2 Tax Structure, Trade Taxes, and Economic Development: An Empirical Investigation Theo Hitiris* Tax structures vary because countries have different economic structures and governments have diverse economic, social, and polit­ ical objectives and varying abilities tO use rhe tax system. During economic development the objectives of tax policy change, a wider range of possible tax bases becomes accessible, and the capability of countries ro administer broad-based tax systems improves. Economic development changes the economic structure, the aims of economic policy, and the tax structure. In theory and practice, the develop­ ment of tax structure means a rise in the share of direct taxes and thus an increase in the progressivity of the tax system. The evolution of trade taxes is frequently presented as typical of the changes of the tax system during economic development. Import and export taxes have an easily identifiable and accessible tax base that provides an administratively convenient and effective source of tax revenue for countries in their early stages of economic develop­ ment (Musgrave ( 1 969)). As the economy grows, government expen­ diture as a ratio of GOP tends to rise. At the same time, the principal role of foreign trade taxes also changes from revenue raising ro protecting domestic production against foreign competition and pro­ moting exports. Consequently, the importance of trade taxes as a source of tax revenue is, arguably, inversely related to the relation­ ship between economic development and tax structure, with particu­ lar emphasis on foreign trade taxes (Carden ( 1 974), Dixit ( 1985)). 'Theo Hitiris is Senior Lecturer in economics at the University of York. England. He was formerly a research economist at the Center of Planning, Athens, Greece. His main writings and research interests are in international trade theory and policy and the economics of the European Community. He is grateful 10 the editor, two referees, and ). P. Hutton for constructive comments on this paper. The remaining errors are his alone. 29 ©International Monetary Fund. Not for Redistribution 30 THEO HITIRIS T h i s paper studies the qua n t i tat ive aspects of an obse rved set o f c ross-se c t i o n data in a n attempt to d iscover empi rical regu l a r i t ies i n the economic developme n t . T h e next section i nvestigates the fiscal importance of trade taxes during economic development <tnd how t hese taxes re late to other macroeconomic variables . The final sec­ tion offers a summary of the res u l t s and conclu ding observa t ions. Tax Structure and Economic Development The analysis is based on a set of observed cross- ect ion data from l 05 industrial and develop i ng cou n t ries. 1 In the data se t , each cou n ­ t r y ' s t a x st ruct u re is desc r i bed by t h e al location of tota l tax revenue among ten di fferent taxe s , comprising fou r broad tax types : Income taxes, I T: personal, IP; social secu rity, IS· corporation a n d o t her, /C. Domestic taxes on goods and services, S T: sa les, SS; excises, SX; other SO. Trade taxes, TT: import dut ie s , TM; export t axe s , TX . Other taxes, O T; weal t h and p r o pe r t y taxe s , O W; o t her, 00.2 In addition to t hese data, the fol lowing variables of t h e cou n t r ies in the sample are also conside red: G O P per capita (in . S . dol lars ) , Y, as a proxy for t h e le ve l of economic developme n t ; t he share of i m ports , exporrs, a n d total trade effe c t i ve t rade tax rat e , V in GDP (M. = X, and T, respect ively); t he trade tax/trade val u e ; and t h e total tax burde n , G = t ot a l tax revenue/G D P. ' h is in f�ct the dt:taik:d sa mp lt: of developing coumries published in Tanzi ( 1 987). augmented hy t h e ad d i t io n of t h e industrial count r i c� - Tanl.i 's sample of 86 developing hascd nn bot h publ ished �nd u npublished data collected by the I nterna­ Fund's B u rea u of tatist ics. From t his sample we h;n·e omit tt:d one nu n , which is considered an oil-exporting economy. and added 20 i ndus­ <.:oun t ry, trial o.:ou n t rit:s for wh ich data comparable 10 t hose of t he developing cou m rit:s are ava i lable (data for japan arc not ;!Vailablc). For cc:nain nf the: countries in the: sam p k the dar� refer to central government revenue only (sn: T:mzi ( 1 987)). For t h t: devel­ oped count ries, the t:tx revenue data come from I nternational Monetary Fund. Cor­ emmenl Financial Statistics Vearbook. Vol . VI (Was h i ngton, 1 9R2), :mel art: :.t\'erage. of t h e years 1 979- 1 982: the sha re of t rad<.: in G O P. effect ive t;tx r:ne. and total tax countries is tional Monetary IJu rdc:n come from International Monet:try Fund, International Financial Stat istics: 7i'ade Statistics, Suppl<.:ment o . -i (Wasl1ingto n . 1 98-), and Supple· me111 011 Co1•emnumt Filla liCe, Supplement N o . I I (Was h i ngto n . 1 981l). For ;til the.: I 06 coumries r h e G P per head v�riable, l', for 1 98 1 at 1 980 m:uket p r i ces in l ' . S . doll:trs, (mid-year popuhn ion) cornt:s from l n te rn a t i on:t l Monetary Fund, lntematio11al Finan­ talislics. Su ppkmt:nt o. H ( Wa�hington. cial SU/Iislics. Supplemenl 011 Output 1 984). F o r t h e i nd u s t rial count ries a n d most of t he non industrial coumries the data are averaged around 1 91l I. 2 0 t her t a xes ;rrc wxes w h i ch cannot be a l l o ca t ee! 10 specific subcategories or are not imporwnt cnough revenue sources to be s t u d i ed on t h e i r own (see Tanzi ( 1 ')87)). Supplement 011 ©International Monetary Fund. Not for Redistribution Tax Structure and Economic Development 31 The cross-section set of data has been arranged in ascending order of the level-of-development variable, Y. This ordered sample places each country in a unique position in the spectrum of development stages, from the least developed (Chad, Y=93) to the most devel­ oped (Switzerland, Y= 16,231 ). Comparative studies of tax structures seem to confirm that, al­ though there may be specific reasons why in some countries particu­ lar taxes have large revenue shares, tax structure is generally associ­ ated with the level of economic development. One notable feature of this pattern is that in developing economies indirect taxes on domes­ tic sales and foreign trade are more important than personal income taxes. In contrast, income taxes are important in industrial econo­ mies, where traditionally they are used as a major instrument for income redistribution. The following analysis examines whether the assumed differences in tax structures can be confirmed by quantita­ tive methods, and whether countries can be classified in distinct groups according to similarities in tax structure and stage of eco­ nomic development. In the comparison of different national tax structures we are con­ cerned with relationships between observed variables-the set of tax revenue shares-and latent processes presumed to be generating the observations. This information is not readily available at the level of the data set, which occupies a matrix of I 0 x 105 elements. What is re�:(Uired is that the underlying factors represented by the set of tax shares be reduced to a comprehensive summary set, but without any substantial loss of the information contained in the original data. This will enable us to identify the important characteristics of each coun­ try's tax revenue structure and to investigate the interrelationships between the tax shares for the presence of common elements. A statistical technique for exploring interdependencies among variates is the principal components analysis that transforms a large set of correlated variables into a smaller set of uncorrelated principal con1ponents that account for the total variance among the original data in the system. The principal components are ordered hierarchically in terms of their variance (Theil ( 1 9 7 1 )). The first principal component is the normalized linear combination of the original variables that contributes a maximum to their variance. The second principal com­ ponent contributes a maximum to the residual variance, and so on until the total variance is accounted for. The weight of each variable in a principal component is the regression coefficient of that variable on rhe principal component; the square of each weight is the propor­ tion of the variance of the original variable accounted for by the corresponding principal component. The results of the principal components analysis are presented in Table I , which gives the variable weights in the composition of each ©International Monetary Fund. Not for Redistribution 32 THEO HITIRIS Table l . Tax Revenue Structure: Variable Weights in the Principal (Components) Component Cl C2 C3 C4 C5 C6 .56 -.40 .68 .45 .00 . 14 -.75 -.47 .27 02 .69 . 14 .25 -.05 .12 -.33 -.49 .04 .28 -.46 -.19 .68 .487 .69 .16 -.37 . 14 -.38 -.3 3 .59 .02 .23 - 28 .06 -.21 .06 .22 -.42 .67 -.26 .78 1 0.924 Tax Variable IP IC IS ss sx so TM TX ow 00 R·SQR Eigenvalues' Tax Type Income, IT -.07 .202 2.019 .95 04 -.87 Sales, ST -. Trade, TT Other, OT R-SQR -.33 Eigenvalues' .442 1 .770 . .24 -.53 - .6 1 -.24 .1 5 - .47 - -.3 5 .365 1.630 1 .220 .3 1 -.97 . 12 -.21 .44 -. 1 7 .758 1 .260 -.21 93 -.37 .35 -.30 -.19 .03 .02 .02 .590 1 .026 -.15 -.46 .18 .18 .08 .26 .689 0.994 1 .000 0.969 'Eigenvalue is the sum of squared weights. The eigenvalue divided by the number of variables gives thl' proportion of the total variance accounted for by that principal component, C,. R-SQfl is the cumulative eigenvalue divided b)• the number of variables. principal component, first, for all the ten taxes and, after aggrega­ tion, for the four broad types of taxes: income (/7), domestic sales (57), trade (TT), and other taxes (07). The size of the weights in the principal components seems to indicate that the structure of tax revenues is primarily determined by the interrelationships between three groups of tax variables: income, domestic sales, and trade taxes. I n the analysis of ten taxes, component 1 (C 1 ) is dominated by a large negative contribution of import taxes (TM), and by two large positive contributions of social security (IS), and personal income (IP) taxes. Although the signs of the weights have no fundamental significance in themselves, as all may be reversed without affecting any of the results, they indicate a classification of the variates into two subgroups (Maxwell ( 1 987)). Then, what C I tells us is that in the total tax revenue of the countries in the sample there is a tendency for high import and export tax shares, TM and TX, to be associated with low personal and social security tax shares, IP and IS, and vice versa. This point accords with the stylized facts and confirms the ©International Monetary Fund. Not for Redistribution Tax Structure and Economic Development 33 findings of previous studies,:\ that trade tax shares are imponant in countries characterized by low income tax shares. Component 2 (C2) is dominated by another of the income taxes, the corporation tax, IC, with a positive effect, and by domestic excise and other taxes, SX and SO, with a negative effect. This again implies that in the data set, countries with high corporation tax shares, /C, display low sales tax shares, SX and SO, and vice versa. The wealth and property taxes, OW, and other taxes, 00, contribute large weights funher down the list of principal components. The R-SQR statistic shows the propor­ tion of variance explained cumulatively by successive principal com­ ponents. Its values show that in general the proponion of variance explained by each principal component is relatively low. This may be taken as evidence that tax structures differ substantially between countries. Table I also shows that at the aggregate level of four tax variables, the total variance is explained completely by just three principal components. C 1 is now dominated by the large effects of two types of taxes operating in opposite directions, income taxes, IT, and trade taxes, TT. This again indicates that in countries of high income tax shares in total tax revenue, the trade tax shares are small. Domestic sales taxes, ST, dominate C2, while the other taxes, OT, dominate component 3 (C3). It is now interesting tO examine whether the identified interrela­ tionships in the structure of tax revenue are in any way associated with the level of economic development and other determinants of the effective tax base. Our problem is to examine whether the sam­ ple of l 05 countries, of which each is described by a number of variables, can be segmented into a number of constituent groups in a way that the members of each group differ from one another as little as possible. This will reduce the data set and thus enable us tO recognize common factors in the composition of total tax revenue which might go some way toward explaining the similarities of tax structures among countries, thus leading tO the formulation of test­ able hypotheses. This objective we pursued by application of cluster analysis, which involves a process of classification based on measure­ ment of the similarities between the countries in the sample with respect to the given set of variates 4 The number of partitions and the elements on which they are based are determined subjectively. We applied the cluster analysis for separation of the sample accord­ ing tO similarities in the set of variables describing the tax revenue structure, the level of economic development, and the degree of openness tO international trade. The analysis shows that the strongest 'On principal components and their interpretation see Maxwell ( 1987). 'For a differem approach to a similar problem see Tait and others ( 1979). ©International Monetary Fund. Not for Redistribution 34 THEO HITIRJS influence in the determination of country groupings is exercised by the GOP-per-head variable, Y, tO the extent that the sample is seg­ mented in blocks across the ordered data series. 5 Accordingly, the results of the analysis show that there are basically four major groups of similar countries (A = underdeveloped, B = developing, C = semideveloped, D = developed), and that further segmentation of the sample to five or six classes results in subdivisions of the first and second groups only (A 1 , A2, B 1 , 82), leaving the last two groups intact. That is, the clusters are shown to be relatively robust. How­ ever, nOte that the sample splits into two groups at observation 87, corresponding to underdeveloped countries, 1-87, and developed/ industrial countries, 88- 105. The variables on which the segmentation of the sample has been based display interrelationships consistent with those detected in the principal components. Thus, on the basis of the means of variables, the comparison between groups in Table 2 shows that, moving from the lower-income tO the higher-income groups: the import tax share in tOtal revenue falls consistently, from 33.2 percent in group A I to 3 . 3 percent in group D; the social security tax share increases consis· tently, from 3.0 percent in group A 1 to 27.3 percent in group 0; the share of direct taxes (personal and corporate income taxes and social security) rises consistently, from 26.6 percent in group A tO 67.6 percent in group 0; and none of the other taxes (including the domestic sales taxes) seem to display a discernible pattern associated with the level of economic development. The results reached by principal components analysis are thus confirmed by the cluster analysis with the addition of a development perspective: the tax revenue structure displays high trade tax shares and low income tax shares in the group of low income countries; in groups of higher­ income countries, the trade tax share falls, and the income tax share rises; and the highest-income group displays the lowest trade tax share and one of the highest income tax shares. These points accord with the observation that economic development entails the substitu­ tion of direct taxes for trade taxes, and therefore a rising share of direct taxes in total tax revenue, reflecting structural changes in the economy and, probably, social and political trends toward redistribu­ tion (Musgrave ( 1 969)). Trade Taxes and Economic Development The statistical analysis so far performed has hinted that trade taxes make a diminishing contribution to total revenue of higher-income 'Without the Y variable. the clusters become funr and mixed. ©International Monetary Fund. Not for Redistribution 35 Tax Structure and Economic Development Table 2. Tax Revenue Structure: Cluster Classification of Countries (Means of Variables) Group Observ:uions AI 1-5R A2 59-62 65-71 71-!12 83-89 90-IOS 10.2 1 3.4 3.0 12.2 13.3 3.7 .33.2 6.3 1 .8 :U .B3 30.7 I 5.6 12.2 19.7 5.4 I 1.1 12.0 5.0 2 1 .8 7.0 2.5 2.9 1 ,026 34.3 25.1 13.1 1 1 .6 14.2 8.5 I 1.2 8.6 196 4.0 33 5.8 1 ,891 34.7 1 8.4 7.9 20.-1 18.0 1.3.9 9.0 5.1 1 7.6 1 .0 3.5 3.2 3.133 49.7 53.9 32.6 16.9 18.1 9.4 7.3 4.2 7.0 0.0 4.1 0.1 5.675 62.4 49.6 26.9 6.9 27 3 17.7 8.9 .3.3 3 3 0.0 2.2 :1.5 1 1 .729 29.8 29.7 Bl fl2 c 0 Tmal 1 - )(IS Variable' IP IC IS ss SX so TM TX OW 00 y M X 1;.7 1;.6 10.8 12.5 1 1 ..3 21.5 ...6 -1.3 ? -) 3.2 3.01 I 36.7 26.7 'The vari:tbles are: //'=personal income tax. IC=corpor:ne and mher income taxes. IS=social secu rity taxes. SS=domcstic s:llcs tax. SX=domcstic excises. SO=other domestic t:txes on goods :mc.l services, 7�\l=impon t:txcs, TX=cxpon taxes. Olfl=weahh :tnd property t:txes. ()()=other tax.-s. !'=income per head (in U.S. dollars). M=imJ>ort v:tlue/GOP. X=cxpon value/GOP. The groups correspond roughly to: ( I ) underdeveloped countries= A I :1nd A2. (2) de\'cloping countries= B l and 02, (5) scmidevcloped countrics=C, and (4) developed countries= D. economies. In this section we explore this point further by cluster and regression analysis. The trade tax share in total revenue, TT, is identically equal to the product of three variables: the effective tax rate, V = trade tax/trade value; the share of trade in GOP, T = trade value/GOP; and the inverse of toral tax burden, G = total tax/GOP: TT = (V X 7)/G. (I) The variables on the right-hand side of the identity in ( I ) are, of course, themselves endogenous. We are now interested in whether the countries in our data set can be classified in distinct groups according tO latent interrelationships between the variables in ( 1 ) and the level of economic development, Y. Application of cluster analysis separated the ordered sample at exactly the same points as the previous analysis, proving once more that the formation of clusters is strongly influ­ enced by the level of per capita GOP. The results for six-group and two-group splits are presented in Table 3. The move from the lower· to higher-income groups shows that the share of trade taxes in total tax revenue, TT, and the effective trade tax rate, V, decline consis­ tently, that the total tax burden, G, rises and that the openness to trade, T, displays no discernible pattern. These points, which are ©International Monetary Fund. Not for Redistribution 36 T H EO HITIRJS Table 3. Trade Tax Shares: Cluster Classification of Countries. (Groups and Mea ns of Variables) Groups Six Groups AI A2 62 c D Observations 1 -38 59-62 63- Il l I 2-1!2 f!.;-!!9 90- 1 0 � Variable TT If T c 39.5 2 .9 23. I 29.9 28.8 18 3 29.7 1 8. 5 23 6 1 6.9 26.6 1 8.7 1 8 .6 1 2 .6 53.9 2 2 .4 7.0 6.3 56.0 30 8 3 . :1 4.4 29.7 344 Two G roups nden.leveloped 1 -87 TT 30.5 If 1 8. 1 T 2 10 c 3 1 .8 Y' (In dollars) 1 ,3 1 7 S1andard deviation ( I ,355 .4) r 0.538 0 238 - 0 405 ' T he variab le.< are 7T= t rade raxlrmal t:ox. Develo p�<.l 1:!8- 1 0 5 3.1 37. 1 4.2 29. 1 Tmal r I - I OS 25.8 1 8. 1 0.962 - 0.265 3 13 2 1 .:1 - 0.665 3,01 1 1 1,19 (2 , 3 7 .4) (4 ,054.7) r 0.664 0 209 - 0 .6 1 3 lf= r rade raxlt radc value, T= trilde value/G OP. G = tmal wxiGDP. Y = pe r capita G D P; r denotes the corrcl:uion coefficient between TT and the other vari· able,. "11<.1 the st• n<.lard deviation is uf mc;m Y. standard propositions in the l i te ra t u re on taxation (M usgrave ( I 969)) and trade (Kindle berger ( 1 962)), are also confirmed by the size and sign of the partial correlation coefficients between TT and the righ t· hand side variables in ( 1 ) , which are presented in Table 3. We can now exam ine whethe r t he pattern of changes in the vari­ able TT t h rough d i ffe rent income count ries can be "explained" by regression analysis. On t h e basis of ide n t ity ( l ) and the i n formation gathe red by data analysis, a reasona b l e hypothesis to make would be that the c rade tax share depends o n t h ree variables, the level of economic developme n t , Y, the measure of openness to trade, T, and the share of income taxes in total tax revenue, IT: TT = f ( Y T, IT ) . (1 . 1 ) The principal components a n d cluster analysis would also imply that (1 . 1 ) is expected to b e : Jy < 0, h > 0, f1r< 0 . The correlation coefficients between the variables V the d i rec t ion of t he relationship in and Y, r = 0 .6, and G and I T, r = 0 . 7 , also suggest that the spec ificat ion in ( 1 . 1 ) is a relati vely close approximation of identity ( I ) . However, in attem pting ro estimate ( 1 . 1 ) by econometric m e t h­ ods, two problems must be considered: fi rst , the distr i b u t ion of the ©International Monetary Fund. Not for Redistribution Tax Structure and Economic Development 37 variables in ( 1 . 1 ), and second, the continuity of the relationship between countries at different stages of economic development. The variables in equation ( 1 . 1 ) are expressed in values which by definition are non-negative. Moreover, the trade taxes and income tax shares, TT and IT, are percentages, that is, positive numbers taking values within the range 0-100. Hence, the values of all the variables in ( l . l ) are constrained and cannm strictly follow a normal distribution. Therefore, the conditions of the classical linear regres­ sion model, that is, homogeneity of variance and normality of errors, are a priori not satisfied (Atkinson ( 1 985)). This means that the esti­ mates of regression coefficients would still be unbiased and best, but their statistical reliability cannot be assessed by the classical tests of significance based on normal distributions. A way out of this prob­ lem is to transform the original data, for example, by taking the logs of the positive variables Y and T and normalizing the proportional variables by the logit transformation, logit(x) log[x/(1 -x)]. The lOglinear form of function ( 1 . 1 ) is: = It = a + b,y + bzy + b3y, b, < 0 , b2>0, b3 <0, (2) where the small case letters in equation (2) denote logs or logit transformations of the variables in ( 1 . 1 ). As we have seen, the sample can be split into several groups characterized by different trade tax ratios. If the differences between the groups are significant, the sample is nonhomogeneous. Conse­ quently, it is possible that structural breaks may exist in the regres­ sion relationship (2) that would render the estimation from the entire set of observations invalid. Therefore, we begin by investigating this possibility, using exploratory methods of data analysis. If signs of departure from constancy become evident, they will be borne in mind during subsequent analysis and assessed by formal statistical tests. A technique for exploring the continuity of a data set is the estima­ tion of cusum and cusum-squares of recursive residuals (Brown and others ( 1 975)). This method generates statistics, which, if they are significant, indicate some departure from constancy over the sample, and provides a likelihood ratio statistic as the means for locating the probable point of break in the stability of the regression (Quandt ( 1 960)). Application of this method tO the ordered sample of I 05 observations6 provided evidence of departure from stability in re''The cusum and cusum-squares methods are applicable to ordered data in time series sequence, that is 10 time series data. We have made use of these techniques here on data ordered in a sequence of income values under the :1ssumption that economic development follows a sequence of stages. Hence, a departure from constancy signi· fies a break i n the ordered series of stages, that is, a significant gap between levels of economic development. ©International Monetary Fund. Not for Redistribution 38 THEO HITIRIS gression (2) at about observation 86 or 87, suggesting the possibility that the sample is not formed by a homogeneous set of observations. Note that cluster analysis has also indicated that, if the data set consists of two groups of countries, the split occurs at observation 87. As will be seen later, the break point identified by these descrip­ tive methods appears to be significant on the basis of standard tests. As a starting point we estimated equation (3) from the entire data set by application of ordinary least squares (OLS): tt = 1 . 561 - 0.754y + 0.5831 - 0.655it, 0. 61 5 , F(3 , 1 0 1 ) = 56.35, s = 1 . 1 64 , RSS = 136.83, n = 105 (Observations 1 - 1 05), x-2(2) = 56.646, F, (6,94) = 4 . 5 1 9 , F2(9,9 1 ) = 3.32, F�(2 ,99) = 4.61, R2 (3) = where in addition to the usual goodness of fit statistics, RSS is the residual sum of squares, and r(.), F1(.), F1(.), F,(.) are diagnostic statistics of model adequacy for testing, respectively, the normality, heteroscedasticity, functional form, and variable specification of the estimated equation. The estimated values of all four diagnostic tests are beyond the acceptance levels, making it clear that there are problems with equation (3) so that it does not seem possible to obtain a satisfactory estimate of the model suggested from the entire set of data. Testing further for the possibility of functional misspecifi­ cation, we failed to find an acceptable form of equation (3) (linear, semi-log, or the like) that could accommodate the entire data set and be judged statistically adequate. Fitting the maintained model to the subset of the first 87 observa­ tions appears to provide satisfactOry results on the basis of the diag­ nostic tests: It = -0.527 - 0.400y + 0.5731 - 0.549il, (0.818) (4.406) (5.023) (5.527) R2 = 0.582, F(3.83) = 38.56, s = 0.680, RSS = 38.78, n = 87, (Observations 1 -87) 0.842, 0,324, F2(9,73) .x2(2) = 4 , 6 5 1 , F1 (6,76) F,(2, 8 1 ) = 0.850. = (4) = The figures in parentheses below the coefficients are estimated /-ratios. The values of :x...! , F,, Fl. and F,, which show that the esti­ mated equation passes all the diagnostic tests, confirm that the model is statistically adequate, and this tends to corroborate the indication of a structural break just after observation 87. The addition of more explanatory variables results in estimations which do not pass the diagnostic tests of model adequacy. For exam­ ple, adding the domestic sales taxes, ST, among the explanatory variables results in rejection of homoscedasticity, functional form, and variable specification, suggesting that in the data set examined ©International Monetary Fund. Not for Redistribution Tax Structure and Economic Development 39 sales taxes are neither close substitutes nor complements of trade taxes. The regression relationship from the last 1 9 observations provided: (5) It = -26,164 + 2.781y - 1 . 193/ - 1 .027it, (I ,443) ( 1 .433) ( 1 .276) ( 1 .745) R2 = 0.338, F(3 , 1 4) = 2,39, s = 1 .784, RSS 44,54, n = 1 9 , (Observations 88-1 05), = with the F-statistic and most of the coefficients statistically nonsignif­ icant, those of income and openness to trade, Y and T, also wrongly signed. These results do not mean that in the data set the group of developed countries has a uniform tax structure (e.g., there are no social security taxes in Australia) nor that trade taxes are negligible in every developed economy (e.g., they account for 2 1 .92 percent of total tax revenue in Iceland). For further evidence concerning the difference between the two groups of countries and the statistical significance of the idemified structural break, we estimated the likelihood ratio (LR) statistic, which tests the equality on both regression coefficients and variances between the two subsamples.7 Its estimated value is LR = 35.913, which, against critical value x1(5) = 1 1 .070 at the 0.05 level, shows that the break at observation 87 is statistically significant and that the structure of the estimated model from the first 87 observations is different from that derived from the last 1 8 observations. Therefore, with regard to the fiscal importance of trade taxes, the data set of 1 05 observations comprises two different subsamples, those of un­ derdeveloped and developed or industrial countries.B These subsam­ ples are heterogeneous, and therefore their data subsets cannot have been originated from the same data generating process. The economic interpretation of these results is that trade taxes are assigned diverse functions in the two different groups of countries. The estimates of equation (4) show that the use of trade taxes for raising tax revenue is associated positively with the relative size of the foreign trade sectOr and negatively with the level of economic development (Corden ( 1 974)). During the process of economic de­ velopment, in their role of raising tax revenue, trade taxes are gradu- "The form of the likelihood ratio test estimated here is; LR = [n ln(RSS/n)-n 1 l n(RSS,!n,)- n,ln(RSSJn!)l = .�5.913. "Not surprisingly, the group of underdeveloped countri<::s detected he;;re;; consists of all the countries traditionally classified under this label. for example. in ·ranzi ( 1 987). with the addition of Ireland and Sp<tin. The cutoff point is Italy's lcvd of incomc, l' = 6,637, which is .3.85 standard deviations above the me;m income of the unclerdevel­ opecl group of countries. ©International Monetary Fund. Not for Redistribution 40 THEO HITIRIS ally replaced by income taxesY Equation (5) shows that, in contrast to the developing countries, industrial countries do not use trade taxes for raising tax revenue: they display neither substitution be­ tween trade taxes and income taxes nor a statistically significanr relationship between trade tax revenue and openness to international trade. These comparisons suggest that, with reference to the fiscal aspects of trade taxes, the developing countries form a group of their own, clearly distinct from that of the industrial countries. The statis­ tical evidence supports the hypothesis that there is a clearly identi­ fied threshold level of economic development that separates two dissimilar groups of countries characterized by different tax struc­ tures and different interrelationships between taxes. Conclusions In this paper, we have examined a cross-section data set in an attempt to identify by quantitative methods the main differences between the tax structures of different countries and to assess the fiscal importance of trade taxes at different stages of economic development. In general, none of the results obtained by our analysis is contrary to predictions. The novelty of the present exercise is that it provides objectively concrete evidence to a number of issues previously con­ sidered to be subjective general statements or speculations. Bearing in mind that cross-section analysis describes general trends, identical terms in tax systems are given different contents in different countries, and the individual characteristics and peculiari­ ties of tax systems cannot be completely captured by purely statisti­ cal summaries (Theil ( 1 9 7 1 )), the present study gives rise to four main conclusions. First, certain taxes or rates of taxes go with the level of economic development and therefore constitute the principal characteristics that make the tax structures of industrial countries different from those of developing countries. We have found that the effective trade tax rate and the trade tax share in total tax revenue fall, while the "1The inverse relationship between trade and income tax shares in tOtal tax revenue i mpl ies that it should be possible w e;:stimate an income tax relationship for the group of less developed coumries. Our data provided: it = 2.904 + 0.271y - 0.41011, (5.702) (3.326) (5.057) R' = 0.478, F(2.84) = 38.47, s = 0.65, n = 87, (Observations 1-87) .\.! = 0.594, F1(4,79) = 1 . 1 9 , F,(5,78) = 1 .76, F,(2,83) = 1 .040, which passes all the econometric tests. As in the case of trade taxes, the generat ion of income tax data of the group of industrial countries cannot be explained by a relatiOn· ship similar to t he one :�bove. ©International Monetary Fund. Not for Redistribution Tax Structure and Economic Development 41 total tax burden and the share of income taxes in total tax revenue rise with economic developmenl. Second, countries can be classified in distinct groups according to tax structure and level of economic development. We have found that four main groups of economic and tax development can be distinguished. The country membership of these groups is deter­ mined by the level of per capita income and the shares of trade taxes and income taxes in tOtal tax revenue; the Iauer two characteristics are related inversely. A second classification of countries according 10 level of economic development and size and composition of trade tax shares in tOtal tax revenue provided an identical classification and country membership. Third, the degree of developing countries' fiscal dependence on trade taxes is shown 10 be associated positively with the relative size of the trade sector in the national economy and negatively with the level of per capita income and the share of income taxes in total tax revenue. Fourth, the statistical evidence seems to support the hypothesis that there is a threshold level of economic development-level of per capita income-beyond which countries cease 10 be fiscally de­ pendent on trade tax revenue. References Atkinson, A. C., Plots. Transformations and Regression (Oxford: Clarendon Press. 1985). Brown. R. L., J. Durbin, and T. M. Evans, "Techniques for Testing the Constancy of Regression Relationships Over Time,"journal of the Royal Statistical Society. Vol. 37 ( 1975), pp. 149-92. Corden, W. Max, Trade Policy and Economic \'(/eljare (Oxford: Clarendon Press. 197"1). Dixit, A., "T:tx Policy in Open Economies," in Handbook of Public Economics, ed. by Alan J. Auerbach and Manin Feldstein (Amsterdam: Nonh-Holland, 1985). Kindleberger, Charles P., Foreign Trade and the National Economy (New Haven: Y:1le University Press, 1962). Maxwell, A. E. . i\1/ultivariate Ana�)>sis in Behauioral Research (London: Chapman and Hall, 1987). Musgrave, Richard A., Fiscal Systems, Studies in Comparative Economics, No. lO (New Haven: Yale University Press, 1969). Quandt, R. E., "Tests of the Hypothesis That a Linear Regression System Obeys Two Separate Regimes," .fourna/ oftbe American Statistical Association, Vol. 60 ( 1960), pp. 324-30. Tait, Alan, Wilfrid L. M. Gratz, and Barry J. Eichengreen, "International Comparisons of Taxation for Selected Developing Coumries. 1972-76." Staff Papers, Internationa l Monetary Fund (Washington), Vol. 26 (March 1979), pp. 123-56. Tanzi, VitO, "Import Taxes and Economic Development," /:.'conomia lntemazionale. Vol. 3 1 ( 1978), pp. 252-69. , "Quantitative Characteristics of the Tax Systems of Developing Countries." in The Tbeoty of Taxation for De11eloping Countries. ed. by David Newbery and Nicholas Stern (Oxford: Oxford University Press, 1987), pp. 205-4 1 . Theil, Henri. Principles ofEconometrics {Amsterdam: North-Holland, 1971 ). ____ ©International Monetary Fund. Not for Redistribution