WORKING PAPERWORLD BANK INSTITUTE

DO TRAINING LEVIES WORK?Malaysia’s HRDF and Its Effects on Training and

Firm-Level Productivity1

July 2001

Hong TanLead Economist, WBIHD

1 1 This paper draws upon the 1997 World Bank Country Study, “Malaysia: Enterprise Training,

Technology and Productivity”, and a 2000 PSD Report to the Economic Planning Unit, PrimeMinister’s Department, “Technology and Skill Needs in Malaysian Manufacturing”.

DO TRAINING LEVIES WORK?:Malaysia’s HRDF and Its Effects on Training & Firm-level Productivity

Hong TanLead Economist, World Bank InstituteTel: (202) 473-3206Email: htan@worldbank.org

I. INTRODUCTION

Many countries, both advanced and developing, have put into place differentpolicies designed to foster increased in-service training among its enterprises, includingpayroll-levy training funds and tax incentives for employer-sponsored training. Theseemployer-targeted training policies take many forms: (i) training levy-grant schemes,where fund administrators use levies to make grants to employers for approved training,as in Singapore and previously in the United Kingdom; (ii) training levy rebate schemes,where employers are partially reimbursed for approved training out of their payrolllevies, as in Malaysia, Nigeria and the Netherlands; (iv) levy exemption schemes whereemployers are exempt from levy payments provided they spend a given percentage oftheir payroll on training, as in France, Korea, and Morocco; and (iv) tax incentives forapprove training paid out of general revenues, as in Chile and previously in Malaysia.

How well have these training policies worked in promoting in-service training?The evidence is mixed.2 In spite of the relative prevalence of these schemes, rigorousevaluations of the effectiveness of levy programs are extremely uncommon. Thescattered evidence suggests that while these schemes have, in general, had a positiveimpact on increasing in-service training, they have been inequitable–large employershave benefited to a greater extent than small or medium-size employers. Employerreaction to these schemes has been mixed, with most (especially the smaller ones) feelingthat the levy is an additional tax that has been imposed on them unjustifiably. Problemsassociated with administering the fund and problems of non-compliance abound,especially with such training schemes in developing countries.

In this paper, we consider Malaysia’s experience with the Human ResourceDevelopment Fund (HRDF), a training levy-reimbursement scheme that has been in placesince 1993. The paper has several objectives: (1) evaluate the impact that enactment ofthe HRDF had on promoting enterprise training, among different groups of firms, andover time; (2) disentangle the training effects of HRDF from that of contemporaneousforces that also influence training, in particular, adoption of new technology; (3) provideinsights into why different firms respond differently to training incentives; and (4)estimate the effects of training on productivity growth. 2 See, for example, the review of training levies by Amit Dar, Sudarshan Canagarajah and Paud Murphy,

“Training Levies: Rationale and Evidence from Evaluations”, The World Bank, November 2000.

The paper addresses these issues using several unique linked data sets. It usesthree large enterprise surveys conducted in 1988, 1994 and 1997 that were linked, notonly to each other, but also to panel establishment-level data covering the period between1985 and 1995. These links were possible because each survey—the 1988 Labor MarketFlexibility Survey (LMFS), the 1994 Malaysia Industrial Training and ProductivitySurvey (MITP), and the 1997 Inter-Firm Linkages and Technology Development Survey(ILTD)—drew upon the manufacturing sector sampling frame of the national statisticaloffice (DOS), which in turn provided linked establishment files from annual surveys ofmanufacturing. Three features of these data make the impact evaluation possible:

• The three enterprise surveys elicited broadly similar data on training and use ofnew technology. This allows us to document the incidence of enterprise trainingboth pre- and post-1993 when HRDF was enacted. It also allows analysis ofchanges in training propensities among different groups of enterprises, taking intoaccount parallel investments in new technologies that also raise skill and trainingrequirements.

• The 1994 and 1997 surveys asked whether or not eligible firms (those with over50 employees) were registered with the HRDF. The fact that registered firmshave (in principle) increased incentives to train so as to recover their payroll levycontributions, while non-registered firms would not, is exploited to disentanglethe incentive effects of HRDF from that of other forces (such as technologicalchange) in the post-1993 period.

• The linkage of the three surveys to annual manufacturing survey data—whichcontain detailed production input and output information—brings togethertraining and productivity information. These linked data permit estimation ofproduction functions to measure the effects of training (and indirectly of theHRDF) on productivity growth.

II. THE HUMAN RESOURCE DEVELOPMENT FUND (HRDF)

The HRDF was established in 1993 with a matching grant from the Government3.It replaced the training tax incentive scheme (the double deduction incentive for training)that had been in operation since 1987, and which was widely acknowledged to have beenrelatively ineffective. The Act created a council (HRDC), with representatives from theprivate sector and from responsible government agencies, and a Secretariat to administerthe HRDF schemes. Eligible employers with 50 employees and above are required tocontribute 1 percent of payroll to the HRDF. Those who have contributed a minimum ofsix months are then eligible to claim a portion of allowable training expenditures up to

3 The Government contributed R48.9 million to match projected company levies in the first year; in each of

the following three years, it will add an additional R16.3 million to the HRDF.

the limit of their total levy payments for any given year. The HRDC set rates ofreimbursement, varying by type of training and generally lower for larger firms.

In 1993, the HRDC introduced three basic training schemes that offered firms agreat of flexibility over their training programs. In the ATP scheme, employers can freelysend employees for approved training courses offered by registered training providerswithout the prior approval of the HRDC, and submit claims on completion of the course.In the SBL scheme, employers submit plans to HRDC for approval of ad hoc inplant orexternal training courses offered by non-registered training providers. These plans mustinclude specific objectives, areas of training, duration, number of trainees, instructors,and means of assessment. In the PLT scheme, which is designed for firms with long-term and predictable training needs, employers submit detailed annual training planscovering at least 10 percent of the company’s workforce and 15 percent of junior levelemployees. In addition, HRDC supported efforts of employers to develop training plansthrough the JURUPLAN scheme.

In 1995-1997, the HRDC introduced several additional schemes, many with afocus on the needs of smaller companies that did not appear to respond to these trainingincentives. The PERLA Scheme (Training Agreement Scheme) is designed to lowerfirms’ training cash outlays by enlisting ATP training providers as their agents, to collectfrom users only that portion of fees for which firms are responsible and claim thereimbursable balance directly from HRDC. The SBL Pre-Approved Scheme gives time-tested in-plant training courses an official pre-approved designation, which not onlyallows training providers to market this training but also simplifies employer claims forreimbursement. The HRDC has also targeted SMEs with Training Needs Analysis(TNA) workshops and clinics to answer questions about different schemes; as well asassistance in the purchase of training aids and for setup of training rooms. More recently,it introduced Joint Training Schemes (JTS) to promote group training for SMEs, and on apilot basis, a Group Training Scheme (GTS) to encourage employer associations to play agreater role in developing training programs for their members.

III. OVERVIEW AND ANALYSIS OF TRAINING TRENDS

In this section, we use the three enterprise surveys to document changes in theincidence of formal training over time—in 1988 prior to HRDF, in 1994 a year or so afterHRDF was enacted, and in 1997 or about 3-4 years after HRDF. For tracking changes intraining patterns over time, we use the three survey samples that can be linked to thepanel establishment data set of the Industrial Survey. The time trends are dramatic—theincidence of manufacturing enterprises providing formal training rose from 47 percent in1988 to 64 percent by 1997. The question we ask is how much of an impact did HRDFhave on influencing these time trends in training?

Our focus is on formal training. We exclude informal on-the-job training becauseit is apparently poorly recalled; other research4 also suggests that informal training rarelyhas any discernible measured impact on productivity. Some differences arise in the waytraining questions were asked in the three surveys. Both the 1994 and 1997 surveysdistinguish between informal on-the-job training, which is provided by co-workers andsupervisors, and formal training—with theory and class-room instruction coupled withon-the-job training—provided either in company programs or by external traininginstitutions. In contrast, the 1988 LMF survey first asked employers about whether theyprovided training to improve workers’ current job performance or retraining to promotethem, and then the main form of training, whether from informal training, formal trainingon-the-job, training in external institutions, or “other” unspecified training. This raisesthe possibility, for example, that while informal OJT is the main training modality,training could also be provided via other more formal channels. At one extreme, usingthe “strict definition” of formal training, only 13.3 percent of LMF companies providedformal training in 1988; at the other extreme, if we assume that all training modalitiesinclude some element of formal training, then 34.1 percent of firms are reported toprovide training. In the balance of this paper, we rely on this latter “loose” definition offormal training for 1988.

Table 1 shows the industry and firm size distributions of the three (linked)surveys, and we note that there are 2,308 observations in the 1988 LMF survey, 2,090observations in the 1994 MITP survey, and 1,653 observations in the 1997 ILTD survey.The second panel shows the percent of firms that report providing formal training toupgrade worker skills and improve performance. Keeping in mind the definition offormal training, note that overall the percent of firms providing training rises over time,from 34.1 percent in 1988 to 40.6 percent in 1994 and to 51.7 percent by 1997. Somesmall part of this change is attributable to subtle changes in sample composition, inparticular a higher fraction of small firms in the 1988 sample which, in general, are lesslikely to train. Even still, a similar rising trend in training is observed separately byindustry and firm size.

To what extent was the 1993 enactment of the Human Resource DevelopmentFund (HRDF) responsible for this rising trend in training? As we noted in Malaysia:Enterprise Training (1997), the training policy prevailing prior to the advent of theHRDF—the Double Deduction Incentive for Training—was used most frequently bylarger companies and MNCs, and was largely ineffective in inducing smaller domesticcompanies to undertake training for employees. The HRDF is thought to have changedemployers’ training incentives by making it mandatory, for firms with 50 or moreworkers, to contribute a training levy of 1 percent to the HRDF, levies that would bereimbursed to employers upon provision of approved training programs under a variety oftraining schemes. However, not all eligible firms registered with the HRDF, and thusonly those firms that did may have faced incentives to train. As such, we restrict analysisto the three samples of firms with 50 or more employees, and in 1994 and 1997, we also

4 World Bank, Malaysia: Enterprise Training, Technology and Productivity, World Bank Country Study,

1997.

distinguish between firms reporting that they were registered with the HRDF (and thuspresumably paying the training levy) and those not registered with HRDF.5

Table 1 Sample Size and Percent of FirmsProviding Formal Training by Industry and Firm Size—1988, 1994, 1997

Number of Observations Percent Formal TrainingIndustry 1988 1994 1997 1988 1994 1997 Food & Beverages 518 399 308 26.25 28.57 39.61 Textiles & Apparel 199 209 174 38.69 34.45 48.28 Wood & Wood Products 234 219 151 20.09 30.14 36.42 Furniture 68 69 76 29.41 26.09 39.47 Paper & Printing 150 126 119 39.33 34.92 52.94 Chemicals & Petroleum 383 339 254 40.99 50.15 65.35 Glass & Ceramics 143 129 95 25.87 44.19 47.37 Basic Metals 71 57 46 29.58 45.61 60.87 Fabricated Metals 141 116 104 39.01 40.52 48.08 General Machinery 121 87 78 28.93 34.48 43.59 Electrical Machinery 138 208 136 63.04 64.42 84.56 Transport Equipment 78 74 47 42.31 50.00 63.83 Other Industry n.e.c. 64 58 65 37.50 58.62 50.77Firm Size Small—Less than 100 1614 1159 894 24.35 25.54 35.12 Medium—100-249 433 535 437 49.88 51.78 63.62 Large—250+ 261 396 322 68.58 69.7 81.68

Total 2308 2090 1653 34.14 40.62 51.72

Table 2 incorporates these two features. The first panel shows a higher proportionof firms reporting training in each year, as might be expected once small firms with fewerthan 50 employees—which typically have lower training propensity—are excluded. Arising trend in training is still apparent, with the percent of firms providing formaltraining rising from 47.2 percent in 1988 to 53.5 percent in 1994, and to 64.5 percent in1997. HRDF registration is also an important determinant of training.

The second panel of Table 2 shows the incidence of training in 1994 and 1997,separately by whether employers were registered with the HRDF or not. In each year,and for each firm size category, the percent of employers reporting training isdramatically higher for the group that was registered with the HRDF. For example, in1994, 60 percent of the registered group provided training as compared to 35 percent ofnon-registered firms; by 1997, the corresponding figures were 77 percent and 34 percent,respectively. Similar trends are found for the balanced panel of firms that appear in all

5 About 75 percent of eligible firms in 1994 and 1997 reported that they were registered with the HRDF.

three surveys.6 These trends tend to support the view that enactment of the HRDF in1993 was important in fostering increased training among employers.

Table 2 Percent of Firms Providing Formal Trainingby Firm Size and Status of HRDF Registration—1988, 1994, and 1997

Firm Size 1988 1994 1997 1994 1997HRDF=0 HRDF=1 HRDF=0 HRDF=1

Small 35.03 39.44 49.10 32.49 46.43 34.54 63.31 Medium 49.88 51.77 63.62 37.06 57.14 33.88 75.00 Large 68.58 69.70 81.68 44.00 71.43 34.21 88.03

Total 47.23 53.47 64.50 35.07 60.48 34.26 77.24 Note: HRDF = 1 indicates registration with the HRDF and payment of training levies, while HRDF = 0 indicates not-registered with HRDF.

Some part of this rising trend in training may also be attributed to technologicalchange. There is a growing body of literature in both advanced and developing countries,including Malaysia, that technological change creates increased demand for skilledworkers and for training to use the new technology.7 We examine this possibility usingdirect measures of changing technology use over time. The first measure, common to allthree surveys, is an indicator variable for whether the firm introduced new product orprocess technologies over the past two years (MITP and ILTD surveys) or three years(LMF survey). About 30.9 percent of LMF firms reported having introduced newtechnology over the recent past; the corresponding figures for the 1994 MITP and 1997ILTD surveys are 42.6 percent and 28.5 percent, respectively. The second measure—anindicator variable for whether the firm is currently using one of five types of informationtechnology (IT)8 in its production processes—was elicited in the 1997 ILTD survey but,because we know when that specific IT technology was adopted, this information can betracked back over time for all firms in the other two surveys that were included in theILTD sample. The proportion of firms using process IT is relatively small in 1988—2.2percent—but it rises to 8.1 percent by 1994, and 20.6 percent by 1997.9

Table 3 shows the percent of firms that report providing formal training by use ofnew technology (TEC), separately by year and by firm size. First, consider Panel A. Ineach year, the likelihood of training is significantly greater in the sample that reported

6 The results are virtually identical when tabulations are repeated, this time restricting the sample to the

balanced panel of 646 firms that appear in each of the three surveys. In 1994, 60 percent of theregistered firms report training as compared to 31 percent of non-registered firms; the correspondingfigures in 1997 are 73 percent and 33 percent, respectively.

7 See Tan and Batra (1995), Bartel and Lichtenberg (1987), and Berman Bound and Machin (1997).

8 These five types of process IT include CAM (computer-assisted manufacturing), robots, CNC(computerized numerical control) machine tools, FMS (flexible manufacturing systems), and CIM(computer-integrated manufacturing).

9 See Hong Tan (2000), and Bresnahan, Erik Brynjolfsson and Lorin Hitt (1999).

introducing new technology in the recent past as compared to those not introducing newtechnology: 53.4 versus 25.5 percent in 1988, 58.3 and 27.5 percent in 1994, and 75.4and 42.3 percent in 1997. In Panel B, the process IT variable in 1988 is not particularlyeffective in discriminating between training and non-training firms, not surprising sincediffusion of advanced process IT was relatively low in 1988. In the more recent period,differences emerge in training propensity by use of advanced process IT—70.4 versus38.0 percent in 1994, and 82.9 versus 43.6 percent in 1997. In tables not reported here,the increased propensity to train when new technology is used is evident even when thesecontingency tables are further broken down by whether or not the firm was registeredwith the HRDF in 1994 and 1997.10

Table 3 Percent of Firms Providing Formal Trainingby Firm Size and Type of New Technology Used—1988, 1994, and 1997

Firm Size 1988 1994 1997TEC=0 TEC=1 TEC=0 TEC=1 TEC=0 TEC=1

A. New technology Small 19.49 41.34 18.25 42.53 29.22 61.21 Medium 43.44 58.20 41.18 62.74 57.29 75.00 Large 59.57 73.65 59.83 73.83 73.25 91.33Total 25.52 53.36 27.50 58.31 42.30 75.37B. Process IT Small 24.36 22.22 24.51 51.11 31.58 64.58 Medium 49.88 50.00 48.75 78.18 57.44 84.16 Large 68.46 70.00 68.19 76.81 71.51 94.41Total 33.81 48.08 38.00 70.41 43.64 82.94 Note: TEC equals 0 or 1 indicates use of each type of new technology, where: A. New technology = introduction of new product/process technology in previous 3 years B. Process IT = use of one or more of 5 types of IT for production process functions.

IV. EXPLAINING CHANGES IN TRAINING (1988-1997)

Given the limitations of contingency tables, we turn to multivariate regressionsfor a more rigorous analysis of the factors that determine firms’ propensity to train. Theunderlying economic model is one in which a firm’s decision to train its employees isshaped by the profitability of this training investment—the discounted future stream ofproductivity increases net of the cost of upgrading worker skills. The firm trains if thenet present value of training is positive; otherwise, it does not. If new technologyrequires skilled workers to fully realize their productivity potential, its introduction canbe a powerful incentive for employers to initiate skills training because training nowbecomes profitable.

10 For example, in the sample not using advanced process IT in 1997, 45.5 percent of firms registered with

the HRDF train as compared to 25.6 percent of non-registered firms; in the sample using advancedprocess IT, 90.7 percent of the sample registered with the HRDF train as compared to 47.5 percentthat are not registered with the HRDF.

The model is a random effects probit of the firm’s training decision, which can bewritten as follows:

Pit = β0 + β1Xit +β2Yi +β3Zt + εit

where the dependent variable, Pit, takes on a value of 1 if firm i provides training in yeart, and 0 otherwise. The probability of training, Pit, depends on its profitability, and weassume that this is reflected in a vector of time-varying variables Xit—highly skilledworkers (professionals, managers, and technicians) and skilled production workers asproportion(s) of the workforce, union status, whether registered with the HRDF, whetherintroduced new technology in the recent past or is using of advanced process IT, whetheran exporting firm, and dummy variables for medium and large firms (some firms grow,others contract over time); a vector Yi of dummy variables for 12 industrial sectors(time-invariant); Zt refers to year dummy variables for 1994 and 1997, with 1988 as the(omitted) base year; and β’s are parameters to be estimated.11

Table 4 Random Effects Probit EstimatesProbability of Formal Training—1988, 1994, and 1997

Dep. variable: P(train)it Model 1 Model 2Explanatory variables: Coef. z-stat Coef. z-stat

Introduced new technology 0.5555 12.74 -- -- Use advanced process IT -- -- 0.4839 6.14 Registered with HRDF 0.6346 11.15 0.6111 10.78 Medium size firm 0.4094 7.71 0.4555 8.62 Large size firm 0.7386 10.57 0.8271 11.88 Proportion highly skilled 1.0244 7.49 0.9906 7.26 Proportion skilled production -0.0046 -0.06 -0.0258 -0.35 Union indicator variable 0.3102 5.87 0.3140 5.97 Export indicator variable 0.1521 3.31 0.2223 4.88 1994 indicator variable -0.2675 -4.90 -0.2138 -3.90 1997 indicator variable 0.1837 3.14 0.1199 1.86 Constant term -1.0995 -8.52 -0.9976 -7.63

Log-likelihood -3281.7 -3347.8Total observations 6,052 6,052Total firms 3,703 3,703

Note: Model includes missing indicator variables for IT, skills, and union status, as well as indicator variables for 12 industrial sectors

Table 4 shows the probit estimates for two model specifications: model 1, wherethe technology variable is introduction of new product or process technologies in therecent past; and model 2, where use of advanced process IT is the technology variable.Consistent with tabulations reported in the previous section, these probit estimates show 11 In addition, indicator variables for missing values were included for union status, skill composition, and

IT use, the latter being available only for those firms that can be linked to the 1997 ILTD survey.

that both enactment of the HRDF in 1993 and use of new technology were importantdeterminants of training, and contributed to the rising trend in training. Both technologymeasures are positive and statistically significant at the 1 percent level.12 As before, thepropensity to train is higher in medium and large firms as compared to the omitted smallfirm size group, suggesting that training is more profitable in larger firms. Trainingpropensity is also positively associated with the proportion of the workforce that is highlyskilled, reflecting the greater ability of more educated workers to benefit from training,and to use new technology; however, the share of skilled production workers is notsignificant. Finally, unionized firms and export-oriented firms are more likely to train.

In Table 5, we report the actual and predicted probabilities of formal training,separately by firm size, and simulations of what training probabilities would be underalternative scenarios. Note that the probit models tend to under-predict levels and the risein training over time for small firms. Compared to actual probabilities of training—35,39 and 49 percent in 1988, 1994 and 1997—the predicted training probabilities are 24,22, and 34 percent, respectively. For medium and large firms, the probit models tend topredict training levels well, but they tend to over-predict the rise in training over time.

Table 5 Actual and Predicted Probabilities of TrainingAnd Simulations With HRDF and TEC—1988, 1994, and 1997

Simulations Actual and predicted training Change relative to BaselineFirm size 1988 1994 1997 1988 1994 1997Small size Firms Actual probability 35.0 39.4 49.1 24.2 22.2 34.2 Predicted baseline 24.2 22.2 34.2 0 0 0 HRDF=0 24.2 16.9 28.2 0 5.2 6.0 TEC=0 20.1 17.2 30.5 4.1 5.0 3.7 HRDF=0 TEC=0 20.1 12.4 24.5 4.1 9.7 9.7Medium size firms Actual probability 49.9 51.8 63.6 47.1 53.1 65.9 Predicted baseline 47.1 53.1 65.9 0 0 0 HRDF=0 47.1 36.3 50.3 0 16.8 15.3 TEC=0 37.9 43.5 59.9 9.1 9.6 6.1 HRDF=0 TEC=0 37.9 26.5 42.8 9.1 26.6 23.2Large size firms Actual probability 68.6 69.7 81.7 65.2 74.8 82.0 Predicted baseline 65.2 74.8 82.0 0 0 0 HRDF=0 65.2 55.1 66.7 0 19.7 15.3 TEC=0 52.4 62.8 76.5 12.8 12.0 5.5 HRDF=0 TEC=0 52.4 40.2 57.9 12.8 33,6 24.0

Note: simulations based upon probit parameters reported in Table 14.

In the simulations, we first set HRDF registration to zero, then set use of newtechnology to zero, and finally, set both to zero. The first counter-factual—no firms are

12 Since the results are quite similar, we focus on the findings of model specification 1.

registered with the HRDF—is tantamount to assuming that the HRDF was not enacted in1993. The second—no introduction of new product/process technology, or no diffusionof advanced process IT—assumes away technological change. These simulationsprovide a feel for how important these two factors have been, quantitatively, in shapingtraining trends, either by themselves or together. To facilitate comparisons, the secondpanel of Table 5 shows the predicted change (decline) in training due to each scenario. In1988, for example, setting HRDF = 0 does not change predicted training since this is apre-HRDF year. Thus, in the second panel, the predicted change in the HRDF=0 row isalso 0. But setting TEC = 0 in 1988 has an impact, reducing predicted training for small,medium and large firms by 4.1, 9.1 and 12.8 percentage points, respectively.

Table 5 suggests that overall the enactment of HRDF had a more importantimpact on increasing training than new technology. To see this, compare the fall intraining probabilities predicted under the no-HRDF and no-technical change scenarios,relative to the baseline probabilities in 1997. For example, for small firms, the 1997baseline probability is 34 percent and this falls to 28 percent in the no-HRDF scenario,and to 30 percent in the no-technical change scenario. For medium size firms, the declineis from 66 to 50 percent in the no-HRDF scenario, as compared to 66 to 60 percent in theno-technical change scenario; for large firms, the corresponding figures are 82 to 66percent in the no-HRDF scenario, and from 82 to 76 percent in the no-technical changescenario. In other words, HRDF had a larger impact on training decisions than didtechnical change.

These two scenarios also have differential impacts on the predicted trainingoutcomes by firm size. HRDF appears to have the largest impact on medium-sizecompanies. To see this, consider 1997. In the no-HRDF scenario, predicted trainingprobabilities decline more in relative terms for medium size firms as compared to smallor large firms. Relative to the baseline probability (see the second panel), the no-HRDFscenario results in a 23 percent drop (15.3/65.9) for medium size firms, and only betweena 17 to 19 percent drop for either small (6.0/34.1) or large firms (15.3/82.0). Oneexplanation is that many medium size firms were close to the training threshold, and theHRDF was instrumental in switching them from non-training to training firms; theHRDF may have been less effective for small firms, perhaps because skills needs wereless pressing, and for large firms, many of whom would have trained employees even inthe absence of HRDF.

In contrast, the impact of the no-technical change scenario on predicted training ismost pronounced for small firms. Relative to the baseline, the drop in predicted trainingprobability is about 11 percent (3.7/34.2) for small firms, and 9 and 7 percent for mediumand large firms, respectively. Unlike small firms, many medium and large size firmsalready have modern equipment installed requiring trained personnel; for them, wemight expect the introduction of new technology in the recent past to have a smallertraining effect. In small firms, on the other hand, we might expect training probabilitiesto be more sensitive to the introduction of new technology since it is a relatively lesscommon event than in other firms.

V. TRAINING AND PRODUCTIVITY GROWTH

The findings reported above are broadly consistent with the view that employerdecisions about training and skills upgrading are shaped by a rational calculus of theprofitability (productivity) of training, especially in the context of technological change.We test this proposition directly by estimating the underlying production function toinvestigate the productivity impact of training, and the magnitude of this productivityeffect both with and without complementary investments in new product and processtechnologies.

We do this in two ways. First, we use the pooled cross-sectional data from the1988, 1994 and 1997 surveys. In the pooled cross-sections, we use predicted trainingprobability as the training measure to address the potential estimation biases that arisefrom including an endogenous training choice variable in the production function.13 Thispredicted training measure, taken from the baseline probit (specification one), yieldsunbiased estimates of the productivity impact of training because it is uncorrelated, byconstruction, with the unmeasured productivity attributes of the firm in the error term.Second, we use the balanced panel of firms appearing in all three surveys to estimate theimpact of multiple training episodes on productivity growth. We define a categoricalvariable that measures the firm’s training history: equals 0 if the firm does not train atall, equals 1 if it trains in just one period, equals 2 if it trains in any two out of the threeperiods, and equals 3 if it trains in all three periods. This allows us to statistically test,within a production function framework, whether productivity growth is enhanced byrepeated episodes of training.

Table 6 reports log-linear Cobb-Douglas production functions estimated using arandom effects model for the panel of establishments that responded to the 1988, 1994and 1997 surveys. 14 In these models, the logarithm of value added is regressed on thelogarithms of fixed capital assets; logarithm of number of employees in four skill groups:(1) professionals, managers, technicians, (2) skilled production workers, (3) semi-skilledproduction workers, and (4) unskilled production and general workers; the predictedprobability of training; indicator variables for ownership; dummy variables for 1994 and1997; and twelve industry dummy variables.

The production function results in Table 6 are fairly typical of those reported inthe literature, and they require little explanation. What is new are the training parameter

13 The decision to train is related not only to measured employer and worker characteristics, but also to

unobserved productivity attributes of the establishment, such as managerial talent, risk-aversion, andthe like. As such, in a second-stage production function, the included training indicator variable islikely to be correlated not only with included employer and worker characteristics, but also with theerror term of the production function, leading to estimation biases. The solution is to use predictedtraining from a first-stage probit that is, by definition, purged of any correlation with the unmeasuredproductivity attributes of the firm. For an extended discussion of this issue, see World Bank (1997),Malaysia: Enterprise Training, Technology and Productivity.

14 The random effects model takes into account the fact that the error term is correlated across cross-sections because the same firms are observed repeatedly in the three survey years.

estimates. Since the predicted probability of training (a continuous variable) is used inplace of a 0,1 training indicator variable, the estimated parameter does not directly yield ameasure of its productivity impact; instead, it has to be calculated by evaluating theestimated parameter at the sample mean of the predicted training probability. Considerthe first panel. Since the training parameter is 0.7016 and the mean predicted trainingprobability is 0.4066, the mean impact of training on value added is 28.5 percent, whichis calculated as (0.7106 x 0.4066 = 0.2853).

Table 6 Production Function Estimates with Predicted TrainingAnd Whether Introduced New Technology Recently

Recently Introduced New Technology?Dependent variable: Log(real value added)

All FirmsNo Yes

Explanatory variables Coef. z-stat Coef. z-stat Coef. z-statProduction Function Log(capital assets) 0.3042 32.51 0.2954 26.49 0.3268 19.88 Log(skill group 1) 0.4035 25.64 0.3962 19.09 0.4023 15.67 Log(skill group 2) 0.1237 16.13 0.1437 13.82 0.0867 8.04 Log(skill group 3) 0.0435 6.66 0.0529 5.81 0.0292 3.20 Log(skill group 4) 0.0757 11.73 0.0865 9.94 0.0465 5.12Training Measure Predicted training 0.7106 9.50 0.7704 5.97 0.8211 5.45 (sample mean)1 (0.4066) (0.3006) (0.6097)Ownership & Time Joint ventures 0.1142 3.22 0.1476 3.28 0.1149 2.21 Foreign owned firms 0.1947 5.15 0.2541 5.14 0.1528 2.96 1994 dummy 0.0728 3.55 0.0761 2.68 0.0658 1.97 1997 dummy -0.0124 -0.50 -0.0511 -1.48 0.0101 0.21Industry dummies Food & beverages 0.5035 6.21 0.4681 4.61 0.6582 5.79 Textiles & apparel 0.3005 3.45 0.2664 2.45 0.3463 2.82 Wood products 0.6540 7.49 0.6474 5.97 0.6281 4.94 Furniture 0.3899 3.93 0.3101 2.47 0.4779 3.41 Paper & printing 0.5462 6.00 0.5202 4.62 0.5975 4.74 Chemicals & rubber 0.4812 5.93 0.4453 4.35 0.5446 4.89 Glass & ceramics 0.2374 2.56 0.2058 1.83 0.3278 2.39 Basic metals 0.2854 2.66 0.2545 1.94 0.3441 2.20 Fabricated metals 0.3017 3.33 0.2996 2.66 0.2478 1.92 General machinery 0.2746 2.67 0.3257 2.42 0.3119 2.26 Electrical machinery 0.2969 3.35 0.1813 1.51 0.3495 3.04 Transport equipment 0.3144 2.85 0.3760 2.77 0.2557 1.71 Constant term 7.9561 61.81 8.0180 51.88 7.7299 35.54

Total observations 5,618 3,640 1,978 Number of firms 3,468 2,614 1,563 Overall R2 0.7896 0.7607 0.7811

Note: 1 sample means of training probabilities predicted using model 1 areenclosed in parentheses. See Table 4.

In the next two panels, the training parameter is estimated at 0.7704 for thesample of firms that did not introduce new technology, and 0.8211 for the sample offirms that did. The implied productivity impacts of training is 23.1 percent (0.7704 x0.3006 = 0.2316) and 50.0 percent (0.8211 x 0.6097 = 0.5006), respectively. In otherwords, on average, the productivity impact of skills training is twice as high in firms withnew technology as it is in firms without new technology. This is dramatic confirmationof the key intermediating role that skills training plays in realizing the productivitypotential of new technology.

Table 7 reports a similar exercise for production functions estimated separatelyfor small firms (first panel) and medium and large firms combined (second panel). In thefirst panel, the estimated training parameter is 0.8371 which, at the sample mean, impliesa training productivity impact in small firms of 22.7 percent (0.8371 x 0.2601 = 0.2271).For medium and large firms, the parameter is 0.6280, indicating an average productivityimpact of 39.7 percent (0.628 x 0.632 = 0.3971). These results provide an explanationfor why training is more common among larger firms than among small firms—trainingis simply more profitable (productive) in larger firms, possibly because larger employersare also more likely to use new technologies requiring skilled and trained workers.

The third panel of Table 7 reports the findings for a balanced panel of 608 firmsthat appear in all three surveys, and that have complete production and training data in allyears. Unlike the previous analyses, the balanced panel allows us to exploit theavailability of training histories over the 1988-1997 period. We categorize all firms asfalling into one of four groups: 0 if the firm does not train in all three periods, 1 if it trainsin any one period, 2 if it trains in any two periods, and 3 if it trains in all three periods.This categorization is a simple, yet parsimonious way of incorporating training historiesinto a panel production function. The results suggest that productivity growth isenhanced by repeated episodes of skills training. There is a strict ordering of productivitygrowth by number of training episodes—compared to the omitted group (who do nottrain), firms are 9 percent, 26 percent, and 31 percent more productive when they train inone, two or in all three periods, respectively. These differentials in productivity growthare statistically significant for firms that train in two or all three periods.

VI. CONCLUDING REMARKS

To summarize, there is strong panel evidence that enactment of HRDF in 1993was instrumental in promoting increased enterprise training in Malaysia. Technologicalchange also had a role in inducing enterprise training, but the overall contribution ofHRDF was much larger, especially among medium-size companies. Smaller companiescontinue to lag behind in training, and more proactive training strategies may be neededto reach this group of enterprises. The resulting increase in training investments, whetherinduced by HRDF or by adoption of new technology, have strong demonstrated impactson productivity growth, especially when training is continuous and not episodic.

Table 7 Production Function Estimates with TrainingBy Firm Size and By Number of Training Episodes

Dependent variable: Log(value added)

Small Firms Medium & LargeFirms

Balanced PanelTraining Episodes

Explanatory variables Coef. z-stat Coef. z-stat Coef. z-statProduction Function Log(capital assets) 0.2575 22.25 0.3558 22.00 0.2942 16.67 Log(skill group 1) 0.4289 19.85 0.3216 14.42 0.4444 16.33 Log(skill group 2) 0.1823 14.74 0.0743 7.73 0.1062 7.79 Log(skill group 3) 0.0851 7.41 0.0091 1.19 0.0316 2.86 Log(skill group 4) 0.1212 11.35 0.0289 3.52 0.0673 6.17Training Measures Predicted training 0.8731 7.54 0.6283 5.66 -- -- (sample means) 1 (0.2601) (0.6320) 1 Training episode -- -- -- -- 0.0908 1.38 2 Training episodes -- -- -- -- 0.2668 3.26 3 Training episodes -- -- -- -- 0.3158 3.05Ownership & Time Joint ventures 0.1191 2.31 0.0530 1.18 0.0638 1.05 Foreign owned firms 0.3737 6.05 0.0556 1.26 0.1809 2.26 1994 dummy 0.1328 4.98 -0.0187 -0.56 0.1522 4.88 1997 dummy -0.0108 -0.33 -0.0405 -1.03 0.0898 2.76Industry dummies Food & beverages 0.5291 4.82 0.6151 5.48 0.2156 1.25 Textiles & apparel 0.3488 2.83 0.2040 1.85 0.0085 -0.04 Wood products 0.6637 5.53 0.5378 4.69 0.2636 1.46 Furniture 0.4210 3.17 0.2300 1.71 0.3433 1.58 Paper & printing 0.6089 4.98 0.3586 2.91 0.3219 1.68 Chemicals & rubber 0.5355 4.75 0.2955 2.82 0.4221 2.43 Glass & ceramics 0.2825 2.28 0.1801 1.39 0.0077 0.04 Basic metals 0.3642 2.49 0.1689 1.19 0.1507 0.60 Fabricated metals 0.3808 3.13 0.0655 0.52 0.0269 0.13 General machinery 0.2268 1.65 0.3028 2.10 0.1166 0.53 Electrical machinery 0.2215 1.51 0.3052 2.89 0.3631 1.81 Transport equipment 0.2986 2.03 0.3038 2.06 0.1947 0.93 Constant term 8.1769 49.69 8.0959 35.86 8.4661 33.16

Total observations 3,359 2,259 1,824 Number of firms 2,212 1,475 608 Overall R2 0.6604 0.6468 0.7973

Note: 1 sample means of training probabilities predicted using model 1 areenclosed in parentheses. See Table 4.

REFERENCES

Bee-Yan Aw and Hong Tan, “Training, Technology and Firm-Level Productivity inTaiwanese Manufacturing, PSD Working Paper, The World Bank, 1994.

Ann Bartel and Frank Lichtenberg, “The Comparative Advantage of Educated Workersin Implementing New Technology”, Review of Economics and Statistics, February 1987,Volume LXIX, 1, pp. 1-11.

Eli Berman, John Bound and Stephen Machin, “Implications of Skill-BiasedTechnological Change: International Evidence”, Centre for Economic Performance,London School of Economics, September 1997.

Timothy Bresnahan, Erik Brynjolfsson and Lorin Hitt, “Information Technology,Workplace Organization, and the Demand for Skilled Labor: Firm Level Evidence”,NBER Working Paper 7136, May 1999.

Amit Dar, Sudarshan Canagarajah and Paud Murphy, “Training Levies: Rationale andEvidence from Evaluations”, The World Bank, November 2000.

S. Machin, A. Ryan and J. Van Reenan, “Technology and Changes in Skill Structure:Evidence from an International Panel of Industries”, Center for Economic Performance,Discussion Paper No. 297, London School of Economics, June 1996.

Hong Tan and Geeta Batra, Enterprise Training in Developing Countries: Incidence,Productivity Effects, and Policy Implications, PSD Book, The World Bank, 1995.

World Bank, Malaysia: Enterprise Training, Technology and Productivity, World BankCountry Study, The World Bank, 1997.

Hong Tan, Technology and Skill Needs in Malaysian Manufacturing, PSD Report to theEconomic Planning Unit (Government of Malaysia), The World Bank, 2000.

OECDS:\Data\Beaudoin\Meetings - other than Fora\FDI & Ed - KOHL Dec 2001\Papers\FINALS\Malaysia HRDF Evaluation Paper-Bkgrnd Hong Tan.doc12-Dec-01 3:42 PM