29
TCG on wages is, therefore, quite robust.
Table 6: Proportion of regressions in which zero effect on wages cannot be rejected. (in %)
educ
age
gndr
urban
dmy
tx Rt
tx Rv
ln TCG
IG
90% CI
93.75
6.25 0 0
0
87.50
0
0 0
95% CI
all
18.75 6.25
12.50
0 all
12.50
0 0
99% CI
all
62.50 18.75 25.00
0 all all 25.00 37.50
for the 16 of 16 regressions they are in
for the 8 of 16 regressions they are in
For the education and tax rate variables, zero effect on wages was in the 95% confidence
intervals for all regressions, so any purported effect of these variables’ effects on wages is highly
questionable.
A summary of the averages of the maximum effects of each variable on wages for all 16
regressions are shown in
Table 7
. From this we see that urban share of the population exerted
the greatest effect on wages at 32.7%, while education and tax rates had low average effects. I
include a Low, Medium, or High robustness assessment from confidence interval data.
25
Table 7: Maximum Effects of Variables on Wage Rates (average of all 16 regressions)
educ
age
gender urban
dmy tax rate tax rev ln TCG
IG
max effect (%)
2.1 -16.0 -19.5 32.7 31.5 -3.1 12.5 -15.0
-19.2
robustness
VL M H H VH VL M
H H
Four of the variables have an effect on wages in the direction we conventionally expect:
as the women-to-men in the workforce ratio increases, wages decrease; as transportation costs
(TCG or IG) increase, wages decrease; as urban share of the population increases, wages
increase. Tax rates and education have meager effects with negligible robustness. Two variables
25
I checked the DW statistic for serial correlation in the residuals because if it is present, then OLS results can
be unreliable. Fortunately, for the 16 regressions it ranged from 1.84 to 2.12 with an average of 2.00,
which is good. The two dU values for 96 observations, 7 independent variables, and a 95% confidence
level are 1.82 for positive and 2.18 for negative serial correlation. The closer DW statistics are to the
highest score of 2.00, the less likely serial correlation is present. All 16 DW statistics are between the two
dU values so we can accept the null hypothesis with a 95% confidence level that there is no serial
correlation in the residuals.
August
2005
Convention