the puzzle of non-participation in continuing...

The Puzzle of Non-Participation in ContinuingTraining Ð An Empirical Study of Chronic vs.Temporary Non-Participation*

Uschi Backes-Gellner, Johannes Mure and Simone N. Tuor

Although participation in continuing vocational training is often found to be associatedwith considerable individual benefits, a puzzlingly large number of people still do nottake part in training. In order to solve the puzzle we distinguish between temporaryand chronic non-participants. Previous studies have shown that training participantsand non-participants differ in unobservable characteristics and therefore self-selectinto training or not. We show that even non-participants cannot be treated as a homo-geneous group: there are those who never take part in training (chronic non-partici-pants) and those who are not currently taking part (temporary (non-)participants).Using a unique data set of non-participants commissioned by the German “ExpertCommission on Financing Lifelong Learning” and covering a very large number ofindividuals not taking part in training, we separate and compare chronic and temporarynon-participants. By estimating a sample selection model using maximum likelihoodestimation we take potential selection effects into account: temporary (non-)partici-pants may be more motivated or may have different inherent skills than chronic non-participants. We find that chronic non-participants would have higher costs than tem-porary (non-)participants and their short-term benefits associated with their currentjobs would be lower. However, in the long run even chronic non-participants wouldbenefit similarly from participation due to improved prospects on the labor market.The results indicate that chronic non-participants either misperceive future develop-ments or suffer from an exceptionally high discount rate, which in turn leads in theirview to a negative cost-benefit ratio for training.

* This paper was released for publication in May 2007.

Contents

1 Introduction

2 Theoretical framework: the investment decision

3 Estimation model and methods

3.1 Costs

3.2 Benefits

4 Data and descriptive statistics

5 Econometric results

5.1 Costs

5.2 Benefits

5.3 Participation decision

6 Conclusions

References

Appendix

ZAF 2 und 3/2007, S. 295Ð311 295

The Puzzle of Non-Participation in Continuing Training Uschi Backes-Gellner, Johannes Mure and Simone N. Tuor

1 Introduction

Lifelong learning and continuing vocational trainingare becoming more and more important Ð a trenddriven primarily by demographics and rapid techno-logical change. On the one hand, the decline in birthrates and an aging workforce may cause a shortageof skilled labor, strengthening the importance ofcontinuing vocational training (Bellmann 2003). Onthe other hand, technological change is stronglyskill-biased, shifting labor demand toward more ed-ucated workers.1 Employees with low levels of edu-cation run the risk of earning lower wages or, evenworse, of being crowded out of the labor market be-cause many jobs are disappearing (Spitz-Oener2006). All this should cause a strong incentive toparticipate in continuing vocational training. More-over, a large number of studies have shown that par-ticipation in continuing vocational training boostsindividual wages.2 Recent research has also foundnon-monetary benefits, such as a reduced risk of be-coming unemployed.3 However, many individualsstill refrain from participating in continuing voca-tional training. Even individuals who leave schoolwith few or no qualifications, and whose only wayto catch up would be to participate in continuingtraining, often decide not to participate in anycontinuing training measure (Schröder/Schiel/Aust2004). Furthermore, Groot/Maassen van den Brink(2003) point to the existence of different trainingtracks, where workers not participating in furthertraining in one year are likely to belong to the groupof non-participants in the next year as well.

The question we raise in our paper is how the puzzlecan be solved, i. e. why is it that we find high returnsto training for all participants on the one hand, anda large number of people not participating in train-ing on the other hand. We argue that the solution isto be found in unobservable characteristics of twodifferent groups of non-participants, namely tempo-rary and chronic non-participants. This distinctionhas not been used before and the results are promis-ing.

There is vast empirical literature devoted to estimat-ing the (positive) impact of training, e.g. on the par-ticipants’ wages, but only a few empirical papers an-alyze the rate of return that non-participants wouldhave if they had taken part in continuing vocational

1 Chennells/Van Reenen (1999) survey the evidence on the corre-lation between technology and skills and come to the conclusionthat the technological change is skill-biased.2 See Frazis/Loewenstein (2005) for US; Bassanini et al. (2005)and Pfeiffer (2000) for Europe.3 Büchel/Pannenberg (2004).

296 ZAF 2 und 3/2007

training. Vignoles/Galindo-Rueda/Feinstein (2004)analyze the effect of work-related training on wagegrowth, but focus only on middle-aged male work-ers. They estimate a selection model Ð taking intoaccount that the training decision could be endoge-nous Ð and find substantial selection effects. Thoseworkers who have received training measures gainsubstantial wage benefits, whereas non-participatingworkers would not have gained higher wages if theyhad participated. Groot (1995) analyzes the impactof an investment in firm-related training on wagesand he finds that participants do have a positive rateof return. But the wage gain a non-participant wouldhave received if he had participated is negative.Moreover, Leuven/Oosterbeek (2002) show that thereturn to work-related training is overestimatedwhen self-selection is not taken into account. Thereis also similar evidence in the field of initial voca-tional training. Wolter/Mühlemann/Schweri (2006)find that most of the apprenticeships in Switzerlandgenerate net benefits for the participating firms. Butthey also show that if non-participating firms wereto start training apprentices, they would incur signif-icantly higher net costs compared to the participat-ing firms, which follows from the fact that non-par-ticipants face much more unfavorable cost-benefitratios. Thus, all these results indicate that due to theabsence of sufficient benefits or due to higher costs,not participating in training could be a rational deci-sion. But the question still remains as to the charac-teristics of those individuals for whom it may notpay to take part in training. It can be assumed thatespecially the distinction between different qualifi-cation levels is fundamental. Highly skilled workersseem to capture a larger share of the productivityeffects of continuing training than low-skilled work-ers (Kuckulenz 2006), influencing the cost-benefitratio of participation in further training.

Bassanini et al. (2005) studied the determinantsdriving the probability of receiving training andfound that educational attainment and the skill-in-tensity of the occupation exert a positive impact.There seem to be complementarities between higherlevels of education on the one hand and continuingtraining on the other. So individuals who are alreadydisadvantaged in schooling and vocational educa-tion are more likely not to participate in lifelonglearning later in life. Jenkins et al. (2003) found thatthe more qualified individuals are, the more likelythey are to return to learning later in life. Moreover,the training probability decreases with age, and bothpart-time and temporary workers are less oftenfound in the group of participants (Bassanini et al.2005). In sum, training participants and non-partici-pants differ in observable as well as in unobservablecharacteristics.

Uschi Backes-Gellner, Johannes Mure and Simone N. Tuor The Puzzle of Non-Participation in Continuing Training

In addition, we argue in this paper that non-partici-pants cannot and also should not be treated as ahomogeneous group. We distinguish betweenchronic and temporary non-participants for the firsttime and study their (potential) returns to continu-ing training separately. We are fortunate in beingable to use a unique data set of non-participantswith more than 1200 individuals.4 This survey allowsthe distinction between individuals never taking partin training (chronic non-participants) and individu-als currently not taking part, i. e. in the year of thesurvey (temporary non-participants). According tothe character of the data set our paper aims to studythe differences between the two groups of non-par-ticipants and particularly to find out why some peo-ple never take part in continuing training. The struc-ture of the paper is as follows: the next section cov-ers the theoretical framework, a simple cost-benefitmodel which mainly serves the purpose of structur-ing our empirical analysis and indicating importantexplanatory variables. Section 3 explains the econo-metric models and the estimation methods applied.The data set used is described in section 4. The re-sults of the estimation of costs and benefits are pre-sented in section 5, and conclusions are drawn in thelast section.

2 Theoretical framework:the investment decision

To structure our empirical analysis of individual de-cisions to participate in training, we use standard hu-man capital theory as a framework.5 We assume thatan individual’s participation decision has to be seenas an investment decision. Individuals bear costs ofparticipating in continuing training and in later peri-ods expect various kinds of benefits in return. If dis-counted individual returns exceed individual costs,it pays to invest; otherwise it is rational not to investin training (see e.g. Borjas 2005: 240 ff. or Lazear1998: 136 ff.).6 Or to be more precise, an individualonly invests in continuing vocational training if thepresent value of his or her individual benefits ex-ceeds the present value of his or her individual costs,i. e. if

4 We thank the Expert Commission on Financing Lifelong Learn-ing for collecting the data and for allowing us to use it.5 Becker (1975).6 Although in many cases the company may also bear some oreven most of the costs of a training measure and therefore getssome of the returns, we only concentrate on the part of the coststhat the individual bears and the parts of the returns that theindividual receives. So if the company and the individual splitcosts and benefits, we would just have the individual’s part in ourempirical analysis.

ZAF 2 und 3/2007 297

�T

t�1

Bijt

(1 � r)tý �

t�z

t�1

Cijt

(1 � r)t(1)

If inequation (1) does not hold, it would not be ra-tional for individuals to invest in training. Accordingto this model we expect chronic non-participants ei-ther to have higher costs and/or to have lower re-turns associated with participating in training thantemporary non-participants. Higher costs could bedue to lower ability, bad learning experience, oremotional distress triggered by schooling situations.Lower returns could be due to systematic differen-ces in their workplaces, in motivation, in effort orcareer prospects.

The total costs (C) of an employee7 (i) participatingin training measure (j) in period (t) can be separatedinto direct costs and indirect or opportunity costs.Direct costs include participation fees, expenses forbooks, transportation costs or expenses for child-care, as well as non-monetary costs such as dislikinglearning or negative feelings attached to trainingdue to bad past learning experience. Opportunitycosts include either the forgone salary an individualcould have earned if he had not taken part in continu-ing training (in the case of unpaid leave) or the lossof leisure. Some direct cost components (e.g. partici-pation fees) may be the same for both chronic non-participants and temporary non-participants. Othercost components Ð especially non-monetary directcosts (such as learning stress) Ð may be substantiallyhigher for chronic non-participants. Opportunitycosts could also differ substantially but the overalleffect is unclear ex ante. Compared to chronic non-participants, temporary non-participants may be in-dividuals with higher ability or motivation, they maylearn more quickly and finish the same trainingmeasure in a shorter time. However, their forgoneincome per time unit may also be higher. At thesame time, for individuals with lower income itmight be more difficult to compensate for forgoneearnings, so even lower absolute amounts of forgoneincome would still prevent them from participating,so the direction of the overall effect is theoreticallynot clear but remains an empirical question.

The benefits (B) also have two components: directlymeasurable benefits such as an increase in pay, andindirect benefits such as an increase in job securityor long-term labor market prospects.

Furthermore, since the decision depends on thepresent value of costs and benefits, the remaining

7 Full-time employees, part-time employees, unemployed persons,training participants as well as unemployed persons who wouldlike to start working (again) in the next two years belong to thisgroup.


time in the labor market (T) and the individual dis-count rate play a crucial role. Individuals with a highdiscount rate r are more present-oriented. As a con-sequence they are less likely to accept costs todayto gain benefits at some time in the future. Thiscould for example be the case for chronic non-par-ticipants and would thus explain their non-participa-tion. Taken together, the decision never to take partcan occur either because chronic non-participantscannot afford the associated costs (monetary and/ornon-monetary), because of low short-term benefitsassociated with training participation, or because ofa high discount rate or a low remaining time in thelabor market.8 Since our cost and return data useonly qualitative variables and are not in currencyunits (such as €) we cannot compute quantitativecost-benefit ratios. Therefore, we have to study costsand benefits separately in the following.

3 Estimation model and methods

In this section we present the estimation model forthe costs that result from taking part in continuingvocational training as well as for various benefitsthat result from participation. In all our estimationswe control for potential selection effects.

3.1 Costs

The basic equation we estimate can be written as:

Costs = �0 + �1 · VocTraining + �2 · ProfStatus +�3 · EmpCharacteristics +�4 · IndCharacteristics + δ · X + u

(2)

Since the dependent variable costs cannot be meas-ured directly because our sample consists of non-participants only, we argue that we can use the indi-vidual’s willingness to pay as a proxy providing uswith a good lower bound for training costs. Sinceour non-participants all decided not to take part intraining it seems reasonable to assume that the ac-tual training costs were higher than their individualwillingness to pay which is why they decided not totake part. Accordingly, willingness to pay providesus with the lower bound of actual training costs andcan in this sense be used as a proxy for trainingcosts.9

8 Another potential reason causing a similar result could be thatchronic non-participants lack necessary long-term informationand thus cannot anticipate future benefits or possible job-offerscorrectly.9 Individuals had to specify the amounts they were willing tospend on continuing training out of five categories (in € p.a.). Weuse the upper limit per category as a proxy for the lower boundof actual training costs. See also section 4.

298 ZAF 2 und 3/2007

The independent variable “VocTraining” includesvarious dummy variables for vocational trainingmeasures, such as apprenticeship, full-time voca-tional school, master craftsman and university.Professional status, represented by “ProfStatus”, in-cludes variables indicating different categories suchas blue-collar worker, white-collar worker andself-employed person.10 “EmpCharacteristics” con-sists of variables describing the employment situa-tion or workplace of the individual, including for ex-ample number of employees in the company ordummies such as full-time employee, use a com-puter at work or at home or knowledge and skillneeds change frequently. Furthermore, we use in-dividual control variables (“IndCharacteristics”)such as net income, age, gender and having chil-dren (kids) and we control for industry-specific ef-fects by adding industry dummies (“X”). In a sec-ond step, we add the marital status and interac-tion terms between the individual characteristicsgender and kids as well as between married andkids. Findings from the German Socio-economicPanel (see Bellmann 2003) show that marriedwomen who live in a household with children areless likely to take part in continuing trainingwhereas men in the same situation are more likelyto participate. In a third step, we include a variablerepresenting an individual’s time preference and es-timate the basic and the extended equation again.Controlling for present- or future-orientedness, it ispossible to consider the impact of costs on the par-ticipation decision exclusively. For a full list of thevariables included in the estimation models and therespective descriptive statistics see Table 1 (in theappendix).

A major methodological problem is that the varia-ble costs can only be observed for people who haveat some time taken part in continuing vocationaltraining. In our sample these are the temporary non-participants, who by definition must have taken partin training at some time in the past but not in thesurvey year (strictly speaking they are not only tem-porary non-participants but also temporary partici-pants, which is what we make use of to estimate ourcost variable). So for temporary (non-)participantswe have (past) training costs, but for chronic non-participants such data obviously cannot be available.So we have a sample selection problem or an inci-dental truncation problem as it is called by Woold-ridge (2003: 587Ð591). Whether we observe the (de-

10 We do not include civil servants because they are in a differentsituation as regards wages and job security. One could also con-sider excluding self-employed persons. But, as the results remainstable with or without this category, we decided not to do so inorder to analyze a larger sample of non-participants.


pendent) variable depends on the participation deci-sion and it is therefore known for temporary (non-)participants only. It can be assumed that individualsare not randomly selected into the two groups11, i. e.temporary (non-)participants presumably have sys-tematically different costs than chronic non-partici-pants (for example due to ability differences). Thus,an OLS regression of equation (2) produces incon-sistent and biased estimates of the coefficients. Theapproach we use to solve this problem is as fol-lows:12

First we model the probability of a person being aparticipant (participation equation):

y*2i = z�i γ + u2i (latent variable y*

2i) (3)

where u2i ~ N (0, 1),; we observe y2i = 1 (an individ-ual took part in continuing vocational training) ify*

2i > 0. In this step we use the whole data set (tem-porary (non-)participants as well as chronic non-participants).

In the second step we examine y1i and take into ac-count that y1i is only observed if y2i = 1 (outcomeequation), meaning that in this step we can only in-clude temporary (non-)participants:

y1i = x�i � + u1i (4)

where u1i ~ N (0, σ21) and y1i is the costs associated

with participation in continuing training. The errorterms u1i and u2i are bivariate normal with correla-tion ρ. In our specification we follow Wooldridge(2003: 589), who strongly recommends that x is astrict subset of z and that there is at least one ele-ment of z that is not also in x (Wooldridge 2003:589). Therefore, any explanatory variable in the re-gression equation should also be an explanatoryvariable in the selection equation. Moreover, weneed at least one variable that affects selection butdoes not have a partial effect on y. The expectedvalue of y1i can then be written as:

E (y1i |z�i, y2i = 1) = x�i � + ρλ (z�i γ) (5)

where λ (z�i γ) is the inverse Mills ratio (Wooldridge2003: 588). If there is no correlation between thetwo equations (ρ = 0), then the participation andoutcome equations are independent and the OLSestimates of � would be unbiased. But if there is acorrelation between the participation decision andthe cost determinants, then there is an omitted vari-

11 Leuven/Oosterbeek (2002).12 As the data set used does not have a panel structure, fixedeffects estimation (Wooldridge, 2003: 461Ð467; Wooldridge 2002:265Ð279), which would also be a possible approach to take poten-tial selection effects into account, cannot be used.

ZAF 2 und 3/2007 299

able problem, λ (z�i γ). That is why we first testwhether there is a selection bias. If the null hypothe-sis (ρ = 0) cannot be rejected, there is no selectionproblem and OLS estimations would be appropri-ate, otherwise selection effects have to be taken intoaccount and the model explained above has to beused. To estimate such a model we can either choosethe two-step estimation method recommended byHeckman (1976) or the maximum likelihood estima-tion recommended by Wooldridge (2003: 591). Thelatter is used in this study, as it is asymptotically un-biased, asymptotically normal and more efficientthan the two-step estimator, provided that the ap-propriate assumptions are met (cf. Wooldridge 2003:591, or Breen 1996: 40).13

3.2 Benefits

The following equation is used to assess the benefitsassociated with taking part in continuing training:

P (Benefit = 1 |x) = Φ (�0 + �1 · VocTraining +�2 · ProfStatus + �3 · EmpCharacteristics +�4 · IndCharacteristics + δ · X + u) (6)

where benefit is a binary response variable. Due todata restrictions and in contrast to the cost analysis,we have to estimate the direct and indirect benefitsof training in separate steps. First we use increasein pay as a dependent variable, indicating a directbenefit. To examine indirect benefits, we use improv-ing job security as well as improving employmentoutlook as dependent variables. The independentvariables are the same as in the cost equation and,as in the cost estimations, we have to solve the prob-lem of unobserved heterogeneity. Individuals whohave taken part in continuing vocational trainingmay also be more motivated or may have differentinherent skills and may therefore have higher earn-ings or better job prospects. Thus, the effect of par-ticipation would be overestimated. It is thereforeimportant to use a maximum-likelihood probit esti-mation with selection instead of a simple probit esti-mation. The estimation procedure follows the sameapproach as that used in the case of incidental trun-cation with a metric dependent variable in chapter3.1. The major difference is that there is a binarydependent variable. Therefore the outcome equa-tion is as follows:

y*1i = x�i � + u1i (7)

where u1i ~ N (0, 1) and y1i = 1 if y*1i > 0.

13 There are other studies preferring the two-step estimator, butthere seems to be no common strategy (see Puhani (1997) for asurvey of criticism of the different methods).


4 Data and descriptive statistics

Our empirical estimation is based on a unique dataset commissioned by the German “Expert Commis-sion on Financing Lifelong Learning”. It covers indi-viduals who did not participate in continuing train-ing in the period from September 2001 to August2002.14 The data set contains information from com-puter-assisted telephone interviews with 1264 em-ployees between 19 and 64 years of age and livingin Germany. The sample is a representative sampleof non-participants which resulted as a by-productof a survey of participants commissioned by theGerman Federal Institute for Vocational Educationand Training (Bundesinstitut für Berufsbildung(BIBB)) for a separate project (cf. Beicht/Krekel/Walden 2006). Unfortunately, the participants’ datawere not available for this study, therefore our ana-lysis is restricted to the non-participants sample asdescribed in Schröder/Schiel/Aust (2004). Since ourmain interest lies in differences between chronic andtemporary non-participants in training, we separatethe respondents into those who have never partici-pated in training15 Ð referred to as chronic non-par-ticipants Ð and those who have taken part in continu-ing vocational training at least once in the past,but not during the survey period Ð referred to astemporary (non-)participants.16

Continuing training is defined very broadly in thisstudy. It is any kind of further learning, be it formal,non-formal and/or informal learning or of a generalor vocational nature, after the completion of initialvocational training (Timmermann et al. 2002: 55).Thus our training measure includes formal trainingprograms that may take place in a company or at acontinuing vocational training institute, it also in-cludes informal on-the-job training (such as e.g.quality circles or job rotation) and even self-directedlearning or conference participation.

The training variable is a dummy variable taking thevalue 1 if someone belongs to the temporary (non-)

14 For details of the survey see Schröder/Schiel/Aust (2004).15 We are aware that the period since entering the labor marketand in which there is an opportunity of taking part in continuingvocational training is shorter for younger employees. But on theother hand, there is the inverse impact on the training probabilityof young and older workers caused by differences in the remain-ing time in the labour market. However, as we control for age, theestimated differences in costs and benefits should not be biased.16 Thus, we deliberately do not differentiate between the variousfurther vocational training measures which the data set contains.Moreover, the duration of continuing vocational training mighthave an impact on returns (Budria/Pereira 2004). As we do nothave duration data, we cannot make these differences. But forour main objective, analyzing why some people never take partin continuing vocational training, this is not crucial.

300 ZAF 2 und 3/2007

participants and 0 if someone is a chronic non-par-ticipant.17 About one third of the people inter-viewed (419) belong to the latter category of chronicnon-participants.

Regarding our dependent variables, people wereasked about the costs and benefits associated withparticipation in continuing vocational training. (Afull list of all variables and their overall means andmeans broken down into temporary and chronicnon-participants is given in Table 1.) Descriptive sta-tistics indicate that chronic non-participants rate thebenefits associated with participation in continuingvocational training considerably lower than tempo-rary (non-)participants. On average, chronic non-participants have a lower willingness to pay thantemporary (non-)participants: the former are willingto invest about € 300, the latter more than € 500 onaverage, which we use as our proxy for training costs(because only the people in the latter group reallyknow what they are talking about). The validity ofthe cost proxy that we obtain by this estimation isbacked up by empirical results from the survey ofparticipants given in Beicht/Krekel/Walden (2006).Based on data from actual participants they findthat participants spend an average of € 502 on theirtraining, which is astonishingly close to our proxy,i. e. the amount that temporary (non-)participantsare willing to pay for continuing vocational training,which is € 520. With regard to our explanatory vari-ables, we find that chronic non-participants are char-acterized by rather low or even no qualifications.For example, 20% of chronic non-participants butonly 5% of temporary (non-)participants have nosecondary school certificate. As a proxy for the dis-count rate we use the strength of an individual’spreference for enjoying life and having enough timefor personal interests and leisure. Descriptive statis-tics show that chronic non-participants are far morepresent-oriented than temporary (non-)participants,which indicates that present-orientedness could in-deed have an influence on the individual trainingdecision.

As explained in section 3, it is important to takeselection effects into account. The equation deter-mining the participation decision should contain aselection variable that satisfies two conditions. Onthe one hand the selection variable has to correlate

17 If people had not participated during the period from Septem-ber 2001 to August 2002, they were asked about participation inthe past. Various types of training were specified and people hadto answer for each type of training separately if they had takenpart before the survey period. Thus, the risk of not rememberinga participation and consequently being assigned to the chronicnon-participants should be very small.


with the participation decision. On the other handit should not have an effect on the outcome of inter-est (the cost or benefit variables). Closely related toother empirical findings for Germany which use thenumber of children to identify the decision to takepart in training because children put higher strainson the time budget (Büchel/Pannenberg 2004), weuse the variable “taking part causes too much stressdue to my job and private obligations” for the selec-tion equation. Slightly less than half of the chronicnon-participants (46%), but only a third of the tem-porary (non-)participants (35%) state that this wasdecisive for their non-participation, so the variableis correlated with the participation decision. In con-trast, it is obvious that the stress a person expectsdue to job or family obligations is not related to theoutcome variables such as direct training costs ormarginal benefits resulting from a training measure.If we look at our estimation results, both assump-tions are confirmed: the identification variable hasa highly significant impact on the probability of be-ing a temporary (non-)participant (cf. section 5.3),but this is neither the case for costs nor for benefits.

After eliminating observations with missing data, asample of 527 individuals is left for analysis.18 Ofthese, 163 observations are censored and are there-fore not included when estimating the outcomeequation separately.19

5 Econometric results

5.1 Costs

Table 2 in the appendix (model 1)20 presents the re-sults of estimating the basic equation. The selectionmodel, using a maximum likelihood estimator, sup-ports the view that there is selection into trainingbased on unobservable characteristics: the hypothe-

18 The non-participants survey contains 1264 observations. As al-ready explained, we do not include civil servants. Moreover, infor-mation about costs and benefits associated with training participa-tion is not known for each non-participant. Finally, only inter-viewed persons for whom the vocational training, as well as spe-cific individual and employment characteristics, are known can beincluded in the empirical analysis.19 Employees who did not choose one of the vocational trainingcategories were generally excluded because they form a verysmall and heterogeneous group. Just a few people are familyworkers, employees subject to social insurance contributions orpeople who have not yet worked. They cannot be included be-cause most of them have missing values in the other variables andwith just one or two observations left, it is not possible to provideevidence.20 The tables only provide limited information about the selectionequations. Full tables are available form the authors upon re-quest.

ZAF 2 und 3/2007 301

sis that there is no sample selection (ρ = 0) is re-jected at the 5% level. Thus, sample selection is aproblem and the costs of training vary based on ob-servable as well as unobservable characteristics. Ascan be seen, the selection effect is negative (ρ =Ð0.1219) which means that temporary (non-)partici-pants have lower costs ceteris paribus than chronicnon-participants, which is a highly plausible result.The unobserved characteristics that increase thelikelihood of ever taking part in continuing voca-tional training are associated with lower actual ex-penditures.

Turning to the predicted costs21, the result shownabove is confirmed: on average chronic non-partici-pants pay slightly more than temporary (non-)par-ticipants. The former have a predicted amount ofabout € 504, which is substantially higher than whatthe chronic non-participants are willing to pay (cf.section 4). A very important determinant seems tobe vocational training: employees with an appren-ticeship or a university degree or equivalent havesignificantly lower costs associated with training par-ticipation than employees without a secondaryschool certificate. Income does not have a significantimpact on willingness to pay, so forgone salary doesnot seem to be crucial in the training decision.Therefore, the loss of leisure and the direct costsmust be the major driving force. Individuals with thelowest level of schooling (no secondary school cer-tificate) can be assumed to have the greatest diffi-culty in learning at school and might therefore alsohave a lower ability and less motivation to keep onlearning later in their life because they would needmuch more time and effort than individuals with ahigher qualification level. In addition, they are as-sumed to be much more averse to learning. Includ-ing a variable representing time preferences (model2) does not yield a substantial change. The resultshown above also holds for the models with addi-tional individual characteristics and interactionterms (models 3 and 4). None of the added variableshas an impact on the costs associated with continu-ing participation in training.

To summarize, individual training costs seem to playa major role in the individual participation deci-sions: the costs that chronic non-participants wouldhave to bear are significantly higher than those actu-ally borne by temporary (non-)participants.

21 Predicted costs and benefits are estimated as follows: given theestimated coefficients of the outcome equation obtained by onlyincluding temporary non-participants but taking into account po-tential selection effects, we can plug in every non-participant inthe regression. Thus, we obtain predicted costs as well as benefitsfor each of them.


5.2 Benefits

In our first estimation we use increase in pay as adependent variable (Table 3 in the appendix). Esti-mating the basic equation, the hypothesis that theoutcome and participation equation can be esti-mated separately cannot be rejected (ρ = 0). There-fore, we can use a simple probit model to estimatethe outcome equation. The results are given in Table3, models 1Ð2. We find that the significant variablesgenerally take their expected signs: white-collarworkers are more likely to have received an increasein pay than blue-collar workers, which is highly rele-vant since the latter are more likely to belong to thegroup of chronic non-participants. This is also truefor employees holding a job characterized by fre-quent changes in knowledge and skill needs. More-over, having children is associated with a lower like-lihood of a wage increase. Calculating the marginaleffects at the means of the independent variablesshows that being a white-collar worker (comparedto a blue-collar worker) is associated with a 9.1 per-centage point higher likelihood of a pay increase,and the need for a frequent change of knowledgeand skills with a 12.8 percentage point higher likeli-hood, while having children is associated with an 8.5percentage point lower likelihood of an increase inpay. Moreover, having a master craftsman’s diplomaenters with a significant negative coefficient. Theyare less likely (9.1 percentage points) to have re-ceived a wage increase than someone without a sec-ondary school leaving certificate. The results remainstable when the time preference variable is added(Table 3, models 3Ð4). Including the variables rep-resenting marital status (married and single mother/father) as well as interaction terms between genderand children and between marriage and childrenyields almost the same results with the exception ofthe variable kids, which no longer has a significantimpact (Table 3, last four models).

To summarize, the predicted probability of havingreceived an increase in pay is about 15 percent: tem-porary (non-)participants have on average a higherand chronic non-participants a lower likelihood ofreceiving a pay rise. On the whole, the results are inline with those of Vignoles/Galindo-Rueda/Fein-stein (2004), who find that only those workers whoactually participate in continuing training are ableto realize wage gains. For those workers who did notparticipate in training, participation would not havebeen associated with a short-term wage gain in theircurrent job.

Secondly, we consider the benefit of training interms of job prospects. We first look at differencesin job security. Again, we assume that individuals

302 ZAF 2 und 3/2007

who occasionally take part in continuing trainingdiffer from those who never take part in training,which is supported by the results of the basic equa-tion estimation. There is a positive selection effectof ρ = 0.8448. The hypothesis that there is no sampleselection (ρ = 0) is rejected at the 5% level (Table 4in the appendix, models 1 and 2). Temporary (non-)participants are less likely to lose their jobs as a re-sult of their inherent ability or other unobservedcharacteristics and not necessarily because of theirparticipation in continuing vocational training. Withrespect to job security, being male, using a computerat work and working in a large company are associ-ated with a lower likelihood of becoming unem-ployed. In calculating the marginal effects at themeans of the independent variables we find that us-ing a computer has a statistically and also an eco-nomically significant positive impact of 13.8 percent-age points on the likelihood of increased job secu-rity. Bearing in mind that employees who use a com-puter at work are more likely to take part in continu-ing training than other workers, it can be assumedthat it is particularly important for these people tokeep on learning later in life in order not to losetheir job. Being a male as compared to being a fe-male increases the likelihood by 11.6 percentagepoints. Finally, firm size affects job security, but theeffect is not practically large: if a firm grows by 1000employees, the likelihood of participation in trainingincreases by only 1.1 percentage points. On averagethe probability of increased job security is 21.4%,whereas the predicted probability for chronic non-participants is slightly lower and the probability fortemporary (non-)participants is higher. The differ-ence between the two groups is about 4%. Evenafter adding the variable representing present-ori-entedness, the above-mentioned results remain sta-ble. The hypothesis that there is no sample selectioncan still be rejected, although only at the 10% level(Table 4, models 3 and 4). The variable does nothave a significant impact on job security, but Ð indi-rectly indicating a usually unobserved characteris-tic Ð reduces the influence of unobserved character-istics which usually make the coefficient of a sepa-rately estimated probit model biased. Nevertheless,estimating a selection model seems preferable. Thelast four models in Table 4 differ insofar as somevariables indicating the marital status and interac-tion terms between gender and kids as well as be-tween marriage and kids are added. This has an in-fluence on the significance of some of the “Individ-ual Characteristics” variables, which are worth look-ing at: the likelihood of an increase in job securityfalls with increasing age but at a diminishing rate.The most notable result is the significant positivecoefficient of having children. The marginal effectis very high Ð at 36.5 percentage points (and 39.2percentage points respectively). Turning to the


added variables, we find that being married has asignificant and positive impact on job security (21.8and 22.7 percentage points respectively). But if amarried couple has children, the likelihood of hav-ing a secure job is reduced by 30.9 (and 32.7) per-centage points. Bearing in mind that the coefficientof the variable kids is also very high, families withchildren are still less likely to lose their jobs.

Turning to prospects on the labor market, we findthat a separate estimation of the outcome equationdoes not yield biased estimates of the coefficients:the hypothesis that there is no sample selection (ρ =0) cannot be rejected at the 10% level irrespectiveof the included explanatory variables. As the regres-sions of the models including marital status and theinteraction terms have a higher significance, we turnto the second part of Table 5 (in the appendix). Inparticular, the coefficients of the variables marriedand married with kids are highly significant. Whilebeing married has a positive effect on the likelihoodof increased prospects on the labor market, havingchildren in addition has a significant negative im-pact; taking into account the coefficient of kids,which is positive but not significant, it is no longerclear whether there is really a difference betweenmarried couples with and without children. Interest-ingly, working full-time, which has neither an influ-ence on costs nor on the likelihood of an increase inpay or job security, has a significantly positive effect.Working full-time instead of part-time (or otherforms) is associated with a marginal effect of 16.9percentage points. As with job security, firm size en-ters with a positive and significant coefficient withan effect that is not practically large. The predictedprobability of continuing training as a necessarycondition for increased prospects on the labor mar-ket is 48.5% for temporary (non-)participants and47.7% for chronic non-participants. Almost half ofeach group benefits from participation in continuingvocational training as far as increased prospects onthe labor market are concerned. Therefore, thisseems to be a highly important and non-negligiblefactor.

To summarize, chronic non-participants would notgain as much as temporary (non-)participants interms of increased job security. Chronic non-partici-pants are more likely to be found in unskilled jobs.Therefore, they are most likely not faced with con-stantly increasing requirements at the workplacedue to technological change, but with the situationthat their job may disappear. Thus, considering onlytheir current job, the decision not to take part intraining seems to be a rational decision in the shortand medium term because chronic non-participantswould indeed not gain much in their current job Ð

ZAF 2 und 3/2007 303

and they know it. However, since it is precisely theseworkers who are at a higher risk of losing theirjobs22 it would be important for them to think morein the long term. As we have seen, continuing voca-tional training would be necessary to improve theiremployment outlook. Participation in training couldprovide them with knowledge enabling them to domore complex or even completely different work,which in turn would make it easier to find new jobsonce they lose their current jobs. Thus, informationasymmetries seem to play a crucial role: chronicnon-participants would benefit as much as tempo-rary (non-)participants in terms of employment out-look, but do not seem to realize that in addition toreturns associated with their current jobs (where re-turns are indeed very low) participation in trainingcould also lead to better prospects on the labor mar-ket and could therefore help them to find new jobsif they are laid off.

Our hypothesis on benefits is therefore only partlyconfirmed. Chronic non-participants would indeedhave lower benefits regarding their current jobs,23

but they would still benefit as much as temporary(non-)participants in terms of long-term labor mar-ket prospects. Another important result is thatchronic and temporary non-participants have to bedistinguished in empirical studies because they arerather different in their costs and benefits. There-fore, studies that do not differentiate betweenchronic and temporary non-participation underesti-mate the returns to education for temporary (non-)participants on the one hand and on the other handoverestimate the returns to education for chronicnon-participants. So if, for example, these twogroups are not differentiated, results on returns totraining cannot be expected to be reliable.

5.3 Participation decision

Given the above results it seems important to distin-guish between chronic and temporary non-partici-pants because they are faced with a significantly dif-ferent cost-benefit structure of training. Accord-ingly, the question is what distinguishes the two non-participation groups. Table 6 (in the appendix) pro-vides probit regression results concerning the proba-bility of taking part in continuing vocational train-ing. The first two models of the Table give results

22 Almost one in five unskilled workers was unemployed in Ger-many in 2003 (iwd 2006).23 This result corresponds with Groot’s (1995) finding that non-participants would have a negative wage gain in the case of partic-ipation.


for the basic equation (model 1) and for the basicequation with a variable representing time prefer-ence added (model 2); the last two models provideboth models with some more individual characteris-tics. The following results are similar in all of themodels. The likelihood of taking part in continuingvocational training is significantly higher for a mas-ter craftsman, a worker who has completed an ap-prenticeship or who holds a university degree orequivalent than for a worker without a secondaryschool certificate. The results are consistent withthose reported in the literature: Bassanini et al.(2005), for example, find that having a lower levelof education exerts a negative impact on the proba-bility of training participation. Surprisingly, there isno significant distinction between different profes-sional statuses. The need for changing knowledgeand skills as well as the use of a computer signifi-cantly increases the likelihood of being a temporary(non-)participant rather than a chronic non-partici-pant. As expected, the likelihood of being a chronicnon-participant is significantly higher for employeeswho consider training to be too much stress in addi-tion to their jobs and private obligations. The regres-sions with the variable representing time preferen-ces indicate that orientation towards the present(high discount rate) is negatively associated with thelikelihood of being a participant. Individuals with ahigh discount rate obviously invest only if they canexpect an immediate gain. Entering variables indi-cating marital status and interaction terms betweendifferent individual characteristics leads to the fol-lowing: contrary to the results usually found in theliterature (see e.g. Bassanini et al. 2005), peoplewho are employed full-time are significantly lesslikely to belong to the group of temporary (non-)participants. Single mothers/fathers as well as peo-ple who are married (no matter whether they havechildren or not) are more likely to take part in con-tinuing vocational training. There is no significantdifference between mothers and fathers.

6 Conclusions

Although continuing training is becoming increas-ingly important, a large proportion of the workforcesurprisingly does not participate in training. Thisseems particularly puzzling since a large number ofstudies demonstrate that participation in trainingleads to substantial positive returns. In our paper weshow that non-participants are not a homogeneousgroup and that the distinction between chronic non-participants and temporary (non-)participants isfundamental to solving the puzzle. We comparechronic non-participants with temporary (non-)par-

304 ZAF 2 und 3/2007

ticipants (employees who have taken part in continu-ing training at least once in the past but are not cur-rently participating). We assume that the two groupsdiffer in observable as well as unobservable charac-teristics and use selection models in our empiricalanalysis in order to account for this problem. Westudy differences in the costs, benefits and/or dis-count rates of the two groups of non-participants.We find that individuals who are chronic non-par-ticipants would have to bear higher costs if theywere to participate. Moreover, the benefits associ-ated with their current job would be lower, i. e. anypay increases or reduction in unemployment riskwould be smaller for chronic non-participants thanfor temporary (non-)participants. However, the re-sults indicate that in the long run chronic non-partic-ipants would benefit from participation in terms ofimproved prospects on the labor market, which indi-cates that the discount rate of chronic non-partici-pants is probably exceptionally high. Although par-ticipation in training would not protect those peoplefrom losing their jobs, it would increase their likeli-hood of finding new jobs once they have becomeunemployed, but this may seem to be too far in thefuture to be important at the present time. Angrist/Lavy (2004) argue similarly with respect to invest-ments in schooling by low-achieving students andsuggest using short-term financial rewards to reducethe problem of exceptionally high discount rates.Based on a randomized trial, they present evidencethat financial incentives do indeed increase highschool certification rates.

Another potential reason could be that chronic non-participants lack the necessary information and cantherefore not anticipate future benefits or possiblejob-offers correctly. Thus, when thinking about pol-icy implications it seems necessary to increase work-ers’ awareness of returns that are not directly associ-ated with their current job, but which might lie farin the future and which might therefore often beneglected. Small financial incentives attached tocompleting training measures could be an optionwith which to overcome the problem of exception-ally high discount rates.

References

Angrist, Joshua D./Lavy, Victor (2004): The Effect of HighStakes High School Achievement Awards: Evidencefrom a Group Randomized Trial. IZA Discussion PaperNumber 1146, IZA, Bonn.

Bassanini, Andrea/Booth, Alison/Brunello, Giorgio/De Pa-ola, Maria/Leuven, Edwin (2005): Workplace Training inEurope. IZA Discussion Paper Number 1640, IZA,Bonn.


Becker, Gary S. (1975): Human Capital. A Theoretical andEmpirical Analysis with Special Reference to Education.New York/London: Columbia University Press.

Beicht, Ursula/Krekel, Elisabeth M./Walden, Günther(2006): Berufliche Weiterbildung Ð Welche Kosten undwelchen Nutzen haben die Teilnehmenden? Bonn: Fed-eral Institute for Vocational Training BIBB.

Bellmann, Lutz (2003): Datenlage und Interpretation derWeiterbildung in Deutschland. Schriftenreihe der Ex-pertenkommission Finanzierung Lebenslangen Lernens,Vol. 2, Bielefeld.

Borjas, George J. (2005): Labor Economics. Third Edition.New York: McGraw-Hill/Irwin.

Breen, Richard (1996): Regression Models: Censored,Sample-Selected, or Truncated Data. In: QuantitativeApplications in the Social Sciences, Vol. 07, Number 111.London: Sage University Papers.

Büchel, Felix/Pannenberg, Markus (2004): Berufliche Wei-terbildung in West- und Ostdeutschland Ð Teilnehmer,Struktur und individueller Ertrag. In: Zeitschrift für Ar-beitsmarktForschung, Vol. 37, Number 2, pp. 73Ð126.

Budria, Santiago/Pereira, Pedro T. (2004): On the Returnsto Training in Portugal. IZA Discussion Paper Number1429, IZA, Bonn.

Chennells, Lucy/Van Reenen, John (1999): Has TechnologyHurt Less Skilled Workers? Institute for Fiscal Studies,Working Paper Series Number W99/27, London.

Frazis, Harley/Loewenstein, Mark A. (2005): Reexaminingthe Returns to Training: Functional Form, Magnitude,and Interpretation. In: Journal of Human Resources,Vol. 40, Number 2, pp. 453Ð674.

Groot, Wim (1995): The Wage Effects of Investments inEnterprise-Related Training. In: Empirical Economics,Vol. 20, Number 1, pp. 133Ð147.

Groot, Wim/Maasen van den Brink, Henriette (2003): Firm-related training tracks: a random effects ordered probitmodel. In: Economics of Education Review, Vol. 22,Number 6, pp. 581Ð589.

Heckman, James J. (1976): The Common Structure of Sta-tistical Models of Truncation, Sample Selection, andLimited Dependent Variables and a Simple Estimatorfor Such Models. In: Annals of Economic and SocialMeasurement, Vol. 5, pp. 475Ð492.

Institut der deutschen Wirtschaft (2006): Geringqualifi-zierte. Anderswo häufiger in Lohn und Brot. In: Infor-mationsdienst des Instituts der deutschen Wirtschaftiwd, Number 15, April 2006.

ZAF 2 und 3/2007 305

Jenkins, Andrew/Vignoles, Anna/Wolf, Alison/Galindo-Rueda, Fernando. (2003): The determinants and labourmarket effects of lifelong learning. In: Applied Econom-ics, Vol. 35, Number 16, pp. 1711Ð1721.

Kuckulenz, Anja (2006): Wage and Productivity Effect ofContinuing Training in Germany: A Sectoral Analysis.Centre for European Economic Research (ZEW), Dis-cussion Paper Number 06Ð025, Mannheim.

Lazear, E. P. (1998): Personnel Economics for Managers.New York: John Wiley&Sons.

Leuven, Edwin/Oosterbeek, Hessel (2002): A New Ap-proach to Estimate the Wage Returns to Work-RelatedTraining. IZA Discussion Paper Number 526, IZA,Bonn.

Pfeiffer, Friedhelm (2000): Training and Individual Per-formance in Europe: Evidence from MicroeconometricStudies. Centre for European Economic Research(ZEW), Discussion Paper Number 00Ð28, Mannheim.

Puhani, Patrick A. (1997): Foul or Fair? The HeckmanCorrection for Sample Selection and Its Critique. AShort Survey. Centre for European Economic Research(ZEW), Discussion Paper Number 97Ð07, Mannheim.

Schröder, Helmut/Schiel, Stefan/Aust, Folkert (2004): Nicht-teilnahme an beruflicher Weiterbildung, Motive, Be-weggründe, Hindernisse. Schriftenreihe der Experten-kommission Finanzierung Lebenslangen Lernens, Vol. 5,Bielefeld.

Spitz-Oener, Alexandra (2006): Technical Change, JobTasks and Rising Educational Demands: Looking Out-side the Wage Structure. In: Journal of Labor Econom-ics, Vol. 24, Number 2, pp. 235Ð270.

Timmermann, Dieter/Färber, Gisela/Backes-Gellner, Uschi/Bosch, Gerhard/Nagel, Bernhard (2002): On the Way toFinancing Lifelong Learning. Interim Report of the Ex-pert Commission on Financing Lifelong Learning. Biele-feld.

Vignoles, Anna/Galindo-Rueda, Fernando/Feinstein, Leon(2004): The Labour Market Impact of Adult Educationand Training: A Cohort Analysis. In: Scottish Journal ofPolitical Economy, Vol. 51, Number 2, pp. 266Ð280.

Wolter, Stefan C./Mühlemann, Samuel/Schweri, Jürg (2006):Why some firms train apprentices and many others donot. In: German Economic Review, Vol. 7, Number 3,pp. 249Ð264.

Wooldridge, Jeffrey M. (2003): Introductory Econometrics:A Modern Approach. Mason (OH): Thomson South-Western.

Wooldridge, Jeffrey M. (2002): Econometric Analysis ofCross Section and Panel Data. Cambridge/Massachu-setts: MIT Press.


Appendix

306 ZAF 2 und 3/2007


ZAF 2 und 3/2007 307


308 ZAF 2 und 3/2007


ZAF 2 und 3/2007 309


310 ZAF 2 und 3/2007


ZAF 2 und 3/2007 311

the puzzle of non-participation in continuing...

Documents