ECOM20001

Econometrics 1

Lecture Note 6
Multiple Linear Regression - Estimation

A/Prof Victoria Baranov

Department of Economics
University of Melbourne

Stock and Watson: Chapter 6

1 / 50
Summary of Key Concepts

▶ Omitted Variable Bias

▶ Population Multiple Linear Regression Model
▶ Control Variables
▶ Heteroskedasticity
▶ OLS Estimator with Multiple Linear Regression
▶ Measures of Model Fit
▶ Least Squares Assumption in Multiple Linear Regression
▶ Perfect Multicollinearity
▶ Imperfect Multicollinearity

2 / 50
Building our Econometric Toolkit

NONLINEAR TIME SERIES

REGRESSION REGRESSION
(Lecture Note 8) (Lecture Note 9)

MULTIPLE LINEAR REGRESSION ESTIMATION AND TESTING

(Lecture Notes 6 & 7)

SINGLE LINEAR REGRESSION ESTIMATION AND TESTING

(Lecture Notes 4 & 5)

PROBABILITY AND STATISTICS

(Lecture Notes 2 & 3)

3 / 50
Student Test Scores and Class Size

100
VCE Economics Class Average Test Score
20 40 60
0 80

0 5 10 15 20 25 30 35
Number of Students in the Class

4 / 50
Student Test Scores and Household Income

100
VCE Economics Class Average Test Score
20 40 60
0 80

20 30 40
Average Household Income (1000's)

5 / 50
Class Size and Household Income

35 30
Number of Students in the Class
10 15 20
5
0 25

20 30 40
Average Household Income (1000's)

6 / 50
Econometrically Modelling Test Scores

▶ We’ve seen that:

▶ (↑ ClassSizei , ↓ TestScorei )
▶ But we also now have that
▶ ↓ Incomei , ↑ ClassSizei AND ↓ Incomei , ↓ TestScorei
So putting it together, we have:
▶ ↓ Incomei → (↑ ClassSizei , ↓ TestScorei )
▶ In words, if some other variable like Incomei varies across
classes, then this automatically creates a negative relationship
between ClassSizei and TestScorei

7 / 50
Econometrically Modelling Test Scores

▶ This has serious implications for interpreting the OLS

coefficient estimate from the following single linear regression:

TestScorei = β0 + β1 ClassSizei + ui

▶ Remember, β1 is meant to capture the individual relationship

(or direct link) between TestScorei and ClassSizei alone

8 / 50
Econometrically Modelling Test Scores

TestScorei = β0 + β1 ClassSizei + ui
▶ The OLS coefficient will fail to isolate the direct link between
ClassSizei and TestScorei
▶ Why?
▶ Because ↓ Incomei → (↑ ClassSizei , ↓ TestScorei )
▶ So the OLS estimate β̂1 from the single linear regression will
be driven by two forces:
1. (↑ ClassSizei , ↓ TestScorei ), the negative direct relationship we
want to determine empirically
2. ↓ Incomei → (↑ ClassSizei , ↓ TestScorei ), a separate negative
indirect correlation between ClassSizei and TestScorei due to
differences in Incomei across classes

9 / 50
Econometrically Modelling Test Scores

▶ In words, what does all of this mean for an econometrician?

▶ Suppose you obtained a statistically significant (from 0) β̂1
▶ Suppose you further concluded from this that it means we can
increase test scores by reducing class sizes.
▶ This interpretation is based on the negative direct relationship
between test scores and class size we have in mind

10 / 50
Econometrically Modelling Test Scores

▶ Once you make this claim, however, someone else who is

watching your presentation says the following:
“you may have a statistically significant (from 0) β̂1 estimate empirically,
but couldn’t that estimate just be driven by the fact that higher income
schools tend to have smaller class sizes, and higher income kids tend to
do better on tests?”
▶ This criticism of the above interpretation is based on the
negative indirect relationship between test scores and class
size that arises because of differences in income across schools
▶ It could be that there is no underlying direct relationship
between test scores and class sizes driving your β̂1 estimate; it
could be completely driven by an indirect relationship driven
by differences income across schools

11 / 50
Econometrically Modelling Test Scores
▶ Should we expect the β̂1 estimate to be bigger or smaller than
the population value of β1 ?
▶ Conceptually, we interpretting what our OLS estimate β̂1
means, we can think of it containing two parts:

β̂1 = β1 + γ
|{z} |{z}
direct indirect

where:
▶ β1 : (↑ ClassSizei , ↓ TestScorei ), the true negative direct class
size – test score relationship we want to determine empirically
▶ γ: ↓ Incomei → (↑ ClassSizei , ↓ TestScorei ), the negative
indirect class size – test score relationship being driven by
differences in income across classes
▶ Given we expect γ < 0, this means that we can expect our
single linear regression estimate to yield β̂1 < β1 , which
means that it gives a biased estimate of the direct class size –
test score relationship
12 / 50
Econometrically Modelling Test Scores

▶ So the magnitude of the OLS estimate of the relationship

between TestScorei and ClassSizei from a single linear
regression will be too large relative to the true value of β1
▶ Intuition: β̂1 captures the true class size – test score
relationship, but is ALSO confounded by the fact that richer
kids are in smaller classes and tend to do better in school for
reasons unrelated to class size

13 / 50
Omitted Variable Bias

▶ The example we have just discussed is an example of omitted

variable bias in econometrics
▶ If the regressor (ClassSizei ) is correlated with a variable that
has been omitted from the analysis (Incomei ) AND that
determines, in part, the dependent variable (TestScorei ), then
the OLS estimator of the effect of interest (test size – class
size relationship) will suffer from omitted variable bias
(causing β̂1 < β in our example)

14 / 50
Omitted Variable Bias

▶ Omitted variable bias occurs when the omitted variable (e.g,.

income) satisfies two conditions:
1. correlated with the included regressor (e.g, class size)
2. helps determine the dependent variable (e.g., test scores)
▶ If omitted variable bias exists, then E [β̂1 ] ̸= β1 , the OLS
estimate of β1 from a single linear regression is now biased,
and all of our machinery for estimating and testing regression
models fails

15 / 50
Omitted Variable Bias and OLS Assumption #1

▶ Omitted variable bias means that the first least squares

assumption of E [ui |Xi ] = 0, fails
▶ Why? Recall ui contains all factors other than Xi that are
determinants of Yi
▶ If one of these factors are correlated with Xi then this means
ui is correlated with Xi
▶ Example: if Incomei is a determinant of TestScorei (e.g., Yi )
and we omit it then it is in ui , AND if Incomei is correlated
with ClassSizei (e.g., Xi ), then ui will be correlated with Xi
▶ Because ui and Xi are correlated in the presence of an
omitted variable, the conditional mean of ui given Xi is not
zero → violates the first OLS assumption!
▶ Recall that if corr (ui , Xi ) ̸= 0 =⇒ E [ui |Xi ] ̸= 0

16 / 50
Formula for Omitted Variable Bias

▶ Suppose least squares assumptions 2 (IID) and 3 (no outliers)

hold, but assumption 1 (independence) does not hold
▶ Let corr (ui , Xi ) = ρXu is the correlation between Xi and ui in
the single linear regression
▶ If omitted variable bias is present, then as n → ∞
σu
β̂1 → β1 + ρXu
σX
▶ If there’s no omitted variable bias, ρXu = 0
▶ However, if there is omitted variable bias ρXu ̸= 0

17 / 50
Implications of Omitted Variable Bias

σu
β̂1 → β1 + ρXu
σX
▶ With omitted variable bias, as n gets large, β̂1 does not get
close to β1 with high probability
▶ The bias term ρXu σσu exists even if n is very large
X
▶ The size of the bias depends on the magnitude of ρXu
▶ The direction of the bias in β̂1 depends on the sign of ρXu
(whether it’s positive or negative)

18 / 50
Signing Omitted Variable Bias
▶ In our example, we had a positive relationship between our
omitted variable Incomei and our outcome variable Y which
was TestScorei .
▶ This means Incomei enters ui in the single linear regression
with a positive sign (+).
▶ Further, there was a negative (-) relationship between our
omitted variable Incomei and our independent variable X
which was ClassSizei
▶ Therefore, the sign of the correlation between X and u is
given by sign[ρXu ]=sign[(+) × (-)]=(-)
▶ Given that
σu
β̂1 → β1 + ρXu
|{z} σX
(-)

there will be a negative bias in β̂1 relative to the true value

β1 , that is: β̂1 < β1
19 / 50
Fixing Omitted Variable Bias

▶ Conceptually, how might we try to fix the omitted variable

bias problem in our example?
▶ Let’s start with the source of the problem: income levels vary
across schools, which is what creates the bias in β̂1
▶ What if instead of using all the schools in our sample, we
focused on a group of schools that had similar income levels?
▶ for example, only look at schools with average household
income between $29,000 and $31,000

20 / 50
Fixing Omitted Variable Bias
Source of the bias: variation in income

35 30
Number of Students in the Class
10 15 20
5
0 25

20 30 40
Average Household Income (1000's)

21 / 50
Fixing Omitted Variable Bias
Fixing the problem: taking a sub-sample with similar income

35
1. Take a sub-sample of schools with
30
average household income between
$29,000 and $31,000
Number of Students in the Class

2. Plot test scores versus class size for this

sub-sample with similar income levels
10 15 20
5
0 25

20 30 40
Average Household Income (1000's)

22 / 50
Fixing Omitted Variable Bias
Test score – class size relationship based on the sub-sample

100
VCE Economics Class Average Test Score
20 40 60
0 80

0 5 10 15 20 25 30 35
Number of Students in the Class

23 / 50
Fixing Omitted Variable Bias

▶ In the sub-sample we focused on, income is now similar across

classes, so if we find a negative relationship between
ClassSizei and TestScorei in the sub-sample, then it is more
likely to be driven by a direct relationship rather than an
indirect relationship because of income differences
▶ This highlights the idea of holding income fixed in estimating
the direct class size – test score relationship

24 / 50
Multiple Linear Regression

▶ The multiple linear regression model extends the single linear

regression model to include additional variables as regressors
▶ The model allows us to estimate the effect on Yi of changing
one variable (X1i ) while holding other regressors
(X2i , X3i , X4i , . . .) constant (or fixed)
▶ In our example, we can use multiple linear regression to isolate
the effect on test scores (e.g., Yi ) of class size (e.g., X1i )
while holding household income (e.g., X2i ) fixed
▶ In this way, multiple linear regression is a tool for eliminating
omitted variable bias

25 / 50
Population Regression Model
▶ Population regression model with k regressors is defined as:

Yi = β0 + β1 X1i + β2 X2i + . . . + βk Xki + ui , i = 1, . . . , n

▶ For expositional purposes, in what follows we will work with a
regression including two regressors, X1i , X2i .
▶ However, everything I discuss immediately extends to the
general case with k regressors X1i , X2i , . . . Xki
▶ Also, X1i and X2i may be any combination of continuous or
dummy variables
▶ Population regression model with k = 2 regressors:

Yi = β0 + β1 X1i + β2 X2i + ui , i = 1, . . . , n

▶ From our test scores example with i = 1, . . . , n classes:

TestScorei = β0 + β1 ClassSizei + β2 Incomei + ui , i = 1, . . . , n

26 / 50
Control Variables

▶ Taking conditional expectations at a point where X1i = x1

and X2i = x2 , the population regression function is:

E (Yi |X1i = x1 , X2i = x2 ) = β0 + β1 x1 + β2 x2

where β1 and β2 are regression coefficients on X1i and X2i

▶ We often refer to some of the regressors in a multiple linear
regression as control variables
▶ Interpreting β1 , we say it is the relationship between X1i on Yi
holding X2i fixed (or, equivalently, controlling for X2i )
▶ From our example, if we used multiple linear regression, β1
would be the relationship between class size and test score,
controlling for income

27 / 50
Coefficient Interpretation

Yi = β0 + β1 X1i + β2 X2i + ui (1)

▶ When we interpret coefficients, we imagine changing only one
regressor at a time, leaving the others fixed
▶ For example, we increase X1i by ∆X1 and leave X2i fixed
▶ Changing X1i we lead to an expected change in Yi , which we
label ∆Y , and is defined from the population regression line:
Yi + ∆Y = β0 + β1 (X1i + ∆X1 ) + β2 X2i (2)
▶ Subtracting equations (1) and (2), we have:
∆Y
β1 = , holding X2 constant
∆X1
▶ β1 is the expected change in Yi from a one-unit change in X1i .
▶ It is often called the partial effect of X1i on Yi , which
emphasises our focus on changing just one regressor while
holding all other regressors fixed
28 / 50
The Constant

Yi = β0 + β1 X1i + β2 X2i + ui
▶ The intercept β0 is called the constant term and it is
interpreted as the average value of Yi when X1i = 0 and
X2i = 0
▶ We can equivalently write the regression including a third
regressor X0i which is a dummy variable that equals one for
all observations:

Yi = β0 X0i + β1 X1i + β2 X2i + ui

where X0i is often called the constant regressor

29 / 50
Heteroskedasticity

▶ The error term in the regression is homoskedastic if the

variance conditional on all of the regressors,
var (ui |X1i , X2i , . . . , Xki ) is constant for i = 1, . . . , n
▶ Otherwise, the error term is heteroskedastic
▶ We will continue to work under the more general assumption
of heteroskedasticity in computing standard errors and
conducting statistical inference (next lecture note)

30 / 50
Student Test Score Example

▶ We could imagine adding additional regressors to our test

score regression:

TestScorei =β0 + β1 ClassSizei + β2 Incomei + β3 ParentEduci +

+ β4 FamilySizei + β5 ShareFemalei + β6 ShareImmig
+ β7 Urbani + β8 Privatei + ui , i = 1, . . . , n

▶ Richer and richer regressions allows us to control for (or hold

fixed) many other variables that predict test scores in avoiding
omitted variable bias to isolate the relationship between
ClassSizei on TestScorei , which is the main effect of interest

31 / 50
OLS Estimation with Multiple Linear Regression
▶ Just like with the singe linear regression, we use the Ordinary
Least Squares (OLS) estimator to estimate the regression
coefficients of a multiple linear regression model
▶ Recall that the OLS estimator aims to find the regression
coefficients that together minimise the mistakes the model
makes in predicting the dependent variable Yi given the k
regressors X1i , X2i , . . . , Xki
▶ For a given set of regression coefficients, b0 , b1 , b2 , . . . , bk ,
the model’s mistake in predicting Yi is:
Yi − b0 − b1 X1i − b2 X2i − . . . − bk Xki
▶ The sum of squared prediction mistakes across all i = 1, . . . , n
observations is:
 2
Xn
Yi − b0 − b1 X1i − b2 X2i − . . . − bk Xki 
| {z }
i=1 prediction mistake

32 / 50
OLS Estimation with Multiple Linear Regression
▶ The OLS estimators of β0 , β1 , β2 , . . . , βk correspond to the
b0 , b1 , b2 , . . . , bk values that together minimise the sum of
squared prediction mistakes
▶ As usual, the OLS estimators are denoted by β̂0 , β̂1 , β̂2 , . . . , β̂k
▶ The OLS regression function is the (k-dimensional) line
constructed using the OLS estimators:

β̂0 + β̂1 X1i + β̂2 X2i + . . . + β̂k Xki

▶ The OLS predicted value of Yi given X1i , X2i , . . . , Xki is:

Ŷi = β̂0 + β̂1 X1i + β̂2 X2i + . . . + β̂k Xki

▶ The OLS residual for observation i is the difference between

Yi in the data and the model’s predicted value for Yi

ûi = Yi − Ŷi
33 / 50
Test Scores and Class Size Example

▶ Single linear regression of test scores on class size

\ i = 105.24 − 2.43 ClassSizei

TestScore
(1.84) (0.12)

Note: slightly different dataset than from previous lectures

▶ Multiple linear regression of test scores on class size and
average income

\ i = 73.99 − 1.96 ClassSizei + 0.82 Incomei

TestScore
(6.15) (0.14) (0.16)

▶ The inclusion of Incomei as a regressor causes the magnitude

of the coefficient on ClassSizei to fall as omitted variable bias
related to Incomei in the regression is accounted for

34 / 50
Measures of Fit in Multiple Linear Regression

▶ R 2 , which recall is the fraction of the sample variance in Yi

that is explained or predicted by the regressors, is again a
common measure of fit for multiple linear regression
▶ It is computed identically in multiple linear regression:

ESS SSR
R2 = =1−
TSS TSS
Pn 2
Pn = i=1 (Ŷi2 − Ȳ )
where the explained sum of squares ESS
and the total sum of squares TSS = i=1 (Yi − Ȳ )

35 / 50
Measures of Fit in Multiple Linear Regression

▶ The R 2 always rises when regressors are added to the

regression unless the estimated coefficient on a regressor is
exactly 0 (which very rarely happens in practice)
▶ Because of this, we often work the adjusted R 2 , denoted R̄ 2 ,
which is a modified version of R 2 that does not necessarily
increase when a new regressor is added
▶ The adjusted R 2 is computed as:

n − 1 SSR s2
R̄ 2 = 1 − = 1 − 2û
n − k − 1 TSS sY

▶ The difference with the standard R 2 = 1 − SSR

TSS formula is the
n−1
multiplication by n−k−1

36 / 50
Measures of Fit in Multiple Linear Regression

n − 1 SSR s2
R̄ 2 = 1 − = 1 − 2û
n − k − 1 TSS sY
▶ R̄ 2 is always less than than R 2 and therefore always less than 1
▶ Adding a regressor to the regression has two effects on R̄ 2 :
▶ SSR falls, which causes R̄ 2 to rise
n−1
▶ n−k−1 (because k goes up), which causes R̄ 2 to fall
▶ R̄ 2 can actually be negative if all the regressors together do
n−1
not decrease enough SSR to offset the n−k−1 factor

37 / 50
Measures of Fit in Multiple Linear Regression

▶ We can also measure fit of the regression using the standard

error of the regression (SER) which estimates the standard
deviation of the error term ui :
Pn 2
i=1 ûi SSR
q
2 2
SER = sû , where sû = =
n−k −1 n−k −1
where n is the number of observations, and k is the number of
regression
P coefficients beyond the constant, and
SSR = ni=1 ûi2 is the sum of squared residuals
▶ The use of n − k − 1 in the denominator in computing SER is
called a degrees of freedom adjustment (one for every one of
the k variables in the model, and one more for the constant)

38 / 50
Test Scores and Class Size Example

▶ Regression results from single linear regression of test scores

on class size

\ i = 105.24 − 2.43 ClassSizei , R 2 = 0.65, R̄ 2 = 0.64

TestScore
(1.84) (0.12)

Note: slightly different dataset than from previous lectures

▶ Regression results from multiple linear regression of test scores
on class size and average income

\ i = 73.99− 1.96 ClassSizei + 0.82 Incomei R 2 = 0.69, R̄ 2 = 0.68

TestScore
(6.15) (0.14) (0.16)

▶ The inclusion of Incomei as a regressor causes the magnitude

of the coefficient on ClassSizei to fall as omitted variable bias
related to Incomei in the regression is accounted for

39 / 50
Beware Interpretations of R 2 and R̄ 2

▶ R̄ 2 is useful because it summarises the extent to which the

regressors explain the variation in the dependent variable
▶ However, “maximising” the R̄ 2 in practice is rarely the goal to
economically or statistically addressing most questions
▶ In fact, if R̄ 2 is very close to 1 is often a sign that there is a
logical problem with the regression model!
▶ We return to strategies for evaluating econometric models
later in the subject, but having an extremely high R̄ 2 is
typically not the goal of econometric analysis
▶ For the remainder of the subject, we will focus on reporting
R̄ 2 for empirical analyses in tutorials and on assignments

40 / 50
The Least Squares Assumptions in Multiple Linear
Regression

▶ As with single linear regression, multiple linear regression

relies on 4 key assumptions that are critical for the sampling
distributions of the OLS estimators, β̂0 , β̂1 , β̂2 , . . . , β̂k
▶ Assumption 1: Independence

E [ui |X1i , X2i , . . . , Xki ] = 0

▶ Assumption 2: (X1i , X2i , . . . , Xki , Yi ), i = 1, . . . , n are IID

▶ Assumption 3: Large Outliers Are Unlikely
▶ Assumption 4: No Perfect Multicollinearity (new)

41 / 50
Perfect Multicollinearity
▶ Two regressors exhibit perfect multicollinearity if one of the
regressors is a perfect linear combination of other regressors
▶ Assumption 4 requires that no regressors exhibit perfect
multicollinearity
▶ Example: suppose you tried to run this regression by accident:

TestScorei = β0 + β1 ClassSizei + β2 ClassSizei + ui

where the two regressors are perfectly collinear

▶ Conceptually, β1 is the impact of ClassSizei (e.g,. the first
regressor) on TestScorei holding ClassSizei (e.g, the second
regressor) fixed, which does not make any sense
▶ Same problem if you attempt to estimate, for example:

TestScorei = β0 +β1 ClassSizei +β2 Incomei +β3 (Incomei /2)+ui

42 / 50
Perfect Multicollinearity

▶ In general, if a group of regressors are perfectly collinear, then

it is impossible to hold one regressor fixed to estimate the
effect of one of the other collinear regressors on the
dependent variable
▶ In practice, statistics software will either drop one of the
perfectly multicollinear variables or will give an error message
if you try to run a regression with perfect multicollinearity
▶ We fix perfect multicollinearity problems by modifying the set
of regressors to eliminate the problem.

43 / 50
Perfect Multicollinearity Example: Huge Classes
▶ Supposed we created a dummy variable HugeClassi which
equals one if a class has more than 35 students and is 0
otherwise. Here is the regression:

TestScorei = β0 +β1 ClassSizei +β2 Incomei +β3 HugeClassi +ui

▶ The regression exhibits perfect multicollinearity. Why?
▶ In our sample, no class has more than 35 students, which
means HugeClassi = 0 for all observations
▶ Recall the constant β0 in the regression is equivalent to having
β0 X0i in the regression, where X0i is the constant regressor
that is always equal to 1
▶ Therefore, HugeClassi = 1 − X0i and we have perfect
multicollinearity
▶ Two important aspects of this example:
1. Perfect multicollinearity can arise because of the constant
2. Perfect multicollinearity is specific to the dataset you have at
hand; we could imagine classes with more than 35 students
44 / 50
Dummy Variable Trap

▶ A possible source of multicollinearity arises when multiple

dummy variables are used as regressors
▶ For example, suppose the schools where split into either being
urban or rural schools, and you created two dummy variables:
▶ Urbani = 1 if school i is in an urban location and is 0 otherwise
▶ Regionali = 1 if school i is in an regional location and is 0
otherwise
▶ For each school, either:
▶ Urbani = 1 and Regionali = 0 (urban school)
OR
▶ Urbani = 0 and Regionali = 1 (regional school)
▶ Therefore, the sum of the two dummy variables equals 1:
Urbani + Regionali = 1 for all i = 1, . . . , n

45 / 50
Dummy Variable Trap

▶ Suppose you tried to run the regression:

TestScorei = β0 +β1 ClassSizei +β2 Incomei +β3 Urbani +β4 Regionali +ui

▶ You again face multicollinearity because of the constant;

Urbani + Regionali = 1 = X0i for i = 1, . . . , n
▶ That is, the urban and regional dummy variables add up to
equal the constant regressor for each observation i = 1, . . . , n

46 / 50
Dummy Variable Trap
▶ The situation when a group of dummy variables add up to
always equal another dummy variable (or the constant
regressor) is called the dummy variable trap
▶ You can avoid the dummy variable trap by dropping one of
the dummy variables (or dropping the constant):

TestScorei = β0 + β1 ClassSizei + β2 Incomei + β3 Urbani + ui

where Regionali is dropped, and β3 is the difference in test

scores in urban schools relative to regional schools,
holding all other regressors constant
▶ In general, if:
▶ there are G dummy variables
▶ each observation falls into one and only category
▶ there is an intercept in the regression
▶ all G binary variables are included in the regression
then the regression fails because of perfect multicollinearity as
you have falled into the dummy variable trap
47 / 50
Avoiding the Dummy Variable Trap
▶ We typically avoid the dummy variable trap by only including
G − 1 of the G dummy variables in the regression
▶ The dummy that we do not include in the regression is the
base category or base group or omitted category
▶ We interpet all the other dummies as the change in the
outcome variable when a given dummy variable is equal to 1
relative to base group, holding all other regressors constant
▶ Alternatively, we can include all G dummies and drop the
constant in the regression (very uncommon)
▶ In sum, when your software indicates that you have perfect
multicollinearity, it is important to eliminate it by:
1. determining the source of perfect multicollinearity
2. creating a base group
3. ensuring you properly interpret regression coefficients for the
dummies in the regression relative to the omitted base group

48 / 50
Multicollinearity

▶ A related issue is imperfect multicollinearity, which arises

when one of the regressors is highly correlated, but not
perfectly correlated, with other regressors
▶ This does not prevent statistics programs from providing OLS
estimates, but it does result in regression coefficients being
estimated imprecisely and having large standard errors, and
therefore statistically insignificant regression coefficients
▶ Intuitively, if two regressors are highly correlated and almost
always co-moving together, it is hard to disentangle their
individual impacts on the dependent variable in the regression
▶ Whereas perfectly multicolliearity typically arises because of a
logical mistake in your regression set-up, imperfect
multicollinearity is not necessarily an error but a feature of
your data, OLS, and the question you are trying to address

49 / 50
Distribution of OLS Estimators in Multiple Linear
Regression
▶ Because random samples vary from one sample to the next,
different samples product different values for the OLS
estimators, β̂0 , β̂1 , β̂2 , . . . , β̂k
▶ That is, these estimators are random variables with a
distribution
▶ Under the 4 least squares assumptions, the OLS estimators
β̂0 , β̂1 , β̂2 , . . . , β̂k are unbiased and consistent estimators of
their population true values β0 , β1 , β2 , . . . , βk
▶ In large samples, the sampling distribution of
β̂0 , β̂1 , β̂2 , . . . , β̂k is well approximated by a multivariate
normal distribution, with each β̂j having a marginal
distribution that is N(βj , σβ̂2 ) for j = 0, 1, 2, . . . , k
j
▶ We can use these results to conduct hypothesis tests with
multiple linear regression models using t-statistics and p-values
similar to what we did with single linear regression models
50 / 50