Panel Data Analysis

1. Pooled OLS Method

OOSC ¿= β1 + β 2 GEE ¿ + β 2 GEE¿ + β3 log ⁡PTR ¿ + β 4 log ⁡GDPPC¿ + β 5 LRF ¿ + β 6 LRM ¿ + β 6 SOI ¿ +α ¿ + μ¿

 α ¿ = individual heterogeneity
 Note that α ¿ + μ¿ =v ¿
 We assume same intercept and constant slope for all variables, therefore all have same
constant, i.e. β
 However, there is unobserved heterogeneity that influences IDV’s affect on DV, such as
culture in countries, institutional structures, effectiveness of policies etc. These variables
are ‘time invariant’
 If error term and IDV are correlated = endogeneity. Regression assumptions say we
shouldn’t have this problem in our model
 Endogeneity leads to biased estimators
 To tackle this issue, we can use

2. Fixed Effect Model using Least Square Dummy Variable (LSDV) technique
 We create dummy variables for each section in the panel data so that the α term is
excluded from the error term. (we have 20 countries, therefore 20 dummy variables)

OOSC ¿=α 1 +α 2 D2 i +α 3 D3 i + α 4 D4 i + α 5 D5 i+ α 6 D6 i +α 7 D7 i +α 8 D 8 i+ α 9 D9 i +α 10 D10 i+ α 11 D11i + α 12 D12 i +

 Issues:
(a) Less degrees of freedom due to high levels of dummies
(b) Multicollinearity due to dummies
(c) Cannot be used with time invariant regressors(IDV) such as gender, race, ethnicity
etc.

3. Random Effects Model

 Assumes that cov ( α ¿ , x¿ )=0 .
If It is 0, we use random effects model.
 If cov ( α ¿ , x¿ ) ≠ 0 , we use fixed effects model
i.e. if there is covariance between individual heterogeneity and IDV, we use fixed effects
model
Which test to use?
 Stata
1. Breusch Pagan Test
. regress OOSC GEE log_PTR log_GDPPC LRF
> LRM SOI  HET test result gave p-value 0.000
>
Source
MS
SS
Number of obs =
df
60
 Therefore, we reject the null hypothesis.
> F(6, 53) = 8.37
 Thus, this data has problem of serial correlation and
Model
> 82.16835
2893.0101
Prob > F =
6 4
0.0000 heteroscedasticity
Residual 3053.87213 53 57
> .6202289 R-squared = 0.4865

> Adj R-squared = 0.4283

Total 5946.88223 59 10
> 0.794614 Root MSE = 7.5908

>
OOSC Coefficient Std. err.
> t
> P>|t|
> [95% con
> f. interval]

>
GEE -.9115177 .7747187
> -1.18
> 0.245
> -2.465407
> .6423713
log_PTR 5.831229 2.876863
> 2.03
> 0.048
> .0609717
> 11.60149
log_GDPPC 1.801764 2.502352
> 0.72
> 0.475
> -3.217319
> 6.820846
LRF -.3651403 .2073577
> -1.76
> 0.084
> -.7810471
> .0507665
LRM .0495209 .2973327
> 0.17
> 0.868
> -.5468529
> .6458947
SOI 4.357614 40.39635
> 0.11
> 0.915
> -76.66719
> 85.38242
_cons -1.447533 27.59109
> -0.05
> 0.958
> -56.78825
> 53.89318

>

. hettest

Breusch–Pagan/Cook–Weisberg test for het

> eroskedasticity
Assumption: Normal error terms
Variable: Fitted values of OOSC

H0: Constant variance

chi2(1) = 24.69
Prob > chi2 = 0.0000

. hettest GEE log_PTR log_GDPPC LRF LRM

> SOI

Breusch–Pagan/Cook–Weisberg test for het

> eroskedasticity
Assumption: Normal error terms
Variables: GEE log_PTR log_GDPPC LRF LRM
SOI

H0: Constant variance

chi2(6) = 39.90
Prob > chi2 = 0.0000

.
2. Fixed Effects Model

. xtreg OOSC GEE log_PTR log_GDPPC LRF LRM SOI, fe

Fixed-effects (within) regression Number of obs = 60

Group variable: ID Number of groups = 20

R-squared: Obs per group:

Within = 0.2270 min = 3
Between = 0.0002 avg = 3.0
Overall = 0.0013 max = 3

F(6,34) = 1.66
corr(u_i, Xb) = -0.3250 Prob > F = 0.1601

OOSC Coefficient Std. err. t P>|t| [95% conf. interval]

GEE .5265818 .6924704 0.76 0.452 -.8806875 1.933851

log_PTR -1.616544 1.932358 -0.84 0.409 -5.543567 2.310479
log_GDPPC -2.189626 3.656131 -0.60 0.553 -9.619779 5.240527
LRF -.3403288 .160679 -2.12 0.042 -.6668677 -.0137898
LRM .5226723 .1944253 2.69 0.011 .1275526 .917792
SOI -100.1496 61.19933 -1.64 0.111 -224.5216 24.22245
_cons 26.19831 34.47826 0.76 0.453 -43.86995 96.26657

sigma_u 10.625388
sigma_e 2.3240226
rho .95434418 (fraction of variance due to u_i)

F test that all u_i=0: F(19, 34) = 27.97 Prob > F = 0.0000

3. Random Effects Model

. xtreg OOSC GEE log_PTR log_GDPPC LRF LRM SOI, re

Random-effects GLS regression Number of obs = 60

Group variable: ID Number of groups = 20

R-squared: Obs per group:

Within = 0.1495 min = 3
Between = 0.3069 avg = 3.0
Overall = 0.2987 max = 3

Wald chi2(6) = 13.36

corr(u_i, X) = 0 (assumed) Prob > chi2 = 0.0376

OOSC Coefficient Std. err. z P>|z| [95% conf. interval]

GEE -.1020126 .6012274 -0.17 0.865 -1.280397 1.076371

log_PTR -.8738602 1.905786 -0.46 0.647 -4.609131 2.861411
log_GDPPC -3.674946 2.720854 -1.35 0.177 -9.007722 1.65783
LRF -.3780298 .1537992 -2.46 0.014 -.6794707 -.076589
LRM .4521638 .192085 2.35 0.019 .075684 .8286435
SOI -35.39953 44.9056 -0.79 0.431 -123.4129 52.61383
_cons 39.95453 26.29957 1.52 0.129 -11.59167 91.50074

sigma_u 8.0202084
sigma_e 2.3240226
rho .92253718 (fraction of variance due to u_i)
3. Hausman Test

 Results state that fixed effect model should be used because p value < 0.05
. hausman fixed random

Coefficients
(b) (B) (b-B) sqrt(diag(V_b-V_B))
fixed random Difference Std. err.

GEE .5265818 -.1020126 .6285944 .3435709

log_PTR -1.616544 -.8738602 -.7426836 .3193546
log_GDPPC -2.189626 -3.674946 1.48532 2.442181
LRF -.3403288 -.3780298 .0377011 .0465139
LRM .5226723 .4521638 .0705085 .0300754
SOI -100.1496 -35.39953 -64.75003 41.57939

b = Consistent under H0 and Ha; obtained from xtreg.

B = Inconsistent under Ha, efficient under H0; obtained from xtreg.

Test of H0: Difference in coefficients not systematic

chi2(6) = (b-B)'[(V_b-V_B)^(-1)](b-B)
= 550.82
Prob > chi2 = 0.0000
(V_b-V_B is not positive definite)

 Using sigmamore option because (V_b-V_B is not positive definite)

Results state that random effects model is more appropriate because p-value > 0.05
. hausman fixed random, sigmamore

Coefficients
(b) (B) (b-B) sqrt(diag(V_b-V_B))
fixed random Difference Std. err.

GEE .5265818 -.1020126 .6285944 .4053487

log_PTR -1.616544 -.8738602 -.7426836 .6799027
log_GDPPC -2.189626 -3.674946 1.48532 2.693327
LRF -.3403288 -.3780298 .0377011 .0682246
LRM .5226723 .4521638 .0705085 .0674672
SOI -100.1496 -35.39953 -64.75003 45.71894

b = Consistent under H0 and Ha; obtained from xtreg.

B = Inconsistent under Ha, efficient under H0; obtained from xtreg.

Test of H0: Difference in coefficients not systematic

chi2(6) = (b-B)'[(V_b-V_B)^(-1)](b-B)
= 10.14
Prob > chi2 = 0.1190