J Prod Anal (2017) 47:33–47 DOI 10.1007/s11123-016-0489-8

Who benefits from job placement services? A two-sided analysis German Blanco

1

Published online: 10 January 2017 © Springer Science+Business Media New York 2016

Abstract In light of additional information market agents would achieve better outcomes, for example, a lower ask price for the buyer and a higher offer price for the seller. I examine this notion in a labor market, where employers and employees do not possess perfect information about wages, and address the question of who benefits from the information provided by job placement services? The empirical strategy considers the two-sided nature of the labor market. Estimates of employee and employer incomplete information are contrasted between users and non-users of placement services provided by Job Corps, America’s largest and most important job training program for youths. Findings suggest that employees that use placement services don’t have more information about better offer wages, relative to non-users. Interestingly, firms that employed users of placement services are better informed about reservation wages relative to firms that employed non-users. Keywords Two-tiered stochastic frontier analysis Job placement services Informational benefits ●

●

JEL classification C15 D83 J30 J20 ●

●

●

1 Introduction A job placement agency is, fundamentally, a provider of information in the labor market. The question that I address

* German Blanco [email protected] 1

Department of Economics, Illinois State University, Normal, IL 61790-4200, USA

is, who benefits from this information? In a labor market with incomplete information, the seller of labor (i.e., the employee) could use the informational advantage provided by the employment agency to achieve better labor market outcomes, e.g., selling labor at a higher price. Naturally, the other potential beneficiary would be the actual buyer of labor (i.e., the employer), who could use information provided by the agency in order to buy labor at a lower price. Consistent with search theory, I analyze these notions by using the variance of labor market prices to measure incomplete information about wages. In doing so, I determine how much of this informational gap is filled with information provided by an employment agency to both the employee and the employer. By and large, empirical studies evaluating placement services focus on measuring success based on the agency’s ability to offer vacancies to job seekers and candidates to employers and the rate at which contracts between both sides are achieved. For example, Black et al. (2003) found that the Worker Profiling and Reemployment Services system reduces the mean weeks of Unemployment Insurance benefit receipt by about 2.2 weeks, and increases subsequent earnings by about $1000. Their analysis suggests that the earning gains resulted primarily from earlier return to work in the treatment group (i.e., the group of participants in the Profiling and Reemployment Services system). While measuring success of placement services in this fashion is of extreme importance, it is not the only way. To my knowledge, one measure of success that has not been analyzed is the benefit that results from providing information about prices to the different parties in the labor market. The empirical analysis of any market with imperfect information should consider the fact that markets are not one-sided and the determination of prices is made jointly

34

between sellers and buyers (Butters 1977). Consistent with this fact, the analysis herein builds on Polachek and Yoon (1987) estimation strategy, the Two-tier Stochastic Frontier (2TSF hereafter). Within an earnings equation framework, they used the 2TSF to estimate two one-sided error terms, that represent the impact of information deficiencies that would lead workers to work for less than a maximum potential wage and employers to pay more than a minimum reservation wage, respectively. I apply this technique to two groups of individuals: users and non-users of job placement services. Importantly, I extend the technique to control for sample selection, which arises due to the endogeneity of choice to use job placement services, by building on the maximum simulated likelihood model proposed by Greene (2010). If indeed the placement services provided significant information about labor market prices to users, then the estimates of the impact of incomplete information would be significantly reduced, relative to the estimates for non-users. This result has important implications from the standpoint of productivity and efficiency in the matching process of labor markets.1 In other words, by providing information to the different parties in the labor market, matches will be generated at a price that is closer to the actual but unknown productivity of the match, improving in this way the quality of matches. Specially in markets characterized by less than perfect information, there may be scope for intermediaries to intercede and improve outcomes. Certainly, lack of information is a major impediment in the efficient pairing of workers and employers in labor markets (Mortensen and Pissarides 1994; Pries 2004), and a job placement agency may help in reducing this insufficiency. The employment agency being analyzed provides its services through the Job Corps program, America’s largest and most important job training program for youths. One of the program’s stated goals is to enhance participants’ labor market opportunities. Within this context, I analyze whether, in line with the program’s goal, the component of job placement services provides significant information to participants that would lead them to identify high-paying jobs. I also analyze whether placement services provide information to employers about Job Corps participants’ reservation wages. By shedding light on these questions, I contribute to the literature evaluating job placement services, an important component of federal employment programs for the disadvantaged in the U.S. and other advance economies (e.g., Black et al. 2003; Sianesi 2004; Winterhager et al. 2006; Autor and Houseman 2010; 1

The labor market search and matching literature is extremely vast. For a review, see Mortensen and Pissarides (1999) and Petrongolo and Pissarides (2001).

J Prod Anal (2017) 47:33–47

Autor et al. 2012). In addition, since my analysis considers both sides of the market being affected by the information provided by an intermediary, I also contribute to the literature on the effects of agent information on market outcomes (e.g., Rutherford et al. 2005; Levitt and Syverson 2008). Estimated results indicate that employees that used job placement services are not more informed about offer wages than employees that did not use the services, while results indicate that firms that employed individuals that used placement services are more informed about reservation wages relative to firms that employed non-users. On the one hand, this evidence is consistent with information distortions on the part of the job placement agency. In the present context, if the job placement agency is better informed about labor market prices than are potential employees and firms, it could have an incentive to exploit this advantage and create more matches between employees and firms in less time. Matches are created quickly if they are based on signaling reservation wages, which are closely related to the worker’s perceived productivity. On the other hand, while employers seem to be the ones benefiting, one has to keep in mind that the matching process is improved, and a wage that closely reflects true productivity of the match reduces the amount of inefficient contracts, which in turn would benefit workers. The rest of the paper is organized as follows. Section 2 presents a simple illustration about the role of information in a two-sided labor market, and how it affects outcomes. Section 3 describes the econometric approach used to estimate the 2TSF to model the role of information transmitted by the job placement agency and presents the extension to deal with potential sample selection. Section 4 describes the data and a discussion of the estimated results is presented in section 5. Section 6 concludes.

2 Modeling incomplete information in the labor market Stigler (1961, 1962), largely responsible for introducing the topic of search theory into economics, noted that in a market with homogeneous goods a unique price will exist only if either buyers or sellers have perfect information about prices. Subsequent studies, building on Stigler’s seminal idea, have derived results characterizing the effects of various changes on optimal-search behavior in markets with imperfect information about prices (see surveys in Rothschild 1973; Lippman and McCall 1976; Mortensen 1986). Recent literature shows that wage dispersion across observably identical workers can be an equilibrium outcome in labor markets with friction (for a review see Mortensen and Pissarides 1999). This recent literature regards search

J Prod Anal (2017) 47:33–47

35

friction as the time required for workers to gather information about wage offers.2 In contrast to earlier studies based on standard models of sequential wage search, the recent literature based on equilibrium models of labor markets explicitly accounts for and emphasizes the role of employers on the demand side of the labor market. What these different strands of the literature have in common is that they acknowledge that both sides of the market have less than perfect information, and thus, there may be scope for Labor Market Intermediaries (e.g., Job Placement Services) to intercede and improve labor market outcomes. As theorized by Stigler (1961), advertisement, which is a method of providing potential buyers (sellers) with knowledge about the identity of sellers (buyers), is expected to reduce the dispersion of prices since it reduces the cost of search. In the present context, within a labor market with imperfect information, an employment agency plays a role that is analogous to that of advertisement. Therefore, in the absence of job placement services one would expect the wage dispersion due to incomplete information about wages in the labor market to be larger. To illustrate, consider a potential employee that is willing to accept the following wage offer (Ws): W s ¼ W p  u;

ð1Þ

where Wp is the maximum potential wage offer and u ≥ 0. The potential wage Wp is only realized if the potential employee has perfect information, which will allow him/her to identify the firm offering such wage. The one-sided residual, u, represents a deviation from Wp due to having incomplete information. Similarly, a potential employer will be offering the following wage (Wb): W b ¼ W r þ w;

ð2Þ

where Wr is the minimum worker reservation wage and w ≥ 0. With perfect information the employer will be able to identify a qualified worker with the minimum reservation wage Wr. The one-sided residual, w, represents a deviation from Wr due to having incomplete information. Clearly, the information provided by a job placement agency should have an inverse relationship with the deviations u and w in (1) and (2), respectively. In other words, additional information will facilitate the identification process, but not to a perfect extent. The actual 2

An alternative branch of the search equilibrium literature is based on the so-called matching function, which is employed in a myriad of modern macroeconomics studies. This literature introduces two-sided frictions in the process of matching trading partners, where agents on both sides of a market make investments in overcoming them (for a review see Petrongolo and Pissarides 2001). Although my study is empirical and micro in nature, the idea of less than complete information in both sides of the market is taken into account.

magnitude and significance of the effects represented by u and w is an empirical question. In order to examine this question, one should focus on the observed equilibrium wage Wo, which is obtained by equating Ws = Wb ≡ Wo shown below: W o ¼ W ðXs ; Xb Þ þ v  u þ w;

ð3Þ

where W(⋅) is a function of exogenous supply and demand factors, Xs and Xb, respectively; and v is the usual random component.3 Determinants in Xs include individual characteristics that enhance human capital as well as other characteristics. Determinants in Xb include firm characteristics along with other market characteristics (e.g., market size). More detail about these determinants is relegated to the data section. Aside from the random noise v, it is evident from (3) that the wage dispersion, v − u + w, encompasses the effects of incomplete information both on the part of employees and employers. There is an important linkage between the wage determination in (3), productivity and information. The wage W(⋅) represents the productivity of the worker-firm match based on observable characteristics, denoted Xs and Xb, and its expected value is used to learn about the true, but unknown, productivity of the match. As long as the implicit participation condition holds, i.e., matches are created when Wr < Wo < Wp, deviations from the productive outcome will be reduced in the presence of more information. This result has important implications for the efficient functioning of the labor market’s sorting process, since more quality matches, that closely reflect true productivity, will be created in the presence of additional information. Theoretical work dating back to the 1970s (e.g., Butters 1977) have stressed the importance of conducting empirical work considering the two-sided nature of markets. Recent theoretical work on two-sided markets is abundant and explicitly models interactions between agents via intermediaries (e.g., Masters 1999; Bloch and Ryder 2000; Armstrong 2006). The next section presents and extends an econometric technique that takes into account the two-sided nature of labor markets. The technique is used to obtain estimates of the unobserved components of the wage 3

In a typical labor market, the observed equilibrium wage is derived after equating the quantity of labor demand and supply. A formal derivation yielding the observed equilibrium as depicted in (3) can be found in Polachek and Yoon (1987). Consistently, in my simplified illustration above, I add u and subtract w from the equality Wp−u = Wo = Wr + w, which yields Wp−w = Wo + u − w = Wr + u. The expression Wo + u − w represents a full information labor market price, which is equal to the function W(Xs, Xb) + v. It follows that under incomplete information then Wo = W(Xs, Xb) + v − u + w. In a reduced form version, models based on bargaining and search (see Mortensen 1986) would yield a similar wage specification as in (3) after allowing for interactions between supply and demand. The latter, for example, was used by Kumbhakar and Parmeter (2009).

36

J Prod Anal (2017) 47:33–47

dispersion in (3), and these estimates are then used to evaluate the role of job placement services as a provider of information about prices in the labor market.

3 The two-tiered earnings frontier Consider the following parametric formulation of Eq. (3): yi ¼ x′i β þ ϵi ;

ð4Þ

where yi is the observed market wage for observation i when the set of supply and demand determinants is xi, β is the vector of parameters attached to this set of determinants, and ϵi, which measures wage dispersion, is a composite error term such that ϵi ¼ vi  ui þ wi ;

ð5Þ

with vi representing a random error, the non-negative components ui and wi measuring the extent of employee and employer incomplete information, respectively.4 The main focus of the paper lies on obtaining estimates for the second and third components of ϵi, which is accomplished via maximum likelihood estimation of the system represented by (4) and (5). Polachek and Yoon (1987) derived an empirically tractable density function for ϵi by imposing the following distributional assumptions about its components: (i) vi is normally distributed with mean and variance ð0; σ 2v Þ, (ii) ui is distributed exponentially with mean and variance ðμu ; σ 2u Þ, and (iii) wi is distributed exponentially with mean and variance ðμw ; σ 2w Þ; in addition to these distributional assumptions, one has to assume that the error components are independent of one another and uncorrelated with the determinants in xi. Based on these assumptions, Polachek and Yoon’s (1987) Two-tier Stochastic Frontier (2TSF) is based on the density of ϵi,
 n
h i σ 2v ϵi ϵi σv 1  exp þ þ f ðϵi Þ ¼  Φ  2 μu þμw μu σv μu 2σ u
 ϵi σv þ Φ σv  μ  w
h

io 1 1 1 ; exp  2 2ϵi μ þ μ þ σ 2v σ12  σ12 u

w

u

w

ð6Þ where Φ(⋅) is the standard normal cumulative distribution function. Estimates of parameters in (4) and (5) for a sample of n observations are obtained by maximizing the likelihood function 4 Hofler and Polachek (1985) and Polachek and Robst (1998) estimated a model of employee ignorance based on Eq. (4) with wi = 0 (i.e., employers have perfect information) by using the conventional stochastic frontier analysis in Aigner et al. (1977). Polachek and Robst (1998) further extended their one-sided analysis by using the technique in Jondrow et al. (1982), to estimate individual specific measures of employee incomplete information.

Q Q with Lðyi jβ; σ v ; μu ; μw Þ ¼ ni¼1 f ðϵi Þ ¼ ni¼1 f ðyi  x′i βÞ, the density f (⋅) as in (6). The assumptions above are needed to obtain a closedform of the likelihood function and tractability of the parameters of interest; unfortunately, the assumptions are untestable. In the present context, however, is not unreasonable to assume that the error components are independent of one another since they pertain to the stock of information about prices that firms and employees have, i.e., the two agents accumulate information independent from each other.5 Regarding the distributional assumptions, these are commonly employed in standard single-tiered stochastic frontier analysis when using maximum likelihood (Kumbhakar and Lovell 2003). Studies that employed the 2TSF relied on the same set of assumptions. For example, to analyze relative levels of information in labor markets, the 2TSF was employed in Polachek and Yoon (1987, 1996), Groot and Oosterbeek (1994), Sharif and Dar (2007); it was used to measure the effects of wage bargaining under production uncertainty in Kumbhakar and Parmeter (2009). The 2TSF has also been used in other settings outside of the labor market, for example, Kumbhakar and Parmeter (2010) used it in a hedonic price model to analyze levels of information in the house market; Gaynor and Polachek (1994) applied and extended the technique to explain variation in the price of physician services; other applications analyzing information asymmetries in the Health Service market include Chawla (2002) and Tomini et al. (2012); also, Ferona and Tsionas (2012) used the 2TSF to analyze under- and over-biding in an auction setting. An alternative that would yield a closed-form solution for the likelihood function was recently proposed by Papadopoulos (2014), where the exponential is replaced by the half-normal distribution to model the one-sided components. Consistent with search theory, the studies above have found that the market price is affected by the relative levels of information that sellers and buyers possess. However, one can be skeptical about the actual role of information given that their estimates are not based on actual measures of information, but are inferred from price data alone, and thus their measures of incomplete information effects can be biased due to unobserved employee and employer heterogeneity. In the present context, I model informational advantage explicitly by defining two groups, that is, one that used job placement services and one that did not, and measure their relative levels of information about labor market prices by employing the 2TSF. With the aim of contributing to the growing literature using this technique, 5

Note that the illustration in the preceding section suggests that u and w are not independent on each other, however, identification is only possible if one employs this assumption. As pointed out by one anonymous referee, in general this is an inherent identification problem of the 2TSF formulation.

J Prod Anal (2017) 47:33–47

37

I now illustrate the process for obtaining observation specific estimates and later present an extension to control and test for sample selection in the 2TSF context. 3.1 Observation specific estimates Kumbhakar and Parmeter (2009) extended the work by Jondrow et al. (1982) to obtain observation specific estimates of the error components ui and wi in the 2TSF. I employ Kumbhakar and Parmeter (2009) expressions for Eðeui jϵi Þ and Eðewi jϵi Þ, which are obtained upon deriving the conditional distributions f(ui|ϵi) and f(wi|ϵi) when the dependent variable is in logarithm form. To avoid notational clutter let ai ¼

σ 2v 2σ 2w

λ¼

1 μu

χ 2i ¼

di ¼ Iðm′i γ þ ϑi > 0Þ;

σ2

 μϵi ; bi ¼ σϵvi  μσ v ; αi ¼ μϵi  2σv2 ; Bi ¼ ðσϵvi  μσ v Þ w

w

u

u

selection in their analysis of technical efficiency for organic and conventional farms. In a recent study, Greene (2010) showed that the familiar approach in which the IMR is added to a nonlinear model such as the stochastic frontier is not appropriate. Furthermore, Kumbhakar et al. (2009) noted that when inefficiency is present no two step approach will work.6 Greene (2010) proposed a method based on maximum simulated likelihood estimation, where the selection model developed by Heckman (1976) is modified to correct for sample selection in the stochastic frontier model by Aigner et al. (1977). Here, I build on Greene (2010) methodology and extend Polachek and Yoon’s (1987) 2TSF to account for sample selection. First, let’s consider the selection equation for using job placement services (di) as:

u

þ μ1 ; χ 1i ¼ Φðbi Þ þ expðαi  ai ÞΦðBi Þ; and w

ϕðBi Þ þ expðai  αi ÞΦðbi Þ then;


2  h i σ λ 1 Eðeui jϵi Þ ¼ 1þλ Φðbi Þ þ expðαi  ai Þexp 2v  σ v Bi ΦðBi  σ v Þ χ 2i

ð7Þ
2  h i σ λ 1 ΦðBi Þ þ expðai  αi Þexp 2v  σ v bi Φðbi  σ v Þ Eðewi jϵi Þ ¼ 1þλ χ 1i

ð8Þ

where Eðeui jϵi Þ and Eðewi jϵi Þ are the measures of efficiency in market wages attributed to employee and employer’s information levels, i.e., the conditional expectations in (7) and (8) are used as the point estimators for eui and ewi , respectively.

ð9Þ

where I(A) is an indicator function equal to 1 if condition A is true and 0 otherwise; mi = [xi zi]′, with xi as in (4) and zi is a vector of identifying instruments, which are variables correlated with di and uncorrelated with ϑi, with no direct impact on yi and uncorrelated with ϵi in (4). Further, assume that ϑi is distributed standard normal, N(0, 1). Sample selection arises if there exists unobservables correlated with both vi and ϑi, from (5) and (9). Formally, selection in the present context assumes that (ϑi, vi) are distributed N2(0, Σ), where the off-diagonal elements of Σ are equal to ρσv.7 Under the conditions above, the conditional density of y given xi, zi, di, and the one-sided errors of interest ui and wi is ( f ðyi jxi ; zi ; di ; ui ; wi Þ

¼ di

 Φ

3.2 The two-tiered earnings frontier with correction for sample selection Sample selection is a well-known and commonly found problem in applied econometrics. In the present context, estimates of the error components in (5) will be compared across users and non-users of job placement services, and thus, selection bias will arise due to the endogenous decision of taking advantage of the informational services provided by the job placement agency. Applications employing single-tiered frontier analysis rely on Heckman’s (1979) two-stage procedure to control for sample selection. For example, Bradford et al. (2000) evaluated cost efficiency of two technologies used to treat coronary artery disease. To control for sample selection, due to patients self selecting into a given treatment, the authors included the socalled inverse Mills’ ratio (IMR) into the single-tiered stochastic frontier model. The same technique was employed by Sipiläinen and Oude Lansink (2005) to deal with sample

expð0:5ðyi x′i βþui wi Þ2 =σ 2v Þ pffiffiffiffi σ v 2π ρðyi x′i βþui wi Þ=σ ε þm′i γ

´

)

pffiffiffiffiffiffiffiffi 1ρ2

þð1  di ÞΦðm′i γÞ: ð10Þ Note that without ui and wi, the conditional density in (10) is the same as that used in the full information maximum likelihood estimator developed by Heckman (1976). 6

An earlier version of the current paper considered a two-stage residual inclusion technique, which is a consistent instrumental variable-based approach for correcting endogeneity in non-linear models estimated via non-linear least squares (Terza et al. 2008). 7 Other approaches for dealing with sample selection in a stochastic frontier model includes Kumbhakar et al. (2009) and Lai (2015). The former approach is similar to the one proposed by Greene (2010), the difference is that the selection equation is also affected by inefficiency. Lai (2015) extends the model in Greene (2010) by replacing the halfnormal distribution for a truncated normal and also considers endogenous switching. While these could be potentially extended and employed in the present context, extending the model in Greene (2010) for the 2TSF is the more straightforward approach. For a recent review of sample selection in stochastic frontier models see Parmeter and Kumbhakar (2014).

38

J Prod Anal (2017) 47:33–47

At this point, the conditional density in (10) is not operational due to the unobserved inefficiency terms ui and wi. In an attempt to integrate out these unobserved variables one will not find a closed-form solution. Instead, I build on Greene (2010) and employ simulation techniques to approximate the integrals of interest. For practical purposes, it is convenient to use the following parameterizations: ui ¼ σ u jUi j

and

wi ¼ σ w jWi j;

where Ui and Wi are both distributed as a standard normal, N(0, 1). Then, the simulated log likelihood function is logLS ðβ; σ u ; σ v ; σ w ; γ; ρÞ ¼ N R h n P P expð0:5ðyi x′i βþσ u jUir jσ w jWir jÞ2 =σ 2v Þ pffiffiffiffi log R1 di ´ σ v 2π n¼1 r¼1   ρðyi x′i βþσ u jUir jσ w jWir jÞ=σ ε þm′i γ pffiffiffiffiffiffiffi2ffi Φ 1ρ  þð1  di ÞΦðm′i γÞ ;

n¼1

σ v 2π

r¼1



Φ



ρðyi x′i βþσ u jUir jσ w jWir jÞ=σ ε þm′i ^γ

pffiffiffiffiffiffiffiffi 1ρ2

3.2.1 Observation specific estimates To characterize efficiency, relative to the full information market wage level, I follow Greene (2010) approach and take advantage of the simulation values of ui and wi during the estimation. Using Bayes’ rule, one can write pðui jϵi Þ

pðϵi jui Þpðui Þ pðεi Þ

¼

pðϵi jwi Þpðwi Þ : pðϵi Þ

pðwi jϵi Þ ¼ ð11Þ

After replacing the parametrizations ui = σu|Ui| and wi = σw|Wi| above, the expectations of interest are R

σ u jUi j ðσ u jUi jÞp½ϵi jðσ u jUi jÞpðσ u jUi jÞdðσ u jUi jÞ

E½ðσ u jUi jÞjϵi  ¼

where Uir and Wir are R random draws from the standard normal distribution. The maximum simulated likelihood estimators of the model are obtained by maximizing (11) with respect to the unknown parameters. Note that this estimation is only slightly more complicated than the single tiered stochastic frontier model proposed by Greene (2010) since it involves the extra inefficiency parameter σw. Estimation is simplified by using a two-step approach, a procedure analogous to the proposed by Greene (2010), where I use the first step probit estimate ^γ instead of γ in (11). To account for the use of the predicted regressor ^γ , the Murphy and Topel (2002) correction for standard errors needs to be implemented. With this simplification, units with di = 0 do not have any of the parameters to be estimated in the second step via maximum simulated likelihood, resulting in the simplified log likelihood logLS ðβ; σ u ; σ v ; σ w ; ρÞ ¼ N R P nexpð0:5ðyi x′i βþσ u jUir jσ w jWir jÞ2 =σ 2v Þ P pffiffiffiffi log R1 ´

when ρ = 0, the model in (12) reduces to a maximum simulated likelihood estimator of the 2TSF without selection and half-normal one sided components. Therefore, a simulated likelihood ratio test can be used to test the selection model specification.8

ð13Þ

and

R

E½ðσ w jWi jÞjϵi  ¼

:

Estimated parameters are obtained using a conventional gradient approach, e.g., the Broyden-Fletcher-GoldfarbShanno (BFGS) algorithm. Olsen (1978) transformation will make estimation of the parameters vastly simpler. Rather than using pseudo-random draws to generate the R pairs of Uir and Wir, estimates are computed using Halton draws. In doing so, one will be able to replicate results, in addition, Halton draws are relatively more effective since fewer are required, resulting in lower computing time (for more details see, for example, Greene 2011; Gourieroux and Monfort 1996; Train 2003). It is important to note that

σ w jWi j ðσ w jWi jÞp½ϵi jðσ w jWi jÞpðσ w jWi jÞdðσ w jWi jÞ

pðϵi Þ

:

ð14Þ

The simulated denominator in expressions (13) and (14) is ^i B

¼

(

R P 1 R

r¼1

 Φ ¼

ð12Þ

pðϵi Þ

1 R

R P

^ σ u jUir j^ expð0:5ðyi x′i βþ^ σ w jWir jÞ2 =^ σv2 Þ pffiffiffiffi σ^v 2π

)

^ σ u jUir j^ ^ ρðyi x′i βþ^ σ w jWir jÞ=^ σ ϵ þm′i ^γ

pffiffiffiffiffiffiffi2ffi 1^ ρ

´ ð15Þ

^f ir

r¼1

while the simulated numerators are R X ^U ¼ 1 A ð^ σ u jUir jÞ^f ir i R r¼1

ð16Þ

and R X ^W ¼ 1 A ð^ σ w jWir jÞ^f ir : i R r¼1

ð17Þ

The estimates of E[(σu|Ui|)|ϵi] and E[(σw|Wi|)|ϵi] are given ^ U =B ^ W =B ^ i and A ^ i , which are computed using the estiby A i i mated parameters, the raw data and the previously used Halton draws. 8

Greene (2010) used the same type of test in the context of a single tiered stochastic frontier model.

J Prod Anal (2017) 47:33–47

4 Data, Job Corps and job placement services I analyze effects of the information provided by job placement services. These services are an important component of the Job Corps (JC) program, America’s largest and most comprehensive education and job training program for youths. This federally funded program is currently administered by the US Department of Labor and it has yearly cost of about $1.5 billion, with an annual enrollment that ascends to 60,000 students (US Department of Labor 2012). The program’s goal is to help disadvantaged young people, ages 16 to 24, improve the quality of their lives by enhancing their labor market opportunities and educational skills set. Eligible participants receive academic, vocational, and social skills training at over 122 centers nationwide, where they typically reside. In addition, participants are offered the opportunity to take advantage of job placement services. A question that has not been addressed within this important job training program is whether, in line with the program’s goal, the component of job placement services provides significant information to participants that lead them to identify high-paying jobs.9 In addition, I analyze a second unexplored question, that is, whether the job placement services provide information to employers about JC participants’ reservation wages. To address these questions, the frontier analysis, described in the previous section, is applied to two groups of employed JC participants. One group is comprised of participants that used job placement services and the second group is comprised of participants that did not use the placement services (users and non-users hereafter). The data is from the National Job Corps Study (NJCS), a randomized evaluation of the JC program. The NJCS assigned eligible applicants randomly to treatment and control groups.10 The sample for the present study is restricted to individuals that are in the treatment group and are employed at the time of the follow up interview conducted one year after random assignment. The focus is on these individuals since the control group was banned from using any of JC’s services, including job placement.11 The second restriction is driven by data limitations. A binary 9

This question is not to be confused with questions about the overall effect of program participation on the probability of employment, which has been reported to be positive, with a magnitude of 4 percentage points (Schochet et al. 2001). 10 Eligibility is based on several criteria, including age, legal US residency, economically disadvantage status, living in a disruptive environment, and in need of additional education or training, among others (see Schochet et al. 2001). From a randomly selected research sample of 15,386 first time eligible applicants, approximately 61 percent were assigned to the treatment group (9409) and 39 percent to the control group (5977). 11 Another option is including control group individuals categorized as non-users, however, one has to be cautious since it is not possible to learn from the data whether they used job placement services outside of JC.

39

indicator recorded information on the usage of job placement services, i.e., equals to 1 if the individual used the services or 0 otherwise. Information about the exact time when the service was provided is not available. I estimate the frontier using earnings recorded one year after randomization since about 80 percent of individuals had finished their actual JC enrollment within that first year, and placement services can only be used within six months after leaving the program (Schochet et al. 2003). Observed wages are a function of exogenous supply and demand determinants (in Eq. (3), Xs and Xb, respectively). Starting with the supply side determinants, I include variables that augment the stock of human capital such as education, as well as other important demographic characteristics. These variables were recorded at baseline, i.e., when randomization took place. Summary statistics of these demographic characteristics for the sample of users and non-users are presented in the upper half of Table 1. I report means and standard deviations for each group to assess the degree of heterogeneity, as well as the difference in means with the associated t-value. Of the demographic variables included there are a few that differ statistically (5 percent level). Average years of education and age for the group of nonusers is 10.42 and 19.33, respectively, so this group is slightly more educated and older than users (t-value for the difference in means is −2.12 and −3.28, respectively). In both groups, females represent 44 percent, Blacks are predominant, about 45 percent, followed by Whites with 32 percent. The proportion of Hispanics differs statistically between users and non-users, where the difference is 3 percentage points (t-value is −2.44). Other proportions that differ statistically are head of household and has a child, with the group of non-users having slightly larger proportions. Finally, non-users also have, on average, more educated fathers and more children, where these differences are statistically significant, albeit small. Now we turn our attention to the demand side determinants. Identification of the actual firms employing individuals in the sample is not possible, hence controls for firm characteristics are proxied with individuals’ occupation. In addition, I include two indicators of metropolitan area classification. The first indicator is for Metropolitan Statistical Areas (MSA), which consists of one or more counties that contain a city with a population larger than 50,000. The second indicator is for Principal Metropolitan Statistical Areas (PMSA), which are areas with a population larger than one million. These indicators are useful to characterize the labor market in terms of its size, which is an important demand side determinant of any market.12 Summary 12

Ideally, this two-sided analysis would benefit from including actual local economic condition indicators, which are important demand side determinants (Hoynes 2000). However, such measures are not available.

40

J Prod Anal (2017) 47:33–47

Table 1 Summary statistics, supply and demand side determinants of earnings

Users Mean

Non-users SD

Mean

Users—non-users SD

Difference

t-value

Demographics Years of education

10.297

1.529

10.424

1.562

−0.126

−2.119

Age

−3.282

19.056

2.179

19.332

2.195

−0.277

Female

0.444

0.497

0.444

0.497

0.000

0.008

White

0.322

0.467

0.324

0.468

−0.003

−0.139

Black

0.484

0.500

0.447

0.497

0.037

1.905

Hispanic

0.134

0.341

0.167

0.373

−0.034

−2.438

Other race

0.060

0.238

0.061

0.239

−0.001

−0.056

Head of household

0.117

0.321

0.150

0.357

−0.033

−2.516

Married

0.022

0.148

0.030

0.170

−0.007

−1.198

Has a child

0.188

0.390

0.227

0.418

−0.039

−2.514

Number of children

0.259

0.615

0.322

0.684

−0.063

−2.508

Mothers education

11.719

2.097

11.562

2.283

0.157

1.858

Fathers education

11.660

2.171

11.445

2.299

0.215

2.494

Occupation at baseline Service

0.222

0.415

0.221

0.415

0.000

0.005

Laborer and construction

0.141

0.348

0.138

0.345

0.003

0.261

Sales

0.127

0.333

0.132

0.338

−0.005

−0.364

Private household

0.059

0.235

0.050

0.219

0.008

0.956

Mechanics

0.045

0.207

0.061

0.239

−0.016

−1.875

Clerical

0.042

0.200

0.055

0.228

−0.013

−1.597

Agriculture

0.036

0.187

0.031

0.174

0.005

0.764

Manufacturing

0.018

0.132

0.016

0.127

0.002

0.310

Other

0.311

0.463

0.296

0.456

0.015

0.860

Metropolitan Statistical Area (MSA)

0.484

0.500

0.491

0.500

−0.007

−0.368

Principal MSA

0.298

0.458

0.296

0.456

0.003

0.147

197.149 145.265 −33.248

−9.390

Market size

Outcome Weekly earnings Number of observations

164.997 131.438 1344

1350

SD standard deviation

statistics for baseline occupations and market size indicators are presented in the lower half of Table 1. Differences between users and non-users in the proportion by occupation are not statistically significant (5 percent level). At baseline, the majority of users and non-users belonged in the service sector (about 22 percent). Other important occupations are Laborer and Construction with about 14 percent, and Sales with about 13 percent. Similarly, no significant differences arise when comparing the proportion of individuals living in MSA and PMSA between users and non-users. For both groups, at baseline, about 49 percent of individuals are located in MSA and about 30 percent are in PMSA. At the bottom of the lower half in Table 1, I present summary statistics for the outcome of interest, average weekly earnings measured 4 quarters after random

assignment. The mean of weekly earnings for users is $164.997 while the mean for non-users is $197.149. The $33.248 difference is statistically significant at a 5 percent level. A naive examination about how information affects the variance of prices can be done by looking at the standard deviations, where the lower variance in earnings for users is consistent with the notion that placement services are providers of information in the labor market. Results from a formal empirical examination is presented in the next section.

5 Results As a point of departure, Table 2 reports linear regression results for the outcome, weekly earnings, in logs. Covariates

J Prod Anal (2017) 47:33–47

41

Table 2 OLS regression of log earnings on supply and demand determinants

Constant Years of education Years of education2 Age Female White Black Hispanic Head of household Married Has a child Number of children Mothers education Fathers education Occupation at baseline Service Laborer and construction Sales Private household Mechanics Clerical Agriculture Manufacturing Market size Metropolitan Statistical Area (MSA) Principal MSA Note: Standard errors in parentheses

Users

Non-users

3.327 (0.596) −0.015 (0.178) 0.001 (0.021) 0.063 (0.011) −0.198 (0.041) 0.162 (0.079) −0.016 (0.078) 0.069 (0.086) 0.065 (0.058) 0.195 (0.114) 0.144 (0.089) −0.086 (0.056) 0.009 (0.009) 0.002 (0.009)

2.736 (0.636) 0.052 (0.106) −0.001 (0.005) 0.081 (0.015) −0.197 (0.058) 0.042 (0.111) −0.114 (0.110) −0.034 (0.119) 0.109 (0.077) 0.085 (0.151) 0.025 (0.120) −0.056 (0.074) 0.012 (0.012) 0.001 (0.012)

0.078 (0.050) 0.231 (0.059) 0.196 (0.061) −0.032 (0.084) 0.264 (0.085) 0.271 (0.088) 0.328 (0.103) 0.125 (0.141)

0.147 (0.071) 0.245 (0.083) 0.290 (0.085) 0.030 (0.122) 0.295 (0.112) 0.198 (0.117) 0.439 (0.150) 0.112 (0.201)

include the set of individual characteristics, controls for JC participants’ occupation and the size of their local market. I find many consistencies with the large literature adopting Mincer type earnings functions, including: the estimated coefficient for age is positive and significant, suggesting that a worker who is one year older has average earnings which are between 6.3 and 8.1 percent higher; females earn significantly less than males, where the gap is around 20 percent; whites earn more relative to the race group other (although these differences are significant for users only), blacks and Hispanics seem to be earning less than whites. Other consistent results are that the variables head of household, married, has a child and parents’ education are positively related to earnings, whereas number of children is negatively related. One inconsistent result is that the coefficient on education is positively related to earnings only for non-users, but statistically insignificant for both users and non-users.13 Regarding the rest of the variables in Table 2, the majority of the included occupations are related to higher earnings relative to occupations classified as other. For example, for users and non-users the difference, relative to occupation other, is larger and significant for the occupation of mechanics, clerical, and laborer and construction, while the difference is small and insignificant for the private household occupation. Finally, individuals living in highly populated metropolitan areas (MSA and PMSA) earn relatively more. We now turn onto the estimation of the effects of incomplete information. Results from the two-tiered earnings frontier model, estimated via maximum likelihood, are presented in Table 3. For brevity, and given that the main focus of the paper is on the estimates of the error components, I omit the discussion of the estimated coefficients in the deterministic part of the frontier.14 Based on the bootstrapped standard errors reported in parentheses for the estimates of μu and μw, i.e., the means of the error components of interest, I find that labor market prices are significantly affected by the relative levels of information that employees and employers have. In addition, the amount of employer’s (buyers) incomplete information is relatively smaller than that of employees (sellers). The latter is evident since the estimated magnitudes for μu are larger than μw, a result that holds for users and non-users. Interestingly, results show a larger (in magnitude) difference between μw’s 13

0.092 (0.047) 0.071 (0.053)

0.111 (0.066) 0.094 (0.074)

The negative coefficient on education of users becomes positive in regression without the square term for education, however it would still be insignificant. 14 Most of the estimated parameters in the deterministic portion of the frontier are qualitatively similar to the OLS results reported in Table 2. It is worth noting that one striking difference, relative to the regression results, is that the estimated coefficient on education is now positive for users, however it remained statistically insignificant.

42

J Prod Anal (2017) 47:33–47

Table 3 Two-tiered earnings frontier estimates

Demographic characteristics Constant Years of education Years of education2 Age Female White Black Hispanic Head of household Married Has a child Number of children Mothers education Fathers education Occupation at baseline Service Laborer and construction Sales Private household Mechanics Clerical Agriculture Manufacturing Market size Metropolitan Statistical Area (MSA) Principal MSA

Table 3 continued

Users

Non-users

4.462 (0.724) 0.110 (0.209) −0.020 (0.025) 0.030 (0.012) −0.171 (0.047) 0.212 (0.093) 0.030 (0.091) 0.060 (0.104) −0.006 (0.069) 0.162 (0.137) 0.157 (0.108) −0.038 (0.071) 0.003 (0.011) 0.005 (0.010)

3.351 (0.410) 0.093 (0.114) −0.001 (0.015) 0.061 (0.011) −0.171 (0.040) 0.095 (0.080) −0.103 (0.078) −0.008 (0.084) 0.127 (0.052) 0.077 (0.095) −0.074 (0.082) 0.016 (0.050) 0.022 (0.008) 0.011 (0.008)

0.010 (0.058) 0.213 (0.068) 0.051 (0.070) 0.004 (0.098) 0.184 (0.103) 0.227 (0.104) 0.167 (0.113) 0.071 (0.155)

0.059 (0.050) 0.142 (0.057) 0.135 (0.057) −0.002 (0.091) 0.174 (0.073) 0.054 (0.080) 0.333 (0.107) 0.054 (0.125)

0.051 (0.055) 0.072 (0.062)

0.032 (0.046) 0.076 (0.052)

Error components Random component Employee incomplete information Employer incomplete information

Users

Non-users

0.404 (0.097) 0.836 (0.046) 0.202 (0.071)

0.162 (0.113) 0.879 (0.030) 0.290 (0.042)

Note: Bootstrapped standard errors in parentheses based on 500 replicates

Table 4 Observation specific estimates of the effects of information on earnings efficiency Users Non-users Difference Confidence interval Estimate of Eðeui jεi Þ Mean

0.541 0.531

0.010

(−.022, .050)

25th percentile

0.369 0.315

0.054

(−.007, .185)

Median

0.610 0.595

0.016

(−.068, .066)

75th percentile

0.733 0.775

−0.042

(−.084, .026)

0.832 0.775

0.057

(.003, .158)

Estimate of Eðewi jεi Þ Mean th

25 percentile

0.824 0.784

0.040

(−.008, .115)

Median

0.850 0.820

0.030

(−.006, .115)

75th percentile

0.860 0.821

0.039

(.003, .117)

Note: Confidence intervals based on biased corrected bootstrap, 500 replicates

across users and non-users, where the differential of −0.088 is consistent with firms hiring users of placement services having more information. This difference, however, is not significant based on the upper 95% confidence interval of 0.008.15 The difference between users and non-users on the supply side is relatively smaller, −.0423, and not significant. A more meaningful characterization of the effects of information on earnings is summarize in Table 4 and a visual display is presented in Fig. 1. Table 4 presents the observation specific levels of efficiency, relative to the full information labor market price, based on Eqs. (7) and (8) for the mean, median, the 25th and 75th percentiles, while Fig. 1 depicts the kernel density estimates. At the mean and median, the stock of information acquired by employees users of placement services allows them to be about 1 percentage point closer to the more efficient outcome, 15

Confidence intervals for the difference were calculated with a biascorrected bootstrap technique (see Efron 1982; Kim et al. 2007). For an alternative, where confidence intervals are obtained based on the estimated conditional distributions in (7) and (8), see Horrace and Schmidt (1996).

J Prod Anal (2017) 47:33–47

43

1 0

.5

Density

1.5

2

Kernel density estimates, Workers

0

.2

.4

Users

.6

.8

Non-Users

20 0

10

Density

30

40

Kernel density estimates, Firms

.2

.4

.6

Users

.8

Non-Users

Fig. 1 Kernel density estimates of observation specific efficiency, based on maximum likelihood estimation of the two-tiered earnings frontier

relative to non-users. Placement services seems to boost the information stock at lower percentiles while the opposite is true for higher ones (this can also be seen upon close examination of Fig. 1). The 1 percentage point differential suggests that the informational advantage provided by job placement services within the Job Corps is relatively small and does not lead to better offer wages, which would have translated to a more efficient match. All of these differences are not distinguishable from zero, based on the 95% biascorrected bootstrap confidence interval. The average level of efficiency for employees in the sample is around 54%, which is low relative to other studies using 2TSF, but one has to keep in mind that the population being study here is comprised of disadvantage young JC graduates.16 Importantly, the lack of informational advantage result is 16 Previously reported levels of employees’ information range from 70 to 85 percent (Polachek and Yoon 1987, 1996; Groot and Oosterbeek 1994; Sharif and Dar 2007; Kumbhakar and Parmeter 2009). In general, the populations studied in these papers are comprised of more educated, older and experienced workers, hence relatively more informed than the population of JC participants.

consistent with findings in the literature evaluating placement services in other federally funded programs, where the tradeoff between quick employment and high wages is evident (e.g., Black et al. 2003).17 On the other hand, employers that hired users have an informational advantage throughout the distribution, where the effect at the mean suggests that they are about 6 percentage points more efficient, relative to firms that hired non-users. On average, this relative difference is associated to having information provided by the job placement services. Furthermore, the differential effect at the mean and the 75th percentile are significant based on the 95% biascorrected bootstrap confidence interval. In general, these results seem to suggest that job placement services played an important role in improving employers’ information about employees’ reservation wages, and as a result improving the efficiency of matches. Note that the average level of efficiency is about 80%. Other studies using 2TSF have reported similar mean effects of employers’ levels of information, with inefficiencies ranging from 20 to 44 percent (Polachek and Yoon 1987, 1996; Groot and Oosterbeek 1994; Sharif and Dar 2007; Kumbhakar and Parmeter 2009).18 Now, sample selection is addressed since individuals in the sample were not randomly assigned to be users and nonusers of job placement services. The component of JC placement services has unrestricted access to those enrolled (or that have graduated within 6 months), so they may choose to participate or not. Here I model this feature within the sample selection framework discussed in Section 3.2 Table 5 presents results from a first stage probit, used to estimate the second stage simulated likelihood in Eq. (12). The identification strategy employs an indicator of whether the individual was assigned to a non-residential slot in the JC center attended as an instrumental variable (zi). The nonresidential assignment was carried out by JC staff based on anticipated program enrollment of eligible applicants prior to baseline and, importantly, more than 98 percent of individuals complied with the assignment. In principle, assignment to a non-residential slot may disincentivize individuals from using certain types of services that are provided within the JC, including job placement. This seems to be the case since the parameter estimate for the Other recent studies within this literature finding similar results include: Autor and Houseman (2010) and Autor et al. (2012). Both analyzed the effect of Detroit’s welfare-to-work job placement on earnings and employment, and concluded that job placements with “direct-hire” employers raise earnings due to a single and continuous job spell. They also find evidence indicating that “temporary-help” job placements do not improve earnings. 18 Other studies not focusing on labor markets report similar mean effects of buyers’ incomplete information on prices, for example, Kumbhakar and Parmeter (2010) report that real estate buyers pay, on average, 30 percent more relative to a perfectly informed buyer. 17

44

J Prod Anal (2017) 47:33–47

Table 5 Probit regression of job placement services on selected covariates Estimated coefficient

t-value

Constant

−0.195

−0.570

Non-residential indicator

−0.223

−3.280

Years of education

−0.012

−0.530

Age

−0.008

−0.560

Female

0.308

0.720

White

0.158

1.150

Black

0.186

1.370

−0.037

−0.240

Hispanic Head of household

−0.076

−0.940

Married

−0.063

−0.260

Has a child

Table 6 Two-tiered earnings frontier estimates via maximum simulated likelihood

0.050

0.200

Number of children

−0.065

−0.350

Mothers education

0.009

0.730

Fathers education

0.022

1.830

−0.014

−0.160

Users Demographic characteristics Constant

3.693

2.599

(0.913)

(0.785)

Years of education

−0.018

0.031

(0.273)

(0.231)

Years of education2

−0.004

0.002

(0.032)

(0.027)

0.041

0.078

Age Female White Black

Occupation at baseline Service Laborer and construction Sales Private household

0.018

0.210

−0.028

−0.280

0.035

0.280

Mechanics

−0.185

−1.520

Clerical

−0.161

−1.310

Agriculture

0.100

0.700

Manufacturing

0.020

0.100

0.025

0.380

0.029

0.400

Market size Metropolitan Statistical Area (MSA) Principal MSA Number of observations

Hispanic Head of household Married Has a child Number of children Mothers education

2694

Note: Right hand side variables include several interactions with the Female indicator, but are not included for the sake of brevity

Fathers education

(0.015)

(0.015)

−0.182

−0.190

(0.058)

(0.057)

0.270

0.043

(0.112)

(0.110)

0.087

−0.097

(0.110)

(0.109)

0.143

−0.043

(0.125)

(0.118)

−0.006

0.100

(0.087)

(0.076)

0.362

0.078

(0.174)

(0.149)

0.260

0.015

(0.133)

(0.119)

−0.124

−0.063

(0.087)

(0.073)

0.007

0.014

(0.013)

(0.012)

0.007

0.004

(0.013)

(0.012)

Occupation at baseline Service

non-residential indicator is −.223 and statistically significant at a 95 percent level of confidence (t-value is −3.28). In addition, there is suggestive evidence indicating that the non-residential assignment is uncorrelated with earnings. For example, using data from the NJCS and administrative earnings records, Schochet et al. (2008) find no effect of the non-residential assignment on earnings of Job Corps participants in the short and medium term. Results from the two-tiered earnings frontier model with correction for sample selection are presented in Table 6. Estimates for the deterministic part of the frontier are very similar in magnitude to their OLS counterpart presented in Table 2. For brevity, these estimates are not discussed here. As stated in Section 3.2, one advantage of this simulated likelihood method is that it allows one to test for the selection specification. The Likelihood Ratio (LR) test, and Wald statistic associated with ρ for users, at the bottom of

Non-users

Laborer and construction Sales Private household Mechanics Clerical Agriculture Manufacturing

0.015

0.144

(0.070)

(0.070)

0.227

0.243

(0.084)

(0.083)

0.087

0.284

(0.087)

(0.084)

−0.098

0.023

(0.115)

(0.120)

0.232

0.292

(0.130)

(0.111)

0.342

0.194

(0.132)

(0.116)

0.228

0.436

(0.141)

(0.148)

0.155

0.103

(0.198)

(0.199)

J Prod Anal (2017) 47:33–47

45 Kernel density estimates, Workers 40

Table 6 continued Non-users

Principal MSA

0.071

0.105

(0.067)

(0.065)

0.050

0.085

(0.075)

(0.073)

0.842

0.843

10

Metropolitan Statistical Area (MSA)

Density

Market size

20

30

Users

Random component Employee incomplete information

0.919

0.919

Employer incomplete information

0.918

0.920

ρ

0.112

0.257

(0.166)

(0.105)

0

Error components

.6

.65

.7

Users

LR test, holding ρ = 0

.75

.8

Non-Users

0.322

3.810

p-value

0.570

0.050 30

LR test

40

Kernel density estimates, Firms

Table 6 do not strongly support the selection model. It would seem that selection, if any, is controlled by the included variables. One can note that the estimates for the error components are not suggestive of a meaningful difference across users and non-users levels of information, this is in contrast with the discussion of the maximum likelihood estimates without the selection correction. A visual analysis of the observation specific estimates, based on the simulated method, is presented in Fig. 2. These Kernel density estimates also show that the users vs nonusers differences in efficiency, relative to the full information labor market price, are not noticeable.

6 Conclusion By and large, studies evaluating job placement services focus on measuring benefits based on how quick matches are created, often neglecting the informational benefit that translates into accessing better labor market outcomes and, as a consequence, improved matching efficiency. In this paper I analyzed the latter, where employees and employers could potentially benefit from the information transmitted by a job placement agency. The analysis is based on a two-tiered earnings stochastic frontier, in which the residual of an earnings function is decomposed into a random component plus two additional measures that represent employee and employer incomplete information about offer and reservation wages, respectively. I applied this technique to users and non-users of job

20 0

10

Estimates of the specification holding ρ = 0 are omitted for brevity. I will point out that they are very similar to those reported in this table

Density

Note: Standard errors in parentheses

.6

.65

.7

Users

.75

.8

Non-Users

Fig. 2 Kernel density estimates of observation specific efficiency, based on maximum simulated likelihood estimation of the two-tiered earnings frontier

placement services provided by Job Corps, America’s largest and most important job training program for youths. The two-tiered method is extended to control for sample selection, which could arise due to the endogeneity of choice to use job placement services. This extension builds on Greene’s (2010) conventional stochastic frontier model with correction for selection, and is also based on maximum simulated likelihood estimation. Evidence strongly suggests that sample selection is not an issue in my application. Results based on the conventional 2TSF model indicate that only employers benefit from the information provided by placement services, where firms that employed individuals that used placement services are more informed about reservation wages relative to firms that employed non-users. This evidence is consistent with information distortions on the part of the job placement agency, where its informational advantage about labor market prices may be exploited to create more matches between employees and firms in less time at the expense of (plausibly) higher wages. Quick matches,

46

J Prod Anal (2017) 47:33–47

however, may not necessarily be a bad thing for workers using placement services, since one has to consider the trade-off between a lower unemployment spell vs. a larger one with a higher pay. In addition, information provided to employers could potentially improve the efficiency of the sorting process, i.e., matches will be generated at a price that closely relates to the actual but unknown productivity of the match, which in turn would benefit employees. Acknowledgements I would like to thank useful comments by Alfonso Flores-Lagunes, Solomon Polachek, Subal Kumbhakar, and the Binghamton University’s Labor Group. In addition, I thank the conference participants at the 2013 Midwest Economics Association meetings. Compliance with ethical standards Conflict of interest interest.

The author declares that he has no conflict of

References Aigner D, Lovell C, Schmidt P (1977) Formulation and estimation of stochastic frontier production models. J Econom 6:21–37 Armstrong M (2006) Competition in two-sided markets. RAND J Econ 37(3):668–691 Autor D, Houseman S (2010) Do temporary-help jobs improve labor market outcomes for low-skilled workers? Evidence from “work first”. Am Econ J: Appl Econ 2:96–128 Autor D, Houseman S, and Kerr S (2012) The effect of work first job placements on the distribution of earnings: an instrumental variables quantile regression approach. NBER Working Paper. Black D, Smith J, Berger M, Noel B (2003) Is the threat of reemployment services more effective than services themselves? Evidence from random assignment in the UI system. Am Econ Rev 93(4):1313–1327 Bloch F, Ryder H (2000) Two-sided search, marriages, and matchmakers. Int Econ Rev 41(1):93–115 Bradford D, Kleit A, Krousel-Wood M, Re R (2000) Stochastic frontier estimation of cost models within the hospital. Rev Econ Stat 83(2):302–309 Butters G (1977) Equilibrium distributions of sales and advertising prices. Rev Econ Stud 44(3):465–491 Chawla M (2002) Estimating the extent of patient ignorance of the health care market. In: Devarajan S, and Roger F (eds), World bank economists forum, World Bank Publications, Washington DC, Vol 2: pp 3–24. Efron B (1982) The Jackknife, the Bootstrap and other resampling plans. Society for Industrial and Applied Mathematics, Philadelphia Ferona A, Tsionas E (2012) Measurements of excess bidding in auctions. Econ Lett 116:377–380 Gaynor M, Polachek S (1994) Measuring information in the market: an application to physician services. South Econ J 60(4):815–831 Greene W (2010) A stochastic frontier model with correction for sample selection. J Prod Anal 34:15–24 Greene W (2011) Econometric analysis. 7th edn. Pearson, New Jersey. Gourieroux C, Monfort A (1996) Simulation-based econometric methods. Oxford University Press, Oxford Groot W, Oosterbeek H (1994) Stochastic reservation and offer wages. Labour Econ 1(3):383–390

Heckman J (1976) Discrete, qualitative and limited dependent variables. Ann Econ Soc Measurement 4(5):475–492 Heckman J (1979) Sample selection bias as a specification error. Econometrica 47:153–162 Hofler R, Polachek S (1985) A new approach for measuring wage ignorance in the labor market. J Econ Bus 37(3):267–276 Horrace W, Schmidt P (1996) Confidence statements for efficiency estimates from stochastic frontier models. J Prod Anal 7 (2):257–282 Hoynes H (2000) Local labor market and welfare spells: do demand conditions matter?. Rev Econ Stat 82(3):351–368 Jondrow J, Lovell C, Materov I, Schmidt P (1982) On the estimation of technical inefficiency in the stochastic frontier production function model. J Econom 19:233–238 Kim M, Kim Y, Schmidt P (2007) On the accuracy of bootstrap confidence intervals for efficiency levels in stochastic frontier models with panel data. J Prod Anal 28(3):165–181 Kumbhakar S, Lovell CA (2003) Stochastic frontier analysis. Cambridge University Press, Cambridge, UK Kumbhakar S, Parmeter C (2009) The effects of match uncertainty and bargaining on labor market outcomes: evidence from firm and worker specific estimates. J Prod Anal 31(1):1–14 Kumbhakar S, Parmeter C (2010) Estimation of hedonic price functions with incomplete information. Empir Econ 39(1):1–25 Kumbhakar S, Tsionas M, Sipiläinen T (2009) Joint estimation of technology choice and technical efficiency: an application to organic and conventional dairy farming. J Prod Anal 31(3):151–162 Lai H (2015) Maximum likelihood estimation of the stochastic frontier model with endogenous switching or sample selection. J Prod Anal 43(1):105–117 Levit S, Syverson C (2008) Market distortions when agents are better informed: the value of information in real state transactions. Rev Econ Stat 90(4):599–611 Lippman S, McCall J (1976) The economics of job search: a survey. Econ Inq 14:113–126 Masters A (1999) Wage posting in two-sided search and the minimum wage. Int Econ Rev 40(4):809–826 Mortensen D (1986) Job search and labor market analysis. In: Ashenfelter O, Layard R (eds) Handbook of labor economics. North-Holland, Amsterdam, pp 849–919 Mortensen D, Pissarides C (1994) Job creation and job destruction in the theory of unemployment. Rev Econ Stud 61:397–415 Mortensen D, Pissarides C (1999) New developments in models of search in the labor market. In: Ashenfelter O, Card D (eds) Handbook of labor economics. North-Holland, Amsterdam, pp 2567–2627 Murphy K, Topel R (2002) Estimation and inference two-step econometric models. J Bus Econ Stat 20:88–97 Olsen R (1978) A note on the uniqueness of the maximum likelihood estimator of the tobit model. Econometrica 46:1211–1215 Papadopoulos A (2014) The half-normal specification for the two-tier stochastic frontier model. J Prod Anal: 43:1–6. Parmeter C, Kumbhakar S (2014) Efficiency analysis: A primer on recent advances. Foundations and Trends in Econometrics 7:191–385 Petrongolo B, Pissarides C (2001) Looking into the black box: A survey of the matching function. J Econ Lit 39:390–431 Polachek S, Robst J (1998) Employee labor market information: comparing direct world of work measures of workers’ knowledge to stochastic frontier estimates. Labour Econ 5:231–242 Polachek S, Yoon B (1987) A two-tiered earnings frontier estimation of employer and employee information in the labor market. Rev Econ Stat 69(2):296–302 Polachek S, Yoon B (1996) Panel estimates of a two-tiered earnings frontier. J Appl Econom 11:169–178 Pries M (2004) Persistence of employment fluctuations: a model of recurring job loss. Rev Econ Stud 71:193–215

J Prod Anal (2017) 47:33–47 Rothschild M (1973) Models of market organization with imperfect information: a survey. J Polit Econ 81(6):1283–1308 Rutherford R, Springer T, Yavas A (2005) Conflicts between principals and agents: evidence from residential brokerage. J Financ Econ 76:627–665 Schochet P, Bellotti J, Cao R, Glazerman S, Grady A, Gritz M, McConnell S, Johnson T, Burghardt J (2003) National Job Corps study: data documentation and public use files. Mathematica Policy Research, Inc., Princeton, NJ, Vol 1 Schochet P, Burghardt J, Glazerman S (2001) National Job Corps study: the impacts of Job Corps on participants’ employment and related outcomes. Mathematica Policy Research, Inc., Princeton, NJ Schochet P, Burghardt J, McConnell S (2008) Does Job Corps work? Impact findings from the national Job Corps study. Am Econ Rev 98(5):1864–1886 Sharif N, Dar A (2007) An empirical investigation of the impact of imperfect information on wages in Canada. Rev App Econ 3:137–155 Sianesi B (2004) An evaluation of the Swedish system of active labor market programs in the 1990s. Rev Econ Stat 86(1):133–155

47 Sipiläinen T and Oude Lansink A (2005) Learning in switching to organic farming. Nordic Association of Agricultural Scientists, NJF Report, Vol 1 Stigler G (1961) The economics of information. J Polit Econ 69 (3):213–225 Stigler G (1962) Information in the labor market. J Polit Econ 70 (5):94–105 Terza J, Basu A, Rathouz P (2008) Two-stage residual inclusion estimation: addressing endogeneity in helath econometric modeling. J Health Econ 27:531–543 Tomini S, Groot W, Pavlova M (2012) Paying informally in the albanian health care sector: a two-tiered stochastic frontier model. Eur J Health Econ 13:777–788 Train KE (2003) Discrete choice methods with simulation. Cambridge University Press, Cambridge US Department of Labor (2012) Job Corps. http://www.dol.gov/dol/ topic/training/jobcorps.htm#doltopics Winterhager H, Heinze A, Spermann A (2006) Deregulating job placement in Europe: a microeconometric evaluation of an innovative voucher scheme in Germany. Labour Econ 13:505–517