Regional wage and productivity in Italy: a spatio-temporal analysis

ABSTRACT

This paper analyses the empirical functioning of the regional labour market in Italy (21 Italian regions over the period 1995–2015). Starting from centralized union bargaining, it derives a micro-founded theoretical long-run relationship between the nominal wage per worker and average labour productivity. The long-run equation is then empirically verified using cointegration analysis, controlling for time and space dependence. It is found that labour productivity has a positive impact on determining the nominal wage per worker. The magnitude of the impact is, on average, 0.28 when considering both direct and indirect (spillover) effects. According to the results, regional policies aimed at reducing local disparities must foster labour productivity, particularly in the south of the country.

KEYWORDS:

• regional labour market
• union behaviour
• panel error-correction model
• spatial analysis

JEL CLASSIFICATIONS:

• C5
• C61
• D7
• J5

ACKNOWLEDGEMENTS

A first version of the paper was presented to the 58th European Association of Regional Sciences in Cork. The present paper benefits from the comments of the participants. The authors thank Giovanni Millo for providing valuable help with the coding R. The usual disclaimers apply.

DISCLOSURE STATEMENT

No potential conflict of interest was reported by the authors.

Notes

1 Database on Institutional Characteristics of Trade Unions, Wage Setting, State Intervention and Social Pacts, ver. 6.1, October 2019.

2 Adjusted bargaining (or union) coverage rate: (0-100) = employees covered by valid collective (wage) bargaining agreements as a proportion of all employed wage and salary earners with the right to bargaining, expressed as a percentage and adjusted for the possibility that some sectors or occupations are excluded from the right to bargain (ICTWSS v.6.1 codebook, p. 15).

3 North: Piemonte, Valle d’Aosta, Liguria, Lombardia, Trentino-Alto Adige (Trento and Bolzano), Veneto, Friuli-Venezia Giulia, Emilia-Romagna; centre: Toscana, Umbria, Marche, Lazio; and south and islands: Abruzzo, Molise, Campania, Puglia, Basilicata, Calabria, Sicilia and Sardegna.

4 Equation (2) uses the nominal wage as a dependent variable coherently with the theoretical model. Nevertheless, in the empirical analysis of the following sections, a referee asked what would have happened had we used the real wage. By using the CPI for worker and employee households, the real wage does not affect the results qualitatively as nominal and real variables share a similar temporal and spatial pattern.

5 All econometric results are performed by R. The authors thank Giovanni Millo for valuable help in coding. The usual disclaimer applies.

6 Table 4 shows average correlation coefficients, whose values are the simple average of the pair-wise cross-section correlation coefficients of the ADF(p) regression residuals:$\hat{\overline{\rho }}=[2/N(N-1){\sum }_{i=1}^{N-1}{\sum }_{j=i+1}^N{\hat{\rho }}_{ij}$where ${\hat{\rho }}_{ij}$ is the correlation coefficient of the ADF(p) regression residuals between i-th and j-th cross.

7 Table 5 shows CD test statistics:$CD=\sqrt{{\displaystyle \frac{2T}{(N-1)}}}\left({\sum }_{i=1}^{N-1}{\sum }_{j=i+1}^N{\hat{\rho }}_{ij}\right)$which tends to $N(0,1)$ under the null hypothesis of no error cross-sectional dependence (Holly et al., 2010).

8 $se$ is the standard error of the estimates $\hat{\beta }$. It is computed as $se=\sigma /\sqrt{n}$, where $\sigma$ is the population standard deviation; and $n$ is the number of observations. se is used to build a confidence interval (CI) for the true parameter $\beta$ around its point estimate $\hat{\beta }$. By choosing a given level of statistical significance $\alpha =5\text{\%},$ the CI is given by $\hat{\beta }\pm 1.96se.$ When the population standard deviation is unknown, then a Student’s t distribution is used. The CI can be used for hypothesis-testing on the true parameter $\beta .$ A practical rule of thumb can be followed: if the point estimate is larger (in absolute value) than 2 se, the null hypothesis $\beta =0$ is rejected (this means that the estimate’s p < 0.05).

9 Table 7 shows the results for the nominal wage; we replicated it in the case of real wage and labour productivity deflated with the CPI. The results are quite similar. In real terms, the coefficient of labour productivity is 0.28, slightly higher than the nominal wage.

10 Generally, a bivariate cointegration analysis is based on Engle and Granger (1987). In the bivariate case, there is at most one cointegrating equation. The thoretical model identifies it in equation (2), which implicitly sets the coefficient of wage per capita to 1.

11 Estimation was carried out by STATA with the package XTDCCE2 by Jan Ditzen (2018). The basic model is augmented with 63 mean group variables (21 lags for three mean variables $LongRu{n}_{t-1}$, $\mathrm{\Delta }RW{2}_{t-1}$ and $\mathrm{\Delta }pro{d}_{t-1}$) and by two lags of the cross-sectional averages for both dependent and independent variables.

12 Following a suggestion by a referee, we tried two alternative weight matrixes, first, by exploiting ship commercial routes that connect islands; and, second, the commercial routes by wheels connecting regions. The results do not change remarkably. Data are available from the authors upon request.

13 As suggested by a referee, we tested the robustness of estimates to alternative specification of $S.$ We tried the road commercial routes provided by Istat; this makes the connectivity mess more dense, but without qualitative differences in results.

14 We also tried to add a separate trend, but it does not change the results.

15 In general, the matrix is a $N\ast TxN\ast T$, with $W(S)$ on its diagonal.