The effectiveness of investment subsidies: evidence from a regression discontinuity design

According to the State Aid Scoreboard of the European Commission, the total non-crisis state aid in the EU27 member states was €67B in 2012, or 0.52 % of GDP. Source: http://​ec.​europa.​eu/​competition/​state_​aid/​scoreboard/​non_​crisis_​en.​html.

Flanders has five provinces, which exhibit only mild differences in income per capita: The lowest income per capita was still 82 % of the highest income per capita in 2009.

The subsidy as % of total investment was generally around  10 % for most applications, see the data section for more details on the maximum %.

For a more detailed discussion on the conditions that determine whether subsidies increase or decrease employment in a standard production function framework with perfect competition in all markets, see Criscuolo et al. (2012).

For projects that ‘enhance the competitiveness or internationalization of an enterprise in the longterm’.

A small enterprise meets the following conditions: It has less than 50 employees, its yearly revenue is lower than €7M (€10M from 2005 onwards), and its balance total is less than €5M (€10M from 2005 onwards), and is independent (<25 % ownership by large companies). A medium-sized enterprise meets the following conditions: It has less than 250 employees, its yearly revenue is lower than €40M (€50M from 2005 onwards), and its balance total is less than €27M (€43M from 2005 onwards), and is independent (<25 % ownership by large companies).

These policy goals are listed, e.g., in the policy note of the Flemish minister of Economy at that time Patricia Ceysens, see http://​docs.​vlaanderen.​be/​portaal/​beleidsbrieven20​07-2008/​ceysens/​beleidsbrief_​ceysens.

More background on this call system can be found in the document Ooghe and Spaenjers (2005).

Banks offer to do the application process for the firm at a cost of €350 per application, with an extra fee of 7.5 % of the subsidy if successful (minimum €500 and maximum €3500). Source: online document of the firm ‘Pylser boekhouding & fiscaliteit BVBA.’

Since the call of June 3, 2005 (decision date), the criterion ‘requested subsidy as a % of investment versus max allowed limit’ was dropped. In addition, since the call of September 15, 2006, the criterion ‘ratio gross wage bill and value added’ was dropped.

See online Appendix A on the subsidy for details on the scaling.

In Dutch: Agentschap Ondernemen.

In Dutch: Groeipremie.

‘Best application’ is defined as closest to the acceptance cutoff. This drops 4701 applications or 29 % of the original 16018 applications that remain after basic data cleaning (dropping observations with missing vat number or missing/erroneous entries for the subsidy amount or total investment amount).

This additionally drops 1689 applications, or 11 % of the original 16018 applications.

This additionally drops 479 applications, or 3 % of the original 16018 applications.

The main results are robust to only keeping the first application of the firms that never received the subsidy instead of keeping the best application, or dropping all firms that received the subsidy and ever got rejected instead of only dropping the unsuccessful applications of these firms.

The accounting system allows for four categories of fixed assets: start-up costs, intangible fixed assets, tangible fixed assets and financial fixed assets. The subsidy can only impact intangible and tangible assets. Depreciation and amortization, used to calculate gross investment, refer to fixed assets without the financial fixed assets. Thus, we define ‘fixed assets’ as fixed assets minus financial fixed assets, as the latter are not included in the subsidy nor impacted by depreciation. Thus, when we use the term (total) fixed assets in the remainder of the text, we actually refer to fixed assets minus financial fixed assets.

These differences, except for value added, turn out to be insignificant when regressing the indicator on a subsidy dummy together with call dummies.

We define a total of 12 aggregated sectors based on the broad structure of the NACE rev2 classification. We do not pursue a narrower sector definition because of the limited number of observations in the data.

A higher-order polynomial is used to take into account possible nonlinearities. Increasing the order does not have an impact on the results, but we prefer to limit the number of terms to avoid over-fitting.

We reduced the polynomial to a second-order polynomial to avoid over-fitting. Note that even then, the number of polynomial parameters to be estimated increases by one compared to the ‘normal’ third-order polynomial, as we estimate a different second-order polynomial on both sides of the treatment.

We limit the sample to firms that have a score that is 0.29 or less from the cutoff score. The number 0.29, which is 1/3 of the SD of the score, balances two objectives: The lower we take the distance to the cutoff score, the more accurate all things equal, but it also implies that we have less observations and therefore lose estimation precision. We report as well results for firms at different distances from the cutoff as a robustness check.

Given the limited number of observations, we report in our base specification only the results with a linear term. Remember that the polynomial should be allowed to differ across the 16 different calls and across treatment groups. For example, a specification with a first-order polynomial requires an estimation of the parameter of interest, 12 sector dummies, 64 polynomial parameters (16 calls × (1 linear term + 1 call dummy) parameters for the polynomials × 2 treatment groups). These are already 77 parameters with approximately 1000 observations. Given that we only keep observations close to the cutoff, a linear control should suffice. We report as well specifications with a higher-order polynomial. However, the Akaike information criterion indicates even that the specification without a control for the score is preferred for observations close to the cutoff.

In our analysis, we use employment input converted to full-time equivalents, which expresses the total labor input in terms of the number of full-time workers that would be needed to work the same amount of hours.

The selection score for the age criterion was age in days, with the score decreasing with age. After 5 years, there was a limit; from this age onward, the score was equal to the score value of an age of 5 years. In the tables on the pre-subsidy differences regressions, we therefore set the age to 5 years from 5 years onward.

The number of observations varies across the different indicators due to data availability and because some indicators were only used for a limited number of calls.

There are no results for the sustainability indicator as there are not enough firms with a sustainability certificate close to the threshold. (<1 % of the firms has a certificate).

Note moreover that at the moment of application, the cutoff score was not known to the firms as it depends on the scores of all firms as well as their requested subsidy amount. The large number of applicants at each call makes collusion as well highly implausible.

Results are displayed for observations within 1 SD of the cutoff. The bin size is equal to 0.05. Similar results are obtained with different values for the bin size. To test for the width of the bin, we follow Lee and Lemieux (2010) and use the idea that if the bins are narrow enough, there should be no systematic relationship between the indicator and the score within each bin. We test for this by adding a set of interactions between the bin dummies and the score to a regression of the indicator on the bin dummies and testing whether the interactions are jointly significant. We do not find this to be the case.

We found similar results when doing a split sample and focusing only on small firms, instead of working with an interaction term.

Given the winsorization, the firms for which the point estimate is negative are in fact as well out of sample.

In the last specification, the estimated effect becomes insignificant, but given the robustness if the estimate over the other specifications, we continue with the assessment that there is a positive effect on employment.

We checked as well whether there is an impact on the exit probability of firms, but failed to find a significant effect.

We checked whether the results changed when including other covariates such as initial value added or initial sales, but found this not to be the case. We have also performed regressions comparing the change in the average value of the indicator before and after the subsidy instead of 4-year cumulative growth and found similar results.

We first standardize the score before including it in the local linear regression. We used the rd module in Stata to execute this analysis (Nichols 2011).

We also included the amount of requested subsidies next to the received subsidies in the regression equation. The idea is that the requested subsidies could control for the endogeneity issue by picking up the relation between the requested subsidies and the planned size of the investment, thereby affecting firm growth. The results of including this extra control are highly similar to the one reported in the paper.

The statistical significance will likely drop due to two reasons. First, estimating a continuous effect is more data-demanding than estimating the coefficient of the simple dummy \(D^{\text{subs}}\) in the baseline specification. Second, the applications for firms that receive the subsidy multiple times are not taken into account, possibly underestimating the subsidy amount received and thus introducing measurement error.

For example, in terms of labor productivity, firms with one employee are approximately 60 % more productive than firms with more than 3 employees. For firms with 2 and 3 employees, these premia are, respectively, approximately 28 and 13 %.

The owner is at least partly paid through the profits of the firm. Some owners choose to be an employee of their firm, e.g., for tax reasons (if they pay themselves a low wage, the tax rate on their wage is lower than the corporate taxes), but not all owners do this. We see, for instance, many firms reporting zero employment in their accounting statements.

More specifically, we add separate dummies for a FTE of respectively one, two and three. We also add log FTE as a general control for the number of FTEs. The details on the exact implementation are explained in online Appendix G.

The results for this robustness check are available on request.

Calculated as \((0.178-0.0594 \cdot \log (10)) \cdot 10.\)

Theoretically it is possible that a capital subsidy has a negative impact on employment growth. However, given the previous results, it seems more likely that the larger firms simply did not invest more due to the subsidy, and hence the effect on employment was also zero.

Cerqua and Pellegrini (2014) estimate the same measure for Law 488 in Italy to be about €62K/job. Criscuolo et al. (2012) estimate a different measure for the RSA program in the UK, i.e., total jobs created (using estimates at the area level rather than the firm level) over the total cost of the program (including administrative costs and taxation). Their estimate is about €6K/job.