Economic Report 48 Original Version: October 1998 Revised: March 1999
Market Power Eects in Evolutionary Labor Markets with Adaptive Search1 Leigh Tesfatsion
Department of Economics Iowa State University, Ames, IA 50011-1070 http://www.econ.iastate.edu/tesfatsi/
[email protected]
Abstract: Real-world labor markets are characterized by highly heterogeneous out-
comes relative to observable heterogeneity in worker and employer attributes. Microfoundation explorations of the possible sources of this outcome heterogeneity are dicult due to analytical intractability and data measurement problems. This study uses an agent-based computational economics model of a labor market to undertake a systematic experimental study of how market power asymmetries between workers and employers can induce persistent heterogeneity in average earnings across employed workers with initially identical observable structural attributes, and similarly for nonvacant employers. This outcome heterogeneity arises because workers and employers adaptively search for worksite partners and evolve their worksite behaviors over time, which can result in persistent path-dependency eects (hysteresis) in the presence of market power asymmetries. Particular attention is focused on the experimental determination of correlations between market power asymmetry and the formation of contractual networks among workers and employers, and between contractual network formation and the types of worksite interactions and earnings outcomes that these contractual networks support.
Keywords: Labor markets; market power; agent-based computational economics;
search and matching; contractual networks; endogenous interactions; adaptation and learning; hysteresis; transactions costs; evolutionary game.
1
The author is grateful to J. Duy, H. Gintis, S. Moss, P. Orazem and R. Reed for helpful comments.
1. INTRODUCTION Real-world labor market outcomes appear to be excessively heterogenous, in the sense that observationally equivalent workers and employers have markedly dierent earnings and employment histories. This puzzle has motivated an extensive analytical and empirical literature directed towards possible resolutions. In analytical studies, much attention has been focused on the possibility of heterogeneous labor market outcomes arising from search and matching processes carried out by imperfectly informed workers and employers. In uential early work in this area includes Jovanovic (1979), Diamond (1982), and Pissarides (1985); more recent work includes Pissarides (1990), Hosios (1990), Mortenson and Pissarides (1994), Aghion and Howitt (1994), and MacLeod and Malcomson (1998). To achieve analytical tractability, however, these studies commonly postulate an aggregate matching function that proxies the complicated process of employer recruitment, worker search, and mutual evaluation. That is, the intense individualistic rivalry among workers and employers that characterizes many real-world labor markets is not modelled. Moreover, the competition of ideas within agents is generally not considered; attention is largely focused on steady-state behavior.2 Empirical studies have typically focused on the extent to which heterogeneous labor market outcomes might arise either from unobservable dierences among workers or from unobservable dierences among employers. In a recent interesting study, however, Abowd et al. (1999) develop an empirical wage equation in which unobservable worker and employer dierences are both explicitly taken into account along with observable dierences. Speci cally, the wage equation developed by Abowd et al. (1999) generates a conditional exImportant exceptions exist, of course. For example, Pissarides (1985) studies the dynamics of adjustment in labor markets in response to shocks. Montgomery (1991) explores the relationship between social networks and labor market outcomes when employers are able to use a stochastically determined social network among workers to relay job oers to the acquaintances of high-ability workers. In addition, motivated in part by an early seminal paper by Myerson (1977), an interesting analytical literature has developed in recent years focusing on the endogenous formation of agent linkages in social networks, which has potentially important implications for labor market modelling. See, for example, Bowles and Gintis (1998), Kirman et al. (1986), van den Nouweland (1993), Jackson and Wolinsky (1996), and Watts and Strogatz (1998). 2
1
pectation for the annual compensation of an individual worker given information on the observable characteristics of the worker, the date of observation, the identity of the individual worker (the \person eect"), and the identity of the employing rm (the \ rm eect"). They demonstrate that the omission of either the person eect or the rm eect from their wage equation can result in a substantially biased estimate for the remaining eect. Applying their wage equation to an extensive longitudinal data set for French workers and employers from 1976 through 1987, they nd that person eects and rm eects are both important as determinants of compensation. However, person eects appear to have substantially more explanatory power than rm eects, particularly with regard to inter-industry wage dierentials. Nevertheless, Abowd et al. (1999, p. 252) warn that computational complexity prevents the full least squares estimation of their wage equation with unobserved heterogeneity in the form of both person and rm eects. They also note that some of their estimated magnitudes (e.g., the estimated eect of seniority on wage rates) appear to be highly sensitive to the particular estimation technique employed. In addition, the underlying labor market is not modelled, and no microeconomics explanation is provided for their estimated person and rm eects. Rather, the authors maintain an assumption of exogenous mobility and state (pp. 255 and 313) that the extent to which person and rm eects arise from considerations such as endogenous mobility, agency, eciency wage, seach/matching, and rent sharing will be the subject of future analyses. As the above remarks indicate, substantial intractability problems are encountered in both analytical and empirical studies when an attempt is made to trace apparently excessive heterogeneity in labor market outcomes back to microeconomic sources. The question then arises whether at least some of these intractability problems might be alleviated by taking an agent-based computational economics (ACE) approach. ACE is the computational study of economies modelled as evolving decentralized systems of autonomous interacting agents.3 The distinctive feature of ACE modelling is that each agent 3
For a variety of downloadable ACE-related resources, including surveys, an annotated syllabus of read-
2
has a cognitive structure consisting of internal behavioral rules acting on internal state variables. Starting from initial conditions set by the modeller, the agents in an ACE-modelled economy evolve their behavioral rules over time in a manner analogous to the growth of a culture in a petri dish. During the course of this growth process, which proceeds without need for any further outside intervention, the modeller can gather detailed information about the economy ranging from realistically obtainable macroeconomic indicators all the way down to the speci c internal state values and behavioral rules for individual agents. An ACE approach thus appears to be particularly well suited for investigating the microeconomic sources of heterogeneity in labor market outcomes. This is especially the case if, as suggested by the ndings of Abowd et al. (1999), a substantial portion of this outcome heterogeneity results from dierences in worker and employer characteristics that are either dicult or impossible to observe in practice. The current study uses an ACE labor market to explore the extent to which asymmetric market power can induce outcome heterogeneity when workers and employers are endogenously mobile and strategically interactive. Speci cally, it is shown that asymmetric market power between workers and employers, as measured by the degree to which potential job oers exceed potential work oers, can induce persistent variations in search and worksite behaviors across workers even when all workers have initially identical structural attributes apart from random variations in their worksite strategies, and similarly for employers. These persistent variations in the search and worksite behaviors of workers and employers can in turn induce persistent heterogeneity in earnings and employment histories. More precisely, the ACE labor market used in this study builds on the trade network game (TNG) model developed in Tesfatsion (1995, 1997a,b) for studying the formation and evolution of buyer-seller trade networks under alternatively speci ed market structures. First developed in ings, software, and pointers to individual researchers and research groups, see the ACE web site maintained at http://www.econ.iastate.edu/tesfatsi/ace.htm.
3
preliminary fashion in Tesfatsion (1998), the ACE labor market comprises workers and employers who repeatedly choose and refuse worksite partners on the basis of continually updated expected utility, engage in worksite interactions modelled as two-person prisoner's dilemma games between worker and employer partners, and evolve their worksite strategies over time on the basis of the earnings obtained from past worksite interactions. The experimental design of the study involves the systematic variation of two treatment factors: basic market structure; and the relative market power of workers as measured by the ratio of potential job openings to potential work oers. Regarding the rst treatment factor, three distinct labor market structures are investigated: two-sided markets comprising pure workers and pure employers; partially uid markets comprising pure workers, pure employers, and \worker-employers," i.e., agents capable of functioning either as workers or as employers; and endogenous-type markets in which each agent is a worker-employer. Regarding the second treatment factor, for each market structure the relative market power of workers is systematically varied from high to low. For each given speci cation of the treatment factors, all workers in the initial period are constructed as structurally identical agents apart from random variations in their worksite strategies, and similarly for employers. Twenty runs (\sample economies") are then generated using twenty dierent pseudo-random number seed values.4 In examining the resulting histories for the sample economies, particular attention is focused on the experimental determination of correlations between the treatment factors and the formation of persistent contractual networks among workers and employers, and between contractual network formations and the types of persistent worksite interactions and earnings outcomes that these contractual networks support. As will be shown in later sections, the ACE labor market permits many aspects of labor markets to be determined endogenously on the basis of local agent interactions that analytical and empirical All labor market experiments reported in this study were implemented using version 105b of the trade network game (TNG) source code developed by McFadzean and Tesfatsion (1997), which in turn is supported by SimBioSys, a general C++ class framework for evolutionary simulations developed by McFadzean (1995). Version 105b of the TNG source code and SimBioSys are both available for downloading as freeware at the current author's web site, along with extensive user instructions. 4
4
labor market studies generally either omit or specify in more restricted ways (e.g., as exogenously given attributes or as assumed aggregate steady-state relationships). Examples of these endogenized aspects include: worker preference orders over potential employers; employer preference orders over potential workers; intensity of job search activity; search opportunity costs; contractual networks among workers and employers; worker and employer worksite behaviors; relative worker and employer earnings; quit rates; ring rates; vacancy rates; and unemployment rates. For example, one interesting nding of the current study is that the intuitively expected correlations between unemployment and vacancy rates and market power asymmetries, while present, are tempered by the endogenous mobility of workers and employers. As the relative market power of workers increases, observed unemployment rates among workers tend to decrease; but workers with greater relative market power tend to evolve worksite strategies with higher defection frequencies, which can lead to increased unemployment due to ring (employers refusing further work oers from predacious workers). Similarly, as the relative market power of employers increases, observed vacancy rates among employers tend to decline; but increased defection frequencies by employers can lead to increased vacancies through quits (exploited workers directing future work oers elsewhere). In addition, the ACE labor market permits the investigation of hysteresis (path-dependency) eects arising from local agent interactions that could help to explain the apparently excessive heterogeneity observed in labor market outcomes. Two basic types of hysteresis eects are encountered in the current study. First, job search in the ACE labor market is costly, in the sense that workers bear the costs of wasted time spent in submitting unsuccessful work oers to employers during the course of job search. In the presence of market power asymmetries, these sequentially incurred transactions costs can induce path-dependent contractual networks among workers and employers that support persistently heterogeneous average earnings across employed workers and nonvacant employers even when each matched worker and employer pair expresses the same worksite behavior (e.g., mutual 5
cooperation). Since the earnings heterogeneity arises from a structural market power asymmetry and not from any de ciency in the worksite strategies of the workers per se, it cannot be remedied by evolutionary selection pressures acting upon worksite strategies. This eect is referred to below as a network hysteresis eect . Second, even when each agent within each agent type engages in the same general pattern of relationships as other agents of his type { e.g., each worker continuously works for just one employer { market power asymmetry together with endogenous mobility can result in persistently dierent expressed worksite behaviors that support persistently heterogeneous average earnings among agents of the same type. These persisently dierent expressed worksite behaviors may be generated by similar underlying worksite strategies (e.g., they may represent dierent branches of the same game tree), so that the resulting earnings heterogeneity cannot be remedied by evolutionary selection pressures acting upon worksite strategies. This eect is referred to below as a behavioral hysteresis eect .
At a more global level, these network and behavioral hysteresis eects result in a one-to-many mapping between treatment factors and experimental outcomes. For each particular treatment, varying the initial random seed value across experimental runs, a small but multiple number of distinct contractual network formations are observed to arise and persist among workers and employers across runs, each supporting a distinct pattern of worksite behaviors and earnings outcomes. This nding is consistent with the many analytical two-sided labor market studies, such as Diamond (1982), that establish the existence of multiple steady-state search equilibria. In the current process study, however, a histogram is obtained for each treatment showing the proportion of runs (sample economies) that evolve each type of network formation, which provides suggestive information regarding the size and importance of their basins of attraction. Although multiple network formations are observed for each treatment, the distribution of these network formations and the nature of the worksite behaviors they support tend to be strongly correlated with the dierential market power of workers and employers as implied by the speci cation 6
of the treatment factors. For example, when workers have relatively low market power, and hence low average match rates, the predominant types of networks that form tend to support some amount of aggressive and predacious employer behavior against pure workers, although this behavior is tempered by the ability of workers to quit an employer or even to exit the workforce altogether. Conversely, when workers have relatively high market power, implying that pure employers have low average match rates, the predominant types of networks that form tend to support a substantial amount of aggressive and predacious worker behavior against pure employers. These ndings provide support for the analytical studies of Pissarides (1985), Rubinstein and Wolinsky (1985), and Wolinsky (1987), among others, who explore search and match models in which the relative returns to the participants of a match are assumed to be aected by their relative match probabilities. In the present context, however, these eects are not symmetrical between workers and employers because of an assumed asymmetry in the job search process: while workers actively make work oers to employers, which the employers either accept or refuse, pure employers do not initiate contacts with workers. Consequently, pure employers are less \mobile" than workers and hence more susceptible to exploitation by workers when workers have relatively high market power, i.e., when there is excess job capacity. Also, the average earnings levels achieved by each agent type (pure workers, pure employers, and worker-employers) across the sample economies generated for each treatment tend to strongly re ect the dierential market power of workers and employers as implied by the speci cation of the treatment factors. These average earnings levels { which re ect both worksite earnings and transactions costs { typically xate within a very short time at approximately constant levels that are maintained throughout the remainder of the experiment. Finally, another experimental nding that may have rami cations for matching research in general concerns the conventional Pareto eciency and stabilility criteria used to evaluate matching mechanisms in static or steady-state settings. These criteria are local optimality criteria in the sense 7
that preference orders over potential partners are taken as given and search costs are ignored. In contrast, in the ACE labor market there is a continual feedback between preference orders over potential partners and the outcomes of agent worksite interactions over time. Moreover, job search costs turn out to have substantial eects over time both on network formation and on attained earnings levels. In consequence, the welfare predictions generated by these local optimality criteria are not globally reliable. This suggests the need to devise global optimality criteria for matching mechanisms that take into account the joint determination of matching networks and interaction behaviors over time as well as the accumulation of search costs. The ACE labor market is described in Section 2. In Section 3, descriptive statistics are constructed for the ex post classi cation of network formations, worksite behaviors, and earnings outcomes. The experimental design of the study is outlined in Section 4, and a detailed discussion of experimental ndings is presented in Section 5. Concluding remarks are given in Section 6.
2. THE ACE LABOR MARKET The ACE labor market consists of three disjoint (and possibly null) subpopulations of agents that separately evolve over time: pure workers who make work oers; pure employers who receive work oers; and worker-employers capable of both making and receiving work oers. The pure workers and worker-employers are collectively referred to as workers , and the worker-employers and pure employers are collectively referred to as employers . Each worker can have no more than wq
work oers outstanding to employers at any given time, and each employer can accept no
more than eq work oers from workers at any given time, where the work oer quota wq and the acceptance quota eq can be any positive integers.5 Although highly simpli ed, these parametric speci cations will be seen in Section 4, below, to permit the study of a variety of labor market structures operating under dierent market power conditions. When wq exceeds 1, the workers can be interpreted as some type of information service provider (broker, consultant, ...) able to provide services to more than one employer at a time. 5
8
int main () f Init(); For (G = 1,...,GMax) f InitGen(); For (I = 1,...,IMax) f MatchTraders(); Trade(); g
UpdateExp();
AssessFitness(); Output(); EvolveGen(); g g
// Construct initial subpopulations of pure workers, // worker-employers, and pure employers with // random worksite strategies // Enter the generation cycle loop. // Generation Cycle: // Con gure all agents with user-supplied // parameter values (initial expected utility // levels, work oer/acceptance quotas,...). // Enter the trade cycle loop. // Trade Cycle: // Determine worksite partners, // given expected utilities, // and record refusal and // wall ower payos. // Engage in worksite interactions // and record worksite payos. // Update expected utilities // using newly recorded payos. // Environmental Step: // Assess agent tness. // Output agent information. // Evolution Step: // Separately evolve the worksite strategies of pure // workers, worker-employers, and pure employers.
Return 0;
Table 1: Logical Flow of the ACE Labor Market As outlined in Table 1, each agent in the initial generation is constructed and assigned a random strategy governing worksite interactions. The agents then enter into a nested pair of generation cycle and trade cycle loops during which they repeatedly determine contractual partnerships, engage in worksite interactions, update their expected utility assessments for worksite partners based on newly recorded payos, and evolve their worksite strategies over time. The ACE labor market facilitates the study of labor markets from an agent-based perspective in two key ways. First, as depicted in Table 2, each agent is instantiated as an autonomous, endogenously-interacting software agent with internally stored state information and with internal behavioral rules. The agents can therefore engage in anticipatory behavior. Moreover, using stored agent addresses together with internalized communication protocols, they can communicate with each other at event-triggered times, a feature not present in standard economic models. Second, as seen in Table 1, the ACE labor market is modular in design. This means that 9
class Agent f
g;
Internal State Information: My physiological attributes; My endowments; My beliefs/preferences; Addresses for other agents; Additional data about other agents. Internal Behavioral Rules: Rules for communicating with other agents; Rules for gathering, processing, and updating information; Rules for determining my contractual partners; Rules for conducting my worksite interactions; Rules for updating my beliefs/preferences; Rules for calculating my tness; Rules for altering my rules.
Table 2: Schematic Description of an Agent experimentation with alternative speci cations for market structure, search and matching among workers and employers, worksite interactions, expectation formation and updating, and evolution of worksite strategies can easily be undertaken | much like changing a lightbulb in a multi-bulb lamp | as long as the interfaces (inputs and outputs) for the modules implementing these speci cations remain unchanged. Moreover, each of these modules can potentially be grounded in agent-initiated actions in the sense that the module is implemented via behavioral rules internal to the agents. Finally, the transitory and longer-run implications of each alternative module speci cation can be studied at three dierent levels: individual characteristics of workers and employers; interactions among workers and employers (network formation); and social welfare as measured by descriptive statistics such as average agent earnings and unemployment rates. A brief description will now be given for the particular module speci cations used in all experiments reported below. See McFadzean and Tesfatsion (1997) for a more careful description. If an employer accepts a work oer from a worker in any given trade cycle, the worker and employer are said to be matched for that trade cycle. Each match constitutes a mutually agreed upon contract stating that the worker shall be employed at the worksite of the employer until the beginning of the next trade cycle. These contracts are risky in that outcomes are not assured. 10
c
Employer d
c
(C,C)
(L,H)
d
(H,L)
(D,D)
Worker
Table 3: Payo Matrix for the Worksite Prisoner's Dilemma Game Speci cally, each matched worker and employer engage in a worksite interaction modelled as a two-person prisoner's dilemma game re ecting the basic eciency wage hypothesis that worker eort levels are aected by overall working conditions (e.g., wage levels, respectful treatment, safety considerations, ...). The worker can either cooperate (exert high work eort) or defect (engage in shirking). Similarly, the employer can either cooperate (provide good working conditions) or defect (provide substandard working conditions). The range of possible payos is assumed to be the same for each match in each trade cycle: namely, as seen in Table 3, a cooperator whose contractual partner defects receives the lowest possible payo L (the sucker payo); a defector whose contractual partner also defects receives a payo D; a cooperator whose contractual partner also cooperates receives a payo C ; and a defector whose contractual partner cooperates receives the highest possible payo H (the temptation payo). The payos are assumed to be measured in utility terms and to be normalized about 0, so that L < D < 0 < C < H . They are also assumed to satisfy the usual regularity condition (L + H )=2