Replication material for paper "Bernauer, Böhmelt (2013): Are Economically "Kinder, Gentler Societies" also Greener? Environmental Science & Technology. DOI: 10.1021/es403362m"

Social Policy Index: a model-based approach Jaya Krishnakumar Juan M. Tellez Minnig Abstract This paper focuses on the measurement and determinants of the social policy input of a country. We concentrate on what countries do in order to accomplish their social objectives and analyze if they are performing satisfactorily with respect to what they could potentially do. The social policy of a country is considered as a latent variable measured by indicators and influenced by some exogenous causes and hence a MIMIC model is specified and estimated using panel data. We compare our index with other additional measures available in the literature. 1 Contents 1. Introduction 2 2. Literature review 4 3. Latent Variable Models 7 3.1. Structural Equation Models 9 3.2. Confirmatory Factor Analysis 12 3.3. MIMIC models 15 4. Theoretical Model 18 5. Data and Exploratory Analysis 20 6. Empirical Model 24 7. Results and Country Rankings 25 8. Conclusions 34 References 36 2 1. Introduction A Social Policy Index (SPI) is a multidimensional measure (input-based) of each country's social policy that intends to incorporate its priorities in terms of development; in order to allow rankings and comparisons between countries. (Prasad, 2005b and 2006). The increasing availability of micro as well as macro economic data has been a major contribution to the construction of social indicators since it has allowed to emphasize the multidimensional aspect by providing a wide range of economic, social and political variables. Several authors have proposed different approaches in order to obtain appropriate measures of development, poverty and inequality and try to explain factors that influence them and evaluate policies that countries adopt to improve their economic and social status. Two very well known tools often used to compare countries in terms of economic development and social progress are the Gross Domestic Product (GDP) per capita and the Human Development Index (HDI). This last index is composed basically of three indicators which are health, education and income; three dimensions that are considered, by its creators, as the most representative elements of economic and social development. Most of these indices prioritize the outcome that is the result that a country obtains by applying a particular social policy. UNRISD, pioneer in the involvement of social variables and the creation of social indicators; defines Social Policy as "state intervention that directly affects social welfare, social institutions and social relations. It involves overarching concerns with redistribution, production, reproduction and protection and works in tandem with economic policy in pursuit of national, social and economic goals" 1 . The next section of this dissertation consists in a literature review of articles that used latent variables models for measuring poverty, multidimensional indices or authors that applied this methodology in other areas. Section 3 presents the theoretical framework of some latent 1 UNRISD 2006. Transformative Social policy: lessons from UNRISD, Research and Policy Brief 5, UNRISD. 3 variable models. The final theoretical model retained is explained in section 4. Section 5 gives a description of the data and Section 6 presents the empirical model, linking the theory explained in section 4 with the available data. Section 7 discusses the results of the estimations and shows some country rankings. Finally, some concluding remarks are presented in Section 8. 4 2. Literature Review Various authors have used latent variables models in fields such as politics, poverty measurement, development measurement and underground economy evaluation. Some articles dealing with empirical applications of these methodologies will be presented. The most common software used to estimate latent variable models is LISREL (Jöreskog and Sörbom 1986). Other software often used are EQS (Bentler 1985) and MPLUS. Dreher, Kotsogiannis and McCorriston (2005) analyze the causes and consequences of the phenomenon of corruption around the world. They propose a MIMIC model where corruption is designed as the latent variable. The indicators chosen for this model were GDP par capita, private credit (as a share of GDP), and consumption of cement and endeavours as well as capital control restrictions. Other indicators such as the size of shadow economy were also taken into account but unfortunately they were not available for all countries. Explanatory variables chosen for the model included school enrolment rate, rule of law (score for the quality of the legal system) and legal origin (French, German, British). They influenced the latent variable directly. Among their main results, the authors confirmed that economic losses due to corruption are considerably high especially in developing countries; therefore corruption is clearly an obstacle to economic development. Dreher, Kotsogiannis and McCorriston (2005) analyze the influence of institutions in shadow economy and corruption for 18 OCDE countries by using a Structural Equation Model. These two phenomena are the latent endogenous variables of the model. The results found by these authors showed that the higher the institutional quality, the smaller is the shadow economy. They also found that corruption is negatively affected by the increase of institutional quality. Krishnakumar (2005) derives multidimensional index of human development by using a Structural Equation Model with the variant of incorporating a set of exogenous variables in the measurement equation (the equation explaining the observed endogenous variables) and another set of exogenous variables in the structural model. The latent endogenous variables are knowledge, health and political freedom, the three most representative dimensions of development. These variables are called capabilities (set of choices). The main result of this paper is that the improvements of political and social environment facilitate the realization of 5 capabilities and raises, very significantly, the level of capabilities themselves. The author concludes by elaborating “capability indices” from latent variables score estimates in order to compare these results with Human Development Index (HDI) and GDP per capita. It is found that countries change significantly their ranking depending on the index chosen. Breusch (2005) uses MIMIC models in order to estimate the size of the underground economy for numerous countries. The author uses three different applications from other papers. He concludes that the results obtained are very different, not convincing enough and that the use of MIMIC models is not appropriate to fit an underground economy. Furthermore, it seems interesting to mention that there are some other studies treating latent variable models. We can cite Hansen, Heckman and Mullen (2003); Heldt, Johansson and Vredin (2005); Jöreskog, K.G. and A.S. Goldberger (1975), Nagar and Basu (2001). Literature offers several approaches for the construction of multidimensional indices. They range from simple or weighted aggregates of different indicators to model-based multidimensional indices incorporating structural relations. In the latter approach, the multidimensional concept that one wants to capture is assumed to be latent, measured through many indicators covering many dimensions. This approach not only yields an aggregate index involving different dimensions but also incorporates the interdependencies and causalities among these dimensions as described by the underlying structural model. This approach which has now become relatively common in well-being measurement (Di Tommaso 2006, Krishnakumar 2007a, 2007b, Krishnakumar and Ballon 2007, Wagle 2005), is also being applied in other areas (Di Tommaso, Raiser and Weeks 2006). We find this methodology to be particularly relevant in our context as the social policy that a country puts in place is largely determined by the availability of resources (economic, political, and institutional). Thus we can speak of a "feasible" policy given the means and an "actual" policy as reflected by the policy input indicators in various dimensions. This way of looking at both the causes and indicators of any social policy fits in well with the MIMIC (multiple indicators multiple causes) approach and will help us evaluate if countries are delivering what is "feasible" given their economic and institutional context, in addition to ranking them in terms of actual policy. 6 Krishnakumar and Ballon (2007) present a framework in order to evaluate countries' social policy effort by using a structural equation model that takes into account the explanations offered by economic theory and the resulting determinants. The authors use the MIMIC model in order to obtain a social policy index that can be used as a basis to compare countries. However, their analysis was only limited to two dimensions of social policy: social spending and social security. Further, the above study only used cross-sectional data whereas we propose to extend the data base to a panel data setting. 7 3. Latent Variable Models Latent variables are a representation of concepts that cannot be observed or measured. In order to have an accurate illustration of a specific latent variable, indicators of this latent variable can be selected. Indicators are observable measures that are assumed to be highly correlated (they should even be perfectly correlated in theory) with latent variables. In some cases indexes are constructed by using various indicators in order to have a better representation of the concept. Consequently, an indicator of a latent variable will always include a systematic measurement error, which defines this indicator as a random variable. Indicators are also known to be called measures, manifest variables and proxies. Latent variables can also be designated as unobserved variables, factors or unmeasured variables. Latent variables can be differentiated by their level of abstractness. Among highly abstract latent variables we have power, social class and intelligence. Education and population size are less abstract latent variables. The measure, for example, of intelligence cannot be directly observed as it has been implicitly mentioned. This is a particular case because a single observed variable is not taken as an indicator, since it would be very restrictive. Hence, a proposed solution consists in constructing a test composed of questions in numerous areas such as mathematics, geography and history. At the end of this test each person receives a numerical result that can be compared with the results obtained by the other individuals. Nevertheless, even though these kinds of tests do not pretend to completely evaluate such a variable as intelligence, they can give an idea of the intellectual coefficient of an individual. Another particular case is the phenomenon of suicide. It should normally be categorized as a directly observed variable. But sometimes some suicides are identified as other forms of death due to the sensitive character of the subject. Therefore, since there seems to be a lack of correct information, suicide cannot be directly observed. Taking into account what has been presented above; we realize that if we are interested in applying econometric estimations to observed variables that are imperfect measures of their corresponding latent variables, customary procedures should not be employed. These procedures give inconsistent estimators because of the presence of measurement errors. 8 Three different models will be developed in this chapter. For each model the definition of variables, assumptions, covariance matrices and methods of estimation will be explained. In the first section the general model or Structural Equation Model (SEM) will be analyzed. Jöreskog and Sörbom have importantly contributed to the popularization of this model in the 1970’s. Later on, in 1980 other authors such as Bentler and Weeks proposed several modifications to the original model. Regarding the second section we shall continue to explore the Confirmatory Factor Analysis Model (CFA). It is a specialization of the general model that is much simpler. This model was initially proposed by Spearman in 1904 in an article concerning the measurement of intelligence. In the third section we shall proceed to explain the Multiple Causes and Multiple Indicators Model (MIMIC). It is also a specialization that is simpler than the Structural Equation Model but more complex than the Factor Analysis Model. The MIMIC models were originally introduced by Jöreskog and Goldberger (1975). This model is the one which will be selected to perform our analysis in this paper. 3.1. Structural Equation Models This is a model where latent endogenous variables are influenced by other latent endogenous variables and by latent exogenous variables. This interdependent system of equations of unobserved variables, called structural model, is completed by measurement equations, called measurement model, where both latent endogenous variables and latent exogenous variables are observed through a set of indicators. This framework constitutes the “general model”. 15 each factor, the quantity of unobserved variables are determined before the analysis and the measurement errors can be correlated over time. 3.3 MIMIC Models The Multiple Indicators and Multiple Causes Models are another particular case of the Structural Equation Models. In the context of this model a single latent variable not only manifests itself through some observed variables but it is caused by other exogenous variables that can be measured with exactitude. The measurement equation contains, as always, measurement errors and the causal relationship (the equation implying exogenous variables that affect the factor) contains disturbance terms as well. The model can be written as: ζη +Γ= x ' εη +Λ= y y ξ = x η denotes a scalar representing a latent variable or factor. ( 1 1 × ). y denotes the vector containing observed indicators of η ( 1 × p ). y Λ denotes the coefficient vector relating y to η ( 1 × p ). ε denotes the vector containing the measurement errors for y ( 1 × p ). Γ denotes the coefficient vector for latent exogenous variables ( 1 × n). ζ denotes the scalar representing the error term in causal relationship ( 1 1 × ). x denotes the vector containing exogenous variables ( 1 × n). Θ denotes the covariance matrix of ε ( p p × ). φ denotes the variance of ζ ( 1 1 × ). 16 The following diagram represents the relations between variables: 1 x 2 x 1 y 3 x 2 y . η 3 y . . . . n x p y It is assumed that the error terms ε and ζ have zero means and are non correlated with each other. In some cases the hypothesis of Θ as a diagonal matrix is done as well. In other words we suppose that measurement errors are non-autocorrelated. A sufficient but not necessary condition for identification of this model is that the number of indicators is equal or greater than two and the number of exogenous variables is equal or greater than one. This means that: 2 ≥ p and 1 ≥ n The expression of the covariance matrix of y is given by: )'()( yyE y =Σ θ )]'')([( ' εηεη +Λ+Λ= yy E )'()'( ' εεηη EE yy +ΛΛ= Θ+ΛΛ= 'yy φ Sometimes by convention the variance of the factor is usually assumed as the unity ( 1 = φ ). 17 So: Θ+ΛΛ=Σ ' )( yyy θ To estimate the parameters the same procedures already exposed in the other sections can be used. For this model the Maximum Likelihood method (ML) is typically employed. For all the models that have until now been explained, alternative estimators have been proposed when numerical, statistical or econometric problems appear. If by applying ML, GLS and ULS the number or iterations are considerable so that the computationally cost is very high, Instrumental Variables estimator (IV/GMM) is an appropriate option for this inconvenience. In cases where the normality hypothesis is restrictive the use of Weighted Least Squares is highly recommended. Both underlying alternatives give consistent estimators. 18 4. Theoretical Model The theoretical model for this framework is a mixed MIMIC model for two-level data. The model will contain a random intercept in order to capture the specific characteristics of each country. Since there are no time effects, we can consider we are in a multi-level model where individuals (countries in our case) are clusters. This is to separate changes within one individual and differences between individuals. Therefore, the whole model can be separated in two parts. The within part (variation across individuals and over time) can be written as follows: (1) jititjwitjjijjit uwfy ++++= ' δλµα Jj ...1 = Tt Ni ...1 ...1 = = (2) ititwit vxf += ' β where (1) represents the measurement equation and (2) represents the causal relationship, J is the number of indicators , N is the number of individuals, T is the number of years, and: jit y denotes a scalar representing indicator j for individual i at time t , j α denotes the specific intercept of indicator j , ji µ denotes a random intercept of indicator j and individual i , j λ denotes the factor loading relating the latent variable wit f to indicator j , wit f denotes the within latent variable for individual i at time t , j δ denotes a vector containing coefficients relating observed exogenous variables to indicator j . it w denotes a vector of observed exogenous variables for indicator j , it u denotes the error term in equation (1). between latent variable within latent variable 19 it x denotes a vector of observed exogenous variables for the within latent variable, β denotes a vector of coefficients relating it x to wit f, it v denotes the error term in (2). The between part of the model (variation only between clusters) can be written as follows: (3) jibijji f ερµ += Jj ...1 = Ni ...1 = (4) iibi emf += ' θ where (3) represents the measurement equation and (4) the causal relationship of this level, and: j ρ denotes the factor loading relating bi f to random intercept ji µ , bi f denotes the between latent variable for each individual i , ji ε denotes the error term in equation (3), θ denotes a vector of coefficients relating i m to bi f , i m denotes a vector containing observed exogenous variables in equation (4) that vary only between clusters (time invariant), i e denotes the error term in equation (4). 20 5. Data and Exploratory Analysis This empirical application needed different types of variables such as economic, social, political and demographic variables. Because of this fact, the data was collected from different sources: World Development Indicators (WDI), Governance Matters, International Social Security Association (ISSA), Country Indicators for Foreign Policy (CIFP), International Country Risk Guide (ICRG), Human Development Reports (HDR), International Labour Organization (ILO), World Institute for Development Economic Research (UNU-WIDER) and KOF Index of Globalization. The main objective was to find data for as many countries and years as possible in view of the fact that one of the interests of this paper is working with cross sectional and time series data and that one can expect to have an important number of missing values in this case. And it was confirmed when the first database was constructed. First of all, some countries had almost no data at all and they were consequently removed. Second, even if there were data from 1998 to 2006, after some exploratory analysis it was decided to drop years 1998, 2005 and 2006 because some variables were completely missing for these periods. Third, countries that had no social security variables at all were also removed, since these variables were available either for a country in all periods or not available in any period. The finally data set consists in 6 years (1999-2004) for 98 countries. Removing all countries with missing data would not have left enough countries to perform our analysis. This could probably create numerical convergence problems for our parameters estimations. Starting by the fact that there is not a single year with complete data for all variables. However, dropping variables was not a solution either because it implies taking away possible explanation for the model without even testing it. Therefore, it was inevitable to contain missing values in the final database in spite of efforts made to reduce this inconvenience as much as possible. This data allows us to consider two policy dimensions: social expenditure and social security. There are more policy dimensions that would have been interesting to study but finding panel data for those other dimensions was a major problem. 21 The latent variables or capabilities are the potential social spending and the potential social security policies. For the first one there are two indicators: public health expenditure as a percentage of GDP and public education expenditure as a percentage of GDP. For the second one there are three indicators being the government's coverage and redistribution of social security program for three categories: old age, disability and survivors, sickness and maternity and work injury. These variables were found in text form and were recoded by using the U.S Social Security Administration Methodology. This system uses a scale from 0 to 9, where 9 represents the maximal coverage and 0 corresponds to an absence of coverage. Table 1 shows the scores of this variable in detail. Table 1 Score Type of coverage and redistribution of the social security program 9 Universal 8 Compulsory social insurance with state subsidy 7.5 Compulsory social insurance with no state subsidy 7 Compulsory insurance with state subsidy, but one or more groups excluded 6.5 Compulsory insurance with no state subsidy, with one or more groups excluded 6 Contributory social insurance, but risk based differentiation with state subsidy 5.5 Contributory social insurance, but risk based differentiation with no state subsidy 5 Private (mandatory) with no state subsidy 4 Provident fund 3 Employer liability 2 Social assitance (means tested) 1 Voluntary private insurance 0 No social policy Exogenous variables that influence capabilities and exogenous variables that influence the indicators were selected from Ballon and Krishnakumar (2006), Baquir (2002), Prasad (2006). These are variables that based on economic theory, determine the government's size and composition. Three sorts of variables were used as exogenous variables. Economic and demographic determinants were used in the structural model; political determinants were included in the measurement model. Table 2 gives a description of all our variables. 22 Table 2 F* Potential social policy y Actual Policy - observed input indicators F1* Potential social spending HEALTH Public Health Expenditure % GDP EDUC Public Education Expenditure % GDP ODS Old age, disability and survivors F2* Potential social security SM Sickness and maternity WI Work injury x Oberved exogenous of endogenous w Observed exogenous of indicators (in the structural model) (in the measurement model) Economic determinants Political Determinants GDP Per capita GDP VA Voice and Accountability GROWTH Growth rate PS Political Stability DEF Fiscal deficit as a % GDP GV Government Effectiveness KOF KOF index of Globalization COR Control of Corruption GINI GINI Inequality Measure Demographic determinants URBAN % of urban population (urban) POP014 % population 14 years old or younger POP65 % population 65 years old or older LIFE Life expectancy (in years) All the exogenous variables are continuous variables. Political determinants are measured in units from about -2.5 to 2.5, where the higher the value of the political score the better the outcome of governance. It is relevant to explain the difference between the variables "Voice and Accountability" that can be considered as a Democracy index by the fact that it measures the participation to elect the government and the freedom of expression and press; whereas "Political Stability" represents how citizens believe that their government can be affected by internal or external conflicts such as terrorism. The KOF index of Globalization takes into account economic, social and political dimensions. These dimensions are measured by using actual economic flows, economic restrictions, data on information flows, data on personal contact and data on cultural proximity. 23 Two interesting facts were found when carrying out an exploratory analysis of the data. The first one is the very small variation for social security variables within clusters. Table 3 shows that the between variance represents at least 90% of the overall variance for these variables. Besides, since the coverage is not a continuous variable, it is almost certain that the within variance does not come from small changes over the years but from one or two medium changes during the whole period. This suggests that this variable is almost invariant within clusters, which could bring some problems when estimating the within model. Table 3 Variable Mean Std. Dev. Min Max ODS overall 6.909014 1.379691 2 9 between 1.354222 2 9 within 0.2919758 5.659014 9.07568 SM overall 6.105442 2.235101 0 9 between 2.217341 0 9 within 0.3477844 1.522109 9.438776 WI overall 6.691327 1.509045 3 9 between 1.402737 3 8 within 0.5712321 3.691327 10.02466 The second fact to take into account is the very high level of correlation between political determinants. Table 4 reports the matrix correlations between the political variables. One can expect that one or two of these determinants are able to capture the effect of the others. Thus, putting all these variables together in the model may give significance only for one or two of them. 24 Table 4 VA PS GOV COR VA 1.0000 PS 0.8656 1.0000 GOV 0.8754 0.8469 1.0000 COR 0.8447 0.8255 0.9706 1.0000 6. Empirical Model The potential or feasible level of social policy that a country can supply is determined by its economic, social and demographic contexts. But in reality this level is not attained because of political and institutional frameworks and other specific characteristics of each country. As a result of this situation, we have a difference between the "potential" social policy, and the "actual" policy performed by the country. The potential social policy is our latent variable (or factor), this level being unobservable, and it is represented by causal relationships in the theoretical model. The difference between the two policy levels is given by measurement equations in the theoretical model. The relations of this model can be described as follows: The potential social expenditure determines health expenditure and education expenditure and is caused by GDP per capita, fiscal deficit, KOF Index of Globalization and Gini Index. Public health expenditure and education expenditure are influenced at the same time by political variables. The potential social security determines the coverage of the three categories mentioned above and is caused by the four demographic variables plus the growth rate. The coverage and distribution of social security programs are influenced at the same time by political variables. 31 Due to the missing values on the covariates of our database, we do not have latent factors for the 98 countries that were included in the analysis either for the six years. The number of countries having scores for each year is unbalanced. The "potential social spending" factor is available for the period 2000-2004 for 29 countries. Table shows the ranking of these 29 countries in term of "potential social spending" from 2000 to 2004. If the country has a (+) sign, it means that it has improved its position compared with the past year. If it is a (-) sign, it means that the country got a worse position compared with the past year. Finally if nothing is written next to the country, it means that its position did not change from one year to the other. Table 11 RANK_00 RANK_01 RANK_02 RANK_03 RANK_04 Luxembourg Luxembourg Luxembourg Luxembourg Luxembourg Denmark Sweden (+) Norway (+) Austria (+) Austria Sweden Norway (+) Sweden (-) Norway (-) France (+) Germany Germany Denmark (+) United Kingdom (+) Norway (-) Belgium Denmark (-) United Kingdom (+) Sweden (-) United Kingdom (-) United Kingdom United Kingdom Germany (-) Denmark (-) Sweden (-) France Belgium (-) France (+) Germany (-) Belgium (+) Norway France (-) Belgium (-) France (-) Denmark (-) Austria Netherlands (+) Netherlands Netherlands Germany (-) Netherlands Finland (+) Finland Belgium (-) Finland (+) Spain Austria (-) Australia (+) Finland (-) Netherlands (-) Australia Spain (-) Spain Spain Portugal (+) Finland Australia (-) Austria (-) Portugal (+) Spain (-) Portugal Portugal Portugal Australia (-) Australia Uruguay Uruguay Uruguay Hungary (+) Slovenia (+) Slovenia Slovenia Slovenia Slovenia Latvia (+) Estonia Czech Republic (+) Hungary (+) Czech Republic (+) Hungary (-) Latvia Hungary (+) Czech Republic (-) Bulgaria (+) Uruguay (+) Czech Republic Poland (+) Bulgaria(+) Lithuania (+) Lithuania Hungary Latvia (-) Poland (-) Ukraine (+) Czech Republic (-) Lithuania Estonia (-) Latvia (-) Poland (-) Ukraine (-) Bulgaria Lithuania (-) Estonia (-) Latvia (-) Bulgaria (-) Poland Bulgaria (-) Ukraine (+) Estonia (-) Poland (-) Ukraine Ukraine Lithuania (-) Uruguay (-) Estonia (-) Venezuela, RB Costa Rica (+) Costa Rica Costa Rica Sri Lanka (+) Sri Lanka Sri Lanka Sri Lanka Sri Lanka Costa Rica (-) Costa Rica Peru (+) Venezuela, RB (+) Venezuela, RB Venezuela, RB Peru Venezuela, RB (-) Peru (-) Peru Guatemala (+) Guatemala Guatemala Guatemala Guatemala Peru (-) There is no a "potential social security" factor varying over the years as this factor was not significant in the estimations. A between factor is available which is invariant over time. 32 The year with the greatest number of factors is 2003. There are estimates for 68 countries. Table 12 reports for the 68 countries in 2003 their ranking of: potential social spending (MIMIC1), potential social spending plus potential social security (MIMIC2) and Human Development Index (HDI). The last two columns of this table show the difference of HDI and MIMIC1 and MIMIC2. If this difference is positive, then the country is not using all its potential and its performance is worse than one could have expected. If this difference is negative, it illustrates that the country's social policy is giving better results than expected given its social, economic demographic and political situation. Table 12 COUNTRY MIMIC1 MIMIC2 HDI DIFFERENCE DIFFERENCE HDI-MIMIC1 HDI-MIMIC2 Norway 3 2 1 -2 -1 Australia 19 39 2 -17 -37 Luxembourg 1 1 3 2 2 Canada 15 19 4 -11 -15 Sweden 6 7 5 -1 -2 Ireland 14 21 6 -8 -15 Belgium 12 8 7 -5 -1 United States 4 16 8 4 -8 Netherlands 11 11 9 -2 -2 Finland 13 18 10 -3 -8 Denmark 7 29 11 4 -18 United Kingdom 5 3 12 7 9 France 10 4 13 3 9 Austria 2 6 14 12 8 Italy 8 5 15 7 10 New Zealand 28 10 16 -12 6 Germany 9 9 17 8 8 Spain 17 12 18 1 6 Israel 20 22 19 -1 -3 Greece 16 13 20 4 7 Singapore 27 61 21 -6 -40 Slovenia 22 14 22 0 8 Portugal 18 28 23 5 -5 Korea, Rep. 37 36 24 -13 -12 Cyprus 23 15 25 2 10 Czech Republic 25 30 26 1 -4 33 COUNTRY MIMIC1 MIMIC2 HDI DIFFERENCE DIFFERENCE HDI-MIMIC1 HDI-MIMIC2 Argentina 34 55 27 -7 -28 Hungary 21 20 28 7 8 Poland 31 32 29 -2 -3 Chile 38 40 30 -8 -10 Estonia 33 25 31 -2 6 Lithuania 29 27 32 3 5 Slovak Republic 36 34 33 -3 -1 Croatia 24 23 34 10 11 Uruguay 35 31 35 0 4 Costa Rica 41 44 36 -5 -8 Latvia 32 24 37 5 13 Bulgaria 26 17 38 12 21 Trinidad and Tobago 42 45 39 -3 -6 Malaysia 45 59 40 -5 -19 Russian Federation 39 33 41 2 8 Colombia 43 37 42 -1 5 Albania 50 38 43 -7 5 Thailand 53 53 44 -9 -9 Venezuela, RB 54 35 45 -9 10 Ukraine 30 26 46 16 20 Peru 55 46 47 -8 1 Philippines 57 47 48 -9 1 China 47 41 49 2 8 Tunisia 52 48 50 -2 2 Jordan 66 62 51 -15 -11 Sri Lanka 48 65 52 4 -13 Jamaica 46 43 53 7 10 Iran, Islamic Rep. 61 51 54 -7 3 Indonesia 60 58 55 -5 -3 Nicaragua 64 52 56 -8 4 Bolivia 44 49 57 13 8 Guatemala 56 42 58 2 16 South Africa 51 63 59 8 -4 Morocco 49 60 60 11 0 India 58 57 61 3 4 Pakistan 65 54 62 -3 8 Nepal 62 66 63 1 -3 Ghana 68 64 64 -4 0 Uganda 59 67 65 6 -2 Madagascar 40 56 66 26 10 Kenya 67 68 67 0 -1 Cote d'Ivoire 63 50 68 5 18 34 Considering only the potential social spending (MIMIC1), the biggest negative differences come from Australia, Canada, New Zealand, Korea, Jordan, Thailand and Venezuela. These countries are doing better than their potential. Meanwhile, the biggest positive differences come from Austria, Bolivia, Croatia, Ukraine, Morocco and Madagascar. These countries are not functioning at their best level. 8. Conclusions In this paper a latent variable methodology was applied to a social policy context. The aims of this dissertation were to provide an accurate theoretical framework in order to obtain scores of potential social policy, which make possible the differentiation of "actual policy" from "feasible policy" (capability), and also to calculate individual effects (random intercepts) for each country given the structure of the data. The model appears as an appropriate choice to this particular context, showing that economic, social, demographic and political variables influence the level of social policy that a country can provide. In the social spending dimension there were no problems during the estimation of the model, while some inconveniences were found in the social security dimension du to the small within variation and the nature of the data. This paper shows data limitations when the number of missing values is important. Because of this problem countries and years have been excluded of the dataset before and during the estimations. In spite of this, some ranking tables were built with a reasonable number of countries that allow comparisons over time and with other indices such as HDI. It revealed that some countries do not perform at their potential while some others do better taking into account their situation, which was one of the objectives of this dissertation. 35 Finally, for future research it will be interesting to fill missing values of the database, to find other pertinent variables that could be add to the empirical model, to analyze more social policy dimensions and to look for economic explanations to the ranking variations. 36 References Ballon, P. and Krishnakumar, J. (2007), “Estimating Basic Capabilities: A Structural Equation Model Applied to Bolivia”, Working paper, World Development (forthcoming). Baquir (2002), "Social Sector Spending in a Panel of Countries", IFM Working paper, WP/02/35. Bartholomew, D.J. and Knott, M. (1999), Latent Variable Models and Factor Analysis, Edward Arnold, U.K. Bollen, K.A. (1989), Structural Equations with Latent Variables, John Wiley & Sons, New York. Breusch, T. (2005), “Estimating the Underground Economy using MIMIC Models”, The Australian National University. Di Tommaso, M. (2006), "Measuring the Well Being of Children using a Capability Approach: An application to Indian data", Journal of Socio Economics (forthcoming). Di Tommaso, M., Raiser, M. and Weeks, M. (2006), "Home Grown or Imported? Initial Conditions, External Anchors, and the Determinants of Institutional Reform in the Transition Economies", Economic Journal (forthcoming). Dreher, A., Kotsogiannis, C. and McCorriston, S. (2005), “How do Institutions Affect Corruption and the Shadow Economy”, University of Exeter. Hansen, K., J. Heckman and K. Mullen (2003), “The Effect of Schooling and Ability on Achievement Test Scores”, Institute for Labour Market Policy Evaluation (IFAU), Working Paper 2003:13. Heldt T., P. Johansson and M. Vredin (2005), “Latent Variables in a Travel Mode Choice Model: Attitudinal and Behavioural Indicator Variables”, Uppsala University. 37 Hsiao, C. (2003), Analysis of Panel Data, Cambridge University Press. Hsiao, C. (1991), “Identification and Estimation of Dichotomous Latent Variables Models Using Panel Data”, Review of Economic Studies, 58, 717-731. Jöreskog, K.G. and Goldberger, A.S. (1975), “Estimation of a Model with multiple Indicators and Multiple Causes of a Single Latent Variable”, Journal of the American Statistical Association, 70, 631-639. Krishnakumar, J. (2007a), “Going Beyond Functionings to Capabilities: An Econometric Model to Explain and Estimate Capabilities”, Journal of Human Development, 7, 39-63. Krishnakumar, J. (2007b), “Multidimensional Measures of Poverty and Well-Being Based on Latent Variable Models”, in N. Kakwani and J. Silber (eds.), Quantitative Approaches to Multidimensional Poverty", Palgrave Macmillan (forthcoming). Lin, X. and Roy, J. (2000), "Latent Variable Models for Longitudinal Data with Multiple Continuous Outcomes", Biometrics, 56 (4), 1047-1054. Muthén, L.K. and Muthén, B.O. (1998-2004), Mplus user's guide. Third Edition. Los Angeles, CA: Muthén & Muthén. Nagar, A. and Basu, S. (2001), “Weighting Socio-Economic Indicators of Human Development (A Latent Variable Approach)”, National Institute of Public Finance and Policy, New Delhi. Prasad, N. (2005a), "Construction of a Social Policy Index", UNRISD Working Paper, UNRISD, Geneva. Prasad, N. (2005b), "Social Policy Index", UNRISD Working Paper, UNRISD, Geneva. Prasad, N. (2006), "Effects of Economic Globalization on Social Spending in Developing Countries", forthcoming. 38 Skrondal, A. and Rabe-Hesketh, S. (2004), Generalized Latent Variable Modeling: Multilevel, Longitudinal, and Structural Equation Models, Chapman and Hall/CRC, Boca Raton, U.S.A. UNDP (2006), Human Development Report 2006. Wagle, H. (2005), "Multidimensional Poverty Measurement with Economic Well-Being, Capability and Social Inclusion: A Case From Kathmandu, Nepal", Journal of Human Development, 6 (3), 301-328. http://www.ilo.org http://hdr.undp.org/en/statistics/ http://www.undp.org/ http://www.carleton.ca/cifp/descriptions.htm http://info.worldbank.org/governance/wgi2007/ http://www.prsgroup.com/ http://www.issa.int/engl/homef.htm http://devdata.worldbank.org/dataonline/ http://www.wider.unu.edu/ http://globalization.kof.ethz.ch/query/ 39