3.4. The determinants of personal income inequality
We move next to the determinants of personal income inequality. Table 5 reports our estimates of
equations (21) and (22) for the largest available sample.17 The 1st column of table 5 abstracts from
country and year fixed effects, which are subsequently included in the 2nd and 3rd columns; a linear time
trend and dummies controlling for changes in definitions are also included.18 We find that all variables
have significant coefficients with the expected signs, with the exception of the unemployment rate. The
labour share has a negative coefficient, while the wage differential has a positive one.19 The
unemployment benefit appears negatively related to income inequality, whereas the unemployment rate
has an insignificant coefficient. However, the unemployment rate has a positive and significant sign when
we move to instrumental variable estimation. The linear time trend bears a negative and significant
coefficient, indicating an unexplained decline in inequality over the sample period.
The comparison between the OLS results obtained in the 3rd and 4th columns and the IV
estimates reported in 5th and 6th columns indicates that OLS-estimation provides downward-biased
estimates of the actual effect of the labour share and wage inequality on income inequality, and an upward
bias for the effect of the unemployment benefit.20 This bias could be merely due to measurement errors,
but it could also indicate that some unobservable variable, which correlates with both income inequality
and labour market institutions — such as the political orientation of the government or the attitude of the
population towards redistribution — has been omitted. It is interesting to note that, while the impact of
passive labour market policies remains significant and negative, the unemployment rate and the time trend
gain statistical significance under IV estimation. As a robustness check, columns 7 and 8 report the same
model estimated in first differences, with and without country fixed effects: the coefficient on the labour
share retains its sign and significance, even if the effect is attenuated, whereas the wage differential is close
to non-significance, while the variables related to unemployment are both insignificant.
17 The sample size hinges crucially on the availability of data on wage differentials. If we concentrate on personal
income inequality only, the available sample is made of 233 observations. When we consider the overlapping with
information on wage differentials, the sample is further reduced to 142 observations. In order not to loose too many
observations, we have replaced the missing observation for the p9010 variable with its country-specific sample mean.
The sample reduction due to the availability of data on wage differentials (2nd and 5th columns) does not affect sign
and significance of the other regressor (details available from the authors). This fictitious enlargement of the sample
allows us to retain relevant information that otherwise would be excluded due to missing observations on earnings
differentials. Fort his reason, in the sequel we will consider this extended sample.
18 The controls for definition include whether the income is gross or net, and whether the recipient is household
equivalent or person equivalent. We also experimented with errors clustered by countries, without significant
changes (available from the authors).
19 Kenworthy (2003) uses household income inequality and personal earnings inequality (proxied by p90/p10 ratio)
computed from LIS (Luxemburg Income Study), with one observation for 14 countries. By regressing the former
onto the latter, he finds a coefficient comprised between 0.61 and 0.68, depending on various specifications, which is
much lower than our figures. But sample size and countries are not comparable.
20 The instrument have been selected from the regressors used in tables 2, 3 and 4 so as to satisfy the Sargan test for
overidentifying restrictions.
19