Can going cashless in the EU have a positive effect on economic growth and if so, through which channel?

(1)

Can going cashless in the EU have a positive effect on

economic growth and if so, through which channel?

15/07/2018

MSc Economics: Public Policy (track)

Author: Matias I. Grinberg

Student number: 11603399

Email: matiasivangrinberg@gmail.com

Supervisor: Naomi Leefmans

Second Reader: Dr. John Lorié

Word count: 14985

(2)

2 STATEMENT OF ORIGINALITY

This document is written by Matias I. Grinberg who declares to take full responsibility for the contents of this document. I declare that the text and the work presented in this document is original and that no sources other than those mentioned in the text and its references have been used in creating it. The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.

(3)

3

Nowadays, it seems that societies are abandoning physical ways of payment settlement for non-physical methods like cards or wire transfers which like any other technological change and trend in habits, poses opportunities and adversities. In the Appendix A, figures 1 to 9 (constructed with data taken from the ECB Data Warehouse at the EU aggregate level) show the trends in the relevance of different payments methods. Figures 1 to 4 have quantity of transactions (standardized to the size of the population) and figures 5 to 9 show the trend in value of transactions (standardized to GDP). These figures show that the reliance on physical methods like cheques or withdrawals of cash has been declining while direct debits and card use has been increasing. In Sweden, the Central Bank is even analyzing the possibility of issuing their own e-currency to promote a safe and efficient payment system (Riksbank, 2017). Even though this change is not a matter of politicians’ decisions but of individuals’ choice, government policies towards it will matter. My research then will try to answer the question “Can going cashless in the EU have a positive effect on economic growth and if so, through which channel?” to therefore be able to provide a reason why governments (if at all) should foster a cashless society. The topic is rather scarce among the previous literature even though there are papers published that aim to measure the direct effects of different payment methods on economic growth (ECB, 2013). Nevertheless, my method is different in the sense that I provide quantitative estimates of how the payment instrument can affect economic growth via indirect channels. In an economy where no transactions are conducted using cash, all transactions become completely traceable which could allow governments to amplify the base of contributors bringing down the average tax rate, or could also keep it constant but improve the public finances (E&Y, 2016). For the same reason (traceability), the transition to cashless might also have a positive effect on transparency (or equivalently a negative effect on corruption) and lastly, if electronic payment methods can increase the pool of merchants for consumers and viceversa, they might foster consumption. The following section after this introduction (section 2), develops the literature review to prove how true is this hypothesis that the avoidance of cash increases consumption while diminishes tax evasion and corruption (see dashed green square in figure 1 below) and how these changes in consumption, transparency/corruption, and tax burden or primary budget/fiscal stance (as a consequence of reducing tax evasion) may impact economic growth (see yellow boxes connected to GDP/capita growth in figure 1 below). Section 3 explains the Methodology of how regressions, that always have a proxy of economic growth as the dependent variable, will be set up to find the coefficients of interest (consumption, tax burden, primary budget and an index of corruption). It should be clear by now, that the effects of going cashless on the boxes inside the dashed square in figure 1 are taken as given from the literature review, but cross-country regressions will be performed to see how these (yellow boxes connected to GDP/capita growth in figure 1 below) translate into economic growth. Section 4 explains the analysis of the data I will be working with to reach some preliminary conclusions, section 5 contains the results from running the regressions along with the robustness tests and the answer to the research question (how can going cashless could positively affect economic growth), to finally make the ultimate conclusions, state the limitations of the research and suggest topics for further research in section 6.

(5)

5 Figure 1. Channels through which Cashless affects Economic Growth

2. Literature Review

The objective of this section is to research theoretically and empirically, firstly (from section 2.1 to 2.3) what are the effects of the abandonment of cash on each of the channels mentioned in the “Introduction” (dashed square in figure 1 above) and secondly (from section 2.4 to 2.7), how these may influence economic growth (yellow boxes in figure 1 above). In section 2.8, I justify the control variables I will be using (those used in cross-country economic growth regressions in the literature covered from section 2.4 to 2.7).

2.1 Effects of Going Cashless on Transparency

The two studies covered in this section aim to discover if there is any significant effect of the choice of payment method on the level of corruption (or interpreted alternatively on transparency). Singh et al. (2017) studied the effects aggregate currency in circulation and specifically large denominated banknotes (standardized to M1) has on corruption, the latter measured with the World Bank’s Control of Corruption Index which gives higher values for those less corrupt. Obviously corruption cannot be measured directly but the index published aims to identify the level of perception. Firstly, they make use of the pooled OLS method and, when adding all their control variables, they found a coefficient of -0.72 on aggregate circulation and -0.81 on large denominated banknotes meaning that increases in cash in circulation translates into a lower index (higher corruption). The adjusted R2_{increased from 0.80 to 0.84 when the regressor was the amount of large banknotes instead of} the aggregate currency in circulation signaling that these exert a stronger influence. To solve endogeneity issues, they firstly used random effects and then the GMM method. In any case, the signs of the coefficients and the difference between the 2 proxies of cash remained the same although decreased in magnitude. There seems to be an inertia effect in the corruption index evidenced by including the lagged value of the index in the regression and finding it statistically significant even at the 1% critical value, probably because changes in the perception of corruption require transformations in institutions and social values which are difficult and time consuming. The final point estimates obtained with the GMM method state that a unit increase in aggregate currency in circulation decreases the corruption index from 0.07 to 0.11 units while a unit increase in the large banknotes ratio has an effect that goes from 0.08 to 0.14. In every model the panel Granger causality test was used to test the direction of the relationship between independent and

(6)

6 dependent variables and their results point out that the aggregate currency in circulation is the one causing corruption but there is a bidirectional relationship between large denominated banknotes and corruption. In the context of the EU it would be that not only 500 euro notes facilitate corruption but also that more corrupt governments favor large note circulation. As the authors suggested by the end, central banks should limit the supply of banknotes of large denomination if they are committed to decreasing corruption.

The Bank of Finland (2011) did an almost similar study measuring how the shares (over the total value of transactions) of transaction instruments used impacted the transparency index provided also by the World Bank and by Transparency.org on an annual basis and by using 3-year moving averages of each of them. Using panel estimation techniques for 12 advanced economies and a “common” set of control variables found in the related literature like economic prosperity, an index of democracy and government size, they find that the choice of transaction instrument matters. These results are especially significant given that the countries in the sample are among the less corrupt and they hold for alternative corruption measures and when allowance is made for endogeneity of payment instruments. Overall, credit cards are consistently associated with lower corruption while cheques and paper credit transfers appear to have a positive link with it, the results for non-paper credit transfers are mixed and direct debits fail to show significant effects. In their final policy recommendation they advise to encourage the entry of foreign banks with sophisticated payment services and minimizing the fees for cards use.

Overall, the literature finds that going cashless increases transparency (reduces corruption).

2.2 Effects of Going Cashless on Consumption

The aim of this section is to describe studies that analyze the effect card use has on consumption patterns. Previous studies measured consumers’ behavior by comparing exclusively credit cards vs cash use and concluded that the first one reinforces the positive feelings of purchases by delaying the pain of paying while cash reinforces cost considerations due to its immediate associated pain. The implication for consumers, with the ongoing digitization of payments, is that they lose some control over their spending. Runnemark et al. (2015) contributed to the literature by conducting a randomized controlled experiment to test if debit cards, which are a perfect substitute for cash that only differs in terms of the format, induces more spending as well. The sample of individuals in the experiment were offered to bid for 3 different products: beer, expensive coffee and regular coffee. Those assigned to the control group were only allowed to pay with cash while those assigned to the treatment were offered the possibility to pay with debit card. The dummy that took a value of 1 for those under the treatment showed that this group’s average bids were 36%, 22% and 52% (percentage) higher for each product respectively. Whenever the “%” sign appears throughout the text means percentage points except when explicitly stated as a percentage change.

Moody’s Analytics (2016) used panel data techniques for a sample of 70 diverse countries over a 5-year period between 2011 and 2015 to determine that 0.01% out of 2.3% of the average annual growth in real consumption experienced in the whole period, is attributable to increased card penetration (percentage of purchases done with cards over the total expenditure) which ultimately helped to sustain GDP growth. In their regressions the dependent variable is the growth rate in consumption and among other variables like income, card penetration is the main explanatory

(7)

7 variable. The reasoning why cards can foster consumption are that they provide consumers with all their available funds or lines of credit whereas with cash or cheques they may be limited. At the same time they give merchants peace of mind as a third-party is backing up the purchase, they can also help smoothing consumption like Friedman’s permanent income hypothesis predicts when faced with shocks in income (Romer), and via e-commerce they provide access to a greater variety of goods at better prices due to increased competition. To successfully spur consumption (and GDP) growth, they warn that a well-developed financial system is a key factor: as consumers feel more comfortable using cards, merchants are more willing to accept them to access the pool of customer suggesting a multiplier effect after a certain threshold. In summary, there exists a behavioral bias: card use risks over-spending (increases consumption).

2.3 Effects of Going Cashless on Tax Evasion

The method of payment is also found to have an impact on tax revenues. Immordino et al. (2017) provide empirical evidence that the payment instrument used for purchases matters for the level of tax evasion. Merging ECB payment statistics with estimates of tax evasion, they construct a regression that has as the dependent variable a proxy of tax evasion explained by the use of payment methods and a vector of control variables. The proxy for tax evasion they use is called the “VAT gap” which is equal to the difference between the theoretical VAT liability (GDP

multiplied by the marginal VAT rate) and the actual VAT revenue collected, standardized to GDP or the theoretical VAT liability. Every regression always included fixed country effects, a time trend and unemployment to control for the business cycle and found similar results for either of the 2 standardizations. The regression of the VAT gap over total card transactions per capita provided a negative and significant coefficient while the same regression ran separately for credit and debit cards found only similar results for the latter with no significant results for credit cards explained by the lack of enough observations. Results were robust when the measures were constructed as the number of transactions per card rather than per capita concluding that the use of cards per se is an effective technique to reduce tax evasion. However, the mere possession of them is not, as when the VAT gap was regressed on the amount of cards (with a payment function) per capita, no significative statistical relationship was found. The regression on the number of POS1_transactions per capita provided the same negative coefficients which was expected given the high correlation between POS and card transactions. On the other hand, the regression of VAT evasion on the number of cash withdrawals from ATMs per capita gave a statistically significant positive estimate signaling that a higher cash use, increases the VAT gap. Finally, there is no evidence of a

relationship between wire transfers (transfers and direct debits) and VAT evasion. The empirical challenge was that the choice of the payment method could be endogenous to tax evasion (e.g. the norm might be to underreport sales by offering discounts if paying with cash). To address this problem, the researchers restored to IV using the number of ATMs per capita as an instrument and obtained the same results. The reasoning behind using the diffusion of ATMs is that they influence the cost of cash payments but are exogenous to tax evasion as they are set up according to banks’ operative and commercial decisions. Given these findings, the authors concluded that the widespread possession of cards per se does not fight tax evasion, as they could be used for

1

(8)

8 cash withdrawals, but policies should aim to promote its use for payments, which does reduce tax evasion.

The research conducted by Madzharova (2014) is similar as it investigates the connection between payment methods, again with data from the ECB, and a slightly changed dependent variable, the VAT compliance/collection performance measured as the ratio of VAT revenue to consumption divided by the marginal VAT rate. The same econometric instruments as before were used and the main explanatory variables are the value of card transactions and the value of ATM cash withdrawals (Over the Counter withdrawals would have been included if they had not been available only for a very limited set of countries and years). Overall, they find ambiguous (sometimes significant with the expected sign and sometimes not) results with respect to card but not to cash, which clearly decreases VAT collection performance. For instance, when the relationship between VAT revenues and cash or cards is assumed to be linear, none of the coefficients is statistically significant but when the quadratic terms are added, the value of cards remains insignificant but cash does not and has a negative coefficient. When the number of POS terminals is used as the regressor, they have a positive impact on VAT collection performance as expected. The results remain consistent across various specifications like excluding the years from 2008 to 2010 which are considered recessionary or splitting the sample according to geographical location or level of income, to the incorporation of additional variables like the unemployment rate to denote the general state of the economy or an index of corruption which can influence the “moral commitment” with paying taxes. To experiment with further methods, researchers calculated the averages of the variables by country and ran a cross-sectional regression. With only 26 observations, standard errors increased substantially with most coefficients becoming statistically insignificant, even though the estimates were virtually unchanged from the longitudinal analysis with fixed effects estimated previously. Overall, we can assure that going cashless is an important tool to reduce tax evasion.

2.4 Effects on Economic Growth from Transparency

Whereas section 2.1 reviewed the impact of the payment method on transparency, this section will discover the effects transparency has on economic growth. From a theoretical point of view, while the efficient grease’ hypothesis argues that corruption enhances efficiency by reducing delays caused by bureaucracy (Leff, 1964) others state that it is prejudicial for the trust in institutions and business climate which hurts investment and income growth. On the empirical side, models used to determine growth rates or levels of GDP in absolute or capita terms, do not vary much: traditional determinants of wealth are generally augmented with one of the famous indices of transparency to measure its effects on income. Mustapha (2014) picks a sample of countries with varying levels of corruption observed from 2003 to 2011 and makes use of pooled OLS, fixed and random effects and IV with the 2SLS estimator to solve the simultaneity bias (post-estimation test confirmed it was necessary) to support a negative impact of corruption on the level of GDP/capita. Hassaballa (2017) also makes use of panel data techniques (IV) for a sample, like most studies on corruption, of 12 developing countries throughout a more extensive time period (1996-2013). The index was taken from Transparency.org and its lagged levels were used as instruments firstly, to then run a second regression using internet penetration as the instrument under the reasoning that it should not be directly related to economic growth and should be relevant for transparency by making society more aware of government actions. Again, results prove that corruption is in the detriment of

(9)

9 income. From the other side of the empirical literature, Rock et al. (2004) is one of the few and most recent studies to outlay the benefits of corruption although their conclusions are only limited to the lately high growing countries in East Asia where they argue that exchanges between the public and private sector seemed to be mutually beneficial. At the same time, they agree with the rest that corruption could be prejudicial for small developing countries. The research covered for this topic fails to provide a determinate answer (either in the theoretical or the empirical level) about whether corruption will damage or benefit economic growth.

2.5 Effects on Economic Growth from Consumption

As section 2.2 reviewed the effects going cashless has on consumption, this section reviews the effects of consumption on economic growth. From a theoretical side, Supply-side and Keynesian economists disagree on what is the component of aggregate demand (Consumption +Investment +Government purchases +eXports –iMports) that leads economic growth. The first group argues that economic growth is dependent on productivity gains, which are achieved via investments in infrastructure, human capital and technology, while consumers’ demand is completely irrelevant. Keynesian theory contrasts this idea and is famous for proposing fiscal and monetary policy as a tool to come out of a recession thanks to the multiplier effect government purchases have on consumption, and hence consumption on investment and lastly GDP growth. To test this empirically, Ramudo et al. (2014) built a Structural VAR model for the Spanish economy during the late financial crisis to see if permanent shifts in the consumption-saving pattern would have permanent effects on investment and unemployment. Their findings were more in line with Keynes’ proposition as shocks to consumption do lead to long lasting effects on investment and unemployment, and the latter was affected to a greater extent by shocks to consumption than shocks to investment. According to their view, recessions should be addressed via measures designed to sustain the consumption patterns if employment gains are desired. Contrarily, Zulkefl et al. (2012) focused on Malaysia and found somewhat similar results but exclusively for the short run. Using a Structural VECM impulse-response model, the empirical results revealed that household consumption (and fixed investment) can influence output growth in the short run, supporting the household consumption-led growth theory, but the effects of both variables diminish in the long run for which a supply-side policy would be needed. For example, by imposing a consumption shock to the system they find that the ln(GDP) responds positively up to the 5th_{quarter (almost a year) but from then up} to the 15th_{, responds negatively to return to the equilibrium path afterwards. Moreover, they did} find a reverse causality effect through which economic growth impacts permanently household consumption (and investment). To sum up, theory and empirical findings contradict each other about whether consumption can be the source of economic growth so no definite conclusion can be made.

2.6 Effects on Economic Growth from Tax Burden

As section 2.3 covered the effects of going cashless on tax evasion and in the Introduction it was mentioned that the increase in the tax base of contributors could be used for reduction of the tax burden or improvement of the fiscal stance, this section reviews the relation between the tax burden and economic growth. Reed (2008) collected data from 1970 to 1999 for the 48 continental

(10)

10 states in the US and set up his regressions using the Cobb-Douglas production from the Solow model augmented with human capital (Mankiw et al. 1992), to find that current period’s increases and lagged levels of taxes are negatively related to income growth, like any supply-side economist would suggest on the basis that government should do the maximum to not interfere in the economy de-regulating as much as possible and cutting down taxes to reduce distortions. Reed made sure to test for robustness of his results modifying variable definitions, time span, the sample of states and the definition of the 5-year intervals in which the variables were organized. However, when Gale et al. (2015) picked up Reed’s model to test its robustness to a variety of straightforward modifications like extending the sample period up to 2006 firstly and then up to 2011, thus including the Great Recession, they found that it reduced the absolute value and even eliminated the statistical significance of the estimates. Even signs changed contradictorily from negative in the first half of the time span to positive in the second. Moreover, while Reed et al. (2008) used the tax burden as the main variable of study, Gale et al. (2015) criticizes the method and decomposes tax revenues into personal income, corporate, sales, property and other miscellaneous, and found contradictory effects with property taxes exerting consistently negative effects and income and corporate taxes usually exerting positive effects. When government spending and a variety of other economic, social, and political variables are controlled for, results remain in place. To sum up, final conclusions of this study disregard the validity of previous conclusions made by Reed to assure that neither tax revenues nor top income tax rates bear stable relations to economic growth, coinciding with the Ricardian equivalence which proposed that government’s financing decisions (in this case taxes) was irrelevant and only the quantity of government purchases could affect agent’s decisions and therefore the economy (Romer). They do find however, that taxes used to fund general expenditures are negatively and significantly associated with income growth, and that raising the top income tax rate by 1% reduces the rate of firm formation by about 0.1% per year although it has no discernible effect on employment (as with income growth). Ojede et al. (2012), like just mentioned, find that property (and sales) taxes negatively affect growth, while corporate and income taxes do not, and that overall tax burdens does affect long-term but not short-term growth. Bania et al. (2007) emphasizes again the relationship between government expenditure, taxes and growth. The relationship between taxes and growth is linearly positive for low levels when they finance productive public investment before turning into a negative quadratic relationship because of the distortions and crowding out of private investment, shaping what they called “growth hills” of taxes. To sum up, like with corruption and consumption, the effect of taxes on economic growth is another popular topic in the research literature with inconclusive effects. What complicates the matter more is the fact that many times government revenues do not vary by themselves but with spending and primary surpluses/deficits so a regression that omits these variables can suffer from omitted variables bias and therefore get unreliable estimates, while one that includes them may suffer from multicollinearity complicating the distinction among individual effects (Stock & Watson).

2.7 Effects on Economic Growth from Fiscal Stance

As shown in the middle channel of figure 1 in the Introduction, if going cashless reduces tax evasion, its final effects on economic growth will depend on if it is used to reduce the tax burden (covered in the preceding section) or to improve the fiscal stance which will be covered in this section. Roubini and Sachs developed the political economy theory that justifies the persistence of deficits due to

(11)

11 weak governments that delay adjustments, even though they might be the best outcome, because of lack of consensus and end up taking the economy to an unstable equilibrium like the one shown in Calvo’s debt model. “Expansionary fiscal adjustments” theorists argue that on the demand side, a current fiscal adjustment per se can induce agents to believe that future adjustments will not be necessary which may be expansionary via a wealth effect that fosters consumption, because consumers anticipate a permanent increase in their lifetime income or because it can lower premium on government bonds leading to an appreciation of assets. On the supply side, how the contraction is composed matters and the “labor market view” states that a cut in expenditure can constrain the job market lowering union’s upward pressure on wages which would lead to profitable investments and growth, while an increase in taxes reduces the net wage of the worker which leads to an increase in the pre-tax wage faced by the employer and the opposite effect on growth (Romer). Alesina et al. (2010) aimed to investigate if size and composition of budget adjustments (or expansions) could increase output while at the same time reduce debt. To rule out cyclical fluctuations, the sample is composed of periods of large changes in fiscal stance done in high income OECD countries from 1970 to 2007. For definition purposes, a period of fiscal adjustment (stimulus) is such that the primary balance improves (deteriorates) by at least 1.5% with respect to GDP, leaving out periods of small changes prolonged for years (for practical reasons), while the composition is measured as the change in the fiscal item with respect to the total size. As the focus is on fiscal policy, the primary balance excludes financial items to make the analysis robust to exogenous changes in interest rates, monetary policy or the collateral effects of the growth path. Reversed causality from the state of the economy towards fiscal policy (either in magnitude or whether to act upon the spending or revenue side) is not considered an issue as those decisions are largely due to political/bargaining reasons and are usually taken in a calendar period precedent to their actual implementation. Regressions are done separately for periods of fiscal stimulus and contractions and control for lagged values of the dependent variable, average growth rates for the rest of the countries, ratio of public debt and a set of fiscal policy variables. In regards of fiscal adjustments, the ones associated with higher GDP growth are those in which a larger share is done via cuts in spending, while a 1% (percentage) increase in tax revenue, for instance, is found to decrease growth by around 0.33%. In years of stimulus, on average among different regression specifications, a 1% (percentage) increase in spending is associated with 0.75% lower growth significant at the 5% level. Instead, cuts in taxes do not have statistically significant effects, and the primary balance worsening per se have a negative coefficient in only 1 among the multiple regressions. Given that composition was found to be more relevant than size in either case of fiscal adjustments or stimulus, the paper suggests that to achieve fiscal stability and economic growth simultaneously, primary spending is to be cut. Otherwise, raising taxes would have recessionary effects that would end up rising spending faster than what tax revenue does. With a similar approach, Favero et al. (2015) use fiscal consolidation episodes identified for 17 OECD countries (all from the EU except for Canada and the US) over a 30-year period to also support that successful consolidations depend on their design whereas spending cuts are much less costly than tax-based ones. The main difference relies on investors’ confidence which does not fall much after an expenditure adjustment and promptly recovers but falls for several years after a tax-based. In addition, Favero et al. (2015) proved that when spending cuts are seen as permanent, wealth effects are larger on the demand-side, outlaying the importance of permanent changes on expectations when compared to transitory. But probably their most innovative contribution to the literature is

(12)

12 testing whether it is a mere coincidence that tax adjustments are chosen during recessions while the expenditure ones during expansions, or if these last ones are successful because they usually come accompanied with market oriented reforms. To carry out their analysis, they use the deviation of output from the Hedrick-Prescott trend as a measure of the cycle and an index of labor market reforms constructed by the OECD, to run a probit regression (method used to measure how the probability of an outcome occurring increases given certain events) that found no evidence of a relation between the cycle or the degree of market reforms and the choice of adjustment.

On the opposite side of the literature, Keynes’ theory points out the benefits of incurring deficits by expanding government expenses to get out of the trough, and has lately received much support from economists who regret not having applied counter-cyclical policies during the last recession. Fatás et al. (2018) provide evidence that hysteresis effects lowered the path of GDP permanently and were caused by the pro-cyclical fiscal policy implemented during the Great Recession which ended up, contrary to its original intention, in a higher debt ratio path. By means of cross-country comparisons that had, as the dependent variable, the long-term change in output growth and the potential output predicted by forecasters, they showed that the fiscal policy implemented had a negative impact on both. To conclude, a positive or negative level or change of the fiscal stance per se does not always have the same effect on economic growth, and what is more, it usually matters how these changes are composed. Therefore, once again we cannot make conclusive asseverations of the relationship between public finances (fiscal stances) and economic growth.

2.8 Control Variables

From section 2.1 to 2.3 (in this chapter) we made a point of how a cashless economy can affect transparency, consumption and tax evasion and from section 2.4 to 2.7 how these can in turn affect economic growth. Throughout the literature covered previously, economic growth regressions (from section 2.4 to 2.7) tended to be set up with rather similar control variables included. Control variables are those we are not particularly interested in but that might be related to the dependent variable and if positively related to one of the regressors of interest as well, might bias its coefficient upward while if they are negatively related might bias it downward. If not included, the regression would suffer from endogeneity caused by omitted variables and regressors’ estimates would be unreliable and therefore useless to make conclusions (Stock & Watson). The most common and significant control variables included in the previous literature review of the effects consumption, transparency, tax burden and the fiscal stance have on economic growth, will also be included in my cross-country growth regressions to obtain the effect (all together) those variables have on economic growth so by the end I will be able to make conclusions if going cashless (already known to have certain effects on these variables) could translate into benefits on economic growth. I will next provide a review of the ones selected.

The primary reference in growth economics, which is used as the baseline model in most of the empirical literature, is the neoclassical paradigm developed by Solow. Given the diminishing returns of the factors of production in a Cobb-Douglas function, Solow stated that countries will eventually reach a steady state determined by their investment and population growth rate where the long-term growth in GDP/capita is delong-termined by the growth rate in technology. Mankiw et al. (1992) added human capital to the model so that when tested empirically, would provide more reliable estimates although it still left the growth of technology explained exogenously. The AK model, the first version of the endogenous growth theory, explained technology as coming from knowledge

(13)

13 spillovers inherited in the accumulation of capital, while Romer proxied technology evolution as the creation of varieties of products inspired on the new theory of international trade which gave countries access to new inputs and knowledge. Finally, Schumpeter said that the turnover of products (“creative destruction”) is its driving force (Aghion & Howitt, 2009). To account for these theories, my model includes the fixed capital investment and population growth rate, an index from the Penn World Table (PWT) that acts as a proxy for human capital and R&D investment to proxy technology. Except for the population growth rate, that is supposed to have a negative relationship with economic growth, all others are supposed to be positively related with economic growth (and are standardized to GDP except for the human capital index). In regards of human capital, due to the fact that most EU members are developed economies, a second possible proxy for it could be the percentage of the population with tertiary education as it was found to have an impact on innovation and consequently on economic growth in those places closer to the technological frontier, while investments in primary and second education (taken into account in PWT’s index) did not (Aghion et al., 2009). Therefore, both proxies for human capital will be used.

Trade, mentioned by Romer and abundant in the literature, is another control variable I will use and is measured as the sum of the absolute values of exports and imports (standardized to GDP) and is expected to have a positive sign on its coefficients according to the Heckscher–Ohlin theorem that says countries can benefit from exporting goods made with their abundant factors and importing those made with their scarce factors, and the Ricardian theory that says countries can benefit from exporting those goods where they have a comparative advantage in its production and importing the rest (Aghion & Howitt, 2009).

As was made clear during the literature review related to fiscal variables, taxes or the fiscal stance might not act alone on GDP growth and its effect might be dependent on government expenditure or the level of indebtedness, which merits to control for these two variables (standardized to GDP) in my regressions. To analyze the relationship between government expenditures and economic growth, Fournier (2016) builds on a baseline convergence model from the neoclassical growth theory, augmented with public investment, primary spending (to account for the budget constraint) and its components for a panel of OECD countries, and reveals a positive growth effect of public investment in defense and education among others, but particularly high in R&D and health, reflecting the plausible increase in worker’s productivity. To allow for non-linearities, the level of public capital stock and investment are combined to show that benefits can be larger in countries with an initially low stock of capital and can increase the speed of convergence for them, but marginal returns of public investment may turn negative at high stock levels. Results remained robust when excluding the crisis years or defining variables in different forms, and the biggest threats to internal validity which were the business cycle and Wagner’s law (richer governments spend more), were solved by using fixed effects and a large set of control variables, and cyclically-adjusted GDP data to attenuate the effects on investment, respectively. In regards of the influence the debt level can exert on economic growth, research done by the ECB (2010) used the Solow’s conditional convergence model augmented with the debt level for a sample of 12 members of the EU over a period of about 40 years, to conclude that no linear effects exist between debt and GDP growth but a concave relationship does exist that does not change for different polynomial forms. The empirical proof coincides with the Debt Overhang model developed by Obstfeld and Rogoff that proposes an inverted U-shaped pattern of debt, for which at high ratios, turns prejudicial for growth as it disincentives investment due to the implicit tax (repayments) on its return. Rother (the author) restores to IV with the 2SLS and the GMM method to solve for the feasible reverse causality from the state of the economy towards the debt level, using as instruments the lagged values of debt and the average level of the rest of the entities, replaces the real for the potential GDP growth rates to avoid fluctuations in the economic cycle, and annual or 5-year specifications to distinguish between

(14)

14 short and long-term effects. The turning point from which debt diminishes growth found in his study is 90%. Results were robust even when including in the regression the ratio of domestic credit to the private sector under the hypothesis that government debt effects could be exacerbated for higher levels of private debt. As said before, accumulated deficits (increases in debt) could have a positive impact on growth if used to finance productive public investment, but because over the past decades in Europe they were related to consumption and transfers, they tend to be associated with uncertainty about future policy together with an upward pressure on interest rates, which may be counterproductive for investment, productivity and growth (ECB, 2010).

The unemployment rate is used in my research to control for the business cycle as used in Stevenson et al. (2011) where they measured how the confidence in different institutions during the last Great Recession reacted to its fluctuations, proxied with the unemployment rate. The study conducted by Bitler et al. (2015) is another one that uses the level of unemployment to account for the economic cycle where they investigate its effects on household income and specifically poverty. According to Okun's law that establishes the relation between employment and growth, the coefficient on this variables is expected to be negative. Lastly, the conditional convergence hypothesis states that as long as countries share the same technology and the same fundamentals, they should have the same steady state and therefore grow faster the further behind they are from it. Those fundamentals are already contained in the set of variables previously described, so to control for this theory, I will include the lagged level of GDP/capita which should have a negative sign (Aghion & Howitt, 2009).

Fixed and mobile broadband and internet penetration are not so often used as control variables in a regression testing the sources of economic growth, but given that the cashless transition is dependent on them, it is sound to include them as to obtain reliable estimates of the channels of impact on GDP growth for the average country, no matter the individual levels of penetration in present or to come. Fixed and mobile broadband allow the remote access of information and are associated with the first wave of digitization (ITU, 2017). The World Bank Development Report (2016) reviews several studies focused on the relationship between broadband connections and GDP growth. A cross-sectional analysis for a sample of 120 developed and developing countries for the period 1980-2006 found that a 10% (percentage) increase in fixed broadband penetration would result in a GDP growth from 1.21% in developed economies to a slightly higher 1.38% in the developing ones, while for a sample of Latin American countries through the period 2003-2009 they found the highest estimate: a 3.2% increase in the growth rate of GDP/capita for a 10% (percentage) increase in penetration. Despite the methodology and the sample, the economic impact from fixed broadband was always positive and discernible after a certain threshold of penetration. In research done by Deloitte (2012), 96 countries from 2008 to 2011, within the usual economic growth equation taken from Solow and controlling for the usual variables, were used to measure how mobile penetration affects economic growth. In an IV estimation using lagged values as instruments because of the risk of simultaneity, they added mobile penetration as an additional variable to find that it had a positive impact on GDP/capita growth, and further analysis concluded that an increase in the speed of connectivity, that is a 10% (percentage) increase in 3G penetration from 2G, could raise up to 0.15% the growth rate of GDP/capita. Moreover, using the same methodology for a different sample of countries they found that not only the change in technologies but doubling the use of 3G data, after controlling for the level of mobile penetration, could increase GDP/capita growth with 0.5%. The channel through which fixed and mobile connections can foster growth is the Total Factor Productivity (TFP), a term used in Solow’s equation with a multiplier effect on the labor force. After this first wave (connectivity) came the second wave of digitization that entails the diffusion of the Internet (ITU, 2017). Choi et al. (2009) used the internet rate of penetration as the proxy for the knowledge spillover introduced in the AK model, for a sample of more than 200

(15)

15 countries during the period the technology exploded (the 90s) to find that 1% (percentage) increase in the ratio accomplished around a 0.05% increase in the growth of GDP/capita.

it

+ X

it

+

α

i

+ δ

t

+ ε

The first four variables after the constant (CONS) are the variables of interest: consumption (Cons), an index of corruption (Corr), the tax burden (T) and the government primary budget (PB), followed by a vector of control variables and fixed effects either at the level of the entity (

α

) or time period (δ). As shown in the literature review, the transition away from cash has the impacts mentioned in the introduction (see figure 1 in that section) over the four variables of interest in the regression. Then, depending on the sign and magnitude of the coefficients I will find for the variables of interest, I will be able to reach conclusions if there are any gains from getting rid of cash in terms of economic growth. Only panel data techniques will be applied as it is the most convenient way to explain how these variables determine the growth in GDP per capita for different entities throughout time. The first model is estimated via Pooled OLS (using the growth rates of the 4 variables of interest measured as Share of GDPt – Share of GDPt-1 or Indext – Indext-1). In terms of how the regressions will be conducted, I plug in the usual variables contained in any regression from the literature of economic growth, which are the fixed set of control variables in my analysis, and augment the model with the 4 explanatory variables, one by one and then all together, to lastly add the second set of control variables (internet, mobile and fixed broadband penetration). Because the sample includes the 28 members of the EU observed from 1990 up to 2014, unless some observations are dropped because of missing values in certain variables, the (maximum) number of 672 observations is enough to afford a large set of control variables. In fact, I could have included more variables used sparsely in the previous literature (e.g. inflation rate) but too many could have reduced the precision (via standard errors) of valid predictors making them insignificant. The reason why regressions are run in steps is to analyze how the sign, magnitude and significance change when explanatory variables

(16)

16 are entered alone by themselves or all at the same time, either with one set of control variables or with both. If two or more of the explanatory variables were correlated between them, the model may suffer from multicollinearity turning coefficients insignificant when it should not. One way to measure multicollinearity is the variance inflation factor (VIF) which will be run after every regression (a value between 5 and 10 indicates high correlation that may be problematic).

Nevertheless, estimates obtained from Pooled OLS are not reliable as they may still suffer from omitted variables bias when unobservables enter the regression through the error term. Fixed effects is the model to overcome the issue and is ran in subsequent steps as previously described while referring to Pooled OLS (and using the same definition of regressors). At the level of the entity, country in this case, fixed effects absorb unobserved characteristics which differ between countries but remain constant through time while time fixed effects absorb everything that varies through time but has the same effect on every country in each time period. A simple command tells Stata to run the regression with entity-level fixed effects but for time fixed effects it is necessary to create a dummy variable for each year except for the first to avoid a “dummy trap”.

Given that we are working with annual data, we should be aware of the vulnerability to measurement error which could violate one of the assumptions of OLS, that is the strict exogeneity of the error term with the variable measured imprecisely, and serial correlation. Over long time periods, measurement errors are likely to cancel out but to allow for serial correlation of the error term, I use in every regression clustered standard errors (at the country level) which are also robust to heteroskedasticity. Once the final framework model is obtained, IV will be used to solve for simultaneity/reverse causality (when the dependent variable is the one causing the change in the independent instead of it being the other way around) with 3 different methods to test the robustness of the results: with and without entity-fixed effects, and with the model estimated in first differences. In the Results section, the instruments used within the 2SLS method and their tests for relevance (are correlated with the endogenous regressor) and exogeneity (not related directly to the dependent variable via the error term) will be described. To test for dynamic effects, the same procedure of plugging in one explanatory variable at the time before all together and with one set of control variables after the other is done again, but this time lagging the values of the explanatory variables. As we will be working with lagged values, simultaneity will not be an issue anymore in these regressions. To conclude, growth rates will also be replaced with levels (shares of GDP or corruption index level) within the original framework model, and then used simultaneously (growth rates and levels) to reach final conclusions, in a process that will be detailed while explaining the results obtained. From this point where we will already be certain about our findings, what comes afterwards are interactions between variables to get further insights, and redefinition or replacement of variables with similar proxies, removal of certain observations and split of the sample by different criteria to test the robustness of the conclusions.

4. Data

As mentioned in the previous Methodology section, the analysis will be carried out for the 28 EU member states during the period 1990-2014 although the first observation of each entity (1990) is lost as the variables used are growth rates. Therefore, observations in the sample cover economies

(17)

17 completely different in terms of development and history and throughout a wide period of crisis and recoveries which is enriching. Obtaining a complete dataset was a challenge given that for some variables, retrieved from a specific source, there were countries entirely missing (e.g. Croatia, Bulgaria, Malta, Romania, Lithuania and Cyprus) or incomplete, particularly when dealing with the furthest away time periods like the beginning of the 1990s. Sometimes missing observations were as much as 16 for a certain variable and others only a few like 2. Notwithstanding, I still believe that after Stata drops those missing observations, the number of observations is enough to obtain reliable estimates (as it still ranges from 281 to 285). In the Appendix B, we can find the detailed list of the sampled countries and descriptive statistics of the variables used throughout the analysis with their definitions. However, it is worth explaining the scores of the “Corruption Indices”: the Transparency.org’s ranges from 0 to 100 and the World Bank’s from -2.5 to 2.5 meaning in both cases from the most to the least corrupt. GDP/capita growth, which will be the dependent variable in every regression to proxy economic growth (others like GDP/labor force resemble productivity instead), and the regressors, that is the change/growth rates of the shares of GDP of consumption, tax revenues, primary budget and changes in both indices of corruption (expressed with “_Dif” in the summary statistics and calculated as Share of GDPt – Share of GDPt-1 or Indext – Indext-1) have a valuable dispersion: observing the standard errors and the range of these variables in the summary statistics, we see that they are always high compared to the absolute value of their means. Their average changes/differences per year are graphed in figures 10 to 12 for ex-Soviet/Communist countries and in figures 13 to 15 for the rest. Most of the volatility in GDP/capita growth rates is due to Eastern European countries who suffered from being dependent on the Soviet Union when it collapsed in the early 90s, as they went into recessions when hyperinflation and the free market economy took over and even the Baltics had to start issuing their own currency. Among the most developed countries, Sweden and Finland saw a welfare and banking type of crisis respectively in the same period, and Greece and Ireland were among the most affected due to the housing crisis started in the US by 2008. Among the data, it also noticeable the spikes in growth, particularly for the same Eastern countries when liberalization of the economy and later (in the early 2000s) entry into the EU occurred, and for some economies in the preceding years right before the 2008 crisis. Even though the mean change in the level of consumption (share of GDP) is almost 0 in every country, volatility was high through time particularly for Bulgaria, Croatia, Latvia, Lithuania and Romania during the early 1990s too, and also for Luxembourg and Malta at the time of the crisis. A similar story is repeated for the changes of the primary budget with its mean around 0.1% but a standard deviation of 2.6%. Volatility is focused mostly on Ireland who experienced large deteriorations in its deficits from 2008 till 2010 to then recover strongly in the following 2 years, and to a lesser extent on Lithuania and Latvia again. In terms of tax revenues, the biggest upward variations were for Sweden and Malta in the mid-90s and the biggest drop for Cyprus in 2009, when their banking industry fell apart, and for Malta by the same time. All other countries seem to have had stable variations through time. Lastly, in regard of the scores provided by Transparency.org to measure the perception of corruption, almost every observation seems to follow a pattern of lost and gain of a few points every year, resembling the inertia effect proved by Singh et al. (2017), while the biggest deterioration happened to be on the years the crisis step in for those most affected by it (Spain, Malta, Italy, Portugal, UK, Greece) as the study from Stevenson et al. (2011) on confidence in institutions suggests.

(18)

18 Scatterplots of each country’s GDP/capita average level against the average level of explanatory variables are drawn in figures 16 to 19. Except for the clear positive relation between GDP/capita and scores in the transparency index meaning that the less corrupt, the wealthier the country, all other relationships seem to be rather dispersed. Nevertheless, countries whose economy tend to rely more on consumption seem to have had through time a slightly lower GDP/capita, as well as those with higher primary deficits and lower tax burdens. Figures 20 to 23 show the scatterplots of the average GDP/capita growth rate against the average growth rates of the same explanatory variables. This time the dots are even more dispersed than before but, if needed to make a prediction, it seems that in line with what was mentioned previously, reductions in consumption accelerates growth in GDP/capita. Small increases in surpluses rather than big, appear to be positive for growth, while its reduction or incurrences in deficits are too dispersed to make any assumption. Variations in the index of corruption do not seem to bear a stable relationship with GDP/capita growth. For tax policy, dots are too dispersed also to make clear guesses. Keeping tax revenues participation over GDP constant, increasing or decreasing it, matches with either low or high growth rates of GDP/capita although this dispersion becomes more pronounced in the case of reductions and the dispersion of increases in the tax burden is among lower growth rates. That said, if some preliminary conclusions are to be done, raising the tax burden slows growth. Correlation matrixes are useful to see how different variables move together (or in opposite direction) but should by no means be taken as a determinant of causation. The first one in the Appendix B has the coefficients between those variables involved in the regressions to be done. It confirms what suggested before by the scatterplots, that wealthier countries (see level of GDP/capita column) are those with sound public finances (primary surpluses), whose economy is less consumption based, where the tax burden is higher and are less corrupt (and with higher R&D investment, internet, broadband and mobile penetration). Moreover, the GDP/capita growth rate is positively correlated with increases in surpluses, transparency and reductions in consumption (see correlation with “_Dif” variables) although the correlation with the change in taxes (“TaxSh_Dif”) is practically 0. In fact, many of the coefficients for the variables of interest’s growth rates (“_Dif” variables) with the rest of the variables in the dataset are almost 0 (not much correlation to discuss) but its levels are not and give stronger support to what just mentioned. A lower level of consumption in GDP can be associated with sound public finances, a higher tax share and less corruption and these last 3 go hand in hand in the same direction. Lastly, the 2 indices of corruption (Transparency and World Bank) have a correlation coefficient of almost 1 so they must be similarly estimated. The second matrix that comes next is between variables (measured as quantity or value of transactions) that proxy the payment methods mentioned in the introduction (cheques, direct debits, credit and debit cards, POS and withdrawals) and its analysis will be necessary for the final conclusion of this thesis. From the variable that proxies the use of cheques onwards, they all measure the preference for cash instead of electronic methods of payment. The correlation coefficient between these variables are positive and close to 1, meaning that a reduction in any of them comes with an almost equal reduction in all the other methods. Besides, the use of POS, direct debits and debit cards (substitutes of cash) have coefficients between them close to 1 and highly negative with those proxies of the use of cash/cheques. Hence, it is reasonable the assumption that a 1% (percentage) increase in any non-physical payment method is accompanied with exactly the same increase in the rest of the methods that belong to the same (non-physical) group and the same reduction in the use of cash.

(19)

19

5. Results

Section 5.1 explains the results of interest for this research: in Appendix C, tables 1 to 7 present the coefficients for the growth rates (in the variables of interest that affect economic growth) using Pooled OLS, Fixed Effects and IV; tables 8 through 12 investigate dynamic effects; and tables 13 and 14 introduce interactions between variables to get further conclusions. Section 5.2 test the robustness of those results: tables 15 to 17 test the robustness to variables definition and table 18 to the time period and the observations included in the sample; table 19 analyzes heterogeneous effects among groups in the sample. Section 5.3 calculates the economic growth gains that can be achieved from going cashless (see table 20 and 21).

5.1 Main Regression Results

Table 1 and 2 in the Appendix C contain the regressions done via pooled OLS. The first one has the regressions done with the common set of control variables plus each variable of interest

individually, and by the end the 4 of them together. The only 2 regressors that happen to be significant both times are the share of consumption and the primary budget and have the same sign as expected. A 1% increase in the share of GDP devoted to consumption or a 1% reduction in deficit or increase in surplus translates into 0.33% slower or faster GDP/capita growth

respectively. Changes in the index of corruption are not significant and even their point estimates are almost null. The tax burden changes sign proving that the first regression including it alone, was suffering from omitted variable bias. In table 2, which adds the connectivity/digital control variables, the explanatory variables keep their significance, sign and magnitude. “Corr_Dif” remains equal to 0 but interestingly now taxes conserve the negative sign though not significant. So far we have a preliminary impression on the channels where the transition from physical to digital settlements can spur economic growth but, as said before, we need a better method to obtain reliable estimates. Table 3 and 4 apply “Fixed Effects” and regressions are set up in the way explained in the Methodology section, which has the same logic just applied with pooled OLS. It is worth noting that the 5th and 6th columns of table 4 test the conditional convergence theory and controls for the business cycle making this last one the most complete of all regressions. In both tables, consumption still has a detrimental effect and corruption is inert, but now the primary budget change is negatively correlated with GDP/capita growth and loses its previous significance, and taxes became the only significant coefficient of interest and a reduction in its burden of 1% is interpreted as an increase in half of it in the growth rate. Results seem consistent across various specifications as (for tax differences) coefficients remain significant, do not change in magnitude and standard errors diminish while the adjusted R2_{, which tells how well the regression fits the} data, increases along up to 0.6 which is considerable. However, from table 3 to 4, “TaxSh_Dif” loses importance as their significance level goes from 1% to 5% when the last 2 variables added capture part of it: recessions (increases in the unemployment rate) reduces growth and countries further away from the steady state do seem to growth faster to converge (negative sign on the lagged level of GDP/capita). One possible threat to the validity of these results could be simultaneity if it is the GDP/capita evolution the one determining the behavior of the RHS variables. As mentioned previously in the “Literature Review”, fiscal variables could behave according to political outcomes and determined in calendar periods precedent to when

implemented, making them exogenous to the state of the economy; there is no sound reason to think that wealth determines the level of transparency, but from the literature review (Zulkefl et

(20)

20 al. (2012) ) and remembering that the consumption function postulated by Keynes was

determined by current disposable income (gross income less taxes) so that, within a country over time, aggregate consumption is proportional to aggregate income (Romer), “ConsSh_Dif” is the one most likely to be endogenous. Nevertheless, we can assume all explanatory variables are endogenous basically to clear doubts. Table 5 through 7 show the results from estimating the same last regression from table 4 (control variables were dropped from the table to focus on the ones of interest) but assuming, one at the time, each explanatory variable endogenous. IV with the 2SLS method is firstly used without fixed effects, secondly with the variables estimated in first differences and lastly with fixed effects (demeaned variables). Under the three specifications, for “ConsSh_Dif” the instruments used were the growth rate in POS, for “PBudSh_Dif” its own lagged values, for “TaxSh_Dif” the lagged values of the changes in share of GDP of the main taxes

(personal, corporate, consumption, property and contributions) as well as for “Corr_Dif” combined with its own lagged values. The justification for picking these ones are that the diffusion of POS gives a merchant the possibility to sell more by accepting wider payment methods but is not expected to affect GDP growth directly and thereby enter the second stage regression through the error term. Lagged values are called internal IVs and were proposed by Arellano and Bond and variations in taxes, because of being just a part of the whole tax burden, should not have a direct impact on GDP growth but does on the tax burden as well as in the level of corruption if more governmental intromission predisposes it. Figures 24 to 27 contain the tests for the instruments. They go in the same order that regressors were supposed endogenous in the tables with the regressions output (e.g. from Consumption to Corruption), and at the top of each figure, the model “without fixed effects” and at the bottom “in first differences”. In every regression, F-statistics are big enough to confirm relevance and the overidentification test (“J statistic”) do not reject the null hypothesis of exogeneity. The Wu-Hausman test, with a null hypothesis of

exogeneity of the theoretically endogenous regressor, was used to test if indeed the regressors were endogenous and was rejected only once when consumption was instrumented and the model was applied in first differences (see figure 24) which confirms Keynes’ proposal (the path of consumption depends on economic growth). So there is no reason to believe that the original model suffered from simultaneity bias and applying IV turns out to be more inefficient in these cases. Nonetheless, even though the coefficients are interpreted differently if the model is

estimated in first differences than like in the other 2 cases, what is most important to notice is that throughout the regressions in tables 5 to 7, consumption is significant in most of the equations and conserves its negative coefficient that oscillates between 0.25 and 0.6, reaching its maximum when it is the one taken as endogenous instead of any other of the 3 explanatory variables

although, as just said, IV estimates should not be taken for granted as the tests of endogeneity did not reject the null. Like in previous regressions, the tax share change is the only other variable that happens to be significant too, again with a coefficient around -0.5. Next, to investigate the

possibility of dynamic effects, the same methodology of plugging in one set of controls after the other and explanatory variables all together after individually, is done but with the lagged values instead of the current periods’. Table 9 is an extension of table 8 in terms of the addition of further controls though some are dropped from the table for clarity. None of the coefficients up to the second lagged value are significant except for corruption, only once. Anyhow, corruption’s

estimate are indiscernible, consumption is still detrimental and this time, strangely, the tax change conserves approximately the magnitude but has the opposite sign than before. To investigate

(21)

21 further dynamic effects, the lagged values of explanatory variables are used again but this time defined in terms of level instead of changes in the share. In the first column of table 10 we see the first lag while in the following the second lag. The only effect is found on the level of tax burden of the previous period, with a similar magnitude and sign than in the tables that used the lagged growth rates (8 and 9) but now consumption, which used to be the second significant variable, is not anymore. To test again the effects of the tax change but in this different context, columns 3 and 4 introduces it instead of the lagged tax level. As earlier, the change in the current period, not in the previous, has a significant impact of around -0.5. What if we test again the effect of the lagged tax level but this time replacing the lagged levels of other variables for their (lagged and current) growth rates? In the first 3 columns of table 11 we see once again that the lagged level of tax is significantly positive (approximately 40% of the change in the tax burden of period t-1 is translated to the current growth rate in GDP/capita) with consistent estimates and standard errors while the lag tax growth rate is not again. Because tax differences were found significant

(exclusively) for the current period but current tax levels have not been tested yet, column 4 casts doubt away replacing the tax growth, in the original model set up with current growth rates, for the level. Column 5 is exactly the same specification but using IV and instrumenting the only possible endogenous regressor, the current tax level, with the lagged growth rates of the different taxes, using fixed effects like always for time periods and countries and clustered standard errors at the country level. The growth rate of consumption is the only (negative) significant estimate though the current tax level’s p-value was 0.11, close enough to be significant at the lowest conventional level, and with a negative sign. The instruments were exogenous, relevant and the null hypothesis of exogeneity of tax revenues was not rejected as can be seen in figure 28.

So far, preliminary conclusions are that increases in consumption and tax participation in GDP is detrimental for growth but, strangely, a higher tax burden in the past accelerates growth in the present. The first half of table 12 is the model with current growth rates and the second half with lagged levels. The first column of each half has the current tax growth and lagged level simultaneously in linear terms, the second column of each half adds the interaction term between the 2 variables and the third column of each half removes the individual linear terms to leave only the interaction. In this way we might be able to reach further insights of what are the mechanics through which past high level of taxes but current reductions foster GDP/capita growth. The current growth rates and lagged levels of consumption, corruption and primary budget are not significant in any specification; not even the current growth rate of consumption which previously was. Before, in table 10 and 11 we saw that the lagged tax level was significant in a regression with other lagged levels or with current growth rates respectively, but in columns 1 and 4 (table 12) where we now introduced the current growth rate in taxes to both same regressions, it loses its significance though the growth rate remains significant and with an equal point estimate; it seems that regressions where the lagged level was significant were suffering from omitted variable (the current growth rate). Columns 2 and 5 add the interaction between the lagged level and current growth rate, and the standard error of the latter variable increases which signals that is not clear who is causing GDP/capita growth so that none coefficient is significant, though the 3 could still be important. Columns 3 and 6 proves this by replacing the individual effects (of lagged level and current growth) for the interaction of the 2, which happens to be significant and proves that the lagged level matters only if it is to be reduced in the current period and not per se. The coefficient of the interaction term is interpreted as the higher the fiscal pressure the economy is at the beginning/end of the previous

Can going cashless in the EU have a positive effect on economic growth and if so, through which channel?