Productivity spillovers through labor flows: productivity gap, multinational experience and industry relatedness

Abstract

Labor flows are important channels for knowledge spillovers between firms; yet competing arguments provide different explanations for this mechanism. Firstly, productivity differences between the source and recipient firms have been found to drive these spillovers; secondly, previous evidence suggests that labor flows from multinational enterprises provide productivity gains for firms; and thirdly, industry relatedness across firms have been found important, because industry-specific skills have an impact on organizational learning and production. In this paper, we aim to disentangle the effects of productivity gap, multinational experience and industry relatedness in a common framework. Hungarian employee–employer linked panel data from 2003–2011 imply that the incoming labor from more productive firms is associated with increasing future productivity. The impact of multinational spillovers cannot be confirmed, once productivity differences between the firms are taken into account. Furthermore, we find that flows from related industries outperform the effect of flows from same and unrelated industries even if we control for the effects of productivity gap and multinational spillovers.

Appendix 1: Detailed description of data management

We have access to the Hungarian administrative data integration database, which is an anonymized employer-employee linked panel dataset created by the matching of five administrative data sources, for the years 2003–2011, developed by the databank of HAS CERS. The database contains a 50% random sample of the population aged 15–74 living in Hungary in 2003 and the involved employees are traced over the period. The most important demographic features of employees (gender, age, place of residence in the year of entry), and information about their employment spells (months worked, ISCO-88 occupation code, monthly wage) as well as company characteristics (4-digit industry code according to the NACE Rev. 2 classification, number of employees, and specific rows of their balance sheets and financial statements including tangible assets, equity owned by private domestic, private foreign, and state owners, sales, pre-tax profits, material-type costs, personnel expenditures, wage bill) are known. All monetary variables are deflated by yearly industry-level producer price indices to calculate their real 2011 value.

The data is managed by the Databank of the Institute of Economics of the Hungarian Academy of Sciences and can be accessed for scientific research upon individual request. For more details consult http://adatbank.krtk.mta.hu/adatbazisok_allamigazgatasi_adatok.

The raw data contains employee-employer links on a monthly basis. We defined the main employer for every worker and for every year as the workplace where the worker spent the highest number of months in the given year, and created yearly matrices of intercompany movements between these main employers. In particular, if an employee switches firm in the second half of year t or first half of year t + 1, the recipient firm will be her employer in year t + 1 and the source firm will be her employer in year t.

However, our models assess the effect of labor mobility on firms’ productivity on a yearly basis, which can lead to an endogenous connection between labor flows and productivity change (not discussed in the main text). The problem is illustrated in Fig. 4; productivity shocks (e.g. purchasing a machine) happening in the first half of year t + 1 can affect the number of new hires in the first half of year t + 1.

Fig. 4

Periods of productivity change and labor mobility

The potential of reverse causality shortly summarized above might distort our analysis. In order to exclude the possibility of such endogeneity, we conduct the analysis only for those new hires that were observed in year t or in January in year t + 1 at the latest, and exclude all the cases of labor mobility that happened between February and June.

A certain time period has to pass for the new employee to exert a significant effect on firm productivity. With new employees working for a short period and not controlling for months worked at the recipient firm, we would underestimate the effect of new hires on yearly productivity growth. Therefore, in the productivity spillover analysis, only those workers were considered as new hires that stayed for at least 6 months with their new employer.

1.1 Appendix 2: Calculation of human capital

As described in the main text, the human capital of each worker is calculated for each year spent in the private sector. The gaps in private sector employment of at most 3 years are filled by linear interpolation. In case of gaps of at least 4 years, or when the worker only worked in the public sector before getting a job in the private sector, human capital is calculated by a wage regression on the subsample of public sector workers. In addition to the multi-dimensional fixed-effects approach, as a robustness check, we also estimated a pooled OLS regression with age, age-squared, gender and skill-levels of workers. Results are presented in Table 9.

Table 9 Wage equations without and with employee fixed effects, separately on private and public sector employees

Figure 5 shows the distribution of human capital calculated without and with employee fixed effects. Version 1 explains 69% of the variation of the log value of wage, whereas version 2 has an R-squared of 84%. The correlation between the two versions of human capital is 0.74. Since fixed effects can control for more individual-specific characteristics, version 2 is a better approximation of the worker’s true human capital. Its closer-to-normal distribution also makes it more desirable for further analysis, therefore we continue with this measure.

Fig. 5

Density plots of human capital without employee FE (version 1) and with employee FE (version 2)

Figures 6 and 7 show the distributions of human capital with employee fixed effects by gender and skill level. Looking at the curves, we can infer that there is no significant difference between the value of the work-related abilities of men and women, although the variation is higher in the case of women. There is a clear difference between the distributions of human capital by skill level, particularly to the advantage of highly skilled workers. These descriptive findings confirm our decision to use human capital calculated with worker fixed effects.

Fig. 6

Distribution of human capital with employee FE by gender

Fig. 7

Distribution of human capital with employee FE by skill levels

High-skilled: worked at least once in an occupation requiring tertiary education; Mid-skilled: worked at least once in an occupation requiring secondary education; Low-skilled: everybody else.

1.2 Appendix 3: Estimating the probability of hiring

Since the hiring decision of the firm might be correlated with productivity, we should control for this endogeneity. We do so by estimating the probability of hiring and including it in the productivity regression. We estimate logit regressions to predict the probability of hiring for period t + 1 based on firm characteristics already known in period t for the whole population of firms.

1.3 Appendix 4: Calculation of the revealed relatedness distances between industries

To construct the revealed relatedness distances between industries in Hungary, we used the Hungarian Prodcom database matched to the firm-level balance sheet panel data from the National Tax and Customs Administration, available at the Central Statistical Office of Hungary for years 2002–2012. Because the Prodcom database covers industrial activities only, we had to restrict this step of the analysis to the mining, manufacturing and energy sectors (NACE Rev 2: 05-35).

The Hungarian Prodcom database, available at Central Statistical Office (CSO) of Hungary has data on the production of manufactured goods in Hungary for years 1996–2012.

The Prodcom survey requires all qualifying firms to record their production activities at the 8-digit Prodcom (PC8) product level. In the first step of data procession, the primary sector of each company was identified. Each company the (4 digit) industry code was assigned, which corresponded in the Prodcom classification to its most important product in sales volume. Second, a directed link from industry A to industry B was created, if a company had the highest sales volume in sector A, but also produced products, which belong to sector B according to the Prodcom classification. By summing the links over the industries we got the co-occurrence network of industries, where each node corresponds to a (4 digit) industry, and each edge weight corresponds to number of the companies producing in both industries. Note, that we have obtained a directed and weighted network this way for each year.

To calculate the expected number of these co-occurrences between the industries, we used a zero-inflated negative binomial regression, following Neffke et al. (2012). This models the dependent variable by a regime selection process, which determines the probability if the outcome is an excess zero or not, and a count data part, which determines the count outcome in case it is not an excess zero, assuming negative binomial distribution:

$$ E\left( {L_{ij} |v_{i} ,w_{j} ,\varepsilon_{ij} } \right) = \left[ {1 - \varPi_{0} \left( {\gamma + v_{i}^{'} \delta_{i} + w_{i}^{'} \delta_{j} } \right)} \right]e^{{\alpha + v_{i}^{'} \beta_{i} + w_{i}^{'} \beta_{j} + \varepsilon_{ij} }} $$

where $\delta_{i}$ and $\delta_{j}$ represent the characteristics of the source and recipient industries respectively, specifically the number of firms, the total revenue, profits, number of employees, and value added. These variables are available in the firm-level balance sheet panel data from the National Tax and Customs Administration (NTCA), which was available at the CSO and was merged to the Prodcom database using the anonymized company identifier.

From the predicted values of the regression, we were able to calculate our revealed relatedness measure:

$$ \widehat{RR}_{ij} = \frac{{L_{ij}^{obs} }}{{k\widehat{{L_{ij} }}}}, $$

(19)

where ^ indicates fitted value and k is a normalizing constant.

1.4 Appendix 5: The effect of relatedness using the product space measure

The product space concept is based on the idea that different productive factors are behind the production of each product. These include labor, land, capital, technological sophistication and institutions (Hidalgo et al. 2007). However, the products are not only different by the necessary level of these factors, but specializing on a product needs specific skills, infrastructure, or institutions: “making cotton shirts does not require more or less skills than making chocolate, but different skills”. The similarity of these factors will increase the likelihood that these products are produced in similar countries, therefore a pair of products will appear in a country’s export portfolio more often, if they require similar production factors. Consequently, the relative co-appearance of the products in the countries’ export portfolios can be used to proxy the similarity of the necessary production factors.

The product space is calculated using the conditional probabilities of being effective exporters of products:

$$ \varphi_{i,j,t} = min\left\{ {P(RCAx_{i,t} |RCAx_{j,t} )} \right.,\left. {P(RCAx_{j,t} |RCAx_{i,t} )} \right\}, $$

where $RCAx_{i,t}$ =1 if a country is an effective exporter of product i in year t, and zero otherwise.

We deployed UN Comtrade country-level data on international trade of 775 products on the 4-digit level of SITC Rev. 2. codes in year 2003⁵ to calculate the measure of product proximity for all countries except Hungary to ensure the exogeneity of the measure (Boschma et al. 2013). We used the official concordance tables of Eurostat RAMON website to translate the codes to the 4-digit NACE Rev. 1.1. system and aggregate them by taking the mean of proximity for each 4-digit industry pair. Assigning a proximity measure was possible between companies in the agriculture, mining, manufacturing and energy sectors (NACE Rev.1.1. 01-40). We calculated the median proximity of effective labor movements, and categorized each movement to “unrelated” for proximity below the median value, “related” for proximity above the median, and „same industry”.

Our regression equation for measuring the effect of relatedness between the source and recipient industry of the mobile workers is similar to Eq. 12. The only difference is that we do not categorize the product space distances to four categories (zero, below median, above median, same industry), but only to three: below median, above median, and same industry, thus the “below median” category became baseline in the regression.

Results indicate a positive effect of incoming labor from related industries (Table 10). This effect does not change when we control for the productivity gap (Table 10. Column B), however, ownership controls result in a small decrease in the coefficient, and the loss of significance at the 5% level (however, it would be significant at 10%) (Column C). Additionally, we do not see significant interaction effect between the productivity gap and the proximity companies in the product space. Altogether, the results based on proximity in the product space are similar to the ones, we obtained with the technological proximity measures (Table 7), however, a somewhat less consistent, as the same industry parameter is significant and positive in the second specification, while the related industry parameter loses its significance at the 5% level.

Table 10 Logit model to estimate the probability of hiring

We can explain these slight differences compared to the previous results on two bases. The conceptual difference to our baseline measure is that product space uses countries, as levels of observations instead of on plants. Therefore, the factors, which are relevant in similarity is wider, most importantly by the inclusion of institutions, which do not tend to be different between firms of a country, but between states. Therefore, this measure includes more noise conceptually, if we want to use it for measuring similarity in skills. This may influence the parameters of related and unrelated industries, but not the parameter of same industry. The difference in its estimate must go back to the difference in industry categorization: when measuring technological proximity (Neffke and Henning 2008) we used the NACE Rev 2. categories, while the proximities in the product space (Hidalgo et al. 2007) are available for the NACE Rev 1.1. classification. The results may differ due to the fact that in the older classification some industries, which are closely related, but not exactly the same (according to the improved classification) belong to the same categories, therefore the positive coefficient of the same industry parameter is due to activities, which are close to each other, however not similar, which is reflected in the improved classification of economic activities.

1.5 Appendix 6: Full regression results reporting all control variables

In the main text, we refrained from reporting full regression tables with all control variables to keep focusing on the effects which are central in answering our research questions and hypotheses. Still, we consider it vital to present the full results of at least our baseline model.

Table 11 shows the results of our baseline regression on the effect of the productivity gap with all control variables in our analysis. Previously unreported controls are the log of total assets and its lag to control for changes in the amount of capital at the firm, ownership and employment size dummies (baseline categories: number of employees: 2–9, ownership: state) of the recipient firm which are essential determinants of growth potential, share of new hires from unemployment to show how picky firms were when hiring new workers, share of workers leaving the firm to control for outflowing workforce, and fluctuation (share of new hires multiplied by the share of leaving workers) to control for the firm-specific dynamics of hiring and firing (Table 12).

Table 11 The effect of the productivity gap and technological proximity on subsequent productivity

Table 12 The effect of the productivity gap: Full regression output