Far away from a skill-biased change: falling educational wage premia in Italy

Abstract

In this article we apply quantile regressions to investigate the evolution of Educational Wage Premia (EWP) in Italy from 1993 to 2004. Using the Survey of the Household Income and Wealth (SHIW, Bank of Italy) and different classifications for educational attainments, we show that, in the private sector, EWP have generally decreased over time, considering both continuous and categorical specifications for education, at all quantiles of the wage distribution. Different patterns are observed in the public sector, where EWP remain basically stable over time. A number of robustness checks and various econometric specifications are also applied in order to address sample selection issues. Our findings provide additional evidence in favour of the thesis that the increasing patterns in inequality and EWP, and the related interpretations concerning skill-biased changes, are much less pronounced in continental Europe than in Anglo-Saxon countries.

Acknowledgements

We wish to thank Andrea Brandolini, Piero Cipollone, Bart Cockx, Muriel Dejemeppe, Claudio Lucifora, Andrea Neri, Paolo Piacentini, Franco Peracchi, Daniela Vuri, Henri Sneessens, Bruno Van der Linden and Robert Waldmann, as well as seminar participants in the Brucchi Luchino 2006 conference, the AIEL 2006 conference, the DSE-ISFOL workshop, and the meeting of the IEIILM research group.

Notes

¹ Peracchi (2006) distinguishes between returns to education, which is a measure of the causal effect of an extra level of schooling on the worker's earnings, and educational wage premia, which is a measure of statistical association between levels of schooling and wages. We make use of this terminology in the article.

² For a detailed explanation on how Instrumental Variables–Local Average Treatment Effect (IV-LATE) estimates can change using different instruments and different groups of compliers, see Angrist et al. (1996). Furthermore, as also stressed by Peracchi (2006), IV estimates of returns to education usually exceed OLS estimates, even though they also tend to be less precise, possibly because of a weak instrument problem. See also Ashenfelter et al. (1999).

³ See Naticchioni and Ricci (2008), who carry out an analysis of the effects of the dynamics of EWP on inequality trends in Italy, in both the private and the public sector.

⁴ For another analysis concerning the Portugal case, see Hartog et al. (2001)

⁵ For analysis concerning other countries, see Magoula and Psacharopoulos (1999) for Greece, Palme and Wright (1998) for Sweden.

⁶ Also note that we exclude 0.05% of the observations in both the right and the left tail.

⁷ We define a public employee using two variables in the database, APSETT and DIMAZ. APSETT provides us with self-declaration of the sector in which the individual works, including the public sector, while DIMAZ refers to the firm size, and it is specified when the employee declares that he/she is employed in the public sector. We consider as public employees those workers who declare for both questions that they are employed in the public sector. Results do not change much when we consider definitions of public employee based on APSETT and DIMAZ separately.

⁸ We also used different definitions for experience. For instance, in order to take into account individuals working during their scholastic career, we also defined the experience variable as: age-years of formal education-6. More specifically, we use this modified experience measure when it is lower than the measure used in this article; otherwise, we continue using the measure in the article. However, since in Italy the practice of working during the university period is much less developed than in other European countries, this measure for experience is very highly correlated with the one used in this article (0.99). For this reason, EWP do not change using this different measure for experience.

⁹ We cannot consider all the other less widespread kinds of tertiary degrees (10% of the graduates in the sample), since the SHIW database put them in the same category, without any distinction. We drop these observations from this kind of analysis.

¹⁰ For further discussion on methodological grounds and techniques used to perform point and interval inference, see Koenker and Basset (1978) and Buchinsky (1994).

¹¹ For further a more general specification of the function Φ, see Arias et al. (2001).

¹² In particular, by applying the probability integral transformation theorem to the conditional quantiles of the ability distribution, i.e. , it is possible to make explicit in Equation 4 the specific effect of unobserved ability for each selected quantiles, that is , where G_a is some monotonic transformation of the ability distribution in the population.

¹³ In a cross-sectional analysis this represents an advantage in using quantile regressions rather than the OLS approach, which can only estimate an average treatment effect of education, i.e. the EWP for an individual with mean ability.

¹⁴ Note also that the intertemporal comparison of OLS estimates of EWP would be unaffected by ability bias if it is assumed that the average ability in the population is time invariant. In this case, however, the results would hold only for the central tendency of the data and cannot be generalized to the dynamics of the whole wage distribution.

¹⁵ In particular, only women born in the 1949–1956 period experienced higher real wages than women belonging to the previous and following cohorts (2.6% in comparison to the both 1932–1948 and 1957–1967 cohorts).

¹⁶ We decided to begin our analysis in 1993 because in 1992–1993 the former wage indexation mechanism (scala mobile) was replaced by a completely new bargaining system. Since then, the bargaining structure of the wage setting has not changed, and can be described as a two-tier system: national contracts are devoted to preserve the purchasing power of wages, whereas decentralized wage bargaining at firm level should be related to rent-sharing, in case of positive surplus.

¹⁷ Education in the continuous specification is computed attributing 5 years for elementary school attainment, 8 years for lower secondary school, 13 years for upper secondary school and 17 for more highly educated workers. Note also that our classification for education is consistent with the International Standard Classification of Education (ISCED).

¹⁸ In particular, we perform simultaneous quantile regressions obtaining an estimate of the variance–covariance matrix via bootstrapping. The SEs are based on the heteroscedastic bootstrap methods where the sample size is equal to the number of observations each year. Further, to validate the heteroscedasticity hypothesis of the quantile regressions we successfully test that the coefficients estimated at different quantiles are statistically different from each other (Koenker and Basset, 1978; Buchinsky, 1994).

¹⁹ Due to limits of space, we do not provide coefficients for all the years after 1993, but only variations. However, the coefficients can easily be computed from the level in 1993 and the variations over time.

²⁰ To test whether the variation over time of EWP coefficients is significant, we assume that each coefficient is distributed normally and that the population in 2004 is independent from the 1993 population. This is quite plausible, since only 10% of the SHIW population in 2004 were also interviewed in 1993, in the panel component of the SHIW.

²¹ It might be argued that carrying out two separate sets of regressions for 1993 and 2004, we cannot control for some aggregate omitted variables that could change over time, such as the business cycle. For this reason, we have also pooled the 2 years together, introducing interacted variables (the time dummy for 2004 times the dummies of lower secondary, upper secondary and tertiary education) along with our standard variables (gender, education and experience). In this framework, the variations in EWP are given by the coefficients of the interacted variables. In such a way, we can control for time dummies, which can be considered as a proxy for unobserved aggregate omitted variables. The results do not change: the interacted terms are negative and significant for all educational attainments, while the time dummy is not significant, suggesting that the business cycle have not so different between 1993 and 2004.

²² Note that the share of individuals with an upper secondary degree increased in the private sector from 33 to 46%, in the period 1993–2004. Hence, the falling EWP for this category might be explained using a demand–supply paradigm.

²³ The estimated coefficients of the educational dummies have to be interpreted as differentials with respect to the omitted category, i.e. having a primary school degree (which includes also those who have not achieved any educational degree). Note that the related share of the omitted dummy decreased over time, since educational levels are increasing in Italy. This means that belonging to this category should be increasingly related to unskilled and low-paid occupations in the labour market. Ceteris paribus, the premia of having lower secondary, upper secondary and tertiary degrees could have increased over time, since the labour market should have rewarded the omitted category even less. This is not the case: lower secondary, upper secondary and tertiary degrees reduce their EWP with respect to the omitted category.

²⁴ The strategy of carrying out several robustness checks for different sub-groups and different specifications allows us to control to some extent for the endogeneity of schooling, as credible IVs or randomized experiment are not available for an analysis over time. As already stressed, previous papers on returns to education in Italy derived convincing IVs in the SHIW data, exploiting information concerning the schooling reforms in the late 1960s (Brunello et al., 2001). However, these IVs become much less convincing when the focus of the analysis is the time dynamic of EWP. Actually, since the effects of schooling reforms change according to the population sub-group involved in the reforms, the group of compliers affected by the instruments changes over time, affecting in turn the dynamic comparison of the estimates. see Angrist et al. (1996).

²⁵ Note that when we include self-employed we have to consider yearly labour income instead of monthly labour income, for both employees and self-employed. For this reason, the coefficients of this column cannot be compared to those of the other columns.

²⁶ In Table A1 in Appendix we report the descriptive statistics for all these addictional variables, except for regional dummies (the sample design of the SHIW data guarantees that they are quite stable over time).

²⁷ Other robustness checks have been carried out by the authors, and are available on request. First, we used as dependent variable the hourly wages instead of monthly wages, in order to control for different working time across individuals and to avoid the correction for part-timers used to compute the monthly wages. Results do not change much, consistently also with the figures showed in column (3) of Table 6, in which the working time has been included as covariate. Second, we used a quadratic specification for experience, instead of the categorical specification. Also in this case results do not change. We prefer to rely on the categorical specification, as in other related papers (Autor et al., 2005) in order to control for possible strong nonlinear trends.

²⁸ See, for instance, Acemoglu (2002) for a detailed survey of all this literature.