Advanced econometrics Winter semester 2009/2010

Homework #3-Solution
Enkhjargal Togtokh

Problem 2

Cobb Douglas production function relationship investigating the productivity of public capital in
private production

ln Y    1 ln K1   2 ln K 2  3 ln L   4 Unemp  u
(1)

where Y is gross state product, K1 is public capital which includes highways and streets, water
and sewer facilities and other public buildings and structures, K2 is the private capital stock
based on US Bureau of Economic Analysis national stock estimates, L is labor input measured as
employment in nonagricultural payrolls. Unemp is the state unemployment rate included to
capture business cycle effects.

The description of the used data set as follows as:

Variables Description

STATE Name of state

Code Numeric code of state

Abbrev State abbreviation

Year Year

Private Private capital stock

Public Public capital stock

GSP Gross state product

Empl Labor input measured by the employment in nonagricultural payrolls

Unemp State unemployment rate

The task is to choose the convenient model that gives the best result.

i) Estimation of parameters in the one way error components fixed effects model and
comments.

Assumption of the fixed effect model to compute parameters is that each and every single state in
our study has its own intercept, and it is done intuitively by including 47 dummy variables. We
assume that the betas pool across states, so in essential, we have 47 parallel regression lines.
Observations in each state across time vary around baseline level specific that state. Fixed effect
model assumes significance of country specific factors.

Fixed-effects estimates using 816 observations Included 48 cross-sectional units

Time-series length = 17
Dependent variable: LGSP

coefficient std. error t-ratio p-value

-------------------------------------------------------------------------------------------------------------------
const 2.35290 0.174813 13.46 3.28E-037 ***

LPublic 0.292007 0.0251197 11.62 7.08E-029 ***

LPrivate -0.0261497 0.0290016 -0.9017 0.3675

LEmpl 0.768159 0.0300917 25.53 2.02E-104 ***

Unemp -0.00529774 0.000988726 -5.358 1.11E-07 ***

-------------------------------------------------------------------------------------------------------------------

Mean of dependent variable = 10.5088 0.00001)

Standard deviation of dep. var. = 1.02113 Durbin-Watson statistic = 0.389711
Sum of squared residuals = 1.11119 Log-likelihood = 1534.53
Standard error of the regression = 0.0381371 Akaike information criterion (AIC) =
Unadjusted R-squared = 0.99869 -2965.06
Adjusted R-squared = 0.99861 Schwarz Bayesian criterion (BIC) =
Within R-squared = 0.94134 -2720.43
Hannan-Quinn criterion (HQC) = -2871.18
F-statistic (51, 764) = 11441.6 (p-value <

Comments

This estimator is called “within estimator” as it regresses y it-mean(yi) on xi-mean(xi) thus it looks
how changes in independent variables affect the dependent variable to vary around its mean
within the state.

Labor input and public capital stocks have both positive and considerable effect on gross state
product, while private stock capital and unemployment negative and moderate effect on the
independent variable.

Due to log-log form of equation we can interpret the results in terms of elasticities. For example
if Empl rises by 10%, GSP goes up by 6.8%. In case of Unemp the form is log-level with
interpretation %Δy = (100β)Δx. Unemp is expressed already in percentages (0-100, not 0.01-1).
So If Unemp rises by 1 percentage point (Δx=1), gross state product declines by 0.5%.

On the top of that private capital stock appear to be statistically insignificant or has no explaining
power in our model.
 ii) Test the hypothesis, that fixed effects are jointly insignificant,

The null hypothesis is that our simple, restrictive model was appropriate, that all of the states
share the common intercept. The alternative is that they vary across units, so the way to test this
is by running both models and then comparing their sum of squares in a joint F-test.

(RRSS−URSS )/(N −1)

F o= F N −1 , N (T −1)−K
URSS /( NT −N−K )
RRSS: residual sum of squares being that of OLS on the pooled model
URSS: residual sum of squares being that of LSDV regression
N: number of observation (48)
T: number of years of observation (17)

H 0 :the¿ effects are jointly insignificant ,i . e . all the states have a common intercept

Gretl automatically reports the F test.

Test statistic: F(47, 764) = 75,8204 with p-value = P(F(47, 764) > 75,8204) = 1,16445e-253.
Since we have a sufficiently low value p value than the usual 1, 5 or 10% level of significance
we claim that the states have no common intercept. That is a result in favor of fixed effects
model – we should get different results in different countries.

iii) Estimation parameters by random effects model and comments on result.

Random-effects (GLS) estimates using 816 observations Included 48 cross-sectional units

Time-series length = 17
Dependent variable: LGSP

coefficient std. error t-ratio p-value

-------------------------------------------------------------------------------------------------------------------
const 2.13541 0.133461 16.00 2.77E-050 ***

LPublic 0.310548 0.0198047 15.68 1.34E-048 ***

LPrivate 0.00443859 0.0234173 0.1895 0.8497

LEmpl 0.729671 0.0249202 29.28 3.66E-129 ***

Unemp -0.00617247 0.000907282 -6.803 1.99E-011 ***

-------------------------------------------------------------------------------------------------------------------

Mean of dependent variable = 10.5088 'Between' variance = 0.00692327

Standard deviation of dep. var. = 1.02113 theta used for quasi-demeaning = 0.888835
Sum of squared residuals = 7.08184 Akaike information criterion (AIC) =
Standard error of the regression = -1547.75
0.0933889 Schwarz Bayesian criterion (BIC) =
'Within' variance = 0.00145444 -1524.22
Hannan-Quinn criterion (HQC) = -1538.72

Comments:
Effect of public capital stock and labor input measured by the employment in nonagricultural
payrolls are both positive and considerably high as in the previous case. The capital stock of
private sector has now positive relationship with the independent variable however in a moderate
scale. Unemployment is significant. AIC got worse.

iv) Hausman specification test, model selection and comment.

Now that we have 2 different estimators of betas from model (1). We use Hausman specification
test in order to decide whether to use Random Effect or Fixed Effect model.

Hausman test assumes:

H0: E(uit/Xit)=0. Null hypothesis is in favour of residuals not being correlated with explanatory
variables.
A: E(uit /Xit) ≠ 0 Alternative is completely the opposite.

Asymptotic test statistic: Chi-square(4) = 9,52542 with p-value = 0,0492276.

Hausman’s null hypothesis is rather critical assumption making fixed effects consistent
inefficient and random effects consistent efficient while alternative makes fixed consistent and
random inconsistent.

The result here is not straightforward, we can decide to reject but also not to reject null
hypothesis. We have seen in ii) fixed to be favoured. Fixed is more stable - always consistent.
The biggest mistake is inefficiency if we decide for null. In case of random effects we can get
inconsistency- that seems worse than inefficiency. Then we decide for fixed effects model.