Panel data analysis - PowerPoint Presentation

Slide1

Panel data analysis

Esman M. NyamongoCentral Bank of Kenya

Econometrics Course organized by the COMESA Monetary Institute (CMI) on 2-11 June 2014, KSMS Nairobi Kenya

1Slide2

Dynamic panel estimation

DAY 9

2Slide3

Dynamics….

…. Economic issues are dynamic in nature and use the panel data structure to understand adjustmentDemand (present demand depends on past demand)Dynamic wage equationEmployment models

Investment of firms3Slide4

Dynamic panel estimation

A dynamic panel model contains at least a lagged variable. Consider the following: with: if i=j and s=t

Here the choice between FE and RE formulation has implications for estimations that are of a different nature than those associated with the static panels.

4Slide5

If the lagged dependent variables also appear as explanatory , strict

exogeneity of the regressors no longer holds.The lagged variable introduces endogeneity problem

The LSDV is no longer consistent when N tends to infinity and T is fixed.In addition, the initial values of a dynamic process raise another problem.It turns out that with a RE formulation, the interpretation of a model depends on the assumptions of initial conditions

The consistency property of MLE and the GLS estimator also depends on this assumption and on the way in which T and N tend to infinity.

5Slide6

The problem with LSDV in DP

The LSDV estimator is consistent for the static model whether the effects are fixed or random.therefore need to show that the LSDV is inconsistent for a dynamic panel data with individual effects, whether the effects are fixed or randomThe bias of the LSDV estimator in a dynamic model is generally known as dynamic bias or

Nickell’s bias (1981)Nickell, S. 1981’ Biases in Dynamic Models with Fixed Effects, Econometrica, 49, 1399-1416.

Proof needed if possible

6Slide7

The LSDV for dynamic individual-effects model remains biased with the introduction of exogenous variables if T is small;

In this case, both estimators and are biased.What is the way out?ML or FIMLFeasible GLS

LSDV bias corrected (Kiviet, 1995)IV approach (Anderson and Hsiao, 1982)GMM approach (Arellano and Bond, 1985)

7Slide8

The GMM

The GMM framework provides a computationally convenient method of performing inferences without the need to specify the likelihood functionSome hints on momentsBackground

materialThe moment conditions

8Slide9

definition of moments

A population moment , v, as the expectation of some continuous function m(.) of a (discrete) random variable z describing the population of interest:The first momentThe population mean, v

1, or first moment about the origin, measures central tendency and is given by:In which case m(.) is the identity function

9Slide10

The second moment

The second population moment is given by:The population variance of z is a measure of spread in a distributionIt is defined as the second moment of z centred about its mean and can be expressed as a function of the first two population moments, v1 and v2:

10Slide11

The above are population moments

However, we rarely obtain information on the entire populationWe therefore use a sample {zi: i=1, 2….T}

In estimation, we therefore need to define the sample momentSample moments is a sample version of the population moments

Expectation replaced by the sample average

11Slide12

Orthogonality /moment conditions

Seeks to set expectations of functions of data, z, and unknown parameter, , to particular valuesUsually zeroMean of the data is:

Giving the population orthogonality or moment condition:Where and contains one unknown parameter, such that

12Slide13

Moment restriction for mean

Stated as sample moment as:This is the method of moments

estimator of Under random sampling, this estimator will be unbiased and consistent for regardless of the other features of the underlying population.

13Slide14

Moment restriction for variance

Given as:This contains 2 unknowns, such that This means that 2 moment conditions are needed. Here we need to include the moment condition for the mean to estimate the variance

Resulting in a system with 2 unknowns and

14Slide15

Moment restriction for covarianceStated as:

It involves 3 unknowns, , resulting in the following system:

15Slide16

….GMM

The GMM also uses the moment or orthogonality conditions and a key ingredient of GMM is the specification of the appropriate moment, or orthogonality condition,

An important approach to model specification testing is to base tests directly on certain conditions that the error terms of a model should satisfy.The moment conditions in GMM are therefore commonly based on the error terms from an economic model.

16Slide17

The intuition:

The basic idea is that if a model is correctly specified, many random quantities which are functions of the error terms should have an expectation of zero.The specification of a model sometimes allows a stronger conclusion, according to which such functions of the error terms have zero expectation conditional on some information set.

17Slide18

Back to our dynamic model

A dynamic panel model contains at least a lagged variable.

with: if i=j and s=t

The dynamic relationship is

characterised

by the presence of lagged dependent variable (Y

it-1

) among the

regressors

Including the lagged var. introduces

endogeneity

problem

Recall in FE, Y is a function of individual effects therefore it lag is also a function of these effects

18Slide19

Therefore Yit-1 is correlated with the error term => OLS cannot solve our problems.

FE cannot manage cos Yit-1 is correlated with individual effectsTo overcome this problem we use GMM.

Arellano and Bond estimatorArellano and Bover estimator

19Slide20

Arellano and bond estimator

To get consistent estimates in GMM for a dynamic panel model, Arellano and Bond appeals to orthogonality condition that exists between Y

it-1 and vit to choose the instrumentsConsider the following simple AR(10 model:

To get a consistent estimate of as N-> infinity with fixed T, we need to

difference

this equation to eliminate individual effects.

20Slide21

Consider t=3

In this case yi1 is a valid instrument of (Yi2-y

i1), since it is highly correlated with (yi2-yi1) and not correlated with (vi3-vi2)Consider t=4

What are the instruments?

21Slide22

For period T, set of instrument (w) will be:

The combination of the instruments could be defined as:Because the instruments are not correlated with the remaining error term, then our moment condition is stated as:

22Slide23

Pre-multiplying our difference equation by w

i yields:Estimating this equation by GLS yields the preliminary Arellano and Bond one-step consistent estimatorIncase there are other regressors

we have:Some practical exercise!

23