Multiple regression model

January 21, 2025 12:04 PM

New Section 1 Page 1

New Section 1 Page 2
Inference in multiple regression
January 21, 2025 3:41 PM

New Section 1 Page 3

New Section 1 Page 4
Asymptotics
January 22, 2025 6:32 PM

Asymptotics is the study that tells us what happens to the estimators,

test statistic and models when Sample size is growing to infinity
We always have finite samples but we might use the asymptotic
normality of estimator due to the large sample
• Consistency means that your sample size is getting close to infinity and
this means that the parameter is close to a true value of the parameter.
For example you try to find the average height in Aarhus and the more
people you measure the closer the average becomes to the true average.
• Consistency is important because if the estimator is not consistent the
estimate would be wrong. For example if we are measuring the heights as
in the example before and our ruler is 3 cm off. It means that it doesn't
matter how many people we are measuring the average will be wrong by
3 cm
• The most important conditions(1-4th assumptions) for this is the zero
correlations of errors since if we leave out an important variable from the
regression and it correlated to one of the included terms it will pull the
regression line in the wrong direction and no amount of extra data will fix
it so it's really important to write a good model that includes relevant
variables and the variables can't be collinear since it would mess up the
calculations.
• We can check the consistency using probability limit definition it shows
us that as the sample size increases the

• And to calculate the degree of inconsistency:

And if the covariance between x and u is 0 (which is true according

the (zero conditional mean assumption)OLS estimator is consistent.
• MLR4' is a weaker form of MLR4 since it does not require for u to be
uncollerated with x1 individually. It's weaker since it's only zero
mean and zero corelation only while MLR4 is zero conditional mean.
We use mlr4' when some of the relationships are nonlinear but you
want linear summary or in settings where full zero conditional mean
asumption cannor be justified.

• Asymptotic normality is the Central limit theorem(n>30), standard errors

shrinks as the sample size increases (1/sqrt(n))

New Section 1 Page 5

 •

Save residuals means to extract difference of the true values of y and predicted values of
the regression.

New Section 1 Page 6

New Section 1 Page 7
Further issues on multiple regression
January 23, 2025 1:11 PM

Data Scaling

Functional form

New Section 1 Page 8

New Section 1 Page 9
 Skip dfc is this

New Section 1 Page 10

Multiple linear regression: qualitative
information(dummy variables)
January 23, 2025 4:21 PM

For example because the coefficiant female -1.8 it means that women earn
1.8 units less than men
After logging the wage we understand the coefficient -0.3011 of female like this:

We can't put two dummy variables for woman and man due to perfect linear relationship they have
Multicollinearity assumption, then one of the variables would have to be ommited, it won't change
The result only the coefficient interpretation

In stata we could use the lincom command to fix the problem of the intercept

Multiple categories of dummies

That one cathegory is going to be the base line

New Section 1 Page 11

 Interactions with dummy variables

It might have different marginal effect since the difference in creditworthiness between
cr1 and cr2 might be different than the difference of credit worthiness between CR2
and CR3

Difference In slopes

New Section 1 Page 12

Heteroscedacity
January 24, 2025 4:07 PM

The heteroscedasticity does not affect the estimators beta orr r^2

New Section 1 Page 13

Testing for Heteroscedasticity

We can also test this visual by looking at the graph

The F test is done on the U(auxiliary regression) but we don't cover it in this course and the statistic we
compare to the LM statistic(or BP statistic) is the chi squared critical value that we find using the
significance level and the independeng variables number.

New Section 1 Page 14

compare to the LM statistic(or BP statistic) is the chi squared critical value that we find using the
significance level and the independeng variables number.

New Section 1 Page 15

New Section 1 Page 16
Specifications and data issues
January 25, 2025 12:58 PM

Myzon and rychard test

New Section 1 Page 17

 For dependent variable y

Y creates biased estimates if everyone of lower income increses their salary in the survey
On the other hand the measurment error in the x is more severe because it lowers the explanitory ower
of x which couses the biased and inconsistant estimatesfor example ven we have error for the x it adds
the noise for the x which weakens the relationship between y and xthis leads to antuiation biaswhich
means that our coefficiants are biased

For explanatory variable x

Regression is less precise

New Section 1 Page 18

New Section 1 Page 19
New Section 1 Page 20
 basics to understand Time series
January 25, 2025 3:06 PM

New Section 1 Page 21

New Section 1 Page 22
Issues in using OLS with time series data
January 25, 2025 3:25 PM

H is change in time
The mean and variance in
different time periods have to be
the same for the regression to be
stationary.

Weakly dependence- our current observations do not effect the future observations as the lag
h increases

We use autocorrelation to check the weakly dependency

New Section 1 Page 23

 Self regression

New Section 1 Page 24

 Here we don't calculate the transformation since our this since the regresion is already stationary

The transformation is when we calculate the first difference since the regresion is non stationary or have
linear trends

New Section 1 Page 25

Serial correlation and heteroskedasticity in ts
January 25, 2025 11:12 PM

White test

How errors are affecting variance under autoregressive model 1

We use this to fix correlation

For g we only take the number before

the comma, we don't round the
number at all.

New Section 1 Page 26

 We don't need to talk about GLS

We use ar

F and t statistic????

New Section 1 Page 27

 White test

Autoregressive conditional heteroskedasticity(when variance is not random but depend on past volatility)

This model is good for using to predict future intrest rates

New Section 1 Page 28

Simple panel data methods
January 27, 2025 4:29 PM

Pooled cross-sectional data

Regression with independently pooled cross sections

Example

Dummy variables
Regression:

Since you suspect the relationship between x and y you create dummy variable

Nr and fr is from the example of the house values that are near incenirator and far from it

It looks how much the coefficient changet in between the pooled cross sectional models of different
times then we run the hypothesis H_0: coefficienat=0 and h1 h0 is not true using the t statistic

Panel data

New Section 1 Page 29

 For panel data we need to remove the variables that are constant over time but cannot be directly
measured to remove ommited variable bias that can lead to incorrect conclusions and we can do it with
First difference estimator but firstly we need to find the unobsserved effects

New Section 1 Page 30

Advanced panel data methods
January 27, 2025 6:09 PM

This topic discusses two methods for estimating unobserved effects panel data
models besides first difference estimator.
Fixed effects estimator
While analyzing panel data fixed effects estimator Is another way to eliminate fixed
effets a_i by differencing the data across time periods
Fixed effects

Does not have intercept because it's absorbed in the ai

Ai is treated as

Random effects estimator

Another way to remove the fixed effects from the regression for more accurate results.
We use it when the fixed effects are uncorrelated wit the regressors
Because there is an intercept we can assume that
ai has the mean of 0
Ai is treated as a random variable

New Section 1 Page 31

 Ai is treated as a random variable

Since we assume that ai seems to be uncorellated

we can put it into an error term

We can use ai because it makes the ai robust and

deletes it later in the regression

When we have a lot time periods and low sample size

New Section 1 Page 32

New Section 1 Page 33
Instrumental variables estimation
January 27, 2025 6:55 PM

Instumental variables estimation is another way how to fix the omitted variable problem. We
are using them when x is correlated with u term violating the zero conditional mean
assumption mlr4.

We can do a simple regression to test it by makin x our explained variable and z

our explaining variable and do the t test(it checks if z explain x)

Simple regression with Instumental variables

New Section 1 Page 34

 Multiple regression with Instrumental variables

Mostly the same as in simple regresion only called two stages least squeres

New Section 1 Page 35

New Section 1 Page 36
Forecasting
January 27, 2025 10:08 PM

In in sample we are looking at the r squared as an criteria the bigger r^2 the better it is

RMSE-root mean squared error, MAE-mean absolute error

We use the data for out sample with data point that were not used in the in sample part.

One step ahead forecasting (AR(1)+X)

New Section 1 Page 37

 Mention forecast error

Granger causality

VAR we want to compare two models

Multiple step ahead forecasting

It's usedwhen we want to forecast more than one period ahead. There are two ways to do it first
one is Intarative forecasting

Direct h-step-ahead forecasts for AR(1)

New Section 1 Page 38

New Section 1 Page 39