Sie sind auf Seite 1von 3

Tarun Singh

Economics 1123, Problem Set 6

Table 1:
Limited Dependent Variable Models of the Determinants of Terrorism
(1) (2) (3) (4) (5) (6)
Dependent ftmpop ftmpop ftmpop ftm_pos ftm_pos ftm_ord
variable:
Regression OLS OLS Tobit OLS probit ordered
method: probit
Data set: ftmpop all all all all all
>0
Regressor:
Lngdppc -.0410 -.0365+ -.0973+ -.0420+ -.1066+ -.1190*
(.0444) (.0205) (.0564) (.0241) (.0623) (.0548)
Intercept .4883** .2540** -.1584 .5164** .0408 __
(.0995) (.0560) (.1035) (.0439) (.1105)
z-cutoff #1 __ __ __ __ __ -.0505
(.1097)
z-cutoff #2 __ __ __ __ __ .8269
(.1231)
N 76 156 156 156 156 156
Notes: Heteroskedasticity-robust standard errors are given in parentheses.
Coefficients are individually statistically significant at the +10%, *5%, **1%
significance level.

Terrorism and GDP


4
3 2
ftmpop
1 0
-1

-2 0 2 4
lngdppc

ftmpop Fitted values


Fitted values Fitted values

Green line=(1)
Red line=(2)
Orange line=(3)
1. a) It makes sense to use the Tobit model for this regression because there
is a lower bound on the number of terrorism fatalities at 0 so therefore this is
basically a floor at 0. The Tobit model will not reveal a latent variable because
we can’t have a negative number of terrorism related fatalities. So the Tobit
model makes sense to use when looking at it mathematically but the use of
the model is qualitatively unmotivated.

b) The reason why would use this regression is to see if there is a relationship
between the level of gdp per capita and the presence of terrorism. This
doesn’t measure the amount of terrorism but instead just looks at the
presence of terrorism. I think this is a good question to ask despite the fact
that it doesn’t measure terrorism levels.

c) Yes, By using the ordered probit regression we are breaking up incidences


of terrorism related fatalities into three ranges which allows us to test the
effects of gdp per capita on terrorism related fatalities without being sensitive
to outliers that would be present had we used the data the same way we did
problem set 4, but gives us more information than from part b above.

d) The coefficients on lngdppc in regressions (1) and (3) are different due to
the OLS regression (1) not including data points when ftmpop=0. The Tobit
regression (3) includes data points when ftmpop=0 and also uncensors the
data by including projections into negative values for ftmpop which would
have been theoretically present had the data not been “censored.”

e) The coefficients on lngdppc in regressions (2) and (3) are different due to
the OLS regression (2) includes values when lnftmpop=0 but has a floor at 0,
meaning it doesn’t allow for negative rates of terrorism related fatalities.
Tobit regression (3) has no floor and includes negative rates of terrorism
related fatalities as it “uncensors” the data. Because the “uncensored” data
includes negative values this will lead to more negative values for data
meaning the line will be pushed down therefore leading to a larger negative
coefficient in regression (3). Thus, a difference between the coefficients in
regressions (2) and (3) can be explained by the “uncensored” data in
regression (3).

f) The reason for the Tobit regression being mainly negative is due to the fact
that the “uncensoring” of the data in the Tobit regression causes negative
projections of terrorism related fatalities when the OLS regression (1) has a
floor at 0 so there are no negative values. The Tobit regression doesn’t make
sense, because it projects a negative amount of terrorism related fatalities,
when we would expect these values to be positive.

g) The standard error on this predicted change is .0241 and the estimated
change in the probability of at least one terrorist fatality occurring is a change
of -.0420.This value is the same as the coefficient in regression (4) because
we have a linear-log specification with a 100% change in X, meaning we have
a change in Y of 100*.01B=B.
h) Using the probit model from regression (5) the change in the predicted
probability of at least one terrorist fatality is stdnorm(-.0658) – stdnorm(-.
1724) = 4.22%.

i) Using the ordered probit model from regression (6), the predicted change of
at least one terrorist fatality is (1-stdnorm(-.1875)) – (1-stdnorm(-.0685)) =
-4.7%.

j) Using the ordered probit model from regression (6), the predicted
probability of a high level of terrorism fatalities is (1-stdnorm(1.0649)) – (1-
stdnorm(.9459))= -2.9%.

2) Including all of the data points and not omitting those with 0 terrorist
fatalities we see that the level of lngdppc does affect the level of terrorist
fatalities. Regression (2-6) all include data when ftmpop=0 and are have
statistically significant coefficients on lngdppc at the 10% level. Regression
(6) includes the most information due to the inclusion of separate z- cut offs
and is significant at the 5% level. Thus, the ordered probit regression (6)
shows a strong relationship between the level of lngdppc and the level of
terrorist fatalities. However, it is important to remember that none of the
regressions in this problem set included the variables such as political
freedoms, ethnic diversity, etc. which were included in problem set 4, so
therefore there is a chance that the regressions in this problem set may be
affected by omitted variable bias.

It is also important to acknowledge that there are more things that need to be
explored in this data. Firstly, we used a lot of Tobit regressions despite the
fact that we determined in question 1 part f) that “uncensoring” the data by
using the Tobit model didn’t make sense, since this would create negative
values for terrorist fatalities. Similarly, we never tested for a quadratic
relationship in the data. For instance, in problem set 4 we determined that
there was a quadratic relationship between lackpf and lnftmpop where
countries with medium levels of political freedoms had the most terrorist
fatalities and this may still be the case. Thus, looking at the potential OVB,
the use of Tobit, and the potential for a non linear relationship in the data, we
can’t safely conclude that there is a negative relationship between lngdppc
and ftmpop. I would suggest further exploration of the abovementioned
sources for error.

Das könnte Ihnen auch gefallen