Skills, innovation, and growth: An agent-based policy analysis | 300 350, Papers of History of Education

Download Skills, innovation, and growth: An agent-based policy analysis | 300 350 and more Papers History of Education in PDF only on Docsity! Skills, innovation, and growth: An agent-based policy analysis ∗ H. Dawid†, S. Gemkow†, P. Harting†, K. Kabus†, M. Neugart‡and K. Wersching† Abstract We develop an agent-based macroeconomic model featuring a dis- tinct geographical dimension and heterogeneous workers with respect to skill types. The model, which will become part of a larger sim- ulation platform for European policymaking (EURACE), allows us to conduct ex-ante evaluations of a wide range of public policy mea- sures and their interaction. In particular, we study the growth and labor market effects of various policy types that promote workers’ gen- eral skill levels. It is examined in how far effects differ if spending is uniformly spread over all regions in the economy or focused in one particular region. We find that the geographic distribution of pol- icy measures significantly affects the effects of the policy even if total spending is kept constant. Focussing training efforts in one region is the worst policy outcome while spreading funds equally across regions generates a larger output in the long-run but not in the short-run. 1 Introduction Normative research in economics has traditionally been, and to a large degree still is based on the development and analysis of highly stylized, analytically tractable models. In particular for macroeconomic issues the models used for policy analysis are typically dynamic general equilibrium models that ∗This research was funded by the European Commission as part of the FP6-STREP project EURACE (‘An agent-based software platform for European economic policy de- sign with heterogeneous interacting agents: new insights from a bottom up approach to economic modeling and simulation’). †Department of Business Administration and Economics, Bielefeld University, Germany ‡Social Science Research Center Berlin (WZB), Germany 1 have been calibrated using empirical data. However, numerous restrictive as- sumptions underly most mainstream analytical models (e.g. homogeneity of individuals, perfect rationality, rational expectations, perfect ex-ante coordi- nation in an equilibrium) and so far there exists almost no general theoretical basis that allows to judge how far findings, obtained under these simplifying assumptions, carry over to scenarios where agents are heterogeneous or out of equilibrium (see e.g. Kirman (1992)). On the other hand, recent devel- opments in computer technology and software engineering have made large scale simulations an increasingly powerful and attractive new approach for understanding the characteristics of economic systems and for deriving eco- nomic policy recommendations. In particular, by explicitly modelling the decentralized interaction of heterogeneous economic agents in systems like markets or organizations, agent-based computational economics (ACE) at- tempts to transcend the limitations of traditional models. The ACE modeling approach is not only well suited for explicit consid- eration of heterogeneity among economic agents, but also allows for a wide range of assumptions concerning the rules that determine the behavior of in- dividual economic agents. Existing ACE models typically consider adaptive individuals who learn how to behave and react based on their own experience and on the available information1. Models of this kind have been developed in many areas of economics. Among others, they have been used to study the emergence of trading behavior on goods-markets and on financial mar- kets, bidding behavior in auctions, numerous issues concerning innovation and industry evolution or the emergence of cooperative behavior in economic systems. Surveys over agent-based research in these areas can be found in Tesfatsion and Judd (2006). A major advantage of the agent-based simulation approach is that the modeler can easily study the effects of changes in the economic framework on various aspects of the resulting economic dynamics and long run out- come. Accordingly, agent-based models have a large potential as a tool to evaluate the impact of public policy measures or of changes in market de- sign. Whereas early ACE-work has been mainly of descriptive character, in recent years there have been a number of projects where agent-based mod- els have been employed to (normatively) address actual market design and public policy questions (see e.g. Dawid and Fagiolo (2007)). The considered policy issues relate to electricity markets (e.g. Bunn and Oliveira (2001), Nicolaisen et al. (2001)), labor market design (e.g. Neugart (2007)), auction design (e.g. Phelps et al. (2002)), patent policy (e.g. Malerba et al. (2001)) 1Work by Cyert and March (1963), Simon (1978, 1983) or Nelson and Winter (1982) has been influential for the development of this line of research. 2 normative analysis of economic policy is limited. An indicator in that respect is that the recently published Handbook of Regional and Urban Economics (Henderson and Thisse (2004)), where a lot of attention is dedicated to NEG, does not include any chapters discussing economic policy. Based on the obvious restrictions imposed by analytical tractability on the spatial analysis of the interplay of technological change and properties of the labor force, the need for an alternative approach for the evaluation of different combinations of economic policy measures seems particularly strong. Agent-based models are well suited to address this need2. In this paper we sketch an agent-based model that is as simple as possible but still apt to address some of the important policy questions raised above. It is a closed model consisting of a consumption goods and an investment (or capital) goods sector. Households and firms are distributed across regions. Some markets (consumption goods) are assumed to be local, some are global (investment goods) and workers might commute for work to employers in neighboring regions at certain costs. Investment goods are supplied at an exogenously given price. The inputs for consumption good production are investment goods and labor. Process innovation improves the quality of in- vestment goods leading to higher productivity of capital in the consumption good production. The labor market hosts workers of different types. They are differentiated along their general as well as their specific skill level. A crucial assumption is that sufficient specific skills of workers are needed to exploit the full potential of the advanced technological level of investment goods like production machines. Put formally, there is complementarity be- tween the average quality of the investment goods of a firm and the average level of specific skills of its employees with respect to the productivity in con- sumption good production. Empirical evidence for such complementarity can for example be found in Griffith et al. (2004). Workers of higher general skills adapt faster in terms of their specific skills needed to produce consumption goods by the use of investment goods of a particular quality. General skills affecting the adaptation of specific skills will allow us to study the effects of government policies improving the general skill level of the workforce on employment and growth. A general problem of agent-based models, that attempt to avoid the (overly) strong assumptions about information and rationality of individuals underlying equilibrium analysis, is the appropriate design of decision rules that govern the behavior of individual agents. Deviation from the intertem- poral (constrained) maximization paradigm opens many degrees of freedom 2See Dawid and Wersching (2006) or Wersching (2007) for agent-based analyses of spatial aspects of industry dynamics. 5 with respect to the type of behavioral rules used and the way behavior is adapted over time. However, as far as firm behavior is concerned for many operational decisions, like pricing, production and inventory choice or market selection decisions, standard decision rules and heuristics have been devel- oped that are well documented in the relevant business and operations man- agement literature. Our ‘philosophy’ in terms of modelling firm behavior is to implement relatively simple decision rules that match standard procedures of real world firms as described in the corresponding management literature. In a similar spirit the decisions of consumers, like the allocation of the available budget between consumption and savings, is modelled according to simple empirically founded rules from the literature. Apart from the fact that behavioral rules of individual agents in the model have to be in accordance with stylized representations of standard decision rules employed by their real-world representatives, it is also important to critically examine the plausibility of the qualitative patterns of simulation results. A widely used approach for this kind of model evaluation in re- cent work in agent-based economics is to compare simulation outcomes with ‘stylized facts’ that have been established using real world data. This kind of comparison is supposed to restrict the range of model parameters and to improve the confidence that the model captures crucial aspects of interac- tions in the sectors considered in the model, see Windrum et al. (2007) for a discussion of approaches to validate agent-based simulation models. Due to space constraints we only sketch such empirical validation of our model here and rather focus on the presentation of the economic logic exhibited by our policy experiments. The strong empirical footing of the developed agent-based model is not only of great importance for purely scientific reasons but is also crucial to establish trust of actual policy makers in the results of the policy analy- sis and the policy recommendations generated by the model. The model and preliminary analysis presented in this paper is part of the EU-funded project (EURACE) ‘An Agent-based Software Platform for European Eco- nomic Policy Design with Heterogeneous Interacting Agents: New Insights from a Bottom-Up Approach to Economic Modeling and Simulation’3. The main purpose of the project is to develop a unified agent-based macroeco- nomic simulation platform that can be used to inform policy makers about expected effects of (combinations of) various economic policy measures. Our focus here is on industrial and labor-market issues, but the goal of the entire project is to cover all main areas of economic policy. 3See http://www.eurace.org or http://www.wiwi.uni-bielefeld.de/̃ dawid/eurace/ for more information about the EURACE project. 6 We proceed as follows. The main features of the simulation model are described in section 2. In section 3 we briefly present key features of simu- lation results generated by the model and then discuss the potential of the framework to evaluate different types of policy measures that aim at faster economic growth. The potential of the model for policy analysis is then illus- trated by comparing the effects of different degrees of spatial concentration of policy measures. We conclude in section 4 with a brief discussion and pointers to future work. 2 The model 2.1 General features Our model consists of a capital good, a consumption good, and a labor mar- ket.4 Capital goods are provided with infinite supply at exogenously given prices. The quality of the capital good improves over time where techno- logical change is driven by a stochastic (innovation) process. Firms in the consumption goods sector use capital goods combined with labor input to produce consumption goods. The labor market is populated with workers that have a finite number of general skill levels and acquire specific skills on- the-job which they need to fully exploit the technological advantages of the capital employed in the production process. Consumption goods are sold at malls. Malls are not treated as profit-oriented enterprises but simply as local market platforms where firms store and offer their products and consumers come to buy goods at posted prices. Thus, two types of active agents and two types of passive agents (in the sense that this type of agent does not take any decisions) are present in the model. Each type of active agent has several ‘roles’ corresponding to its activities in the different markets. Table 1 summarizes these roles. The economy consists of R = 2 regions and each agent is located in one of these regions. Some actions occur locally, such as the agents’ consump- tion, others occur globally including the sale of the investment good or labor supply. Generally, the minimal unit of time is a day, however almost all the inter- actions and decisions are repeated on a monthly basis.5 Therefore, whenever 4In the fully fledged EURACE model, a financial and a credit market will be added, and an exogenous energy market will constitute a proxy for the link to the ‘rest-of-the-world’ by affecting the production costs in the capital goods market. 5In the model each week consists of 5 days and each month of 4 weeks. Accordingly, each year has 240 days. 7 The quantities actually delivered to the malls, Di,r,t, are adjusted propor- tional to the intended quantities D̃i,r,t, so that Di,r,t = D̃i,r,t∑R r′=1 D̃i,r′,t ·Qi,t. Production times of consumption goods are not explicitly taken into account and the produced quantities are delivered on the same day when production takes place. The local stock levels at the malls are updated accordingly. 2.3.2 Factor demand Consumption good producers, denoted by i, need physical capital and labor to produce the consumption goods. The accumulation of physical capital by a consumption good producer follows Ki,t+1 = (1− δ)Ki,t + Ii,t where Ki(0) = 0 and Ii,t > 0 is the gross investment. Every worker w has a level of general skills bgenw ∈ {1, . . . , bgenmax} and a level of specific skills bw,t. The specific skills of worker w indicate how effi- ciently the corresponding technology is exploited by the individual worker. Building up those specific skills depends on collecting experience by using the technology in the production process. There is vast empirical evidence for such adjustment processes (see e.g. Argote and Epple (1990)). The shape of the evolution of productivity follows a concave curve, the so-called learning curve, when the organizational productivity is recorded after implementing a new production method or introducing a new good. Concavity in this context means that the productivity rises with proceeding use of the produc- tion method or production of the new good, but this increase emerges at a decreasing rate. We transfer this pattern of organizational learning on the individual level and assume that the development of individual productivity follows a learning curve. The specific skills are updated once in each produc- tion cycle of one month. Further, we assume that updating takes place at the end of the cycle. A crucial assumption is the positive relationship between the general skills bgenw of a worker and his ability to utilize his experiences. Building up worker’s technology specific skills depends on a worker’s level of general skills, i.e. his education and the other abilities which are not directly linked to the particular technology. Taking the relevance of the general skill level into account the specific skills of a worker w for technology j is assumed to evolve according to bw,t+1 = bw,t + χ(b gen w ) · (Ai,t − bw,t) , 10 where we denote with Ai,t the average quality of the capital stock. The function χ is increasing in the general skill level of the worker. Note that this formulation captures the fact that in the absence of technology improve- ments marginal learning curve effects per time unit decrease as experience is accumulated and the specific skills of the worker approaches the current technological frontier. The production technology in the consumption goods sector is repre- sented by a Cobb-Douglas type production function with complementarities between the quality of the investment good and the specific skills of employ- ees for using that type of technology. Factor productivity is determined by the minimum of the average quality of physical capital and the average level of relevant specific skills of the workers. Capital and labor input is substi- tutable with a constant elasticity and we assume constant returns to scale. Accordingly, output for a consumption goods producer is given by Qi,t = min[Bi,t, Ai,t]× Lαi,tK β i,t, where Bi,t denotes the average specific skill level in firms and α+ β = 1. Firms aim to realize a capital to labor ratio according to the standard rule for CES production functions. That is a ratio of quantity to price of the two factors proportional to the corresponding intensity parameter. Accordingly, K̃i,t pinv / L̃i,t wet = β α . Taking into account the above production function this yields under the assumption of positive investments ˜̃Ki,t = (βwet ) αQ̃i,t (αpinv)α min[Ai,t, Bi,t] ˜̃Li,t = (αpinv)βQ̃i,t (βwet ) β min[Ai,t, Bi,t] and if ˜̃Ki,t ≥ (1−δ)Ki,t−1 the desired capital and labor stocks read K̃i,t = ˜̃Ki,t and L̃i,t = ˜̃Li,t. Otherwise, we have K̃i,t = (1− δ)Ki,t−1 L̃i,t = ( (Q̃i,t ((1− δ)Ki,t−1)β min[Ai,t, Bi,t] )1/α . 11 For simplicity credit constraints are not incorporated in this version of the model.6 All desired investments can be financed. The monthly realized profit of a consumption goods producer is the dif- ference of sales revenues achieved in the malls during the previous period and costs as well as investments (i.e. labor costs and capital good investments) borne for production in the current period. In cases of positive profits, the firm pays dividends to its stockholders and the remaining profits, as well as losses, are entered on an account Acci,t. Similar to the capital goods pro- ducer, we assume that all households hold equal shares in all consumption goods producers, consequently the dividends are equally distributed to the households. In order to avoid exceeding accumulations of savings as well as excessive indebtedness, we employ a simple dividend policy that provides different dividend rates depending on the current balance of saving account Acci,t. The rule states that a firm pays no dividends, if the balance is negative and the debt is on a scale above the last monthly revenue. If the balance is positive and savings are above the monthly revenue, the firm disburses all profits. In the remaining case, if the balance is between these critical levels, a fixed proportion div ∈ [0, 1] of profits is paid out. Since there are no constraints on the credit market and there is infinite supply of the investment good, the consumption goods producers are never rationed on the investment goods market. Wages for the full month are paid to all workers at the day when the firm updates its labor force. Investment goods are paid at the day when they are delivered. 2.3.3 Pricing Consumption good producers employs a standard approach from the man- agement literature, the so-called ‘break-even analysis’ (see Nagle (1987)), to set their prices. The break-even formula determines at what point the change in sales becomes large enough to make a price reduction profitable and at what point the decrease in sales becomes small enough to justify a rise in the price. Basically, this managerial pricing rule corresponds to standard elasticity based pricing. Assuming that all firms have constant expectations εei < −1 of the elas- ticity of their demand, they set the price according to the standard rule pi,t = c̄i,t−1 1 + 1/εei , where c̄i,t−1 denotes unit costs in production of firm i in the previous period. 6In contrast, in the fully fledged EURACE platform, there is an explicit credit market model which can be appropriately linked to the real sectors considered here. 12 where p̄k,t−1 is the average price of the goods consumer k has consumed in t − 1. The consumer selects one good i ∈ Gk,weekt , where the selection probability for i reads Probk,i,t = Exp[λconsk vk(pi,t)] ςnonek Exp[λ cons k v none k (p̄k,t−1)] + ∑ i′∈Gk,weekt Exp[λconsk vk(pi′,t)] . Thus, consumers prefer cheaper products and the intensity of competition in the market is parameterized by λconsk . Once the consumer has selected a good he spends his entire budget Bconsk,weekt for that good if the stock at the mall is sufficiently large. In case the consumer cannot spend all his budget on the product selected first, he spends as much as possible, removes that product from the list Gk,weekt , updates the logit values and selects another product to spend the remaining consumption budget there. If he is rationed again, he spends as much as possible on the second selected product, rolls over the remaining budget to the following week and finishes the visit to the mall. 2.5 Labor market 2.5.1 Labor demand Labor demand is determined in the consumption goods market. If the firms plan to extend the production they post vacancies and corresponding wage offers. The wage offer wOi,t keeps unchanged as long as the firm can fill its vacancies, otherwise the firm updates the wage offer by a parameterized fraction. In case of downsizing the incumbent workforce, the firm dismisses workers with lowest general skill levels first. 2.5.2 Labor supply Job seekers consist of a randomly determined fraction φ of employed workers who search on-the-job and the unemployed. A worker k only takes the posted wage offer into consideration and compares it with his reservation wage wRk,t. A worker will not apply at a firm that makes a wage offer which is lower than his reservation wage. The level of the reservation wage is determined by the current wage if the worker is employed, and in case of an unemployed by his adjusted past wage. That is an unemployed worker will reduce his reservation wage with the duration of unemployment. When a worker applies he sends information about his general as well as his specific skill level to the firm. 15 2.5.3 Matching algorithm According to the procedures described in the previous sections consumption goods producers review once a month whether to post vacancies for pro- duction workers. Job seekers check for vacancies. The matching between vacancies and job seekers works in the following way: Step 1: The firms post vacancies including wage offers. Step 2: Every job seeker extracts from the list of vacancies those postings to which he fits in terms of his reservation wage. Job seekers rank the suitable vacancies. The vacancy which offers the highest wage is ranked on position one and so on. If the wage offers that come with the posting are equal, vacancies are ranked by chance. Step 3: Every firm ranks the applicants. Applicants with higher general skill bgen levels are ranked higher. If there are two or more applicants with equal general skill levels, but different specific skill levels, the applicant with the higher specific skill level is ranked higher. Based on their rank- ing firms send job offers to as many applicants as they have vacancies to fill. Step 4: Each worker ranks the incoming job offers according to the wages net of commuting costs (comm > 0) that may arise if he was to accept a job in the region where he does not live. Each worker accepts the highest ranked job offer at the advertised wage rate. After acceptance a worker refuses all other job offers and outstanding applications. Step 5: Vacancies’ lists and applications’ lists are adjusted for filled jobs. If a firm received refusals, these applicants are dropped from the list of applicants. If all vacancies of a firm have been filled the firm refuses the other applicants and the algorithm for this firm ends. Step 6: If the number of vacancies not filled exceeds some threshold v > 0 the firm raises the wage offer by a fraction ϕi such that w O i,t+1 = (1+ϕi)w O i,t. If an unemployed job seeker did not find a job he reduces his reservation wage by a fraction ψk, that is (w R k,t+1 = (1 − ψk)wRk,t). There exists a lower bound to the reservation wage wRmin which may be a function of unemployment benefits, opportunities for black market activity or the value of leisure. If a worker finds a job then his new reservation wage is the actual wage, i.e. wRk,t = wi,t. Go to step 1. This cycle is aborted after two iterations even if not all firms may have satis- fied their demand for labor. As indicated above this might lead to rationing 16 of firms on the labor market and therefore to deviations of actual output quantities from the planned quantities. In such a case the quantities de- livered by the consumption good producer to the malls it serves is reduced proportionally. This results in lower stock levels and therefore increases the expected planned production quantities in the following period. 3 Simulation results and policy experiments 3.1 The base case Before we illustrate the potential of our model for carrying out policy exper- iments with respect to the spatial distribution of policy measures we show that it generates time series of key economic variables with very plausible features. To that end we consider a base scenario which we will refer to as the uniform low skill scenario. Throughout the paper we assume that there are bgenmax = 5 levels of general skills. The function χ(b gen w ), which governs the speed of specific skill improvement, is chosen such that the time workers with general skill 3 need to cut the gap between their specific skill and the firm’s technology level in half is the mean of the corresponding time needed by a skill level 1 and a skill level 5 worker. An analogous linear relationship also determines the adjustment speed of workers with general skill levels 2 and 4. In a low skill region the skill distribution is such that 80% of workers have the lowest general skill level, whereas the remaining workers are equally distributed across the other four levels of general skills. Analogously, a re- gion is a medium skill or high skill regions if 80% of workers have general skill level 3 respectively 5. In the uniform low skill scenario both regions are low skill regions. In addition we assume that there is a 10% probability of a quality improving innovation in the investment goods sector per month and each innovation on average increases the quality of the investment good by 5%. This corresponds to an average productivity growth rate of 6% if specific skills were sufficiently large to fully exploit all new technologies. Simulations are run for 4000 days which corresponds to about 17 years. The full set of pa- rameters employed in this simulation is given in the Appendix. As discussed above, the main purpose of this exercise is to demonstrate the potential of our approach for carrying out spatial policy experiments and at this point no serious calibration of parameters using empirical data has been carried out. In figure 1 we show the time series for a single run of output, unemployment, (skill-dependent) wages and specific skills along the technology frontier in a typical run in the uniform low-skill scenario. Output is increasing over time with some fluctuations. Unemployment 17 lic spending between primary and secondary education, tertiary eduction, life-long learning measures and dual apprenticeships. An additional aspect in the policy debate about the optimal design of educational policies fostering innovation and growth is the question in how far such policies should depend on the technological distance of the firms in the economy from the technological frontier. Claims have been made that in regions far from the frontier the main focus should be on primary and sec- ondary education improving the lower end of the skill distribution, whereas regions at the frontier profit more from higher investments in the tertiary sector (see Aghion (2007)). From a dynamic perspective this raises the ques- tions in how far the effectiveness of different policy approaches depends on the speed at which the technological frontier is moving, which itself is in- fluenced by measures in the area of innovation and technology policy. Such measures include among others direct funding of basic and applied research, financial incentives for R&D efforts by firms, providing infrastructure and incentives for R&D cooperation and intensive knowledge exchange between research institutions and firms. The version of our framework presented here captures the effects of such policies in reduced form through the parameters γinv and ∆qinv. Policy efforts aiming at a speed-up of the movement of the technological frontier should lead to an increase in γinv, whereas attempts to foster basic research and to move from incremental to more fundamental innovations in the investment goods sector should trigger an increase in ∆qinv. Our framework allows to examine how such policies would influence the relative performance of different policies that influence the skill distribution. Understanding such cross-effects is essential to coordinate the different policy measures and to design a well balanced package of measures where the negative interaction terms between the different parts are avoided. We illustrate the potential insights from such policy experiments by con- sidering a simple example. Starting from the uniform low skill scenario we assume that a policy-maker responsible for economic development in both regions considers to launch a campaign in order to increase the general skill level of workers in the region. Such a campaign might involve improvements in the system of lower and higher education, providing incentives for firms to train workers, e-learning activities and so on. Furthermore, it is assumed that using the budget assigned to this campaign the policy-maker can either improve both regions from low-skill to medium-skill regions8 or, if all efforts are invested in only one region, improve one region to a high-skill region with 8According to our definition above in a medium-skill region 80% of workers have general skill level 3 and the remaining 20% are equally distributed among the other four skill levels. 20 0 50 100 150 200 30 0 35 0 40 0 45 0 50 0 55 0 Months O ut pu t Figure 2: Batch run for outputs in uniform low scenario (dashed line), uni- form medium scenario (solid line), and low/high scenario (dotted line) the other region remaining low-skill. Since in both cases the total number of general-skill units added to the work force is identical, it is a-priori a non- trivial question whether the two policy options yield identical results and, if not, in which way the effects will differ. In order to address this question we have run batches of 50 simulation runs for the uniform medium and high-low scenarios and compare them with each other and with the base case of uniform low-skill regions. In figure 2 we depict mean trajectories over the 50 runs in the three scenarios. As can be seen, the difference in the effect of the two policy measures is quite striking. Focussing the resources to improve the general skill level of workers in one region only (low/high scenario) yields worse results in terms of output if compared to the uniform low skill scenario. In the short run there is also a negative effect of the policy that spreads resources equally across the two regions. However, in the longer run raising the general skill level of workers in both regions yields an output that is higher than in the uniform low skill case and the low/high scenario. The large impact of the spatial allocation of policy measures in this simple example highlights the strongly non-linear and path dependent nature of a micro-founded agent-based model which leads to 21 non-trivial relationships between policy intervention and emergent results. A more detailed consideration of the mechanisms underlying the different effects of these policy measures shows that the interplay of wage-dynamics, wage-driven price dynamics, and barriers to skill transfer through commut- ing workers that arises in the high-low scenario is responsible for the larger growth of output and employment in the uniform medium scenario. In the low/high scenario depicted in the upper left panel of figure 3 the average specific skills increase more in the region at which the policy measure is targeted. Obviously, workers learn the specific skills more quickly as their general skills improved while no such process arises in the region without improvements in the general skills of the workers. Allocating resources to improve general skills equally across regions yields the expected results that the improvement of average specific skills is equal in both regions (upper right panel). The lower two panels give insight into the workings of the model by showing the diverging patterns of output across the two regions comparing the two policies. In the low/high scenario output increases for the region with the higher average specific skill level. Clearly, firms in this region take advantage of the more skilled workforce which allows them to exploit the technological improvement of their capital stock to a larger degree than the firms in the other region. Here, we actually observe a decline in output. Adding up regional outputs yields an increasing total output. However, total output does not increase so strongly as in the uniform medium skill scenario depicted in the lower right panel. Here, as average specific skills in both regions have increased somehow due to the equal distribution of policy efforts, output increased also in both regions. In total this has a larger impact on output in the longer run which explains why targeting the policies to only one region turns out to be the inferior policy with respect to output gains. Looking at prices and commuting employees in the low/high scenario corroborates the analysis of our major finding of the policy experiment. Due to the faster adaptation of specific worker skills to the improving technological frontier firms in the high-skill region after a short initial phase produce with lower unit costs compared to those in the low skill region. It should be noted that this observation holds true in spite of the fact that wages in the high skill region are higher than those in the low skill region (not shown in our figures). As can be seen in the left panel of figure 4 the cost advantage translates to a lower price of the goods offered by producers from the high- skill region and this shifts demand away from producers in the low skill region towards producers in the high skill region. However producers in the high skill regions are limited in their output expansion by the lack of additional local workers. They start hiring workers from the low skill region (see the lower solid line in the right panel of figure 4), but since this is associated 22 4 Conclusions Agent-based models are not only a useful tool to transcend borders set by standard analytical models for the study of economic systems but also open up promising avenues for economic policy design. We developed a still par- simonious micro-based macroeconomic model to study the role of innovation and skills for economic growth and employment in a regional setting. More effort has to be put into seriously calibrating our model. But still our policy experiments – which as we believe are ultimately of high relevance for policy- makers concerned about the right mix of policies targeted toward improving specific or general skills, and giving firms incentives to innovate – yield very plausible insights, highlighting the role of non-trivial interaction effects in a dynamic setting which are hard to analyze in more standard static frame- works. Without stressing the results of our policy experiments too strongly there occurs to be a point in allocating resources to improve general skills equally across regions to improve output in the long run. The next steps towards providing a software platform for policymakers that eventually will allow for ex-ante policy evaluations are to first take more serious steps towards calibrating the model. On the list of extensions to the model introducing the possibility to study product versus process innovations looms high. Finally, as pointed out earlier, this project is part of a larger endeavor. Within EURACE we aim at integrating models for the credit and financial markets build by our partners to finally arrive at a fully fledged closed macroeconomic model. Obviously, this will add many more possibilities to conduct policy analyses of the type we were trying to promote in our attempt here. Acknowledgements This work is part of the EURACE research consortium financed by the 6th Framework of the European Commission. The aim of the research project is to build a software platform for evaluating European economic policies. We are thankful to our colleagues for making valuable contributions to this part of the research project. The model has been implemented in FLAME, an agent-based simulation environment developed by project partners working in the area of software engineering. A list of project participants and further information on EURACE is available at www.eurace.org. 25 Appendix In table 2 the parameter setting for the simulation is summarized. Table 2: Parameter settings Description Parameter Value Number of households: 400 Number of firms 10 Number of regions R 2 Labor Intensity of Production α 0.75 Labor Intensity of Production β 0.25 Innovation probability γinv 0.1 Depreciation rate of capital δ 0.05 Monthly Discount factor ρ 0.995 Production smoothing ξ 0.5 Mark-up factor 1|ei |−1 0.15 Wage update φi 0.02 Reservation wage update ψk 0.01 Minimal reservation wage wRmin 1 Marginal saving propensity κ 0.1 Intensity of choice by consumers λconsk 30 Vacancy threshold v 2 Commuting costs comm 0.2 References Aghion, P. (2007). Growth and the financing and governance of education. Keynote Lecture for the 2007 Meeting of the German Economic Associa- tion. Argote, L. and Epple, D. (1990). Learning curves in manufacturing. Science, 247:920–924. Audretsch, D. and Feldman, M. (2004). Knowledge spillovers and the geog- raphy of innovation. In Henderson, J. and Thisse, J.-F., editors, Handbook of Regional and Urban Economics Volume 4: Cities and Geography, pages 2715–2739. North-Holland. 26 Bassanini, A. (2004). Improving skills for more and better jobs? the quest for efficient policies to promote adult education and training. European Economy: Special Reports, 4(3):103–137. Bassanini, A. and Scarpetta, S. (2002). Does human capital matter for growth in oecd countries?: A pooled mean group approach. Economics Letters, 74(3):399–405. Bosker, M., Brakman, S., Garresten, H., and Schramm, M. (2007). Adding geography to the new economic geography. Working Paper No. 2038, Ce- sifo. Bunn, D. W. and Oliveira, F. (2001). An application of agent-based simu- lation to the new electricity trading arrangements of england and wales. IEEE Transactions on Evolutionary Computation, 5(5):493–503. Chiaromonte, F. and Dosi, G. (1993). Heterogeneity, competition and macroeconomic dynamics. Structural Change and Economic Dynamics, 4:39–63. Cyert, R. M. and March, J. G. (1963). A Behavioral Theory of the Firm. Prentice Hall, Englewood Cliffs. Dawid, H. and Fagiolo, G. (2007). Agent based models for economic policy design. Forthcoming special issue of the Journal of Economic Behavior and Organization. Dawid, H. and Wersching, K. (2006). On technological specialization in industrial clusters: an agent-based analysis. In Rennard, J.-P., editor, Handbook of Research on Nature Inspired Computing for Economics and Management, pages 367–378. Idea. Deaton, A. (1991). Saving and liquidity constraints. Econometrica, 59:1221– 1248. Deaton, A. (1992). Houshold saving in ldcs: Credit markets, insurance and welfare. Scandinavian Journal of Economics, 94:253–273. Delli Gatti, D., Guilmi, C., Gaffeo, E., Giulioni, G., Gallegati, M., and Palestrini, A. (2005). A new approach to business fluctuations: hetero- geneous interacting agents, scaling laws and financial fragility. Journal of Economic Behavior and Organization, 56:489–512. 27