Null and Alternative)
• Select the significance level
• Find the critical value and use it to set up the Decision Rule
• Calculate the test statistic, i.e., the statistic
• Conduct the test, i.e., apply the Decision Rule, i.e., compare the
statistic and the critical value

• Draw conclusion
 test: An example
Are US baseball players’ salaries affected by their performance?

଴ ଵ ଶ ଷ ସ ହ

annual salary in USD

years in league
average number of games played per year
career batting average
number of home runs per year
number of runs batted per year

Formulate the hypothesis that corresponds to the question and the model
 test: An example
Are US baseball players’ salaries affected by their performance?

଴ ଵ ଶ ଷ ସ ହ

annual salary in USD

years in league
average number of games played per year
career batting average
number of home runs per year
number of runs batted per year

Formulate the hypothesis that corresponds to the question and the model
଴: ଷ ସ ହ , i.e., performance has no effect on salary
ଵ: ଷ and/or ସ and/or ହ , i.e., at least one aspect of
performance has an effect on salary
Test required: a joint test of three exclusion restrictions, i.e., an test
Selected significance level:
The step-by-step guide to conducting tests
• Formulate the hypothesis (Null and Alternative)
• Select the significance level
• Find the critical value and use it to set up the Decision Rule
• Calculate the test statistic, i.e., the statistic
• Conduct the test, i.e., apply the Decision Rule, i.e., compare the
statistic and the critical value

• Draw conclusion
 test: An example ଴ ଵ ଶ ଷ ସ ହ

Applying OLS yields:

ଶ
 test: An example
Applying OLS yields:
distribution
ଷ,ଷସ଻

଴: ଷ ସ ହ , i.e., performance has no effect on salary

The restricted model is ଴ ଵ ଶ

Applying OLS to this yields:

ଶ
23
 test: An example
Applying OLS yields:
distribution
3, 347

଴: ଷ ସ ହ , i.e., performance has no effect on salary

The restricted model is ଴ ଵ ଶ

Applying OLS to this yields:

ଶ
 test: An example
Applying OLS yields:
distribution
3, 347

଴: ଷ ସ ହ , i.e., performance has no effect on salary

The restricted model is ଴ ଵ ଶ

Applying OLS to this yields:

ଶ
 test: An example
Applying OLS yields:
distribution
3, 347

଴: ଷ ସ ହ , i.e., performance has no effect on salary

The restricted model is ଴ ଵ ଶ

Applying OLS to this yields:

ଶ
 test: An example
Applying OLS yields:
distribution
3, 347

଴: ଷ ସ ହ , i.e., performance has no effect on salary

The restricted model is ଴ ଵ ଶ

Applying OLS to this yields:

ଶ
 test: An example
Applying OLS yields:
distribution
3, 347

଴: ଷ ସ ହ , i.e., performance has no effect on salary

The restricted model is ଴ ଵ ଶ

Applying OLS to this yields:

ଶ
 test: An example ଴ ଵ ଶ ଷ ସ ହ

Applying OLS yields:

଴: ଷ ସ ହ , i.e., performance has no effect on salary

The restricted model is ଴ ଵ ଶ

Applying OLS to this yields:

So,
௥ ௨
௨
 test: An example
Applying OLS yields:
distribution
3, 347

Reject ૙ in favour of ଵ at the 5% significance

level (result would have
ଶ been the same if we’d
pre-selected the 1% significance level)
଴: ଷ ସ ହ , i.e., performance has noPerformance
Conclusion: effect on salary
does affect salary

The restricted model is ଴ ଵ ଶ

Applying OLS to this yields:

So,
௥ ௨
௨
 test: An example
Applying OLS yields:

଴: ଷ ସ ହ , i.e., performance has no effect on salary, is rejected

Conclusion: Performance does affect salary…

Even though:
• the coefficient on is insignificant;
• the coefficient on is insignificant; and
• the coefficient on is insignificant

Apparent disagreement between tests and test is a sign of multicollinearity,

i.e., high correlation between , , and
 tests, tests and the Ballentine Venn
଴ ଵ ଵ ଶ ଶ

Variation in Variation in

Variation
in ଵ Variation
in ଵ

Variation in ଶ Variation in ଶ
ଶ
Across these two cases and hence significance of the regression ( stat) are similar
Multicollinearity:
relatively low relatively high
Lots of unique information available Little unique information available
to estimate ଵ and ଶ => t stats large to estimate ଵ and ଶ => t stats small
 tests vs. tests
• When stats are low and a corresponding stat is high it is a sign of
multicollinearity

• When groups of regressors are co-linear, an test may be superior to a set

of individual t tests

• When a group of regressors are all measuring closely related dimensions

of the same thing and are co-linear (as a consequence), an test is
superior to a set of individual t tests
 test: An example
Applying OLS yields:

଴: ଷ ସ ହ , i.e., performance has no effect on salary, is rejected

Conclusion: Performance does affect salary…

Even though:
• the coefficient on is insignificant;
• the coefficient on is insignificant; and
• the coefficient on is insignificant

Apparent disagreement between tests and test is a sign of multicollinearity,

i.e., high correlation between , , and
 tests vs. tests
• When stats are low and a corresponding stat is high it is a sign of
multicollinearity

• When groups of regressors are co-linear, an test may be superior to a set

of individual t tests

• When a group of regressors are all measuring closely related dimensions

of the same thing and are co-linear (as a consequence), an test is
superior to a set of individual t tests

• However, it is important that the test is appropriate given the hypothesis

Finally,…
• The test is very versatile and is used in many, varied situations