The dependent variable and control variables

Following Temple (Reference Temple1999: 131–132), we run panel regressions with time- and country-fixed effects and growth of real gross domestic product (GDP) per capita as the dependent variable, averaged over five-year periods. There is no complete agreement on what control variables to include in growth regressions, but we use an extensive set including initial GDP, investment rate, openness (as measured by trade shares), government size, inflation, life expectancy and labour force growth. This includes the most commonly used control variables in the growth literature except education, omitted here to economize on data.Footnote ⁸ (In section 5, we do include education as a robustness test, and note that it does not affect our main results.) This full set of control variables is included in all regressions, even when not shown to save space. Table 1 gives variable description and sources for the data we use, and Table A1 in Appendix A contains descriptive statistics. In the next section, we describe our variables of interest, measuring institutional quality and instability.

Table 1.

Variable definitions

Variables of interest: institutional quality and institutional instability

Aron (Reference Aron2000: 115) stresses the importance of using institutional measures carefully, as many studies in the growth literature employ an ‘often-arbitrary aggregation of different components’ (see de Haan, Reference de Haan2007). We share this concern, and as will be described, we use PCA to minimize this problem.

To construct a measure of institutional quality and instability, we use the ICRG (International Country Risk Guide, 2008), which is the only measure of institutional quality that suits the requirements to test the theory. Containing yearly data since 1984 for a large number of countries, the data allow us to quantify instability using the coefficient of variation over time within four five-year periods; note that this metric is scale invariant. The full dataset from the ICRG consists of three dimensions, quantifying political risk, economic risk and financial risk. Because the latter two consist mainly of economic outcomes such as international GDP ranking, inflation, foreign debt and current account balance, we use the political risk index to construct measures of institutional quality.

The overall political risk index is composed of 12 components listed in Table 2. These are aggregated with equal weights into a single index.Footnote ⁹ As stressed by Aron (Reference Aron2000), aggregating different components is inappropriate given the multidimensionality of institutional quality: some of the 12 different components differ substantially from each, and a growth effect from instability in the aggregated index would not reveal what is driving the result. On the other hand, some of the 12 components are conceptually similar and highly correlated, and it is not advisable to test these against each other. These problems can in principle be alleviated in two different ways: (1) by manually separating components into theoretically informed groups from which to form conceptually separate indices; and (2) by using an algorithm exploiting the empirical associations between components to form indices that are statistically separable. On the one hand, solution (1) has the benefit of providing readily interpretable data, as they are based on the theoretical preconception of its author. However, this solution does not solve the problems of statistical separability and suffers from necessarily being constructed from an arbitrary weighting scheme, and from relying on the validity of the constitutive theoretical conception. Additionally, solution (1) may tempt the researcher to cherry-pick components that generate interesting results. Solution (2), on the other hand, may under some circumstances fail to provide meaningful index structures. The ideal solution is, obviously, when solution (2) yields an index structure and dimensionality that makes theoretical sense. We therefore choose solution (2) and argue that the solution indeed is theoretically consistent and interpretable.Footnote ¹⁰

Table 2.

The components of the political risk index of the ICRG

To avoid imposing a one-dimensional structure with a potentially arbitrary weighting scheme on the data, we therefore use PCA to form a number of institutional indicators from the 12 components in the political risk index. By doing so, we maximize variation and avoid testing partially correlated indices against each other. Using PCA lets the structure of the data determine how components are pooled to form separate indices instead of forcing a specific organization on the data. The results of the PCA are reported in Table 3.

Table 3.

Principal components analysis (PCA): loadings and uniqueness

Notes: Loadings in bold are referred to in the text as ‘heavy’ loadings, i.e. the major influences on the PCA scores. Loadings in italics refer to indices with intermediate influence. The component solution has been rotated using the Varimax technique.

Table 3 shows that the 12 components of the political risk index do not load onto a single factor but split quite nicely into three underlying dimensions explaining approximately 70% of the variation of the original data. We thus avoid one of the main problems of choosing solution (2).

The use of PCA entails two potential problems: (1) that one risks throwing away valuable information even though the analysis provides a best fit of the data; and (2) that component solutions may be difficult to interpret by not conforming to any established theory. Yet, given that the fourth component has an eigenvalue of only 0.8 and the constituting components of the political risk index derive from dichotomous data, the precision of the PCA – an R ² of 0.7 – can be deemed satisfactory.Footnote ¹¹ As for the second potential problem, that the results from the PCA may be difficult to interpret, we note that, in our case, the three dimensions turn out to be rather informative.

The first dimension is interpreted quite easily, as it loads heavily on Law and order, Democratic accountability, Military in politics, Socioeconomic conditions, Corruption and Bureaucratic quality, all of which either measure the quality and capacity of the legal system or consequences and reflections of such quality and capacity. Furthermore, the correlation between our first dimension and the second area of the Economic Freedom of the World Index (EFI), Legal structure and security of property rights (often treated as the most transparent and arguably the ‘cleanest’ measure of the rule of law; see Gwartney and Lawson, Reference Gwartney and Lawson2007), is 0.77, making it intuitively sensible to interpret this dimension as a ‘legal dimension’ of institutional quality.

The second dimension includes heavy loadings of countries’ Investment profile and their Government stability. The correlation between this dimension and area five of the EFI, Regulation of credit, labour and business, is 0.42, while the partial correlation, when controlling for area two of the EFI, is 0.34. Adding the two areas of the EFI to the PCA shows that area two exclusively loads onto a factor including the same ICRG components as the first dimension (with a loading of 0.78), which we term a legal dimension, while area five loads moderately onto the first dimension and heavily onto the second dimension (loading 0.63). We therefore interpret the second dimension as a measure of the quality of regulatory policy, in short a ‘policy dimension’.

In Table A2 in Appendix A, we present the complete correlations between our two PCA dimensions and areas two and five of the EFI. The first dimension, which we interpret as the legal dimension, correlates highly with the corresponding dimension of the Economic Freedom Index not only in levels but also in variation.Footnote ¹² Furthermore, our second dimension, the policy dimension, correlates highly with the corresponding dimension of the Economic Freedom Index (EFI5), and our third dimension is uncorrelated with these, as it should be.

The third dimension, finally, consists of heavy loadings of the ICRG components on External and Internal conflict, Religious and Ethnic tensions and Law and order, and correlates at –0.37 with the ethnic diversity index from Alesina et al. (Reference Alesina, Devleeschauwer, Easterly, Kurlat and Wazciarg2003). This final index can therefore readily be interpreted as a measure of both actual and latent conflicts and tensions in society, including socio-political instability and social unrest (see Alesina and Perotti, Reference Alesina and Perotti1996). We thus call this dimension ‘social congruence’.

Finally, to arrive at a set of measures of institutional instability, we calculate the coefficients of variation of the resulting principal components within each five-year period using the variance and averages of institutional quality. Thereby we also allow the heterogeneity of the instability inherent in the data to determine our indicators.

An illustration

As an illustration of what the three indices obtained from the PCA actually measure, we explore their correlations with the well-known Gastil index. The three indices all correlate with the Gastil index at 0.59, 0.45 and 0.27, respectively, and the Gastil index in turn correlates with the overall political risk index at 0.77.

The PCA indices can be used to clarify the potential pitfalls of not treating institutional quality as a multidimensional concept, and the need to separate quality and instability. To take an example, Denmark receives the second-largest score in the latest period for legal quality and is the fifth most stable country in that area. However, it is only number 71 in terms of the quality of policy and number 80 in terms of social congruence, and receives relatively unstable scores on both these dimensions, placed at number 50 and 71, respectively. Panama, on the other hand, is placed at number 52 on the legal dimension but has the seventh most stable legal environment. These countries exemplify how quality and stability are only imperfectly associated: the correlation between legal quality and legal instability is –0.50, that between policy quality and policy instability is –0.61, and that between the level of social congruence and its stability is –0.40.

The main points of our strategy are illustrated in Figure 1, where we plot the scores of legal quality for Denmark, Malaysia, the USA and Venezuela. First, Danish legal quality has been high and very stable across the entire period 1984–2004, while American legal quality has been of almost the same quality, but as the figure illustrates, somewhat more volatile. Simply comparing quality at the beginning of the period may therefore give a slightly biased impression of actual institutional performance in the two countries, although the differences may seem relatively minor. Second, comparing Malaysia and Venezuela accentuates this point, as the two countries had almost equal legal quality around 1990. The legal quality of the Venezuelan system has, on the other hand, been less stable than its Malaysian counterpart across the entire period and has obviously been characterized by a long-run downward trend. Yet, if either the initial level of 1990 or the average is used, one is likely to overestimate the positive impact of Venezuela's legal institutions compared with Malaysia, a disparity reflected in the difference between the Venezuelan average annual growth rate during 1984–2004 of –0.36% and the Malaysian average of 3.2% in the same period. Likewise, comparing the instability of the institutions of the two countries can be misleading, as Malaysia has seen instability around a relatively stable long-run level while the instability of Venezuelan institutions is a reflection of a steady deterioration. One therefore ideally has to take into account both the level, the medium-run trend of the quality of such institutions as well as its instability in order to get a full estimate of the institutional impact.

Figure 1.

Legal quality 1984–2004, giving four examples. For interpretative convenience, we have rescaled indices in this figure to be within the same interval as the original International Country Risk Guide (ICRG) components

Estimation strategy

We estimate regressions as in equation 1, where Gr is the growth rate of real GDP per capita, and X is a set of standard controls; D are time- and country-fixed effects and ε is a noise term. In order to separate the potential effects as discussed above, we include three groups of variables:

When interpreting these effects, one must therefore keep in mind that what our trends measure is strictly categorical and allows only for separate effects between situations where the trend is positive, i.e. conditional on institutions improving (trend = 1), when the trend is negative, i.e. where institutions are worsening (trend = –1), or when the trend is roughly constant (trend = 0).

(1)

In further analysis, we expand the specification to equation 2 and add an interaction term between CV_Q and TR_Q as specified in equation 2.Footnote ¹³

(2)

Although our main focus is on CV_Q, we need to include Q and TR_Q in the specification at all times. As the correlations noted above suggest, these elements (variation, level and trend) are statistically separable, but they remain sufficiently strongly associated that excluding one or both would be likely to cause an omitted variable bias. We thus note that this problem means that previous estimates in the literature may suffer from such a bias. In addition, by including the trend we gain more precise information about under what conditions institutional instability matters for growth.

The control variables in our specification are factors that are broadly used in the empirical growth literature. In all regressions, the X vector consists of the logarithm of initial GDP per capita to account for conditional convergence, government expenditures as percentage of total GDP, openness (imports plus exports as percentage of total GDP), the investment share of GDP, inflation, life expectancy and labour force growth. As such, we capture the most important non-institutional determinants of economic growth while still keeping the specification sufficiently parsimonious to include a large and diverse set of countries (in line with Barro, Reference Barro1997). As we are thereby running the risk of spurious results due to omitted variables bias, we offer a set of robustness tests in which we include five additional variables.

Our full sample covers 127 countries with a political risk rating in at least one of the four time periods 1984–1989, 1989–1994, 1994–1999 and 1999–2004; the countries are listed in Table A3 of Appendix A. Growth is measured as the five-year average, as are all control variables except initial GDP per capita. Forty of these countries have a GDP per capita above 14,000 USD in at least one period, which we define as our high-income subsample for which determinants of growth and institutional impacts may differ from the full sample and the poor subsample (see Keefer and Knack, Reference Knack and Keefer1995; de Haan and Siermann, Reference de Haan and Siermann1996). We split the sample, as citizens and market actors in high-income countries have access to more complete insurance markets, financial instruments in deeper markets as well as better market information, and are therefore substantially better suited to handle institutional instability without real losses in the short to medium run.Footnote ¹⁴ The rich subsample roughly corresponds to the current group of Organisation for Economic Co-operation and Development (OECD) member countries.

Additional indices and outliers

First, we test whether the main results are robust to including the level and coefficient of variation of four alternative indicators of institutional quality: the Gastil index of political rights and civil liberties, Henisz's (Reference Henisz2002) ‘Political Constraints V’ indicator of veto player strength, the Polity IV index of democracy, and the Herfindahl index of the legislature as an index of the level of political competition. These alternatives mainly pick up variations in political institutions, while we argue that the ICRG indices mainly capture economic and judicial institutions (see Munck and Verkuilen, Reference Munck and Verkuilen2002). By including alternative indicators with established interpretations we test whether our results simply proxy for effects of, e.g., democracy or constraints on policy-makers, although we also note that the simultaneous inclusion of alternative institutional measures most probably causes some variance inflation. We limit ourselves to these indices, as they are the only other institutional measures with a sufficient amount of years covered.

We perform all types of robustness analysis with the full specification, although we only report the institutional coefficients in Tables A4 and A5 of Appendix A; these should be compared with columns 4 and 6 of Table 4. For the poor countries, out of 36 coefficients (nine for each of the four indices), 25 have the same sign and statistical significance as before. For the rich countries, the corresponding figure is 20 out of 36, implying less robustness, although we should stress that the problems of variance inflation when including multiple institutional indicators seem particularly acute in this sample. The result that only legal quality is significant in rich countries while both legal quality and policy quality are significant in poor countries also turns out to be robust when excluding outliers, as do the results pertaining to instability and institutional trends. The tables show that most main findings are largely robust to including the quality and coefficient of variation of the Gastil index, the Political Constraints V index, the Polity IV index and the Herfindahl index.Footnote ¹⁷ The only non-robust result in rich countries appears to be that the positive effects of legal instability, as evaluated at the sample mean, are not robust to including the Herfindahl index, although neither its level, trend nor its instability are close to significance.

Second, we test what happens when potential outlier observations are removed from the sample and whether the results are robust to excluding the observations with the best and worst institutions. We use a jack-knife exercise in which we exclude single regions and countries with few observations; in general, the main results are reconfirmed. In the full sample, the effect of the instability of policy quality fails significance when excluding observations from either the post-communist countries, Sub-Saharan Africa, the Middle East and North African region, Asia or countries with less than three observations in the dataset. What is more, excluding the Sub-Saharan African countries – i.e., the absolutely poorest countries in the sample – yields the legal quality index insignificant. However, in the rich subsample, all results remain robust in a jack-knife.

Third, we acknowledge a potential problem of omitted variables bias with the use of a parsimonious model specification. In Tables A6–A8 we therefore check the robustness of our results to the inclusion of further controls (the standard baseline of control variables still included): investment price, which we include instead of the investment rate; two human capital variables (the share of population with at least secondary schooling, and average years of schooling in the population); fertility; and a dummy for terms of trade crisis. We find that terms of trade crises are important (and negative) for growth in poor countries, while fertility is weakly significant in the rich subsample, also with a negative sign.

In general, our main results pertaining to institutional quality are robust – including the positive growth effects from policy quality trends and negative effects of trends in social congruence. In rich countries, results pertaining to policy and social congruence instability remain robust while the effects of legal instability appear less robust.

Handling possible endogeneity

As a final issue, we try to control for possible endogeneity and simultaneity in two ways.Footnote ¹⁸ As is almost always the case, we note that the institutional measures may lag rather than lead growth rates for several reasons. First, simple arguments could be made why institutions might improve when the economy grows (see Chong and Calderón, Reference Chong and Calderón2000). For example, the quality of legal systems and public bureaucracies could be constrained by available resources, in which case growth would lead to better institutions in the long run by alleviating this constraint. Second, we note the risk when using subjective or quasi-subjective indices that evaluations of institutional quality are affected by expectations of economic growth in the immediate future. If these expectations are on average correct, higher growth rates in the short run would simply be reflected in our measures of institutional quality instead of causing actual quality. In this case, we would expect this reflection to show up in higher investment rates to the extent that the expectations are shared by the market to which the ICRG primarily delivers its risk assessments.

To investigate causality in a tentative way, we first include lagged growth rates, based on the simple argument that if higher growth rates cause rather than follow higher institutional quality and affect institutional stability, a lagged dependent variable would pick up at least some of this effect by being the actual cause of institutional differences. If some estimates are due to simultaneity or reverse causality, we would expect to see those estimates become smaller and statistically weaker. Yet, the estimates, which we report in Table A9 in Appendix A, in general do not suggest that endogeneity is a major concern even though the inclusion of a lagged dependent variable induces a degree of downwards Nickell bias. In the full sample, we find no significant differences although the point estimates of trends in policy quality and social congruence are slightly smaller. The results in the rich subsample are entirely unaffected while the instability of policy quality in the poor subsample is rendered insignificant. With few exceptions, this exercise therefore does not suggest major endogeneity problems. With respect to the possibility that our estimates suffer from simultaneity bias due to institutional indices reflecting market expectations, the exclusion of investment rates does not affect our estimates of institutional effects (not shown). Given that such expectations would most probably show up in the investment rate instead of affecting productivity, we do not believe that this is a major worry.

Our second test is an attempt to instrument for our variables of interest. We must note that, as is often the case, our search for valid instrumental variables that account for the variation of institutional quality and instability over time has proven to be unsuccessful. In particular, as all our main estimates are obtained with country- and time-fixed effects, we cannot rely on the advances in instrumentation from the recent literature on long-run development, as all potential instrumental variables would need to define medium-run institutional changes. The best instrumental variables we could find proved to be lagged measures of institutional quality derived from the PCA and lagged growth rates; but these primarily explain the cross-country variation while leaving almost all within-country variation unexplained. The same problem pertained when we switched to either a random-effects generalized least squares (GLS) estimator or pooled ordinary least squares (OLS) with panel-corrected standard errors. As these choices allowed us to identify institutional differences by time-invariant factors, it proved substantially easier to find statistically valid candidates for instruments. However, while legal quality can be instrumented satisfactorily in more than one way, our search for instrumental variables for instability measures was unsuccessful. In particular, identification across the spectrum of trends – what appeared as a mediating factor in Table 5 – was not possible, implying that these instruments exhibited a significant bias towards zero (Dunning, Reference Dunning2008).Footnote ¹⁹

We nevertheless found one set of additional practicable instruments for legal quality and its instability in the rich subsample. We therefore report the instrumental variable estimates of legal quality, policy quality and social congruence in column 4 of Table A9, where instruments are lagged growth, lagged institutional quality and voting patterns in the United Nations General Assembly (from Voeten, Reference Voeten2004). We find that only legal quality appears important; as usual, the instrumented estimate is somewhat larger despite good identification statistics, although we cannot reject that it is the same as the simple estimate in previous tables. In column 5, we instead provide the instrumented estimates of legal quality and its instability, instrumented by lagged growth, lagged legal quality and social congruence (hence excluded) and the investment price level. We again find insignificantly larger but robust estimates with good identification statistics. As such, these exercises do not point to major endogeneity problems – if anything, our simpler estimates may arguably give relatively conservative estimates of the importance of institutional quality and instability.

As the main results are relatively robust to a set of feasible tests, especially with regard to the important role of institutional quality, and especially in rich countries, we move on to discussing the implications of the findings in the final section.