03_Linear-Regression

Report
Econometrics
Session 3 – Linear Regression
Amine Ouazad,
Asst. Prof. of Economics
Econometrics
Session 3 – Linear Regression
Amine Ouazad,
Asst. Prof. of Economics
Outline of the course
1. Introduction: Identification
2. Introduction: Inference
3. Linear Regression
4. Identification Issues in Linear Regressions
5. Inference Issues in Linear Regressions
This session
Introduction: Linear Regression
• What is the effect of X on Y?
• Hands-on problems:
– What is the effect of the death of the CEO (X) on
firm performance (Y)? (Morten Bennedsen)
– What is the effect of child safety seats (X) on the
probability of death (Y)? (Steve Levitt)
This session:
Linear Regression
1. Notations.
2. Assumptions.
3. The OLS estimator.
– Implementation in STATA.
4. The OLS estimator is CAN.
Consistent and Asymptotically Normal
5. The OLS estimator is BLUE.*
Best Linear Unbiased Estimator (BLUE)*
6. Essential statistics: t-stat, R squared, Adjusted R
Squared, F stat, Confidence intervals.
7. Tricky questions.
*Conditions apply
Session 3 – Linear Regression
1. NOTATIONS
Notations
• The effect of X on Y.
• What is X?
– K covariates (including the constant)
– N observations
– X is an NxK matrix.
• What is Y?
– N observations.
– Y is an N-vector.
Notations
• Relationship between y and the xs.
y=f(x1,x2,x3,x4,…,xK)+e
• f: a function K variables.
• e: the unobservables (a scalar).
Session 3 – Linear Regression
2. ASSUMPTIONS
Assumptions
•
•
•
•
A1: Linearity
A2: Full Rank
A3: Exogeneity of the covariates
A4: Homoskedasticity and
nonautocorrelation
• A5: Exogenously generated covariates.
• A6: Normality of the residuals
Assumption A1: Linearity
• y = f(x1,x2,x3,…,xK)+e
• y = x1 b1 + x2 b2 + …+xK bK + e
• In ‘plain English’:
– The effect of xk is constant.
– The effect of xk does not depend on the value of
xk’.
• Not satisfied if :
– squares/higher powers of x matter.
– Interaction terms matter.
Notations
1. Data generating process
2. Scalar notation
3. Matrix version #1
4. Matrix version #2
Assumption A2: Full Rank
• We assume that X’X is invertible.
• Notes:
– A2 may be satisfied in the data generating
process but not for the observed.
• Examples:
– Month of the year dummies/Year dummies,
Country dummies, Gender dummies.
Assumption A3: Exogeneity
• i.e. mean independence of the residual and the
covariates.
• E(e|x1,…,xK) = 0.
• This is a property of the data generating
process.
• Link with selection bias in Session 1?
Dealing with Endogeneity
• You’re assuming that there is no covariate
correlated with the Xs that has an effect on Y.
– If it is only correlated with X with no effect on Y, it’s
OK.
– If it is not correlated with X and has an effect on Y,
it’s OK.
• Example of a problem:
– Health and Hospital stays.
– What covariate should you add?
• Conclusion: Be creative !! Think about
unobservables !!
Assumption A4: Homoskedasticity
and Non Autocorrelation
• Var(e|x1,…,xK) = s2.
• Corr(ei, ej|X) = 0.
• Visible on a scatterplot?
• Link with t-tests of session 2?
• Examples: correlated/random effects.
Assumption A5
Exogenously generated covariates
1. Instead of requiring the mean
independence of the residual and the
covariates, we might require their
independence.
– (Recall X and e independent if f(X,e)=f(X)f(e))
2. Sometimes we will think of X as fixed rather
than exogenously generated.
Assumption A6:
Normality of the Residuals
• The asymptotic properties of OLS (to be
discussed below) do not depend on the
normality of the residuals: semi-parametric
approach.
• But for results with a fixed number of
observations, we need the normality of the
residuals for the OLS to have nice properties
(to be defined below).
Session 3 – Linear Regression
3. THE ORDINARY LEAST SQUARES
ESTIMATOR
The OLS Estimator
• Formula:
• Two interpretations:
– Minimization of sum of squares (Gauss’s
interpretation).
– Coefficient beta which makes the observed X
and epsilons mean independent (according to
A3).
OLS estimator
• Exercise: Find the OLS estimator in the case
where both y and x are scalars (i.e. not
vectors). Learn the formula by heart (if
correct !).
Implementation in Stata
• STATA regress command.
– regress y x1 x2 x3 x4 x5 …
• What does Stata do?
– drops variables that are perfectly correlated. (to
make sure A2 is satisfied). Always check the
number of observations !
• Options will be seen in the following
sessions.
• Dummies (e.g. for years) can be included
using « xi: i.year ». Again A2 must be
satisfied.
First things first: Desc. Stats
• Each variable used in the analysis:
Mean, standard deviation for the
sample and the subsamples.
• Other possible outputs: min max,
median (only if you care).
• Source of the dataset.
• Why??
• Show the reader the variables
are “well behaved”: no outlier
driving the regression,
consistent with intuition.
• Number of observations should
be constant across regressions
(next slide).
Reading a table …
from the Levitt paper (2006 wp)
Other important advice
1. As a best practice always start by
regressing y on x with no controls except
the most essential ones.
•
No effect? Then maybe you should think twice
about going further.
2. Then add controls one by one, or group by
group.
•
Explain why coefficient of interest changes
from one column to the next. (See next
session)
Stata tricks
• Output the estimation results using estout
or outreg.
– Display stars for coefficients’ significance.
– Outputs the essential statistics (F, R2, t test).
– Stacks the columns of regression output for
regressions with different sets of covariates.
• Formats: LaTeX and text (Microsoft Word).
Session 3 – Linear Regression
4. LARGE SAMPLE PROPERTIES OF
THE OLS ESTIMATOR
The OLS estimator is CAN
• CAN :
– Consistent
– Asymptotically Normal
  = 
( − ) → (0, )
• Proof:
1. Use ‘true’ relationship between y and X to show
that b = b + (1/N (X’X)-1 )(1/N (X’e)).
2. Use Slutsky theorem and A3 to show consistency.
3. Use CLT and A3 to show asymptotic normality.
4. V = plim (1/N (X’X)) -1
OLS is CAN: numerical simulation
• Typical design of a study:
1. Recruit X% of a population (for instance a random
sample of students at INSEAD).
2. Collect the data.
3. Perform the regression and get the OLS estimator.
• If you perform these steps independently a
large number of times (thought
experiment), then you will get a normal
distribution of parameters.
Important assumptions
• A1, A2, A3 are needed to solve the identification
problem:
– With them, estimator is consistent.
• A4 is needed
– A4 affects the variance covariance matrix.
• Violations of A3? Next session (identif. Issues)
• Violations of A4? Session on inference issues.
Session 3 – Linear Regression
5. FINITE SAMPLE PROPERTIES OF
THE OLS ESTIMATOR
The OLS Estimator is BLUE
• BLUE:
– Best
…
– Linear
…
X and Y
– Unbiased …
– Estimator …
observations
i.e. has minimum variance
i.e. is a linear function of the
i.e.   = 
i.e. it is just a function of the
• Proof (a.k.a. the Gauss Markov Theorem):
OLS is BLUE
• Steps of the proof:
– OLS is LUE because of A1 and A3.
– OLS is Best…
1.
2.
3.
4.
For any other LUE, such as Cy, CX=Id.
Then take the difference Dy= Cy-b. (b is the OLS)
Show that Var(b0|X) = Var(b|X) + s2 D’D.
The result follows from s2 D’D > 0.
Finite sample distribution
• The OLS estimator is normally distributed for
a fixed N, as long as one assumes the
normality of the residuals (A6).
• What is “large” N?
– Small: e.g. Acemoglu, Johnson and Robinson
– Large: e.g. Bennedsen and Perez Gonzalez.
– Statistical question: rate of convergence of the
law of large numbers.
This is small N
Other examples
Large N
• Compustat (1,000s + observations)
• Execucomp
• Scanner data
Small N
• Cross-country regressions (< 100 points)
Session 3 – Linear Regression
6. STATISTICS FOR READING THE
OUTPUT OF OLS ESTIMATION
Statistics
• R squared
– What share of the variance of the outcome variable is
explained by the covariates?
• t-test
– Is the coefficient on the variable of interest significant?
• Confidence intervals
– What interval includes the true coefficient with
probability 95%?
• F statistic.
– Is the model better than random noise?
Reading Stata Output
R Squared
• Measures the share of the variance of Y (the
dependent variable) explained by the model
Xb, hence R2 = var(Xb)/var(Y).
• Note that if you regress Y on itself, the R2 is
100%. The R2 is not a good indicator of the
quality of a model.
Tricky Question
• Should I choose the model with the highest
R squared?
1. Adding a variable mechanically raises the R
squared.
2. A model with endogenous variables (thus not
interpretable nor causal) can have a high R
square.
Adjusted R-Square
• Corrects for the number of variables in the
regression.
−1
 2 = 1 − − (1 − 2)
• Proposition: When adding a variable to a
regression model, the adjusted R-square
increases if and only if the square of the tstatistic is greater than 1.
• Adj-R2: arbitrary (1, why 1?) but still
interesting.
t-test and p value
=

 2 
→ ( − )
• p-value: significance level for the coefficient.
• Significance at 95% : pvalue lower than 0.05.
– Typical value for t is 1.96 (when N is large, t is normal).
• Significance at X% : pvalue lower than 1-X.
• Important significance levels: 10%, 5%, 1%.
– Depending on the size of the dataset.
• t-test is valid asymptotically under A1,A2,A3,A4.
• t-test is valid at finite distance with A6.
• Small sample t-tests… see Wooldridge NBER conference,
“Recent advances in Econometrics.”
F Statistic
• Is the model as a whole significant?
• Hypothesis H0: all coefficients are equal to
zero, except the constant.
• Alternative hypothesis: at least one
coefficient is nonzero.
• Under the null hypothesis, in distribution:
2
  − 1,  −  =  − 1 → ( − 1,  − )
1 − 2
−
Session 3 – Linear Regression
7. TRICKY QUESTIONS
Tricky Questions
• Can I drop a non significant variable?
• What if two variables are very strongly
correlated (but not perfectly correlated)?
• How do I deal (simply) with
missing/miscoded data?
• How do I identify influential observations?
Tricky Questions
• Can I drop a non significant variable?
– A variable may be non significant but still have a
significant correlation with other covariates…
– Dropping the non significant covariate may
unduly increase the significance of the
coefficient of interest. (recently seen in an OECD
working paper).
• Conclusion: controls stay.
Tricky Questions
• What if two variables are very strongly
correlated (but not perfectly)?
– One coefficient tends to be very significant and
positive…
– While the coefficient of the other variable is very
significant and negative!
• Beware of multicollinearity.
Tricky Questions
• How do I deal (simply) with missing data?
– Create dummies for missing covariates instead of
dropping them from the regression.
– If it is the dependent variable, focus on the subset
of non missing dependents.
– Argue in the paper that it is missing at random (if
possible).
• For more advanced material, see session on
Heckman selection model.
How do I identify influential points?
• Run the regression with the dataset except
the point in question.
• Identify influential observations by making a
scatterplot of the dependent variable and
the prediction Xb.
Tricky Questions
• Can I drop the constant in the model?
– No.
• Can I include an interaction term (or a
square) without the simple terms?
– No.
Session 3 – Linear Regression
NEXT SESSIONS …
LOOKING FORWARD
Next session
• What if some of my covariates are measured
with error?
– Income, degrees, performance, network.
• What if some variable is not included
(because you forgot or don’t have it) and
still has an impact on y?
– « Omitted variable bias »
Important points from this session
• REMEMBER A1 to A6 by heart.
– Which assumptions are crucial for the
asymptotics?
– Which assumptions are crucial for the finite
sample validity of the OLS estimator?
• START REGRESSING IN STATA TODAY !
– regress and outreg2

similar documents