Country Heterogeneity in the Effect of Capital Taxation on Economic Growth

Country Heterogeneity in the Effect of

Capital Taxation on Economic Growth

R

M

T

Fabian ten Kate

August 21, 2015

Abstract

The literature on the effects of capital taxation on economic growth has largely ignored the possibility of country heterogeneity. This paper exam-ines a large cross-section of 115 countries over the period 2005-2013 and finds that capital taxation appears to have a different effect on growth in different countries. In the least developed countries a significantly nega-tive effect on economic growth is found, whereas the effect is insignificant in more developed countries. This key result is obtained under a number of alternative specifications, using a wide variety of control variables and including controls for spatial dependencies in the data. The analysis is further extended to find a possible explanation for the observed country heterogeneity. Capital taxation’s effect on growth varies with measures that capture how modern an economy is. Countries with low electric-ity usage, a large dependence on agriculture, and little manufacturing capabilities seem to be particularly harmed by the taxation of capital.

Keywords: Capital Taxation; Economic Growth, Heterogeneity

1

Introduction

C

be taxed as it would harm economic growth (Mankiw et al. 2009). While this view is widely accepted on a theoretical basis, the empirical literature paints a more mixed picture. One explanation for the different em-pirical findings is that there is country heterogeneity, which leads to different conclusions depending on which countries are included in the analysis. In particular, it may be the case that there are differences between developed and developing countries in the effect that capital taxation has on economic growth. In this paper I examine whether this is indeed the case.

The notion that capital taxation is harmful for economic growth was for-malized by the theoretical work of Chamley (1986) and Judd (1985). In the context of a neoclassical growth model these authors have independently shown that the optimal tax rate on capital is zero. Over the years various authors have sought to add qualifications to this result, yet it has been found to remarkably robust. Atkeson et al. (1999) note that so long as a number of standard assumptions are made, the result will almost always be arrived at.

The empirical literature on the other hand is not quite as clear in its conclu-sions. A number of papers have found that for samples of primarily (highly) developed countries capital taxation does not significantly affect economic growth (Mendoza et al. 1994, Widmalm 2001, Gemmel and Sanz 2013). Ex-ceptions are the work of Arnold et al. (2011) who report a negative growth effect, and Ten Kate and Milionis (2014) who report a positive effect. At the same time, however, for a wider sample of countries Lee and Gordon (2005) do find negative growth effects of capital taxation. One explanation for this is country heterogeneity. If the taxation of capital harms economic growth only in developing countries but not in developed ones, the above pattern of results may be observed.

Consequently the analysis is extended to examine which factors might explain the observed country heterogeneity. By introducing conditioning variables into the regression model the marginal effect of capital taxation on growth is allowed to differ across countries depending on the value of the conditioning variable. Using this method with a complementing graphical marginal effect analysis I find standard measures of economic development or institutional quality do not properly explain cross-country differences. Instead factors related to how technologically advanced a country are found to perform particularly well. Specifically, countries with very low electricity usage, a great dependence on agriculture and hunting, and with little by means of manufacturing capabilities seem to be particularly harmed by the taxation of capital.

This paper will continue as follows. The next section briefly discusses the key literature on capital taxation. Section 3 then continues with the empirical approach and a description of the data. Section 4 presents and discusses the results, followed by an extensive examination of their robustness in section 5. Section 6 extends the analysis by considering which empirical factors may explain the observed heterogeneity, followed by a discussion in section 7. Lastly, section 8 concludes.

2

Capital Taxation in the Literature

The notion that capital taxation is detrimental for economic growth was first formalized in the work of Judd (1985) and Chamley (1986). Within the con-text of a neoclassical growth model these authors show independently that a positive tax rate on capital cannot be optimal in the long run. Intuitively this occurs because a tax rate on capital distorts the relative price of consumption between today and some period in the future, leading to more consumption and less saving in the present day (Salanié 2003). Many authors have subse-quently sought to add qualifications to the Chamley-Judd result, but it has proved remarkably robust. Atkeson et al. (1999) note that as long as a number of baseline assumptions are maintained the result will typically be obtained. Specifically, if taxes are restricted to be linear, and if governments are assumed to be both benevolent and unable to commit to future tax rates, then a zero optimal tax on capital will typically be found.

rates. For example, there are property taxes, corporate taxes, capital gains taxed, and dividend taxes.

The literature has dealt with this issues in various ways, with different methods having different strenghts and weaknesses. First of all, Mendoza et al. (1994) employ detailed national account statistics for a very limited set of countries to compute average rates of capital taxation. They do not find any evidence that higher rates of capital taxation are associated with lower growth rates of output for the countries in question, all of which are highly developed OECD member countries. The main issue of this approach is that the detailed data necessary is simply not available for most countries, so that the analysis cannot be easily extended to a broader set of countries.

A number of other papers have instead opted to focus on a specific element of capital taxation, being the taxation of corporations. Data on corporate taxation is typically available for a larger sample, which allows for more extensive analyses. Here Lee and Gordon (2005) study the impact of capital taxation on growth by using statutory corporate tax rates. For a global sample of 70 countries they find that the taxation of capital is harmful for economic growth. A clear disadvantage of this method, however, is that statutory rates can be quite different from effective rates, as they do not take into account possible tax deductions and such.

same effects in every country, rather its effects may depend on for example the institutional framework (Rodrik 2006). There is no reason to assume that tax policy is any different.

3

Empirical Strategy

The general empirical strategy employed in this paper is very similar to that of Barro (1991), which in turn is broadly based on the Solow (1954) framework. It relies on a very standard regression in which economic growth is regressed on initial income, population growth and measures of physical and human capital accumulation. To this regression additional variables of interest can then be added in order to ascertain whether they affect economic growth.

In this particular case the main relationship of interest is that between capital taxation and economic growth. This gives way to a specification where the growth rate of output per worker, gi, in country i, is regressed on the

logarithm of initial output per worker, yi,t−1, the share of investment in output,

invi, the tertiary enrolment rate, enri and the growth rate of the workforce,

ni. The effect of capital taxation will be assessed by means of a variable that

captures the taxation of corporations, tcorpi, described in more detail below.

To allow for country heterogeneity in this effect a low income dummy, dumi

specified below, is introduced which will be interacted with the corporate taxation variable. Additional controls, Zi are also considered in extensions of

this basic specification. This can be summarized as:

g_i = β0+β1ln yi,t−1+β2invi+β3enri+β4ni+β5tcorpi (1) +β6dumi+β7tcorpi×dumi+Ziθ+ε_i.

For developed countries the effect of corporate taxation is thus given by the coefficient estimate β5and for lesser developed countries by the sum of

β5 and β7. In order to improve the noise to signal ratio all variables used

are taken as averages over the entire time period (2005-2013), as specified in some more detail below. This leads to a cross-sectional analysis, which should reduce the impact of business cycles and unique events generally.

3.1 Data

Table 1: Summary Statistics

Variable Obs Mean Std. Dev. Min Max Output Growth 115 2.657 2.193 -2.072 11.677 Log Initial Output 115 8.170 1.621 5.009 11.075 Investment Share 115 23.635 6.886 10.502 57.426 Population Growth 115 1.552 1.235 -1.046 3.678 Tert. Enrol. Rate 115 33.501 25.586 0.508 93.892 Corporate Tax Rate 115 40.405 11.690 8.411 69.267

All variables, except initial output, are averages over the period 2005-2013.

program, satisfies these criteria, but does so at the cost of being available for only a small number of years. The resulting sample covers a time period ranging from 2005 to 2013 and includes 115 countries, a list of which can be found in the appendix in table A2.

This measure captures the amount of taxes paid by businesses expressed as a percentage of their commercial profit. As such, it can be seen as an effective rate of taxation. It is, however, a very broad measure that includes all taxes that affect a firm’s bottom line and which accounts for possible deductions and exemptions. The taxes included fall into four main categories: corporate income taxes, social security contributions and labour taxes paid by the employer, property taxes, and turnover taxes (The World Bank 2015). Note that this definition purposely excludes a number of taxes paid by firms, such as value added taxes, because these do not affect the firm’s profit directly. The amount of taxes paid is then divided by the amount of commercial income, which measures net profit before taxation. This measure leads to effective rates of taxation between 8 and 70%.1

From the World Bank I furthermore employ an additional number of Development Indicators. First of all, data on real GDP per capita is used to compute the GDP growth rate as well as the initial level of GDP. The ratio of gross fixed capital formation to GDP is used as a measure for the investment share. Moreover, the growth rate of the working age population (i.e. age 15 to 64) is also included. To capture the impact of investment in human capital I include the tertiary enrolment rate. While other, arguably more preferable, measures of human capital accumulation exist, these were not available for as

Table 2: Correlation Matrix

1 2 3 4 5 6

1. Output Growth 1.000

2. Log Initial Output -0.466 1.000

3. Investment Share 0.291 -0.035 1.000

4. Population Growth 0.157 -0.659 0.060 1.000

5. Tert. Enrol. Rate -0.282 0.783 -0.093 -0.763 1.000

6. Corporate Tax Rate -0.054 -0.271 -0.109 0.167 -0.159 1.000

All variables, except initial output, are averages over the period 2005-2013.

many countries in the relevant time period.

For the low income dummy three alternative specifications are considered based on the World Bank’s classification of developing countries. The World Bank’s low income classification captures countries with income less than $1045 per person. Similarly, low-middle income countries are those with per capita income above $1045 but below $4125, and high-middle income countries are those between $4125 and $12746. Based on this I consider three alternative specifications of the low income dummy, being incomes below $1045, below $4125 and below $12746. The three specifications correspond to approxmately 25%, 50%, and 75% of the countries included in the analysis respectively.

In order to minimize the amount of noise in the data I employ averages of all the variables in question. Specifically, all variables employed are averages over the period 2005 to 2013, with the exception of the initial GDP measure, which is simply real GDP per capita as measured in 2005. Data on the tertiary enrolment ratio was available for most countries, but not for all years. To maximize the countries that can be included in the analysis the variable that is used is the average of any and all available data. If, for example, a country only has data on this measure for 2005, 2008, and 2012, then the measure used is the average of those three observations.

rate. Correlations between the dependent and the independent variables are mostly quite high, with the exception of the corporate tax rate.

Throughout this paper a number of additional control and conditioning variables are introduced. In particular these are the unemployment and infla-tion rate, the amount of government revenue, the Fraser Institute’s economic freedom index, the Polity score, the ICRG’s quality of government index, the GDP share of agriculture and manufacturing, electricity usage, and the number of internet users in a country. All of these variables and their sources are discussed in some detail in the appendix.

4

Results

The results of an OLS estimation of equation 1 are reported in table 3. The first column reports the results for a simple cross-sectional growth regression that accounts for initial income, investment, human capital accumulation, and population growth. The results are quite reasonable and in line with what other research on this topic has found (e.g. Mankiw et al. 2009). One issue, however, is that the measure for human capital, the tertiary enrolment rate, is not found to have a significant effect on growth, which it would be expected to have. Even so, on the whole the regression seems reasonably well-behaved, and the coefficients are comparable to those of similar long-run analyses. This is reassuring in particular because of the relatively short period of time covered by this analysis.

The second column adds the corporate taxation measure. It enters the results with a significantly negative sign offering support for a hypothesis that the taxation of capital is generally harmful. The size of the effect seems reasonable, suggesting that a ten percent increase in the corporate tax rate (broadly defined, as noted above) may lead to a modest 0.0113 percentage point decrease in the average annual growth rate. Furthermore, its inclusion in the analysis does not appear to exert undue influence on the other coefficient estimates.

Table 3: OLS Growth Regressions

(1) (2) (3) (4) (5)

OLS OLS OLS OLS OLS

Log Initial Output -0.968*** -1.046*** -1.290*** -1.126*** -1.009*** (0.158) (0.153) (0.169) (0.217) (0.158) Investment Share 0.0695** 0.0877*** 0.0625** 0.0681** 0.0676**

(0.0285) (0.0301) (0.0245) (0.0279) (0.0288) Tert. Enrol. Rate 0.0106 0.0126 0.0164 0.0129 0.0123

(0.0109) (0.0106) (0.0105) (0.0109) (0.0114) Pop. Growth Rate -0.295* -0.410** -0.0699 -0.327* -0.312*

(0.175) (0.185) (0.181) (0.179) (0.183) Corporate Tax Rate -0.0113** -0.00746 -0.0309** -0.0278 (0.00505) (0.0128) (0.0154) (0.0173)

Low Inc. Dummy 1 2.304

(1.671)

Low Inc. Dummy 2 -1.241

(1.077)

Low Inc. Dummy 3 -0.675

(1.052) Low Inc. Dummy -0.111*** 0.0193 0.0252

×Corporate Tax Rate (0.0390) (0.0261) (0.0252)

Observations 115 130 115 115 115

R-squared 0.35620 0.368 0.484 0.37543 0.377 Log Likelihood -227.65900 -260.60182 -214.90730 -225.91544 -225.75788

Robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. All variables, except initial output, are averages over the period 2005-2013.

sum of the coefficients of the corporate tax rate and the interaction term is significantly negative. This suggests that in more developed countries there is an insignificant effect of capital taxation on growth, whereas the effect in the least developed countries is negative.

In columns four and five a higher threshold is chosen for the low income dummy. Using a higher threshold causes the dummy interaction term to lose all of its significance and even to change signs. This suggests that the observed effect is driven by the poorest countries in the sample.

The conclusions from this simple regression exercise are thus as follows. On the whole, the taxation of corporations broadly seems to reduce the growth rate of output per capita. This is generally in agreement with standard eco-nomic theory, which postulates the strong negative effects of the taxation of capital. Looking closer at the data, however, one finds that this negative effect on growth is only observed in the least developed countries, being those in the World Bank’s low income category corresponding with an income level below $1045. This is in line with the notion that the same government policy may have different effects in different countries (Rodrik 2006). The low income dummy used in the remainder of this paper is the one based on this threshold of $1045.

5

Alternative Specifications

This section considers a number of alternative specifications in order to verify the robustness of the key result discussed above. First of all, I reconsider the OLS estimation results by using a different time period for most of the variables, and by including a number of additional control variables that are typically linked to economic growth. Second, I apply a number of spatial econometric techniques to allow for possible spatial dependencies in the data.

5.1 OLS Robustness Checks

Table 4 reports a number of additional robustness checks on the basis of the OLS estimation results reported above in table 3.2 In order to avoid any undue influence from the global financial crisis and the post-crisis years it would be interesting to analyse a different time period. If the corporate tax rate only changes slowly over time, then one reasonable approach would be to use the rate of taxation in the period 2005-2013 as a proxy for that in earlier periods. The first column of table 4 reports the results for the period

2000-2009. All variables are now taken as the average over this time period, where the exception is the rate of corporate taxation which is over the period 2005-2013. The actual coefficient estimates for all variables are very close to those previously observed and the tertiary enrolment rate is now found to be significant at the 1% level, whereas the population growth rate is no longer significant. Most importantly, however, the dummy interaction term is still marginally significant at the 10% level. Even though the rate of taxation over 2005-2013 is clearly an imperfect measure for the taxation rate over 2000-2009, the coefficient size and significance are still close to those seen in the other regressions. Using an earlier time period than 2000-2009, however, does not yield any results. It is important to note that the rate of corporate taxation does change over time, and as such is only poorly captured by the tax rate in a later period.

Columns 2 through 7 of table 4 focus again on the initial time period 2005-2013 and examine the effects of other variables typically linked to economic growth. The first two variables capture the effect of macroeconomic stability and are the unemployment and inflation rates. As columns 2 and 3 show, however, neither the unemployment rate nor the inflation rate is found to sig-nificantly affect economic growth in this period. Furthermore, their inclusion is not found to impact the other coefficient estimates much. The population growth rate loses its marginal significance that it initially had in both cases, and when inflation is included the tertiary enrolment rate gains a single star. The interaction term is still highly significant and of similar size as before.

Column 4 then includes the total amount of taxation collected by the government, as a share of GDP. Bergh and Karlsson (2010) summarize the literature on this topic, and note that by and large a negative relationship is obtained between the economic growth rate and the amount of revenue collected by the government. If the size of the government is related to any of the independent variables as well, then not including it would induce omitted variable bias. Given the ability of a government to levy certain taxes can vary with its size, due to scale effects, it could be important to control for it here (Rothstein 2001). The results indicate, however, that the coefficient for government revenue is not significant and that it does not affect the other estimates much.

As can be seen in column 5 of table 4, however, the index is not found to be statistically significant in this analysis. Moreover, its effect on the coefficient estimates is modest. In this case, however, the dummy interaction term does lose some significance but is still significant at the 5% level.

The last variable considered is the Polity score of the Polity IV project, which is a measure of whether a country is a democracy or autocracy. A sub-stantial body of research has investigated whether autocracies or democracies tend to grow faster, and whether there might be other economic differences between them (e.g. Tavares and Wacziarg 2001). It is not unimaginable that democracies and autocracies have different tax policies, and as such not in-cluding this variable could again cause omitted variable bias. That does not appear to be the case here, as the results in columnn 6 are not much different from those observed before.

To be entirely thorough, the last column of table 4 includes all of these additional control variables. This too is not found to affect the results much, though the dummy interaction term’s coefficient is attenuated a bit and loses some significance. On the whole, however, the specifications considered in this section seem to indicate that the results are reasonably robust. Capital taxation does indeed seem to have a different effect on growth in the least developed countries, as indicated by the consistently significantly negative dummy interaction term found in all of considered regression specifications.

5.2 Spatial Analysis

A standard analysis, such as that done above, implicitly assumes that the units of analysis are unrelated objects. In reality, however, the countries in question have geographical and economical relationships with one another, so that what happens in one country is likely to affect what happens in another. If such spatial effects are present in the data, then not including them means risking omitted variable bias. To account explicitly for this possibility, this section reports the results of various spatial estimation methods.

Table 5: Spatial Growth Regressions

(1) (2) (3)

SAR SLX SDM

Key Variables

Log Initial Output -1.056*** -1.000*** -0.976*** (0.172) (0.173) (0.171) Investment Share 0.0602*** 0.0522* 0.0529**

(0.0204) (0.0269) (0.0206) Tert. Enrol. Rate 0.0110 0.00859 0.00650 (0.0103) (0.0114) (0.0110) Pop. Growth Rate -0.0411 0.165 0.158

(0.176) (0.279) (0.207) Corporate Tax Rate -0.00506 -0.00299 -0.00160

(0.0128) (0.0110) (0.0130) Low Inc. Dummy 2.095 1.644 1.703

(1.581) (1.621) (1.547) Low Inc. Dummy -0.104*** -0.0971*** -0.0964***

×Corporate Tax Rate (0.0375) (0.0369) (0.0366) Spatial Variables

Output Growth×W 0.335*** 0.130 (2.623) (0.139) Log Initial Output×W -1.1129*** -0.8765**

(0.3119) (0.3746) Investment Share×W 0.0270 0.0183

(0.0446) (0.0422) Tert. Enrol. Rate×W 0.0439** 0.0371** (0.0174) (0.0182) Pop. Growth Rate×W -0.460 -0.418

(0.427) (0.318) Corporate Tax Rate×W -0.0295 -0.0264 (0.0275) (0.0248)

Observations 115 115 115

Log likelihood -209.126 -204.962 -204.536 Standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. All variables, except initial output, are averages over the period

way, however, OLS is biased and a maximum likelihood estimation technique must be used (see e.g. LeSage and Pace 2009). The third possibility is spatial dependency in the error terms, meaning that random shocks in one country could affect the dependent variable in related countries. This case, however, is less relevant, as ignoring it does not give way to potential biases. If the true data generating process includes a spatial dependency in the error term, then standard estimation methods are still unbiased and consistent, and only lack efficiency (Greene 2005). As such, this type of spatial dependency is ignored in this paper.

A spatial analysis is conducted by including spatial lags of a variable in the regression. Such a lag of a variable is effectively a weighted average of the value of that variable in related countries. For example, for the Netherlands the spatially lagged GDP growth rate may be the average of the GDP growth rates of Belgium, Germany, and the United Kingdom. This weighted average is created by means of a spatial weighting matrix, which specifies explicitly how regions or countries are related to one another. Problematically, however, there is no clear consensus in the literature on how such a spatial weighting matrix should be specified exactly and as such various methods exist (Leenders 2002). The main approach taken in this paper is to use a so called inverse distance matrix, where the weights are given by the inverse of the geographical distance between countries. This matrix is created on the basis of computations of the distance between the average coordinates of countries. By normalizing the values in this matrix over each row a spatial weighting matrix is created that is in line with the standard spatial econometric literature (LeSage and Pace 2009, Elhorst 2010).

The results of the spatial estimation exercise are reported in table 5. This table presents three different types of spatial estimation models. The first two of these account for the two main types of spatial dependencies discussed above, and the third accounts for both at the same time. The first column reports the results for the model that includes a spatially lagged dependent variable (also known as the Spatial Autoregressive model, or SAR). Examining first the key variables, in the sense that these were also present in the previ-ous analysis, it is clear that they are very similar to the estimates obtained before. All estimates are a bit more attenuated, yet they tend to become more significant. The spatially lagged output growth variable appears to be highly significant, suggesting that growth in one country indeed depends on the growth rate of its neighbours.

insignificant. The spatial part of the regression shows that only spatial lags of initial output and the tertiary enrolment rate are significant. Particularly the latter is interesting, as it suggests that education can have potential spill over effects across borders. The analysis also suggests that the degree of capital taxation in neighbouring countries does not directly affect a country’s growth rate, as it might in case of international tax competition. The analysis can also be conducted employing spatial lags of only a subset of variables, for example those for which there is clear reason to expect an effect. This does not, however, meaningfully impact the results and as such is not reported here.

The third and last column reports the results for a model that combines the first two, as it includes both spatially lagged dependent and independent variables (called the Spatial Durbin Model, or SDM). The estimates for the key variables are again very similar to those obtained before. The spatial part of the regression looks very similar to those of the second column. However, the spatially lagged output growth rate is not found to be significant when included alongside with lags of the independent variables. At the same time, however, the independent spatial variables are not jointly significant either. A simple testing procedure as suggested by LeSage and Place (2009) and Elhorst (2010) indicates that the model with the spatially lagged output growth variable is most likely the relevant one. One implication of this model is that all variables are effectively allowed to have indirect spill over effects. For example, the level of investment in one country affects that country’s growth rate, which in turn affects the growth rate of its neighbours.

5.2.1 Alternative Spatial Weighting Matrix

Given the fact that there is no clear consensus on what the actual preferred spatial weighting matrix is, the standard approach in the spatial econometric literature is to employ various weighting matrices. As noted above, the SAR model, being the one with a spatially lagged dependent variable, is the preferred specification. For this specification table 5 reports the estimation results for a number of different weighting matrices.3 Instead of using an inverse distance matrix, here a so called “K nearest neighbours matrix” is used. While under the inverse distance matrix a country is linked to all other countries, with the strength of this link being being larger if the distance is small, under a nearest neighbour matrix a country is only linked to a relatively small number of other countries. Specifically, each country is related to a fixed number of neighbours that are close by. With three neighbours, for example, each country has a link with the three countries that are closest geographically and no link with all the other countries. Moreover, all countries to which

Table 6: Spatial Growth Regressions: K Nearest Neighbours Matrix

(1) (2) (3) (4)

SAR SAR SAR SAR

Key Variables

Log Initial Output -1.078*** -1.048*** -1.092*** -1.104*** (0.177) (0.175) (0.175) (0.172) Investment Share 0.0608*** 0.0554*** 0.0548*** 0.0547***

(0.0208) (0.0206) (0.0209) (0.0209) Tert. Enrol. Rate 0.0114 0.0114 0.0116 0.0120

(0.0105) (0.0104) (0.0105) (0.0105) Pop. Growth Rate -0.0247 -0.0317 -0.0782 -0.0709

(0.180) (0.177) (0.179) (0.179) Corporate Tax Rate -0.00323 -0.00400 -0.00376 -0.00234

(0.0131) (0.0129) (0.0131) (0.0131) Low Inc. Dummy 2.497 2.283 2.264 1.979

(1.611) (1.589) (1.611) (1.611) Low Inc. Dummy -0.115*** -0.108*** -0.107*** -0.102***

×Corporate Tax Rate (0.0381) (0.0376) (0.0382) (0.0382) Spatial Variables Output Growth×W 0.264*** 0.352*** 0.343*** 0.362*** (0.0886) (0.101) (0.113) (0.114) Observations 115 115 115 115 Log likelihood -210.83425 -209.53128 -210.75724 -210.35386 Number of Neighbours 3 5 7 10

Standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. All variables, except initial output, are averages over the period

a country is linked in this fashion are given an equal weight. Intuitively, a weighting matrix created in this way simply creates an average value of a variable in a country’s neighbours. If three neighbours are used, then a spatially lagged variable would, for example, for the Netherlands contain the (unweighted) average of that variable for Belgium, Germany, and the United Kingdom.

Reported in table 6 are the estimation results of the model with only a spatially lagged dependent variable for four different values of K, the num-ber of neighbours, that are in line with values typically used in the spatial econometric literature (LeSage and Pace 2009). What is immediately clear from this table is that the results are virtually unchanged compared to those of the inverse distance matrix. As such, the results are robust to a different specification of the weighting matrix and the conclusion reached is the same in all cases.

In summary, this section has conducted a spatial econometric growth analysis. Some forms of spatial econometric effects are found to be significant, and as such can possibly cause omitted variable bias if excluded from the analysis. In this case, however, these spatial effects do not seem to influence the results overly much. Allowing for relations in either, or both, the independent and dependent variables across countries is not found to change the conclusion of the simple OLS approach. As such this section supports the conclusion reached above, which is that capital taxation seems to have a different effect in the least developed countries vis-à-vis the rest of the world. Specifically, the effect is clearly negative in these poorest countries, whereas it is insignificant for the others.

6

Explaining Country Heterogeneity

The analysis thus far largely indicates that there is indeed country heterogene-ity in the effect that capital taxation has on economic growth. In the least developed countries the taxation of capital seems to affect growth negatively, but not significantly in the result of the world. This effect appears to be quite robust across various specifications. This section seeks to offer a number of empirical explanations for this finding. The main approach used is to interact the corporate taxation measure with various conditioning variables and to analyse the resulting marginal effects graphically.

Table 7: OLS Conditioning Variables: GDP and Institutional Indicators

(1) (2) (3) (4) (5)

OLS OLS OLS OLS OLS

Key Variables

Log Initial Output -0.893** -0.986*** -0.862*** -1.010*** -0.579** (0.415) (0.172) (0.207) (0.191) (0.277) Investment Share 0.0695*** 0.0598** 0.150*** 0.0617** 0.141***

(0.0246) (0.0254) (0.0359) (0.0263) (0.0415) Pop. Growth Rate -0.328 -0.416* -0.328 -0.323 -0.132

(0.207) (0.216) (0.209) (0.210) (0.239) Tert. Enrol. Rate 0.0118 0.0120 -0.00277 0.0114 0.00498 (0.0123) (0.0129) (0.0146) (0.0134) (0.0156) Corporate Tax Rate -0.00113 -0.0276 0.0621 -0.000759 -0.00341 (0.0844) (0.0200) (0.112) (0.0443) (0.0596) Conditioning Variables

Polity Score -0.124

(0.109)

Economic Freedom Index 0.771 (0.786) Government Revenue 0.0272 (0.0595) Quality of Government -1.340 (4.339) Interaction Term1 -0.00252 0.00208 -0.0117 -0.000545 -0.0127 (0.0101) (0.00269) (0.0165) (0.00132) (0.0968) Observations 115 102 98 112 93 R-squared 0.370 0.366 0.436 0.35444 0.38300 Log Likelihood -226.416 -219.455 -205.817 -218.931 -175.197

1_{The interaction term is the product of the corporate tax rate the conditioning}

variable. For model (1) the conditioning variable is Log Initial Output. Robust standard errors in parentheses. *** p<0.01, ** p<0.05,* p<0.1 All variables,

variable. The significance of estimated coefficient on this interaction term is an indication of whether the variable works as a conditioning variable. As noted by Brambor et al. (2006) a proper examination of the marginal effects can be done by means of a graphical analysis. Moreover, these authors recommend the inclusion of all constituent terms in the analysis, meaning that if an interaction term is included, so too should both measures that are interacted.

Table 7 shows the results for a simple regression analysis that includes a number of possible conditioning variables. In the first column no additional conditioning variable is introduced, but rather an interaction is created be-tween the corporate tax rate and the initial output variable. This allows the effect of corporate taxation to vary with the (initial) amount of GDP per capita. The results indicate that this measure does not seem to adequately explain cross-country differences, as the interaction term is not significant. A proper analysis of the marginal effect can be done graphically. Figure 1 shows a marginal effect plot for each of the columns of table 7. In all cases a number of objects are plotted in each graph. Most importantly the solid line is the point estimate of the marginal effect for different values of the conditioning variable. The dotted lines denote a 95% confidence interval. Additionally, a histogram is also plotted which shows the distribution of the conditioning variable on the right hand axis. Lastly, a number of countries are marked in the graph to provide an intuition as to what type of country can be associated with a certain value of the conditoning variable.

The first graph thus shows that the effect of corporate taxation on GDP growth is negative for the entire range of initial output. Moreover, the poor fit observed in the table can also be seen in this graph as the 95% confidence interval includes zero.

HUN LAO NER _NGA PRY _RUS SVN THA GBR -.1 0 .1

Marginal Effect of Corporate Taxation on Real GDP Growth 0

5

10

15

Percentage (Histogram)

5 6 7 8 9 10 11

Log GDP Per Capita

DZA BLZ BEN FRA LAO MAR NER RWA SAU ZWE -.1 -.05 0 .05

Marginal Effect of Corporate Taxation on Real GDP Growth 0

10 20 30 Percentage (Histogram) -10 -5 0 5 10 Polity Score COG FIN MEX NER PAK POL SWE CHE -.1 0 .1

Marginal Effect of Corporate Taxation on Real GDP Growth 0

10

20

Percentage (Histogram)

4 5 6 7 8 9

Economic Freedom Index

AUT BEN CAN NGA PAK SWE CHE -.1 -.05 0 .05

Marginal Effect of Corporate Taxation on Real GDP Growth 0

10 20 Percentage (Histogram) 10 20 30 40 50 60 Government Revenue, % of GDP BRA CIV FRA KEN MEX MAR

NLD SVN GBR

-.1

0

.1

Marginal Effect of Corporate Taxation on Real GDP Growth 0

10

20

30

Another measure that may capture institutional quality is the economic freedom index of the Fraser Institute, which was also used above. This mea-sure captures very broadly the “rules of the game,” in the sense that it seeks to measure to which extent citizens are free to pursue their own economic interests. The index ranges from 0 to 10, although the extreme values are not observed in this sample of countries. Column 3 reports the results with this interaction included, which is again found to be insignificant. The third graph in figure 1 suggests that under higher levels of economic freedom, capital taxation is more harmful. Here again, however, the effect is not at all sta-tistically significant and so this measure does not seem to properly explain cross-country differences.

The theoretical model of Aghion et al. (2013) suggests that capital taxation might have the standard negative effect in countries with a small government, whereas the effect could be positive when the government is large. In this analysis, however, no evidence is found in favour of this proposition. In column 4 of table 7 an interaction is introduced with the amount of government revenue, as a percentage of GDP, which is a measure of government size. The results indicate that this variable also does not properly explain the observed country heterogeneity. Moreover, the interaction term is actually negative, rather than the positive sign that would be expected. The corresponding graph in figure 1 also clearly shows that the marginal effect falls with government size, although it is never statistically different from zero.

Instead of the size of government the relevant factor may be its quality. The International Country Risk Guide (ICRG) publishes a measure that seeks to quantify the quality of government, and is based on measures of corruption, law and order and bureaucratic quality. The measure ranges from zero to one. As is clear from the last column of table 7, however, the interaction term is not significant. The last graph in figure 1 shows a rather flat line, suggesting that indeed the effect of capital taxation on growth does not vary much with the ICRG’s measure.

The above results seem to indicate that standard measures of institutional quality and economic development do a poor job of explaining the cross-country differences observed in the data. A number of additional indicators are considered in table 8. These variables are selected to capture the difference between the least developed countries and those slightly more developed. As such, they mostly relate to how “modern” and technologically advanced a country is. In addition, the marginal effects are plotted in figure 2.

very large, as they are very dependent on (primarily) agriculture. As the first column of table 8 shows, when this variable is introduced to the regression it creates a (marginally) significant interaction effect. The marginal effect is plotted in the first graph in figure 2. Here it is clearly seen that for countries that do not depend overly much on agriculture the marginal effect of corporate taxation is around zero. When the share of agriculture, hunting, forestry and fishing exceeds 15% or so, the effect becomes significantly negative.

Table 8: OLS Conditioning Variables: Other Indicators

(1) (2) (3) (4)

OLS OLS OLS OLS

Key Variables

Log Initial Output -1.483*** -0.995*** -0.0116*** -1.229*** (0.262) (0.173) (0.00195) (0.232) Investment Share 0.0775** 0.0769*** 0.00162*** 0.0843***

(0.0319) (0.0254) (0.000298) (0.0244) Pop. Growth Rate -0.221 -0.264 -0.154 -0.375*

(0.210) (0.216) (0.112) (0.211) Tert. Enrol. Rate 0.0194 0.0117 -7.40e-05 0.00753 (0.0121) (0.0127) (0.000112) (0.0139) Corporate Tax Rate 0.00382 -0.0730** -0.00161*** -0.0128**

(0.0161) (0.0287) (0.000493) (0.00555) Conditioning Variables AHFF1_{Share in GDP 0.00244} (0.0511) Manufacturing Share in GDP -0.158* (0.0867)

Log Electricity Usage -0.00430 (0.00290) Internet Usage 0.00669 (0.0209) Interaction Term2 -0.00187* 0.00434** 0.000206*** 0.000265 (0.000970) (0.00193) (6.70e-05) (0.000382) Observations 112 112 115 111 R-squared 0.402 0.40454 0.54638 0.36338 Log Likelihood -236.89169 -228.28173 314.48612 -252.26754

1_{Agriculture, Hunting, Forestry, and Fishing.} 2_{The interaction term is the}

product of the corporate tax rate the conditioning variable. Robust standard errors in parentheses. *** p<0.01, ** p<0.05,* p<0.1 All variables, except

BLZ BEN BGR ETH NER PAK TZA USA ZWE -.15 -.1 -.05 0 .05

Marginal Effect of Corporate Taxation on Real GDP Growth 0

10 20 30 40 Percent 0 20 40 60

GDP Share of Agriculture, Hunting, Forrestry and Fishing

CZE GNQ FIN DEU MWI PAK RUS UGA VUT -.1 0 .1 .2

Marginal Effect of Corporate Taxation on Real GDP Growth 0

10 20 Percentage (Histogram) 0 10 20 30 Manufacturing Share in GDP IDN CHE TUR BEN BGR COG ETH GHA NOR PRY -.1 -.05 0 .05 .1

Marginal Effect of Corporate Taxation on Real GDP Growth 0

10

20

Percentage (Histogram)

4 5 6 7 8 9 10 11

Log Electricity Use per Capita (kwh)

BRA COG HRV CZE ETH FRA MEX NGA NOR UGA USA -.1 0 .1

Marginal Effect of Corporate Taxation on Real GDP Growth 0

10

20

30

Percentage (Histogram)

0 20 40 60 80 100

The last measure considered here and that also captures modernity is internet usage. The fourth and last column of table 8 creates an interaction with the share of a country’s population that uses the internet. Examining the regression results, the interaction is not found to be significant. The graph, however, indicates that the marginal effect is significant at some point. For those countries where the share of internet users is really low, below the 5% mark, the effect of corporate taxation is negative. The effect then slowly in-creases and becomes insignificant as its confidence interval simultaneously widens. The fit, however, is not as good as for the other technological indica-tors considered here, which may be because the least developed countries all have very similar values on this measure.

All in all there are twelve countries for which a significantly negative effect is obtained for each of the four conditioning variables considered in table 8. These countries are Benin, Burkina Faso, Ethiopia, Ghana, Lao, Mauritania, Nepal, Niger, Rwanda, Tanzania, Uganda, and Vanuatu. For these countries it seems particularly likely that the taxation of capital reduces the growth of GDP per capita.

In summary, this section has considered which variables might explain the observed country heterogeneity in the effect of capital taxation on economic growth. Standard measures of institutional quality, such as the size and quality of the government, or the Polity score perform poorly in this regard. Moreover, the heterogeneity cannot be explained solely by differences in GDP per capita. Instead, measures that capture in some sense how technologically advanced a country is seem to work much better. In countries with a high dependence on agriculture, a low dependence on manufacturing, low electricity usage and few internet users the effect of capital taxation on economic growth is more likely to be negative.

7

Discussion

where agriculture, hunting, forestry and fishing make up a large part of GDP, the effect of capital taxation also seems to be mostly negative. The same is true for countries where manufacturing is only a very small part of GDP.

This is in line with the conclusion of a theoretical paper by Ten Kate (2014), which combines insights from Aghion et al. (2013) and Acemoglu et al. (2006). In this model firms in a country either copy foreign technologies (adaptation) or create new technologies themselves (innovation). Which approach is more feasible depends on how close a country operates to the technological frontier in the sense of Acemoglu et al. (2006). When a country is close to the technolog-ical frontier, innovation is more likely to be profitable, as little is to be gained by copying existing technologies (which are very similar to the technologies already available to the firm). Conversely, for less technologically developed countries adaptation becomes more attractive compared to innovation.

When countries are close to the technological frontier the model continues as in Aghion et al. (2013) and technological progress, and thus economic growth, is the result from R&D driven innovation. As shown in Aghion et al. this gives way to a scenario in which the effect of capital taxation is in principle indeterminate, as it depends on other factors, primarily the size of the government. For technologically advanced countries the taxation of capital can thus have either a positive or a negative effect, which may explain the general null result observed in this paper for these countries.

If, however, countries operate far from the technological frontier, then technological progress results from firms copying foreign technologies. As such, technological progress is exogenously determined (from the perspective of a country that employs an adaptation strategy) and a simple neoclassical model with exogenous growth is obtained as in Solow (1956). For such a model the standard Chamley (1986) and Judd (1985) result applies and the taxation of capital is thus found to harm economic growth. As such, a country that is technologically backwards is predicted to suffer negative growth effects from the taxation of capital, which is also what is observed in the analysis conducted in this paper.

More technologically advanced countries are thus more likely to have technological progress based on innovation. As in Aghion et al. (2013) growth is thus endogenous and the effect of capital taxation is indeterminate. For less technologically advanced countries technological progress seems more likely to be exogenous, as it depends on the speed at which the technological frontier is advancing. In some sense these less technologically advanced countries can be said to have economies that are better described by neoclassical models. The “more neoclassical” an economy is, the more likely it is that capital taxation

8

Conclusion

This paper has examined the impact that capital taxation has on economic growth for a large sample of 115 countries in recent years. The main finding is that the taxation of capital seems to be particularly harmful for economic growth for the least developed countries. For countries that do not list among the poorest quartile, however, no significant effect of capital taxation on eco-nomic growth was observed, which include a large number of lesser developed countries. This finding is obtained under a substantial number of alternative specifications. The inclusion of additional control variables or controling for spatial dependencies in the data does not generally impact the results.

This finding is particularly interesting because it may explain why em-pirical research on this topic has found some mixed results. Authors that have studied mostly developed countries (Mendoza et al. 1994, Widmalm 2001,Gemmel et al. 2013) typically did not find a negative effect of capital taxation on economic growth. For the world as a whole, however, a negative effect was observed in Lee and Gordon (2005). A simple explanation based on the findings of this paper is that the effect is simply different in developed vis-à-vis developing countries. When no effort is made to differentiate be-tween these two groups a negative effect of capital taxation on growth is also observed in this paper. This effect, however, is found to be driven by the least developed countries in the sample.

From a theoretical perspective it may be the case that these least developed countries most resemble a neoclassical economy, where technological progress is exogenous. Standard macroeconomic results would thus apply to these countries, and the optimal tax rate on capital would be zero. On the other hand when countries are more technologically advanced, technological progress may be innovation driven and the effect of capital taxation is indeterminate as in Aghion et al. (2013). This argument is presented more formally in Ten Kate (2014), which combines insights from Aghion et al. (2013) and Acemoglu et al. (2006) to suggest a negative growth effect of capital taxation in lesser developed countries and an indeterminate effect in more developed ones.

are also found to be harmed by capital taxation, whereas this is not the case for countries with moderate to high shares of manufacturing in GDP. The variable that seems to best describe cross-country differences is electricity usage (in per capita terms). For countries with very low electricity consumption, such as for example Haiti or Ethiopia, capital taxation is found to significantly reduce economic growth. When electricity consumption is more moderate, as it is in Mexico or Turkey, the effect is insignificantly different from zero. For countries with very high levels of electricity consumption on the other hand, such as Kuwait or Iceland, the effect of capital taxation even becomes positive, and significantly so. Insofar as these variables capture how technologically advanced a country is, this offers some support for the argument presented in Ten Kate (2014).

References

1. Aghion, Philippe, Ufuk Akcigit, and Jesús Fernández-Villaverde. Optimal Capital versus Labor Taxation with Innovation-Led Growth. NBER Working Paper Series, 2013: http://www.nber.org/papers/w19086.

2. Arnold, Jens Matthias, Bert Brys, Christopher Heady, Asa Johansson, Cyrille Schwellnus and Laura Vartia. Tax Policy for Economic Recovery and Growth. The Economic Journal, 2011: F59-F80.

3. Atkeson, Andrews, Varadarajan V. Chari, and Patrick Kehoe. Taxing Capital Income: A Bad Idea. Federal Reserve Bank of Minneapolis Quarterly Review, 1999: 3-17.

4. Barro, Robert J. Economic Growth in a Cross Section of Countries. Quarterly Journal of Economics, 1991: 407-443.

5. Bergh, Andreas, and Martin Karlsson. Government Size and Growth: Account-ing for Economic Freedom and Globalization. Public Choice, 2010: 195-213. 6. Brambor, Thomas, William Roberts Clark, and Matt Golder. Understanding

Interaction Models: Improving Empirical Analysis. Political Analysis, 2006: 63-82.

7. Chamley, Christophe. Optimal Taxation of Capital Income in General Equilib-rium with Infinite Lives. Econometrica, 1986: 607-622.

8. Dahlberg, Stefan, Sören Holmberg, Bo Rothstein, Felix Hartmann, and Richard Svensson. The Quality of Government Dataset, version Jan15. Uni-versity of Gothenburg: The Quality of Government Institute, 2015: http://www.qog.pol.gu.se

10. Ferguson, Ross, William Wilkinson and Robert Hill. Electricity Use and Economic Development. Energy Policy, 2000: 923-934.

11. Gemmel, Norman, Richard Kneller and Ismael Sanz. The Growth Effects of Tax Rates in the OECD. Working Papers in Public Finance, Victoria Business School, 2013.

12. Greene, William H. Econometric Analysis, 6th edition. Upper Saddle River, Pearson Prentice Hall, 2005.

13. Gwartney, James, and Robert Lawson. The Concept and Measurement of Economic Freedom. European Journal of Political Economy, 2003: 405-430. 14. Judd, Kenneth L. Redistributive Taxation in a Simple Perfect Foresight Model.

Journal of Public Economics, 1985: 59-83.

15. Lee, Young, and Roger H. Gordon. Tax Structure and Economic Growth. Journal of Public Economics, 2005: 1027-1043.

16. Leenders, Rogers T.A.J. Modelling Social Influence Through Network Autocor-relation: Constructing the Weight Matrix. Social Networks, 2002: 21-47. 17. LeSage, James P. and Robert K. Pace. Introduction to Spatial Econometrics.

Boca Raton, Taylor and Francis, 2009.

18. Mankiw, N. Gregory, Matthew Charles Weinzierl, and Danny Ferris Yagan. Optimal Taxation in Theory and Practice. Journal of Economic Perspectives, 2009: 147-174.

19. Mauro, Paolo. Corruption and Growth. The Quarterly Journal of Economics, 1995: 681-712.

20. Mendoza, Enrique G., Assaf Razin, and Linda L. Tesar. Effective Tax Rates in Macroeconomics Cross-Country Estimates of Tax Rates on Factor Incomes and Consumption. Journal of Monetary Economics, 1994: 297-323.

21. OECD. Tax Burdens: Alternative Measures. OECD Tax Policy Studies 2, 2000. 22. Rodrik, Dani. Growth Strategies. In Handbook of Economic Growth,

Vol-ume 1A, by Philippe Aghion and Steven H. Durlauf. Amsterdam, 2006. 23. Rothstein, Bo. Trust, Social Dilemmas and Collective Memories. Journal of

Theoretical Politics, 2001: 477-501.

24. Salanié, Bernard The Economics of Taxation. MIT Press, 2003.

25. Solow, Robert M. A Contribution to the Theory of Economic Growth. The Quarterly Journal of Economics, 1956: 65-94.

26. Tavares, José, and Romain Wacziarg. How Democracy Affects Growth. Euro-pean Economic Review, 2001: 1341-1378.

28. Ten Kate, Fabian, and Petros Milionis. Does Capital Taxation Harm Economic Growth? Working paper, 2014. Available on request.

29. The World Bank. World Development Indicatators. 2015: http://data.worldbank.org/

30. Vega, Solmaria Halleck and Paul Elhorst. The SLX Model. Working paper. 31. Widmalm, Frida. Tax Structure and Growth: Are Some Taxes Better Than

Appendix

Additional Variables and Data Sources

Besides the main variables described in the data section a number of additional variables are used as part of the robustness checks and as conditioning vari-ables. These variables were all taken from a dataset created by the Quality of Government Institute (Dahlberg et al. 2015), and which in turn originate from various other sources. All variables discussed in this section have their key statistics summarized in table 9. As is clear from the number of observations in this table, not every measure is available for all countries. This in turn limits the number of observations that can be used in the corresponding regression.

The robustness analysis introduced a number of additional control vari-ables to the analysis whose sources are described here. The unemployment rate variable is from the same source as most of the data used in this research, being the World Bank’s development indicators. Both the inflation rate and the government revenue variables are from the International Monetary Fund. Government revenue is expressed as a percentage of GDP. The Economic Freedom index is from the Fraser Institute and is created on the basis of a substantial number of indicators that seek to capture whether citizens are free to pursue their own economic interests. The index ranges from a theoretical 0 (no freedom) to 10 (perfect freedom), yet such extreme score are not typically observed. For this sample all countries score between a 3.7 and an 8.4. The Polity index is from the Polity IV project and is a measure of whether a country is more democratic or more autocratic. A perfect democracy has a score of 10, whereas a perfect autocracy has a -10. Most developed countries score a 10 on this measure.

Table A1: Summary Statistics Additional Variables

Variable Obs Mean Std. Dev. Min Max

Unemployment Rate 107 7.682 5.427 0.800 33.338 Inflation Rate 110 1.464 0.996 0.053 7.444 Government Revenue 112 30.659 11.493 8.300 57.169 Economic Freedom Index 98 6.899 0.811 3.707 8.410

Polity Score 102 4.999 6.102 -10 10

Quality of Government Index 93 0.794 1.595 0.2448 1 AHFF Share in GDP 112 12.352 11.299 0.644 46.970 Manufacturing Share in GDP 112 13.936 6.354 0.154 40.792 Log Electricity Use 115 7.486 1.674 0.223 10.670 Internet Users, % of Total Population 111 31.754 26.688 0.653 91.941

All variables are averages over the period 2005-2013.

Table A2: Country List

Afghanistan Georgia Nigeria

Albania Germany Norway

Angola Ghana Pakistan

Antigua and Barbuda Greece Paraguay

Armenia Grenada Peru

Australia Guatemala Philippines

Austria Honduras Poland

Azerbaijan Hong Kong, China Portugal Bangladesh Hungary Puerto Rico

Belgium Iceland Romania

Belize India Russian Federation

Benin Indonesia Rwanda

Bhutan Ireland Sao Tome and Principe Bosnia and Herzegovina Israel Saudi Arabia

Botswana Japan Senegal

Brazil Jordan Seychelles

Brunei Darussalam Kazakhstan Slovak Republic

Bulgaria Kenya Slovenia

Burkina Faso Korea, Rep. South Africa Cambodia Kyrgyz Republic Spain Cameroon Lao PDR Sri Lanka

Canada Lebanon St. Lucia

Chile Macedonia Sudan

Congo, Rep. Madagascar Swaziland

Costa Rica Malawi Sweden

Cote d’Ivoire Malaysia Switzerland Croatia Mauritius Tanzania Czech Republic Mexico Thailand

Denmark Moldova Tunisia

Dominican Republic Mongolia Turkey

Ecuador Montenegro Uganda

Egypt Morocco Ukraine

El Salvador Mozambique United Kingdom Equatorial Guinea Namibia United States

Estonia Nepal Uruguay

Ethiopia Netherlands Vanuatu Finland New Zealand Vietnam

France Niger Zimbabwe