Social divisions and institutions: Considering cross-country institutional parameter heterogeneity

Transcription

1 Social divisions and institutions: Considering cross-country institutional parameter heterogeneity Ann-Sofie Isaksson Göteborg University Work in progress March 2007 Abstract This paper investigates the hypothesis that the association between institutional quality and economic performance is weaker in countries marked by social divisions, and whether this contributes to a weaker institutional parameter in the African sub-sample. The results of the empirical estimations support the social division hypothesis, and indicate that it could have some relevance for explaining the identified weaker institutional parameter in the African sub-sample. Keywords: Institutions, social divisions, parameter heterogeneity JEL classification: O10, O17, P14, P26 1 Introduction By now the positive association between institutional quality and economic performance is well documented 1, and few would disagree with assertions like institutions matter. Aiming beyond this arguably uncontroversial conclusion this paper examines whether the relation between institutions and economic performance differs systematically with country circumstances. More specifically it investigates the hypothesis that the association between institutional quality and economic performance is weaker in countries marked by social divisions, and whether this contributes to a weaker institutional parameter in the African sub-sample. The results of the empirical estimations support the social division hypothesis. Furthermore they suggest a weaker institutional coefficient in the African sub-sample, and indicate that social divisions could have some relevance for explaining this regional variation in the institutions parameter. Several considerations motivate this focus. First, there is a methodological debate pointing to the hazards of not taking account of parameter heterogeneity in empirical studies of economic performance. Parameter heterogeneity can be defined as systematic and group-wise parameter variation in cross-section data (Zietz 2005). In this sense, parameter heterogeneity is one form of regression misspecification; surely revealing variation that is systematic, or non-random, is necessary in order to uncover the true parameters of a regression model. Still, this type of misspecification is much less considered than other forms, such as endogeneity and measurement error (Temple 2000). Furthermore, it seems fair to argue that neglecting parameter heterogeneity is especially problematic when dealing with cross-country analysis. In the words of Department of Economics, Göteborg University, Box 640, Göteborg, Sweden. ann-sofie.isaksson@economics.gu.se, Tel. +46-(0) See for example Acemoglu, Johnson and Robinson (2001), Hall and Jones (1999), Knack and Keefer (1995), or Rodrik et al. (2004). 1

2 Brock and Durlauf (2001) the assumption of parameter homogeneity is particularly inappropriate in studying complex heterogeneous objects, such as countries (p. 236). Second, although context specific effects of institutions are rarely allowed for, several authors still acknowledge their existence. For instance, Djankov et al. (2003), Mukand and Rodrik (2005), and Rodrik et al. (2004) all claim that different institutions are appropriate in different contexts, and thus that efficient institutional design depends on the specific characteristics of societies. Going back to North (1994) the message is similar; the same institutional setup will result in different performance in different countries due to variation in enforcement strategies and informal institutions. It thus seems plausible to argue that economic effects of specific institutions should work differently under different circumstances. These concerns motivate the need for taking account of potential institutional parameter heterogeneity in empirical studies of economic performance. Still, the institutions literature contains surprisingly few examples of studies evaluating, or even allowing for, institutional parameter variation 2. Eicher and Leukert (2006) point out that the literature on the economic effects of institutions tends to examine either the global sample 3 or developing countries 4. While the latter alternative makes parameter comparisons more difficult (in terms of degrees of freedom, and since one of the extremes the economically advanced countries is now left out), the former alternative certainly leaves room for taking account of parameter heterogeneity. A couple of recent studies do in fact allow for institutional parameter variation, even though the issue is not their main focus. One of these is the paper by Mehlum et al. (2006) 5 investigating the so called resource curse. In their growth regression, they find a positive interaction effect between institutions and resource abundance, and thus support for their prediction that resource abundance is harmful to growth only when institutions are not well developed. Their discussion is not about the potential for a differential impact of institutions in resource rich countries (which their results could be taken to support), however, but rather deals with how the effects of resource abundance on growth varies with institutional setup. Another study touching on institutional parameter heterogeneity is that of Rodrik (1999), who argues that the magnitude of a growth collapse following a negative external shock to the economy depends on social divisions and the strength of domestic institutions of conflict management. When there are deep social divisions, and when the institutions of conflict management are weak, he argues that the effects of exogenous shocks are likely to be magnified by distributional conflicts. There should thus exist an interaction effect between external shocks, social divisions and the strength of conflict management institutions on growth disruptions. This hypothesis is supported by his findings. Again, however, the focus is not on variation in the institutional parameter, but rather on how the growth effects of external shocks are affected by social divisions and institutional protection 6. 2 In the general growth literature, on the other hand, there are a few examples of studies taking parameter heterogeneity seriously. Worth mentioning in this context are the papers of Block (2001), Canarella and Pollard (2004), Collier and Gunning (1999), Durlauf and Johnson (1995) Masanjala and Papageorgiou (2003a and 2003b). 3 See for example Hall and Jones (1999) 4 Here Eicher and Leukert refer to the sample of former colonies used by for example Acemoglu, Johnson and Robinson (2001 and 2005) 5 For similar analysis see also the paper by Boschini et al. (2003). 6 See also the study of Baliamoune-Lutz (2005), in which the author finds a positive interaction effect between measures of social capital and institutions when looking at African panel data. 2

3 To my knowledge, the only research effort that keeps institutional parameter variation in main focus is that of Eicher and Leukert (2006), who pose the question of whether institutional conclusions with regard to economic performance hold in subsamples of more advanced economies. One of their central findings is that parameter heterogeneity is so marked that when splitting their sample into OECD and non- OECD countries the influential instruments used by Hall and Jones (1999) are invalid. Using an alternative instrumentation strategy Eicher and Leukert find that their institutional indicator has explanatory power in both the OECD and the non-oecd samples, but that the evidence for parameter heterogeneity is strong, with institutions being more important in non-oecd than in OECD countries. A discussion of regional variation in the institutional parameter can be revealing, but it is more interesting if it comes with a theoretical motivation. This paper is, as already noted, particularly concerned with African circumstances. In the regression setup, being interested in to what extent the general conclusions on the association between institutions and economic performance apply in the African sub-sample one could simply evaluate the coefficient of an interaction term between the Africa dummy and the institutional measure in focus. Unfortunately this does not answer the more intriguing question of why the association between institutional quality and economic performance should, or should not, work differently in Africa. This paper seeks to evaluate the underlying reasons for regional variation in the institutional parameter, examining the hypothesis that the association between institutional quality and economic performance is weaker in countries with deep social divisions, and that this gives a weaker institutions-income relation in the African sub-sample. Hence, this study contributes to the literature not only by explicitly focusing on parameter heterogeneity in the institutions-income relation, which is a relatively novel approach in itself, but also by aiming beyond identification of simple regional heterogeneity and evaluating the influence of social divisions on the institutional parameter. The next section seeks to operationalize the concept of social divisions, and to explain the mechanisms whereby they should act to weaken the association between institutional quality and economic performance, and why this should be relevant for Africa. Section 3 outlines the empirical strategy of the paper, including basic econometric specification and choice of variables and data, section 4 presents the results of the empirical estimations, and section 5 sums up the discussion. 2 The link between social divisions and institutional payoffs Why should social divisions act to weaken the positive association between institutional quality and economic performance? In order to be able to clarify the hypothesised linkages we first need to make clear what we mean by social divisions, as well as by institutions. Social divisions could refer to the depth of societal dividing lines along several potential dimensions, such as income, class, ethnicity, or gender. This paper considers social divisions, and their effect on the institutional parameter, along an income dimension, as proxied by measures of income inequality, and along an ethnic dimension, as captured by ethnic fractionalisation indicators. Turning to the concept of institutions this could, along the lines of North (1990), be defined as the formal and informal rules in society. Economists usually interpret the concept more narrowly, however, with the quality of institutions taken to refer to how conducive these rules are to desirable economic behaviour (Rodrik et al. 2004). 3

4 In practice this often translates into studying property rights institutions 7. Considering that this study is partly motivated by the previous neglect of institutional parameter heterogeneity in the literature often focusing on property rights institutions, this paper will follow in this tradition. I am well aware that property rights institutions, let alone the particular indicator examined here 8, far from capture the entire institutional spectrum as it is defined by North, but out of convenience I will at times speak simply of institutions or institutional quality. Speaking of property rights institutions and their effect on economic performance an immediate question arises; property rights for whom? Rich and poor, men and women, people of different ethnic origin, large scale corporations and small scale peasants are all offered the same protection? These questions relate to how well the property rights institutions incorporate the different segments of economic actors in society, what I will refer to as the inclusiveness of the institutional framework. Acemoglu et al. (2002) argue that good institutions should secure property rights for a broad cross section of society. The inclusiveness of institutions, which should not only depend on legal formulations but also on factors like enforcement, or the lack thereof, has to do with the extent to which they live up to this criterion. I view the inclusiveness of the institutional configuration as the most likely mechanism whereby social divisions should affect the strength of the association between institutions and economic performance. It seems reasonable to suggest that in a society marked by social divisions, as opposed to a relatively homogenous community, property rights institutions are more likely to protect, or to be perceived as protecting, one group more than another. Importantly, the perceived inclusiveness of the institutional framework should in this context be at least as important as its actual inclusiveness, considering that this is what influences the economic decisions of individuals. Actual and perceived inclusiveness are likely to be highly related, however, and it seems plausible to argue that social divisions should have a negative influence on both. Arguably, perceived lack of inclusiveness, be it founded in actual circumstances or not, then acts to weaken the positive association between institutional quality and economic performance. Two mechanisms appear important here. First of all, strong property rights institutions are usually argued to induce desirable economic behaviour, such as investment and specialisation 9. It seems reasonable that these behavioural effects should increase with perceived institutional coverage. That is, if broad segments of society consider, rightly or otherwise, that the existing property rights institutions offer them no protection this should make the effects of these institutions on economic behaviour less widespread. As pointed out by Acemoglu et al. (2002), if we only secure the property rights of a small elite then much of entrepreneurial capacity and investment opportunity will be among those without effective property rights protection. To illustrate, if strong property rights institutions encourage investment, then the greater the number of people that feel that they are protected by existing property rights institutions, the greater should be the number of people that do in fact invest. Second, if citizens consider that the institutional framework does not protect their interests this is likely to have a negative effect on compliance with its formal rules. For instance, if property rights institutions are seen as protecting the property of one group more than that of another, this should reduce the legitimacy of those institutions 7 See for example Acemoglu et al. (2001, 2002), and Knack and Keefer (1995). 8 Again, this will be discussed in section 3. 9 See for example Knack and Keefer (1995). 4

5 in the eyes of the group which perceives itself as disadvantaged. Reasonably this group should then be less willing to live by the regulations put forward. A reactance effect of this type, which affects compliance with formal rules, could undermine the rules themselves, and hinder society from fully experiencing their effects. These two mechanisms, operating via the perceived inclusiveness of the institutional framework, should thus work in the same direction to weaken the association between economic performance and the strength of property rights institutions in a country marked by social divisions. So why should this be relevant for Africa? Well, social divisions, as they are measured in this paper, seem to be a reality for many African countries. On the ethnic fractionalisation measure in main focus 10, which ranges from 0-1 with 1 indicating the highest possible ethnic fractionalisation, the mean African score is 0.75 (which is the highest regional average in the sample) to be compared with the non-african mean of Similarly, on the income inequality measure (the Gini index) the mean African score is 46 as compared with the non-african mean of According to the social division hypothesis put forward above, Africa should thus have a weaker institutional parameter than the rest of the sample. To what extent social divisions affect the association between institutional quality and economic performance, and whether they contribute to a weaker institutional parameter in the African sub-sample, will be investigated in section 4. First, however, let us turn to how this question can be approached empirically. 3 Empirical estimation The empirical issue of whether the association between institutional quality and economic performance varies with context (here social divisions) can be approached by regressing the measure of economic performance on explanatory variables including an interaction term between our institutional indicator and the contextual indicator along which the institutions parameter is allowed to vary. The standard OLS benchmark regression will thus take the form: (1) log yi = α + β 1 Insti + β 2 Conti + β 3 Insti Conti + β 4 Xi + ε i where y i is income per capita in country i, Inst i is our institutions indicator, Cont i is the contextual characteristic in focus, Inst i Cont i is the interaction term allowing the institutional parameter to vary along the contextual characteristic in focus, X i is a vector of control variables, and ε i is the random error term. The existence of institutional parameter heterogeneity along the selected dimensions can be evaluated by interpretation of the interaction term parameters, marginal effects, and various sample splits. Being interested in the extent to which social divisions can help explain a weaker institutional parameter in the African subsample we can examine to what degree the coefficient of a regional interaction term (between the Africa dummy and the institutional indicator) survives the inclusion of the interaction terms between the institutional indicator and the respective social divisions variables. 10 The variables used are discussed in section 3.2, and are summarised in table 1 in the appendix. 11 See table 3 in the appendix. 5

6 3.1 A note on endogeneity In specifications like the OLS setup above endogeneity in the income-institutions relation is certainly a concern. The aim of this paper, however, is not to test to what extent institutions affect income, or the other way around. That is, the objective is not to establish the general degree to which institutions matter for economic development. That institutional quality is important for economic development this study takes as given 12, and in reasoning on why social divisions should contribute to institutional parameter heterogeneity I thus allow myself to speak in terms of variation in impacts or effects. When discussing the specific findings of this paper, however, one has to be a bit more careful. The fact that getting around the endogeneity problem is not in focus in this paper could be seen as problematic, but could also be viewed as a way of avoiding results to be blurred by invalid instruments, leaving at least a clean correlation pattern for interpretation. There is a trade-off involved here, and being concerned with dimensions of cross-country variation in the institutions parameter, rather than the coefficient as such, using an invalid instrument comes across as potentially more problematic than the endogeneity issue itself 13. Against this background we can, in interpreting the parameter heterogeneity results, speak of variation in strength of the institutions-income relation, which is indeed the very focus of this paper, but will have to leave discussions of causality to investigations concerned with establishing the causal impact of institutional quality on economic performance. 3.2 Variables and data As dependent variable I use log GDP per capita (in PPP terms) in 2000 obtained from the Penn World Tables. Income levels are arguably more exposed to problems of endogeneity than is growth, but on the other hand one could argue that they provide a better indicator of development, and that the transitory nature of growth rates are not suitable when studying slow structures like institutions 14. Hence income levels rather than growth rates are used 15. As noted the focus is on property rights institutions. To proxy for these I use the measure of protection against risk of expropriation, developed by the International Country Risk Guides (ICRG). This indicator is a subjective assessment of the risk to foreign investors of outright confiscation and forced nationalization of property, ranging from 1-10, with higher values meaning less risk of expropriation. Even though this measure focuses on the risks to foreign investors, it is commonly used to proxy for property rights institutions in a more general sense, which in turn are suggested to have wide reaching economic effects, well beyond those from foreign investment 16. This might well be the best available proxy for general property rights institutions. In any case, the extensive use of measures such as this one highlights the 12 Although sceptical of many econometric attempts to get around the problem of endogeneity in the institutions literature, a vast number of studies do, based on theoretical reasoning, strong correlation patterns, and the quite diverse range of IV based estimations, at this stage point in the same direction institutions are important for determining economic performance. 13 It is quite possible that the paper will be developed by including estimations where institutions are instrumented, but if so, this will be for robustness considerations rather than as benchmark estimation. 14 For reasoning on this issue see Hall and Jones (1999). 15 Growth rates averaged over longer periods could be considered in this context, and I plan to examine to what extent my findings apply using this measure as well. 16 See for example Acemoglu et al. (2001) and Knack and Keefer (1995). 6

7 need for taking account of institutional parameter heterogeneity, since the institutional framework protecting big actors, such as foreign investors, need not to the same extent look after the rights of people who are economically active on a smaller scale 17. According to the reasoning above, the economic impact of property rights institutions, as they are captured by this measure, should be affected by their degree of inclusiveness that is, by how well they incorporate different segments of economic actors in society. Let us now turn to the variables along which the institutional parameter will be allowed to vary. As a first step in the analysis the institutional parameter is allowed to differ depending on the country belonging to Sub-Saharan Africa (henceforth simply Africa) or not. Hence an Africa dummy is included and interacted with the institutional indicator. When it comes to social divisions two main dimensions are considered, namely ethnic fractionalisation and income inequality. The ethnic fractionalisation variable primarily used is the one put forward by Alesina et al. (2003). Just as the classic ethno-linguistic fractionalisation measure used by Easterly and Levine (1997), this indicator gives the probability that two individuals selected randomly from the population belong to different groups. However, the measure of Alesina and his colleagues to a greater extent takes account of ethnic origins when distinguishing between groups 18, and has the advantage that it includes a greater number of observations. Ethnicity is a far from clear-cut concept and ethnic classifications are likely to contain ambiguities. To evaluate the sensitivity of results to different ethnic indicators I therefore consider alternative measures used by Easterly and Levine (1997), Fearon (2003), Alesina et al. (2003), and Montalvo and Reynal-Querol (2005). As a measure of income inequality the Gini index is used. Again, however, alternative indicators are used for robustness considerations 19. These variables, included to capture the depth of social divisions, are then interacted with the institutional indicator. As control variables, the constituent variables of the interaction terms of course need to be incorporated, not only as part of interaction terms. In addition, I control for a number of geographical characteristics and a trade indicator 20, seeing that these 17 For instance, Acemoglu and his colleagues (2002) argue that good institutions should secure property rights for a broad cross section of society, but at the same time they use the ICRG risk of expropriation variable, measuring risk of expropriation to foreign investors, as one of their major institutional indicators. If the institutional measure does not pick up the strength of property rights applying to a broad segment of society, but rather evaluates institutional quality from the perspective of large investors (in principle, a country with virtually non-existent property rights for the majority of its population could still score 10 on this measure if giving preferential treatment to foreign investors), this definition of what constitutes good institutions surely seems to imply that the economic effect of institutions, as they are measure here, should vary between countries depending on their coverage. 18 As opposed to primarily focusing on linguistic distinctions, according to which for example black and white people in the US would be classified as belonging to the same ethno-linguistic group. 19 More specifically, I also consider the ratio of income or consumption of the richest 10 (and 20) percent of the population to the poorest 10 (and 20) percent of the population, the share of income or consumption of the poorest 10 (and 20) percent of the population, and the share of income or consumption of the richest 10 (and 20) percent of the population. 20 In addition, the openness measure of Sachs and Warner (1995) is in some estimations included as a control for policy (the measure does not only look at direct trade policy but also incorporates estimations of black market premium and socialist rule etc.). However, using this measure you loose a number of observations, so in order to be able to stick to my benchmark sample I choose not to present these estimations. Furthermore, for lack of space I do not present the estimations including a variety of regional dummies, but nevertheless allow myself to briefly comment on the results. 7

8 factors, just like institutions, are viewed as standard determinants of economic performance 21. The benchmark sample consists of 87 countries from all over the world, and is only limited by data availability. 21 of these countries are located in Sub-Saharan Africa. For variable definitions and data sources, descriptive statistics for the benchmark sample, and for some summary statistics of key variables, see tables 1-3 in the appendix. 4 Results This section begins by investigating whether Africa does in fact stand out in terms of the association between institutional quality and economic performance. Next, it explores to what extent social divisions affect the institutional parameter by letting the institutional slope term vary with ethnic fractionalisation and income inequality measures. Finally, to get a picture of whether social divisions contribute to a differential institutions-income relation in the African sample it examines the extent to which the coefficient of the regional interaction term survives the inclusion of the social division interaction terms. 4.1 Regional variation in the institutional parameter: does Africa stand out? With Africa scoring high on the main social division variables included 22 one would, in line with the social division hypothesis advanced in this paper, predict that Africa should have a weaker institutional parameter than the rest of the sample. To investigate whether this is so, the first round of regressions gives Africa its own intercept as well as permits its institutional slope term to differ from that of the rest of the sample. Table 5 presents the results of these regressions. First of all one should note that throughout all full sample regressions, including a variety of control variables and the regional as well as the social division interaction terms (as we will see in section 4.2), the coefficient of the institutions indicator is positive and statistically significant, thus seemingly indicating a robust positive association between institutional quality, as it is measured here, and income. That said, let us turn to the potential regional variation in the institutional slope term. From column 1 of table 5 one could note that controlling for a wide range of variables (the institutional indicator, the geographical variables, the trade indicator, and the main social divisions variables), being an African country is still associated with a strong and statistically significant negative effect on economic performance. When including the interaction term between the Africa dummy and the institutional indicator as in column 2-5, however, the parameter of the Africa dummy does not retain its statistical significance. Considering the very high correlation between the regional interaction term and the Africa dummy this is not very surprising. What is interesting in this context is that the interaction term coefficient is negative and (despite the strong correlation with the Africa dummy) statistically significant throughout the regressions, thus seemingly suggesting a weaker positive association between institutional quality and income in African countries See for example Rodrik et al. (2004) who seek to evaluate the relative explanatory power of institutions, geography and trade for determining cross-country patterns in income levels. 22 See table 3 23 Furthermore the regional institutions-africa interaction term parameter remains negative and statistically significant when including all the alternative regional dummies (East Asia Pacific, Europe 8

9 The predicted marginal effect of a change in the institutional index on log GDP log y = inst β + β Africa. From regression 5 [ ] per person is given by: ( ) i. inst Inst Africa we can see that for a one unit improvement in the institutions index the model predicts that income per capita in African countries should increase by approximately 18 percent, whereas in the non-african countries incomes per person are predicted to increase by about 64 percent. For non-african countries the positive marginal effect of a one unit increase in the institutional index (which when the Africa dummy equals zero simply reduces to the institutional parameter) is, as we can see, statistically significant. For African countries, on the other hand, the impact is comparatively small and not statistically significant at the ten percent level 24. Splitting the sample and running separate regressions for African and non-african countries, thereby allowing all slope terms to vary along this regional dividing line, confirms this picture. In the non-african sample we can observe a positive and statistically significant institutions parameter, whereas in the African sample the coefficient is closer to zero (less than half the size of that in the non-african sample) and not statistically significant. The results thus seem to suggest that the positive association between institutional quality and income is less pronounced in Africa than elsewhere. Hence the findings so far do not contradict the social division hypothesis put forward. It remains to be seen how the displayed regional variation in the institutional parameter compares with that along the social divisions dimension. 4.2 Social divisions Let us now turn to the social division hypothesis, postulating a weaker association between institutional quality and economic performance in countries marked by social divisions. To evaluate this idea the institutional parameter is allowed to vary with the measures of ethnic fractionalisation and income inequality, included to capture social divisions in society Ethnic fractionalisation Table 6 presents the results of regressions allowing the institutional coefficient to be conditional on level of ethnic fractionalisation. As exemplified in specification 1 (the same pattern holds when including a range of control variables), before including the interaction term the parameter of the ethnic fractionalisation variable is negative, but not statistically significant. When including the interaction term between ethnic fractionalisation and the institutional measure in specification 2, the parameter of the ethnic variable switches sign to positive, but is still not statistically significant. The coefficient of the interaction term, on the other hand, is negative and almost statistically significant at the five percent level (it has a p-value of 0.059). When including further controls the negative parameter of the interaction term gains in and Central Asia, Middle East and North Africa, South Asia, Western Europe and North America), suggesting that it is relatively robust to controls for level effects originating in structural differences between regions. In addition, including the openness measure of Sachs and Warner (1995) in order to control for policy differences between countries the regional interaction term parameter remains negative and statistically significant. 24 Based on an F-test we cannot reject the hypothesis that β inst +β inst-africa =0 (with a p-value of 0.11 we are not far from being able to do so at the ten percent level, however). 9

10 statistical significance 25. In specification 4 and 5 the positive coefficient of the ethnic variable is also statistically significant, whereas in regression 6, when a squared ethnic term is included to allow for the possibility of a non-linear relationship between ethnic fractionalisation and income, it does not retain this significance (which is not very surprising considering the degree of multicollinearity that is likely to result from three different variables incorporating the same ethnic measure). The statistically significant interaction term parameter indicates that the impact of the two constituent variables (ethnic fractionalisation and institutions) each depend on the value of the other, and thus that they cannot be interpreted in isolation. In fact, in the presence of a significant interaction effect the respective parameters of the component variables do not depict general effects, but rather tell us the impact of a change in one variable when the other indicator equals zero 26. To get a picture of the marginal effect of a change in institutions predicted by the model one thus has to consider both the institutional parameter, the parameter of the interaction term, and the level of the other component (here ethnic fractionalisation) in the interaction term. That said, we can note that in line with the social division hypothesis the negative interaction term coefficient seems to suggest that in countries with a higher degree of ethnic fractionalisation the positive relationship between institutions and income is weaker. To get a picture of the magnitude of change in the institutions-income relation resulting from differences in degree of ethnic fractionalisation, consider the marginal effect of a change in institutions: log yi. = inst[ β inst + β Inst Ethnic ( Ethnic) ]. Based on regression 5 we can see that with ethnic fractionalisation at its mean level, a one unit improvement in the institutions index is predicted to give a 47 percent increase in income per capita. If instead having ethnic fractionalisation at a level one standard deviation above its mean, the same institutional improvement instead generates a 33 percent increase in income per capita. And correspondingly, with ethnic fractionalisation at one standard deviation below its mean the marginal effect of the institutional improvement is an increase in income of about 64 percent. Hence, the lower the degree of ethnic fractionalisation the higher is the predicted marginal effect of an institutional improvement. Similarly, if we split the sample at the mean level of ethnic fractionalisation and run regressions separately for groups with higher and lower ethnic fractionalisation, the institutional parameter in the less fractionalised group is 0.51, to be compared with 25 In addition to the geographical controls and the trade indicator I also include a variable capturing if the country has been struck by civil war in the period between 1960 and 1999 (see table 1), considering that a potential negative influence of ethnic fractionalisation on income could work via this mechanism. Moreover, a squared ethnic term is included to allow for the possibility of a non-linear relationship between ethnic fractionalisation and income. The reasoning of several authors seems to suggest that such a relationship exists. Montalvo and Reynal-Querol (2005) for example argue that the relationship between ethnic diversity and conflict should be non-linear, with less conflict in highly homogenous and highly heterogeneous societies, and the highest risk of conflict occurring in the middle range of ethnic diversity. Similarly, Collier (2001) put forward that ethnic fractionalisation should be less problematic for economic performance than a situation of ethnic dominance, where one group constitutes the majority. These arguments suggest that one should not expect a monotonic relationship between the number of ethnic groups and economic performance, and that factors such as the size of groups, and distance between groups also need to be taken into account. Although a crude approach for dealing with this issue, for the time being, this seems to motivate including a squared ethnic fractionalisation term as control variable, even though it comes out far from statistically significant in this estimation. In addition to specifications with these controls, which appear in table 6, the interaction term parameter is robust to the inclusion of the alternative regional dummies, as well as to the openness measure. 26 see the reasoning of Braumoeller

11 0.27 in the more fractionalised group (both estimates being statistically significant at the one percent level). If, for the purpose of comparison, instead splitting the sample at the mean level in the institutions index, the ethnic parameter in the good institutions group is -0.19, while that in the bad institutions group is These estimates are far from statistically significant, however. In line with the social division hypothesis, the results seem to indicate that the positive association between institutional quality and income is weaker in more ethnically fractionalised societies. As noted, however, ethnicity is a far from unambiguous notion. Hence, to evaluate if the institutional parameter heterogeneity identified is contingent on choice of ethnic fractionalisation measure, the same regressions are carried out for a number of alternative fractionalisation indicators 27. Using the ethno-linguistic fractionalisation variable used by for example Easterly and Levine (1997), the ethnic measure of Fearon (2003), and the language fractionalisation measure of Alesina et. al. (2003) produces similar results. Also, if using Fearon s (2003) measure of cultural diversity, aiming to capture the cultural distance between ethnic groups by estimating the proximity between their languages, the results displayed show a similar pattern. Most importantly, the negative parameter of the interaction term between the ethnic and institutional measures remains. However, if instead of using a fractionalisation measure we use the polarisation index of Montalvo and Reynal-Querol (2005), which aims to capture how far the distribution of ethnic groups is from the bipolar distribution which is taken to represent the highest level of polarisation, the interaction term parameter (which still is of the expected sign) is not statistically significant. Hence, while the hypothesised weaker relationship between institutions and economic performance in ethnically fractionalised societies is seemingly robust to alternative fractionalisation measures, including one attempting to take account of cultural distance between groups, there is no evidence that it applies to ethnically polarised societies Income inequality Turning to income inequality as measured by the Gini index, table 7 presents regressions where the institutional indicator is allowed to vary with this variable. As could be seen from column 1, without the interaction term the parameter of the Gini measure comes out positive and statistically significant. When including the interaction term between our institutional measure and the Gini indicator, as in column 2-6, the Gini parameter remains positive and is now statistically significant at the one percent level. As it turns out, however, the interaction term parameter is, facing the alternative controls 28, also consistently statistically significant at the one percent level, but negative 29. First of all, this seems to suggest that the positive association between institutional quality and economic performance is stronger in societies where income is distributed more equally, as was postulated by the fractionalisation hypothesis. More generally, the statistically significant interaction 27 The results are available upon request. 28 Those displayed in table 7, as well as the alternative regional dummies and the openness measure of Sachs and Warner (1995). 29 As could be seen in column 6, if in line with the hypothesis that the relationship between income and income inequality is characterised by an inverted U-shape including a squared Gini term this pattern remains, with the interaction term parameter negative and the coefficient of the Gini estimate positive (both being statistically significant at the one percent level). The squared Gini term parameter comes out negative, as postulated by the inverted U-shape hypothesis, but is not quite statistically significant at the ten percent level (it has a p-value of 0.117). 11

12 effect implies that the parameters of the constituent variables (institutions and income inequality) cannot be interpreted in isolation. To get a picture of the magnitude of variation in the institutions-income relation depending on the level of income inequality, consider the marginal effect of a change in institutions at different levels of income inequality, as given by log yi. = inst[ βinst + β Inst Gini ( Gini) ]. Characterising the mean Gini score as normal inequality, a Gini one standard deviation above the mean as high inequality, and correspondingly a Gini one standard deviation below the mean as low inequality, we can (based on regression 5) see that with normal inequality the model predicts that a one unit improvement in the institutional index will generate an 46 percent improvement in income per capita. With low inequality, on the other hand, the same institutional improvement should instead give a 67 percent increase in income, whereas with high inequality it should only result in a 28 percent improvement in income. To illustrate the same point somewhat differently, if we split the sample at the mean Gini level and run regressions separately for groups with higher and lower Gini, we can see that the institutional parameter in the less equal group is 0.18, to be compared with 0.51 in the more equal group (the 95% confidence intervals of these estimates do not overlap). For comparison, if instead splitting the sample around the approximate mean level in the institutions index, one can see that the impact of having a more unequal income distribution is 0.03, significant at the five percent level, in the bad institutions group, whereas in the good institutions group the estimate is not statistically significant. Could the variation in the strength of the institutions-income relation at different levels of income inequality depend on choice of inequality measure? Six alternative inequality indicators are considered 30 ; the ratio of income or consumption of the richest 10 (and 20) percent of the population to the poorest 10 (and 20) percent of the population, the share of income or consumption of the poorest 10 (and 20) percent of the population, and the share of income or consumption of the richest 10 (and 20) percent of the population. Using all these measures the parameter of the interaction term between the inequality indicator and the institutional measure was statistically significant and of the expected sign, supporting the finding that the positive association between institutional quality and income is stronger in more equal societies. 4.3 The regional interaction term revisited To get a picture of whether social divisions contribute to the weaker institutional parameter identified in the African sub-sample we can examine the extent to which the coefficient of the regional interaction term survives the inclusion of the social division interaction variables. Table 8 presents the results from this undertaking. All specifications include the constituent variables of the regional and the social division interaction terms. That is, they include the institutional indicator, the Africa dummy, the Gini measure and the ethnic fractionalisation variable. For the purpose of comparisons regressions 1-3 include the regional and social division interaction terms one at the time. Next, regressions 4-6 include the different interaction term in pairs, and regression 7 include them all in combination. 30 The results are available upon request. 12

13 As could be seen from regression 1, the regional interaction term parameter is negative and statistically significant at the one percent level, thus surviving the inclusion of the constituent variables. When including the interaction term between the institutional indicator and the ethnic variable (as in regression 4), however, the size of the regional interaction term coefficient drops by approximately one third (in absolute value), and it is not quite statistically significant (it has a p-value of 0.11). Neither is the negative parameter of the ethnic interaction term, the absolute size of which is reduced by more than one third compared to the case when it was the sole interaction variable (regression 2). When faced with the interaction term between the institutions indicator and the Gini index, as in regression 5, the absolute size of the regional interaction term parameter again drops by about one third (as compared to regression 1), but remains statistically significant at the ten percent level. The coefficient of the Gini interaction is still negative and statistically significant, but compared to the specification where it was the only interaction term included (regression 3) it is now approximately one fourth smaller. When including the regional interaction term and both the social division interaction variables in combination, as in regression 7, the size of the regional interaction term parameter is down to about half the size it had when not controlling for the social division interaction variable (regression 1), and it is no longer statistically significant. Neither is the coefficient of the ethnic interaction variable, which in addition has dropped to less than half of its initial (regression 2) size. The only interaction term parameter that remains statistically significant in this setup, and then only at the ten percent level, is the coefficient of the Gini interaction. The coefficient of the Gini interaction term is also the one most stable in terms of size. Including the interaction terms in combination comes across as potentially problematic considering the likely existence of multicollinearity giving little independent variation in the explanatory variables, and thus large variances of the OLS estimates (which should still be unbiased, however). Multicollinearity is in a way a small sample problem that seems difficult to avoid in cross-country analysis of this type, and even more so when using interaction terms. In the regressions along the social division dimensions (table 6 and 7) in section 4.2 we can note that in spite of potentially inflated variances due to multicollinearity the social division interaction terms, as well as their constituent variables, were statistically significant. The collinearity should become more of an issue, however, when including several interaction terms and their constituent variables in combination. Judging from the variance inflation factors 31 of the regressors in estimation 7, which are all high, this concern appears justified. Looking at the correlation coefficients in table 4, we can first of all note that as might be suspected the correlations between the interaction terms and their respective constituent variables are high. However, with the possible exception of the correlation between the regional interaction term and the ethnic interaction variable, which is quite marked (0.54), the correlations between the interaction terms do not stand out as being particularly high compared to those between the other explanatory variables. Moreover, we should note that the regional interaction term does in fact survive the inclusion of a great number of collinear control variables 32, but when allowing the institutional slope term to vary with degree of ethnic fractionalisation and income inequality, the size of the regional 31 which give the impact of collinearity among the explanatory variables on the precision of the estimation. The variance inflation factor (VIF) for variable X j is 1/(1- R 2 j), where R 2 j is the R 2 we get from regressing the independent variable X j on all the other independent variables. 32 see table 5 as well as table 8 13

14 interaction term parameter drops notably and it is no longer statistically significant. Hence, while problems of multicollinearity should certainly make us cautious when interpreting the results in this section, they should not make us disregard the findings altogether. Based on the estimations put forward it seems fair to argue that social divisions, acting to weaken the association between institutional quality and income, could bear some relevance for explaining the discrepancy between the non-african and the African institutional parameter. Let us consider an illustration. Regression 6 allows the institutional slope term to differ with degree of ethnic fractionalisation and level of income inequality (but not along the regional dividing line between Africa and non-africa). The marginal effect of a one unit improvement in institutions index is in this setup given by: log yi. = β 1 inst + β Inst Ethnic ( Ethnic) + β Inst Gini ( Gini). Evaluating this at the African and non-african mean levels of ethnic fractionalisation and income inequality we get the predicted marginal effects of the institutional improvement in the respective regions. Based on the African mean levels of ethnic fractionalisation and income inequality the model predicts that for a one unit improvement in the institutional index log income per capita should increase by approximately 0.332, whereas based on the mean levels of ethic fractionalisation and income inequality in non-africa the same institutional improvement is predicted to raise log incomes per person by about Judging from their different average levels of ethnic fractionalisation and income inequality the model thus predicts that the marginal effect of a one unit improvement in institutions on log income per person should be smaller in Africa than in the rest of the sample. Compare this to regression 1, where the institutional slope term is only allowed to vary along the regional dividing line between Africa and non-africa. In this setup, the interaction term parameter gives the difference in predicted marginal effect of a one unit improvement in institutions between Africa and the rest of the sample 33. According to regression 1 the marginal effect of a one unit improvement in institutions on log income per capita should thus be smaller in Africa than elsewhere. As seen above, roughly half of this regional difference could be predicted from differences in ethnic fractionalisation and income inequality, in turn affecting the institutional slope term. To sum up, we can note that the weaker African institutional parameter suggested by the regional interaction term coefficient corresponds with the weaker association between institutional quality and economic performance implied by Africa s higher levels of ethnic fractionalisation and income inequality. Unfortunately, too much variance makes it difficult to say anything about the exact magnitude of parameters, but we can note that while surviving the inclusion of a great number of collinear control variables (including all the constituent variables), the regional interaction term parameter cannot to the same extent handle the inclusion of the social division interactions. 33 For African countries the marginal effect of a one unit change in the institutional index is obtained by adding the institutional parameter and the interaction term parameter. For non-african countries (when the Africa dummy equals zero), on the other hand, the marginal effect of a one unit change in the institutional index is simply given by the institutional parameter. 14