2.8.1
Data and Explanatory Variables
I study the influence of international outsourcing, measured by foreign capital accumulation, on the relative demand for skills in Poland in the manufactur- ing industry in Poland. The sample consists of an unbalanced panel18 of 23 NACE (2-digit) industries over a 9 years’ period (1994-2002).19
The dependent variable, relative demand for high skilled workers, is prox- ied by the high skilled labor wage share, measured as the non-production workers’ wage share in the total wage bill. One component of the dependent variable, the employment of high skilled (low skilled) workers, is measured by annual average employment of non-production (production) workers. The second component, the wage of high skilled (low skilled) workers is measured by an annual average gross wage of non-production (production) workers. Unfortunately, especially at the level of disaggregation necessary for econo- metric analysis, no better proxies for high-skilled and low-skilled labor are available.
The data set obtained from Polish Statistical Office (PSO)20 allows for direct measurement of foreign capital and also enables separation of foreign and domestic owned fixed assets.21 However, I cannot distinguish foreign
affiliates that explicitly engage in outsourcing, or more precisely offshoring22, from other foreign subsidiaries. Therefore, I treat all foreign firms as off- shoring firms, even if the latter actually constitute only a subset in my data
18Some numbers are not made public for confidentiality reasons.
19As mentioned in Section 2.2 Bruno, Crino and Falzoni (2004) examine a similar ques-
tion for Poland, the Czech Republic and Hungary. However, they have data on 6 ISIC industries and in the case of Poland they cover the period 1994 to 2001.
20Data has been partly collected from various publications of PSO and partly obtained
from PSO in electronic form. For details see Appendix.
21Feenstra and Hanson (1997) for lack of data could not directly measure the capital
stock in foreign ownership and thus used the number of foreign firms as a proxy. Bruno, Crino and Falzoni (2004) measure foreign capital with foreign direct investment stock.
22As previously noted, offshoring can be defined as international outsourcing of activities
set. (1 +KiF DI KD
i
) andKD representing the ratio of foreign to domestic capital and domestic capital, respectively, compose my basic specification.
In addition to the basic variables I include several control variables. Feen- stra and Hanson (2001) argue that one should include any structural variables that capture other factors that might influence the production costs. In or- der to account for the restructuring processes in Polish manufacturing I use a measure of privatization (the share of private firms in the total number of firms). I assume that private enterprises have stronger incentives to ratio- nalize and modernize their production than their public counterparts so that their activities might have affected the relative high skilled labor demand.
Furthermore, it is necessary to include variables that, following theoretical and empirical literature, could also have an impact on relative demand for high skilled labor. For this purpose I include the share ofR&Dexpenditures in sales in order to account for technological improvement, and import and export penetration ratios to control for potential influence of international integration and of exposure to international competition.
It is common practice to include output in this type of the regression, as the variable cost function condition on total output. However, due to high correlation between output (measured by sales) and domestic fixed assets, which enter the regression in levels, I excluded the output variable from regression. Thus my modified estimating equation is:
W BSitHS = β1+β4ln( witHS wLS it ) +β2ln(1 + KitF DI KD it ) +β3lnKitD +β5lnP RIV F IRM/F IRMit+β6lnR&D/Yit
+β7lnIM P/Yit+β8lnEXP/Yit+it (2.6)
where P RIV F IRM/F IRM denotes the share of private firms in the total
number of firms, R&D/Y is defined as the R&D expenditures over sales, IM P/Y represents import share in sales and EXP/Y - export share in sales.
2.8.2
Estimation Strategy
The above regression will be estimated with fixed effects, since any variation between units not accounted for by the independent variables creates unob- served heterogeneity in the model. Given that industries differ from each other in terms of size or skilled labor and capital intensities, estimating with OLS would relegate the unobserved heterogeneity to the error term and the coefficients would be biased.23
Furthermore, I also incorporate time fixed effects. There are two impor- tant reasons for doing so. First, I have neglected the fact, that foreign capital might be determined by some foreign factors. Due to obvious reasons I can- not include these variables in the regression. By inclusion of time dummies, I assume that the impact of foreign variables is the same across industries and varies only over time. Second, one should not forget that Poland is a tran- sition economy with institutions and the economic system as a whole being still “work in progress”. Hence, there might exist some aggregate exogenous factors that are correlated with the industry-level relative labor demand. Accounting for industry and time fixed effects helps also to resolve potential problems arising from omitting output in the regression.
Not surprisingly, statistical tests show that there is a heteroscedasticity problem plaguing our data. In order to assure the efficiency of diagnostic tests all standard errors reported in the results are robust to heteroscedasticity. Finally, the relative wages of high skilled workers are likely to be endogenous in the wage share regression, and failure to control for this may lead to simultaneity bias. I am avoiding this problem by excluding the relative wages variable while estimating with OLS. This in turn may cause omitted variable bias. It is therefore necessary to verify the robustness of the OLS estimates by instrumental variables method.
23The big advantage of the fixed effects versus random effects is that any potential
correlation of the explanatory variables with the individual effects is rendered harmless since the fixed effects and therefore their correlation with the explanatory variables are annihilated. Additionally, the Hausman test rejects the null hypothesis that the estimates from the two models are the same, that is, the random effects estimator is not a viable solution and fixed effects should be more efficient (Beck and Katz (1995)).