of the stochastic component of value added data.24
6.2 Workers’ earnings
For workers’ earnings we consider a logarithmic specihcation of the process (7), in which the
Hrst difference of log annual net real earnings is regressed on a set of observable attributes:
a fourth-order polynomial in age, education (here proxied by a set of occupation dummies),
gender, area of residence dummies, and time dummies (the vector aijt). As noted above,
nominal gross earnings are Hrst transformed into nominal earnings net of taxes and social
security contributions (using the rules coded in the Italian tax system at each point in time),
and then deflated by the CPI to 1991 prices. We use the available data for all workers rather
than just those in the matched sub-sample. Our estimated regression is:
A(T,p)Δln wij t = Δa,ijtλ + Δωijt (24)
Recall from Section 3 that under our hypothesis, the length of the AR process for Hrm
performance carries over to the length of the AR process for workers’ earnings. Moreover,
under the same hypothesis, Δωijt is no longer an MA(1) process but an MA(2) process if
p = 1, as we have assessed in the previous section. We thus impose p = 1 and estimate the
speciHcation (24) by IV using the orthogonality condition:
E (Δωij t∣Ωt-3) = 0 (25)
The results from estimating equation (24) are reported in Table 4. The AR(1) coefficient
takes on a value of 0.33, with a standard error of 0.03. The reduced form regression (not
reported here) shows that the instruments have sufficient predictive power: the p-value of
the F test is in fact below 0.1 percent.
24Covariances tend to decay rapidly even when estimated on a year-by-year basis. This exercise, however,
reveals that a distinctive feature of the data is covariance non-stationarity, in particular around the strong
recessionary episode of 1993. This recession was particularly anomalous because it was characterized by a
sharp devaluation and a major tax increase. The former was advantageous only for exporters, while the latter
bore on all Hrms. Before 1992, however, stationarity would not be an extremely unlikely characterization of
the data. The full matrix of estimated autocovariances is available on request.
21