Upload
paer-sjoelander
View
219
Download
0
Embed Size (px)
Citation preview
8/12/2019 Panel Lecture - Gujarati
1/26
01/06/2014 1
Applied Econometrics:Panel Lecture
Chapter 16, GujaratiIn Panel Data, the Same Cross-sectional Unit Is Surveyed Over
Time.
8/12/2019 Panel Lecture - Gujarati
2/26
01/06/2014 2
Introduction
Today we will examine paneleconometrics.
One of the biggest areas of modern
econometrics. See Gujarati for introduction.
More advanced panel textbooksoutside the scope of our discussion.
Today we will introduce these topicsand look at some empiricalexamples.
8/12/2019 Panel Lecture - Gujarati
3/26
01/06/2014 3
Topics of Discussion
Types of data cross sectional/timeseries/panel.
Reexaming economic models. Types of panel.
Example of panel.
Estimating panel models. Fixed effects versus random effects.
8/12/2019 Panel Lecture - Gujarati
4/26
01/06/2014 4
Main Point of Lecture
Panel econometrics is a huge fieldand to cover everything would notbe possible. This lecture aims tointroduce you to panel econometricsusing research examples.Particularly, I want to discuss when
and why you would use fixed versusrandom effects models.
8/12/2019 Panel Lecture - Gujarati
5/26
01/06/2014 5
Terminology
Time series cross section data/micropanel data/longitudanaldata/cohort analysis/ event historyanalysis.
Many authors use the above termsinterchangeably. Will talk about
some potential distinctions.
8/12/2019 Panel Lecture - Gujarati
6/26
01/06/2014 6
Why Panel Data?
Allows one to analyse issues such aspersistency (e.g. Unemployment).
Allows one to examine complexbehavioural models (e.g. Life cycle modelsof saving/consumption).
Allows one to explicitly take in to account
individual heterogeneity.
8/12/2019 Panel Lecture - Gujarati
7/26
01/06/2014 7
Limitations of Panel Data
(i) Measurement Error that can arise in surveydata
(ii) Self-Selection in to the panel
(iii) Non-Response
(iv) Attrition
(v) Typically, micro-panels cover a short timedimension e.g. the Living in Ireland covers onlyfrom 1994 to 2001.
8/12/2019 Panel Lecture - Gujarati
8/26
01/06/2014 8
Some Examples of Panel
Labour economics, welfare economics andseveral other fields rely heavily on householdpanel studies.
Panel study of income dynamics (Michigan). Industrial economics utilises very large panel
datasets (e.g. Amadeus).
In Ireland, the ESRI ran the living in Ireland
survey panel from 1994 to 2001 (willexamine some of this later in the class).
8/12/2019 Panel Lecture - Gujarati
9/26
01/06/2014 9
Famous Panel Studies
Canadian Survey of Labour IncomeDynamics
Japanese Panel on Consumers Korea Labor and Income Panel Surveys
Household Income and Labor Dynamicsin Australia
Indonesia Family Life Surveys
8/12/2019 Panel Lecture - Gujarati
10/26
01/06/2014 10
Economic Modelling
Very similar principles of economicmodelling apply to panel as appliedto the other techniques we haveexamined.
Specification issues e.g. Omittedvariables or irrelevant included
variables. Distributional issues e.g. Non-
normality.
8/12/2019 Panel Lecture - Gujarati
11/26
01/06/2014 11
An Example of a Panel Model (1)
Grunfelds (1958) investment data.
A little dated but will help us to
work through the stuff in gujarati. Q: how does investment depend on
the real value of the firm and thereal capital stock?
Data: 4 companies for 20 years(from 1935 1954).
8/12/2019 Panel Lecture - Gujarati
12/26
01/06/2014 12
5 Ways of Estimating a Panel
Model
Assume that intercepts and slopes arethe same over time and individuals.
Assume that slopes are constant but
that intercepts vary over individuals. Assume that slopes are constant but
that intercepts vary over time andindividuals.
Assume that all coefficients vary overindividuals.
Assume that all coefficients vary over
individuals and time.
8/12/2019 Panel Lecture - Gujarati
13/26
01/06/2014 13
reg grossinv valuefirm capstock
Source | SS df MS Number of obs = 80
-------------+------------------------------ F( 2, 77) = 119.63
Model | 4849457.37 2 2424728.69 Prob > F = 0.0000
Residual | 1560689.67 77 20268.697 R-squared = 0.7565
-------------+------------------------------ Adj R-squared = 0.7502
Total | 6410147.04 79 81141.1018 Root MSE = 142.37
------------------------------------------------------------------------------
grossinv | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
valuefirm | .1100955 .0137297 8.02 0.000 .0827563 .1374348capstock | .3033932 .0492957 6.15 0.000 .2052328 .4015535
_cons | -63.30413 29.6142 -2.14 0.036 -122.2735 -4.334735
------------------------------------------------------------------------------
8/12/2019 Panel Lecture - Gujarati
14/26
01/06/2014 14
. xtreg grossinv valuefirm capstock, fe
Fixed-effects (within) regression Number of obs = 80
Group variable (i): nofirm Number of groups = 4
R-sq: within = 0.8068 Obs per group: min = 20between = 0.7304 avg = 20.0
overall = 0.7554 max = 20
F(2,74) = 154.53
corr(u_i, Xb) = -0.1001 Prob > F = 0.0000
grossinv | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------valuefirm | .1079481 .0175089 6.17 0.000 .0730608 .1428354
capstock | .3461617 .0266645 12.98 0.000 .2930315 .3992918
_cons | -73.84946 37.52291 -1.97 0.053 -148.6155 .9165748
-------------+----------------------------------------------------------------
sigma_u | 139.05116
sigma_e | 75.288894rho | .77329633 (fraction of variance due to u_i)
------------------------------------------------------------------------------
F test that all u_i=0: F(3, 74) = 67.11 Prob > F = 0.0000
8/12/2019 Panel Lecture - Gujarati
15/26
01/06/2014 15
. xtreg grossinv valuefirm capstock, reRandom-effects GLS regression Number of obs = 80
Group variable (i): nofirm Number of groups = 4
R-sq: within = 0.8068 Obs per group: min = 20
between = 0.7303 avg = 20.0
overall = 0.7554 max = 20
Random effects u_i ~ Gaussian Wald chi2(2) = 317.79corr(u_i, X) = 0 (assumed) Prob > chi2 = 0.0000
------------------------------------------------------------------------------
grossinv | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
valuefirm | .1076555 .0168169 6.40 0.000 .0746949 .140616
capstock | .3457104 .0265451 13.02 0.000 .2936829 .3977378
_cons | -73.03529 83.94957 -0.87 0.384 -237.5734 91.50284-------------+----------------------------------------------------------------
sigma_u | 152.15823
sigma_e | 75.288894
rho | .80332024 (fraction of variance due to u_i)
8/12/2019 Panel Lecture - Gujarati
16/26
01/06/2014 16
Fixed Effect Panel Model
The intercept in the regression isallowed to differ among individualsin recognition of the fact that eachindividual (unit) may havecharacteristics of their own.
Also known as the least squares
dummy variable model.
8/12/2019 Panel Lecture - Gujarati
17/26
01/06/2014 17
Why use Fixed Effects
Fixed Effects are generally usedwhen there is a correlation betweenthe individual intercept and theindependent variables.
Generally used when n is relativelysmall and t is relatively large.
8/12/2019 Panel Lecture - Gujarati
18/26
01/06/2014 18
Random Effects Model
Random Effects models assumethat the intercept of an individualunit is a random drawing from amuch larger population with aconstant mean value.
Also (less frequently) know as the
Error Components model.
8/12/2019 Panel Lecture - Gujarati
19/26
01/06/2014 19
Fixed Effects or Random Effects
IF N is large and T is small, and if theassumptions underlying RE hold, the RE
are more efficient estimators. Use Fixed Effects if the errors and the
observations are correlated (e.g.countries).
The Hausman test is distributed Chi-Squared Asymptotic around the nullhypothesis that Random Effects is
appropriate.
8/12/2019 Panel Lecture - Gujarati
20/26
01/06/2014 20
Hausman Test
Hausman (1978).
The null hypothesis is that the FE and REdo not differ substantially.
Test is distributed asymptotically chi-squared.
FE is consistent under both the null andthe alternative.
RE is consistent under the null andinconsistent under the alternative.
We can test the appropriateness of REusing critical values.
8/12/2019 Panel Lecture - Gujarati
21/26
01/06/2014 21
. hausman fixed
---- Coefficients ----
| (b) (B) (b-B) sqrt(diag(V_b-V_B))
| fixed . Difference S.E.
-------------+----------------------------------------------------------------
valuefirm | .1079481 .1076555 .0002926 .0048738
capstock | .3461617 .3457104 .0004513 .0025204------------------------------------------------------------------------------
b = consistent under Ho and Ha; obtained from xtreg
B = inconsistent under Ha, efficient under Ho; obtained from xtreg
Test: Ho: difference in coefficients not systematic
chi2(2) = (b-B)'[(V_b-V_B)^(-1)](b-B)
= 0.07
Prob>chi2 = 0.9678
8/12/2019 Panel Lecture - Gujarati
22/26
01/06/2014 22
An Example of a Panel Model (2)
Did people get better off over thecourse of the Celtic tiger?
We certainly became richer but howabout happier or more satisfied withour income?
To address this question I have been
analysing a panel of people trackedfrom 1994 to 2001.
8/12/2019 Panel Lecture - Gujarati
23/26
01/06/2014 23
An Example of a Panel Model (3)
What are the determinants of international disputes.
Beck et. al. (1998) in their paper on international
conflict, and consist of time-varying data on 827politically-relevant dyads in the international
system. Each dyad has one observation for each yea
from 19501985, inclusive. Omitting observations
with ongoing conflicts, this yields a total N = 20448
This data is taken from Chris Zorns Easter
Workshop on Panel Econometrics at Oxford.
8/12/2019 Panel Lecture - Gujarati
24/26
01/06/2014 24
Dyadid: The dyad identification number.
Year: The year identifier.
Dispute: 1 if a militarized interstate dispute occurred betweenthe members of that dyad in that year, 0 otherwise.
Start: The starting counter variable.
Duration: The duration variable.
Democ: Rescaled POLITY democracy variable ( [-1,1]). Growth: Lagged measure of growth, as a proportion of GDP.
Allies: 1 if the dyad members are allied, 0 otherwise.
contig : 1if the members of the dyad are geographically
contiguous, 0 otherwise. Capratio: The natural log of the ratio of the two states military
capacities, as measured by the
Correlates of War (COW) data.
Trade: The ratio of bilateral trade to GDP, in constant USdollars.
8/12/2019 Panel Lecture - Gujarati
25/26
01/06/2014 25
Outside the Scope (1)
Generalised Panels can utilise many differentdistributional forms with many different types ofdata.
Quasi-Maximum Likelihood Techniques thatmake less assumptions about the full distributionhave become increasingly utilised.
STATA package of choice for most researchersbut other packages such as LIMDEP and R havebecome increasingly used.
8/12/2019 Panel Lecture - Gujarati
26/26
01/06/2014 26
Outside the Scope (2)
A great deal of work has beenconducted on what to do with theproblem of unbalanced panel data.
Sample selection models to dealwith panel attrition have alsobecome increasingly important in
the literature.