The Joint Determination of Union Status and Union Wage Effects: Some Tests of Alternative Models Chris Robinson The Journal of Political Economy, Vol. 97, No. 3. (Jun., 1989), pp. 639-667. Stable URL: http://links.jstor.org/sici?sici=0022-3808%28198906%2997%3A3%3C639%3ATJDOUS%3E2.0.CO%3B2-D The Journal of Political Economy is currently published by The University of Chicago Press. Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at http://www.jstor.org/about/terms.html. JSTOR's Terms and Conditions of Use provides, in part, that unless you have obtained prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you may use content in the JSTOR archive only for your personal, non-commercial use. Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at http://www.jstor.org/journals/ucpress.html. Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed page of such transmission. JSTOR is an independent not-for-profit organization dedicated to and preserving a digital archive of scholarly journals. For more information regarding JSTOR, please contact support@jstor.org. http://www.jstor.org Wed May 30 15:27:14 2007 The Joint Determination of Union Status and Union Wage Effects: Some Tests of Alternative Models - - Chris Robinson Universzty of Western Orltario The problems of estimation and interpretation of union wage differentials are examined. T h e properties of cross-section and longitudinal estimators are compared. Estimates are presented and those in the literature summarized. Conflicting results are obtained. Longitudinal estimators typically produce results smaller than those of OLS, while cross-section methods (instrumental variables or inverse Mills ratio) raise the estimate. T h e paper offers a reconciliation of these results. It supports a more optimistic conclusion than that reached in reviews by Freeman and Lewis, who argued that little has been learned from attempts to deal with the endogeneity issue. Comparisons between estimators are used to throw light on the process governing union status and to suggest interpretations of "union differentials" consistent with the current evidence. I. Introduction In recent discussions of the estimation of "union differentials" there appears to be a quite general agreement that union status is not exogI wish to thank without implicating Robin Carter, James Heckman, Peter Kuhn, Glenn MacDonald, Aman Ullah, and Mike Veall fbr helpful discussions on the issues addressed in this paper. Helpful comments on an earlier draft were received from workshop participants at the Universities of Chicago, Guelph, Western Ontario, and Wilfrid Laurier. An anonymous referee provided many useful suggestions that are reflected in the present version. Finally, thanks are also due to Paul Rilstone for making available his program for nonparametricjoint density estimation. Financial support for this research was provided by the Social Science and Humanities Research Council of Canada and by the Center for the Study of the Economy and the State. University of Chicago, through a grant from the Lilly Endowment. [ J o u m l of Poltt~ralE ror~omj,1989, vol 9 i , no 31 0 1989 b) T h e University of Chlcago. All rights reser\ed 0022-5808!89/9i03-0007$01 50 J O U K N $1. OF POLITIC ,$L ECONC)M'I' 640 enous (Freeman 1984; Robinson and .I'onies 1984; Duncan and Leigh 1985; Le~vis1986). Given that some endogenous process determines union status, further questions have naturally arisen, along lvith discussions about ~ v h a tunion differentials measure in the endogenous union status context and h o ~ vthey should be interpreted. Unfortunately, there has been much less agreemerit over these issues. Follo~ving a variety of attempts to "deal with" the endogeneity, several authors, notably Freeman and Xledoff (1982) arid Le~vis(1986), have asked ~vhetheranything has been learned f'rom these exercises. 'I'he different estimates of union differentials obtained from these attempts have been compared. 'I'lie pessimistic conclusion that has usually been d r a ~ v nis that there is no discernible pattern to the estimates-many of ~vliichare considered unreasonably high or lo~vand that therefore no improvement in our understanding of the union differential problem has been made (1,elvis 1986, pp. 58-59). A subgroup of these estimates that appeared to hold out great promise and that dici offer more "stable" estimates Ivere those obtained from longituciinal estimators. Panel estimates of union differentials have been presented by hlincer (1981), C:hamberlain (1982), Jakubson (1984), and others. Typically the longitudinal estimates are much smaller than the comparable cr-oss-section estimates. Freeman (1984) argues that these results are d u e to the differential effects of both measurement error- and selectivity effects across the two types o f data sets. In particular, either singly or (less restrictively) together, measurement error and selectivity problems bias longitudinal estimates do~vn~var-d. Some evidence is provided of severe measurement error problenls ~vithlongitudinal (relative to cross-section) data that, Freeman argues, considerably reduce their value. Indeed, put into context, Freeman (1984) and Freelnari ancl lledoff (1982) argue that longitudinal estimates are a yet further manifestation of'the great disparity in union differential estimates that occur ~ v h e r eany attempt is made to deal lvith the union endogeneity problem. I'hese estimates are thus characterized as being "less useful" than the "more stable" estimates obtained from ordinary least squares (OLS) cross-section estimates. AAt best the longitudinal estimates may serve as a lolver bound on the "true" differential (Freeman 1984). T h e present paper offers evidence that supports a more optimistic view than that of Freeman and l,e\vis. In Section 11, a tlvo-period model of unionization is specified as a frarne~vork~vithin~vhichboth cr-oss-section and longitudinal estinlators may be compared. Since both interpretation and properties of the estimators depencl crucially on the ~inderlyirigendogenous process, a detailed, explicit characterization is provided for the stochastic structure. T ~ v obasic sources of UNION W A G E EFFECT'S 641 endogeneity are identified: the behavior of the agents (~vorkersor firms) and measurement error. Three alternative approaches to solving these pi-oblems are considered: control function methods (e.g., inverse Mills ratio), instrumental variables, and longitudinal differencing. T h e different properties of these estimators under various assumptions are used to thro\v light on the process governing union status. Section I11 conducts exogerieity tests and presents both "corrected" and "uncoi-rected" estimates of the union differential from both cross-section and longitudinal data sets. Exogeneity of union status is generally rejected. T h e union rvage differential estimates provide substantial evidence of a definite pattern to biases induced both by selection problems and by measurement error. T h e test results and the pattern of differentials are compared ~viththe existing literature in Section IV. It is argued that the pessimistic \-ie\v of Freeman and Le~vis,that \ve have little to learn from alternative (to OLS) estimates of union differentials, may be unjustified. Section V asks the follo~vingquestions: What kind of interpretation is necessary for the reconciliation of the results of the different estimators! Does it require the imposition of ~vhatniight be regarded as unr-easonable assumptions governing the under-lying str-ucture? Does it imply other conclusions that are at variance lvith our observations? In Section VI, comparison of alternative estirriators designed to deal ~viththe endogeneity issue is used to test hypotheses about the union status process. A simple "fixed effects" model (of union status) is contrasted ~vitha "1.andom effects" model. Comparison betlveen instrumental variables estimators and contr-ol function estimators suggests that the simpler fixed effects rnodel may be a good approximation for the union probleni. Some conclusions are presented in Section VII. 11. A Framework for Union Wage Differential Estimates T h e purpose of this section is to extend the cross-section nod el of Robinson and Tomes (1984) to t ~ v operiods, keeping the specification as simple as possible. This cross-section model was estimated on 1979 Canadian data that contained high-quality Lvage data for hourly paid workers. second year of data (1981) is now available. 'The primary advantages of the additional year of data are the larger sample size and the possibility of testing a Lvider variety of hypotheses regarding union status. T h e simple t~vo-periodmodel of this section is designed to capture only a very limited extension of the one-period setting. Wages in each sector (union [('I or nonunion [N]) in each period 642 JOURNAL O F POLITICAL ECONOMY are as specified in Robinson and Tomes (1984): where the superscript t indicates period t values; Xj is a vector of exogenous regressors. T h e precise nature of the disturbances E:., and E,;., determines the difference in the properties of the estimators considered below and ' I An explicit characterization that would generate (1)-(4) is as follows. Let workers have observable inputs ( f I , ) that they cotivert into output ( 0 , )in the union and nonunion sectors. according to individual functions that may be approximated t)y indi\idual proportionality factors: where C , I if worker I is in the union sector. Variation across individuals in = +,.,(or +,-,) could simply model the unobser\able fraction of individual 1 ' 5 total skills. For a given individual I , I$, ., and +,-,ma) differ to allow for the utilization of diff'erent t)pes 01, skills or the operation of different production processes in the two sectors. Payment to each worker is. in general. some function o f output that ma) differ by sector; assume that these may be approximated as I-', (0,)= RI.+,.,fI, it C', = 1, T h e log wage (payment) form is then sirnpl) where y:' ET,, = In In M',., = y: + In W,-, = y,* + X,y X,y + E:.,, + ec,, In K,. + y,,, y: = In K,- + y,,. X , y = C:= I ylX,,. It1 H , = y,, + I:_,ylX,,, ,, and ,E: = In +.,-,. Assume that X, and the +'s are independent. Then = +, E(ln +, lH) +, ) = €7- = E(ln = E(ln I$,-) = e: and E(In +.,-/fI) In general, ET and E-: rameteri7ing. we get will be nonzero so that y:. and y: = y,. In K',., + X , y + E,.,. In W,., + = y,- X,y + are not identified. Kepa- t,-,. where y , = y: + E:., y, = y:. + E:, E ( . , = e:, e:, and E,-, = E:, - E:.. Thedifference, in terms o f structural parameters, that unions make to observed earnings comes from two sources. First, there is a difference in the payment functic~n(y?. y:.) associated with union "rents"; second, there may be difference5 in the production processes, implying differences in the relationship between observed Inputs and output (E:. - E: ), associated with union productivit) effects. T h e "union effect" that ma) be estimated from the log wage equations is - - 7,. - 7,- = (y: - yX-) + (e:. - E:.). Since, in general, e? Z E:., the separate "rent" and "productivity" effects cannot be identified, only an overall union effrct. U N I O N \VAC;E: E:FEFCTS (343 the interpretation o f t h e estiniates obtained. Since two periods of data are available, the disturbances may usefully be decomposed into tirneinvariant components and remainders: T h e time-invariant components, O 1 . and 0,,-,, niay be thought of as aspects of omitted skills that persist for a given ~vorkerover time. T h e remainders r({ and T/,,, then represent time-varying omitted chax-acteristics or, more generally, aspects of the production process o r payments process that may be represented as raridom drawings for each individual and that may differ over- time.' If the vectors y,\. and y l . differ only in their first elements, (1) and (2) may he combined to form the staridard OL,S esti~natirlgequation for the union problem, that is, an earnings equation that includes a dumtny mr-iable for union status:" 'I'he standard "union effect" estimates (Freeman 1982; L,e\vis 1986) are OL,S estimates of the parameter (yl - y,\) in equation (3). If union status, l':, is uncorrelated with the disturbance \':, this pariinleter is consistently estimated by OLS and represents the earnings difference any given ~vorkerobtains in the union sector relative to the nonunion sector. '1-he problem ~ v i t l itreating (3) as a standard cross section that [nay be estimated by OLS is that union status ia likely to be correlated lvith the disturbance. T h e r e are t ~ v osources of this correlation. T h e first source is the correlation induced by the behavior of some of the agents in the model. For example, there is the argument that en~ployerschoose, frotn some union pool, those indivitluals tvith higher (unobserved) aljility (C),,,, 0[ ). 'I'here is also the nlore general selection argument that the ~vorkersknow (:it least part o f ) the disturbance iind select thenlselves into their most p r e k r r e d sector taking this into account. T h e second source is the correliitioti iliduced bv For example, o u t p u t at the incli\idual le\el nra\ I)e \to( h,~\tic.. ~ n dthe ~ratllr-eo t tllc stochastic- proce5s rna\ ditfer tn sector hut l)e \t;~hleo i e i time. T h r 111(11\idl1al tlr:~\\.i~rg\ \vill diff'cr acro5s i ~ i d i \ i d u , ~ l wctors, s, ntid t i ~ n r . Tests o n the clata 5ets ernplovecl in Sec. 111 hrlo\\ cannot jiener,iIl\ reject ecll~;ll~t\ of the earnings function slope coefficients. ' 644 JOUKKAI. OF POLITICAL. ECONO>I.I1 errors in the measurement of union status. Freeman (1984) has argued that measurement error is snlall enough to be ignored in crosssection estimators. Nevertheless, OI,S estimation of (3) for individual years o r as a pooled cross sectionlti~neseries ~vouldstill be inconsistent hecause of the first source of correlation incluced by the choice behavior of Lvorkers o r firms. T h e possibility of a more coniplex error structure than that assumed for consistency of OLS raises both estiniatiorl and identification problems. Without further restrictions on V:, eq~latiorl(3) represents a random effects model of union status on earnings. A4 ~vorker'searnings in the union sector differ fi-om ~ v h a they t ~vouldbe in the nonunion sector by t ~ v oseparate terms in (5).'I'11e first is (yl. y,,), ~vhichis common to all ~vorkers;the second is (E:. - EL,),representing a random effect of union status. If this random component of the ~ l n i o neffect is suppressed by imposing E:. = E;, for all i , then a fixed effects model of union status results: In the fixed effects model, the parameter (y,. - y,.) represents the union effect, as in the 0 1 , s studies, though the parameters \\.ill not be consistently estimated by OLS if cov(i', E , \ ) # 0. In the random effects model, the randoni component complicates the identification of the union effect. T h e parameter (y,. - y,.) represents the difference in earnings bet\\,een union and nonunion sectors ~ v h e n~vor-kersa re randonily assigned to sectors, as in the fixed effects nod el. Ho~vever, contrary to the fixed effects model, (y,. - y v ) does not represent the difference in earnings expected if a union ~vorker(i.e., one having been allocated to the union sector by the union status determination process) Lvas transferred to the nonunion sector. Under the OLS assumptions, o r mor-e generally under fixed effects assuniptions, any experiment that moves a ~vorkerbet~veensectors results in a fixed change in the ~vorker'searnings given by (yl- - y,.). Under ranctorn effects, union Lvorker i ~ n o v e dto the nonunion sector Lvill experience the earnings change (y,. - y,,.) + (E, - E,.). T h e expected change is (Y,, - Y . ~ )+ E[(E[.,- E,\.)/(', = 11. LVith data o n earnings alone (i.e., in the absence of, say, a 1-egressor in the process determining union status), all workers have the same \.slue of E[(E,.,- E \ - ) I L', = 11. AS a result, (yl - y,,) cannot be separately identified. This is apparent if equation ( 5 ) is reparameterized to 615 UNION W A G E E F F E C T S Equation (7) is econometrically equivalent to the fixed effects model (6) but identifies a different union effect.' T h e effect ( y , . - y,.)* does not have the interpretation of the union differential resulting from random assignment. Alterncrtzve Estzmators of C'nzon Effects T h e problem of union status endogeneity, cov(C':, 1':) # 0, has resulted in several alternative approaches to estimating ( y l- - y v ) . Two alternative approaches to the problem for single cross sections are instruniental variables (I\') and control function methods. Consider first the I\' method. This requires the existence of instruments for C': that are uncorrelated r\.ith the disturbance in ( 5 ) .Since in general the disturbance contains C':, E:.,, and E:, some special assumptions are required. example is given in Duncan and Leigh (198.5). T h e validity of the assumptions necessary for the 'I.\ method to be consistent depends on the underlying process generating union status. This issue is discussed below. T h e alternative class of estimators considered is the control function type, of ~vhichthe inverse hlills (IhI) ratio method has been most widely used. In this case, union status occurs rvhen 1: > 0, where I: is the net-of-"costs" ( C : ) rvage difference from choosing the union sector: I," = 111 W:., - In W!v, - C: = Z:*i + E:, c: = s:s + Given joint normality of E:, E:.,, and follorving modification to (5): E: E:,, E: = E:., - E: - E: . + E(.~[(c; = E{('j[(e;, = 0. - - e;.,~ - e',,) - (8) (9) ( l ) , ('L), ( 8 ) , and (9) yield the * 111 the fixed effects model ( 6 )the distur-bance iu e',,. H \ assumptiorr, Ee: the disturbance is I.':*, which also has rer-(1 mean: El.':* = Ee;, ,, E ( e ; , - c; E ( E ; , - e',.,i(lr: = = 0. Ill ( 7 ) = 11 l]}Pr(( , = 1 ) See Heckrnan a n d Robb (198.3) for- m o r e detailed diucussionu in the context of estin~atirig effects of tr-aining programs. J O U R S A L OF POLITICAL ECOh'Ohll and A:., anti A(, are the relevant inverse hlills ratios. T h e tfisturbarice in this case al~vayshas a zero mean and is uncorrelated ~vithC':; hence n o instrument is necessary. This method will produce consistent estimates fhr any type of assumption governing the disturbances, O's, q's, and E(.'s provideti that there is a regressor in the selection rule and that the conditionai ciisturbances k1(8[. + q:.)I l', = 1 and E(O,v, + q\- ) I I', = 0 can be consistently estimated. T h e availability of longitudinal data for equation (5) suggests an alternative to the cross-section approach to the union endogeneity probleni, by exploiting the tirne invariance of the person-specific 8's. Unfortunately, the longitudi~laldifferencing approach rvill not ~ v o r k in general. T h e person-specific effect is not elimi~latedbecause of the differences no-oss sectors. In adtfition, the different q's across sectors each period result in 1,'; and c',' both appearing in the error term. T h e standard longitudinal approach has been to impose 8[. = O.y and = T( , t = 1, 2 (Freeman 1983, pp. 3-4, eqq. 1-3). I n that case the (individual) mean difference form of ( 5 ) is rvhere = xi. This 12.ithin-groupregression, ~ v h e r ethe "group" is the data for both time periods for- the same individual, rvill provide unbiased estimates if I', and the q's are uncorrelated. that is, if the permanent components (0's) are the o~liysource o f the encioge~leity problenl. ..lltr? ~ i a t ~ .Ilocicl, ~rp of tkc Df~tfv?)iinatlo~i of ( 'nzon Strrtuc T h e ~alidityof the assumptions necessary for consistency for the estimators considered above depends on the nature of' the process b!, tvhich individuals become union members and the error rvith rvhich union status is measur-ed. Unfortunately there is little agreement in the current literature on the union process. hlost of the discussions, ho~vever,involve sonle form of choice by rvorkers or- firms based on t3oth obser-vable and unobservable factors (Lee 1978; Abolzd and Fart3er 1982; Farber 19830; Robinson and Tomes 1984). O n e general point that emerges frorn these discussions is that rvhere workers know some of their characteristics that the investigator does not, the most sel-ere econometric problem arises. T h e corollar-y to this, pointed out t x Heckman and Kobb (1985) in the training model context and exploited earlier by Zellner, Kmenta, and Dreze (1966), is that the econometric problems are often reduced if the agent's perfect knorcledge is replaced by uncertainty. ?'he agents then have to make their decisions on. tor example. union status, not or1 the basis of unobserved (by the investigator) individual characteristics but rather- on expectations the agents torm about horc rvell they would do in each sector. Under appropriate assunlptions these expectations, conditioned on observables, tvill be the same across i~idividuals,thus removing the correlation b e t ~ c e e ~uriion i indistatus and ~~nobservable victual characteristics present in the certaint) case. 1)iffel-e~lcesacross estimators may theretore be used to construct tests o n alternative hypotheses of' the union status process. '1.0 facilitate tlisc.ussion of the t3asic issues, a simplified structure for union choice over ttvo periods is specified and used to contrast the properties o f the alternative estimators. Assume that two sectors are available to each rvorker, unio11 o r nonunion, in each period and that the rvage rates in each sector are given by (1) and (2). T h e r e is a cost attached to union status, including psychic costs of' union membership, together ~ c i t hresources spent on actually gaining membership anti any differences in expected \cork duration o r rvorking conditions. Assurne that this cost, f'or simplicity, may be expressed, relative to the nonunion tcage, 1)y C:, as specified in (9). Consider first the case of per-f'ect certainty for the agents. 'I'he individual's objecti1.e is to maximize net tvages over the ttvo periods. 111the simplified framercork of this section, this requires choosing the sector that niaximizes net rvages in each period. Hence the rule tor each period is choose C': = 1 i f f I: > 0, t = 1, 2 Given this process, longitudinal differencing methods generally fail, even in the absence of measurement error, because since I: depends on the pil's, CT: and the q's rcill t) pically be correlated. 'll1e available of the instr-urnents for CT, are the Z, (or some knorcn transforrn;~tio~i Z,). Consistency f'or the I\' estimation thus reyuir-es that (some transformation of) the Z, be uncorrelated ~ c i t h\', in the limit. ,\ detailed discussion and precise conditions for consistency are available in Heckman and Robt) (1983). Duncan and Leigh (1985) satisfy the consistenc), requirement by assumi~lgthat the joint distributions of (e( , e) and ( E , ~ E, ) are the same. This implies that E;17,1Z,= E17, = 0. Since I', = ev + I',(E[ - E , ~ , )a nd EE,,, = 0, EI', = 0 +L;Lr, (el - E , , ) = 0. Thus union status and (eI. - E Y , ) are required to be uncorrelated. This is guaranteed by the fixed effects (of union status) rnodel, rchich sets el. = E.\. , for all i. More generally in the random effects frarne- 'j49 L'NION W A G E EFFECTS tained by probit estimation o f t h e probability of union status. Consistency in this case will therefore fail if normality fails. An obvious alternative to perfect certainty at the level of the agent within the present framework is to assume that agents kno~vthe 0's but that they have the same information as the investigator regarciing the q's. In this case 0 may be thought of as omitted individual characteristics o f the same type as schooling and experience whose value in each sector is common knowledge. T h e q's then represent individual variation in earnings across sectors that arise from stochastic elements in the sectors themselves on different realizations to individual character-istics whose value in each sector is uncertain, for example, the ability to function well in a regulated union environment. Let the criterion for choosing the union sector then be that the expected net wage in each period is positive, that is, choose C': Given Eq: = Eq:, Elf = = = 0, 1 E(Z:n = 1 iff El: > 0, 1 = 1, 2. 1, 2, + e:) = Zfn + 0( - 0, - e: . Thus e: reduces to - O., , - c::~). T h u s L': will be correlated with ( O ( . - O.v, - E::,), but not with o r q;.,, assuming that e:,, is uncorrelated with the q's. T h e longitudinal estimator remains inconsistent unless O1., = O,, and C': is measured without error. If (01- - 0.\.,) # 0, this term remains in the disturbance and is correlated with CT:. T h e I\' approach now has to contend with the term (0(., - O.\.,) in the criterion function rather than (E(., - e . ~ , )Suppose . that EL',(qc., - q.,,,) = 0 so that E\~,Iz,= EC',(O1, - O.,z) = E(Oc IC', = l)Pr(CTl= 1, Z,) Since indiviciuals know ( O r - O,v,) when the union decision is taken, this will not, in general, be zero on the same grounds as before. Special assumptions, analogous to those made in the perfect certainty case, are thus required for consistency. T h e Ihl approach would yield consistent estimators subject to the normality assumption as before. T h e only change would be the interpretation of the parameter estimates for the coefficients on the inverse hlills ratios. Finally, a more extreme form of depai-tur-e from perfect certainty assumes that the agent knows only what the investigator knows, except, possibly, for ec, In that case the expected wage difference is El: = Zfn + EeC,,and E, thus reduces to (at most) €I:,. In this case, C': thus depends, given Z f , only o n ec,. Since (el., - E.,,~) is now eliminated from 650 J O U R N A L O F POLII'ICAL ECONOSlY the criterion function, there is no reason to suppose a nonzero cot-re, E-Y,). T h e "special" assumptions required lation between C', and ( E ~ in the previous section no longer appear restrictive, given the choice model generating union status. From the econometric point of view, this form of the random effects model presents estimation problen~s equivalent to those of the fixed effects n~odel.' T h e Ihl model remains consistent under any intormation assumptions, subject only to the distributional requirement. T h e imperfectforesight assumption reduces the problems for the longitudinal estimator. However, consistency still requires the difference in union status to be uncorrelated with (rl' - -qj). Ho~vever,if union status is generally uncorrelated lvit11 the ~ ' s the11 , all the estimators in the imperfect-foresight case, including OLS, will be consistent.' I'his raises the question of the gain from longitudinal estimators in this context. Consistency of the longitudinal estimator- requires stringent assumptions. T h e only source of the endogeneity problem must be time-invariant individual components that are the same across sectors. In that case, horvever, given obsei-ved characteristics, the union status changes over time have to be random, tvhich is an unattractive behavioral model for the evolution of union status. Differences across estimators may be exploited to test alternative hypotheses about the process generating union status. In the next section, union differential esti~~lates and exogeneity tests based o n the alternative estimators are presented. I n Section I V , the results are compared rvith the existing literature. 111. Empirical Results on the Alternative Estimates of Union Wage Differentials and Exogeneity Tests T h e estinlates of union rvage differentials presented in this section are based on the Canadian Quality of Life Sui-vey descl-ibed in Robinson and To111es (1984). Table 1 presents summary measures of the data. T h e alternative estimators ciiscussed in the previous section tvere used to construct exogeneity tests and to obtain estimates of ( y l . - y.,.). T h e ' T h e hxed effects rnodel \ i ~ p p o s e se i , = E \ , \o that I., = E , , . T h e random etfect\ model u n d e r uncertaint> has I., = E , + C ' , ( E , , E \ . , ) b u t E C ' , ( t l t.\ ) = 0 . so that the additional tern1 C ' , ( t l , - E , , ) presents rlo further econometric problern tli,un tlie o n c that was present with the flxed effects model (Heckrnan a n d Kobb 1085). ' T h e dihtur-ba~iceis [ ( H I + T; , ) l ' :+ ( 0 , ) + r(', ) ( I - (./)I. Since I./ a n d the 8's and r('\ a r e unco~-I-elated, E{.} = 0 : - Hut ( :' = L': since C: takes o n the ~ a l u er e r o 01 - o n e hence DATADEFINITIONS, MEANS, A N D STANDARD DEVIATIONS 1979 DATA VARIABLE NAME Atlantic Quebec Ontario Prairies Yrssch Expr Tenure Male POW Ln u! Union Part time Married Public Plantsize DEFINITION Dummy variable = 1 if resident in Atlantic province* Dummy variable = 1 if resident in Quebec Dummy variable = 1 if resident in Ontario Dummy variable = 1 if resident in Prairies Years of schooling Experience = age yrssch - 6 Number of years with same firm Dummy variable = 1 if male Percentage of organized workers in respondent's industry Satural logarithm of respondent's hourl! wage (1980 dollars) Dummy variable = 1 if individual belongs to a union Dummy variable = 1 if respondent works fewer than 30 hours per week Dummy variable = 1 if respondent is married Dummy variable = 1 if respondent works in public sector Mean size of plant in individual's industryt Mean Standard Deviation 1981 DATA Mean Standard Deviation .063 (.074) ,320 (.288) ,407 (.436) ,104 (.093) 11.353 (1 1.249) 18.1 17 (19.237) 5.632 (6.622) ,508 (377) 38.459 (39.447) 1.877 (1.946) ,464 (.542) ,257 (.228) ,620 (.683) ,241 (.256) 27.109 (27.842) NOTE.-Data for longitudinal sample are In parentheses * Reference group 1s Br~tishLolulnbid 'Computed from respondents' answers to question on number of enrplo\ees ar r-esl~ondcnt'iplace of work (see Roblnson and Tomec 1984). Units are 10 emploveer 632 .JOLIRiVAL O F P O L I T I C A L E(:ONOhlY 1 ABLE 2 EXOGE:NEIT 1. 7 EST S FOR UNIONS1'h.l'~'~ Test a n d D'ttn Set Keault .A. Tests Based on I\' versus OLS Corn1)arisons M'u-Hausman test: F-test o n ,joint significarlce of i~lstrurnentsin instru~llent-augmented regression: 1079 cross section 198 1 cross section Pooled 1979-8 1 cross sectiorl F F F = = = 1.883 vs. critical F ( 5 9 ) = 2.10 2.367 vs. critical F(.5R i = 2.10 3.623 vs. critical F ( 5 % ) = 2.10 B. Tests Based o n IRl Lersus OLS (:oniparisons F-test o n joint signif~canceof the i n ~ e r s e Xlills ratios: 1979 cross sectiori 198 1 croaa section Pooled 19711-8 1 cross section F F F = = = 4.613 vs. critical F(.5'/;) = 2 9 9 4.162 vs. critical F(.5%)= 2.99 6.175 \s. critical F ( 5 R ) = 2.99 (;. Tests Based on Longitudinal Lersus Cross-Section Estimates \Vu-Hausman test: t-test on the difference in the union coefficient: witliin-group estimator L s. NerloL e-Hausman-Ta)lor estiniator: 1979-8 I longitud~nalsample 1 = 3.335 results of the exogeneity tests are presented in table 2. Panel A reports the results based on I\.' versus OLS con~parisons.T h e I\.' approach was used to construct a Wu-Hausman test similar to that employed in Duncan a n d Leigh (1985). Conditional 011 the observed unio11 status, under the null hypothesis of'exogeneity of union status, for a single year of data, OLS applied to the separate union and nonunion sectors is unbiased and efficient. T h e r e are n o common parameters to be estimated between the sectors. If, however, the union and tlotlunion disturbance variarlces are equal-[u'(Ol~) + u~(~;.= ) ] [u2(0.,,) + u2(~<-)]-then E V ~= [u2(0.,,) + and OLS applied to the pooled union and nonunion sector samples will be unbiased and efficient. T o test for exogeneity using a single cross section, the OLS estimates may be compared with the estimates obtained using an instrument for C':. 'The natural instruments are 'The probability was estimated using probit estimates for n and assuming no~.malitjfor E:. T h u s CTiVIOiV WAGE EFFECTS 633 T h e term 0: does not have to be a consistent estimator for (,':;Ahence violation of the normality assumption here does not invalidate C': as an instrument. Under the null hypothesis of exogeneity. I\' is consistent but not efficient. Under the alternative hypothesis, OLS is inconsistent. T h e I\' estimator may o r may not be consistent, depending on the assumptions on the nature of the disturbances. However, in general under the alternative. plim 4ol.s f plim T h e tests reported in panel A utilize the "altel-native form" of Hausman (1978). This consists of a test of the significance of auxiliary regressors in the OLS estimation of' equation (5), perrrlitting interactions in the standard earnings function variables." T h e auxiliary regressors are the vector of instruments for union status and its rele\;ant interactions. This test was performed for both 1979 and 1981 cross sections. Under the null hypothesis of exogeneity, the true covariance matrix is consistently estimated by the standard OLS covariance matrix, M-hich is used to construct the F-test on the significance of the auxiliary regressors. These are reported in the first two ~ O \ V Sof panel A. Tests were also performed on a sample that pooled the independent observations from 1979 and 198 1. This consists of the 1979 cross section plus those individuals added to the sample in 198 1. T h e covariance matrix in this case remains proportional to the identity matrix under the null hypothesis of exogeneity. T h e F-test based o n it is reported in the third row in panel A. Exogeneity is rejected at the 5 percent level for 1981 and the pooled sample. It is rejected at the 7 percent level for 1979. Panel B reports the results for tests based on I M versus OLS comparisons. T h e null hypothesis of exogeneity is equivalent to the hypothesis that the inverse Mills ratio terms d o not enter the regressions. Under the null hypothesis of exogeneity, OLS consistently estimates the covariance matrix for equation (8). This covariance matrix is used to construct the tests in panel B. ( T h e results are very similar if the covariance matrix is corrected for the heteroscedasticity and the use of estimated \fills ratios that would follow under the alternative hypothesis of significant selection.) T h e F-tests were constructed on the joint significance of the inverse Mills ratio terms. Exogeneity is always rejected in these tests. Finally, panel C presents the test results based on the longitudinal data. T h e presence of a n individual component, assumed constant over time, creates a nonzero covariance in the disturbances for the ' .As rloted in n. 3, zero slope interactio~~s cannot generally be rejected. Howeler, the pattern of slope coefficierlt differences between union arld nonunion sectors is consistent with that found in other studies. e.g., a lower schoolillg coefficient in the union sector. 1 he model kcas theretore estimated pernlittillg slope illteractions for the standard earnings function variables. same individual in different time periods. However, the pooled (2year) model may be consistently and efficiently estimated by OLS, under the null hypothesis of exogeneity, provided that an appropriate weighting scheme is used to transform the covariance matrix to one that is proportional to the identity matrix. T h e appropriate transformation for the case of OI., = O,,. and = rl:., is quite straightf'orward and, as described in Nerlove (1971) and Hausman and Taylor (1981), is equivalent to a differencing of the observations as follows: In Wj - (1 - - w)ln 1VJ = [ciXj - (1 - - w)L',X,]yl, where o = { ~ r " ( ~ ) / [ c ~+' ( ~202(0)])' ) '. Equation (13) is a matrix weighted average of the within-group regression ( 1 1) and the between-groups regression delirled by averaging (-5) across t for each irldividual i : In M7,= y,. + (yl - y,.)c + Fy+ c. (13) ?'he longitudinal aspect of the sample permits a test of' independence of' union status from 0, under a maintained assumption of independence of L:; and 77:. provided 0(., = O.,.,. Under the null hypotheses of exogeneity, OLS estimates of (13) will be consistent and efficient."' ?'he OLS estimates of (1 1) will also be consistent, but since they use none of the information from the between-groups regression, they will be inefficient. Thus a \Vu-Hausman test may be constructed f'i.on1 a compai-ison of' these estimators. 'I'he test reported in panel (: was constructed from the within-group estimators and the efficient (under exogeneity) weighted OLS (Nerlove-13ausmanTaylor) estimator. Again exogeneity is rejected. '' I" T h e e\timate requires a conslslent e5timate of w. '1-lii\ ol)t;~inedfrom the fact that where SSK(LV(;) and SSRIBC;) are the surii\ of \cluared residuals froni the within-g~.oup and I)et\\een-groups regression, (Ilausnian ant1 Ta\Ior l!IHl). I I In general, 0, , f 0,. and ~i , # q; . In this caw tlie 1Vi1-Ilau\man tests nia\ \till I)e t ~ , ~ n \ l o r m a performed as indicated abo\e. 'The major dittere~iceis in the ,~ppropri;~te tion for eq. (13) and the estin>ation ot w. An added cornplicatiou ctelns from ctle tact that not all of the i~idi\idualsIn tlie 1979 ant1 1981 ianlple\ \\-ereinte~.vie\\edin botli years. If the ditterent data aiailahilit) is randoni, then the 5aml)les ma) tle reatiilv estimated together in a form analogous to eq. (13). TABLE. :3 - Data Set Zero Slope 11iter;rctions Earnirigs Function Slope Intcractions A. 1979 cross section (.V = 6851: O1.S Iv I hl B. 1981 cross section (.V = 85.5): OLS I \. IS1 C;. Pooled independent 197Cl-81 cross rectlons = 1,220): 0I.S Iv I hi D. Lotigitudi~lalsample. pooled 1979-8 I data ( S = 621): OLS I \' I hL E:. Lotigitutlinal s a ~ ~ i p l~ntli\idual-average e. data. between-groups estimate5 (.V = 3 12): 0I.S 11' I hi F. Longitirdinal sample, mean-differencecl data. within-group estimates (.V = 6001: OLS 111table 3 the alternative estimators are used to contrast the alternative estimates of union wage differentials, calculated as exp[(?( ?,,)XI - 1. T w o fhrms of the basic model ( 5 ) are estimated. ?'he first imposes ~ e r ionteractions between C': and all the elemenrs of' X: apart from the constant; the second permits slope interactions in the earnings function (see n. 9).T h e first three panels, A-(:, report the results of single-year cross sections individually and pooled. T h e ) show O12S estimates o f approximately 'LO percent for all the cases. T h e I\.' estimators use as instruments the predicted values of union status obtained f.1-om maximum likelihood estimates of a probit model for equation (8).12T h e IhI estimator is as described in the previous section. Both the I\.' and 111 estimates are much larger than the OLS estimates. T h e estimates range fiom 27 percent to 33 percent. T h u s " Instrurne~itstrom a lineal. prol)at)ilit\ n~odel5 x1-eal\o useti and procluced almost identical results. 636 ]OC'Kh'.AL OF 1'OI.ITICAL E C O N O M Y attempts to correct for endogeneit? by either the I\.' or I M method have the effect of raising the estimated union wage differential. 'I'he lo~lgitudinalresults are presented in panels D-F. T h e longitudinal estimator employed is the within-group regression defined by (1 1). This estimator requires data on the same individuals in both the I979 and 1981 cross sections. L~lfortunately,in the Quality of Life data, not all individuals in the 1979 cross section appear in the 1981 cross section and vice versa. In order to produce an OLS cross-section estimate from the same sample used to compute the within-group estimate, both the 1979 and 1981 data sets were reduced to include only individuals working both years. T h e result is termed the longitudinal sample.':' Because of the sr:laller sample size for this analysis, only the model imposing zero slope interactions was estimated. First, panel D uses both the 1979 and 1981 observatio~isfrom the longitudi~lalsample and computes pooled estimates equivalent to those in panel C. 'I'he observations in this sample are not in general independent, but the coefficients are consistently estimated by OLS. As in the full pooled cross-section sample, the corrected estimates are approximately 30 percent larger than the OLS estimates, though the general size of the differentials is snlaller. In panel El, estimates based on the between-groups regressio~~ (14) are presented. Inlposing O ( . , = O.,., and T i - , = T.;,. t = 1,2, as in the case of the within-group estimator yields 1',= 0, + 7,. T h e disturbance thus retains all the potential problems encountered with the cross sections. T h e results are therefore expected to be similar to those in panel D, as in fact they are. These may be used as the cross-section standard by rvhich to compare the longitudinal within-group estimates presented in panel F. 'I'he first row of panel F presents the estimated differential from the standard within-group regression represented by e q u a t i o ~(1 ~1). This shows a union differential of 11.5 percent, substantially lower than the 13.8 percent from the between-groups regression or the 13.7 percent from the OLS r e g r e s s i o ~on ~ the pooled longitudinal data. There is thus an apparent conflict between the results obtained by the alternative methods of addressing the endogeneity problem. Crosssection methods (I\' and IhI) result in upward adjustment; longitudinal (differencing) methods result in downward adjustment. ?'he remaining estimates in panel F are discussed in Section \' below. IV. Comparison with the Existing Literature T h e rejection of exogeneity of union status is in general agreement with results presented in the literature (see, e.g., Duncan and Leigh 13 Some o1)vious prol~lemsof selectiorl bias ma) arise because ot this restriction; these are discussed below. U K I O N %'AGE EFFECTS TABLE 1 Data Stud\ Reqult A. Sample Selection Bias KLS Older Men SEO PSID Duncan and Leigh (1980) Lee (1978) Heckman and Neurna~ln (1977) Ro1)insoti and Tomes (1981) SWH Kumar and Stengos (1985) LCWS Simpson (1985) NLS Yot~tigMen Leigh (1980) KLS Young Men and Womeri QOE QOE Blau and Kahn (1983) Farber (19836) Atrostic (1981) CPS Wessels (1981) Grant, Swidinsky, and X'ariderkamp ( 1 98.5) Podgursky (1980) Sul~stantialrise (all workers) I'ery slight decline (operatives) Moderate or substantial rise (all workers) Moderate risc (hourl) paid workers) Moderate or sul~stantialrise (all workers) Moderate or substantial rise (all workers) Moderate or sut)stantial rise (all workers) Moderate rise (all workers) Moderate rise Substantial rise (blue-collar, rnanuhct~~ring) Small rise Substantial rise No change or- substantial fall B. Instrumental I'ariables arid Other Simultaneous Equation Rlethods QoL Farber (1980) Duncan and Stafford ( 1 980) Robinson and Tomes (1984) QoL Leigh (1978) CPS Podgul-sky (1980) NLS Y o t ~ n gMen PSID Substantial rise (all workers) \loderate rise (all workers) Moderate rise (hourly paid workers) Rise (2oung men); fill (older men) Highly ~ a r i a h l e NOTE.-NLS: Ndtlonal Longlrud~nalS u r \ e \ (Cnlted States), SEO S u r t r * of Fconomlc Opportun~tk1L n ~ t e d States), PSID. Panel Stud) of Income D)narnlc\ (L'nlted Surer), QOI Quallt\ of Llfe Surte, iLanada). S\( tl S u n e v of Work Histor) (Canada), LCWS. Labour Canada \cages Sur\e* ((,anadd), I'lC r>KD L n e n ~ p l o ~ r n r n t Insurance C o m m ~ s s ~ odata n ~ l t matched h Satwnal Re\enue Department data ((.anadd). QOF Qualll, of Frnplo\ment ( L n ~ t e dStates), CPS Current Populat~onSurbei ( L nlted States) 1985). Few investigators appear to doubt that union status is endoge~louslydetermined. There is, however, major disagreement about what has been learned from attempts to deal with the endogeneity problem. It has bee11 forcefully argued by Freeman and Medof'f' (1982) and Lewis (1986) that the 11' and IM methods yield no discernible pattern relative to OLS. This may be true for data with any unit of observation, but for the large individual-level data sets there is considerable uniformity. Table 4 displays the estimates from the studies using such data sets listed in Freeman and Medoff (1982, table 2) and Lewis J O L K N A I OF I'OI.ITIC A L E(,O~YO.M\ 638 (1986, tables 4.1,4.2),together with some post-1982 studies, using the I\.' or IS1 method. Except for Podgursky's (1980) \iork with the Current I'opulation Survey files, these studies provide substantial evidence of a corisistent rise in the union differential relative to OLS estimates when the endogeneit) of' union status is addressed by the I\' o r Ihl method. T h e estimates presented in Section I11 above for the current data set reinforce this pattern fol-I\.' ant1 IhI estimators. T h e longitudinal estimator approach to solving the union endogeneity problern has also exhibited considerable ~ ~ n i f o r m i tof' y results. Ho~vever,in this case it is to r r d z ~ cthe ~ estimated differential relative to OLS I-ather than to increase it. Again, the results fhr- the current data set presented above conform to this pattern. I'he reduction found for the longitudinal estimate relative to OLS in table 3 is comparable to the magnitude reported in Freeman (198.1) and the studies cited therein. Some investigators using longitudinal data have attributed this reduction to fixed effects of higher-quality workers being in unions. However, this is not consistent with the cross-section studies that deal with such effects b y the I\.' o r IhI method. V. Reconciling the Results of Alternative Approaches to the Union Endogeneity Problem T h e conflict reported in Section 111 and, as argued above, found more generally in the literature was investigatetl further using the longitudinal sample. If the longitudinal results are to be reconciled ~\.iththe cross-section corrections hi. endogeneit), tlifferenceinduced Ineasurenient error o r selection effects must reverse the direction of' endogeneit) bias from positive to negative. Freeman (1984) has provided some evidence to suggest that while the number of rnisclassified union members may he small enough to produce n o discernible effect on cross-section estimates based on levels, it is large enough to have a major impact on longitudinal estiniators based o n differences because of the relatively small nunil)er of union status changes. If measurement error of the magnitude reported b) Freeman is generally present, then even if' selection effects in cross sections result in differentials larger than OLS differentials, sufficiently large difference-induced measurement error could produce an estimate lo\\-er than OLS cross-section estimates. Further, even if selection effects result in "corrected" estimates of ( y ( , - y,) based on cross-section or lrz~rls-basedrstiniates being larger than 0 1 3 estimates, the same selec- '' I 1 For- a n an;ll>sis of the PI-obleni of measurement rrl-01- i ~ cr lichotomous \alial~lcs. Tee, e.g., Cochr-an il!If 8) and Xigrler- ( lLI'i3). tion process may )ield dzffrrenced-based corrected estiniates of (yl y,.) that are biased downward arid may be lo~verthan O1,S estimates. This issue is addressed in the last t ~ . orows of panel F in table 3. T o examine tlie measurement error and selection issues iri the longitudinal estimates, the within-group regression as reestimated in two forms. First, an instrument was used tor the change in union status. T h e individual mean-differenced observations for union status Fvere regressed o n the union status regressors in mean-differenced form and tlie predicted values m-ere used as instruments. Second, the regressors in the rvithin-group regression b.ere augmented b) the inclusion of the inverse Slills I-atio terms. T h e IV estiniate is 22.7 percent and the IhI estiniate is 16.2 pel-cent. Both of these estimates are in excess riot only o f t h e staridard within-group estimate but also of the OLS estimates in panels D and E. T h e IV estimate is in fact very similar to the IV estimates in panels D and E. T h e I'LI estimate is lo\\.er than the IS1 estirrlate in panels D and E. T h u s there is evidence in these data that measurement error o r selection effects result in a downward bias in tlie longitudinal esti~nator. T h e results above should be vie\\.ed with caution since the longitudinal data set of hourly paid workers is not large arid is sub-ject to some potential attrition bias pr-oblems that, though perhaps mitigated, are not necessarily solved by the inclusion of the conditional disturbance terrns in the analysis. Ho~+.ever, they d o provide some evidence fol- a reconciliation of cross-section and longitudinal results. They are also in conformity rvitli a study by Chowdhury and Nickell (1982), cited in Freenian (1984), \\.ho also found a substantial increase in their longitudinal estimate ( f r o ~ n.10 to .30) when an instrument was used for union status. T h e results d o not support tlie "hounding" results of Freeman. This assumes that agent-induced endogeneit) o r selection will bias OLS cross-section estimates upward, \\.hich conforms rvith many researchers' priors, and quite persuasive arguments have heen advanced in its favor. At the sirnplest theoretical level, these arguments are based on the notion that "better" workers \\.ill want to join unions. It might also be argued that the large differentials i.eported in the corrected results suggest large rents that should induce widespread and obvious queuing activity that, it is suggested, is not universally observed. T h e results in table 3 d o sho~vlarge differentials and do suggest that the corrected differentials are larger than the OLS differentials. In the renniinder of this section a structure is outlined that is consistent with these results and that at the same tirne attempts to deal ~+.ith tlie "bounding-result" arguments sketched above. It appears to he quite generally agreed that union status is not exogenous. 111 that case, what corrected estimates attempt to measure is the outcome of 660 J O U R N A L O F POI.ITICAI. E C O N O ~ I Y an experiment conceptually different from what OLS estimates measure. In particular, corrected estimates measure the difference in earnings (conditional on the exogenous variables) of individuals randornl) assigned to the union and nonunion sectors. T h e OLS estiniates measure the difference given the nonr-andom outcome of the sector assignment process. T o simplify the comparison, consider an economy of individuals with exogenously endowed human capital that is unobserved. (Assunie no observed exogenous variables.) T h e endowment of human capital permits certain tasks to be performed. Let the nonunion sector be characterized by production processes in rvhich individual outputs are readily observed, and one in which there is scope for individual "initiative" to be productive. Let the rvage rate then depend on the individual output: In = E Z , Let the union sector be characterized by production processes at the other extreme rvhere (for simplicity) a teani structure for the process renders all operatives o n the production line equally pr-oductive. Let the wage I-ate in the union sector thus be the same for all: In W [ , ,= EL,, = EL.. Consider the union differential estimated by one of the methods (e.g., IM) that correct for the endogeneit) of union status. This differential corresponds to that obtained from a randotn assignment to sector. T h e expected value of this differential is T h e differential obtained from OLS would be the difference between the average wage of those actually in the union sector and the average of those in the nonunion sector. Its expected value is T h e expected difference betmeen the corrected estimator and OLS is therefore This will be positive, provided E(E.,.~C', = 0) > EE,\.. There are two mechanisms that determine E(E,,.~C', = 0). T h e first is the choice by Lvorkers of offering themselves to the union sector. This ensures that E , ~ ,< . This mechanism therefore tends to raise E(E,,.\ C', = 0) by removing the upper tail off'(^.,.) from the union queue. T h e second mechanism is the choice by eniployers of workers from the union queue. Given the simplifying assumption that all workers are equally productive in the union sector, so that employers would be indiffer- C S I O K IVAGE EFFECTS FIL. 1.-Urliorl tents a n d differentials ent about who works in the union sector, the first effect would determine the outcome. If nwrkers were randomly selectetl fronn the queue o r those with the most to gain are most likely to be in the union, then E ( E , ~ ?CT, I = 0) > E E , ~ T . h e outcome of this case is an estirrlated union differential that is larger \\.hen endogeneity is accounted for than the one that occurs with OLS. Even if employers choose only the ~ i o r k e r s with the highest values of E . ~fl-om , the queue, it may still be the case that E ( E , ~ ,CT, I = 0) > E E .because ~ of the removal of the upper tail off‘(^,,) by the first mechanism. Consider next the problem of rent size and queues. T h e differential estimated by correction methods is not an estimated rent for the marginal individual. Suppose that in equilibrium E I - > E:, where E,: = rnin(es, I C', = 0). T h e n this individual ~\.ouldgain E( - E,: by joining the union sector, hut he is not permitted to d o so because of the queue rationing process. This is also the maximutr~rent that could be earned by any individual by assumption. H o ~ r e v e reven , this rnaxirnum rent map actually be smaller than the corrected union differential, as illustrated in figure 1. If those with t h r largest gain enter "first," the fraction unionized is F(E,$), \\.here F ( . ) is the cumulative distribution function of'^.^. T h e union wage is E ~ . T. h u s the corrected differential is length b; the OLS differential is length a , and the maximum rent is c, where c < b. T h e average gain for those indi\,iduals rlot in the 662 JOURNAI O F POI.ITl(,.AI. ECONOhl\ union, but \\.ishing to join, is This highly si~nplifiedexample illustrates that, given a nonhierand E , ~ the , corrected estimates of union archical interpretation of differentials may be higher than the 01,s estirnates. hloreover, the concerns over apparently large rents to union rne~nbershipassociated with the large differentials estimated by endogeneity-corrected methods may he unfounded since these differentials will not in general he rents in the standard sense. VI. Testing Alternative Hypotheses Regarding the Union Process T h e results of the earlier sections, along ~ i t hother results in the literature, reject the hypothesis of exogeneity for union status. T h e task remains of identifying the specific endogenous process. As noted in Section 11, the IV and I M estimators are based on different specifications of the underlying process. X comparison of these estimators may therefore provide evidence on some of these processes. A Wu-Hausman test \%.asp e r h r m e d on a comparison of the union coefficients estimated by IS1 and IV for the zero slope interactions model. T h e results shotv n o significant difference in the 1x1 and IV estimators. For 1979 the difference between the coefficients was .0150 with standard. error ,0605. For 1981 the difference was .012f ~ . i t h standard err-or ,0682. T h e similarity of'the IV and IM estimates suggests that the simpler fixed effects (of union status) framework may suffice for the union problem. T h e IV estimator is consistent in the fixed effects framework and also in the random effects frarne~%.ork under uncertainty, \\.hich yields the same econonletric properties as the fixed effects framework. T h e IhI estimator is consistent under normality. Consistency of the IV estimator follows if \', satisfies E1',IZ, = E ( E [, ( C ,= l)Pr(C, = 1, Z,) Under normalit\ this become, 4 E;(E,\,IC',= O)Pr(C, = 0, Z,) = k. (Xi3 UKION \VAGE EFFECTS Thus, in general, E17,1Z,depends on Z, unless a(€(, E) that case, howe\ er, k = 0 and = a ( € \ E, ) ; in T h e select~onterms must therefore be opposlte In sign. T h ~ follor\.s s ~mnied~atel\, in the fixed effects fr'inieuork where E I = E \ for all I T h e restriction k = 0 mav be tested b\ exnmirling the equl\ d e n t null h\ pothes~c: For 1979, D = ,2544 with standard error .07 17; for 1981, L) = ,2445 rvith standard error ,0566. Since this test is carried out under the assunlption of union endogeneity, the standard errors used are those obtained from the covariance matrix corrected tor heterogeneity and estimated inverse Mills terms. Thus, for both 197'3 and 1981, k > 0. Positive selection into both sectors, under no1-rnality, however, is not compatible with EI;,IZ, = El', and hence is not con1p;itible with a consistent I V estimator based o n Z,. An alternative interpretation of the similarit) of the I V and IS1 estimates is that normality fails but nevertheless approximates the true distribution. Under normality, the conditional distul-bances are pr-oportional to the hazard f'(.)!F(.) or f ( . ) /[ 1 - F(.)]. Suppose instead that E(E( I CT, = 1, Z,) 1 C', = o, z,) = = aI Pr(CT, = 1 , Z,) and E(E\ a\ Pr(C', = 0, Z,)' '4s \\.ith f ( . ) ! F ( . ) , l!Pr(C', = 1, Z,) decreases as Z, is varied to increase the probability of sanlple inclusion. If this holds, so that I V will be consistent. For the control function approach, all that is needed for consistency is a consistent estimate of Pr(CT, = 1 , Z,). Since the distribution is unknown, this may be estimated nonparanletrically. To examine this interpretation, nonparanietric estimates of Pr(Cr, = 1 , Z,) were obtained, and the selection terms [Pr(C', = I)]- and [Pr(C', = O ) ] for union and nonunion workers, respectively, were used for the control f ~ ~ n c t i oapproach n in place of the inverse YIills ratio terms.'"he results of' this analysis are presented in table 5 . '' T h e nonp;iramrtric estimates rvere obtained trorn a progt-an1 \\-I-itcen),t Paul KIIstone that use5 a kernel eqtlmatlon approach described in Ullah a n d S i ~ ~ g(1989) li - - E c t ~ r n a t ~ oRlethod n 1970 cross section: Nonparanletric control function 1x1 I \' 198 1 c l - o s section: Nonpara11irt1-ic contl-ol function I \1 I\' 9\ k 4; 1 .04 13 1 04!)0 .02(i5 1 Oti78 ... 1.0433 1 .(I027 ... . . ... ... 1 .O<IXCI .08.57 1. 12!)0 ... ... ... 1.1520 -%-?? ,3591 .2$11)3 ,284.5 .33 8 .5 ,2706 ,2579 Under this interpretation, I\' estimation of equation ( 1 1) yields a consistent estinlate of y.\. = (y.,, + k ) as the constant term. T h e control fr~nctionyields an estimate of' y,,, as the constant term and k = a ( , q v as the sum of the coefficients o n the selection ternis in the union arld nonunion sectors. T h e control function and IV estimates may thus be compared for co~npatibilityin this regard, together with their respecti\-e estimates of the rnairl parameter of' interest ( y ( . - y,,,). T h e results in table 5 show that the new controi function estimates are compatible with the IV estimates, both ~\.ithrespect to the estimate o f t h e union dif'ferential (7,. - y,,,) and \iith respect to the predicted difference in the constant terms. T h e esti~natesof k for both years are positive. T h e difference between the IV constant, y;,, and the nonparametric control function constant, y,,,, should thus be positive. This is true for both 1979 and 1518 1 data. For 1979 the estimate of k accounts for one-half of the difference in the constant terms; for 198 1 the estimate of k accounts f i ~ four-fifths r of the constant term difference. For both years the estimates of the union differential (y,, - y,,,) are close. Finally, conlparison with the IM control function estimator indicates that the selection terrns computed f i x this esti~natorl\.ere sufficiently close approxi~nationsto the alternative selection terms to yield similar results for both (y,. - y , ) and y,\ . T h e positive values for k are also consistent ~\.iththe tests for D > 0 above. + VII. Conclusions O n surveying the recent literature on union differentials, Lewis (1986) reached the pessirllistic conclusion that, in essence, nothing had heen learned about union effects on wage rates from the many scridies designed to deal with the endogeneity of' union status. T h e same conclusion is reached in Freeman and Medoff (1982). T h e basic U N I O K WAGE EFFECTS 665 objection of these authors to the "other-than-OLS" studies is that they exhibit no stable pattern, that they fluctuate ~\.ildly,and that therefore they are unreliable. In this paper, existing evidence has been reviewed and new evidence provided that supports a much more positive view of what can be learned fronn the attempts to deal with union endogeneity. With the large individual data sets, there is a strong pattern to the "corrections" for endogeneity. T h e unconditional differential (for instance, that arising fronn workers randomly assigned to sectors) estimated by correction methods is systematicall) larger than the conditional (on union status) differentials estimated by OLS. This pattern holds ~\.hetherthe attempt to deal with endogeneity uses cross-section methods such as I V o r control function approaches o r longitudinal estimators. T h e results presented in Section 111 imply that the cross-section OLS estimates are biased downward hy selectiorl effects. This conflicts with the priors of man) researchers, on the basis of the notion that "better" workers will be picked from a queue by unionized firms. This, however, will result only in an upward bias for OLS under a hierarchical notion of omitted "abilit)"; if this is abandoned in favor of a comparative advantage specification, OLS may be biased downward. T h e idea that the corrected estimates are larger than 0 1 , s estinnates is also considered a problem by many researchers because of the large implied "rents." If union status I\.as exogenous, large estimated difftrentials would be very troublesome because of their association with rents in that setting. Large rents, especially in the absence of obvious and widespread queuing activity, would clearly be difficult to explain. However, if union status is endogenously determined, the relation between the estinlated differentials and rents in the standard sense is much more complex: in particular, a large differential does not imply large rents. Finally, the Canadian longitudinal data used in Section I11 were also employed to construct extensive tests on the exogeneity of union status. With exogeneity rejected, Section \.'I exploits the conflicting properties of different estimators to examine hypotheses on the process governing union status. An insignificant difference bet~veenIV and IM estimators suggests that simple "fixed effects" (of union status) models may be good approximations for the union problenn. References Ahowd, John M., and Farber, Henry S. "Joh Queues and the Union Status of Workers." Indus. and Lubor Relations Rev. 35 (April 1982): 334-67. Aigner, Dennis J. "Regression with a Binary Independent Variahle Subject to Errors of Observation." J. Econom~trics1 (March 1973): 49-59. Atrostic, B. K. "Alternative Compensation Measures: Effects o n Estimates of' \Yage a n d Compensation Differentials." h'lanuscript. Washington: Bur. Labor Statis., November 198 1. Blau, Francine D., a n d Kahn, Lawrence hl. "Job Search and U n i o n i ~ e dEmployment." Econ. Inquiq 21 (July 1983): 412-30. Chamberlain, Gary. "h~lultivariate Regression hlodels for Panel Data." J. econometric^ 18 (January 1982): 5-46. Cholvcihury, G., and Nickell, Stephen J . "Individual Earnings in the U.S.: Another Look at Unionization, Schooling, Sickness and Unemployment Using Panel Data." Discussion Paper no. 14 1. London: London School Econ., Centre Labour Econ., November 1982. Cochran, W. G . "Errors of hleasurement in Statistics." Trchnom~tl-icc 10 (November 1968): 637-66. Duncan, Greg J . , a n d Stafford, Frank P. "Do Union hlembers Receive Cornpensating Ll'age Differentials?" A . E . R . 70 ( J u n e 1980): 3.35-71. Duncan, Gregory hl., a n d Leigh, Duane E. "\l'age Determination in the Union a n d Nonunion Sectors: '4 Sarnple Selectivity Approach." Indus. crnd Labor Relations Rev. 34 (October 1980): 24-34. . " T h e Endogeneity of Union Status: .An Empirical Test." J . Labor Econ. 3 (July 1985): 385-402. Farber, Henry S. "Unionism, Labor Turnover, a n d LVages of Young hlen." In Rrsrclrch i n Lubor Economics, vol. 3, edited by Ronald E. Ehrenberg. Greenwich, Conn.: J A l , 1980. . "The Determinatio~lof the Union Stat~lsof Lt'orkers." Eco~~omrtrica 51 (Septernber 1983): 1417-37. ( a ) . "Worker Preferences for Uniorl Representation." I n RPtm,-ch in Labor Econornicc, suppl. 2, S e w Approtichrs to Labor C'nions, edited by l o s e p h D. Reid, J r . Greenwich, Conn.: J A I . 1C183. (6) Freernan, Richard IS. "Longitudinal Analyses of the Effects of T r a d e Unions." J. Labor Econ. 2 (January 1984): 1-26. Freernan, Richard IS., a n d hledoff, James L. " T h e Impact of Collective Bargaining: Can the New Facts Be Explained by Monopoly Unionisnl?" \\'arking Paper no. 837. Cambridge, hlass.: NBER, January 1982. Grant, E. K.; Swidinsky, R.; a n d Vanderkamp, J . "Estinlates of Union LVage Gaps Using Canadian 1,ongitudinal hlicro Data, 1969-1970.'' Manuscript. Guelph, O n t . : Univ. Guelph, April 1085. Hausrnan, Jerry A. "Specification Tests in Econometrics." Economrtric(~ 46 (November 1978): 1251-71. Hausrnan, Jerry A., a n d Taylor, William E. "Panel Data and Unobservable Individual Effects." E ~ o n o m ~ t r i c49 u (Novernber 1981): 1377-98. Heckrnan, James J., a n d Neumann, George. "Union \Vage Differentials a n d the Decision to Join Cnions." hlanuscript. Chicago: Univ. Chicago, 1977. Heckman, James J . , a n d Robb, R. "Alternative hlethods fbr Evaluating the Irnpact of Interventions." In Lor~gitrrdinctl Ar~aljsis of Lnbor .Ilarkr.t Ilcrtcc, edited by Jatnes J . Heckrnan and Burton Singer. Econometric Society Monograph Series. New York: Cambridge C'niv. Press, 1985. Jakubson, George H . "'I'he Effects of Unions o n \Vages: Estirnation f r o ~ n Panel Data." hlanuscript. Ithaca, N.Y.: Cornell Univ., 1984. Kurnar, P., and Stengos, T. "hlicro Estimates of the Union-Nonunion \\'age Differential in Canada, Using the Sarnple Selectivity Approach." hlanu- script. Kingston, Ont.: Queen's Univ., 1985. Lee, Lung-Fei. "Unionism a n d \Vage Rates: .A Simultaneous Equations hlodel UNION W A G E EFFECTS 667 with Qualitative a n d Limited Dependent \'ariables." I n ( ~ 1 - n u /k;co~l. . R PLJ.154 (June 1978): 413-33. Leigh, Duane E. "An Analysis of the Interrelatio~ibetween Unions, Race, and il'age a n d Kon-Ll'age Con~pensation."Final Report u n d e r Research a n d Development Grant no. C)1-3:'I-77-06, Employment and Training Admiriistration. il'ashington: Dept. Labor, April 1978. . "Racial Differentials in the Union Relative il'age Effects: A Simultaneous Equations Approach." J. I.crbor Res. 1 (Spring 1980): 93-1 14. Lewis, H . Gregg. C'lzio,~ R ~ ~ l c t t i z I~ ~r - ( I ~E'fl~rt.\: P :I S Z O Y ~ ~ (:liicago: JJ. Univ. Chicago Press, 1986. hlincer, Jacob A. "Union Effects: Ll'ages, Turnover, a n d J o b Training." Working Paper no. 808. Cambridge. Mass.: NBER, November 1981. Nerlove, hiarc. "A Note o n Error (:omponents \lotlels." Etotlorr~rtricn 35) (March 1971): 383-96. Podgursky, hiichael J . "Trade Unions a n d Income Inequality." Ph.D. dissertation, Univ. il'isconsin-hladis011, 1980. Robinson, Chris. "Union Endogeneity a n d Self Selection." J. 1.crbor Erotc. 7 (1989), in press. Robinson, Chris, a n d Tomes, Nigel. ''Union il'age Dif'f'erentials in the Public and Private Sectors: A Sirn~lltaneousEquations Specification." J. 1.11bo1 Econ. 2 (January 1984): 106-27. Simpson, il'ayne. "The Impact of Unions o n the Structure of (:anadinn il'ages: An Empirical Analysis w ~ t hhlicrodata." (;cr~lc~clicru J. ~ . : I ~ I I . 18 (Fellruary 1985): 164-8 1. Ullah, A . , and Singh, R. S. "Estimation of a Probability Density Fu11~tio11 \\.ith Applications to Nonpararnetric Inference in Econorrletrics." In Corti~iblrtions i n Econornrtric.r cold .Ilorlt,llirlg. edited b) B. Raj. Boston: Klu\ver, 1989. Ll'essels, ivalter J . "Econornic Effects of Right to il'ork Laws.",/. I.(rbo~ Krs. 'L (Spring 1981): 33-73. and E s t ~ m a Zellner, Arnold; Krnenta, J a n ; and D r e z e . J a c q ~ ~ e"Specification s. tion of Cobb-Do~lglasProduction Function hiodels." k;cono~nrtl-irct:'I4 (October 1966): 784-93. http://www.jstor.org LINKED CITATIONS - Page 1 of 3 - You have printed the following article: The Joint Determination of Union Status and Union Wage Effects: Some Tests of Alternative Models Chris Robinson The Journal of Political Economy, Vol. 97, No. 3. (Jun., 1989), pp. 639-667. Stable URL: http://links.jstor.org/sici?sici=0022-3808%28198906%2997%3A3%3C639%3ATJDOUS%3E2.0.CO%3B2-D This article references the following linked citations. If you are trying to access articles from an off-campus location, you may be required to first logon via your library web site to access JSTOR. Please visit your library's website or contact a librarian to learn about options for remote access to JSTOR. [Footnotes] 6 The Endogeneity of Union Status: An Empirical Test Gregory M. Duncan; Duane E. Leigh Journal of Labor Economics, Vol. 3, No. 3. (Jul., 1985), pp. 385-402. Stable URL: http://links.jstor.org/sici?sici=0734-306X%28198507%293%3A3%3C385%3ATEOUSA%3E2.0.CO%3B2-I 10 Panel Data and Unobservable Individual Effects Jerry A. Hausman; William E. Taylor Econometrica, Vol. 49, No. 6. (Nov., 1981), pp. 1377-1398. Stable URL: http://links.jstor.org/sici?sici=0012-9682%28198111%2949%3A6%3C1377%3APDAUIE%3E2.0.CO%3B2-3 14 Errors of Measurement in Statistics W. G. Cochran Technometrics, Vol. 10, No. 4. (Nov., 1968), pp. 637-666. Stable URL: http://links.jstor.org/sici?sici=0040-1706%28196811%2910%3A4%3C637%3AEOMIS%3E2.0.CO%3B2-C References NOTE: The reference numbering from the original has been maintained in this citation list. http://www.jstor.org LINKED CITATIONS - Page 2 of 3 - Errors of Measurement in Statistics W. G. Cochran Technometrics, Vol. 10, No. 4. (Nov., 1968), pp. 637-666. Stable URL: http://links.jstor.org/sici?sici=0040-1706%28196811%2910%3A4%3C637%3AEOMIS%3E2.0.CO%3B2-C The Endogeneity of Union Status: An Empirical Test Gregory M. Duncan; Duane E. Leigh Journal of Labor Economics, Vol. 3, No. 3. (Jul., 1985), pp. 385-402. Stable URL: http://links.jstor.org/sici?sici=0734-306X%28198507%293%3A3%3C385%3ATEOUSA%3E2.0.CO%3B2-I The Determination of the Union Status of Workers Henry S. Farber Econometrica, Vol. 51, No. 5. (Sep., 1983), pp. 1417-1437. Stable URL: http://links.jstor.org/sici?sici=0012-9682%28198309%2951%3A5%3C1417%3ATDOTUS%3E2.0.CO%3B2-7 Longitudinal Analyses of the Effects of Trade Unions Richard B. Freeman Journal of Labor Economics, Vol. 2, No. 1. (Jan., 1984), pp. 1-26. Stable URL: http://links.jstor.org/sici?sici=0734-306X%28198401%292%3A1%3C1%3ALAOTEO%3E2.0.CO%3B2-H Specification Tests in Econometrics J. A. Hausman Econometrica, Vol. 46, No. 6. (Nov., 1978), pp. 1251-1271. Stable URL: http://links.jstor.org/sici?sici=0012-9682%28197811%2946%3A6%3C1251%3ASTIE%3E2.0.CO%3B2-X Panel Data and Unobservable Individual Effects Jerry A. Hausman; William E. Taylor Econometrica, Vol. 49, No. 6. (Nov., 1981), pp. 1377-1398. Stable URL: http://links.jstor.org/sici?sici=0012-9682%28198111%2949%3A6%3C1377%3APDAUIE%3E2.0.CO%3B2-3 NOTE: The reference numbering from the original has been maintained in this citation list. http://www.jstor.org LINKED CITATIONS - Page 3 of 3 - A Note on Error Components Models Marc Nerlove Econometrica, Vol. 39, No. 2. (Mar., 1971), pp. 383-396. Stable URL: http://links.jstor.org/sici?sici=0012-9682%28197103%2939%3A2%3C383%3AANOECM%3E2.0.CO%3B2-D Union Wage Differentials in the Public and Private Sectors: A Simultaneous Equations Specification Chris Robinson; Nigel Tomes Journal of Labor Economics, Vol. 2, No. 1. (Jan., 1984), pp. 106-127. Stable URL: http://links.jstor.org/sici?sici=0734-306X%28198401%292%3A1%3C106%3AUWDITP%3E2.0.CO%3B2-%23 The Impact of Unions on the Structure of Canadian Wages: An Empirical Analysis with Microdata Wayne Simpson The Canadian Journal of Economics / Revue canadienne d'Economique, Vol. 18, No. 1, Econometrics Special. (Feb., 1985), pp. 164-181. Stable URL: http://links.jstor.org/sici?sici=0008-4085%28198502%2918%3A1%3C164%3ATIOUOT%3E2.0.CO%3B2-8 Specification and Estimation of Cobb-Douglas Production Function Models A. Zellner; J. Kmenta; J. Drèze Econometrica, Vol. 34, No. 4. (Oct., 1966), pp. 784-795. Stable URL: http://links.jstor.org/sici?sici=0012-9682%28196610%2934%3A4%3C784%3ASAEOCP%3E2.0.CO%3B2-Y NOTE: The reference numbering from the original has been maintained in this citation list.