The labour market effects of offshoring of small and medium-sized firms: micro-level evidence for Germany

ABSTRACT

Small and medium-sized enterprises (SMEs) are the economic powerhouse of many OECD countries (perhaps most prominently so in Germany). Yet, the labour market dynamics caused by the internationalization of their production activities are largely unexplored. We use survey-based micro-level data for Germany to explore the employment effects of offshoring of SMEs, relying on propensity score matching and difference-in-difference analysis. We find evidence for a downsizing effect in the immediate aftermath of offshoring whereas, initially, job creation is not spurred. In the medium run, we find evidence for a slowing down of employment dynamics of offshoring firms (that tend to belong to the better performing SMEs in Germany) relative to non-offshoring firms. Even though our results do not point to a net employment loss in the medium run, our evidence suggests that offshoring may lead to less jobs being created. This conclusion cannot be confirmed for large companies.

KEYWORDS:

• Offshoring
• SME
• employment
• difference-in-difference analysis
• Germany

JEL:

• F23
• F16
• C22

Disclosure statement

No potential conflict of interest was reported by the authors.

Notes

¹ Musteen, Ahsan and Park (2017), Canham and Hamilton (2013) for New Zealand and the case study for 13 Canadian SMEs by Mohiuddin and Su (2013) are exceptions to the rule.

² SMEs play an important role in other countries, too. For example, throughout the recent economic downturn, SMEs in 27 EU member states have maintained their role as the backbone of the European economy, with some 20.7 million firms accounting for more than 98% of all enterprises (Ecorys 2012).

³ Using plant level data for the Irish electronics sector between 1990 and 1995, Görg and Hanley (2004) conclude that only plants that are considerably larger than the mean size gain from offshoring. They measure the benefits of offshoring as enhanced profitability.

⁴ We follow the definition used by, e.g. the European Commission (http://ec.europa.eu/growth/smes/business-friendly-environment/sme-definition_en).

⁵ Unfortunately, our data do not allow us to differentiate between these two cases of offshoring in our empirical analysis.

⁶ We follow the approach recently adopted by Mitze and Kreutzer (2017) in a different context.

⁷ See, e.g. Becker, Ekholm and Muendler (2013), Feenstra and Hanson (1999), Desai et al. (2009) and Munch (2010) for the manufacturing sector, Amiti and Wei (2005, 2009), Crinò (2010) and Hijzen, Inui and Todo (2010) for services offshoring and Ebenstein et al. (2014) for an analysis across sectors.

⁸ Grossman and Rossi-Hansberg (2008) liken this effect to labour-augmenting technological progress.

⁹ Moser, Urban and Weder Di Mauro (2015) also show that the employment effect of offshoring hinges critically on the offshoring motive. If the offshoring decision is part of a restructuring process in which parts of the production facilities are shut down or sold off, the employment effect is likely to be negative.

¹⁰ They illustrate their point by showing that the gap between such a measure and a direct survey-based measure for the manufacturing sector in Canada can amount to 16 percentage points.

¹¹ Using the mean outcome for the group of untreated (non-offshoring) firms as counterfactual would introduce a bias into the estimation because offshoring and non-offshoring firms can be expected to differ even in the absence of treatment.

¹² Caliendo and Kopeinig (2008) provide an in-depth discussion of PSM.

¹³ The variables included in the set of covariates must not be influenced by the treatment and are measured before the treatment.

¹⁴ This does not apply to time-varying unobservables.

¹⁵ The DiD estimator is conditioned on the firm’s propensity score, i.e. their offshoring probability. This variant of the DiD analysis is therefore sometimes referred to as conditional DiD (see, for instance, Heckman, Ichimura, and Todd 1997; Lechner 2010).

¹⁶ The STATA routine diff. is used for the calculations.

¹⁷ The data were accessed with controlled data teleprocessing by the FDZ (Research Data Centre at the Institute for Employment Research).

¹⁸ The panel comprises all establishments with at least one employee liable to social security with a response rate varying between 63% and 84% (Fischer et al. 2009).

¹⁹ It should be noted that the choice of covariates is limited by data availability as discussed above.

²⁰ This idea is based on the Proximity Concentration Model by Brainard (1997). The model posits that if the fixed costs of exports (monetary and non-monetary costs) exceed the costs of FDI, the firm will pursue an offshoring strategy.

²¹ In a study using plant level data for Ireland, however, Görg and Hanley (2005) find that labour demand is significantly reduced by offshoring in the short run.

²² The nearest neighbour algorithm with five neighbours is used to conduct the PSM matching exercise. After the matching, no significant differences in any of the covariates between the treatment and the control group persist. Another indicator for the matching quality is that the pseudo R² improves after matching. In other words, the matching algorithm found statistical twins for the offshoring firms. Other matching algorithms (kernel, calliper, radius) confirm the results. Details of the PSM marching exercise are not shown here due to space constraints but are available upon request. Exemplarily, we present detailed matching results for the medium run analysis in the next section.

²³ We obtain the same result when we test Hypothesis 1 for 2007 and 2010, the other years for which our sample provides unambiguous offshoring data.

²⁴ Again, we use a nearest neighbour matching algorithm with five neighbours and are able to find good matching partners for our treatment groups for both hypotheses The PSM has to be repeated for every hypothesis. Not all firms respond to all questions so that the sample changes for every outcome variable.

²⁵ Details are available upon request.

²⁶ Due to data restrictions – not all firms responded to all questions every year and not all covariates used for our short run analysis were included in every annual survey – the list of covariates used for the panel analysis differs from that of our short run analysis. Due to the mechanics of the matching procedure where close statistical twins are assigned a higher weight the empirical results may seem to differ sometimes from what a quick look at the descriptive analysis (where the whole sample is considered) may suggest.

²⁷ Formally, the standardized percentage bias is calculated as the difference of the sample means of treated and comparison firms as a percentage of the square root of the average of the sample variances in the treated and comparison group (Rosenbaum and Rubin, 1983).

²⁸ As the matching results are similar to those displayed in Table B1, they are not shown here, but available upon request.

²⁹ Note that the period for which employment effects of offshoring may occur (2007–2013) overlapped with the global financial crisis and the European debt crisis. Even though employment was not as adversely affected in Germany as in many other countries, recruitment went down.

³⁰ Again, a valid control group could be found (Table B2). The t-tests for most variables are no longer significant after the matching process. The pseudo R² of the probit regressions drops significantly after re-estimating the propensity score on the matched sample, the χ² test points in the same direction.

³¹ It should be noted that we focus on the effects of offshoring on the offshoring firms only. The analysis of the economy-wide employment effect of offshoring would have to take the repercussions on e.g. domestic competitors and suppliers into account and is left for future research.

³² We again follow the categorization used by e.g. the European Commission (http://ec.europa.eu/growth/smes/business-friendly-environment/sme-definition_de).

³³ Caliendo and Kopeinig (2008) discuss a range of matching estimators suggested in literature such as nearest neighbour, kernel, calliper and radius. The various matching estimators differ essentially in the way they define the acceptable range of matching 1075 partners for the treated firm and in the weight they attach to them.