DOES THE EFFECT OF CAPITAL TAXATION ON PRODUCTIVE INVESTMENT VARY WITH THE LEVEL OF STOCK MARKET SPECULATION?

(1)

Faculty of Economics and Business

DOES THE EFFECT OF CAPITAL TAXATION ON PRODUCTIVE INVESTMENT VARY WITH THE LEVEL OF STOCK MARKET SPECULATION?

Master’s Thesis – International Economics and Business

Jonathan Parisi S3808971

j.d.parisi@student.rug.nl

Date: 18/06/2019

Supervisor: prof. dr. D. J. (Dirk) Bezemer

Co-assessor: prof. dr. R. C. (Robert) Inklaar

(2)

Abstract

Using data from 14 OECD countries over the period 1950-2015, this paper analyzes the effects of changes in average effective capital taxation rates on investment efficiency (efficiency of real capital investment) during speculative stock market booms. We measure the investment efficiency during “boom” periods by evaluating total factor productivity (TFP) growth two years into the future, allowing time for investments to become productive. We find strong empirical evidence of a negative effect of stock market speculation on investment efficiency at low rates of effective capital taxation (below approximately 20%). We find limited evidence that during stock market booms, capital taxation provides a more beneficial effect to investment efficiency relative to non-boom periods, since the marginal effect of capital taxation on investment efficiency is less negative.

(3)

The role of this paper is to explore the connection between incentives in the tax structure and the efficiency of capital investment in different stock market environments. We seek to add a qualification to the large body of theory suggesting, with some empirical support, that capital taxation has a negative effect on economic growth. That is, we expect that the extent of this relationship may be dependent on the prevalence of speculative activity in the economy.

In particular, while patterns in the allocation of financial capital may negatively affect the efficiency of real capital investment during stock market booms, we hypothesize that this effect differs with the level of capital taxation. We hypothesize that since capital taxation reduces the return on financial investment and speculation, it can improve the efficiency of capital investment by directing financial capital back towards the real economy during speculative boom periods.

Our research question, then, is to determine whether the effect of capital income taxation on productive real investment varies with the level of stock market speculation prevailing in the economy in a given period. We are interested in the taxation of overall capital income, of which capital gains makes up one part alongside corporate profits, interest and dividends.

Although this study primarily focuses on the relationship between tax structure and efficient real investment, it is motivated in part by recent work on the increasing financialization of the advanced economies over the past several decades. While a strong positive association has historically been documented between the size of a country’s financial system and economic growth (Levine, 1997; Levine, Loayza and Beck, 2000; Valickova et al, 2015), this effect has virtually disappeared since the 1990’s (Bezemer et al, 2016; Beck et al, 2012). The latter researchers show the differential growth effects of “unproductive” loans to sectors driven by financial asset investment (to which lending has grown explosively) compared to “productive” loans to the private non-financial sector, which have remained largely constant as a share of GDP.

Rather than looking at bank credit specifically, in this study we focus on general flows of financial capital, whatever their source, within the economy, hypothesizing that increasing attraction of funds to financial securities markets during “boom” periods reduces the efficiency of investment in the real economy.

During periods not including one or more speculative “boom” years, our results demonstrate a negative average marginal effect of capital taxation on investment efficiency, which we measure through future total factor productivity growth. We also provide reasonably strong evidence that at low levels of capital taxation, the effect of higher stock market speculation on investment efficiency is negative, but this effect becomes indistinguishable from zero at effective capital tax rates higher than approximately 20%. Finally, we present limited evidence that during boom periods, the marginal effect of capital taxation on investment efficiency is less negative than in non-boom periods, suggesting a relatively beneficial impact of higher capital taxation during stock booms.

(5)

II. Literature Review

Capital allocation

The efficient allocation of capital has long been considered by economists as a requirement for optimal economic growth. According to Jorgenson & Yun (1986), “efficiency in the allocation of capital requires that the addition to wealth generated by one dollar’s worth of investment must be the same for all assets.” This follows from the idea that prices of assets adjust perfectly to their expected return. While we refer to real capital here when we reference capital allocation, in practice there exist considerable links to the financial system, given that real capital investments are very often financed. Indeed, Schumpeter (1911), argued for a role of the financial system in directing financial capital to entrepreneurs who would stimulate growth through innovation, leading to the process of creative destruction.

Along these lines, capital allocation has been studied by Rajan & Zingales (1996), Beck and Levine (2002) and Wurgler (2000) as a channel through which financial development affects growth. A positive relationship between financial development and growth had long been established dating back to Cameron (1967), Goldsmith (1969) and McKinnon (1973). However, the direction of causality remained ambiguous, with reverse causality highly plausible as well as bi-directional causality. The three well-regarded studies on capital allocation mentioned above attempted to address this causality problem. The authors focused on capital allocation as a channel and were able to show empirically that financial development causally affected the efficiency of capital investment.

Rajan and Zingales (1996) estimated the external dependence on finance at the industry level within 43 countries from 1980-1990, using data from listed U.S. firms (this was calculated as capital expenditures less cash flow from operations, divided by capital expenditures). They posited that this dependence would be the same for a given industry in all countries in the sample. For each country they calculated the difference in growth in value-added for more externally dependent industries relative to industries with average external dependence. Using the interaction between external dependence and country level financial development (credit-to-GDP and stock market size-to- GDP), they showed that more externally dependent industries grew faster relative to average industries in countries with higher levels of financial development. This was theorized to confirm the role of financial sector development in lowering the cost of external finance, thereby facilitating more efficient real capital allocation.

Wurgler (2000) follows a similar methodology. Within 28 industries in 65 countries over 1963-1995 he estimates the difference in gross fixed capital formation between industries with high growth of value-added and industries in which value added is declining. He refers to this measure as elasticity, which will be positive when capital allocation is more efficient (directed more toward growing industries). He finds that in countries with higher levels of financial development, this elasticity is larger than in countries with a less developed financial system. Beck and Levine (2002) built on this work in a study to evaluate whether bank-based or market-based systems foster greater efficiency of capital allocation. In one of their models, they use Wurgler’s elasticity of investment as their dependent variable. Their data covers 36 industries in 42 countries from 1980-1995. While the authors do not find evidence supporting the superiority of either type of system, they do confirm many conclusions of Rajan and Zingales and also that higher financial development predicts higher levels of Wurgler’s elasticity of investment.

(6)

This paper builds on this work, studying capital allocation from a different angle. Specifically, we look at the relationship between the taxation of financial capital and the allocation of real capital, paying particular attention to the role of stock price booms and taxation of the overall returns to stock ownership. This research fits within the literature on taxation and growth, introduced below, and evaluates real capital allocation as a transmission channel.

Taxation and growth

The literature on the aggregate level of taxation and its relationship to growth is rather inconclusive. Overall taxation is thought to influence growth mainly through channels of physical and human capital accumulation - in line with the neoclassical Solow model (1956), and through labor vs. leisure choices (Acosta-Ormaechea and Yoo, 2012). These effects are analyzed in several key studies using endogenous growth models (Barro, 1990; King and Rebelo, 1990;

and Jones et al, 1993). One recent study by Gemell et al (2014) confirms however that in advanced economies, rather than operating through capital accumulation, tax effects on growth operate through factor productivity. The impact of the progressivity of taxation has also been studied and related to risk-taking activity (Gentry and Hubbard, 2000) and human capital decisions (Heckman et al, 1998).

Empirically, some studies have found a negative relationship between aggregate government spending and growth (Barro 1991; Folster and Henrekson 2001; Kneller et al 1999). Others have found inconclusive results using various measures of tax rates or other related fiscal variables (Koester and Kormendi, 1989; Easterly and Rebelo, 1993;

Levine and Renelt, 1992; Slemrod, 1995; Mendoza 1997; Agell et al, 2006). The composition of spending funded by taxation also matters, of course. For example, Gemell et al (2014) found that in OECD countries, allocating spending to infrastructure and education was more growth-enhancing than spending on social welfare.

Tax structure and growth

More relevant to this study is the literature on tax structure and growth. To set the groundwork for a more detailed discussion, we briefly discuss the data landscape underlying the variety of studies that have been completed to date.

Due partly to data limitations, several different classifications of the tax structure have been utilized. Some studies have focused on contrasting corporate income taxes with personal income taxes (Lee and Gordon, 2005; Gordon, 1998), meaning that taxation on all forms of income (e.g. profits, labor, interest, dividends) received by each group was aggregated and compared. Others additionally included consumption tax and property tax as separate categories (Arnold, 2008; Widmalm 2001) alongside aggregate personal income tax and corporate income tax. Other researchers (such as Gemmell et al, 2011) have looked along the lines of the distortionary/non-distortionary breakdown initially developed by Bleaney, Gemmell, and Kneller (2001), who categorized all taxation outside of consumption tax (income and profit, social security, payroll and property) as distortionary. Finally, thanks to an approach developed by Mendoza et al (1994) and later modified by Carey and Rabesona (2003) and McDaniel (2007), another classification among capital, labor and consumption taxation has also been utilized (McDaniel, 2011;

Mendoza et al, 1997; and Gemmell et al, 2014), and this is the approach followed in this paper. This classification involves mapping each specific category of tax receipts (from OECD Revenue Statistics or from the national accounts) to its specific counterpart in income from the national accounts, and then categorizing each such tax among capital, labor and consumption (at times additional categories are also introduced).

(7)

The literature is also replete with a wide variety of measurement strategies for tax variables, including the use of average statutory rates, marginal statutory rates (for which income distribution data is necessary), the share of each tax in total tax revenue (or GDP), and average effective tax rates (AETR’s), which Mendoza, Razin and Tesar (hereafter “MRT”) first developed in their 1994 paper. These authors, building on earlier work by Lucas (1990, 1991) and Razin and Sadka (1993), provided a solution for the key problem of separating taxation on capital income reported at the personal level from that of other income such as labor earnings, making it possible to do so across a wide variety of countries with consistent and available data.

The intention to argue that tax structure may influence the efficiency of real investment motivates our choice of MRT- style average effective tax rates to measure taxation. The definitive split of capital and labor income among households is unique to this approach and is key for our hypothesized channel, through which we expect such influence to take place. Before discussing our own hypothesis in detail, we first provide a brief overview of the relevant findings to date relative to the various tax categorizations.

Theory and evidence on tax structure and capital taxation

Capital taxation, including corporate income tax, is thought to have substantial negative effects on capital accumulation, since it reduces the return to capital investments (Johansson et al, 2008). Several well-known studies have concluded that the long-term optimal rate of capital taxation is actually zero in the long run, perhaps the most widely cited being Chamley (1986) and Judd (1985). Both studies are theoretical and use dynamic general equilibrium models to explore optimal taxation. Mankiw (2009) surveys the optimal taxation literature and reiterates the conclusion that capital taxation should be close to zero under its recommendations. However, the arguments cited in Mankiw’s survey often appear to conflate taxation of financial capital (our interest in this paper) with taxation of real capital. For example, the argument is made that current taxation of capital is essentially a tax on future consumption, since capital goods are used to generate consumption in the future. In this paper, we attempt to maintain a clear distinction between financial investment in securities and what is considered real capital formation under economic theory.

With respect to empirical studies, clear evidence has been difficult to generate, largely due to persistent data challenges making it difficult to derive an estimate for capital income. Mendoza et al (1997) were the first (to the knowledge of the author) to employ MRT-style tax rates in a cross-country empirical study of tax structure. They expand the AETR’s from MRT 1994 to cover 18 OECD countries from 1965-1991 and test the effect of taxation on the investment rate and then, separately, long-term growth. They use data in five-year averages, as is seen frequently in the studies cited here, in order to account for short-term fluctuation and focus on the long-term. They aim to test Harberger’s “superneutrality conjecture”, that tax changes affect investment but have little impact on growth (which their results support). Capital and labor taxation are shown to significantly and negatively affect investment, while the coefficient on consumption tax is significant and positive. The panel regressions on long-run growth show all the three tax categories to be individually insignificant. However, Wald tests show that they are not jointly insignificant, as this can be rejected at the 1% level. The authors also note that there is some evidence that taxes are significant in a panel of annual data as opposed to five-year averages, although the authors attribute this to demonstrating transitional effects rather than long-term effects. The annual results come from an earlier CEPR-working paper version of the study (Mendoza et al, 1996a). The coefficients indeed appear very small (as characterized by the authors) and capital taxation is insignificant at the 5% level, with a t-statistic of -1.50.

(8)

Another highly relevant study is by Gemmell, Kneller and Sanz (2014), hereafter “GKS”, who use tax rate data from McDaniel (2007) to study the effect of tax structure on growth. The McDaniel data expands the original MRT rates for 15 OECD countries to cover the period 1950-2003. The GKS study looks at both macro-level growth effects using country data and also micro-level effects using marginal rates at the personal and corporate level. We are most interested in the macro study, which shows ambiguous results for capital taxation AETR’s, with the sign of the coefficient changing depending on the model specification. The paper features a noteworthy distinction from earlier studies in that the authors introduce foreign corporate tax rates in order to better capture the reality of an open economy with tax competition and any associated effects. Finally, alternative specifications use statutory/marginal corporate income tax rates and these are found to have a significant negative effect on growth.

The GKS methodology uses a pooled mean group (PMG) regression, which (mathematically) allows short run effects to vary across panel units but assumes that long run effects (i.e. the reported coefficients) are identical. This approach differs from mean group (MG) estimation, in which separate regressions are conducted for each panel unit and the average coefficients are used, while allowing all parameters to vary freely. It also differs from the fixed/random effects approach, where the coefficients are assumed to be identical for each unit in the long and short term. PMG estimation pursues a middle ground (Pesaran, 1997). The approach enables the authors to address some findings in earlier literature, such as in Mendoza, et al (1997), where some negative effects of taxation on growth were assumed to be transitory adjustments not reflective of long-run impacts. The authors also state that PMG estimation “makes full use of the available time-series information and provides estimates of long-run parameters without the need for long lag structures” (GKS, 2014).

The macro regression model results discussed above do show a significant negative effect of capital taxation using a mean group (MG) estimator (although the coefficient is of much lower magnitude than that of labor taxes). This is the only model specification in which AETR’s for labor, capital and consumption are included without other tax-related controls that may confound their effect. Other measures such as top marginal personal income tax rates and top corporate tax rates (as well as foreign taxes) were more central to the paper, and were generally negatively associated with growth. Their inclusion in the PMG models tended to flip the sign of the coefficient of the AETR on capital and most often resulted in its statistical insignificance. This is difficult to interpret, since the variables overlap substantially.

A recent study by Ten Kate & Milionis (2019) instead showed a positive growth effect of both corporate income tax and capital taxation in general. The study covered 77 countries over the period 1965-2010 (with 23 countries having data for the entire period). In studying capital taxation, the authors held constant the total level of taxation, which implies that increases in capital taxation would be replaced by equivalent decreases in other taxes. As noted above, the authors include property taxation as a capital tax, which we do not do in this paper. In a similar result, Jones et al (1993) found that capital taxation specifically could enhance growth if the resulting revenues were spent productively.

Several studies have also focused specifically on the corporate income tax rate or other subcategories of capital taxation. Lee and Gordon (2005) found a significant negative effect of corporate taxation on growth using statutory rates, while Arnold (2008) came to the same conclusion using realized share of tax revenue (tax receipts from CIT as a % of total tax receipts). We exclude here studies that lump personal and corporate income together as “income

(9)

taxes” or “distortionary taxes”. Other studies have shown either an insignificant effect or ambiguous results depending on model specification (Widmalm, 2001; Acosta-Ormaechea and Yoo, 2012).

Meanwhile, Arnold (2008) ranked various tax categories (personal, corporate, consumption and property) by growth- friendliness and found property taxation (especially recurring property tax) to have a positive and significant growth coefficient and to be the most growth-friendly tax. Consumption tax was also found to have a positive effect, while the effect of corporate income tax was negative. Acosta-Ormaechea and Yoo (2012) found a similar result on property and consumption tax, although the latter’s effect was not completely unambiguous.

Capital gains taxation vs. capital taxation

In an ideal scenario, in order to properly measure the tax incentives relating to the ownership of financial securities, we would have access to detailed data on capital gains income and taxation. In preparing to undertake this research, we did identify statutory capital gains tax rates for between 9 and 11 countries at various points in time (1990, 2000, 2012, 2014, 2016). However, even if this data represented an accurate measurement of tax incentives (it does not), it still would be insufficient to incorporate into a panel study. A key problem with this data is that statutory capital gains tax rates may bear little to no resemblance to the economic reality of capital gains taxation across countries. Each country’s tax code (we refer to countries within the OECD) is littered with various rules, qualifications, and exemptions that make it extremely difficult to estimate the tax base to which a statutory rate might be applied. Additionally, many advanced countries, if not all, allow for trading of financial securities tax-free within pension or retirement accounts.

Capital gains are often also treated differently if the seller has owned the asset for longer than some holding period, which may be as brief as one year.

Extending the analysis to the housing market, in most OECD countries the sale of an owner-occupied home is exempted completely from capital gains taxation. On top of that, taxpayers may be tempted to misclassify properties as owner-occupied even if they are fundamentally financial investments, in order to avoid taxation.

Since we focus here on speculative activity within the financial markets (specifically the stock market in this paper), an ideal analysis would include a measure of all financial asset sales and the capital gains tax revenue received from such sales. Even a measure of all taxable asset sales and the associated capital gains tax revenue would represent a substantial improvement over what is now available. Either of these would likely require detailed data from national tax authorities. There have been occasional studies on one country in isolation, for example in the United States, but sufficient data for a cross-country study does not appear to exist in the public realm. The reason capital gains taxation would be so beneficial as opposed to the measure we are forced to use (overall capital taxation) is that we hypothesize that a rush of activity towards the capital markets occurs during speculative periods. We reason that the average investment holding period is lower during these times. Because of the rapid increase in securities prices, over short holding periods the recurring taxes such as corporate income tax and dividend tax likely deliver much less impact to the overall return than taxation of the capital gain.

However, we expect theoretically that changes in overall capital taxation will still influence investment efficiency during periods of booming stock prices. It is just that the measurement of the effect will be far less precise than it might be

(10)

given adequate data specific to capital gains income and tax receipts. We justify our study’s use of the broader category of capital taxation based on the following reasoning.

The favored tax treatment of capital gains has important consequences for asset prices and speculative activity. By way of illustration, we take an example from the literature on the housing market. Looking at the cyclical fluctuation of leverage in real estate markets, Haughwout and Lee (2011) discuss the concept of the “user cost” of housing, representing the “annual flow cost to the owner per dollar of house price, taking into account after-tax financing costs, property taxes and insurance, maintenance and depreciation costs and the expected risk-adjusted return to owning the house.” House prices typically show a close relationship to the level of user costs (Himmelberg et al, 2005) and the former are sensitive to changes in the latter (Haughwout and Lee, 2011).

Although their paper does not explicitly reference capital gains taxes (only “property taxes”), the authors’ model contains a variable for “the tax environment” as well as a “risk adjusted return” to housing, the latter of which captures price appreciation. Since capital gains tax, when assessed, clearly affects the after-tax return on capital for a housing investment (or any other taxable investment), we assume that this is captured by one or the other of these variables and constitutes part of the user cost. It may not specifically be mentioned because of the frequent exclusion of owner- occupied housing from capital gains taxes under tax policies in most developed countries.

Although the preceding discussion refers to the housing market, the same fundamental logic applies to any other capital investment, with respect to the cost of ownership and after-tax return. For example, for a share in a corporation, we can consider taxation of corporate profit and the taxation of interest and dividends received to be the “user cost”

of ownership of the asset. Meanwhile the sum of interest and dividends received, plus the after-tax gain on the ultimate sale, represent its return. Lower capital taxation then implies a lower user cost, which should boost the price of assets, thus encouraging additional direction of financial capital toward speculation. This is consistent with the findings of Lang (2000) that capital gains taxes are indeed capitalized into share prices. Thus, we expect to find relatively more financial capital devoted to financial securities when capital taxation is lower.

Our Hypotheses

Our main premise is that at times in which financial speculation is particularly prevalent in the economy, allocation of financial capital will be geared increasingly towards pursuit of capital gains in financial and asset markets relative to production of real capital for use in the economy. This means that productive investments that would normally be undertaken in other times may be passed up on in these periods. Their financing cost may increase, for example, since investors expect higher than normal returns in financial markets. Additionally, it may be that the flood of liquidity in the financial markets allows for firms in struggling industries, or simply unproductive firms in general, to finance their (relatively unproductive) investments when they otherwise would not be able to.

We expect to see a negative relationship between the presence of higher speculative activity in the stock market and productivity growth in the economy once the investment during the speculative period is able to take its effect on production (we assume a two-year delay). However, we posit that higher capital taxation rates reduce the allure of speculation and should reduce the magnitude of this inefficient investment pattern when stock price appreciation is

(11)

high. In other words, capital taxation provides a social benefit during such periods that partially offsets its typical negative effect.

This leads to our main hypothesis:

1) we expect the interaction between an indicator for stock price “boom” periods and average effective capital tax rates to have a positive effect on future total factor productivity growth.

III. Methodology and Data

This section explains the model(s) used to test our hypotheses regarding the relationship between the average effective tax rate on capital income (“AETR capital”) and changes in investment efficiency during stock market booms.

Since we cannot observe investment efficiency directly, we search for its effect using growth of total factor productivity.

The initial model specification is as follows:

dTFPi, t +2 = 𝞪i + 𝛃1Ktaxi,t + 𝛃2Boomi,t + 𝛃3(Ktax x Boom)i,t + 𝛃4Xi,t + 𝛃5Pdt + ɛi,t (1)

Where a boom year is defined as:

= 1 if ΔSi,t > LT Avg(ΔSi) + 0.9*SD(ΔSi) Boomi,t =

(year) = 0 otherwise

Where ΔS represents the annual change in real stock prices, and the long-term average and standard deviation of ΔS is calculated over the period 1950-2015 for each country individually.

The variable dTFPi, t +2 represents the annual change in total factor productivity in country i two years following the current year, in which the investment and stock appreciation take place. This model is specified with fixed effects, with each individual country intercept represented by 𝞪i. Ktaxi,t represents the average effective rate of capital taxation within the economy. Boomi,t is an indicator variable that takes a value of one based on domestic stocks appreciating by more than a threshold level determined by the long-term average and standard deviation of stock returns in country I, as shown above. It is calculated at the annual level. The third independent variable is the interaction of AETR capital (Ktax) and the boom indicator. Xi,t is a vector of control variables, Pdt refers to time fixed effects (one dummy variable per period) and ɛi,t is the error term. Each variable is discussed in detail in the next section.

The regression based on formula (1) is carried out using three-year averages of the data. Many studies completed on taxation and growth have been conducted using five-year averages, in order to rule out short-term fluctuations and business cycle effects. We employ three-year averages for our main specification for two reasons. First, we use an indicator variable (i.e. Boom) that is based on annual stock price data, since we find that reliably evaluating

“speculative activity” over a period of three of five years carries many pitfalls. In our model, the classification of a period as a boom period refers to the presence of one or more boom years within that period. Using five-year averages, we naturally obtain much more variation in where a boom year falls within a period, and what stock price

(12)

behavior occurs in the surrounding years. This makes boom periods less comparable to each other, and may cloud our results. Secondly, we increase our sample size using three-year averages. The result of this choice is explored in the robustness section, where the results of five-year periods tests are presented and discussed.

III-a. Main variables and data Dependent variable

We use the annual change in total factor productivity (dTFPi, t+2) two years into the future in order to measure the efficiency of investment over the current period. That is:

dTFPi, t+2 = % TFP growth in year t + 2

All remaining variables are all measured in the current year. We assume a two-year lag between the time of investment and the effect on productivity. This is in line with Wurgler (2000), who cites an average period of two years established in prior literature for fixed investment to become productive. He specifically references the work of Mayer (1960) and Hall (1977).

If, as we theorize, investment decisions are skewed away from the real sector of the economy during periods of high stock appreciation, and profitable real investment opportunities are passed up on, technology spillover effects from investment should be less impactful after such periods, and we should observe a negative effect on TFP growth.

Data for this variable comes from a database created by Bergeaud et al (2016), which covers productivity variables for advanced economies from 1890-2016. As noted earlier, the data are consolidated into three-year averages before running the regressions, and dependent variable dTFP is a future period value at t + 2. For example, in our models we evaluate the impact of independent variable Ktax during the period 1950-1952, when the investment takes place, by measuring dependent variable dTFP in the period 1952-1954.

Independent variables

Capital tax rate (Average effective tax rate on capital income)

Ktax = Tax receipts from capital income / Capital income from national accounts

This variable represents the average effective tax rate on capital income for each country. We use this rate, rather than the ratio of capital tax receipts to GDP, since the AETR reflects (holding position in the business cycle constant) the degree to which capital investment, including investment in financial securities such as stocks, is faced with an incentive or disincentive from tax policy.

Data for this variable comes from taxation rates broken down into the categories capital, consumption, and labor by McDaniel (2007). All interest, dividends, corporate profits, and capital gains from asset sales, whether earned by

(13)

individuals, corporations or the self-employed, are included as capital income in the calculation of AETR capital.

According to McDaniel’s 2007 paper, property taxation paid by households is considered a consumption tax, rather than a capital tax, since housing services are considered consumption in the national accounts (property tax paid by other entities is considered capital tax). This differs from the treatment of property tax in other studies, including that of Ten Kate & Milionis (2019), in which all property tax is included in capital taxation. The absence of property taxes from capital tax suits our purposes in this paper well, as we focus on taxation of the returns from share ownership rather than real estate. This is a meaningful difference, since Arnold (2008) cites immovable property taxes as the most growth friendly tax. Finally, taxation of direct investment (e.g. import duties or VAT on capital equipment purchases) is also considered separately from capital taxation. This is most closely aligned to consumption tax.

Wages and social security taxes are considered as part of the labor taxation category.

An advantage of using AETR’s as a measure of taxation is that they “correspond to aggregate realized tax rates…

which aggregate the information on statutory taxes, credits, and deductions implicit in national accounts and revenue statistics in a manner that captures the overall tax burden from each tax and maintains consistency with the representative agent framework” (MRT, 1994). The authors note that empirical work by Lucas and by Razin and Sadka suggest that the AETR’s “are useful approximations to the taxes that distort economic decisions in dynamic macroeconomic models” (MRT, 1994).

A note about depreciation: the capital tax rates utilized in this study are calculated on gross income excluding depreciation. Per McDaniel (2007), this has to do with the lack of comparability between the national accounts measure for depreciation (consumption of fixed capital) and the amount of depreciation deducted from income for tax purposes by businesses. As noted in Gorter and Mooij (2001), “international differences between depreciation allowances for machinery and buildings… cause differences between the taxable corporate income of two otherwise equivalent corporations.” For example, a set of rules allowing rapid depreciation of capital assets might allow for habitual large capital purchases which continually reduce accounting profits on an annual basis and result in a persistently low tax liability. The fact that our capital tax rates are calculated gross of depreciation is useful in this regard. The rates themselves may be understated, but since they are based on actual tax receipts and gross income, they should be more comparable across countries and should also encapsulate differences in depreciation policies.

Stock “boom” years

The classification of a boom is calculated at the annual level as specified in formula 1. However, since our model is run in three-year averages of data, the actual variable ‘Boom’ in the regression takes on a value of ‘1’ if any year in the period meets the criteria for a boom year. Time series for the domestic stock index of each country, as well as CPI, are taken from Jorda et al (2017).

Adopting an approach of Bordo and Jeanne (2002), we define a boom year (in the “base case”) as any year in which average stock prices in a country appreciate in excess of the long run average annual change plus 0.9 times the long run standard deviation of returns. The multiplier applied to the standard deviation is not fixed by theory but rather is meant to be calibrated based on the characteristics of the unit being measured. Oliviero et al (2019) used a multiplier of 1.1 in calculating booms and busts in property tax (representing tax policy changes), for example.

(14)

We populate this indicator variable based on annual returns instead of the overall change in prices over each period, in order to better capture speculative periods. We wish specifically to study periods of unusually high appreciation.

For example, we would rather look at a period with one year of 20% appreciation, one year of (5%) decline and another of 5% gain than a period of stable, modest returns summing to the same overall growth for the period. We believe the use of annual returns does a better job of capturing periods reflecting a higher degree of speculative activity. Aside from our standard approach of an annual boom indicator, a model with an independent variable (and its interaction with Ktax) based on the average stock appreciation in a given period is also tested and discussed later in this section.

Interaction Ktax x Boom

We hypothesize that although capital taxation may have a negative effect in general on productivity growth, this effect varies depending on the environment in the financial markets. More specifically, if speculative activity is increased, and we expect some investment at the margins to be misallocated as a result, we believe the rate of capital taxation will impact the degree of misallocation. Since capital taxation reduces the returns of financial asset investment and speculation for capital gains, we expect misallocation to be smaller during asset booms when higher capital taxation rates are in force. Therefore, we expect a positive coefficient for the interaction term.

Control variables

Time trend

We start with a discussion of the time variable ‘pd’ (for ‘period’) due to the sensitivity of our results to the construction of the three-year average periods. Our results differ, sometimes substantially, depending on how the overall sample period of 1950-2015 is divided: for example, whether the first period covers 1950-1952, 1951-1953 or 1952-1954, with the rest of the periods shifted accordingly. Because of the lagged and future values used in the construction of some model variables, we sometimes need to drop one or more partial periods. Which exact years are covered in each specification is presented alongside the regression results in each table. All models have been run with each possible starting year for the first period.

Real GDP growth

Annual growth in total GDP is included for each country and averaged over each three-year period. This variable is generally highly related to our dependent variable of growth in total factor productivity when considered concurrently.

There is clearly a strong relationship between the TFP residual and GDP, but we focus on TFP growth since it is in theory not reflective of changes in capital and labor inputs (as GDP is), but rather reflects the efficiency with which those inputs are employed.

The numerical correlation is noteworthy even with the two-year time difference, although the degree of correlation is naturally stronger in the five-year period specification. With three-year average periods, only one year included in the average future TFP growth is also included in the calculation of current GDP growth for that period. For example, GDP growth from 2010-2012 is tested against TFP growth from 2012-2014, so 2012 is the only year included in both averages. In the five-year specification, there are three such years included in each average value for a given period.

This of course influences the magnitude of the correlation.

(15)

This variable is highly significant in most (but not all) specifications of the model, with the expected positive coefficient.

In terms of collinearity, however, the correlation is only moderately high at 0.54 in the three-year models, rising to 0.70 in the five-year models (we report here the highest values among the various period specifications, created from different starting years).

Size of the financial system

We use total private bank loans to GDP as a measure of the size of the financial system and include this as a control variable. This is in accordance with Wurgler (2000) and Beck and Levine (2002), who showed that financial development positively influenced Wurgler’s “investment elasticity”, which is very similar to what we are trying to measure in this paper. It is noteworthy that this earlier research included both advanced and developing nations and therefore a distinction based on the size of the financial system is likely to have been more meaningful between these countries than between our sample of developed OECD nations.

The correlation of credit-to-GDP with future TFP growth is moderate at -0.45 based on the pooled three-year average data, and its correlation with AETR capital is negligible (less than 0.10). However, the correlation with AETR capital rises to 0.84 when the annual data are averaged across all countries, yielding one observation per year (the correlation of credit-to-GDP with future TFP growth rises in magnitude to -0.72). Collinearity issues are discussed in Appendix A-1. We observe no evidence that inclusion of this variable negatively impacts the precision of our results.

Stock index level (log)

In addition to using the annual change in stock prices for population of our indicator variables, we include the logged average level of the stock market index in each country as an additional control variable. We use logs since the distribution of this variable is not normal. To avoid negative values, we first multiply our index measure (in which the 1990 index value = 1) by 100 before taking the log. We believe it is important to include this variable since the AETR on capital income is somewhat correlated to the stock index level. When the data are averaged across all countries, yielding one observation per year, this variable’s correlation to AETR capital is 0.62.

Since corporate profits are a key metric watched by investors to estimate the value of companies, it is intuitive that stock prices increase when corporate profits increase. Since corporate profits fluctuate more over the economic cycle than GDP, we see that tax receipts (the numerator in the AETR) are more volatile than the national accounts income in the denominator. Therefore, our variable Ktax is sensitive to the business cycle. By controlling for the stock index level, we are able to focus on the portion of Ktax that is not influenced by cyclical changes in corporate profits, to the extent these are measured by cyclical changes in the stock market index.

Overall tax level

taxGDPi = Total government revenues/GDP

The variable taxGDP represents total government revenue as a % of GDP, with data coming from the macrohistory database created by Jorda et al (2017). This database covers a number of macroeconomic variables for 17 OECD countries from 1870-2016.

(16)

We expect a high correlation between this variable and AETR capital, since capital taxation is a component of overall taxation. The correlation is modest at 0.33 in the pooled sample, but rises above 0.9 when the data are averaged across all countries, yielding one observation per year. Collinearity issues are discussed in Appendix A-1. We observe no evidence that inclusion of this variable negatively impacts the precision of our results.

Since Ktax is an average effective tax rate, an increase in its value does not imply a definitive change in the overall level of taxation. By controlling for overall taxation, we effectively set in place that an increase in Ktax implies a decrease in one or more AETR’s on other categories of income.

It may be a potential weakness that we control for the overall level of taxation without evaluating the effectiveness of the associated incremental government spending. However, attempting to measure this is outside the scope of this paper. Also, since we focus on investment efficiency while controlling for GDP, the theoretical imperative for controlling for expenditure is reduced here relative to pure growth studies.

Period (time) fixed effects

We include period fixed effects in all model specifications in order to account for economic shocks that affect our sample countries similarly. Since we deal with OECD countries, these are relatively common. Examples include the oil shortages of the 1970’s or the effects of the global financial crisis in 2008-2009. We include dummy variables for each period, which costs us 21 degrees of freedom in the three-year models (12 in the five-year models). When the period dummies are included, the time trend variable is omitted due to collinearity.

III-b. Alternative model specifications Dealing with stock price volatility

Since we aim in this paper to test the effect of capital taxation on investment efficiency in periods of high speculation in financial markets, the main focus is on periods in which stock prices rise rapidly in one or more years. However, one potential flaw with our measure for speculation is that given the degree of fluctuation of stock prices, a large increase in stock prices in any given year could be a function of price behavior in the preceding year (or years).

Treating a rebound from a stretch of substantial stock declines the same as a rapid stock price increase following a few years of modest price gains could lead to biased results. Therefore, we attempt to distinguish between such episodes by incorporating a measure of the position of a country’s stock index value relative to the long-term trend at the time of a stock boom.

Specifically, we introduce a Hodrick-Prescott filter to the (CPI deflated) stock price index data for each country. We apply a smoothing parameter of 100, which is typically appropriate for annual data (Hodrick & Prescott, 1997).

Analysis of the filtered data confirms our suspicion that boom years may be more likely when stock values are below trend, for example after one or more years of price declines. Figure 1 shows the number of boom years that occurred within various intervals of the HP filter’s cyclical component value (this is discussed in more detail below). We can see the distribution is heavily skewed to values below zero.

(17)

Figure 1 – Frequency of boom years (boom = 1) at values of hp cyclical indicator

Model 2 – Using the Hodrick-Prescott filter to address stock price cyclicality

In order to account for years in which stock-price growth may simply represent a cyclical return to trend, we impose the condition that for a year to be considered a boom year, our cyclical indicator from the HP filter must be greater than -0.20 at the end of the previous year. This model is specified in formula 2.

dTFPi, t +2 = 𝞪i + 𝛃1Ktaxi,t + 𝛃2Boomi,t + 𝛃3(Ktax x Boom)i,t + 𝛃4Xi,t + 𝛃5Pdt + ɛi,t (2)

Where a boom year is here defined as follows:

= 1 if {ΔSi,t > LT Avg(ΔSi) + 0.9*SD(ΔSi)}

AND Boomi,t = {Ct-1 > -0.20}

(year)

= 0 otherwise

Where Ct-1 represents the value of the cyclical component of the stock index (from the HP filter) for country i at the end of the year prior to the current year.

The HP filter decomposes our time series of the real stock index into a smoothed long-term trend ψ, and C, which measures the deviation of the current value of stk_level (the stock index value), and the smoothed long-term trend from the HP filter.

(18)

That is: Stk_level = ψ + C

Our stock index variable has a value of 1 at the end of the year 1990. Across all sample country-years, the index level ranges from 0.04 – 9.06, with the 25^th percentile at 0.61, the median at 1.04 and the 75^th percentile at 1.85. The interpretation of units for this metric is unwieldy. One unit of this measure represents the value of a country’s stock index at the end of 1990, so that a value of -0.20 for Ct-1 indicates that the stock index value at the end of the prior year is below the smoothed long-term trend by 20% of the value of the stock exchange in 1990. A more intuitive way to understand it can be explained using the graphs below. These figures depict the current value of the stock index fluctuating around the smoothed trend. The value of (-0.20) can be understood as a position on the line for the current index value below the smoothed long-term trend, beyond which approximately 19.5% of our sample data fall (across all country-years). The distribution of years over values of the HP filter cyclical component can be viewed in appendix figure B3.

Figure 2a – Stock trend by country

Y-axis: Index of real stock prices (1990 = 1)

(19)

Figure 2b – Stock trend by country

Y-axis: Index of real stock prices (1990 = 1)

Model 3 - An alternative measure for speculation

Partly to address the variation of results across period specifications, we also substitute a period-based measure of stock price appreciation as an alternative measure for speculative activity. In this model, we use formula 3:

dTFPi, t +2 = 𝞪i + 𝛃1Ktaxi,t + 𝛃2BoomPdi,t + 𝛃3(Ktax x BoomPd)i,t + 𝛃4Xi,t + 𝛃5Pdt + ɛi,t (3)

Where a boom period is here defined as follows:

= 1 if {ΔSi,p > LT Avg(ΔSi,p) + 0.8*SD(ΔSi,p)}

BoomPdi,p = AND

(period) {Cp-1 > -0.30}

= 0 otherwise

(20)

Where ΔSi,p is the average annual change in real stock prices over period p, and the long-term average and standard deviation of ΔSi,p is calculated over the period 1950-2015 for each country individually. Cp-1 represents the value of HP cyclical component of the stock index for country i at the end of the year prior to the start of each period.

Results for all models are presented in section IV.

III-c. Sample

Our panel data consist of 14 OECD countries and cover the 66-year period 1950-2015, resulting in 924 total annual observations. The panel is completely balanced. All growth rate variables are calculated using the one-year lagged value and our dependent variable uses a future value from two years forward. This reduces the data set to 882 usable observations. After aggregating the data into three-year average periods and varying the start year of the first period (to test for consistent results), we typically end up with 21 full three-year periods for 14 countries (12 periods when five-year averages are used), for a total of 294 observations (168 with five-year averages). Sample countries include Australia, Belgium, Canada, Switzerland, Germany, Spain, Finland, France, United Kingdom, Italy, Japan, the Netherlands, Sweden, and the United States.

Descriptive Statistics

In figure 3 below, the average effective tax rate on capital income, a key independent variable in our study, is shown for each country. In figure 4, the overall average capital tax rate for all countries in aggregate is shown over time.

Figure 5 shows the mean and standard deviation of each country’s AETR capital, TFP level and TFP growth over the entire sample period.

In figure 4, we see that on average, effective capital taxation rates grew from the early 1960’s into the 1990’s across our sample countries, but have entered a declining trend since that time, and with much wider fluctuation. The graph provides clear visual evidence that capital AETR’s move to some degree in concert with the broader economy and perhaps even more closely with corporate profits and stock price values. The metric grows strongly in the late 1980’s, the second half of the 1990’s, and the early 2000’s, all periods where economic growth and stock prices rose strongly in the majority of OECD countries. In each case this was followed by a decline, as recession and stock declines were also relatively widespread in sample countries after these periods. This suggests the AETR on capital is sensitive to business cycle fluctuations. We account for this by controlling for the logged level of stock price index in each country in our models.

(21)

Figure 3 - Capital tax rates by country

AETR capital: total tax receipts on capital income/total capital income from national accounts

Figure 4 - Capital tax rates over time (aggregate across sample)

(22)

Figure 5 - Capital tax rates, TFP, and annual TFP growth by country (1950-2015)

TFP calculation: GDP/F(K,L) in 2010 $USD on PPP basis (Bergeaud et al, 2014)

The dependent variable is charted in figure 6 as an average across all sample countries over time. We see clearly the historical downward trend in productivity growth in the advanced countries over time, including a significant downward spike to slightly below zero in the aftermath of the global financial crisis.

Figure 6 - Dependent variable (in three-year averages)

In figure 7, we show the trajectory of the effective tax rates applied to different categories of economic activity in advanced countries since 1950, using rates as calculated by McDaniel (2007), except for overall taxation (own calculation). On average, effective labor taxation grew rapidly from 1950 to the mid-1980’s. Effective consumption tax has also risen considerably, albeit from a low base. Meanwhile effective capital taxation has leveled off and declined somewhat since the end of the 1990’s/early 2000’s. As noted above, effective capital tax rates also appear sensitive to business cycle fluctuations (tax receipts appear to fluctuate more than the associated national accounts income).

(23)

Figure 7 - Decomposed tax rates, average all countries

Data in three-year periods (e.g. 2010 reflects 2010-2012)

Figure 8 shows the classification of boom periods and its variation depending on the choice of start period. Such differences may plausibly affect our results, since a boom year may occur anywhere within a period (e.g. the first or last year in the period). If the change in period cutoffs moves a boom year (say, the first year of period x) into an earlier period (now the last year of period x-1), and period x now has a value of 0 for boom, this clearly changes which years of capital taxation are compared to future TFP growth (also in the interaction with boom). We attempt to address this by including time fixed effects (period dummies) in our models. This is partially successful but does not eliminate material variation of the model results given the years chosen as the starting period.

Figure 9 provides information on the historical average real annual stock returns for our sample countries, and threshold values depending on the choice of the standard deviation multiplier in the definition of the boom variable (the baseline choice is 0.9, shown in bold). It also shows the total number of booms per country.

(24)

Figure 8 - Occurrence of boom years* (@ 0.9*SD multiplier)

*As defined in model 1.

Figure 9 - Historical stock return information by country

Correlations

The correlation matrix below, with the data in three-year averages, shows real GDP growth to be significantly correlated with our dependent variable at 0.54. The correlation is relatively strong even though the TFP growth variable reflects values two years into the future. Credit to GDP is also relatively highly (negatively) correlated at a maximum magnitude of -0.45, reflecting the deterioration of this metric as a positive factor correlated with growth, as discussed earlier. The period (time trend) variable shows a correlation of -0.61, matching the pattern in TFP growth seen in figure 6. Other notable correlations are those between taxGDP (overall tax rate) and the capital tax

(25)

rate, which is modest at approximately 0.33, and the correlation of 0.73 between credit-to-GDP and the time trend variable.

In order to give the reader a more complete picture of the correlations between our variables, we must note a few facts. First, the correlations with the data in three-year averages vary somewhat depending on the years chosen as the first period. For example, a lower correlation of 0.42 is observed when the data are averaged with the first period constituting 1951-1953. This exemplifies the sensitivity of our model results to the choice of the starting period. This appears to be a quirk of the data and reflects differences in the period averages deriving from the cutoff of each period. It also reflects our limited sample size. Secondly, these correlations are based on the pooled data from all countries. The relationships between several variables appear much stronger (perhaps more in line with intuitive expectations) when the data are averaged across countries, giving one observation per variable per year.

This approach also more closely mirrors the correlations we see when testing the data one country at a time.

A full presentation of the variable correlations using different aggregations and period start years is included as appendix section A-1. We also describe there our efforts to investigate correlations with potentially problematic magnitudes, including the results of variance inflation factor analysis. Our analysis led us to conclude that there is no evidence that multicollinearity is affecting the precision of our results.

Figure 10 - Correlation matrix, three-year average periods (first period 1950-1952)

Definition of boom is from model 1.

Panel data tests – fixed vs. random effects

The Hausman test was run for our model with the starting year of the three-year period specification varied. The test returned p-values ranging from 0.12 to 0.21, such that we could not reject the null hypothesis that the random effects and fixed effects models were equally efficient. However, we used the xtoverid test within Stata as a second check and received p-values below .05 for all period specifications. The theoretical case is also strong that

unobservable, time invariant factors are correlated with our explanatory variables. Therefore, we use fixed effects.

Using five-year average periods, the Hausman test results are completely inconclusive, with p-values ranging from 0.27 to 0.48. However, per the xtoverid test, we can reject that the random effects and fixed effects models are

(26)

equally efficient with a confidence level greater than 99%. Since the theoretical case is also strong, we continue to use fixed effects.

IV. Results

The main regression results for the three-year average periods are displayed in table 1 below. Columns 1, 2 and 3 represent the same model run with a different starting year for composing the average periods, as the model was found to be sensitive to this decision. The starting period for each specification can be found at the top of each column.

Recall that the dependent variable measures TFP growth two years into the future, such that the value for 2013 is the TFP growth in 2015. As such, column 3 contains one fewer period and thus 14 fewer observations, as the final period must be dropped. The final period would otherwise be 2012-2014, and there is no value for TFP growth in 2016 in our dataset.

Three-year results, @ 0.9 SD Multiplier

In the base model (model 1) results, as seen in columns 1, 2 and 3, the coefficient for the effective capital tax rate is negative and significant at the 5% level for all period specifications. This demonstrates a negative association between the intensity of capital income taxation in the economy and future total factor productivity growth, after a two- year period during which investment decisions have time to take effect.

In column 1, the coefficient of -0.0719 indicates that in a period containing zero ‘boom’ years, an increase of 10 percentage points in the AETR on capital income would result in a 0.72% reduction in the average annual growth rate of TFP over the period beginning two years into the future.

The coefficient on ‘boom’ is negative and significant at the 5% level in columns 2 and 3 of table 1, but only at the 10%

level in column 1. The magnitude is similar across each column. The column one coefficient value of -0.0092 indicates that the intercept for each country is reduced by the amount of the coefficient in periods in which there are one or more boom years (which we hereafter refer to as “boom periods”). The overall effect on future TFP growth is dependent on the level of capital taxation.

The interaction of the effective capital tax rate with boom is positive in all period specifications and is significant at the 10% level in two of three columns, but at the 5% level only in column 2. The results therefore provide limited support for our hypothesis that higher capital taxation positively influences investment efficiency during times of increased speculative activity. The positive and statistically significant column 2 coefficient of 0.0419 indicates that in this period specification, during boom periods, the marginal effect of capital taxation on future TFP growth becomes less negative. Across all period specifications, however, our results fail to demonstrate sufficient statistical significance to assert the existence of this effect overall.

The overall effect of higher speculation on (real capital) investment efficiency also includes the negative change in intercept when boom = 1. Thus, whether the overall effect of boom periods can be shown to be negative is dependent on the level of effective capital tax rates. We include a graph of the average effect of boom periods at different levels

(27)

of effective capital taxation in figure 11. All other variables are held constant. We use the column 3 specification since it depicts a rather common pattern in the results of our various models.

Table 1 - Base model results (with unit and time FE)

(1) (2) (3)

VARIABLES S 1950-52 S 1951-53 S 1952-54

Ktax -0.0719** -0.0643** -0.0802**

(0.0259) (0.0223) (0.0237)

boom -0.00922* -0.0101*** -0.00975**

(0.0888) (0.00903) (0.0120)

boom x Ktax 0.0336 0.0419** 0.0349*

(0.139) (0.0160) (0.0932)

dGDP 0.149* 0.155** 0.157**

(0.0738) (0.0267) (0.0123)

cGDP 0.00399 0.00499 0.00583

(0.374) (0.272) (0.224)

Stk_lvl 0.000660 0.00123 0.00192

(0.746) (0.475) (0.306)

taxGDP 0.0767** 0.0503 0.0716

(0.0208) (0.123) (0.108)

Constant -0.00589 -0.00271 -0.00919

(0.688) (0.817) (0.533)

Observations 294 294 280

R-squared 0.543 0.615 0.583

Robust p-values in parentheses; *** p<0.01, ** p<0.05, * p<0.1

Figure 11 shows that in the column 3 specification of our base model, the effect of speculative activity on future TFP growth is negative with 95% confidence at effective capital tax rates below approximately 20%. At rates higher than this level, the effect of a boom period cannot be distinguished from zero. Additionally, we have insufficient evidence (except based on the specification in column 2) to conclude that the effect of a boom period is different at any two levels of AETR capital. This can be observed visually in figure 11, as there is no area of the graph along the x-axis in which the (discretely calculated) 95% confidence intervals for the boom effect do not overlap. This of course is a corollary to the lack of statistical significance of the interaction coefficient in column 3 at the 5% level (p-value of .093).

To ensure accurate interpretation of the interaction effect shown in figure 11, we analyzed the distribution of capital tax rates for the matching column 3 specification. The widening of the confidence interval (in figure 11) above the value of approximately 20% for AETR capital does not appear to be caused by a lack of common support from data in the sample. There is a similar amount of observations (country-periods) between capital tax rates of 10% and 20%

as exists between rates of 25% and 35%. Boom years are also approximately normally distributed across values of the effective capital taxation rate. Histograms are provided in appendix figures B1 and B2.

(28)

Figure 11 - Interaction effect - column 3, starting period 1952-1954

Shaded band represents 95% confidence interval

Sensitivity of the model to “boom” stock price increase threshold

Given our model design, our results may be sensitive to the threshold chosen as the cutoff in annual stock price appreciation beyond which a year is considered a boom year. Recall that the base calculation for the boom variable is as follows:

= 1 if ΔSi,t > LT Avg(ΔSi) + 0.9*SD(ΔSi) Boomi,t =

(year) = 0 otherwise

Figure 12 shows some key model coefficients and their associated p-values across varying levels of the SD multiplier used in our “boom” calculation. The effective capital tax rate variable maintains its sign and statistical significance at the 5% level at a reasonably wide range of values for the SD multiplier. Our model fits the sample data best for multipliers of 0.9 and 1.0. There is no particular theoretical justification for our choice of 0.9 in the base model. This table shows that our results are quite sensitive to this choice. However, the sign of each estimated coefficient is uniform throughout the specifications and p-values are typically under 0.20 throughout, with only a few exceptions.

(29)

Figure 12 - Key coefficients with varying period specification and SD multiplier

Control variables

GDP Growth

With respect to the control variables, average real annual GDP growth is positive and significant at the 5% level in columns 2 and 3, but only at the 10% level in column 1. This result appears to be mainly the result of chance, based on how the period cutoffs fall in the data.

Credit-to-GDP

The coefficient for our credit-to-GDP variable does not show as significant in any specification in table 1. We cannot assert with any meaningful degree of confidence that the coefficient differs from zero. The positive estimate for the coefficient contradicts the strong negative correlation of cGDP with dTFP2 as shown in appendix figure A7 (with the data aggregated across all countries). It is likely to be affected by strong positive correlations with AETR capital (perhaps due to business cycle impacts on tax receipts) and the stock index level, as well as the time trend.

Period (Time) Dummies

Period fixed effects are included in all models.

IV-a. Alternative Specifications

We now show the results of two alternative model specifications. In model 2, we aim to address a key potential weakness in our model pertaining to the classification of boom periods. In model 3, we use an alternative measure for speculative activity that is determined at the period level, unlike model 1, which evaluates boom conditions at the annual level.

Table 2 shows that our key coefficient estimates are largely consistent with the base model, except that statistical significance weakens across all key variables, especially in column 1, where the boom variable now becomes

(30)

insignificant at the 10% level. There are some modest reductions to many coefficient estimates, but the results follow a similar pattern, with no sign changes in the estimates. These results generally support the view that our base model specification is not unduly influenced by classification of boom years when the gains in that year are largely a function of a rebound from a cyclically low level of stock prices.

Table 2 - Model 2 results - Addressing rebounds from cyclical lows using the Hodrick-Prescott filter

Robust p-values in parentheses; *** p<0.01, ** p<0.05, * p<0.1; Years excluded from “boom”

classification if HP cyclical value < -0.20 at the start of the year.

Since our use of a threshold of -0.20 for the cyclical component has no particular theoretical justification, we provide figure 13, which shows the effect of varying this choice. The figure also shows the number of booms removed in each period specification by introducing this condition.

(1) (2) (3)

VARIABLES S 1950-52 S 1951-1953 S 1952-1954

Ktax -0.0647** -0.0579** -0.0745**

(0.0400) (0.0355) (0.0295)

boom -0.00901 -0.0116** -0.0116**

(0.164) (0.0121) (0.0353)

boom x Ktax 0.0280 0.0411** 0.0350

(0.321) (0.0306) (0.147)

dGDP 0.162** 0.165** 0.161**

(0.0500) (0.0197) (0.0123)

cGDP 0.00351 0.00433 0.00510

(0.411) (0.321) (0.262)

Stk_lvl 0.000627 0.00114 0.00183

(0.753) (0.532) (0.314)

taxGDP 0.0754** 0.0545 0.0722*

(0.0371) (0.109) (0.0927)

Constant -0.00690 -0.00383 -0.00964

(0.615) (0.769) (0.479)

Observations 294 294 280

R-squared 0.542 0.616 0.588

DOES THE EFFECT OF CAPITAL TAXATION ON PRODUCTIVE INVESTMENT VARY WITH THE LEVEL OF STOCK MARKET SPECULATION?

Faculty of Economics and Business