Policy Optimization with a CGE Model Martín Cicowiez (CEDLAS-Universidad Nacional de La Plata), Bernard Decaluwé (Université Laval) and Mustapha Nabli this version: May 9, 2016 VERY PRELIMINARY DRAFT; do not cite 1. Introduction Generally speaking, with CGE models, policy analysis is mostly carried on as “shock analysis”. That is, changing the values of selected policy variables and computing the corresponding model results. However, today’s computer and software developments allow us to perform optimal policy exercises, even for relatively large models. That is, for the specification of an objective function and the computation of the corresponding optimal values for selected policy variables, and where the CGE model operates as the constraint of the optimization exercise.1 In formal terms, the problem can be stated as max u J g x* s.t. F x, u, z, where x = vector of endogenous variables x* = vector of policy variables; subset of x z = vector of exogenous variables = vector of parameters 1 In Mercado et al. (1998) an approach similar to the one proposed here is applied in the context of a small macroeconomic model. -1- u = vector of policy variables, whose value is endogenously determined as the result of the optimization problem Moreover, parameter uncertainty may also be taken into account in a sophisticated way when performing optimal policy exercises, something with a long tradition in the realm of empirical macroeconomics, and to which CGE modeling may perhaps begin to converge. The CGE literature is not abundant in this sort of applications. In Böhringer and Rutherford (2002) a static multi-country CGE model is used to determine optimal environmental tax, whereas Bovenberg and Goulder (1996) performed a similar analysis applied to the case of the United States.2 In turn, Kim (2004) uses a CGE linear model in the context of a stochastic control problem that incorporates the uncertainty about the value of certain parameters of the model. In this paper, however, we propose to solve a more general problem where the function to be optimized is a loss function that can incorporate various objectives. In fact, we will allow the use of different policy instruments, not only tax rates. To illustrate the potential of our approach, we develop an exercise with real data from Argentina, a developing country with a relatively large agri-food sector. 2. The Model In this paper we ask the CGE model what would be the best policy mix, given an objective function for the policy maker. In practical terms, we will start by using the small open economy CGE model known as PEP-1-1 (Decaluwé et al., 2012) as a constraint to an optimization problem. Specifically, we implement a loss function that the policy maker minimizes. Mathematically, u x min L i *i 1 j *j 1 u i j xi j 2 2 2 The two articles can be framed within the theory of optimal taxation (see Myles (1995), Stiglitz (1987), Auerbach and Hines (2002)). -2- s.t. F x, u, z, where L = loss function xi = endogenous variables u j = policy instruments available in the model (e.g., tax rates) * * In addition, xi and ui represent values reached by xi and u j , respectively, in the reference scenario. As an alternative, xi could represent a given policy objective. In the implementation below, we ran the static version of the model. Thus, departures of policy instruments from their target (initial) values are penalized. PEP-1-1-OPT Model: Mathematical Statement As said before, as a first step, we extended the PEP-1-1 model (Decaluwé et al., 2012) in order to incorporate endogenous unemployment – through a wage curve -- and a loss function for the policy maker.3 In the model, the loss function is written as 3 In Appendix A we provide additional details on the extensions introduced to PEP-1-1. -3- RGDPFC LOSS wtrgdpfc 1 * RGDPFC UR wtur * 1 UR 2 2 RSG wtrsg 1 * RSG 2 CAB wtcab 1 * CAB 2 TTDHADJ wttdh 1 * TTDHADJ TTXADJ wttx 1 * TTXADJ 2 2 CGADJ wtgovcon 1 * CGADJ 2 and the wage curve is written as W WO UR PINDEX URO PINDEXO where UR = unemployment rate; URO refers to the initial unemployment rate RSG = real government savings RGDPFC = real GDP at factor cost TTDHADJ = scaling/adjustment factor for direct tax rates TTXADJ = scaling/adjustment factor for indirect tax rates CGADJ = scaling/adjustment factor for government consumption wtopt = weight of the opt argument in the loss function; certainly, not all weights in the loss function will be simultaneously different from zero W = wage PINDEX = consumer price index -4- = wage curve elasticity and the variables with a star represent the policy objectives. 3. Illustrative Application In this section, we show three example applications of the proposed approach, implemented over the PEP-1-1 model. As will be shown, the proposed approach can be applied for three different purposes: optimal policy response to a negative shock, optimal selection of macro closure rule, and policy optimization – in the last case, starting from the base scenario. Social Accounting Matrix and Other Data For illustrative purposes, we will implement our extension of the PEP-1-1 CGE model using Argentina data. Like other CGE models, our CGE model uses a base-year SAM (in this case for 2012), to define base-year values for the bulk of the model parameters, including production technologies, sources of commodity supplies (domestic output or imports), demand patterns (for household and government consumption, investment and exports), transfers between different institutions, and tax rates. The disaggregation of the Argentina SAM coincides with that of the rest of the model database. As shown in Table 3.1, it is disaggregated into 17 sectors (activities and commodities) – 1 in agriculture, 2 in mining, 7 in manufacturing, and 7 in services – with each activity producing one or more commodities. The factors are split into labor, capital, and natural resources (5 types: agricultural land, forestry land, fishing resources, and two natural resources used in extractive industries). The institutions are split into households, enterprises, government, and the rest of world. A set of tax accounts cover the different tax instruments. A stylized (Macro-)SAM for Argentina is provided in Table 3.2. Argentina GDP reached 2,766 million pesos in 2012. In 2012, the government current account surplus was around 0.9 percent -5- of GDP and government current consumption was 15.1 of GDP. In 2013, exports and imports were 15.5 and 13.7 percent of GDP, respectively. Table 3.1: disaggregation of Argentina PEP-1-1-OPT and SAM Category - # Primary (3) Sectors (activities and comm) (17) Manufact (7) Services (7) Item Agriculture, forest and fish Other mining Petroleum and gas Food Textiles and apparel Petroleum products Chemicals, rubber and plast Metals, mach and equip Vehicles Other manufacturing Elect, gas and water Construction Trade Transport and comm Other services Public administration Education and health Source: Authors’ elaboration. -6- Category - # Item Factors (4) Labor Capital Land Natural resources Taxes (4) Taxes on sales Tariffs Taxes on exports Taxes on income Institutions Households (4) Enterprises Government Rest of the world SavingsSavings Investment Investment, private (4) Investment, gov Stock change Table 3.2: Macro-SAM for Argentina 2012 act com f-lab f-cap tax-vat tax-com tax-imp tax-exp cssoc tax-dir hhd ent gov row sav invng invg dstk total 159.2 73.4 47.1 38.7 66.1 15.1 15.5 0.0 0.4 14.8 2.3 14.8 2.3 -0.1 8.1 76.8 35.7 31.6 16.8 17.0 14.8 2.3 6.9 4.5 0.6 2.2 6.5 3.1 40.6 13.7 35.4 1.2 0.0 2.5 0.0 159.2 187.1 47.1 39.1 6.9 6.9 4.5 4.5 0.6 0.6 2.2 2.2 6.5 6.5 8.1 1.3 0.2 6.1 5.0 20.8 15.3 0.2 0.3 9.6 0.0 0.9 0.1 0.1 0.3 0.4 to ta l ds tk g in v ng in v sa v ro w go v en t ir hh d ta xd om ta xim p ta xex p cs so c at ta xc ta xv ap f-c ab f-l co m ac t (percent of GDP) 159.2 -0.1 187.1 47.1 39.1 6.9 4.5 0.6 2.2 6.5 8.1 76.8 35.7 31.6 16.8 17.0 14.8 2.3 -0.1 -0.1 where act = activities, com = commodities, f-lab = labor, tax-vat = value added tax, taxcom = sales tax, tax = tariffs, tax-exp = export tax, cssoc = social security contributions, taxdir = direct tax, hhd = households, ent = enterprises, gov = government, row = rest of the world, sav = savings, invng = non-government investment, invg = government investment, dstk = changes in stocks. Source: Authors’ calculations based on Argentina SAM. On the basis of SAM data, Table 3.3 summarizes the sectoral structure of Argentina’s economy in 2012: sectoral shares in value-added, production, employment, exports and imports, as well as the split of domestic sectoral supplies between exports and domestic sales, and domestic sectoral demands between imports and domestic output. For instance, while (primary) agriculture represents a significant share of exports (around 16 percent), its shares of value added (VA) and production are much smaller (in the range of 6.2 percent). The share of its output that is exported is around 16.8 percent while only some 1.4 percent of domestic demands are met via imports. -7- Table 3.3: sectoral structure of Argentina’s economy in 2012 (percent) Sector Agriculture, forest and fish Other mining Petroleum and gas Food Textiles and apparel Petroleum products Chemicals, rubber and plast Metals, mach and equip Vehicles Other manufacturing Elect, gas and water Construction Trade Transport and comm Other services Public administration Education and health Total VAshr PRDshr 6.2 6.2 1.4 1.1 1.9 2.0 4.7 9.8 1.4 1.9 0.6 2.3 3.3 5.3 3.3 5.0 0.7 1.8 1.7 2.1 3.3 2.9 6.2 6.8 11.6 9.0 5.9 6.8 27.8 22.8 7.6 5.8 12.3 8.3 100.0 100.0 EXPIMPEXPshr OUTshr IMPshr DEMshr 16.1 16.8 0.8 1.4 5.0 40.0 1.4 15.2 2.8 6.2 4.8 18.0 28.1 15.4 1.7 1.8 2.9 9.9 3.1 14.2 4.3 7.9 5.0 17.2 8.8 12.2 17.6 25.6 5.8 9.3 29.0 36.6 10.3 40.8 14.7 55.1 1.0 3.7 3.0 12.1 0.2 0.5 0.5 1.4 0.1 0.1 0.2 0.2 0.0 0.0 0.0 0.0 3.3 4.7 4.5 5.6 11.2 4.8 13.7 5.2 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 100.0 6.9 100.0 9.5 where VAshr = value-added share (%); PRDshr = production share (%); EMPshr = share in total employment (%); EXPshr = sector share in total exports (%); EXP-OUTshr = exports as share in sector output (%); IMPshr = sector share in total imports (%); IMP-DEMshr = imports as share of domestic demand (%). Source: Authors’ calculations based on Argentina SAM and employment data. Table 3.4 shows the factor shares in total sectoral value added. For example, the table shows that agriculture is relatively intensive in the use of capital and land. On the other hand government services and education and health are relatively intensive in the use -8- (salaried) labor.4 Of course, this information will be useful to analyze the results from the CGE simulations. Table 3.4: sectoral factor intensity in 2012 (percent) Sector Agriculture, forest and fish Other mining Petroleum and gas Food Textiles and apparel Petroleum products Chemicals, rubber and plast Metals, mach and equip Vehicles Other manufacturing Elect, gas and water Construction Trade Transport and comm Other services Public administration Education and health Total Labor 31.6 21.1 21.1 45.0 45.0 45.0 45.0 45.0 45.0 45.0 26.4 42.1 43.2 56.4 41.0 100.0 69.6 48.8 Capital 32.9 78.8 78.8 48.4 48.4 48.4 48.4 48.4 48.4 48.4 73.6 45.7 42.5 38.4 53.2 0.0 26.5 42.7 Land 35.5 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 2.4 Total 100.0 99.9 99.9 93.4 93.4 93.4 93.4 93.4 93.4 93.4 100.0 87.7 85.7 94.8 94.2 100.0 96.0 93.9 Source: Authors’ calculations based on Argentina SAM. In addition to the SAM, our CGE model also requires (a) base-year estimates for sectoral employment levels and unemployment estimates for the different labor types, and (b) a set of elasticities (for production, consumption and trade). In order to estimate sectoral employment we combined population data with estimates for sectoral employment shares in broad sectoral categories from the Encuesta Permanente de Hogares (EPH), the 4 As shown in the table, due to the lack of reliable data, we to assumed that factor shares in all manufacturing sectors are the same. -9- main household survey in Argentina. In turn, elasticities were given a value based on the available evidence for comparable countries. For elasticities, the following values were used: (a) the elasticity of substitution among factors is in the 0.2–1.15 range, relatively low for primary sectors and relatively high for manufactures and services (see Narayanan et al. 2015); (b) the expenditure elasticities for household consumption were obtained from Seale et al. (2003); and (c) trade elasticities are 2 for both Armington and CET elasticities. Given the uncertainty with respect to our elasticity values, in Appendix B we conduct a systematic sensitivity analysis of our simulation results with respect to their values. Optimal Policy Response to a Negative Shock In this case, we compute the optimal policy response to a negative shock, imposing a loss function with different weights. Firstly, we simulate a 25% decrease in all (immobile) sectoral capital stocks, without computing the optimal policy response. The macro closure for this scenario is the following: the government balance clears through endogenous savings, exogenous/fixed saving rates for households with endogenous real investment, and fixed current account balance (in foreign currency) with flexible real exchange rate. The results of this scenario are shown under column non-opt (i.e., no policy optimization) in Table 3.2. As expected, unemployment increases -- from 16.5 to 25.6 percent-- at the same time that (real) government savings decrease – from 0.9 to -3.8 percent of GDP. Secondly, we solve a variant of the following optimization problem with different sets of weights in the loss function: 2 UR RSG min LOSS wtUR 1 wt RSG 1 0 0 UR RSG 2 s.t. all equations in the CGE model, including CGi CGi0 CGADJ and using government consumption as the policy instrument; i.e., in this set of simulations the variable CGADJ is endogenous. So, we have one policy variable and two objectives. where -10- wtUR = weight of (percent deviation in) the unemployment in the loss function wt RSG = weight of (percent deviation in) the real government savings in the loss function UR = unemployment rate RSG = real government savings UR 0 = base year unemployment rate RSG 0 = base year real government savings CGi = base year government consumption CGi0 = base year government consumption CGADJ = adjustment factor government consumption Of course, restriction on the policy instrument could have been introduced. For example, CGADJ0*0.90 <= CGADJ <= CGADJ0*1.1 In fact, one can solve the same optimization problem with the government using more than one (constrained) policy instrument. In practice, in order to solve the above problem it is useful to normalize the different policy objectives so that their values lie between zero and one. To that end, we start by computing the so-called pay-off matrix by solving two optimization problems. Firstly, we solve min LOSS wtUR UR 2 where UR UR UR0 1 subject to all the equations in the CGE model and considering government as the decision/policy variable. -11- From solving this problem we obtain (a) the minimum attainable value for the percentage UR deviation of the unemployment rate ( min ), and, given the conflict that exist between both policy objectives, (b) the maximum value for the percentage deviation of the real RSG government savings ( max ). Secondly, we solve min LOSS wtUR RSG 1 2 where RSG RSG RSG 0 1 subject to all the equations in the CGE model and considering government as the decision/policy variable. From solving this problem we obtain (a) the minimum attainable value for the percentage RSG deviation of the real government savings ( min ), and (b) the maximum value for the UR percentage deviation of the unemployment rate ( max ). In Table 3.1 we show the pay-off matrix for this particular exercise. In the second row, the results from solving the first problem show that, after the negative shock, it is possible to obtain an unemployment rate of 22 percent with negative government savings of -376.9. Similarly, the results from solving the second problem show that it is possible to obtain real government savings of $25.7 (i.e., no change relative to the base) together with an unemployment rate of 30.2 percent. In other words, the values in the main diagonal of Table 3.1 show the best attainable results when only one policy objective is considered. Table 3.1: pay-off matrix; unemployment versus government savings; 25% decrease in all the sectoral capital stocks Scenario base min δ_UR min δ_RSG UR 16.5 22.0 30.2 -12- RSG 25.7 -376.9 25.7 From Table 3.1, we compute the ideal and anti-ideal values for the policy objectives in the loss function; i.e., RSG RSG UR UR min 0% , max 1,566.8% , min 33.2% , max 82.8% Next, we re-define our loss function as 2 UR RSG UR min RSG min LOSS wtUR d wt RSG RSG d RSG max min max min 2 Thus, this normalization eliminates any units of measurement, and the weights wt in the above formula are easier to interpret. In this case, five alternative weighting schemes in the loss function are considered, as shown in columns (3)-(7) of Table 3.2. As can be seen (see variable CGADJ), a trade-off exists between both policy objectives: compensating for the negative impacts of the shock implies (a) decreasing government consumption when wtUR 0 (see column [3]), and (b) increasing government consumption when wtRSG 0 (see column [7]). In other words, comparison of columns (3) and (7) shows the degree of conflict between the two policy objectives considered in the loss function. Of course, results in columns (3) and (7) of Table 3.2 are consistent with the pay-off matrix in Table 3.1. Not surprisingly, reductions in unemployment relative to the non-opt (i.e., no policy optimization) scenario cannot be attained without further decreasing in (real) government savings (compare columns [2] and [7]). In fact, note that under wtUR 1 and wtRSG 0 real gross fixed capital formation is zero (again, see column [7]), meaning that the government has exhausted the funds available to finance its deficit; in other words, the crowding out effect is 100 percent. -13- Table 3.2: simulation results; optimal policy response to a negative shock; 25% decrease in all the sectoral capital stocks Item UR RSG CGADJ RGFCF CAB REXR LOSS Units (%) ($) (index) ($) (FCU) (index) base (1) 16.5 25.7 1.000 473.5 -11.3 1.000 N.A. nonopt (2) 27.7 -115.5 1.000 273.0 -11.3 0.985 N.A. weights in loss fn UR=0 UR=0.25 RSG=1 RSG=0.75 (3) (4) 30.2 28.5 25.7 -69.2 0.694 0.901 392.8 313.3 -11.3 -11.3 1.003 0.990 0.000 0.201 UR=0.5 UR=0.75 RSG=0.5 RSG=0.25 (5) (6) 26.4 24.0 -183.1 -298.4 1.142 1.380 211.5 95.8 -11.3 -11.3 0.977 0.966 0.278 0.205 UR=1 RSG=0 (7) 22.0 -376.9 1.542 0.0 -11.3 0.963 0.000 where UR = unemployment rate, RSG = real gov savings, CGADJ = scaling factor gov consumption, RGFCF = real gross fixed capital formation, CAB = current account balance, REXR = real exchange rate, and FCU = foreign currency units. Source: Authors’ calculations. Optimal Selection of Macro Closure Rule Usually, a CGE application entails the selection of a macroeconomic closure rule. In contrast, our approach allows the “optimal selection” of the macro closure rule. As an example, we first simulate a 25% increase in government consumption assuming, as before, that government budget clears through changes in government savings – for savings-investment and rest of the world we also keep the same assumptions as before (see column non-opt in Table 3.3). Then, we simulate the same increase in government consumption but assuming that the government optimally selects the mix between government savings and foreign savings used to finance the increase in government consumption (see columns (3)-(7) in Table 3.3). Analytically, the optimal selection of the (government) closure rule implies solving the mathematical program 2 RSG CAB min LOSS wt RSG 1 wtUR 1 0 0 RSG CAB -14- 2 s.t. all equations in the CGE model, including CGi CGi0 CGADJ and optimally selecting RSG and CAB in order to finance the increase in CGi brought about by the simulated 25% increase in CG0. where wt RSG = weight of (percent deviation in) the real government savings in the loss function wtCAB = weight of (percent deviation in) the current account balance in the loss function RSG = real government savings CAB = current account balance expressed in foreign currency Table 3.3: simulation results; optimal selection of macro closure rule; 25% increase in government consumption Item UR RSG CGADJ RGFCF CAB REXR LOSS Units (%) ($) (index) ($) (FCU) (index) base (1) 16.5 25.7 1.000 473.5 -11.3 1.000 N.A. nonopt (2) 14.9 -98.7 1.250 365.7 -11.3 0.986 N.A. weights in loss fn CAB=0 CAB=0.25 RSG=1 RSG=0.75 (3) (4) 2.7 4.5 -9.7 -31.0 1.250 1.250 1,199.1 1,073.6 -1,914.5 -1,350.9 0.343 0.448 0.000 0.167 CAB=0.5 CAB=0.75 RSG=0.5 RSG=0.25 (5) (6) 6.3 9.4 -48.5 -73.3 1.250 1.250 964.5 778.8 -1,007.9 -591.2 0.540 0.693 0.232 0.197 CAB=1 RSG=0 (7) 14.9 -98.7 1.250 365.7 -11.3 0.986 0.000 where UR = unemployment rate, RSG = real gov savings, CGADJ = scaling factor gov consumption, RGFCF = real gross fixed capital formation, CAB = current account balance, REXR = real exchange rate, and FCU = foreign currency units. Source: Authors’ calculations. -15- By construction, the case where wtRSG 0 and wtCAB 1 is equivalent to the non-optimal selection of closure rule; i.e., columns (2) and (3) are equivalent. In addition, we see that the higher (lower) the weight on RSG (CAB), the larger the decrease in CAB (RSG). Not surprisingly, the decrease in CAB brings about a real exchange rate appreciation. Thus, our proposed approach can be used to simulate increases in government spending that are financed with more than one source of resources, whose mix is optimally selected given an objective function for the policy maker. Policy Optimization In this set of simulations, we compute the optimal change in government consumption to reduce unemployment by (ideally) 95% (i.e., from 16.5 to 0.8 percent), given a loss function that also penalizes changes in government savings. Thus, in this case, all non-base simulations are run under the policy optimization assumption. Analytically, we solve the problem 2 UR RSG min LOSS wtUR * 1 wt RSG 1 * UR RSG 2 s.t. UR* 0.05UR0 0.008 RSG * RSG 0 all equations in the CGE model, including CGi CGi0 CGADJ and using government consumption as the policy instrument; i.e., in this set of simulations the variable CGADJ is endogenous. So, we have one policy variable and two policy objectives: decrease unemployment by increasing government consumption financed with domestic resources, but taking into account the negative impact on government savings. The results are shown in column (2)-(6) of Table 3.4. -16- Table 3.4: simulation results; optimal change in government consumption to reduce unemployment Item UR RSG CGADJ RGFCF CAB REXR LOSS Units (%) ($) (index) ($) (FCU) (index) base (1) 16.5 25.7 1.000 473.5 -11.3 1.000 N.A. weights in loss fn UR=0 UR=0.25 RSG=1 RSG=0.75 (2) (3) 16.5 15.1 25.7 -86.8 1.000 1.222 473.5 376.3 -11.3 -11.3 1.000 0.987 0.000 0.196 UR=0.5 UR=0.75 RSG=0.5 RSG=0.25 (4) (5) 13.5 11.8 -211.7 -339.7 1.461 1.697 260.6 127.8 -11.3 -11.3 0.974 0.962 0.267 0.200 UR=1 RSG=0 (6) 10.2 -440.0 1.878 0.0 -11.3 0.957 0.000 where UR = unemployment rate, RSG = real gov savings, CGADJ = scaling factor gov consumption, RGFCF = real gross fixed capital formation, CAB = current account balance, REXR = real exchange rate, and FCU = foreign currency units. Source: Authors’ calculations. By construction, there are no changes in the first non-base simulation (see column [2]), as wtUR 0 and wt RSG 1 (i.e., only deviations of RSG from its base year are penalized). The other columns show the trade-off between increased government consumption/decrease in unemployment and decrease in government savings. Source: Authors’ calculations. NOTES: Certainly, we could have selected more than one policy instrument in each simulation. For example, taxes could also be optimally selected. In the optimization exercise, we can restrict the tax rates to vary less than 5% with respect to their benchmark values (taxrat0); -17- that is, the following constraints could be imposed to the model: 0.95 taxrat0 <= taxrat <= 1.05 taxrat0. As shown in the last set of experiments, there is no need to assume that policy targets in the loss function correspond to base year values. 4. Concluding Remarks In this paper, we have embedded a computable general equilibrium model within a programming problem for policy simulation. In other words, policy design is seen as a decision problem with multiple conflicting objectives. Certainly, we could have selected more than one policy instrument in each simulation for example, taxes could also be optimally selected can restrict the tax rates to vary by less than 5% with respect to their benchmark values Next, we plan to (a) apply the approach to a relevant policy issue in Argentina and/or elsewhere, and (b) implement dynamic version of the approach, over a recursive dynamic CGE model and assuming that the government is a forward-looking agent. -18- Appendix A: Extensions to PEP-1-1 In this appendix we present the modifications introduced to the single-country static PEP model PEP-1-1 v2.1. Exports In the PEP 1-1 Standard Model, the world demand for exports of product i is (62) e.PWX i EXDi EXDOi FOB PEi iXD In case iXD , equation (64) simplifies to (62’) e.PWX i PEiFOB which represents the “pure” form of the small-country hypothesis; producers can always sell as much as they want on the world market at the (exogenous) current price, PWX i . To simulate a change in the world export demand of a given commodity exported by a given industry keeping the small country assumption (see scenario edem-txt), we introduce the following changes to the model: (1) again, replace equation (62) by (62’), and (2) replace equation (61) (i.e., the relative supply of exports and local commodity) by equation (61’) for the selected commodity and industry pair(s). In addition, we drop the first order condition of the CET function that determines domestic and export sales. (61’) EX j ,i EXO j ,i Current Account BoP Equation (RW1) defines the current account balance in foreign currency. Equations (RW2) and (RW3) define the index for domestic producer prices and the real exchange rate, respectively. As we be shown, variables CAB_FCU and REXR are used to select the macroeconomic closure rule for the model. (RW1) CAB FCU CAB e -19- (RW2) DPI dwtsi PLi i (RW3) REXR e DPI where CAB_FCUO = current account balance in foreign currency units DPI = index for domestic producer prices (PL-based) REXR = real exchange rate dwts(i) = domestic sales price weights Government In the PEP Standard Model, government consumption of commodity i is determined by the following equation (see equation (55) in Decaluwé et al. (2010)). (55) PCi CGi iGVT G with G (i.e., current government expenditures on goods and services) fixed and equal to its initial value (i.e., G GO ). As an alternative, we modified the government behavior assuming that the real government spending can be exogenous (i.e., all the CGi variables) while G is endogenous. Specifically, we dropped equation (55) from the model and added equations (55’) and (55’’), (56’) CGi cgbari CGADJ (56’’) G PCi CGi i where CGADJ = adjustment factor for CG cgbar(i) = base-year CG(i) Equation (G1) defines real government savings, as the ratio between nominal government savings and the GDP deflator. -20- (G1) SG REAL SG PIBGDP where SG_REAL = real government savings Tax Rates (T1) TTDH1h ttdh1hTTDHADJ (T1) TTICi ttic hTTICADJ where TTDHADJ = adjustment factor for TTDH TTICADJ = adjustment factor for TTIC ttdh1bar(h) = base-year TTDH1 tticbar(i) = base-year TTI Household Savings Equation (S1) defines the marginal propensity to save of households. Its structure is the same as that of equations (T1) and (T2) for tax rates and (56’) for government consumption. (43). In fact, whether MPSADJ is flexible depends on the closure rule for the savings-investment balance. (S1) sh1h sh1h MPSADJ where MPSADJ = savings rate scaling factor sh1h = base-year sh1h -21- Calibration using Employment by Sector In PEP-1-1 it is assumed that all sectors pay the same wage. In the extended PEP-1-1, the analyst can complement the SAM with data on number of workers by sectors. To do so, the remuneration to labor type l paid by the activity j is computed as Wl wdist l , j 1 ttiwl , j where wdist l , j is a “distortion” factor applied to for labor type l in industry j that allows modeling cases in which the factor remuneration differs across activities. In other words, each activity pays an activity-specific wage that is the product of the economy-wide wage and an activity-specific wage (distortion) term. To calibrate wdist l , j , the model dataset must provide physical labor quantities. In implementing this extension, the following equations of the original model were modified. (11) YHL h Wh,lL Wl wdist l , j LDl , j l (37) j TIWl , j ttiwl , jWl wdist l , j LDl , j YROW e PWM i IM i (44) k i RK row, k R k, j KDk , j j WrowL ,l Wl wdist l , j LDk , j l j TRrow, agd agd (70) WTI l , j Wl wdist l , j 1 ttiwl , j (92) GDP _ IB Wl wdist l , j LDl , j RK k , j KDl , j TPRODN TPRCTS l, j k, j Wage Curve The PEP Standard Model assumes full employment of the labor force. As explained above, we introduced endogenous unemployment by means of a wage curve. Specifically, we add to the model equation (WC) and the endogenous variable UERAT (unemployment rate). -22- The value of the phillips parameter (i.e., the wage curve elasticity) was set at -0.1 based on international evidence documented in Blanchflower and Oswald (2005). Of course, the equilibrium condition for labor market was adjusted accordingly (see equation 85). (WC) UERATl Wl WOl PIXCON PIXCONO UERATOl (85) LS l 1 UERATl LDl , j phillipsl j where UERAT(l) = unemployment rate for type l labor phillips(l) = elasticity of real wage with respect to unemployment rate Policy Optimization In its general form, the loss function for the policy maker is written as GDPBP, REAL LOSS wtGDPBP , REAL 1 * GDPBP, REAL UERATl wtUERATl 1 * UERAT l l SG wt SGREAL REAL 1 * SGREAL 2 2 2 CABFCU wtCABFCU 1 * CABFCU 2 TTDHADJ wtTTDHADJ 1 * TTDHADJ TTICADJ wtTTICADJ 1 * TTICADJ CGADJ wtTTICADJ 1 * CGADJ 2 2 2 where -23- wt(iopt) = weights in the policy optimization objective function gdp_bp_realstar = GDP BP REAL in optimization objective function ueratstar(l) = UERAT in optimization objective function sg_realstar = SG_REAL in optimization objective function cab_fcustar = CAB_FCU in optimization objective function ttdhadjstar = TTDHADJ in optimization objective function tticadjstar = TTICADJ in optimization objective function cgadjstar = CGADJ in optimization objective function -24- Appendix B: Sensitivity Analysis As usual, the results from the PEP-1-1-OPT model are a function of (i) the model structure (e.g., functional forms used to model production and consumption decisions, macroeconomic closure rule, among other elements); (ii) the base year data used for model calibration (i.e., the SAM), and; (iii) the values assigned to the model elasticities or, more generally, to the model’s free parameters. Certainly, the elasticities used in this study implicitly carry an estimation error, as in any similar model. Consequently, we have performed a systematic sensitivity analysis of the results with respect to the value assigned to the model elasticities. Hence, if the conclusions of the analysis are robust to changes in the set of elasticities used for model calibration, we will have greater confidence in the results presented above. In order to perform the systematic sensitivity analysis, it is assumed that each of the model elasticities is uniformly distributed around the central value used to obtain the results. The range of variation allowed for each elasticity is +/- 80%; that is, a wide range of variation for each model elasticity is considered. Then, a variant of the method originally proposed by Harrison and Vinod (1992) is implemented, which allows for performing a systematic sensitivity analysis. In short, the aim is to solve the model iteratively with different sets of elasticities. Thus, a distribution of results is obtained to build confidence intervals for each of the model results. The steps for implementing the systematic sensitivity analysis are as follows. Step 1. In the first step, the distribution (i.e., lower and upper bound) for each of the model parameter that will be modified as part of the systematic sensitivity analysis is computed: elasticities of substitution between primary factor of production, trade-related elasticities, expenditure elasticities, and unemployment elasticities for the wage curves. Step 2. In the second step, the model is solved repeatedly, each time employing a different set of elasticities; it is, therefore, a Monte Carlo type of simulation. First, the value for each model elasticity is randomly selected. Second, the model is calibrated using the selected elasticities. Third, the same counterfactual scenarios as previously described -25- are conducted. Then, the preceding steps are repeated several times, 500 in this case, with sampling with replacement for the value assigned to the elasticities. Table B.1 shows the percentage change in private consumption estimated (i) under the central elasticities, and; (ii) as the average of the 500 observations generated by the sensitivity analysis. For the second case, the upper and lower bounds under the normality assumption were also computed; notice that all runs from the Monte Carlo experiment receive the same weight. As can be seen, the results reported above are significant, while estimates presented in Tables 1-3 are within the confidence intervals reported in Table B.1. For example, there is virtual certainty that the pwefood scenario has a negative effect on private consumption in Argentina. Table B.2: sensitivity analysis; real private consumption percent deviation from base 95% confidence interval under normality assumption Source: Authors’ elaboration. Figure B.2 shows non-parametric estimates of the density function for the percentage change in private consumption in the pwefood scenario. Again, the sign of the results (i.e., positive) is not changed when model elasticities are allowed to differ in +/- 80% of their “central” value. Figure B.1: sensitivity analysis, real private consumption deviation from base in 2030 -26- References Böhringer, Christoph and Thomas F. Rutherford (2002). In Search of a Rationale for Differentiated Environmental Taxes. ZEW Discussion Paper No. 02-30. Bovenberg, A. Lans and Goulder, Lawrence H. (1996). Optimal Environmental Taxation in the Presence of Other Taxes: General Equilibrium Analysis. The American Economic Review 86 (4): 985-1006. Decaluwé, Bernard, André Lemelin, Véronique Robichaud and Hélène Maisonnave (2012). The PEP Standard Computable General Equilibrium Model Single‐Country, Static Version PEP‐1‐1. Partnership for Economic Policy (PEP). MPIA Research Network. Decaluwé, Bernard, André Martens and Luc Savard (2001). La Politique Economique du Développement et les Modéles d'Equilibre Général Calculable. Montréal, Canada: Les Presses de l'Université de Montréal. Kim, Seung-Rae (2004). Uncertainty, Political Preferences, and Stabilization: Stochastic Control Using Dynamic CGE Models. Computational Economics 24 (2): 97-116. Mercado, P. R., D. Kendrick and H. Amman (1998). Teaching Macroeconomics with GAMS. Computational Economics 12 (2): 125-149. -27-
© Copyright 2026 Paperzz