Journal Issue
Share
Article

The Role of Domestic Debt Markets in Economic Growth: An Empirical Investigation for Low-Income Countries and Emerging Markets

Author(s):
International Monetary Fund. Research Dept.
Published Date:
March 2010
Share
  • ShareShare
Show Summary Details

Although there is a large body of literature on the effects of public debt on capital accumulation and growth in industrialized countries—dating to Diamond (1965) on the welfare effects of national debt, and Barro (1974) on debt neutrality—the study of public debt and growth in lower-income countries (LICs) and emerging markets (EMs) has mostly happened in the narrower context of external debt.1 The debt Laffer curve literature a la Sachs (1989) and Husain (1997), and empirical investigations thereof—including most recently, by Pattillo, Poirson, and Ricci (2002)—have all concentrated on the growth impact of “external” debt. Even here, the focus has been the country’s total external debt, rather than its publicly owed counterpart.

Indeed, little formal academic or policy interest has been shown in understanding the possible relationship(s) between public “domestic” debt and economic growth in LICs and EMs.2 Although some work on local bond markets has been initiated by EMs and international financial institutions of late, this predominantly reflects a response to the East Asian financial crisis, which was seen as stemming from, inter alia, weak financial markets and an excessive reliance on bank-intermediated external finance. The arguments made in this literature focus on the benefits of local bond markets generally and, while the role of government bond markets is stressed, the implications for domestic debt, as a whole, and/or fiscal financing strategies are neither clearly derived, nor highlighted.

Consequently, surprisingly basic fiscal and public debt policy questions have remained unanswered to date, such as: Is there a causal link between domestic debt and growth in general, and, in LICs and EMs specifically? Controlling for endogeneity, does domestic debt contribute to economic growth? If yes, what are the relevant channels: financial development, capital accumulation, investment efficiency or institutional? Is the impact linear or nonlinear? If nonlinear, can a notion of an optimal (growth-maximizing) domestic debt share in GDP or deposits be developed? To what extent might such “optimal share” calculus depend on the quality of the underlying debt, or the macroeconomic, financial sector and institutional environment obtaining in a country?

Such questions are more pertinent than ever given the increased scope for expanding domestic debt in many LICs and EMs following external debt reduction initiatives and a surge in international portfolio interest in local currency bond markets. A meaningful policy response, however, remains constrained by a dearth of comprehensive empirical studies that examine domestic debt’s impact on savings, investment, financial deepening, institutions and, hence, growth.

The lack of interest in formally studying the impact of domestic debt on growth could be attributed to (1) data unavailability—reliable datasets on domestic debt either do not exist or are not amenable to empirical analysis; (2) a wide-spread perception that domestic debt is “endogenous” rather than an exogenous policy choice variable that governments can tweak to affect macro-financial outcomes: countries’ domestic debt issuance capacity is “determined” entirely by their level of income, pool of savings and institutional quality; and (3) the relatively small size of domestic debt relative to external public debt in most LICs and EMs. These factors have, arguably, combined over the years to “crowd out” the amount of attention paid to domestic debt.

The aforementioned “attention deficit” has, however, not prevented policymakers from taking a strong view on the desirability, or otherwise, of domestic debt in developing countries. In most cases, policy advice from international financial institutions has sought to limit the accumulation of domestic debt through zero or negative net domestic financing conditionalities. Shallow financial markets, financial repression propensities and poor debt management capacity, which are found in many LICs and even some EMs, have led to the belief that domestic debt expansion will have significant negative implications for private investment, fiscal sustainability and, ultimately, economic growth and poverty reduction. In addition, given that many LICs have access to very cheap external finance, in the form of concessionary loans and grants, the idea of borrowing domestically at market rates appears expensive. In summary, low domestic debt issuance has been generally considered beneficial for economic development.

In recent years, however, some researchers have begun to echo the positive view of many market participants regarding the importance of domestic debt instruments for monetary and financial systems, as well as the development of political institutions. Compared to other forms of budgetary finance, market-based domestic borrowing is seen to contribute more to macroeconomic stability—low inflation and reduced vulnerability to external real and domestic monetary shocks—domestic savings generation and private investment. This seems to be supported by the experience of fast growing EMs such as China, India and Chile, which have maintained relatively low external indebtedness and avoided major financial or fiscal crises.3

The objective of this paper is to fill this void in the literature by bringing together the various arguments for and against domestic debt issuance currently scattered across the literatures on capital markets, public finance, debt management and fiscal sustainability (Section I); compiling a new domestic debt database spanning the period 1975–2004 for 93 LICs and EMs, as well as consolidating existing databases on domestic debt (Section II); using panel econometric techniques to examine the endogeneity of domestic debt and its impact on growth with a view to obtaining a sense of the optimal size and quality of domestic debt (Section III); and presenting empirically-grounded policy conclusions on domestic debt to guide macro-financial practitioners, especially in LICs (Section IV).

I. Existing Theoretical and Empirical Studies

Pros and Cons of Domestic Debt

Issuing domestic debt, whether to finance the fiscal deficit or to mop up monetary liquidity, involves a complex evaluation of the costs and benefits to the economy. Although practitioners’ views on the subject abound, the academic literature on the pros and cons of domestic debt issuance and the channels through which this type of financing can affect public finances, the financial sector and the real economy is limited. Critics of domestic debt are concerned with the repercussions on private sector lending, fiscal and debt sustainability, weakening bank efficiency and inflationary risks.

The most prominent concern about domestic debt is the crowding out effect on private investment. When governments borrow domestically, they use up domestic private savings that would otherwise have been available for private sector lending. In turn, the smaller residual pool of loanable funds in the market raises the cost of capital for private borrowers, reducing private investment demand, and hence capital accumulation, growth and welfare (Diamond, 1965). In shallow financial markets, especially where firms have limited access to international finance, domestic debt issuance can lead to both swift and severe crowding out of private lending.

Second, critics of domestic debt are also concerned with repercussions on fiscal and debt sustainability. Domestic debt is viewed as more expensive than concessionary external financing (Beaugrand, Loko, and Mlachila, 2002).4As a result, the interest burden of domestic debt may absorb significant government revenues and thereby crowd out pro-poor and growth enhancing spending. In addition, reliance on domestic financing may also delay tax mobilization efforts, which may be necessary but politically costly. Given the short-term structure of the domestic debt portfolio in many LICs and EMs—see, for example, Christensen (2004) on sub-Saharan Africa—governments also face a significant liquidity risk from having to constantly roll-over large amounts of debt.

Third, the cost of domestic debt may rise sharply due to time inconsistency problems when government credibility is low. If the state has weak (direct) tax collection, as is the case in most LICs, the state will have a strong incentive to monetize deficits and use the net domestic financing window to both, generate seigniorage, and, reduce the real burden of existing domestic debt. Under these circumstances, the government faces a classic time inconsistency problem and, therefore, either cannot issue nominal debt at all, or has to pay a significant premium to compensate investors for the potential risk of surprise inflation.5

Fourth, high-yielding government domestic debt held by banks can make them complacent about costs and reduce their drive to mobilize deposits and fund private sector projects. The incentive to provide credit to the private sector is often weakened by a poor credit environment. Hence, from a risk weighted perspective, government debt is highly attractive, providing a constant flow of earnings, so that banks have less incentive to expand credit to riskier private borrowers or cut their overheads (Hauner, 2006).6

Proponents of domestic debt stress its positive impact on growth, inflation and savings from deeper and more sophisticated capital markets, which enhance the volume and efficiency of private investment. Consequently, they question the wisdom of forever pursuing zero net domestic financing policies in countries where marketable domestic debt is already small and capital markets under-developed. Such policies, by reducing the size of domestic debt relative to GDP and deposits, could exert a negative impact on financial market development, and complicate the exit from foreign aid dependency.

Domestic debt markets can help strengthen money and financial markets, boost private savings and stimulate investment. First, government securities are a vital instrument for the conduct of indirect monetary policy operations and collateralized lending in interbank markets; the latter helps banks manage their own liquidity more effectively, reducing the need for frequent central bank interventions.7 Consequently, central banks operating in well-developed domestic debt markets do not have to rely as much on direct controls like credit ceilings, interest rate controls and high reserve requirements, all of which distort financial sector decisions and lead to financial disintermediation at the expense of private sector savings and investments (Gulde, Pattillo, and Christensen, 2006). Second, yields on government securities can serve as a pricing benchmark for long-term private debt issued by banks or enterprises and, hence, promote the development of a corporate bond market that boosts competition in the baking sector (Fabella and Madhur, 2003). Third, the availability of domestic debt instruments can provide savers with an attractive alternative to capital flight as well as lure back savings from the nonmonetary sector into the formal financial system (IMF, 2001). The possible benefits here can go beyond saving mobilization and extend to a reduction in the size of the black economy, widened tax base, increased financial depth, de-dollarization and improved perceptions of currency and country risk.

In addition to enhancing the volume of investment, domestic debt can also improve the efficiency of investment and help increase total factor productivity. Banks in many developing countries face an inherently risky and unpredictable business environment, which makes them reluctant to engage with the private sector. As a result, banks play only a very limited role in providing longer-term financing to important strategic sectors, such as agriculture and manufacturing, and prefer instead to finance consumption-related trade activities (in the case of Africa, see Gulde, Pattillo, and Christensen, 2006). In providing banks with a steady and safe source of income, holdings of government securities may serve as collateral and encourage lending to riskier sectors. In other words, holdings of government debt may compensate for the lack of strong legal and corporate environments (Kumhof, 2004; Kumhof and Tanner, 2005). The collateral function of domestic debt may be particularly important when bank overheads cannot be reduced further, and lending risks remain high due to asymmetric information and/or weak contract enforcement (including of foreclosure laws).8

Third, in the longer term, nominal debt contracts enhance political accountability and help governments build a track record to access international capital markets. Increasing the reliance on domestic financing may help mitigate the problems of external borrowing, which has been found to crowd out domestic institutions by weakening the state’s dependence on its citizenry and hence severing the accountability channel that forces domestic institutional reform (Abbas, 2005; Moss, Pettersson, and de Walle, 2006). Furthermore, developing a track record may promote access to international financial markets. Research shows that countries that have successfully issued sovereign bonds on international markets have typically had a long prior experience with issuing domestic government bonds in their own markets (Kahn, 2005).

Empirical Survey

Studies on domestic debt have been constrained by a lack of reliable data, especially time series data for a large enough panel of countries. Fry’s (1997) is the only panel study on the impact of alternative deficit-financing strategies on economic growth in LICs and EMs. For more than 66 LICs and EMs over 1979–93, Fry finds market-based domestic debt issuance to be the least costly method of financing the budget deficit, as opposed to external borrowing, seigniorage and financial repression, all of which are eventually seen to stifle growth, reduce domestic saving and fuel inflation. Indeed, the real question, according to him, is not “whether” countries should switch to market based domestic financing, but “how” they should do so.

Several studies have examined the impact of domestic financing on bank efficiency and private sector lending. Using bank-level data on 73 middle-income countries over the post-1990 period, Hauner (2006) finds that banks, which allocate more credit to the government, are more profitable, but less efficient. However, applying aggregate country-level data on commercial bank holdings of domestic debt, the results are mixed: domestic debt only begins to harm financial development at very high levels. As Hauner’s sample excludes sub-Saharan Africa and other poorer LICs, which typically have low domestic debt, his results are already somewhat biased towards finding a low residual domestic debt capacity. Furthermore, the study does not take into account the fact that the extent to which banks can “sit on” government bond interest income or “pass them on” to depositors and borrowing firms depends on the nature of competition in the financial sector.

Moreover, Hauner (2006) does not consider the possibility that banks’ decision to hold domestic debt may be economically efficient from a risk-diversification perspective. For instance if in the long term, banks’ real return on private lending were negatively correlated with their income from government securities, the overall risk of the bank portfolio would fall through a risk-diversification effect. This will lower depositors’ required return, enabling banks to lower their lending rate for any given intermediation margin. Abbas (forthcoming) demonstrates the theoretical plausibility and empirical support for this negative correlation between private and government returns.9

Empirical evidence on the crowding out effects of domestic debt at the macroeconomic level is mixed. In a study of the determinants of financial depth, such as loans and deposits scaled to GDP, Detragiache, Tressel, and Gupta (2005) include government domestic interest payments as a proxy for domestic debt in 82 LICs and EMs over 1990–2001. The coefficient on interest payments is found to be significantly negative, although not robustly so in regressions of bank assets scaled to GDP, thereby suggesting a standard crowding out effect, at first glance. However, domestic interest payments enter the loans to GDP and deposits to GDP regressions positively, significantly and robustly, suggesting a crowding in effect in line with Kumhof and Tanner’s (2005) collateral argument.

IMF (2005) explicitly examines the impact of domestic debt on private sector credit in the context of 40 LICs (including 15 mature stabilizers) over 1993–2002.10 Overall, the study finds “limited evidence of government recourse to domestic financing crowding out private sector borrowing in the mature stabilizers” (IMF, 2005, p. 34). Higher “levels” of domestic debt are found to be associated with lower levels of corporate lending, but the relationship breaks down when first differences of the variables are used. The report also finds no robust evidence of a negative correlation between real treasury bill rates and changes in domestic debt for either the mature stabilizers or the broader LIC group. However, the report notes that crowding out may occur through channels other than interest rates, such as credit rationing, and cautions against a rapid buildup in domestic debt, especially in the context of the availability of concessionary external financing.

Testable Hypotheses

The foregoing suggests a complex cost-benefit calculus for domestic debt and a series of plausible Hypotheses 1–3 must be tested in order to unravel it. For instance:

  • Hypothesis 1: Domestic debt may have either a positive or negative net impact on growth. Furthermore, the impact of domestic debt on growth may well be “nonlinear.”

Domestic debt could both have positive and negative effects on economic growth. These contrary views may be bridged by the existence of a nonlinear impact of debt: at moderate levels, domestic debt boosts growth but beyond a certain level, more traditional crowding out concerns may dominate. Hence, it may be important to identify whether such threshold exists, which could help evaluate whether debt in a given country has reached inappropriate levels.

  • Hypothesis 2: The macroeconomic impact of domestic debt may work primarily through the investment efficiency channel rather than capital accumulation.

Domestic debt may both boost the pool of savings and enhance the volume of investment in the economy. In addition, the positive spillover effects from domestic debt markets to broader capital markets may promote more risk taking and support better allocation of capital to productive sectors. If these sectors have been underfinanced in the past, domestic debt markets will help raise total factor productivity and expand the economy’s production frontier.

  • Hypothesis 3: The institutional environment as well as the quality and span of domestic debt markets may have a significant bearing on the growth impact of domestic debt.

The institutional environment could have a complex interaction with domestic debt. On one hand, better institutions can imply a competent policy framework, featuring optimal use of fiscal resources for the provision of public services, infrastructure development, maintenance of law and order, and property rights protection. This would tend to make the growth impact of domestic debt, or any source of budgetary finance for that matter, higher. On the other hand, domestic debt markets may be less important in stable institutional environments as the collateral or risk-diversification function that domestic debt performs on banks’ balance sheets will play a smaller role. Furthermore, the risk-diversification benefit will become less important as the overall magnitude of risk in the economy falls. A priori, it is not clear which of the two effects dominates: that is, whether good institutions complement domestic debt for public service provision, or whether they substitute for the collateral and risk-underwriting functions that domestic debt performs on banks’ balance sheets.

The quality and breadth of domestic debt markets should also be important. It is relevant to investigate how (1) the composition of domestic debt in terms of arrears and overdrafts vs. auctioned securities (marketable treasury bills and bonds) and (2) the holding of domestic debt in terms of banks vs. nonbank sectors, affect the growth impact of domestic debt. Indeed, many of the benefits of domestic debt discussed above: safe asset and collateral functions, monetary policy and liquidity management benefits, and benchmark yield curve for private lending, all clearly apply to securitized domestic debt and not to debt issued in captive markets or accumulated due to fiscal irresponsibility.

Commercial bank holdings of domestic debt are likely to be associated with lower financial system efficiency and greater crowding out, than when debt is held by the nonbank sector.11 As indicated in Christensen (2004) and Gulde, Pattillo, and Christensen (2006), the ability of LICs, in particular in sub-Saharan Africa, to expand domestic debt without crowding out bank lending to the private sector partly depends on the importance of the contractual savings sector (pension and insurance companies, and so on). The participation by individual and contractual savings institutions in the government securities market boosts competition in the financial sector alleviating some of the concerns by Hauner (2006).12

II. Data and Econometric Framework

New Domestic Debt Database

As mentioned earlier, reliable domestic debt data has been, and still is, a serious problem in LICs and some EMs. There are only a handful of LICs that maintain and report public domestic debt data in an organized and regular manner. Even among this small subset, regular reporting has been instituted only recently and consistent time series on domestic debt are not available for a decent stretch of time. The absence of such data has also effectively precluded, in our view, serious research on domestic debt, the consequent emergence of a “total” public debt (that is, domestic + external) view on debt management and fiscal policy, and an understanding of how debt structure choices are affected by and affect macro, fiscal, financial and institutional variables.

Researchers at IMF have recently attempted to collect domestic debt data on subsets of LICs and EMs.13Christensen (2004) collected annual data on central government (CG) domestic securities from 1980–2000 on 26 sub-Saharan African countries; however the data has many gaps, effectively covering only 20 countries. Mellor and Guscina (forthcoming) collected securitized and unsecuritized CG domestic debt on 70 IMF “program” countries from 1996–2004. The data is also usefully disaggregated by holder (banking system vs. nonbank sectors) and securitized vs. unsecuritized. A third database introduced by Jeanne and Guscina (2006) compiles securitized CG securities on 19 EMs, disaggregated by maturity and currency since 1980. We make selective use of the Mellor and Guscina and Christensen databases in this paper.

Our main data source, however, is Abbas (2007), who extracts from the IFS monetary survey data a domestic debt series spanning 93 LICs and EMs over 30 years (1975–2004).14 The definition used is commercial banks’ gross claims on the CG plus central bank liquidity paper.15 The series is then scaled to both GDP (DOMdebt) and commercial bank deposits (DD2dep). Table 1 provides a list of all the countries in the sample as well as a country-by-country breakdown of the evolution of these ratios over three decades, starting in 1975.

Table 1.Public Domestic Debt Trends in Low-Income Countries and Emerging Markets, 1975–2004
Domestic Debt/GDP (Percent)Domestic Debt/Deposits (Percent)
1975–861987–951996–20041975–20041975–861987–951996–20041975–2004
Sub-Saharan Africa Average3.23.95.74.122.420.823.322.2
Benin0.52.51.81.52.914.611.18.9
Botswana2.94.36.44.415.427.026.522.2
Burkina Faso2.21.81.31.824.215.411.317.7
Burundi0.61.11.71.19.110.213.310.7
Cameroon1.93.42.52.512.526.123.619.9
Cape Verde19.021.620.335.638.537.0
Central African Rep2.50.81.01.636.219.426.328.2
Chad0.10.80.50.41.420.512.010.3
Comoros1.00.50.30.720.48.36.112.5
Congo, Dem.0.20.10.30.28.53.19.27.1
Rep
Congo, Rep.4.64.41.43.638.438.319.532.7
Cote d’Ivoire0.84.84.73.2
Djibouti2.00.91.43.71.81.7
Ethiopia4.811.211.28.631.350.729.736.7
Gambia, The2.95.211.56.219.029.846.230.4
Ghana2.73.49.85.028.626.454.935.8
Guinea0.21.71.04.124.414.3
Guinea-Bissau0.70.20.427.52.615.0
Kenya4.36.79.16.521.533.629.327.5
Lesotho8.87.56.97.824.722.128.425.0
Madagascar1.41.43.01.911.79.520.013.5
Malawi5.85.16.05.734.430.441.835.4
Mali0.40.71.00.75.06.66.96.1
Mauritania0.20.32.00.8
Mozambique4.60.53.33.022.95.613.915.0
Namibia1.84.63.25.912.79.3
Niger2.52.11.01.929.317.918.022.5
Nigeria6.23.34.54.835.122.128.929.4
Rwanda1.22.41.31.617.423.410.017.0
Senegal2.01.83.22.311.311.117.113.0
Seychelles6.532.869.733.325.176.779.456.9
Sierra Leone4.71.74.73.838.224.150.037.5
Sudan3.00.10.31.422.10.93.910.3
Swaziland2.00.81.11.47.12.95.75.4
Tanzania10.12.94.66.340.918.232.131.4
Togo0.80.71.40.93.93.38.85.2
Tunisia5.65.44.35.217.614.09.614.1
Uganda1.50.35.02.223.83.837.221.8
Zambia8.66.75.17.040.536.929.736.2
Zimbabwe58.760.145.155.0
EM average7.88.514.310.025.825.632.327.7
Argentina5.98.112.38.532.063.025.839.4
Brazil6.18.634.115.254.325.480.753.5
Chile14.115.110.713.423.349.929.133.0
China1.08.64.83.919.111.5
Colombia1.70.95.52.613.06.724.314.5
Cyprus5.812.020.312.09.516.218.714.3
India6.310.217.210.723.027.936.128.4
Indonesia0.21.918.76.31.75.741.614.9
Korea, Rep.5.85.49.06.619.916.115.117.3
Malaysia31.918.28.920.973.430.910.241.7
Mauritius10.019.519.215.626.431.324.727.4
Mexico10.412.718.813.637.536.771.347.4
Pakistan7.49.810.89.234.245.337.238.5
Philippines3.57.213.57.617.123.225.621.5
South Africa6.84.76.06.014.59.710.811.9
Thailand6.24.24.65.117.17.15.110.5
Turkey3.25.724.410.315.535.973.238.9
All LICs5.55.68.46.421.420.223.721.7
OTHER average6.96.18.77.218.117.120.118.4
Algeria4.58.813.38.413.527.341.125.9
Bahrain3.39.111.67.66.514.616.912.1
Bangladesh2.13.25.33.315.515.018.316.2
Bolivia0.33.61.91.09.85.4
Cambodia0.00.10.10.60.70.7
Costa Rica4.32.48.44.914.07.924.515.3
Dominican Rep3.31.24.12.919.86.513.513.9
Ecuador0.10.32.81.01.11.915.65.7
Egypt, Arab Rep15.517.118.416.945.424.326.033.3
El Salvador1.63.57.13.84.810.517.010.2
Fiji4.65.55.15.016.312.113.214.1
Guatemala1.42.22.52.07.812.713.611.0
Guyana46.128.322.433.748.950.438.346.2
Haiti0.40.12.50.92.80.78.03.7
Honduras5.46.91.24.627.930.03.521.2
Iran, Islamic Rep7.03.31.24.120.17.13.511.2
Jamaica8.88.415.010.626.923.240.830.0
Lao PDR1.91.01.4
Lebanon41.432.277.049.312.524.343.525.3
Libya7.517.38.310.731.142.822.732.1
Morocco11.519.518.115.943.750.929.341.5
Myanmar
Nepal2.75.15.84.317.422.116.218.5
Nicaragua2.60.110.84.310.50.430.913.6
Panama7.72.41.54.325.46.32.312.7
Papua New G.3.86.911.97.217.323.441.226.3
Paraguay0.20.22.60.91.20.910.53.9
Peru7.44.12.54.945.528.010.129.6
Sri Lanka1.83.66.13.69.314.817.913.5
Suriname0.50.00.00.2
Syrian Arab Rep2.53.99.24.910.814.521.815.2
Trinidad & Tobago3.25.38.75.59.913.120.414.0
Uruguay5.15.15.75.313.112.211.512.3
Venezuela, RB2.04.95.03.86.619.130.817.6
Vietnam1.52.52.045.429.337.4
Yemen, Rep.0.24.12.10.921.411.1
Note: (i) The white lines separate the three groups of low-income countries (LICs): Sub-Saharan African excluding South Africa), OTHER (some Asian, North African, Middle Eastern and Latin American LICs) and EM (emerging markets).(ii) “Domestic debt”=[Banking sector’s claims on central government (IFS 22a + 42a) + securitized claims on central bank (IFS 20c + 40c)] divided by GDP at current market prices (IFS line item 99b).(iii) “Deposits” include current, time, fcy and saving deposits.
Note: (i) The white lines separate the three groups of low-income countries (LICs): Sub-Saharan African excluding South Africa), OTHER (some Asian, North African, Middle Eastern and Latin American LICs) and EM (emerging markets).(ii) “Domestic debt”=[Banking sector’s claims on central government (IFS 22a + 42a) + securitized claims on central bank (IFS 20c + 40c)] divided by GDP at current market prices (IFS line item 99b).(iii) “Deposits” include current, time, fcy and saving deposits.

Domestic debt, as a share of GDP, appears to have risen over time from 5.5 percent to 8.4 percent, but, as a share of deposits, has remained relatively stable around 21.5 percent. The ratios for both DOMdebt and domestic debt/deposits ratio (DD2dep) are higher in EMs at 10 percent and 27.7 percent, respectively, compared with sub-Saharan Africa’s 4.1 percent and 22.2 percent, respectively. As substantial scope for economic expansion and financial deepening remains in sub-Saharan Africa, the implied domestic debt issuance capacity may be significant, going forward. Also, although the distribution of key domestic debt ratios across sub-Saharan Africa countries has been stable over time, the same in EMs has become significantly more dispersed, indicating increased heterogeneity among EMs; Figure 1 compares the ratios for the pre- and post-1990 periods for both sub-Saharan Africa countries and EMs.16

Figure 1.Public Domestic Debt Trends, 1975–1989 versus 1990–2004

(dashed line)

Note: Outliers removed.

Controls and Causality Variables

For Granger-causality regressions to investigate the endogeniety of domestic debt and the channels through which it may affect the economy, we use the following variables: per capita income, private savings rate, institutions and financial development. As reliable series on the latter two variables do not exist in international financial institution databanks, we invoke suitable proxies. For institutions, we use the International Country Risk Guide (ICRG) “composite index,” which tracks countries’ political economic and financial risks over time; the index rises as the risk reduces and stability increases. The series runs from 1986 onwards, is available for most of the 93 countries and is denoted as “stability.” Although we use it as a proxy for institutional quality, it can be construed as such only insofar as good institutions affect risk and stability.

To proxy for financial development, a “financial depth index” was constructed using the approach in Huang and Temple (2005). The index was developed from three underlying series—liquid liabilities of the financial system; private sector credit; and commercial bank share in banking system assets, using principal component analysis techniques (see Appendix II for details on extraction methodology).17 With private sector credit included as an integral part of the index, the latter’s response to domestic debt would also shed light on any crowding out effects.

Our control variables for the growth regressions are similar to those used in Pattillo, Poirson, and Ricci (2002): lagged income, population growth, investment, budget balance, openness to trade and terms of trade growth (gTOT) and the additional controls, inflation and external debt.18 The summary statistics, definitions and correlation matrix for the main regression variables are presented in Appendix I. Our choice and number of countries is similar to Pattillo, Poirson, and Ricci (2002): 93, but our time period is 1975–2004 divided into 10 three-year periods.

Econometric Specification for Growth Regressions

Preliminary support for the nonlinear relationship hypothesized (in Section I) between domestic debt and growth is provided by the scatter plots in Figure 2.19 The plots suggest that growth may have a Laffer curve relationship with DD2dep, and a linear relationship with DOMdebt. This appears realistic in low financial depth contexts, where the financial size of a country would place a more binding constraint on domestic debt capacity than economic size.

Figure 2.Growth—Domestic Debt Scatter Plots

(Country means over the period 1975–2004)

Note: Dashed line is linear prediction and unbroken line is quadratic prediction.

The empirical analysis is modeled on Pattillo, Poirson, and Ricci (2002), who investigate the nonlinear growth effects of external debt on a panel of 93 developing countries over the 1969–98 period, using FIBVE-year averaged data and a conditional convergence framework. Like them, we employ fixed effects (FE), system generalized method of moments (GMM) and pooled ordinary least square (OLS) regressions of purchasing power parity (PPP) per capita real income growth on linear and nonlinear debt terms and an elaborate set of controls.20 For Hypothesis 1 identified in Section I, we estimate the following two equations:

where g is growth in PPP GDP per capita, X is a vector of control variables, DOMdebt is the domestic debt/GDP ratio, αi captures country heterogeneity and φt are period dummies. Similar regressions are run for the DD2dep.

Our specification differs from that of Pattillo, Poirson, and Ricci (2002) in that we work with the actual domestic debt ratios (as opposed to logs) while using quartile dummy interaction variables (instead of squared terms) to study nonlinear growth effects.21 For the case of DOMdebt, the corresponding quartile dummy is named quart_DOM, and for DD2dep, quart_DD2dep. Our choice of specification is driven by the particular constraints posed by the domestic debt database. For instance, many of the domestic debt ratios in the sample—especially for LICs—were less than “1.00” (that is, less than 1 percent). This precluded taking logs (which would have produced negative values), or squaring the nonlogged ratios (as the squaring numbers less than 1.00 yields smaller not larger values). Moreover, given the high dispersion of the domestic debt ratios—see Appendix I—squaring the ratios would have increased outlier problems.

For Hypothesis 2 of Section I: whether domestic debt impacts growth through investment efficiency of capital accumulation, investment is removed from the control set and the difference in results from the with-investment regressions observed.

And finally, for Hypothesis 3: the extent to which the growth impact of domestic debt depends on the institutional environment in which the debt is issued, we have:

The regressions with attributes of domestic debt were similar in structure to the above, except that, instead of quart_STABILITY, the following ratios were used for interaction:

  • Share of securities in total domestic debt stock (SDD2DD) [range 0–100].

  • Share of domestic debt held by banking system (shBANK) [range 0–100].

  • A period dummy (ERA) taking the value of 0 for pre-1990 observations (corresponding roughly to financial repression years) and 1 for 1990 onwards (corresponding to the financial liberalization years).

  • Dummy variable (REALi) taking the value of 0 for observations where the real interest rate (deposit rate minus inflation rate) was zero or negative, and 1 when it was positive.

We also run regressions using Christensen’s (2004) domestic debt data on 20 sub-Saharan Africa countries over 1980–2000. This dataset covers CG securities (unsecuritized debts excluded) and both bank and nonbank holdings of this debt, enabling us to indirectly test if these features have the expected positive impact on any observed growth effect of domestic debt.

The priors on the coefficients of our control variable follow from the large number of empirical growth studies. GDP per capita growth should, in accordance with Solow’s convergence hypothesis, have a negative impact on growth. High inflation and population growth rates are also expected to undermine real economic growth. Robust empirical evidence (Elbadawi, Ndulu, and Ndung’u, 1997; Pattillo, Poirson, and Ricci, 2002) suggests that external debt impacts growth negatively. In contrast, gross fixed capital formation, fiscal balance (FISBAL), gTOT and openness should have positive effects on growth.

The primary attraction of using panel data methods in these crosscountry regressions is their ability to deal with time-invariant individual effects (αi). If the effects are random, we can use the random effects (RE) estimator for unbiased and efficient estimation. However, if the effects are fixed, or if they are correlated with the regressors, RE is inconsistent, and FE methods, which wipe out the individual heterogeneity altogether, must be employed to recover consistent estimates of β, γ and δ.

FE methods, however, are biased and inconsistent in dynamic panel data models of the type we are estimating. In particular, the coefficient on the lagged dependent variable (lnY_1) will be severely downward biased (numerically).22 Reverting to OLS and RE for estimating this coefficient is also unhelpful as both are severely upward biased, as discussed in Bond, Hoeffler, and Temple (2002). Secondly, FE models (like OLS or RE) cannot deal with endogenous regressors, a key concern in the present context. For these reasons, we rely, in the main, on system GMM estimation of our regressions, which can simultaneously address the problems of endogeneity and lagged dependent variable.

III. Empirical Results

Granger-Causality Tests on the Endogeneity of Domestic Debt

As a precursor to the growth regressions, we run a battery of Granger-causality panel regressions to study the extent to which domestic debt is endogenous to, or drives, income, private savings, institutions (politico-economic stability) and financial development. Although these tests are very widely used in a range of contexts, it must be acknowledged that they are judgments on statistical causality and may not necessarily imply economic causality. This disclaimer applies equally to the causality inferences derived below.

Appendix II details the econometric methodology underpinning the Granger causality regressions whereas Table 2 presents the results. The latter are also summarized in the causality map (Figure 3), and seem to suggest support for two-way statistical causality links between domestic debt and the other variables.23 Institutions are not causal, income and financial depth are weakly causal, but private savings are strongly causal for domestic debt. Evidence on reverse-causality suggests that domestic debt is an important explanatory variable for private savings and institutions and, to a lesser extent, for financial development and income. Overall, this appears to weaken the case for approaching domestic debt as a purely endogenous variable.24

Table 2.Granger-Causality Tests for Domestic Debt, Financial Depth, Stability, Income, and Saving
A. Financial Depth and Domestic DebtB. Stability (Institutions) and Domestic Debt
Dependent

variable
(1)

FINDEPTH
(2)

DOMdebt
(3)

DOMdebt
Dependent

variable
(1)

STABILITY
(2)

STABILITY
(3)

DOMdebt
RegressorsRegressors
DOMdebt_1–0.04331.0033**0.8648**DOMdebt_10.2474**0.2873**0.7709**
–1.567.107.803.242.585.69
DOMdebt_20.1272*–0.3748**–0.3851**DOMdebt_2–0.0297–0.0708–0.1835**
1.91–3.26–4.13–0.60–1.53–2.71
FINDEPTH_11.3448**0.3339**0.2547*STABILITY_10.7758**0.7532**0.0718
18.742.541.847.369.370.83
FINDEPTH_2–0.3551**–0.1523–0.1352STABILITY_2–0.1942**–0.2180**–0.0207
–2.93–1.54–1.21–3.21–3.30–0.20
INCOME_10.4055**INCOME_10.1926**
2.082.11
Hansen’s chi2 test (prob > chi2)0.3710.4320.174Hansen’s chi2 test (prob > chi2)0.9410.2230.619
AR(1) test (prob>z)0.006(−)0.006Π0.007(−)AR(1) test (prob>z)<0.001(−)0.001 (−)0.011 (−)
AR(2) test (prob > z)0.885 (−)0.713(+)0.366(+)AR(2) test (prob > z)0.491 (−)0.694(−)0.306(−)
Joint test for H0: β1 = β2 = 0 (prob > chi2)0.1580.019**0.143Joint test for H0: β1 = β2 = 0 (prob > chi2)0.003**0.034**0.701
H0: β1 + β2 = 0; prob > chi20.085*0.2350.426H0: β1 + β2 = 0; prob > chi20.023**0.029**0.699
βLR = (β1 + β2)/(1-α12)8.150.490.23βLR = (β1 + β2)/(1-α12)0.520.470.12
Roots; for stability2.77; 1.021.34±0.94i1.12±1.16iRoots; for stability2.00±1.08i1.73±1.26i2.10±1.02i
StableStableStableStableStableStable
Note: (i) INCOME is PPP GDP per capita from WEO; DOMdebt is 100 × domestic debt/GDP (see text); FINDEPTH index is developed using Beck, Demirgüç-Kunt, and Levine (2000; updated in 2006) data (see text); prSAVING rate is from IMF: WEO database; STABILITY is ICRG composite risk index capturing a country’s political, economic and financial risk (higher value denotes lower risk).(ii) z-statistics in italics, unless otherwise stated; constant and time dummies included in all regressions.(iii) **significant at 5 percent, *at 10 percent.(iv) Data spans 93 countries and 10 three-year time periods constructed from 30-year annual data (1975–2004).(v) GMM results reported here were obtained using the xtabond2 two-step command in STATA; z-statistics are based on heteroskedasticity-consistent errors and the finite-sample adjustment of Windmeijer (2000). AR(1) and AR(2) are tests for first-order and second-order serial correlation. First-order (negative) serial correlation is expected due to first-differencing, but identification of the models relies on the absence of second-order correlation. Hansen’s chi-squared tests the additional moment conditions used by the system GMM estimator.(vi) GMM instrumentation: for the difference equation, levels dated t−2 and t–3 are used as instruments; while for the moment conditions in levels, first differences dated t−1 are used as instruments.(vii) The joint test for significance of βs relates, in any given column, to the coefficients appearing in bold in that column.(viii) For regressions in panel C, variables enter in nonnormalized form.(ix) βLR captures the LR effect of the explanatory variable on the regressand; where βLR > 1, the beta and the associated β1 + β2 = 0 test are not reported; the coefficients on regression lags are denoted by α1 and α2.(x) Stability requires that the roots of the polynomial 1–α1L–α2L2 = 0 are outside the unit circle.(xi) The test H0: β1 + β2 = 0 checks if there is support for the long-run “levels” relationship (captured by βLR). Failure to reject the null implies β1 = –β2, so that Yit = β1(Xit–1Xit–2); in other words Y depends on changes in X, not the level of X; the relationship, therefore, is not long term but short term. A rejection of the null implies the opposite, and provides support for a long-run “levels” relationship.
C. Private Saving and Domestic DebtD. Income and Domestic DebtE: Financial Depth and Income
Dependent

variable
(1)

prSAVING
(2)

DOMdebt
Dependent

variable
(1)

INCOME
(2)

DOMdebt
Dependent

variable
(1)

FINDEPTH
(2)

INCOME
RegressorsRegressorsRegressors
DOMdebt_11.0631**0.1518**DOMdebt_10.0671**0.7436**INCOME_10.2998**1.4868**
11.073.032.456.772.3126.53
DOMdebt_2–0.3140.0417DOMdebt_20.0195–0.141**INCOME_2–0.3806–0.4857**
–4.850.740.98–2.31–0.96–6.90
prSAVING_10.1430*0.6400**INCOME_11.3568**0.2920*FINDEPTH_11.3435**0.0009
1.697.316.131.6518.610.05
prSAVING_20.0638*0.0116INCOME_2–0.3838**–0.0044FINDEPTH_2–0.5504**0.0038
1.650.25–3.74–0.03–4.460.21
Hansen’s chi2 test (prob > chi2)0.5280.778Hansen’s chi2 test (prob > chi2)0.075*0.344Hansen’s chi2 test (prob > chi2)0.4140.316
AR(1) test (prob>z)0.004(−)0.005(−)AR(1) test (prob>z)<0.001(−)0.002(−)AR(1) test (prob>z)0.009(−)<0.001(−)
AR(2) test (prob > z)0.399(−)0.372(−)AR(2) test (prob > z)0.659(−)0.394(−)AR(2) test (prob > z)0.253(+)0.373(+)
Joint test for H0: β1 = β2 = 0 (prob > chi2)0.089*0.003**Joint test for H0: β1 = β2 = 0 (prob > chi2)0.048**0.098*Joint test for H0: β1 = β2 = 0 (prob > chi2)0.059*0.858
H0: β1 + β2 = 0; prob > chi20.035**0.003**H0: β1 + β2 = 0; prob > chi20.031**0.04**
βLR = (β1 + β2)/(1–α1–α2)0.820.56βLR = (β1 + β2)/(1–α1–α2)3.2070.72
Roots; for stability1.69±0.57i1.52; -56.69Roots; for stability2.49; 1.052.64±0.38i
StableStableStableStable
Note: (i) INCOME is PPP GDP per capita from WEO; DOMdebt is 100 × domestic debt/GDP (see text); FINDEPTH index is developed using Beck, Demirgüç-Kunt, and Levine (2000; updated in 2006) data (see text); prSAVING rate is from IMF: WEO database; STABILITY is ICRG composite risk index capturing a country’s political, economic and financial risk (higher value denotes lower risk).(ii) z-statistics in italics, unless otherwise stated; constant and time dummies included in all regressions.(iii) **significant at 5 percent, *at 10 percent.(iv) Data spans 93 countries and 10 three-year time periods constructed from 30-year annual data (1975–2004).(v) GMM results reported here were obtained using the xtabond2 two-step command in STATA; z-statistics are based on heteroskedasticity-consistent errors and the finite-sample adjustment of Windmeijer (2000). AR(1) and AR(2) are tests for first-order and second-order serial correlation. First-order (negative) serial correlation is expected due to first-differencing, but identification of the models relies on the absence of second-order correlation. Hansen’s chi-squared tests the additional moment conditions used by the system GMM estimator.(vi) GMM instrumentation: for the difference equation, levels dated t−2 and t–3 are used as instruments; while for the moment conditions in levels, first differences dated t−1 are used as instruments.(vii) The joint test for significance of βs relates, in any given column, to the coefficients appearing in bold in that column.(viii) For regressions in panel C, variables enter in nonnormalized form.(ix) βLR captures the LR effect of the explanatory variable on the regressand; where βLR > 1, the beta and the associated β1 + β2 = 0 test are not reported; the coefficients on regression lags are denoted by α1 and α2.(x) Stability requires that the roots of the polynomial 1–α1L–α2L2 = 0 are outside the unit circle.(xi) The test H0: β1 + β2 = 0 checks if there is support for the long-run “levels” relationship (captured by βLR). Failure to reject the null implies β1 = –β2, so that Yit = β1(Xit–1Xit–2); in other words Y depends on changes in X, not the level of X; the relationship, therefore, is not long term but short term. A rejection of the null implies the opposite, and provides support for a long-run “levels” relationship.
Note: (i) INCOME is PPP GDP per capita from WEO; DOMdebt is 100 × domestic debt/GDP (see text); FINDEPTH index is developed using Beck, Demirgüç-Kunt, and Levine (2000; updated in 2006) data (see text); prSAVING rate is from IMF: WEO database; STABILITY is ICRG composite risk index capturing a country’s political, economic and financial risk (higher value denotes lower risk).(ii) z-statistics in italics, unless otherwise stated; constant and time dummies included in all regressions.(iii) **significant at 5 percent, *at 10 percent.(iv) Data spans 93 countries and 10 three-year time periods constructed from 30-year annual data (1975–2004).(v) GMM results reported here were obtained using the xtabond2 two-step command in STATA; z-statistics are based on heteroskedasticity-consistent errors and the finite-sample adjustment of Windmeijer (2000). AR(1) and AR(2) are tests for first-order and second-order serial correlation. First-order (negative) serial correlation is expected due to first-differencing, but identification of the models relies on the absence of second-order correlation. Hansen’s chi-squared tests the additional moment conditions used by the system GMM estimator.(vi) GMM instrumentation: for the difference equation, levels dated t−2 and t–3 are used as instruments; while for the moment conditions in levels, first differences dated t−1 are used as instruments.(vii) The joint test for significance of βs relates, in any given column, to the coefficients appearing in bold in that column.(viii) For regressions in panel C, variables enter in nonnormalized form.(ix) βLR captures the LR effect of the explanatory variable on the regressand; where βLR > 1, the beta and the associated β1 + β2 = 0 test are not reported; the coefficients on regression lags are denoted by α1 and α2.(x) Stability requires that the roots of the polynomial 1–α1L–α2L2 = 0 are outside the unit circle.(xi) The test H0: β1 + β2 = 0 checks if there is support for the long-run “levels” relationship (captured by βLR). Failure to reject the null implies β1 = –β2, so that Yit = β1(Xit–1Xit–2); in other words Y depends on changes in X, not the level of X; the relationship, therefore, is not long term but short term. A rejection of the null implies the opposite, and provides support for a long-run “levels” relationship.

Figure 3.Causality Map

Note: An arrow from X to Y implies that the null of “X does not Granger-cause Y” is rejected at the 5 percent level of significance; at 10 percent in case of broken arrow.

Domestic debt and private savings were found to be closely associated. Higher private savings increase the scope for domestic debt issuance while a larger supply of domestic debt instruments provides incentives to increase private savings. Strengthening and expanding domestic debt markets can, therefore, form a potentially virtuous cycle of higher private savings and stronger capital markets. Note that the size of the long-run marginal effect of domestic debt on private saving, βLR[=0.82] in panel C of Table 2, far exceeds the rather low estimates [around 0.5] for the Ricardian offset ratios floating in the savings literature, ruling out a pure Ricardian explanation for the positive association (see Masson, Bayoumi, and Samiei, 1998).

Domestic debt was found to weakly and positively Granger-cause financial depth positively. However, financial depth had a surprisingly weak causal contribution to income (Table 2, panel E), which seems at odds with other empirical studies that find a significant impact of financial development on economic growth.25 The inconsistency can be resolved, partly, by noting that financial depth—which is highly sensitive to short-term credit and deposit booms—is only a crude proxy for financial development, which may be regarded as a “longer-term” concept. In that context, insofar as expanding domestic debt, markets are also a long-term phenomenon (especially when measured in relation to GDP), domestic debt can serve as a better proxy for financial development than financial depth. Indeed, the financial development dataset of Beck, Demirgüç-Kunt, and Levine (2006) includes both data on financial depth and local bond market capitalization. To the extent that the latter is driven primarily by outstanding government bonds in LICs and EMs—see World Bank (2006, Figure 2)—this seems like an implicit acknowledgement that the development of domestic debt markets is, in and by itself, an integral part of the process of financial development.26

The Behavior of Control Variables in Growth Regressions on Domestic Debt

Coefficient signs and magnitudes for the control variables all appear to be empirically plausible, and broadly in line with our stated priors (Tables 3 and 4).

Table 3.Growth Regressions on DOMdebt (Domestic Debt/GDP)Dependent variable: per capita income growth (gY=100 × growth rate in PPP per capita income)
A. LINEARB. NONLINEAR

(Using Interaction of DOMdebt with a DOMdebt Quartile Dum
(1) OLS(2) RE(3) FE(4) GMM-sys(5) OLS(6) RE(7) FE(8) GMM-sys
lnY_1–0.8275**–1.0733**–7.2849**–2.3040**lnY_1–0.8253**–1.0574**–6.3615**–1.9714**
–4.17–4.43–8.86–2.49–4.11–4.31–9.00–3.24
gPOP–0.5021**–0.4342**–0.1671–0.6990gPOP–0.4825**–0.4396**–0.1888–0.5526
–3.36–2.71–0.91–1.33–3.17–2.73–1.07–1.15
INFLATION–0.0237**–0.0236**–0.0005–0.0156INFLATION–0.0220**–0.0233**–0.0317**–0.0092
–2.47–2.26–1.55–0.33–2.25–2.21–2.59–0.22
lnINVEST3.4373**3.3271**2.0932**5.7036**lnINVEST3.3317**3.3219**2.4884**3.7973**
9.007.903.533.118.657.874.832.99
FISBAL0.1159**0.1305**0.1394**0.1611*FISBAL0.1126**0.1265**0.1032**0.1575**
3.774.023.731.833.543.822.912.43
gTOT–0.0015–0.0031–0.0063–0.0114gTOT–0.0003–0.0034–0.00500.0024
–0.11–0.23–0.46–0.61–0.02–0.25–0.390.12
lnOPEN0.04530.28513.2017**2.2311*lnOPEN0.05680.28552.6352**2.1346**
0.160.834.631.820.200.834.262.15
EXTdebt–0.0091**–0.0095**–0.0226**0.0063EXTdebt–0.0089**–0.0093**–0.0140**–0.0079
–3.38–3.08–4.570.51–3.24–3.02–3.06–1.54
DOMdebt0.0406**0.0724**0.0637*0.0742*DOMdebt0.0909**0.0894**0.0667*0.0981*
2.222.951.731.772.402.411.791.93
DOMdebt × (quart_DOM)–0.0075–0.0070–0.0155–0.0054
–0.65–0.62–1.36–0.25
R2 overall0.420.420.070.203Hansen’s χ2: prob >χ2R2 overall0.420.410.070.656Hansen’s χ2: prob >χ2
R2 within0.410.51<0.001(−)A-Bond AR(1): prob > zR2 within0.410.48<0.001(−)A-Bond AR(1): prob > z
R2 between0.47<0.010.329A-Bond AR(2): prob > zR2 between0.46<0.010.277A-Bond AR(2): prob > z
Note: (i) t-statistics (FE and OLS) and z-statistics (RE and GMM) in italics.(ii) Constant; time dummies included in all regressions.(iii) **significant at 5 percent, *at 10 percent.(iv) Data spans 93 countries and 10 three-year time periods constructed from 30-year annual data (1975–2004).(v) Results obtained using STATA’s reg (OLS), xtreg (RE; FE) and xtabond2 (GMM-system; two-step) commands.(vi) Domestic debt regressors are lagged one period.(vii) z-statistics for GMM are heteroskedasticity-consistent.(viii) Arellano-Bond AR(1) and AR(2) tests are for first- and second-order serial correlation in errors. First-order negative serial correlation is expected due to first-differencing; model identification requires absence of second-order correlation.(ix) Hansen’s chi-squared test checks if the moment conditions used by the system GMM estimator are valid.(x) GMM instrumentation: (a) gTOT and gPOP were assumed exogenous, while the domestic debt variables, lnINVEST, lnOPEN and FISBAL were treated as endogenous; (b) instruments used for the difference equation: Xt–2, where X denotes an endogenous variable; (c) additional instruments used for the levels equation: Xt–1, where X is an endogenous variable.(xi) “quart_DOM” dummy = 0 if DOMdebt falls in 1st quartile, 1 if DOMdebt falls in 2nd quartile, 2 if DOMdebt falls in 3rd quartile, and 3 if DOMdebt falls in the 4th quartile; relevant DOMdebt percentiles are: p25 = 1.26 percent, p50 = 4.61 percent and p75 = 7.36 percent.relevant DOMdebt percentiles are: p25= 1.26 percent, p50 = 4.61 percent and p75 = 7.36 percent.The key coefficients are denoted in bold.
Note: (i) t-statistics (FE and OLS) and z-statistics (RE and GMM) in italics.(ii) Constant; time dummies included in all regressions.(iii) **significant at 5 percent, *at 10 percent.(iv) Data spans 93 countries and 10 three-year time periods constructed from 30-year annual data (1975–2004).(v) Results obtained using STATA’s reg (OLS), xtreg (RE; FE) and xtabond2 (GMM-system; two-step) commands.(vi) Domestic debt regressors are lagged one period.(vii) z-statistics for GMM are heteroskedasticity-consistent.(viii) Arellano-Bond AR(1) and AR(2) tests are for first- and second-order serial correlation in errors. First-order negative serial correlation is expected due to first-differencing; model identification requires absence of second-order correlation.(ix) Hansen’s chi-squared test checks if the moment conditions used by the system GMM estimator are valid.(x) GMM instrumentation: (a) gTOT and gPOP were assumed exogenous, while the domestic debt variables, lnINVEST, lnOPEN and FISBAL were treated as endogenous; (b) instruments used for the difference equation: Xt–2, where X denotes an endogenous variable; (c) additional instruments used for the levels equation: Xt–1, where X is an endogenous variable.(xi) “quart_DOM” dummy = 0 if DOMdebt falls in 1st quartile, 1 if DOMdebt falls in 2nd quartile, 2 if DOMdebt falls in 3rd quartile, and 3 if DOMdebt falls in the 4th quartile; relevant DOMdebt percentiles are: p25 = 1.26 percent, p50 = 4.61 percent and p75 = 7.36 percent.relevant DOMdebt percentiles are: p25= 1.26 percent, p50 = 4.61 percent and p75 = 7.36 percent.The key coefficients are denoted in bold.
Table 4.Growth Regressions on Domestic Debt to Deposits Ratio [DD2dep: Domestic Debt/(Current, Time, and Saving Deposits)]Dependent variable: per capita income growth (gY= 100 × growth rate in PPP per capita income)
A. LINEARB. NONLINEAR(Using Interaction of DD2dep with a DD2dep Quartile Dummy)
(1) OLS(2) RE(3) FE(4) GMM-sys(5) OLS(6) RE(7) FE(8) GMM-sys
lnY_1–0.8382**–1.0932**–6.6924**–1.8499**lnY_1–0.8415**–1.1021**–6.6446**–1.7925**
–4.27–4.55–9.58–2.42–4.19–4.48–9.35–2.24
gPOP–0.5338**–0.4788**–0.1802–0.4083gPOP–0.5044**–0.4517**–0.1460–0.4699
–3.52–2.98–1.01–1.12–3.3–2.78–0.81–1.34
INFLATION–0.0289**–0.0291**–0.0339**–0.0500INFLATION–0.0256**–0.0272**–0.0387**0.0023
–2.95–2.74–2.73–1.47–2.54–2.5–3.000.05
lnINVEST3.6083**3.6611**2.9434**5.5998**lnINVEST3.6396**3.6777**2.8784**6.7292**
9.338.635.783.449.368.65.594.73
FISBAL0.1145**0.1254**0.1003**0.0924FISBAL0.1018**0.1116**0.0754**0.2407**
3.643.842.891.333.123.292.082.58
gTOT0.00250.0003–0.00010.0013gTOT0.00350.0009–0.0037–0.0052
0.180.02–0.010.080.250.07–0.30–0.27
lnOPEN0.05810.25982.3124**0.8511lnOPEN0.06010.27682.4235**1.5334
0.200.763.810.640.210.83.920.93
EXTdebt–0.0095**–0.0099**–0.0142**0.0028EXTdebt–0.0100**–0.0104**–0.0131**0.0019
–3.48–3.20–3.090.2–3.60–3.33–2.750.19
DD2dep0.0146**0.0133*0.00420.0322**DD2dep0.0280**0.0258**0.0188*0.0513**
2.031.710.472.122.562.381.702.13
DD2dep*(quart_DD2dep)–0.0056*–0.0056*–0.0067**–0.0151*
–1.67–1.68–1.97–1.65
Maxima @DD2dep percentileout-of-sampleout-of-sample83rd97th
@ DD2dep ratio-do--do-35.71%65.40%
R2 overall0.420.420.070.215Hansen’s χ2: prob >χ2R2 overall0.420.420.060.294Hansen’s χ2: prob >χ2
R2 within0.410.49<0.001(−)A-Bond AR(1): prob > zR2 within0.420.47<0.001(−)A-Bond AR(1): prob > z
R2 between0.46<0.010.440A-Bond AR(2): prob > zR2 between0.44<0.010.469A-Bond AR(2): prob > z
Note: (i) t-statistics (FE and OLS) and z-statistics (RE and GMM) in italics.(ii) Constant; time dummies included in all regressions.(iii) **significant at 5 percent, *at 10 percent.(iv) Data spans 93 countries and 10 three-year time periods constructed from 30-year annual data (1975–2004).(v) Results obtained using STATA’s reg (OLS), xtreg (RE; FE) and xtabond2 (GMM-system; two-step) commands.(vi) Domestic debt regressors are lagged one period.(vii) z-statistics for GMM are heteroskedasticity-consistent.(viii) Arellano-Bond AR(1) and AR(2) tests are for first- and second-order serial correlation in errors. First-order negative serial correlation is expected due to first-differencing; model identification requires absence of second-order correlation.(ix) Hansen’s chi-squared test checks if the moment conditions used by the system GMM estimator are valid.(x) GMM instrumentation: (a) gTOT and gPOP were assumed exogenous, while the domestic debt variables, lnINVEST, lnOPEN and FISBAL were treated as endogenous; (b) instruments used for the difference equation: Xt–2, where X denotes an endogenous variable; (c) additional instruments used for the levels equation: Xt–1, where X is an endogenous variable.(xi) “quart_DD2dep” dummy = 0 if DD2dep falls in 1st quartile, 1 if DD2dep falls in 2nd quartile, 2 if DD2dep falls in 3rd quartile, and 3 if DD2dep falls in the 4th quartile; relevant DD2dep percentiles are: p25 = 8.54 percent; p50 = 17.43 percent; p75 = 29.93 percent.The key coefficients are denoted in bold.
Note: (i) t-statistics (FE and OLS) and z-statistics (RE and GMM) in italics.(ii) Constant; time dummies included in all regressions.(iii) **significant at 5 percent, *at 10 percent.(iv) Data spans 93 countries and 10 three-year time periods constructed from 30-year annual data (1975–2004).(v) Results obtained using STATA’s reg (OLS), xtreg (RE; FE) and xtabond2 (GMM-system; two-step) commands.(vi) Domestic debt regressors are lagged one period.(vii) z-statistics for GMM are heteroskedasticity-consistent.(viii) Arellano-Bond AR(1) and AR(2) tests are for first- and second-order serial correlation in errors. First-order negative serial correlation is expected due to first-differencing; model identification requires absence of second-order correlation.(ix) Hansen’s chi-squared test checks if the moment conditions used by the system GMM estimator are valid.(x) GMM instrumentation: (a) gTOT and gPOP were assumed exogenous, while the domestic debt variables, lnINVEST, lnOPEN and FISBAL were treated as endogenous; (b) instruments used for the difference equation: Xt–2, where X denotes an endogenous variable; (c) additional instruments used for the levels equation: Xt–1, where X is an endogenous variable.(xi) “quart_DD2dep” dummy = 0 if DD2dep falls in 1st quartile, 1 if DD2dep falls in 2nd quartile, 2 if DD2dep falls in 3rd quartile, and 3 if DD2dep falls in the 4th quartile; relevant DD2dep percentiles are: p25 = 8.54 percent; p50 = 17.43 percent; p75 = 29.93 percent.The key coefficients are denoted in bold.

For lagged income (lnY_1) and population growth (gPOP), a 1 percentage point rise corresponds to a decrease in per capita growth of about ½ of a percentage point. The INFLATION coefficient is negative and significant (although not in all regressions), confirming conventional wisdom that low inflation is a pre-condition for lasting growth. Gross fixed capital formation or investment (lnINVEST) is highly significant in all regressions and has the usual high semi-elasticity of around 3.5—similar to Pattillo, Poirson, and Ricci (2002), but significantly higher than the range of 2.1 to 2.2 of Mankiw, Romer, and Weil (1992). The coefficient on FISBAL is, expectedly, significant and positive in all regressions hovering in the 0.1–0.2 range, and virtually identical to the range found in Pattillo, Poirson, and Ricci (2002). The benefits of fiscal austerity underlined here will inform the policy implications on domestic debt derived in Section IV. Finally, the results on EXTdebt are also in line with expectations to the extent that the sign on all significant coefficients thereof is negative—consistent with the finding in Pattillo, Poirson, and Ricci (2002).27

The Growth Impact and Optimal Size of Domestic Debt

Results here suggest broad support for a positive overall contribution of “moderate” domestic debt levels to economic growth. For the first domestic debt ratio measured in percent of GDP, DOMdebt, we find a positive significant linear coefficient. The range for the coefficient was between 0.04 with OLS, 0.06 for FE and 0.07 for RE and GMM (Table 3, panel A). Taking 0.06 as the average, increasing domestic debt by one standard deviation (9.70 percent), implies an increase in the growth rate by 0.58 percentage points (0.13 standard deviations). The nonlinear specification in this case (Table 3, panel B) neither adds to overall explanatory power, nor throws up significant nonlinear effects. That said, the linear coefficients are higher and the sign on the DOMdebt × (quart_DOM) term consistently negative in all four regressions, so that the possibility of a Laffer curve relationship between growth and domestic debt, perhaps in a slightly richer group of countries, cannot be ruled out. In the current sample, however, with the fourth DOMdebt quartile beginning at 7.36 percent, there does not appear to be any evidence of diminishing returns to public domestic debt.

However, regressions with DD2dep suggest a nonlinear growth impact. The linear specification in Table 4, panel A produces positive and significant coefficients for DD2dep in all but the FE regression, with the coefficients being smaller than those obtained for DOMdebt. This may partly be because of stronger in-sample nonlinearities in the growth-DD2dep relationship compared with DOMdebt, in line with the scatter plots in Figure 2. The results on the nonlinear specification (Table 4, panel B) do indeed suggest support for this hypothesis. The coefficient of the linear DD2dep term strengthens noticeably compared with its counterpart in Table 4, panel A, whereas the nonlinear interaction term DD2dep × quart_DD2dep is negative and significant in all regressions. The turning points, or growth-maximizing levels of DD2dep, in the OLS and RE regressions are out-of-sample, but for the FE and GMM regressions are 35.7 and 65.4 percent, respectively. The FE maxima also appear closer to the 35–40 percent turning point suggested by the growth-DD2dep scatter plot in Figure 2.28

The Channels of Influence: Investment Volume vs. Efficiency

The foregoing raises important questions about the channels through which DOMdebt might affect growth. A causality-centered treatment of this question proffered important causal links from domestic debt to institutions, savings and financial depth. As far as growth is concerned, these channels could feature both volume effects that work primarily through the quantity of investment, and efficiency (quality of investment) effects that work through total factor productivity.

By including both investment and domestic debt in our growth specifications, we have thus far been focusing on the efficiency contribution of domestic debt, rather than its investment volume effect, currently picked up by the investment coefficient. To establish the relative weight of the volume and the efficiency contributions, we run regressions excluding investment as a regressor and study the difference. As can be seen from Table 5, the DOMdebt coefficient is consistently higher in such regressions. The ratio of the with-investment to the without-investment DOMdebt coefficients is 78.6 percent (average across all four regressions), indicating that the primary contribution of domestic debt is through investment efficiency, mirroring the conclusions of Pattillo, Poirson, and Ricci (2002) on external debt.29, 30 This, in turn, suggests, that should other determinants that constrain the quality of investment improve, such as private sector risk, the contribution of domestic debt to growth will weaken. Some evidence of this emerges below.

Table 5.Linear Specifications Excluding Investment (lnINVEST)Dependent variable: per capita income growth (gY=100 × growth rate in PPP per capita income)
(1) OLS(2) RE(3) GMM-sys
lnY_1–0.5199**–7.4156**–1.9597**
–2.51–7.06–2.06
gPOP–0.5829**–0.1444–0.7959
–3.67–0.72–1.24
INFLATION–0.0263**–0.0006*–0.0110
–2.69–1.92–0.19
lnINVEST
FISBAL0.1450**0.1628**0.2033**
4.523.942.62
gTOT–0.00150.0009–0.0173
–0.100.06–0.73
lnOPEN0.7101**3.3762**2.4579
2.584.821.63
EXTdebt–0.0097**–0.0206**0.0054
–3.45–3.740.36
DOMdebt0.0542**0.0801**0.0990**
2.802.102.31
memo: DOMdebt0.0406**0.0637*0.0742*
with lnINVEST included2.221.731.77
R2 overall0.340.050.151
R2 within0.48<0.001(−)
R2 between<0.010.260
Note: (i) t-statistics (for OLS and FE) and z-statistics (for RE and GMM) in italics.(ii) Constant; time dummies included in all regressions.(iii) **significant at 5 percent, *at 10 percent.(iv) Data spans 93 countries and 10 three-year time periods constructed from 30-year annual data (1975–2004).(v) Results obtained using STATA’s reg (OLS) xtreg (FE; RE) and xtabond2 (GMM-system; two-step) commands.(vi) Domestic debt regressors are lagged one period.(vii) z-statistics for GMM are heteroskedasticity-consistent.(viii) Arellano-Bond AR(1) and AR(2) tests are for first- and second-order serial correlation in errors. First-order negative serial correlation is expected due to first-differencing; model identification requires absence of second-order correlation.(ix) Hansen’s chi-squared test checks if the moment conditions used by the system GMM estimator are valid.(x) GMM instrumentation: (a) gTOT and gPOP were assumed exogenous, while the domestic debt variables, lnINVEST, lnOPEN and FISBAL were treated as endogenous; (b) instruments used for the difference equation: Xt–2, where X denotes an endogenous variable; (c) additional instruments used for the levels equation: Xt–1, where X is an endogenous variable.(xi) The “quart_STABILITY” dummy = 0 if the STABILITY variable (ICRG composite index) falls in the 1st quartile, 1 if it falls in 2nd quartile, 2 if it falls in 3rd quartile, and 3 if it falls in the 4th quartile; the relevant “STABILITY” percentiles are as follows: STABILITY (ICRG composite index) p25 = 50.8 percent; p50 = 59.15 percent; p75 = 66.9 percent.The key coefficients are denoted in bold.
Note: (i) t-statistics (for OLS and FE) and z-statistics (for RE and GMM) in italics.(ii) Constant; time dummies included in all regressions.(iii) **significant at 5 percent, *at 10 percent.(iv) Data spans 93 countries and 10 three-year time periods constructed from 30-year annual data (1975–2004).(v) Results obtained using STATA’s reg (OLS) xtreg (FE; RE) and xtabond2 (GMM-system; two-step) commands.(vi) Domestic debt regressors are lagged one period.(vii) z-statistics for GMM are heteroskedasticity-consistent.(viii) Arellano-Bond AR(1) and AR(2) tests are for first- and second-order serial correlation in errors. First-order negative serial correlation is expected due to first-differencing; model identification requires absence of second-order correlation.(ix) Hansen’s chi-squared test checks if the moment conditions used by the system GMM estimator are valid.(x) GMM instrumentation: (a) gTOT and gPOP were assumed exogenous, while the domestic debt variables, lnINVEST, lnOPEN and FISBAL were treated as endogenous; (b) instruments used for the difference equation: Xt–2, where X denotes an endogenous variable; (c) additional instruments used for the levels equation: Xt–1, where X is an endogenous variable.(xi) The “quart_STABILITY” dummy = 0 if the STABILITY variable (ICRG composite index) falls in the 1st quartile, 1 if it falls in 2nd quartile, 2 if it falls in 3rd quartile, and 3 if it falls in the 4th quartile; the relevant “STABILITY” percentiles are as follows: STABILITY (ICRG composite index) p25 = 50.8 percent; p50 = 59.15 percent; p75 = 66.9 percent.The key coefficients are denoted in bold.

Does Domestic Debt Complement, or Substitute for, Good Institutions?

Results on the interaction of domestic debt with institutions (STABILITY), suggest a substitutive rather than complementary relationship. For our preferred GMM estimation (regression 3, Table 6), the marginal growth effect of domestic debt becomes negative at the 60th percentile (ICRG index = 62), indicating a nonlinear relationship in the variable. Interpreted in economic terms, this suggests that the collateral and risk-diversification functions of domestic debt might be more relevant in high-risk countries where banks cannot lend to the private sector as freely as they would wish to. To further understand this result, we look at the composition of the subsample for which the STABILITY index was greater than its optimal threshold of 62.

Table 6.GMM Regressions including DOMdebt Interactions with STABILITYDependent variable: per capita income growth (gY=100 × growth rate in PPP per capita income)
(1) OLS(2) FE(3)

GMM-sys
lnY_1−0.9877**−5.2696**−1.4576*
–4.17–7.44–1.76
gPOP−0.8217**−0.4041−1.2352**
–3.65–1.07–2.70
INFLATION−0.0006**−0.0402**−0.0001
–2.14–3.22–0.12
lnINVEST3.7385**2.5725**5.1745**
7.693.474.47
FISBAL0.1628**0.1636**0.2918**
4.123.463.05
gTOT−0.0007−0.0222−0.0265
–0.04–1.31–1.23
lnOPEN−0.36192.7658**1.5609
–1.053.100.98
EXTdebt−0.0078**−0.0045−0.0027
–2.54–0.79–0.24
DOMdebt0.1204**0.0757**0.1568*
3.561.941.83
DOMdebt*(quart_STABILITY)–0.0288**–0.0143–0.0826*
–2.1–0.89–1.69
Maxima @ STABILITY percentileout-of-sampleout-of-sample60th
@ STABILITY value62
R2 overall0.300.030.258Hansen’s χ2: prob> χ2
R2 within0.30<0.001A-Bond AR(1): prob>z
R2 between<0.010.947A-Bond AR(2): prob>z
Note: (i) t-statistics (for OLS and FE) and z-statistics (for RE and GMM) in italics.(ii) Constant; time dummies included in all regressions.(iii) **significant at 5 percent, *at 10 percent.(iv) Data spans 93 countries and 10 three-year time periods constructed from 30-year annual data (1975–2004).(v) Results obtained using STATA’s reg (OLS) xtreg (FE; RE) and xtabond2 (GMM-system; two-step) commands.(vi) Domestic debt regressors are lagged one period.(vii) z-statistics for GMM are heteroskedasticity-consistent.(viii) Arellano-Bond AR(1) and AR(2) tests are for first- and second-order serial correlation in errors. First-order negative serial correlation is expected due to first-differencing; model identification requires absence of second-order correlation.(ix) Hansen’s chi-squared test checks if the moment conditions used by the system GMM estimator are valid.(x) GMM instrumentation: (a) gTOT and gPOP were assumed exogenous, while the domestic debt variables, lnINVEST, lnOPEN and FISBAL were treated as endogenous; (b) instruments used for the difference equation: Xt–2, where X denotes an endogenous variable; (c) additional instruments used for the levels equation: Xt–1, where X is an endogenous variable.(xi) The “quart_STABILITY” dummy = 0 if the STABILITY variable (ICRG composite index) falls in the 1st quartile, 1 if it falls in 2nd quartile, 2 if it falls in 3rd quartile and 3 if it falls in the 4th quartile; the relevant “STABILITY” percentiles are as follows: STABILITY (ICRG composite index) p25 = 50.8 percent; p50 = 59.15 percent; p75 = 66.9 percent.The key coefficients are denoted in bold.
Note: (i) t-statistics (for OLS and FE) and z-statistics (for RE and GMM) in italics.(ii) Constant; time dummies included in all regressions.(iii) **significant at 5 percent, *at 10 percent.(iv) Data spans 93 countries and 10 three-year time periods constructed from 30-year annual data (1975–2004).(v) Results obtained using STATA’s reg (OLS) xtreg (FE; RE) and xtabond2 (GMM-system; two-step) commands.(vi) Domestic debt regressors are lagged one period.(vii) z-statistics for GMM are heteroskedasticity-consistent.(viii) Arellano-Bond AR(1) and AR(2) tests are for first- and second-order serial correlation in errors. First-order negative serial correlation is expected due to first-differencing; model identification requires absence of second-order correlation.(ix) Hansen’s chi-squared test checks if the moment conditions used by the system GMM estimator are valid.(x) GMM instrumentation: (a) gTOT and gPOP were assumed exogenous, while the domestic debt variables, lnINVEST, lnOPEN and FISBAL were treated as endogenous; (b) instruments used for the difference equation: Xt–2, where X denotes an endogenous variable; (c) additional instruments used for the levels equation: Xt–1, where X is an endogenous variable.(xi) The “quart_STABILITY” dummy = 0 if the STABILITY variable (ICRG composite index) falls in the 1st quartile, 1 if it falls in 2nd quartile, 2 if it falls in 3rd quartile and 3 if it falls in the 4th quartile; the relevant “STABILITY” percentiles are as follows: STABILITY (ICRG composite index) p25 = 50.8 percent; p50 = 59.15 percent; p75 = 66.9 percent.The key coefficients are denoted in bold.
Proportion in

Full Sample

(percent)
Proportion in Subsample:

ICRG>62

(percent)
Sub-Saharan Africa4319
Emerging Markets1833

Sub-Saharan Africa countries make up 43 percent of the total observations in the full sample (930 = 93 × 10), but only 19 percent of the subsample for which STABILITY>62. By contrast, EMs are significantly over-represented in this subsample. This seems to confirm Kumhof and Tanner’s (2005) argument that domestic debt has more of a positive role to play when there are structural and institutional factors constraining good quality lending. If one of these factors is the high undiversifiable risk arising from politico-economic instability, then a country that reduces such instability through improved governance and stronger domestic institutions is less likely to need, and/or benefit from, domestic debt.

The Impact of Domestic Debt Quality on Its Optimal Size

The signs on the relevant interaction regressors employed here appear to underscore the importance of domestic debt quality for its growth impact. Debt that is securitized, bears positive real interest rates and is diversely held is found robustly friendlier to growth. Some of these results, summarized in Table 7 (panels A-B), are obtained from data spanning 70 IMF program countries over the 1996–2004 period (the Mellor database). Similarly, regressions (panels E-F) employ data on 20 sub-Saharan Africa countries since 1980 (Christensen’s, 2004 database). Less than 200 observations were available for each of these four regressions, so the results, especially the coefficient “sizes,” should be interpreted with caution.

Table 7.Interactions with Attributes of Domestic Debt (DD) QualityDependent variable: per capita income growth (gY= 100 × growth rate in PPP per capita income)
Attribute of DD

Being Tested

through Interaction

(INT Variables = >)
Share of

Securities

in DD

(SDD2DD)
Banking

System

Share in DD

(shBANK)
Financial

Repression

vs.

Financial

Liberalization

(ERA)
Bearing

Positive or

Negative

Real

Interest

Rate

(REALi)
Regressions with Christensen’s (2004)

DD Data on 20 Sub-Saharan

Countries: DD includes CG Securities

Held by Central Bank, Banks, and

Nonbanks (DDSSA)
ABanking system’s share

in DD (shBANK)
CDE: as percent

of GDP
F: as percent% of

deposits
Control variables:
lnY_1–6.7032**–5.4523*–6.2995**–7.0100**–8.4418**–7.0812**
–2.37–1.85–14.86–9.15–4.05–3.43
gPOP–0.23290.0846–0.2860–0.1282–0.2416–0.2571
–0.850.20–1.58–0.71–1.11–1.12
INFLATION0.00440.0414–0.0007**–0.0014–0.0645*–0.0535
1.000.65–2.66–1.59–1.86–1.48
lnINVEST4.1701**1.9955*3.3646**2.8014**3.5150**4.3390**
7.151.856.525.332.432.66
FISBAL–0.02400.07350.1288**0.0820**0.02880.0452
–0.420.833.702.190.340.49
gTOT–0.0135–0.0196–0.00060.0077–0.0079–0.0027
–0.70–0.73–0.050.58–0.30–0.10
lnOPEN–0.7008–1.59802.4805**3.3450**2.7712*2.0907
–1.48–0.934.035.171.781.22
EXTdebt–0.0070*–0.0042–0.0049–0.0188**–0.0164–0.0095
–1.950.41–1.34–3.79–1.33–0.72
DD variables:
ALLonCG–0.0187
–1.64
ALLonCG*SDD2DD0.1677**
2.57
DOMdebt0.0515*0.0319–0.0307
1.66–0.91–1.04
DOMdebt*INT–0.0017**0.0587*0.0470*
–3.601.721.74
DDSSA0.1590**0.0214**
3.092.04
DDSSA × quart_DDSSA–0.01140.0002
–0.82–0.06
No. of observations:186176623558130124
Summary Statistics for New Variables Introduced on DD Quality
In PercentSDD2DDShBANKREALiDDSSA/GDPDDSSA/Deposits
p2505.29–6.013.9634.35
P5026.8450.75–0.0110.5865.55
P7574.7085.223.7624.9996.02
Note: (i) FE regressions reported only (due to space considerations); t-statistics in italics.(ii) Constant; time dummies included in all regressions.(iii) **significant at 5 percent, *at 10 percent.(iv) For regressions A and B, data spans 70 IMF program countries over 1996–2004, that is, four three-year periods (using IMF Policy Development and Review Department database); for regressions C and D, the full 93 country (1975–2004) sample was used; for regressions E and F, Christensen’s (2004) domestic debt data on 20 sub-Saharan Africa countries (1980–2000) was used.(v) ALLonCG = 100 × [all banking system claims on central government/GDP]. SDD2DD = Share of “securities” in total CG DD [IMF Policy Development and Review Department database on 74 IMF program countries over the period 1996–2004]; CG denotes central government.shBANK: Banking system’s (including central bank’s] share of total CG DD [-do-].ERA: Dummy variable taking the value 0 for the “financial repression” era (1975–89) and 1 for the “financial liberalization” period (1990–2004).REAL i: Deposit rate minus inflation [WDI and IFS].DDSSA: CG securities [Christensen’s, 2004 database on 20 SSA countries; 1980–2000].The key coefficients are denoted in bold.
Note: (i) FE regressions reported only (due to space considerations); t-statistics in italics.(ii) Constant; time dummies included in all regressions.(iii) **significant at 5 percent, *at 10 percent.(iv) For regressions A and B, data spans 70 IMF program countries over 1996–2004, that is, four three-year periods (using IMF Policy Development and Review Department database); for regressions C and D, the full 93 country (1975–2004) sample was used; for regressions E and F, Christensen’s (2004) domestic debt data on 20 sub-Saharan Africa countries (1980–2000) was used.(v) ALLonCG = 100 × [all banking system claims on central government/GDP]. SDD2DD = Share of “securities” in total CG DD [IMF Policy Development and Review Department database on 74 IMF program countries over the period 1996–2004]; CG denotes central government.shBANK: Banking system’s (including central bank’s] share of total CG DD [-do-].ERA: Dummy variable taking the value 0 for the “financial repression” era (1975–89) and 1 for the “financial liberalization” period (1990–2004).REAL i: Deposit rate minus inflation [WDI and IFS].DDSSA: CG securities [Christensen’s, 2004 database on 20 SSA countries; 1980–2000].The key coefficients are denoted in bold.

The panel A regression tests the interaction of the share of securitized domestic debt in total CG domestic debt (SDD2DD) with all banking system claims on the CG (ALLonCG). The linear ALLonCG term, which includes the CG’s inflationary overdrafts from the central bank is negative and almost significantly so. However, the securitized component of CG domestic debt [ALLonCG × SDD2DD] has a strongly positive coefficient, indicating the benefits of issuing domestic debt as marketable securities. The positive coefficients on the interaction terms with ERA (financial liberalization post-1990 = 1) and REALi (positive real interest rates = 1) in panel regressions (A) and (D), respectively, provide further confirmatory evidence of this result.31Further as can be seen from the summary statistics on SDD2DD (right panel, Table 7), only about 27 percent of CG domestic debt in IMF program countries is securitized, representing substantial scope for marketization, going forward.

The careful selection and interaction of terms in panel regression B test the hypothesis that the growth contribution of DOMdebt decreases in the share of domestic debt held by the banking system.32 Summary statistics on this shBANK series indicate a median share of 50 percent and an interquartile range of 5.3–85.2 percent. The results indicate that DOMdebt becomes less growth enhancing as shBANK rises.33 The most obvious policy implication of this result is that public debt issuers should attempt to diversify debt holdings beyond commercial banks by encouraging participation from institutions (pension funds, and so on), the retail sector, and if appropriate, foreign investors. Fortunately, with private domestic savings rebounding in LICs and EMs, contractual saving institutions expanding and foreign interest in their domestic debt markets increasing, the conditions are quite conducive for undertaking such diversification.

Panel regressions E and F suggest that the result on positive overall growth payoff of domestic debt documented earlier remains robust to the sub-Saharan Africa subgroup and with an important alternative definition of domestic debt, that is, “all central government securities.”34 The estimated linear marginal effect for the domestic debt sub-Saharan Africa (DDSSA)/GDP ratio is 0.16 and for DDSSA/deposits is 0.02, matching the earlier pattern of higher growth payoffs to DOMdebt compared with DD2dep. The quadratic terms are negative but insignificant in both regressions, suggesting that current perceptions of domestic debt capacity in sub-Saharan Africa may be unnecessarily bearish. The fact that the quadratic term is not significant—even for DDSSA/deposits—may partly reflect (1) the exclusion from the DDSSA measure of a less desirable component of domestic debt: unsecuritized liabilities, overdrafts of the central bank, and so on; and (2) the inclusion in the measure of a relatively desirable component: nonbank-held domestic debt.

Indirectly, therefore, these results support the same hypotheses that panel regressions A and B lean towards: domestic debt is more growth-friendly when issued as marketable securities and, to a diverse investor base, including the nonbank sectors.

Selected Robustness Tests

Before discussing any policy conclusions, the results reported here should be tested for robustness (1) across estimation methods, (2) over different horizons and country subgroups, and (3) after removing outliers. By using OLS, RE, FE and system GMM and establishing the stability of the results over this broad range of estimation techniques, (1) has already been addressed. For (2), the panel regressions E and F on the sub-Saharan Africa subsample partly addresses the issue of robustness over country-groups. Further, to test for robustness over horizon length, three-year data are aggregated into six-year data, to make sure that any residual cyclical effects are also smoothed out. As a result of the conversion, the total number of observation halves. Table 8 summarizes the results of OLS, FE and system GMM regressions of growth on DD/GDP (DOMdebt, linear, panel A) and DD/deposits (DD2dep, nonlinear, panel B) using six-year data. As can be seen, the DOMdebt coefficient strengthens compared with the three-year case. By contrast, the evidence for a nonlinear growth impact of DD2dep weakens: all the nonlinear terms are insignificant whereas the linear coefficients are lower than their three-year counterparts in two of the three regressions.

Table 8.Robustness Check 1: Selected Regressions with Six-Year DataDependent variable: per capita income growth (gY=100 × growth rate in PPP per capita income)
A. DOMdebt (Domestic Debt/tGDP)B. DD2dep (Domestic Debt/Deposits)
LINEARNONLINEAR
(Counterpart of Table 3, Panel A)(Counterpart of Table 4, Panel B)
(1) OLS(2) FE(3)

GMM-sys
(1) OLS(2) FE(3)

GMM-sys
lnY_1−0.9394**−5.4120**−1.7637**−0.8094**−5.1443**−1.5416
–4.09–6.21–2.22–3.86–6.21–1.59
gPOP−0.6717**−0.3216−1.8077**−0.7903**0.0310−0.9176**
–3.04–0.94–2.08–3.820.09–2.05
INFLATION0.00000.00020.0001−0.0013**−0.0012**−0.0023
–0.090.420.04–3.02–2.54–0.69
lnINVEST3.8625**1.8400**5.2615**2.8337**0.98834.0645**
8.652.703.776.821.602.72
FISBAL0.1683**0.1094**0.11530.1202**0.0941**0.1016
4.442.361.353.112.091.50
gTOT−0.00520.00040.05690.0448**0.00240.0393
–0.270.020.711.980.111.64
lnOPEN−0.5231**2.2988**−0.78150.10942.5828**1.8763
–2.063.07–1.050.383.531.24
EXTdebt−0.0054*−0.0102**0.0006−0.0072**−0.0058−0.0031
–1.89–1.980.04–2.71–1.19–0.25
Domestic debt0.0732**0.0587*0.1302*0.0208**0.0262**0.0421**
2.921.661.762.132.392.09
Domestic debt × quartile–0.0018–0.00400.0036
–0.60–1.150.33
Memo (with three-year data):
Domestic debt0.0406**0.0637*0.0742*0.0280**0.0188*0.0513**
Domestic debt × quartile–0.0056*–0.0067**–0.0151*
R2 overall; Hansen’s χ20.400.030.4020.390.020.090
R2within; AR(1)0.41<0.001(−)0.410.057(−)
R2between; Ar(2)0.010.495< 0.010.595
Note: (i) t-statistics (FE and OLS) and z-statistics (GMM) in italics.(ii) Constant; time dummies included in all regressions.(iii) **significant at 5 percent, *at 10 percent.(iv) Data spans 93 countries and five six-year time periods constructed from 10 three-year period data (1975–2004).(v) Results obtained using STATA’s reg (OLS), xtreg (RE; FE) and xtabond2 (GMM-system; two-step) commands.(vi) Regressor domestic debt means DOMdebt (domestic debt/GDP) for panel a regressions, and DD2dep (domestic debt/deposits) for panel b regressions; quartile means quartiles of DD2dep.(vii) z-statistics for GMM are heteroskedasticity-consistent.(viii) Arellano-Bond AR(1) and AR(2) tests for GMM regressions check for first- and second-order serial correlation in errors. First-order negative serial correlation is expected due to first-differencing; model identification requires absence of second-order correlation.(ix) Hansen’s chi-squared test checks if the moment conditions used by the system GMM estimator are valid.(x) GMM instrumentation: (a) gTOT and gPOP were assumed exogenous, while the domestic debt variables, lnINVEST, lnOPEN and FISBAL were treated as endogenous; (b) instruments used for the difference equation: Xt–2, where X denotes an endogenous variable; (c) additional instruments used for the levels equation: Xt–1, where X is an endogenous variable.The key coefficients are denoted in bold.
Note: (i) t-statistics (FE and OLS) and z-statistics (GMM) in italics.(ii) Constant; time dummies included in all regressions.(iii) **significant at 5 percent, *at 10 percent.(iv) Data spans 93 countries and five six-year time periods constructed from 10 three-year period data (1975–2004).(v) Results obtained using STATA’s reg (OLS), xtreg (RE; FE) and xtabond2 (GMM-system; two-step) commands.(vi) Regressor domestic debt means DOMdebt (domestic debt/GDP) for panel a regressions, and DD2dep (domestic debt/deposits) for panel b regressions; quartile means quartiles of DD2dep.(vii) z-statistics for GMM are heteroskedasticity-consistent.(viii) Arellano-Bond AR(1) and AR(2) tests for GMM regressions check for first- and second-order serial correlation in errors. First-order negative serial correlation is expected due to first-differencing; model identification requires absence of second-order correlation.(ix) Hansen’s chi-squared test checks if the moment conditions used by the system GMM estimator are valid.(x) GMM instrumentation: (a) gTOT and gPOP were assumed exogenous, while the domestic debt variables, lnINVEST, lnOPEN and FISBAL were treated as endogenous; (b) instruments used for the difference equation: Xt–2, where X denotes an endogenous variable; (c) additional instruments used for the levels equation: Xt–1, where X is an endogenous variable.The key coefficients are denoted in bold.

The final robustness check—sensitivity to outliers—based on the DFBETA post-estimation command in STATA is also green.35 The command works with OLS and LSDV (FE) regressions, and computes the influence of each observation (country-period) on the coefficient of interest. DFBETA series for all three measures of domestic debt were generated after running their corresponding OLS and FE regressions. Observations with |DFBETA|>√2/N (N being the total number of observations) were then dropped from the sample and the regressions re-run on the new smaller samples. Table 9 reports the results from these outlier-cleansed regressions. As can be seen, in the case of DOMdebt, the coefficient size rises significantly for both the OLS and FE cases. For DD2dep (panel B), evidence for a nonlinear growth impact endures, with a turning point for DD2dep at 35.4 percent for the FE case, very similar to the result obtained earlier.36

Table 9.Robustness Check 2: Selected Regressions with DFBETA Outliers RemovedDependent variable: per capita income growth (gY=100 × growth rate in PPP per capita income)
A. DOMdebt (Domestic Debt/GDP)B. DD2dep (Domestic Debt/Deposits)
LINEARNONLINEAR
(Counterpart of Table 3, Panel A)(Counterpart of Table 4, Panel B)
(1) OLS(2) FE(1) OLS(2) FE
lnY_1−0.7929**−8.2050**−0.7213**−6.6612**
–4.05–7.91–3.69–9.37
gPOP−0.7148**−0.1369−0.4616**−0.1611
–4.03–0.75–3.10–0.89
INFLATION−0.0288**−0.0004−0.0255**−0.0368**
–3.11–1.46–2.61–2.83
lnINVEST3.2133**2.2811**3.6362**2.8442**
8.603.209.535.50
FISBAL0.1243**0.1499**0.0889**0.0796**
4.093.652.782.17
gTOT−0.01110.0032−0.00320.0033
–0.840.22–0.230.25
lnOPEN−0.11223.3060**0.04632.4173**
–0.414.120.163.91
EXTdebt, TOTdebt−0.0073**−0.0244*−0.0102**−0.0129
–2.79–4.36–3.77–2.72
Domestic debt0.0541**0.1213*0.0225**0.0178
2.101.932.101.61
Domestic debt*quartile–0.0056*–0.0064*
–1.69–1.89
Memo (without DFBETA outliers removed):
Domestic debt0.0406**0.0637*0.0280**0.0188*
Domestic debt*quartile–0.0056*–0.0067**
R2 overall0.440.060.440.06
R2 within0.530.47
R2 between<0.01<0.01
Obs. in original regressions (Tables 57)618530606505
|DFBETA threshold|0.08050.08690.08120.0890
Note: (i) t-statistics in italics.(ii) Constant; time dummies included in all regressions.(iii) **significant at 5 percent, *=at 10 percent.(iv) Data spans 93 countries and 10 three-year time periods constructed from 30-year annual data (1975–2004).(v) Results obtained using STATA’s reg (OLS) command; as DFBETA is a post-estimation command that works only after reg, the FE regressions had to be simulated using least squared dummy variable specifications (that is, by including country dummies)(vi) DFBETA series for each domestic debt coefficient in a particular regression were generated; DFBETA outliers were then defined as those observations for which ¡DFBETA¡> (2/On) where n is the total number of observations in the original regression. For example, take regression a1 above: the original regression (Table 3, panel A1) corresponding to regression a1 had 618 observations, implying a DFBETA threshold of±2/Ö618 = ±0.0805. All observations with DFBETA outside this range (24 in this case) were construed as “outliers.” Regression A1 (reported here) was run after dropping these outlier observations.(vii) Regressor Domestic debt means DOMdebt (domestic debt/GDP) for panel a regressions and DD2dep (domestic debt/deposits) for panel b regressions; quartile means quartiles of DD2dep.The key coefficients are denoted in bold.
Note: (i) t-statistics in italics.(ii) Constant; time dummies included in all regressions.(iii) **significant at 5 percent, *=at 10 percent.(iv) Data spans 93 countries and 10 three-year time periods constructed from 30-year annual data (1975–2004).(v) Results obtained using STATA’s reg (OLS) command; as DFBETA is a post-estimation command that works only after reg, the FE regressions had to be simulated using least squared dummy variable specifications (that is, by including country dummies)(vi) DFBETA series for each domestic debt coefficient in a particular regression were generated; DFBETA outliers were then defined as those observations for which ¡DFBETA¡> (2/On) where n is the total number of observations in the original regression. For example, take regression a1 above: the original regression (Table 3, panel A1) corresponding to regression a1 had 618 observations, implying a DFBETA threshold of±2/Ö618 = ±0.0805. All observations with DFBETA outside this range (24 in this case) were construed as “outliers.” Regression A1 (reported here) was run after dropping these outlier observations.(vii) Regressor Domestic debt means DOMdebt (domestic debt/GDP) for panel a regressions and DD2dep (domestic debt/deposits) for panel b regressions; quartile means quartiles of DD2dep.The key coefficients are denoted in bold.

Overall, the foregoing permits a more confident restatement of our earlier results on the growth contribution of domestic debt, that is, that the contribution is generally positive for DOMdebt, with evidence of diminishing returns to DD2dep above a ratio of 35 percent.

IV. Conclusions and the Way Forward

As noted at the start of this paper, deriving the policy implications of domestic debt in LICs has traditionally been complicated by one of four factors: (1) traditional pessimism about the role of domestic debt in a shallow market and/or financially repressed economies; (2) lack of domestic debt data amenable to empirical analysis; (3) the perception that a government’s domestic debt capacity is fully endogenous to the country’s level of income, financial development, saving and institutions (including government credibility), and cannot be chosen independently by policymakers; and (4) fears that active market-based domestic borrowing in small LIC markets would prove expensive relative to concessionary external finance, risking fiscal sustainability and leading to crowding out of private sector activity.

Against this backdrop, this paper’s contribution has been to develop a comparable cross-country database on domestic debt and use it to empirically analyze and size its growth impact, and study the sensitivity thereof to the quality of debt issued, and the macroeconomic, financial and institutional circumstances of issuance. A battery of panel growth regressions were run using data on 93 LICs over the 1975–2004 period (10 three-year periods), using extensive controls, different estimation methods (OLS, FE and GMM) and rigorous robustness tests including a six-year horizon for averaging, outlier treatment and regressions over country subsamples like sub-Saharan Africa.

The basic regressions suggest a robust positive linear growth contribution of domestic debt when scaled to GDP (DOMdebt). The standardized marginal effects suggest nontrivial orders of magnitude, with a 1 standard deviation increase in DOMdebt driving a 0.1 standard deviation increase in the per capita growth rate. The growth contribution of domestic debt scaled to deposits (DD2dep) appears more complex, with domestic debt seen to support growth up to a ratio of 35 percent but strangling it at higher levels. This lends some credence to the crowding out argument against domestic debt, but also proffers a sense of what the optimal level of domestic debt is. Given an average DD2dep ratio of 21 percent in LICs, and considering significant scope for financial deepening and nonbank and foreign participation in LIC domestic debt markets, the result suggests more substantial domestic debt issuance capacity than is currently perceived.

Importantly, the optimal size of domestic debt was found to be highly sensitive to its quality. A higher level of domestic debt can likely be sustained without compromising growth if domestic debt is issued in the form of marketable securities, bears positive real interest rates and is issued to the nonbank sectors (contractual saving institutions and retail investors, and so on). The latter result supports the hypothesis that nonbank participation in the government securities market boosts competition in the financial sector, both on the deposit-taking side—as banks have to compete with government for nonbank deposits—and, on the investment side, as banks compete with nonbanks in public securities auctions. This increased competition should put downward pressure on banks’ overheads and intermediation margins, partly alleviating the efficiency concerns of high bank holdings of domestic debt highlighted by Hauner (2006).

Additional specifications studying the channels through which domestic debt affects growth of influence reveal that three-quarters of the impact likely occurs through the investment efficiency or factor productivity channel, rather than the volume of capital accumulation itself. Intuitively, this could imply that the main function domestic debt performs is that of protecting banks’ profitability against the possibility of downside risks, thus permitting more aggressive risk-taking vis-à-vis private sector lending. If risks are already manageable in the economy, say, due to a stable political economy, high private saving rates (reduced external reliance) and a good degree of financial development and economic openness, the benefits of domestic debt should be more nuanced. This seems to be confirmed by a series of interaction regressions, where the growth effect of domestic debt was found to be diminishing in the above-noted variables, especially institutions.

The foregoing seems to imply that countries that have the greatest capacity for domestic debt issuance (high savings and financial depth, better institutions, and so on) probably have the least need for it; whereas countries that are unable to expand domestic debt issuance have probably the most to gain from it. Moreover, as the coefficient on FISBAL is consistently positive in all our growth regressions, such countries appear additionally constrained to find nondeficit increasing methods of expanding domestic debt (through issuance of liquidity paper, or converting unsecuritized debt into marketable securities). Thus, caution in the issuance of domestic debt is well-prescribed in contexts where debt sustainability or fiscal control remain important concerns. However, countries that have established a track record of fiscal responsibility and evolved an effective institutional framework for debt management could be encouraged to adopt a proactive program to direct budgetary financing needs towards the local currency domestic debt market.

Looking ahead, the hope is that the paper will spawn further research into the reasons why agents (especially banks) voluntarily hold domestic government debt in LICs and EMs. This will also serve to clarify if the reduced-form macro returns from domestic debt we find have clear structural underpinnings, and whether microeconomic channels like the collateral function of domestic debt ultimately find support in “micro” econometric evidence. Fortunately, there are databases (such as Bankscope) that now offer a partitioning of banks’ balance sheets into government securities and private risk assets for a number of countries and over a decent time span (1994 to date). Moreover, as the general question of banks’ portfolio choice, as it related to public and private risk, has not been formally asked in the context of LICs and EMs, it would make a naturally interesting research topic.

Appendix I. Description of Data

See Tables A1, A2 and A3.

Table A1.Summary Statistics for Main Regression Variables
NMinMaxMeanSDCVMedian
gY930−16.2023.475.004.500.904.57
Y93017120261266927641.041644
gPOP928−8.638.202.250.990.442.32
INFLATION807−10.8698.8713.9516.021.159.33
INVEST8522.0861.5120.327.380.3619.75
FISBAL930−63.5427.03−4.786.02−1.26−4.03
gTOT922−50.0178.440.8610.6012.31−0.07
OPEN9190.69239.3564.4338.710.6054.85
DOMdebt844089.726.349.701.533.61
DD2dep843094.4021.2117.170.8117.43
EXTdebt7916.06384.6174.2159.150.8057.64
prSAVING733−52.0541.8013.299.450.7113.43
STABILITY52623.8081.7058.2211.700.2059.15
Note: N=no. of observations; SD=standard deviation; CV = coefficient of variation; time period: 1975–2004 (three-year averages); “DD2dep” is domestic debt/deposits (in percent).
Note: N=no. of observations; SD=standard deviation; CV = coefficient of variation; time period: 1975–2004 (three-year averages); “DD2dep” is domestic debt/deposits (in percent).
Table A2.Correlation Matrix for Main Regression Variables
gYlnY_lgPOPINFLATIONlnINVESTFISBALgTOTlnOPENDOMdebtDD2depEXTdebtprSAVINGSTABILITY
gY1
lnY_l–0.101
GPOP–0.04–0.431
INFLATION–0.160.010.041
lnINVEST0.380.31–0.16–0.131
FISBAL0.090.24–0.08–0.110.131
gTOT0.05–0.01–0.04–0.020.000.091
lnOPEN0.070.26–0.10–0.100.400.090.021
DOMdebt0.030.33–0.23–0.040.22–0.13–0.050.241
DD2dep–0.010.07–0.05–0.030.04–0.21–0.020.020.681
EXTdebt–0.32–0.240.140.10–0.13–0.17–0.040.20–0.050.031
prSAVING0.190.44–0.230.030.450.010.080.140.260.14–0.311
STABILITY0.250.60–0.39–0.160.450.380.030.360.280.02–0.380.381
Note: As there were noticeably fewer observations for prSAVING and STABILITY, the correlation matrix for the “remaining variables” was computed by excluding the two variables.
Note: As there were noticeably fewer observations for prSAVING and STABILITY, the correlation matrix for the “remaining variables” was computed by excluding the two variables.
Table A3.Descriptions and Sources of Variables
VariableDescriptionSource
gY100 × (per capita PPP GDP growth)World Bank: WDI
lnY_1log of [lagged per capita PPP GDP]World Bank: WDI
gPOP100 × (growth rate of population)World Bank: WDI
INFLATION100 × [π/(1 + π)], where π is the annual percent change in CPIWorld Bank: WDI
lnINVESTlog of [100 × (gross fixed capital formation/GDP)]IMF: IFS 93e
FISBAL100 × (fiscal balance/GDP)IMF: WEO
gTOT100 × [growth in terms of trade (goods)]IMF: WEO
lnOPENlog of [100 × (trade/GDP)]World Bank: WDI
DOMdebt100 × [(banks’ claims on CG) + (central bank securities)/GDP] [(22a + 42a) + (20c + 40c)]IMF: IFS
DD2dep100 × [DOMdebt/All bank deposits (current, time, saving)]Deposits: World Bank: WDI
EXTdebt100 × [(public + private) external debt/GDP]World Bank: GDF
prSAVING100 × (Private savings/GDP)IMF: WEO
STABILITYICRG “composite” risk index to proxy for politico-economic stability and quality of domestic institutions; high value of index indicates low riskPRS group:www.icrgonline.com
Appendix II. Financial Depth Index and Granger-Causality Regressions

Extracting the Financial Depth Index using Principal Component Analysis

Huang and Temple (2005) use liquid liabilities of the financial system/GDP (LLY), private sector credit provided by commercial (and other) banks/GDP (PRIVO) and commercial bank assets as a ratio of total banking system assets (BTOT) as the principal components of the underlying latent financial depth variable.37 They also constructed other measures of financial development related to financial intermediation efficiency, and the existence and size of stock and bond markets, but the data on these had too many gaps for the countries we are interested in, that is, LICs and EMs. As such, we focus on financial depth, which, indeed, is also the variable they use in their panel regressions.

Principal component analysis consists in taking N specific indicators (with LLY, PRIVO, BTOT, N=3) and solving for their uncorrelated principal components (P1,…, PN) that capture different dimensions of the underlying series. We only use the first component (P1), formally defined by a vector of weights α = (α1, α2, αN)’ on the (standardized) indicators such that α’X has the maximum variance for any possible weights, subject to the constraint α’α = 1.

The method is applied to the “log normalized” LLY, PRIVO and BTOT series to obtain the principal components (below). As can be seen, the weights are roughly similar for the three series, indicating that they are indeed highly correlated and hopefully capturing the same underlying latent variable, financial depth. P1 explains about 84.6 percent of the variation in the series, and therefore sufficient for our purposes.

Principal Components (Weights)
P1P2P3
LLY0.5610.7760.288
PRIVO0.575−0.6170.539
BTOT0.596−0.137−0.791

Weights P1 were applied to LLY, PRIVO and BTOT and used to construct the financial depth index referred to in the text.

Econometric Framework to Test for Granger-Causality

The starting point for Granger-causality tests is dynamic single equation panel data regressions of the form:

where y and x denote, respectively, the endogenous and exogenous variables of interest; ni denotes unobserved country heterogeneity; φt period dummies;38 νit the error term; i = 1,2… 93; and for the chosen lag length of 2, t = 3 … 10.39 A joint significance Wald test on β1 and β2 [=0] helps ascertain if y is Granger-caused by x. As panel regressions of this form involve a lagged dependent variable, it is problematic to employ standard FE estimators to eliminate ni.40 Instrumental variable estimators, like the GMM offer a robust solution to these problems by first-differencing (A.1) to produce:

and using appropriate lags of y and x to instrument for Δy and Δx.41 The problem with these simple “difference GMM” estimators is that lagged levels of regressors are often weak instruments of the differenced variables. This is especially true when the underlying series are persistent, or the variance of the individual effects (ni) is high relative to the variance of the transient shocks (νit). These conditions are likely to be met for the data we are using: the time series process for income (or GDP per capita) is known to be highly persistent; the variance of country heterogeneity is likely to be very high in our sample as it includes Asian EMs like China, Latin American oil-producers like Venezuela and very poor sub-Saharan Africa countries like the Democratic Republic of Congo.

For precisely such cases, Arellano and Bover (1995) and Blundell and Bond (1998) have developed “system GMM” estimators, which can deliver significant improvements to model identification. Such estimators utilize additional assumptions about the initial conditions of the data process. In the context of growth regressions of the type we will be running later, the additional assumption pertains to there being no correlation between output growth and the country-specific effect in the absence of conditioning on other variables. Such an assumption is consistent with Solow’s conditional convergence growth framework, and its violation would tend to have implausible long-run implications.

The system GMM estimator then uses lagged differences to instrument the level variables appearing in the extra moment conditions permitted by the additional initial condition assumptions. Simulations have suggested that system GMM deals with weak instrument biases more robustly than difference GMM. As a result, the former has become increasingly popular in cross-country panel econometric studies.

References

S.M. Ali Abbas is an economist with the IMF Fiscal Affairs Department. Jakob E. Christensen is an advisor with the Central Bank of Denmark. The authors are grateful to Christopher Adam, David Bevan, Thorsten Beck, Anne-Marie Gulde-Wolf, Jay Peiris, Hans Weisfeld, Tsidi Tsikata, and participants at an IMF African Department Informal Seminar Series where an earlier draft was presented.

The debt neutrality literature developed in industrialized countries remains agnostic to the external-domestic debt distinction. This is understandable, given that industrialized country governments borrow predominantly from their own citizens and in their own currencies. However, as the “original sin” literature has made clear, the distinction is nontrivial for LICs (see Eichengreen and Hausmann, 2005).

Since this paper is concerned with domestic debt “markets,” we define domestic debt as the domestic currency indebtedness of a country’s consolidated public sector to its citizens; thus, liquidity absorbing paper issued by the central bank is included, but government securities held by the central bank are excluded (see Section II for details). We recognize that some governments, especially in Latin American, have issued debt in foreign currency (or debt that is index linked) to domestic residents. This “foreign currency” domestic debt is analytically quite different from our “domestic currency” domestic debt, but the data do not permit us to discriminate between the two types.

For evidence favoring internal over external finance reliance in LICs, see Aizenmann, Pinto, and Radziwill (2004).

Abbas (2005) questions the conclusion of Beaugrand, Loko, and Mlachila (2002) that domestic debt is always more expensive than concessionary external debt, noting that they do not take account of (1) the impact of the higher variance of external debt service (due to currency risk) on the present value burden of external debt; and (2) the fact that the implicit interest rate on external debt—a large part whereof is in default—is not comparable to the implicit interest rate on domestic debt, which is rarely defaulted on.

See Agenor and Montiel (1999, Chapter 5) on how government incentives to extract seigniorage through high inflation leads to an erosion of the underlying nominal tax base.

The complacency effect should, however, diminish over time as competition reduces both yields in government securities auctions, and profits in the banking sector. Indeed, yields have fallen significantly lately as international investors have increased competition in African bond markets.

See Chirwa and Mlachila (2004); Barajas, Steiner, and Salazar (1999, 2000); and Brock and Rojas-Suarez (2000).

His theoretical argument, in summary, is that in an economic downturn, when private sector returns are falling, the government’s domestic tax revenue and foreign aid receipts would also likely fall, leading to a widening of the fiscal gap. To the extent that the latter is financed domestically, yields on government paper will rise, boosting bank profitability and, thus, militating against the contemporaneous lower returns from private lending.

No explicit definition of domestic debt is provided in the report; however, it appears that, like Hauner (2006), commercial bank credit to government is used as a proxy.

See Hauner (2006) on the possible efficiency concerns associated with commercial bank holdings of government securities.

Other interesting attributes of domestic debt could also matter for its growth impact, but are not covered here. These include whether the debt is short- vs. long-term, index-linked, held by local or foreign residents, fixed or floating, backed by an active secondary market and issued in benchmark maturities to enhance instrument liquidity. For the purposes of this paper, however, and given the data constraints, we are able to focus only on the marketability and holder-profile of domestic debt.

These data have usually been extracted from IMF’s country statistical appendices and staff reports, which contain useful data on domestic government securities for IMF member countries. However, it is quite difficult to extract continuous and consistent domestic debt series from these documents, as the coverage and timeliness of the appendices and staff reports varies significantly from program to nonprogram countries and from program to nonprogram “periods” for each individual country.

Abbas (2007) also features a full discussion of the merits and demerits of alternative domestic debt definitions.

Formally, the definition is the following: Public sector domestic debt (DD) = DMB & OBI claims on CG + DMB & OBI holdings of liquidity paper; or the lines in IFS: DD = (22a & 42a) + (20c&40c), where DMB is deposit money banks, OBI is other banking institutions, and CG and CB denote central government and central bank, respectively.

These periods correspond broadly to the financial repression and financial liberalization eras, even though we recognize that the latter is an ongoing process.

The three series are reported from 1960 onwards for all countries in our sample in the Financial Structure Database of Beck, Demirgüç-Kunt, and Levine (2006; updated from 2000).

Education (lnEDU) was dropped as a control for two reasons: (1) due to the poor measure of education (gross enrollment rates) the variable has been found insignificant in many cross-country growth regressions. This was the unfortunate result we obtained as well, and it persisted over lagged specifications (that is, lnEDU_1 and lnEDU_2); and (2) there were many data gaps in the schooling series, so that the variable’s inclusion would have resulted in a reduction of over 100 degrees of freedom.

Although scatter plots are useful for “prior” formation, they capture unconditional relationships between country means (over time) and, thus, should not be over-interpreted.

Note that, like Pattillo, Poirson, and Ricci (2002), we also lag the domestic debt variables and their interaction terms to avoid possible regressor endogeneity.

The dummy variable takes the value 0 if the relevant domestic debt ratio is in the bottom (that is, first) quartile, 1 if it is in the 2nd quartile, 2 if in the 3rd quartile and 3 if in the 4th quartile.

Dynamic panel data models feature a lagged dependent variable as regressor. In the context of our income growth equation, lnY_1 can be viewed as the lagged dependent variable. This is because growth is simply lnY–lnY_1; so that regressions of growth on alnY_1 + bX can equivalently be written as lnY–lnY_1 = alnY_1 + bX or lnY = (a + 1)lnY_1 + bX, which is clearly a lagged dependent specification.

Key diagnostic tests are explained in the footnotes to Table 2 and appear to be fine for the reported regressions. The joint significance test on the βs of the causal (or x) variables is the Granger-causality test. Formally, our null hypothesis is: H0 = x does not Granger cause y (joint β test is insignificant); so that a low p-value on the joint β test allows us to reject the null in favor of the alternative, that is, a causal channel exists from x to y.

The intermediating channels from domestic debt to private saving in particular could be complex, ranging from (1) Ricardian equivalence to (2) widened pool of investment grade instruments to (3) a strong collateral function of domestic debt on bank balance sheets luring in private savings to the financial system, and (4) strengthened accountability channels leading to greater policy credibility and increased public confidence in the economy.

We see this as a puzzle because (1) financial development is generally believed to be an important causal variable for income, and (2) domestic debt, which we find to be causal for income, is likely to owe this at least partly to the agency of financial development.

We could not incorporate the series on bond market capitalization in our financial development/depth index as continuous data on the former were only available for a handful of LICs.

Other variables had unstable and insignificant coefficients. gTOT has an unstable sign and is insignificant in most regressions. The result appears to corroborate both the earlier skepticism on optimal management of commodity price booms and the more recent concerns on natural resource curses in lower- and middle-income countries. The results on openness or trade/GDP (lnOPEN) are complicated, and vary across both specifications and estimation methodologies.

This suggests, in line with earlier discussion, that the supply of bank deposits may act as a more binding constraint, than economic size (GDP), on a government’s capacity to issue domestic debt without severely crowding out the private sector. To that extent, therefore, regressions with DD2dep may be more insightful than those with DOMdebt regarding the optimal level of domestic debt in an economy. Note, however, that this line of reasoning does not take into account possible endogeneity of financial depth (and the supply of deposits) to the domestic debt stock, as suggested by the Granger-causality regressions (Table 2, panel A).

The context, of course, in Pattillo, Poirson, and Ricci (2002) is how external debt reduces (not increases) growth.

One must be careful with this inference, however, as our proxy for investment, gross fixed capital formation/GDP, only captures physical capital accumulation (and possibly with measurement error), while excluding the human, public and institutional dimensions to capital.

The real interest rate at which the marginal impact of domestic debt becomes positive is + 0.66 percent [= 0.0307÷0.047], which is not too far off from the −0.01 percent median observation for REALi in the right panel of Table 7.

Formally, the expression domestic debt in this sentence refers to all CG domestic debt (including CG securities held by the central bank). Liquidity paper is naturally excluded.

Since the data on shBANK are not broken into central bank and commercial banks, we cannot ascertain whether this result reflects (1) the negative effects of higher inflationary finance by the central bank, or (2) the adverse efficiency and crowding out effects of high bank holdings of government debt. As the latter would be inconsistent with the positive growth effects of bank-held domestic debt documented in Tables 13, we are inclined to attribute the negative coefficient on shBANK to (1).

The summary statistics for this DDSSA variable, reported in Table 7, indicate median DDSSA/GDP and DDSSA/Deposits ratios of 11 percent and 66 percent, respectively. These are quite high, respectively, in relation to the DOMdebt ratio (2 percent) and the DD2dep ratio (19 percent) for sub-Saharan Africa reported in Table 1, the difference caused by large central bank holdings of CG securities in these countries.

This robustness check was applied to the thee-year averaged data.

Unfortunately, we do not have the benefit of DFBETA-adjusted GMM regressions, which would have enabled a fuller comparison.

Combining the individual variables can also help alleviate measurement errors and outlier problems that might arise if only a single variable is used.

Period dummies are extremely important in these regressions to control for the financial repression “years” and other common shocks, such as the intermittent debt and financial crises.

The first time period is 1975–77 and the 10th is 2002–04.

The within-transformed lagged dependent regressor becomes correlated with the transformed error term, rendering the FE estimator biased.

The Anderson-Hsiao difference estimator, can also circumvent the fixed effects bias, but performs badly with highly persistent series, such as income.

Other Resources Citing This Publication