Desarrollo y sκκgλ
SEPTlEMBtE M 19„ ““"Ж" ЗВ
for robberies, whereas it is close to zero and very far from significance for
homicides. This goes in the direction suggested by simple theory.
As recalled above, the ambiguity of pure cross-sectional estimates is well-
known. One way of eliminating it is to use panel data and to control for
country fixed-effects. This is what the study of FLL does. However, it also
explicitly takes into account the hysteresis effect of criminality we referred
to in the previous section by explicitly allowing the crime rate of a given
year to depend on that of the previous year. This rules out standard fixed-
effect estimation and requires estimating an auto-regressive model in first
differences. They do so on reduced samples of countries defined by the
availability of all variables of interest after taking first differences and
lags. They also instrument some of the explanatory variables by lagged
values of the variables of the model so as to avoid endogeneity problems.
The resulting estimates are reproduced in Table 1.
Table 1 Panel Regressions of Crime Rates: First Difference
Auto-regessive Models01
(ρ-values in italics)
Explanatory variables
Hnmicide rate (growth rate)
Robbery rate (growth rate)
Difference in.
|
Gini coefficient1’' |
0.036 0.000 |
0.011 0.009 |
|
Urbanization rate |
0.004 0.063 |
0.011 0.000 |
|
GDP per capita (log) |
•0.207 O.(XX1 |
-0.045 0.035 |
|
GDP growth rateb' |
-0.036 0.001 |
•0.072 0.000 |
|
Drug possession crime rate |
0.001 0.(M7 |
0.001 0.079 |
|
Secondan, enrollment rate |
0.009 0.(XX) |
0.002 0.101 |
|
Lagged homicide rate |
0.640 0.000 |
0.839 0.000 |
|
Number of observations (countries) |
5b(2O) |
50 (17) |
a, GMM estimates. Second lags and third lags of dependent and independent variables used as
instruments with the exception of the lagged crime rate for which third lag is used as an instrument.
b∙, Strictly exogeneous.
Source: Fakjnzylber, Lederman and Loayza (1998).
87