Basic Assumptions
E(ui2/X) =𝞼2 : Homoskedatic :)
- 𝞼i2: heteroskedastic :( -> ;eads to wrong
statical inferences ( 90%, 95%, 99%)
Endogenity- 1) x not independent or 2)bicausal
- All Xs are independent= > no multicollineaity and no endogeneity.
Consumption= b0 + b1 income + b2 wages
→ wage is a part of income. Which means the two independent variables income and wages are not
independent of each other– > VIOLATION → This violation leds to endogeneity in the model.
GDP = b0 +b1 Investment
→ Invetsment (indepant var) impacts gdp(dep var)
But at the same time
Investment= a0 +a1 GDP
—> BICAUSALITY when independe fafctes depend but also dependent affects independent.
→ ENDOGENITY
As n(no. of obsv) increase beta variance
Sources of Misspecification increases significantly . Consistency is key
other statistical inference gets impacted.
hetero
Linear v/s non linear.
E(u/x) is not 0 :(
Not including an important X variable
that affects Y eg. no including work ex
when find wages= b0 +b1 education.
Implication E(u/x) not equal to 0 :(
Proxy data eg. using exam marks
instead of iq
Parameters (Coefficnet) differ
significantly if I run the same regression
Cons= 10 + 0.2 Income very different form the true value (cooreclt mode :)
Corect model will give true estimates
Incorrectly specified
- Beta wil be there but not TRUE beta
- Convergence of beta will happen but not converge to true beta
—> pseudo-true value of PARAMETER
Key assumption for unbiased results
E(u/x) =0 =1/N*∑xiui = 0 : Algebra terms
Y=b0 +b1x1+u → E(u/x)=0 for unbiased
Bias if E(u/x) =0 is not true i.e
this assumption is violated :(
TRUE form: non linear ESTIMATED form: Linear
Coefient . Coefficient .
How does y change if x changes? How does y change if x
changes?
dy/dx= g’(x).=
dy/dx= β
If β converges to g’(x): unbiased and consistent
If β not converges to g’(x): biased and inconsistent Error terms from functional form misspeciation. We
minimise this error square
GLS will ensure that even β converges to g’(x)
actual estimate
Dependent variable is :
OLS - What is the consumption levels based on income:
Consumption = b0 +b1 income + ui. —> Consumption is a continuous
Non-OLS- The number of cars i own
No. of Cars owned= b0 +b1 income. → No. of cars is discrete (whole
numbers)
- whether or not I get a job ( 1 is yes, 0 if no)
Non-OLS
Job dummy ( 1 yes job, 0 no job) = b0 + b1 educ → Depent variable is
binary ( only takes 1 or 0 value(
E.g i wrongfully put OLS when dep var was binary. Regression estimate→ Job dummy = 0.8 + 0.3 edu
Now what would be the answer if education is 10 year.
0.8 + (0.3)*(10) = 0.8+3= 3.8 so estimated value is 3.8
Instrumental Regression
Y = b0 +b1x1 +b2x2 +b3x3 +ei
Weight = b0 +b1height +b2food intake +b3familyBMI+ ei
Now what if my x2 is endogenous coz x2 depends on x1
→ food intake depends on height
Food intake = a0 + a1 Height +ui
IV regression
Weight = b0 +b1height +b2(a0 + a1 Height +ui) +b3familyBMI+ ei
Weight = b0 +b1height +b2a0 + b2a1 height + b2ui +b3familyBMI+ ei
Weight =b0+b2a0+ b1height + b2a1 Height+ b3familyBMI+b2ui+ei
Weight = G0 +(b1+b2a1) Height + b3familyBMI+ vi
Weight = G0 + G1Height +b3familyBMI+ vi : STRUCTURAL EQuation
Gdp = b0 + b1techonological progress + b2 Cocumption + b3x3 +ei
Countries mix of countries such as
Sudan
India FIXED EFFECTS
Bangladesh MODEL TO TAKE
Vanatau CARE OF COUTEY
FIXED EFFECTS
Skorea
AND ENDOGENEITY
Japan
US
UK
Omitted variable Bias
Crop = b0 +b1rainfall + b2subsidies +b3soil quality+ ei
E(u/x) =/= 0: Biased and
inconsistent estimates
(N^-1)=1/N
(N^-1)-1= N
(X’X)^-1 (X’X)B + (X’X)^-1(X’Za) + (X’X)^-1X’v
=/= 0
N 1/N
y = 0.8 + 0.2x1—> OVB and the estimated betas were inconsistent and biased.
y = 0.8 - 0.1x1 +0.9x2
E. g returns to schooling and ability
Wages = b0 + b. schoolingyears + (a. ability + ei)
It follows that δα > 0, so the omitted variables bias is positive in this
example.
ESTIM:
Earning = b0 +
b1(s) schooling
Correl between schooling and ability
TRE MODEL:
Earning = b0 + b1(
‘) schooling+b2ability
