ECO441K Introduction to Econometrics

Chapter 4: The Multiple Regression Model — Inference

University of Texas

(University of Texas) Chapter 4: Inference 1 / 53

Lecture 10 — Outline

The Multiple Regression Model — Inference I

1. Sampling Distribution of OLS Estimators
2. Testing hypotheses about 1 population parameter (t tests and
p-values)
3. Confidence Intervals

(University of Texas) Chapter 4: Inference 2 / 53

What is inference?

log(wage ) = β0 + β1 educ + β2 exper + u

We’ve learned how to estimate parameters using OLS, e.g.:

log\
(wage ) = β̂0 + β̂1 educ + β̂2 exper
β̂0 = .217, β̂1 = .0979, β̂2 = .0103

Now we focus on specific hypotheses we can formally test using

statistical inference, e.g.
Is β1 = 0? (Does education have zero effect on wage?)
Is β1 = β2 ? (Does an extra year of education increase wage the same
amount as an extra year of experience?)

(University of Texas) Chapter 4: Inference 3 / 53

What is inference?

We can’t answer these questions using only the estimates β̂1 , β̂2 !
Why? because β̂1 and β̂2 are random variables that depend on our
random sample.
Even if β1 = 0, it is possible for β̂1 = .0979 in a particular sample
(actually, if β1 = 0, it is exceedingly unlikely to get a β̂1 of exactly 0)
So to make inferences and statistically test hypotheses, we will need
to calculate things like:
“If the true β1 is zero, how likely is it that we would get a β̂1 of at
least .0979?” (i.e. to get β̂1 this far away from zero (note the absolute
value))
If it is very unlikely, we will reject the null hypothesis that β1 = 0.

(University of Texas) Chapter 4: Inference 4 / 53

What is inference?

Or, we might want to calculate

”If β1 − β2 = 0 (i.e. β1 = β2 ), how likely is it that we would get
a β̂1 − β̂2 of at least 0.0876 (= 0.0979 − 0.0103).” (i.e. to get
β̂1 − β̂2 this far away from zero)
Again, if it is very unlikely, we will reject the null hypothesis that
β1 = β2 .
To answer these questions, we need to know the entire sampling
distribution of β̂1 (and β̂2 ). In other words, we need to know more
than just the means and variances of β̂1 and β̂2 to do this.

(University of Texas) Chapter 4: Inference 5 / 53

One More Assumption

The sampling distribution of β̂’s depends on the underlying

distribution of the errors (u). For now we will make the additional
assumption that the u are normally distributed in the population.
Assumption MLR.6 (Normality) The population error u is independent of
the explanatory variables x and is normally distributed with mean 0 and
variance σ2 :
u ∼ Normal (0, σ2 )

Note that MLR.6 implies MLR.4 (ZCM) and MLR.5

(Homoskedasticity)
MLR.6 is a quite strong assumption, often unrealistic
But in a couple of lectures we will show we can drop MLR.6 and
today’s results are still good approximations.

(University of Texas) Chapter 4: Inference 6 / 53

Classical Linear Model (CLM), graphically

(University of Texas) Chapter 4: Inference 7 / 53

Classical Linear Model (CLM)

Together Assumptions MLR.1-MLR.6 are known as the classical

linear model (CLM) assumptions
Under these assumptions the OLS estimators are the best (minimum
variance) unbiased estimators.
This is a little stronger than the Gauss-Markov result from prior lecture
(since we are not limiting ourselves to estimators which are linear in
the yi )

(University of Texas) Chapter 4: Inference 8 / 53

Normal Sampling Distribution

The normality assumption (MLR.6) implies that the OLS estimators are
themselves normality distributed! (intuition), i.e.

Theorem (Normal Sampling Distribution of β̂j ) Under the CLM

assumptions (MLR.1-MLR.6),

β̂j ∼ Normal ( βj , Var ( β̂j )) (1)

where Var ( β̂j ) is given by the formula from the previous lecture notes.

(University of Texas) Chapter 4: Inference 9 / 53

Normal Sampling Distribution

As a result, we can standardize (de-mean and divide by the stdev) to

get:
( β̂j − βj )
∼ Normal (0, 1) (2)
sd ( β̂j )

Problem: Recall from notes that we don’t know sd ( β̂j ) since we don’t
know σ2

(University of Texas) Chapter 4: Inference 10 / 53

t Distribution because sd ( β̂j ) is not known
Theorem (t Distribution for Standardized Estimators) Under the CLM
assumptions (MLR.1-MLR.6),

( β̂j − βj )
∼ tn−k −1 (3)
se ( β̂j )

Note that (3) is different from (2) — se ( β̂j ) vs sd ( β̂j ) (σ̂2 vs σ2 )

This replacement changes the distribution from a normal distribution
to a t distribution (with n − k − 1 degrees of freedom).
Facts about t distributions:
(i) they are symmetric about zero (bell-shaped like a normal),
(ii) they have fatter tails than the Normal (0, 1) (the “standard normal
distribution”), and
(iii) as n gets large, tn−k −1 starts looking more and more like the
Normal (0, 1).
(University of Texas) Chapter 4: Inference 11 / 53
t Distribution

(University of Texas) Chapter 4: Inference 12 / 53

Testing Hypotheses about a Single Parameter

Consider testing whether a single parameter βj is equal to some number a.

The null hypothesis:
H0 : β j = a
The (two-sided) alternative hypothesis:

H1 : β j ̸ = a

The specific case of a = 0 is often of interest:

H0 : βj = 0 (xj does not belong in the MLR model)

H1 : βj ̸= 0 (xj does belong; is a “statistically significant” determinant of y )

(University of Texas) Chapter 4: Inference 13 / 53

Example

Example: Wages

Consider the model:

log(wage ) = β0 + β1 educ + β2 exper + β3 tenure + u

The null hypothesis H0 : β2 = 0 means that once education and tenure
have been accounted for, the number of years of labor-market experience
has no effect on the wage. If H1 : β2 ̸= 0, then prior work experience does
have an effect on wages (ceteris-paribus).

(University of Texas) Chapter 4: Inference 14 / 53

Testing Hypotheses about a Single Parameter

Two Important Points:

1 First, these are statistical hypothesis tests. Since there is inherent
”statistical noise” due to random sampling, statistical tests are not
going to be 100% accurate.
For example, even if β2 = 0, there is some small probability that we
draw a really crazy sample and b β2 is extremely large. In this case we
might incorrectly conclude that the null hypothesis that β2 = 0 is not
true.
However, statistical theory will allow us to precisely quantify the
possibility of making such errors. In particular, by choosing what is
called a ”significance level”, we can ”control” the probability that we
will incorrectly reject the null hypothesis H0 when it is actually true.

(University of Texas) Chapter 4: Inference 15 / 53

Testing Hypotheses about a Single Parameter

1 We will be covering three equivalent methods of doing a hypothesis

test
1) Using a t-statistic (t-stat)
2) Using a p-value
3) Using a confidence interval
These are all equivalent - they look different because the decision rule
is applied w.r.t. differently transformed metrics.
This makes things confusing, so why do we teach all three?
.....because unfortunately, all three methods get used in different
situations by different researchers in different fields, so it is important
to be familiar with all the methods.

(University of Texas) Chapter 4: Inference 16 / 53

The t-test

First, lets pretend that the null hypothesis is true (H0 : βj = a).

If this is the case (we sometimes say ”under H0 ”), then based on the
above Theorem, we have

( β̂j − a)
t∗ ≡ ∼ tn−k −1 .
se ( β̂j )

t ∗ called the t-statistic (or the t-stat). The t-statistic tells us how many
standard errors the estimator β̂j is from the hypothesized value a.

Note: The t-statistic is itself a random variable because β̂j is a random

variable (as is se ( β̂j )).

(University of Texas) Chapter 4: Inference 17 / 53

The t-test: a = 0 case

Testing the null hypothesis H0 : βj = 0 is very common. In this case a = 0

and the t-stat is simply
β̂j
t∗ ≡ ,
se ( β̂j )

i.e. it tells how ”far” β̂j is from zero (measured in terms of number of
standard errors).

Stata reports the t-statistic for H0 : βj = 0 for each estimate

β̂0 , β̂1 , . . . , β̂k (Column titled ”t”)

(University of Texas) Chapter 4: Inference 18 / 53

The t-test procedure
Steps for a (two-sided) t-test (H0 : βj = a, H1 : βj ̸= a):
( β̂j −a)
1 Compute t ∗ ≡ se ( β̂j )
.
2 Note that |t ∗ | is what is relevant, since, e.g. t ∗ = 4 and t ∗ = −4
both involve β̂j being 4 se ( β̂j )’s away from a
3 Pick a significance level (siglevel, e.g. 0.05 (5%)) for the test. (The
significance level tells us how often we will incorrectly reject the null
hypothesis H0 when it’s actually true.). This determines the critical
value for the test, tn−k −1,siglevel /2 , e.g. tn−k −1,0.025
4 Rejection rule:
Reject H0 if |t ∗ | > tn−k −1,0.025 .
Do not reject H0 if |t ∗ | < tn−k −1,0.025 . (NEVER say “accept H0 ”!)
Note: The critical value tn−k −1,0.025 is the value for which a tn−k −1
distributed random variable has only a 2.5% chance of being above
(and because symmetric, −tn−k −1,0.025 is the value at which it has
only a 2.5% chance of being below).
(University of Texas) Chapter 4: Inference 19 / 53
The t-test procedure — graphically

(University of Texas) Chapter 4: Inference 20 / 53

Where do we get t-distribution Critical Values
To get the critical value corresponding to any significance level (siglevel)
we can use Stata: disp invttail(n − k − 1, siglevel /2) ), e.g.
If you want significance level 0.05 (5%) and n − k − 1=15, then the
critical value tn−k −1,0.025 = 2.131 (= disp invttail(15,0.025)) so we
reject H0 when |t ∗ | > 2.131.
Intuition: We are rejecting H0 when |t ∗ | > 2.131 because if
|t ∗ | > 2.131, the probability of getting β̂j ”so far” away from a
(when H0 is true, i.e. βj = a) is only 5%
If you want significance level 0.01 (1%) and n − k − 1=15, then the
critical value tn−k −1,0.005 = 2.947 (= disp invttail(15,0.005)) so we
reject H0 when |t ∗ | > 2.947.
Intuition: We are rejecting H0 when |t ∗ | > 2.947 because if
|t ∗ | > 2.947, the probability of getting β̂j ”so far” away from a
(when H0 is true, i.e. βj = a) is only 1%
Think about why siglevel controls how often we incorrectly reject the
null hypothesis H0 when it is true. We get to choose it!
(University of Texas) Chapter 4: Inference 21 / 53
p-value for a t-test

The p-value associated with a (two-sided) t-test is defined as

p − value ≡ Pr(tn−k −1 ≥ |t ∗ |).

The p-value directly tells us how likely it would be to see a random draw
from a tn−k −1 distribution be larger in magnitude than the observed
t-statistic. (In other words, it tells us whether the t ∗ value ”unusually”
large?)

(University of Texas) Chapter 4: Inference 22 / 53

p-value for a t-test — graphically

(University of Texas) Chapter 4: Inference 23 / 53

p-value for a t-test

The p-value is particularly convenient because it tells us whether or not we

would reject H0 at any chosen significance level.

Rejection rule: For any chosen significance level ℓ, if ℓ > p-value you
reject H0 at that significance level ℓ
e.g. if p-value = 0.086, you would reject H0 at significance level 0.10,

but you would not reject H0 at significance level 0.05 or 0.01.

Stata also reports the p-value associated with H0 : βj = 0 (Column
entitled ”P > |t |”)

How do we get a p-value when a ̸= 0?

First calculate t ∗ . Then Stata: disp ttail(n − k − 1, t ∗ ) and multiply
by 2 (because 2 sided p-value)
[Link]

(University of Texas) Chapter 4: Inference 24 / 53

Example
Test whether the return to education, after controlling for experience and
tenure, in population is 0 against the alternative it is not equal to 0.
Testing:

H0 : βeduc = 0
H1 : βeduc ̸= 0

Test statistic is:

β̂educ 0.092029
t∗ = =
se ( β̂educ ) 0.00733
= 12.56

For a test at the 1% level of significance, the critical value is 2.576

(n − k − 1 = 526 − 3 ≈ ∞) and since |t ∗ | exceeds this value, we reject
the null in favor of the alternative hypothesis (that the return to education
is not equal to 0).
(University of Texas) Chapter 4: Inference 25 / 53
Examples
2. Housing Prices and Air Pollution.
Using a sample of 506 communities in a large US city, a model was
estimated which related median house price in the community (price) to
various community characteristics: nox is the amount of nitrous oxide in
the air (in part per million), dist is the average distance of the community
from 5 employment centres, rooms is the average number of rooms in
houses in the community, and stratio is the average student-teacher ratio
in the community. The population model is:

log(price ) = β0 + β1 log(nox ) + β2 log(dist ) + β3 rooms + β4 stratio + u

Hence β1 is the elasticity of price with respect to nox. We wish to test

H0 : β 1 = − 1
H1 : β 1 ̸ = − 1
( β̂1 +1)
The t-statistic for doing this test is t ∗ = se ( β̂1 )
(University of Texas) Chapter 4: Inference 26 / 53
Example

Using the data, the estimated model is:

log\
(price ) = 11.08 − 0.954 log(nox ) − 0.134 log(dist )
(0.32) (0.117) (0.043)
+0.255rooms − 0.052stratio
(0.019) (0.006)
2
n = 506, R = 0.581

Each slope coefficient is significantly different from 0 even at very small

significance levels
But we want to test H0 : β1 = −1 against H1 : β1 ̸= −1.
For this test, the t-statistic is (−0.954 + 1)/0.117 = 0.393, which is very
small in magnitude. We would not reject H0 at any conventional
significance level (such as 5%, where the critical value is 1.96). There is
little evidence that the elasticity is different from −1.
(University of Texas) Chapter 4: Inference 27 / 53
Analogous p-value approach

In Example 1, t ∗ = 12.56 with n − k − 1 ≈ ∞.

Corresponding p-value is <0.0001(from Stata). If βeduc = 0, then
with a sample of this size there is almost zero (<0.0001) probability of
getting b
βeduc as far away from zero as 0.092029.
Hence, we would reject H0 at any of the conventional significance
levels.
In Example 2, for H0 : β1 = −1, t ∗ = 0.393 with n − k − 1 ≈ ∞.
Corresponding p-value is 0.3472. If β1 = −1, then with a sample of
this size there is about a 35% chance of getting b
β1 as far away from
−1 as −0.954.
With this high a probability we would not reject H0 at any of the
conventional significance levels.
What about Example 2 but a test of H0 : β1 = 0?

(University of Texas) Chapter 4: Inference 28 / 53

Economic vs Statistical Significance

Economic vs Statistical Significance

We have been focusing on statistical significance, which is determined
by the t-statistic (or p-value).
However, the economic (or practical) significance also depends on
the magnitude of β̂j itself.
Too much focus on statistical significance can lead to the false
conclusion that a variable is important for explaining y even though
its estimated effect is very small.
Make sure you consider this.

(University of Texas) Chapter 4: Inference 29 / 53

Confidence Intervals
Recall that MLR.1-MLR.6 imply

( β̂j − βj )
∼ tn − k − 1 .
se ( β̂j )

We know that
!
( β̂j − βj )
Pr −tn−k −1,0.025 < < tn−k −1,0.025 = 0.95.
se ( β̂j )

Re-arranging to isolate βj yields

 
Pr β̂j − tn−k −1,0.025 · se ( β̂j ) < βj < β̂j + tn−k −1,0.025 · se ( β̂j ) = 0.95.

We say that the 95% confidence interval for β̂j is given by

β̂j ± tn−k −1,0.025 · se ( β̂j ).

(University of Texas) Chapter 4: Inference 30 / 53

Confidence Intervals

Interpretation of the 95% confidence interval (CI): If random samples were

repeatedly obtained (with a confidence interval computed each time), the
population value βj would lie in the confidence interval for 95% of the
samples.

95% is not the only confidence interval! Just change the critical value to
get any confidence interval you would like:

90% confidence interval: β̂j ± tn−k −1,0.05 · se ( β̂j )

80% confidence interval: β̂j ± tn−k −1,0.10 · se ( β̂j )

(University of Texas) Chapter 4: Inference 31 / 53

Confidence Intervals

The link between a confidence interval and a two-sided t-test: We reject

H0 : βj = a at the 0.05 (5%) significance level if and only if a is not
contained in the 95% CI. (Similar for other levels.)

For a 95% confidence interval with large n, note that the critical value is
1.96 (close to 2), so we have

β̂j ± 1.96 · se ( β̂j ) ≈ β̂j ± 2 · se ( β̂j )

There are also one-sided hypothesis tests (eg Eco 329). Same basic
idea as two-sided tests but not as commonly used and we will not
cover in this course.
More examples (lect10 [Link])

(University of Texas) Chapter 4: Inference 32 / 53

Lecture 11 — Outline

Multiple Linear Regression Model: Inference II

1 Testing single linear combination of parameters
2 Testing multiple linear restrictions (The F test)

(University of Texas) Chapter 4: Inference 33 / 53

Single Linear Combination of Parameters

Consider a single hypothesis involving more than one parameter.

Example: Wages

Consider the standard wage equation

log(wage ) = β0 + β1 educ + β2 exper + u (4)

Suppose we are interested in determining whether another year of formal

education (educ) has the same effect on the log (wage ) as another year of
general workforce experience (exper ).

(University of Texas) Chapter 4: Inference 34 / 53

Single Linear Combination of Parameters

The null and alternative hypotheses are

H0 : β 1 = β 2 (5)
H1 : β 1 ̸ = β 2

Re-write the hypotheses as:

H0 : β 1 − β 2 = 0 (6)
H1 : β 1 − β 2 ̸ = 0

(University of Texas) Chapter 4: Inference 35 / 53

Single Linear Combination of Parameters

The t-statistic is based on whether β̂1 − β̂2 is sufficiently different

from 0 to reject H0 .
Linear combinations of variables that are normally distributed also
have normal distributions. Hence we can proceed as prior lecture.
Standardize by dividing by the standard error of the difference:

( β̂1 − β̂2 )
t∗ = (22)
se ( β̂1 − β̂2 )

But how do we calculate se ( β̂1 − β̂2 )?

(University of Texas) Chapter 4: Inference 36 / 53

Single Linear Combination of Parameters

Note: se ( β̂1 − β̂2 ) ̸= se ( β̂1 ) − se ( β̂2 )

Instead:

Var ( β̂1 − β̂2 ) = Var ( β̂1 ) + Var ( β̂2 ) − 2Cov ( β̂1 , β̂2 ),
d ( β̂ − β̂ ) = Var
Var d ( β̂ ) + Vard ( β̂ ) − 2Covd ( β̂ , β̂ ),
1 2 1 2 1 2
q
and se ( β̂1 − β̂2 ) = d ( β̂ − β̂ )
Var 1 2

The lincom command in Stata can calculate this for us (wage

example).

(University of Texas) Chapter 4: Inference 37 / 53

Single Linear Combination of Parameters

An alternative is to estimate a ”different” model that directly

provides the se.
Define a ”new” parameter θ ≡ β1 − β2 and rewrite the model as

log(wage ) = β0 + (θ + β2 )educ + β2 exper + u (7)

= β0 + θeduc + β2 (educ + exper ) + u

This means that we can alternatively regress log(wage ) on educ and

the ”new” variable (educ + exper )
This produces estimates bθ and bβ2 (instead of b β1 and b
β2 ) - but it is
essentially the ”same thing” since the bβ2 ’s are numerically the same
θ=b
and b β −b β .
1 2

(University of Texas) Chapter 4: Inference 38 / 53

Single Linear Combination of Parameters

The convenient thing about this alternative ”equivalent”

representation of the model is that it directly gives both θ̂ and the
se (θ̂ ) as part of the Stata OLS output.
Thus the t-statistic and p-value for testing

H0 : θ = 0
H1 : θ ̸ = 0

are also provided, i.e.

θ̂
t=
se (θ̂ )

(University of Texas) Chapter 4: Inference 39 / 53

Multiple Linear Restrictions: The F Test

Testing Joint Significance of a Set of Variables

Suppose that you want to test whether a subset of the x variables belongs
in the model (once you’ve controlled for the other x variables in the
model).
Example: Parents’ Education and Child Birthweight
Consider the following model to explain child birthweight:

bwght = β0 + β1 cigs + β2 parity + β3 faminc + (8)

β4 motheduc + β5 fatheduc + u

where bwght = birthweight

cigs = average # cigs the mother smoked per day during pregnancy
parity = birth order
faminc = family income
motheduc = years of education for mother
fatheduc = years of education for father
(University of Texas) Chapter 4: Inference 40 / 53
Multiple Linear Restrictions

Suppose interested in testing:

H0 : β4 = 0, β5 = 0 (9)
H1 : β4 ̸= 0 or β5 ̸= 0

If H0 is true, then motheduc and fatheduc should be excluded from the

model after the other variables have been controlled for.

(University of Texas) Chapter 4: Inference 41 / 53

Multiple Linear Restrictions

This is an example of multiple restrictions because we are putting

more than one restriction on the model parameters. The associated
test is called a joint hypothesis test.
The alternative hypothesis holds if β4 or β5 is non-zero. The result of
the test will not, however, tell us which one is non-zero.
We will not be testing this joint hypothesis by considering the
individual t-statistics on the coefficients on motheduc and fatheduc.
It would be statistically incorrect to do this.
Instead we will develop a test based on how SSR (or the R 2 statistic)
changes when the model is estimated with the restrictions of H0
imposed (the restricted model) as compared to no restrictions
imposed (the unrestricted model).

(University of Texas) Chapter 4: Inference 42 / 53

Multiple Linear Restrictions

Terminology
The full model with all of the original parameters is known as the
unrestricted model:

bwght = β0 + β1 cigs + β2 parity + β3 faminc + (10)

β4 motheduc + β5 fatheduc + u

The restricted model is obtained by imposing the restrictions in H0 :

bwght = β0 + β1 cigs + β2 parity + β3 faminc + u (11)

We know that SSR for (11), called SSRr , will be larger than SSR for (10),
called SSRur . Similarly, we know that Rr2 (R-squared restricted) is smaller
2 (R-squared unrestricted).
than Rur

(University of Texas) Chapter 4: Inference 43 / 53

Multiple Linear Restrictions

Need a test statistic whose distribution under the null H 0 is known so

the significance level can be set
=⇒ for this example, the appropriate test statistic is:
(SSRr − SSRur )/2
F =
SSRur /(n − k − 1)
2 − R 2 ) /2
(Rur r
= 2 ) / (n − k − 1)
∼ F2,n−k −1
(1 − Rur

where F represents the ”F distribution” (note two different degrees of

freedom values).

(University of Texas) Chapter 4: Inference 44 / 53

Multiple Linear Restrictions

Running OLS on both models we get Rur 2 = 0.0387, R 2 = 0.0364,

r
and n − k − 1 =1191-5-1=1185. Hence F = 1.42. Since the 10%
critical value for F2,∞ = 2.3 (see table), we do not reject H0 .
Note that we would have rejected H0 if F had been sufficiently large,
2
which corresponds to a situation where the difference between Rur
2
and Rr is bigger, i.e. a situation where adding motheduc and
fatheduc increase R 2 by more.
This is intuitive - i.e. if adding more variables increases R 2 by more, it
is stronger evidence that they do enter the model (i.e. determine y )

(University of Texas) Chapter 4: Inference 45 / 53

General F Test

General Case:
In general, the unrestricted model has k independent variables:

y = β0 + β1 x1 + · · · + βk xk + u

Suppose we want to test the null hypothesis that q of the k variables

have coefficients equal to 0 (for notation, assume these are the last q
variables in the list of explanatory variables):

H0 : βk −q +1 = 0, . . . , βk = 0

The alternative is
H1 : H0 is false

(University of Texas) Chapter 4: Inference 46 / 53

The F Test

When we impose the restrictions under H0 , the restricted model is

then
y = β0 + β1 x1 + · · · + βk −q xk −q + u
and the F -statistic (or F ratio) is now given by

(SSRr − SSRur )/q

F = ∼ Fq,n−k −1
SSRur /(n − k − 1)

Under the null hypothesis (assuming the CLM assumptions are true),
the F -statistic is distributed as an F random variable with
(q, n − k − 1) degrees of freedom (recall that q = number of
restrictions).
The F -statistic measures the relative increase in the SSR as we move
from model ur to r (and this will always be positive)

(University of Texas) Chapter 4: Inference 47 / 53

F and R-squared

The R 2 Form of the F Statistic

Recall that R 2 = 1 − SSR/SST , and hence SSR = SST (1 − R 2 )
=⇒ the F statistic can be rewritten as:
2 − R 2 ) /q
(Rur r
F = 2 ) / (n − k − 1)
∼ Fq,n−k −1 (33)
(1 − Rur

(University of Texas) Chapter 4: Inference 48 / 53

The F Test

Steps for conducting an F test manually (later we will see Stata).

1 Calculate F using the formula above.
2 ∗
For the chosen significance level ℓ, obtain the critical value Fq,n −k −1,ℓ
3 Rejection rule:
Reject H0 if F > Fq∗,n−k −1,ℓ . (We say that xk −q +1 , ..., xk are jointly
statistically significant.)
Do not reject H0 if F < Fq∗,n−k −1,ℓ . (We say that xk −q +1 , ..., xk are
jointly statistically insignificant, which justifies dropping them from the
model.)
Important: It is possible that for a group of variables, each are individually
insignificant (t tests), but the group is jointly statistically significant (F
test) - e.g. if the variables are highly correlated with each other.

(University of Texas) Chapter 4: Inference 49 / 53

The F Test — graphically

(University of Texas) Chapter 4: Inference 50 / 53

P-values

Computing p-values for F-tests

We are not even going to explicitly do the above (i.e. comparing F to
∗
Fq,n −k −1,ℓ ) - instead we will use the more convenient p-value version
of the test (p-values are reported by STATA as illustrated below).
In an F test context, the p-value is given by:

p − value = Pr(Fq,n−k −1 > F ) (34)

where Fq,n−k −1 is an F random variable with df = (q, n − k − 1)

Just like with t tests, once the p-value has been calculated the F test
can be carried out at any level of significance. We would reject H0 for
all significance levels greater than the p-value.

(University of Texas) Chapter 4: Inference 51 / 53

Overall Significance of a Regression
One hypothesis that is tested automatically by most regression
packages is that all the coefficients (except the constant term) equal
zero, i.e.
H0 : β1 = 0, β2 = 0, . . . , βk = 0
In other words, the null hypothesis is that no explanatory variable has
an effect on y (and the alternative hypothesis is that at least one βj
is different from 0).
The restricted model is y = β0 + u.
Rr2 for the restricted model is 0! (None of the variation in y is
explained by the x variables.)
The F statistic (sometimes called the full F-statistic) for testing this
null simplifies to

R 2 /k
F = ∼ Fk , n − k − 1
(1 − R 2 ) / (n − k − 1)

where R 2 is the usual R 2 from the regression of y on all the x’s.

(University of Texas) Chapter 4: Inference 52 / 53
General Linear Restrictions
Although testing exclusion restrictions is the most important
application of the F test, sometimes we are interested in testing more
complicated hypotheses.
Suppose we have the unrestricted model
y = β0 + β1 x1 + β2 x2 + β3 x3 + u
and want to test
H0 : β1 = 1, β2 = 0, β3 = 0
The restricted model would be
y = β0 + x1 + u
There are ways to do this by hand..... but Stata is much easier! (test
command examples in Stata, e.g. H0 : β1 = 1, β2 = 0, β3 = 0, or
H0 : β1 = β2 , β3 = 2β4 in a model with 4 x variables, and
lect11 [Link])
(University of Texas) Chapter 4: Inference 53 / 53