Sie sind auf Seite 1von 5

Last par 3sls

Is 3SLSL estimating twice?


How to apply the equations in 3sls and fiml?
Compare 3SLS and FIML
Consists usage of hats for estimated
Do we use 2sls or ols to estimate errors in 3sls compares to sur?
Last para 3sls
Three stage squares is a systems method (full information method) that
combines 2SLS and SUR and it was developed by Zellner and Theil in 1962. 2SLS
corrects for endogeneity and SUR corrects for contemporaneous correlation
between the error terms in the system.
Consider a general linear model containing G jointly dependent endogenous
variables, with K predetermined variables where the ith equation is:
(1)
(2)

yi=Yibi+ Xii+ ui

i=1,.,G

yi=Zii+ui

Where Zi =
And

Yi is an nx1 vector of sample observations on the dependent variable in the ith


equation, Yi is an nxgi matrix of endogenous variables in the equations and Xi is
an nxki matrix of exogenous variables in the ith equation.
3SLS begins by estimating each structural equation of the model separately by
2sls. We premultiply each equation in (2) by X (the matrix of all the
predetermined variables in the model).
(3)

X ' yi= X ' Zii+ X ' ui

i=1,.,G

The variance, covariance matrix of the disturbance term is (using the standard
assumption that E(uiui)= i iI
(4)

and X is fixed)

E ( X ' uiui ' X )=X ' E ( uiui ' ) X

Since we have heteroskedasticity, we can estimate each equation using GLS or


OLS on suitably transformed variables(equivalent methods). For GLS we use the
standard formula:
Bgls

E ( uiui ' )= i 2

With

where here

= X ' X

Hence the GLS estimator is:

( 5 ) i =
Which is the same as writing the 2SLS estimator of (2).
We can also apply OLS on the transformed variables:
It can be shown that if a non singular matrix P exists such that:

(X ' X )1 = PP where PXXP = I


Where we can name P the square root matrix of

(X ' X )1 and pre-multiply

equation (3) by P
6

PXyi =

Which we can rewrite as


7

Wi = Wi +

Where the variance for the transformed disturbance term (vi=(PXui) is


E(vivi) = E(PXviviXP)
Therefore we can apply OLS to the equation as the variance term for the
transformed error term is homoscedastic (7) gives:
(8)

OLS

This can be seen as the 2SLS estimator in 5.


These 2SLS estimates of the coefficients are then used to calculate the residuals
in each stochastic structural equation.
Now that we have suitably corrected for endogeneity we can now apply SUR.
Using (8) on each structural equation in the systems gives us G structural
equations in total, this gives:
(9)
And in matrix form:

They are seemingly unrelated in the sense that there is no longer endogeneity
between them, but the error terms between the different equations may be
correlated. OLS on each equation would give unbiased but inefficient estimates
as we do not take into account the full use of information available.

We use the matrix as in (10).


We know that the v vector is not homoscedastic and it is made up of g different
equations where
E(U )
.
Var (v) = E(vv) = E
(11)

Where we can see that each of the terms in (11) are matrices in their own right.
We make the assumption that each structural equation has homoscedastic, non
autocorrelated error terms and the disturbance terms in the G structural
equations may be contemporaneously correlated.
Therefore we may have for example:
E( V1V2) =
Where I is the identity matrix and
is the contemporaneous correlation between
the error terms from the first and second equations in the system.
Collecting the variances and covariances together gives:

And therefore using the assumptions we can get


V = matrix

Where v =

Type equation here .

Provided that at least some of the

i j

are non zero, then we should use SURE

as there is an efficiency gain over 2SLS.


Because we have heteroskedasticity, we must use GLS, with the V defined up.
The only difficulty is that the matrix

is unknown and Zellner and Theil

suggest that we should first estimate each structure equation by 2SLS, obtaining
the residuals (equation)
:
The elements of the matrix
sij =

are then estimated by

ML estimates of sij will give biased but consistent results. Arguably, we must
correct for that using

(nki)1 /2 (nkj)1/ 2

instead of the n in the denominator

where ki is the number of parameters estimated in the ith equation.


We can also do iterative GLS until the error terms stop changing.
We now get:

^ =
V

And therefore the 3SLS estimator of:

3SLS = (W
A necessary condition for the superior efficiency of 3SLS over 2SLS s that the
specification of the model should be correct. Granted this 3SLS will be equivalent
to 2SLS if:
a)
b) if all eqns in the system are exactly identified
(CHECK THE REST OF 3SLS END PARA)
Comparison 3SLS and FIML
Theyre both full information systems of estimation
FIML sometimes has difficulty converging to a solutions but 3SLS is matrix
algebra so does not
3SLS simpler than FIML
Both allow for endogeneity and patterns among residuals
FIML must assume errors are normally distributed, 3SLS is robust to non
normality
Different numerically
FIML computationally more expensive
FIML non linear so need an iterative process to solve it
Normal errors then FIML is efficient among all estimators
To make a truly valid comparison we need to know the true parameters
If there is a misspecifcation in either then if spreads to all equations
(unlike 2SLS or OLS when it is contained within the equation)
3SLS better in small samples

As with 3SLS, this is a complete system method of estimation, but is


computationally more expensive as it involves the solution of non linear
equations. Consider a linear model containing G structural equations, where the
ith relation at time t may be written by:
Where yit denotes the endogenous variables at time t and the xit indicate the
exogenous variables. Some of the betas may equal zero as may some of the
gammas. In general notation, one of the bis (here bii) is equal to one. The model
may then be regarded as a theory explaining the determination of G jointly
dependent variables in terms of the pretermined variables and the disturbances.
The model may be written in the form:

Das könnte Ihnen auch gefallen