PROCESO DE EXPORTACIÓN

2. The covariance structure of male wages

through subsequent years of the panel is reported above the diagonal, where the number of observations used in the computation of each cell of the covariance matrix for the whole sample is collected: as can be seen, this path is, with few exceptions, decreasing.

The evolution of some sample features over the period is reported in Table 2.2. As can be seen, apart from a tendency to white collar undersampling in 1976 and 1977, the occupational structure tends to stay constant from 1974 to 1988, while the cohort composition shows how the weight of younger cohorts grows through years. The last two rows of the table report the evolution of the cross-sectional distribution of logarithmic wages over the period: mean logarithmic wages recorded a 5% growth over the period (23% in levels) while wage dispersion dropped of almost 20% from 1974 to 1982 and grew thereafter, the value in 1988 being approximately equal to that in the starting year. We turn now to the longitudinal analysis of such inequality dynamics.

2.4 The covariance structure of Italian male wages

This section contains the results of the empirical analysis of Italian male wages. The analysis has been carried out for the whole sample (blue collars, white collars and managers) and for blue and white collar workers separately; this last exercise seems particularly relevant, given that, as seen in the Section 1.3, the use of flexible pay determination schemes during the 1980s developed mainly for non-manual workers.

In order to construct the wages covariance matrix, real wages25 have first been adjusted for year, age and cohort effects; the three effects are intended to capture

Nominal figures have been deflated with the CPI (1985=100).

2. The covariance structure o f male wages

business cycle, life-cycle and productivity growth effects respectively and to identify the last two separately, cross-sections have been pooled over years. Age effects are controlled for by a quadratic function of age, while year and cohort effects are specified as dummy variables. Figure 2.1 plots kernel density estimates of the cross- sectional distribution of residuals from this regression for the whole sample: a noticeable feature is the concentration of frequencies around the mean taking place from 1977 to 1984, which results in an increase of the corresponding density height; this tendency reverts thereafter, especially during the last two years of the sample, when the peak’s height decreases.

2.4.1 The empirical covariance structure

For each of the 3 groups considered (i.e. whole sample, blue collars and white collars26) residuals from such initial regressions have then been utilised to construct the respective covariance matrices; Table 2.3 reports the one for the whole sample together with each element's standard error, while the corresponding correlation coefficients are reported below the diagonal.

As can be noted in the table, the covariance structure fails to asymptote to a long run level, a feature which has been observed both in UK and US data (see Dickens [1996] and Moffitt and Gottschalk [1993] respectively): a marked historic­ time dependence seems to be present in the data, with covariances peaking in 1985 and especially in 1988.

Figures 2.2 to 2.4 provide some further element of discussion by reporting the evolution of the variance (i.e. the diagonal of the covariance matrix) and of some

26 The blue collar/white collar split has been carried out by considering the worker's occupational status of each year. This means that if a worker is classified as blue collar in year t and as white collar In year s (i.e. if a change in occupation takes place between t and s) he will not contribute to E(u„ u„) for either

2. The covariance structure of male wages

wage persistence and immobility indicators, namely the correlation coefficient and the ventile quasi-immobility ratio.

The first panel of Figure 2.2 displays the well-known picture of trends in Italian wage inequality (see, for example, Dell’Aringa and Lucifora [1994]), with a strong compression of differentials during the 1977-82 period, in which the egalitarian system of indexation was fully effective, and a reopening thereafter, particularly marked in 1988. The remaining two panels of Figure 2.2 help in assessing the evolution of inequality within each occupational group; the reopening of wage differentials affects white collars data since 1984, while the decrease of blue collars' wages variance continues, although at a diminished pace, until 1987. If compared with the evidence in the top left corner of the figure, this suggests that the widening of overall wage differentials was, to a large extent, driven by inequality within white collar workers, for which it seems that the relaxation of equalising forces had quicker effects.

Patterns of wage persistence are described in Figure 2.3 by means of the correlation coefficient, which is plotted for wages one and five years apart. The graph in the top left corner shows a short term measure which tends to cycle before 1982 and to slightly decrease in the years of growing dispersion, with the exception of the last year of the sample; the picture is different for the medium term measure, for which a marked drop can be observed from 1985 onwards. Taking manual wages into account, it can be seen how correlation tends to be concentrated in the central- final part of the sample period, depending upon the lag width considered; both the decrease in short term correlation since the early 1980s and the drastic drop in medium term correlation from 1985 onwards are still evident. The data for white collar workers data tell a somewhat different story, with correlation which increases both in the short and the medium term during the last years of the data, with a drop in 1988.

2. The covariance structure of male wages

Figure 2.4 plots the values of the ventile quasi-immobility ratio for wages one and five years apart; compared with an usual immobility ratio (i.e. the average, across classes, proportion of cases not changing wage class during the transition considered) this measure considers immobile also those individuals moving only to the adjacent class and is thus robust to the effect of small wage “pushes".27 Compared to the correlation coefficient, which is based on co-deviations from the marginal mean and picks-up absolute changes at each point of the wage range, quantile mobility indices measure variations in relative ranks and ignore within classes movements (see also Jarvis and Jenkins [1998]).

The graph shows, for the whole sample and blue collar workers, a tendency for the frequency of transitions among ventiles to initially drop and then level-off towards the central years of the data, especially in the short term. The pattern is slightly different for the white collars' sample, where immobility steps up in correspondence of the increase in autocorrelation observed above.

2.4.2 Variance components models for the covariance structure

Table 2.428 presents results for the EWMD estimator applied to the whole sample of male wages. In this and the subsequent tables the fourth moment matrix has been utilised to correct asymptotic standard errors (reported in parentheses) for the presence of both heteroskedasticity and serial correlation in second moments (see the Appendix). Two goodness of fit measures are reported: the sum of squared

The measure ranges from .145 in the case of “perfect mobility' (stochastic independence of earnings in the years delimiting the transition considered), to 1 in the case of ‘ complete immobility' (no changes in ventile ranks).

2. The covariance structure of male wages

residuals and the sum of squared residuals weighted by the inverse of the fourth moment matrix29 30

As a starting point for the analysis, columns 1 to 7 present results obtained without allowing for time varying loading factors within each component. A first thing to note in the table is that the permanent wage component explains about 75% of residual variance in the basic model (model 1, with constant permanent and white noise transitory component), a split similar to that obtained in previous studies (see Dickens [1996]). By letting the transitory component admit some form of serial correlation (models 2, 3 and 4, where the transitory component is AR(1), MA(1) and ARMA(1,1), respectively, while the permanent wage is held constant), we can see how the additional parameter capturing the dispersion of transitory wages in the first year of the panel required by the AR part impacts on the estimate of permanent variance, which tends to be lower in such cases; moreover, in the case of the AR(1) transitory component (model 2) this also affects the estimate of the base transitory

variance (o^ ).

When the permanent wage component is allowed to be a linear trend (models 5 to 7), intercepts and slopes of such individual trends negatively covary: the covariance parameter ( tr>iy) thus captures the compression of differentials which has been observed in figure 2 and indicates the presence of forces equalising wage

levels within the permanent wage component. 29 *

29 Moffitt and Gottschalk [1993] adopt the unweighted sum of square residuals, while Dickens [1996]