How to overcome rent seeking in Tanzania ’ s skills sector? Exploring feasible reforms through discrete choice experiments

Skills gaps and mismatches are widely documented as a hindrance to inclusive structural transformation across developing countries, especially in Africa. What is often overlooked, however, is the fact that skills development is a complex political economy process challenged by institutional and financing problems on the supply side, and inadequate demand, that is, a shortage of firms that can organise skilled labour and provide on-the-job training effectively. In such adverse contexts, rent seeking and corruption may arise from conflicting objectives, trade-offs and mis-aligned incentives among stakeholders – public sector skills providers and firms. With a focus on Tanzania, we (i) analyse the incentive structures underlying such rule-breaking behaviours and processes

Tanzania is no exception here, and a substantial skills gap has been identified as a key bottleneck for productivity-enhancing structural change and good employment generation, especially among the youths (LO/FTF Council, 2016;Tan, Bashir, & Tanaka, 2016).While high levels of academic education provide youths with higher returns than those available from vocational education, the vocational return can exceed the academic return at lower levels (Kahyarara & Teal, 2008;Attanasio, Guarín, Medina, & Meghir, 2017).Among more recent studies, experimental research has focused on the returns to primary education across different employment activities, finding higher returns in non-farm selfemployment activities (Delesalle, 2021).
Most studies and policies addressing the skills gap tend to have a 'supply-side' bias (Amsden, 2010) and a relatively 'linear view' of the skills development process.A growing body of literature has explored interacting dynamics linking supply and demand for skills and has highlighted several reasons why standard vocational education and training (VET) programmes have had limited success in developing countries (Carranza & McKenzie, 2024;McKenzie, 2017).One key explanation offered is that even workers who have access to VET programmes (and manage to develop high-level technical skills) might find it difficult to get the jobs that need their skills due to 'search and matching frictions' in the market (Carranza, Garlick, Orkin, & Rankin, challenge, highly subsidised VET programmes are used to keep youth out of the labour market for longer than actually needed, to compensate for poor basic education and to provide youth with essential skills for employment in the formal but mostly informal sector.The funding to provide for subsidised VET courses is generated through the Skills Development Levy (SDL).The SDL is a tax payable by employers as a percentage of payroll and is administered by the public VET Authority (VETA).
The SDL has been highly contested in Tanzania since its introduction and featured in several government-business negotiating tables (Andreoni, 2018).Over the years, the private sector has raised concerns about the size of the levy, the usefulness of the subsidised training provided, and misallocation and misuse of the levy by the government.The government, on the other hand, has accused companies of hiring foreign workers without work permits or so-called succession plans, which is a violation of the Non-Citizens Employment Act.The political economy of the SDL has hence caused significant mistrust between the public and the private sector, with evidence of rule bending and breaking behaviour on both sides (Andreoni, 2018).As a result, the skills sector is affected by conflicting dynamics and allegations of resource misallocations (on the public-sector side) and tax avoidance (on the private-sector side), despite a strong joint interest in developing skills between governments and companies.This impasse results in skills deficiency and a low-productivity skills trap.
The aim of this paper is twofold.First, we evaluate the incentive structures that result in the political economy problems highlighted above and that is conducive to rule-breaking behaviour.Second, we empirically test institutional design strategies that would better align the incentives of different private-and public-sector stakeholders and thereby reduce misallocation of resources, increase employers' compliance with tax and labour law, and, eventually, overcome the lowproductivity skills trap.We hypothesise that the successful alignment of incentive structures must take into account the different and potential conflicting objectives of skills provision pursued by public-sector stakeholders and must also account for the heterogeneities in skills needs and capabilities of different types of productive organisations in Tanzania.
For the first objective, we rely on over 30 in-depth stakeholder interviews and data obtained from various government sources throughout 2018 and 2019.For the second objective, we conduct three Discrete Choice Experiments (DCEs) complemented by a comprehensive questionnaire of a sample of 209 Tanzanian firms collected in 2019.This approach enables us to uncover latent preference structures (i.e., not directly observable 'wants' which guide agents' behaviour), differentiated by observable firm characteristics.Knowledge about these latent preference structures then enables us to navigate differences across firms and their incentives, so to formulate feasible policy solutions that reflect such incentives and preferences, as well as differences in capabilities to effectively respond to such incentives.
Results confirm substantial heterogeneity in the preferences across firms.Differences arise mainly due to diverging technical capabilities, existing training provision and the size of the organisations.Larger and more technically capable organisations prefer to be involved in the design of the VET curriculum and prefer substantial parts of the SDL to be spent on sector-specific training provision.These organisations are also less concerned about the size of the levy or financial rewards for investing into skills provision, and, in many cases, already engage in skills provision through in-house training.In contrast, smaller and less technologically and organisationally capable organisations prefer a generalist curriculum and subsidised courses to be shorter in duration.
Section 2 outlines the potentially conflicting economic and social objectives pursued by the public sector in skills provision schemes and details the existing incentive structures for both private-and publicsector stakeholders in Tanzania.Section 3 describes our research method, research design and data collection.Section 4 reports on our research findings by firstly summarising information collected about the A. Andreoni et al. extent of skills provision among the firms surveyed and the perceived skills gap, and secondly by reporting the results of the DCEs.Section 5 concludes with a summary of our findings and policy suggestions.

The political economy of the Tanzania's skills sector: Conflicting objectives, different skills needs, and rent seeking
The provision of VET schemes in Tanzania is motivated by conflicting objectives, which result in trade-offs and various forms of rule bending and breaking behaviouri.e.rent seeking. 1 In this section, we carefully review the different objectives and incentive structures underpinning private and public stakeholder behaviour, focusing on VET financing and institutions on the skills supply side first, and on firms demanding skills after.The SDL financing modelthe source and allocation of these resourcesis central to understand the current misalignment of interests and behaviours between supply and demand.We conclude this analysis with two main hypotheses around ways to realign incentives among different actors.

The institutional context and financing of VET
VET activities and services in Tanzania are run by around 520 providers and promoted under more than 30 programmes and public-private partnership (PPP) schemes.VET providers include public VET centres, 28 of which are owned by VETA.VETA is both a VET provider and the regulatory body for the entire skills sector.The National Council for Technical Education (NACTE) is responsible for regulating technical education in Tanzania, but, unlike VETA, it does not provide any technical education and training directly.
VETA is funded through the SDL, which, since the fiscal year 2016-17, is charged at 4.5 per cent of the total emoluments paid to all employees per month.Any employer with four or more members of staff must pay the SDL.At 4.5 per cent, Tanzania's SDL is significantly higher than in other countries which collect a levy, such as Kenya and South Africa, and consequently there are worries among private-sector stakeholders that this can create a competitive disadvantage and discourage formal employment.
Tanzania's existing VET centres are extremely heterogeneous with respect to their capacity, training quality, and the type of courses provided.The duration of VET courses is determined by the content of the curriculum and ranges between one and three years for 'long courses' (LCs) including on-the job placements ('field attachments') and from three to nine months for 'short courses' (SCs).LCs offered by VETAowned centres are heavily subsidised through the SDL.SCs are offered by both VETA-owned and private VET centres and offer VET centres the opportunity to generate income at the local-centre level. 2n this paper, we focus on institutional-based VET activities, that is, those conducted in a VETA-owned VET centre or a registered, privately owned VET centre, the largest cluster of VET activities in Tanzania.
The SDL is a significant levy across tax departments in Tanzania, with the total funds collected across small, medium and large taxpayers/ employers accounting for 3.26 per cent of total domestic revenue.The SDL is collected by the Tanzania Revenue Authority (TRA) under the VET Act and Income Tax Act to promote skills development in the country.According to the revised VET Act 2006, two thirds of the SDL collected should go to the Treasury, and one third to the VET Fund.The Treasury then allocates parts of the funds to the Ministry of Education, Science and Technology, which in turn transfers the resources to VETA, based on an agreed annual budget.
Since 2016 the Treasury has applied a ceiling of 53 billion TShs on the VETA budgets, in part a response to worries over wasteful expenditure at the level of individual VETA centres, as well as a desire to use part of the SDL for other education-related budget items.Such ceiling remained in place throughout the period of the study.The shift in the allocation process for the levy has resulted in a reduction of resources allocated to VETA centres, which, in turn, has stimulated an expansion in the provision of SCs by VETA-owned VET providers.
While private VET providers charge the full economic cost plus a profit margin for their LCs, VETA centres do not make any margin on LCs.This means that only SCs offer public VET providers an opportunity for internal resource generation, which can be used to increase teachers' salaries.The excessive use of SCs for income-generating purposes is problematic as resources funded by the SDL are used to deliver these courses, and so they potentially compete with LCs over rooms, machinery and teaching staff.Furthermore, teachers are drawn to centres that offer opportunities to deliver SCs, which thus reduces the number of qualified teachers available for the subsidised LCs.

Different skills needs and different training incentives
Over 97 per cent of the Tanzanian industrial sector comprises small firms with fewer than 10 employees.In the 2013 firm census, the last one conducted, only 1,322 large firms (with 10 + employees) were registered, of which 998 are manufacturing companies (NBS, 2016).These large firms operate in a handful of industries, mainly mining and manufacturing of food products and beverages, and are geographically concentrated in Dar es Salam and Morogoro.
While it is imperative to provide these large firms with a skilled workforce for productivity-enhancing structural transformation to be achievable, the majority of employers have very different skills needs.Hence, the quality and type of skills developed among the workforce do not necessarily meet the requirements of large productive organisations, while at the same time there are very few firms capable of organising and deploying skills effectively.The current VET system has been designed to satisfy social and economic objectives, some of whichbut not allalign with privatesector interests.The social objective arises around high youth unemployment.An estimated 850,000 young people enter the Tanzanian job market annually, but only about 50,000-60,000 formal-sector jobs are created each year.Given the limited absorption capacity of the formal sector, underemployment and informal-sector employment have seen a rising trend (LO/FTF Council, 2016).
The heavily subsidised LCs aim to address this issue of youth unemployment by keeping young people in education for an additional two to three years and by providing skills for both formal-and informalsector employment.LCs are also used to address other challenges arising from poor quality primary education in the country, especially for the most disadvantaged segments of society.While LCs provide a temporal fix, SCs are often preferred, especially by youth who want to establish their own, mainly informal, businesses (hairdressers, tailors, etc.).
Therefore, SCs have proliferated in Tanzania over recent years, both in response to demand from the private sector, and because of the existing incentive structure mentioned earlier. 3From the perspective of employers this practice has a bitter taste as many are preferring SCs for the provision of their own training needs over the subsidised LCs for which they then have to pay full fees on top of the SDL.

Rents capture and heterogeneity
The governance of existing VET schemes is made particularly difficult by the fact that private organisations hardly see concrete results from their financial contribution via the SDL.In many cases, this has been imputed to a problem of rents capture. 4Since mid-and late 2000, companies and their representative bodiesthe Association of Tanzania Employers (ATE), Tanzania Private Sector Foundation (TPSF), the Confederation of Tanzania Industries (Cti) and the Chief Executive Officer (CEO) Roundtablehave expressed dissatisfaction with the fact that the SDL is not used for vocational training only.They also reported that employers look at this situation as one of double taxation.
This claim stems from three main facts.First, the SDL is not ringfenced and tends to disappear in the Education Fund where it complements (and is potentially a substitute for) other central budget allocations for education.Second, companies pay additional costs if they want to access more focused SCs for training, upgrading and/or retraining their existing workers.Finally, employers perceive that VETA graduates are often not ready to work and that the amount of on-the-job training required to compensate for this lack of readiness is an additional cost to employers.At the same time, VET centres complain about the quality of their students when beginning their training and use this to justify the diversion of parts of the SDL to provide basic educational activities; see also Munishi (2016).
Addressing these critical challenges is complicated by the fact that not all firms engage with the skills gap problem and the skills institutions and financing process in the same way.Indeed, there is a heterogeneity problem alongside the problem of rents capture.
In many cases, private organisations hardly see the results of their financial contribution to VETA because firms themselves do not develop the managerial and organisational capabilities needed for them to be able to benefit from the improved skills; see Higuchi, Mhede, and Sonobe (2019) for an experiment on the impacts of management training in Tanzania.This is reflected in the conflicting claims around the SDL over the last decade in Tanzania (ATE (Association of Tanzania Employers), 2011; interviews in 2018 with Cti, TPSF and the CEO Roundtable).Among smaller and less productive enterprises, references to 'soft skills' are often made because employers themselves lack the organisational capabilities to assess the specific types of skills that are required to reach certain products and raise quality standards.This explains why the World Bank (2017) (2017) results do not appear to point to a skills crisis in Tanzania.On the contrary, in key areas such as problem-solving and job-specific skills, most firms reported that their workers reach the required skill levels. 5oreover, while a few larger companies have invested in on-the-job training, the majority expect VET centres to provide ready-made skills.There appears to be widespread lack of awareness that companies/employers in all countries have to provide on-the-job training. 6This lack of ready-made skills in Tanzania's domestic market has often pushed a number of companies in different directions.Some have lobbied for SCs that can provide more tailored skills training.Other companies have simply ignored regulations on work permits and have employed foreign workers to circumvent the lack of domestic skills.
The fact that private-sector representative organisations have been unable to first, challenge the government with hard evidence on the total amount of SDL collected from employers and second, call the government to account for the funds allocated to VETA, suggests that many employers are concerned that they will expose their own potential lack of SDL compliance in the process. 7Anecdotal evidence suggests that firms underestimate their total payroll and engage in employment strategies that reduce their payroll to avoid paying the full SDL, with potential implications for increasing formal employment.
While these differences reveal a plurality of interests and attitudes among businesses and can be seen as a problem, they can also be seen as opportunities for feasible change.So far, the debate around reforming the skills sector has been mainly driven by mistrust between government and businesses and the idea that the only solution is a reduction in the SDL.However, while this reduction in the levy might reduce (or shift) the problem of rents capture, it would result in a decline in the provision of VET programmes and their inclusiveness.Instead of implementing a one-size-fits-all solution to rents capture in the skills sector, we explore potential strategies that take into account firms' differences in capabilities and leverage their different incentives.

Realigning incentives to account for differences: Our hypotheses
Even though both the government and companies have a strong joint interest in developing skills, conflicting dynamics and allegations of resource misallocations, corruption and tax avoidance have prevailed for years.We make two main hypotheses around the existence of latent preferences (and interests) across manufacturing firms in Tanzania which can be leveraged to escape this low-productivity skills trap.Specifically, based on their different preferences and underlying interests, firms will respond to different 'incentive packages'.We call it a 'package' in the sense that each incentive package combines different features of the VET policyfor example its financing or delivery models to which different firms are expected to react differently based on their characteristics.
We hypothesise that preferences differ by firm characteristics and broadly distinguish between two main types of firmsi.e.'high capabilitymore productive' and 'low capabilityless productive' firms.This categorisation is based on 6 different firm characteristics which are linked to productivity: size, technical capabilities, demand for skilled labour, human resources or organisational capabilities, and labour intensity of the production process.For each firm in our study, we collect evidence on these characteristics.This allows us to match revealed preferences from the experiment with firms' characteristics.

H1: High capabilitymore productive (HCMP) firms are willing to pay the SDL in exchange for tangible improvements in the quality of the skills supply coming from VET centres.
The reason for this is that HCMP firms are not concerned about reducing their tax base per se: they are relatively more concerned about getting access to adequate skills domestically, and hence, to see an improvement in the VETA-run system.It is indeed in their interest to have better skills, because these firms know how to deploy them effectively and profit.Such firms can also reward higher skills because they deliver productivity increases.
We expect that these HCMP firms are more willing to commit resources jointly with VETA in exchange for changes in the design of the schemefor example, more sector-specific and technology-tailored trainings.More specifically, we expect that there are some capable firms that are willing to collaborate with VETA in the design of training, including designing 'long-enough' courses, contributing to on-the-job training and responding positively to ex-post incentives.Based on the existence of a skills gap; e.g.see Tan et al (2016). 6Unfortunately, there has been no attempt to estimate the extra on-the-job training that companies have to provide in different sectors in Africa.We have attempted this in our analysis to allow comparison with common practices in other countries and across sectors; see section 4.1.
positive experience of some VET centres (e.g., Moshi VET 8 ), we also assume that there are VET-centres willing to adapt to these changes and that this has not happened so far because the SDL financing mechanism does not allow for such realignment of incentives.

H2: Low capabilityless productive (LCLP) firms have a strong preference for a reduction in the SDL and are unwilling to discount relatively higher levies for more co-designed solutions or sectorspecific training.
Although relatively less capable firms might have latent preferences that are significant for different skills-related dimensions, we expect that these organisations' main concern is about a reduction in the SDL, stemming from the fact that they cannot really benefit from higher-level skills as they do not know how to deploy them effectively, and they are relatively more cash constrained due to smaller profit margins.This means that these LCLP firms will reveal an incentive package that is always skewed towards a reduction in direct contributions (monetary and in-kind) that has no conditionality attached.
For LCLP firms, an incentive package designed for HCMP firms is not a feasible strategy to reduce rents seeking.If we design the same incentive package for both types of firms, it will result in debasing the SDL, further weakening VET centres and no improvement in collaborations between productive organisations and the centres.The rents that are allegedly captured by the state and used unproductively will simply be shifted towards the private sector with no positive development outcome.First experiences in this direction can be seen in action, in relation to some schemes in which adverse selection has led to crowdingin the 'wrong' type of firm, that is, those with the wrong type of incentives.For example, in some sectors, LCLP firms joined the government promoted on-the-job training pilot schemes (National apprenticeship guidelines, 2017) for the wrong reason, that is, simply to replace their low-skilled workers with unskilled trainees whose minimal salary was subsidised.These companies had no real interest and capability to train the trainees and, indeed, hire them after their training period.

Research design
A prominent tool to uncover latent preference structures among different stakeholder groups is the Discrete Choice Experiment (DCE).This method is widely used in transport and health care studies, but less so in social science.A standard random utility model builds the theoretical foundation for the DCEs (Hole, 2007;Lancsar, Fiebig, & Hole, 2017).Assume that the utility accruing to decision-maker n if choosing alternative j can be described as: with V nj being a function of observed attributes of the alternative j(x nj ) and the decision-maker n(z n ) and ε nj being assumed to be random.Then the probability that the decision maker n choses alternative i over j is given as: The utility is approximated by a function that is linear in parameters, so that: A mixed logit model is chosen, due to the possible heterogeneity in preferences in the sample.Mixed logit models account for latent scale and preference heterogeneity by allowing estimated parameters to vary around their mean, so that β in (3) is defined as: with η j being the random variation around the parameter mean β (Lancsar et al., 2017).The mixed logit model has the additional advantages that it accounts for the panel structure of the DCE data, provides more reliable standard errors and moves away from the proportional substitution assumption that is often violated (Hole, 2007).In a second step, indices, that capture different dimensions of firms' productive capabilities are constructed by use of principal component analysis (PCA).The indices are then interacted with attribute levels to capture heterogeneity in preferences across observed characteristics, in line with our hypotheses.
It is important to note, that estimation results are indicative of preferences for a set of characteristics that could define policies, however, one cannot infer behavioural changes from these revealed preferences (Brathwaite & Walker, 2018).This means, that once a specific policy is implemented, the uptake and effect of the policy might differ from the revealed preference.However, the aim of the DCEs is to identify policy options that are feasible for and desired by different types of firms and therefore assist the identification and design of potentially effective policy packages.

Experimental design
DCEs require the identification of relevant attributes (e.g. the size of the skills levy) and attribute levels (e.g.2.5 %, 3.5 %, 4.5 %, 5.5 %) to construct choice packages relevant for the research objectives.In this study, the identification of attributes has been guided by the research team's knowledge of the Tanzanian skills sector as well as months of semi-structured and unstructured interviews throughout 2018 and 2019 with business organisations, government officials and firm representatives in Tanzania (See Appendix A1 for a complete account of this preliminary analysis).
Three separate DCEs are created that focus on different contested policy elements, namely: (1) the design of the training, (2) the nature of the skills levy, and (3) the design of incentives to firms and VETA for 8 In 2009 the Tanzanian Chamber of Minerals and Energy (TCME) in partnership with its members -African Barrick Gold and Anglo Gold Ashanti initially, later joined by other six sponsors (Sandvik Mining, Kabanga Nickel, Pan African Mining Services, Shanta Mining, Mantra Resources (Uranium One) and Atlas Copco) -and the VETA Moshi campus launched the Integrated Mining Technical Training (IMTT) project.TCME member companies made US $2.4 million available towards the project implementation with a focus on: a) converting/refurbishing some of the existing VETA facilities; b) building a completely new training workshop and several classrooms; c) retooling/ upgrading of the existing welding, electrical and motor vehicle mechanic workshops; d) training of trainees (six sent to South Africa to upgrade their skills by attending three-month tailormade 'Train the Trainer' programmes); e) exposing Trainers to higher training standards by undergoing staff development at mines; f) recruiting a project manager from an overseas partner organisation to oversee implementation of the programme; and g) guaranteed maintenance of equipment by TCME.The IMTT initiative started in response to the shortage of skills required by the highly technical mining operations in Tanzania and the increasing salaries that qualified miners were managing to secure with the opening of new mines.The programme provides a three-year training course for students who have successfully passed lower secondary school as a minimum.After a generic four-month skills course in the use of hand and measuring tools, basic machine and drawing, students start on a continuous rotation between theoretical, practical and on-the-job training, which continues until they qualify at the end of the third year.Successful students receive a NACTE level four certificate at the end of the course in their area of specialisation.Five areas are offered: electrical, fitting/fitter mechanics, welding/plater welding, diesel mechanics/heavy moving equipment, and auto electrical.TCME manages implementation of the programme and student scholarships.Companies provide scholarships of US$1,500 for each student, of which US$600 are transferred to VETA Moshi to cover full board and a small salary.The remaining resources are retained by TCME to maintain and upgrade training equipment, to train trainers and to provide training incentives.Since 2013, the programme has increased its student enrolment numbers to 200 annually, and many graduates are in high demand in other sectors as well (Andreoni, 2017).
A. Andreoni et al. skills provision.Attributes (in bold with grey background) and attribute levels (bullet points) are summarised in Table 1.The Size attribute is part of all three choice experiments.This allows direct comparison of the preferences identified by estimation of the relative willingness to pay (WTP).Size and Length are coded as continuous variables in the model estimation.Non-linearities are explored by considering quadratic forms of the continuous variables.
We generated a seven-task d-efficient design for each DCE with zero priors.Participants were asked to rank three alternatives in each task.This brings the DCEs to a total of 21 tasks each.The design was motivated by an effort to balance boredom with statistical power and enable participants to take breaks between the DCEs.Entries were dummy coded so that coefficient estimates measure the strength of preference of the attribute level relative to the omitted level of the attribute.
Participants were further asked to rank levels of each attribute before conducting the DCEs.The purpose of the ranking was (i) to give participants the opportunity to familiarise themselves with the different attribute levels that were subsequently used in the DCEs and (ii) to provide a quick sense check on revealed preferences as well as an easy opportunity to gauge firm characteristics across which preferences might differ.Enumerators were advised to probe participants on their understanding of attribute levels during the ranking exercise by asking them to provide reasons for their ranking choice.
The ranking exercise and DCEs were further complemented by a comprehensive questionnaire to identify firm characteristics.The questionnaire was divided into three parts which required completion by different personnel.Part A covered general characteristics of the organisation including location, main activities, and ownership structure.Part B contained the ranking exercise and the DCEs.Part C covered characteristics of the company around: (a) the company's productive capabilities and competitiveness and (b) employment and skills.
Building on our two main hypotheses presented in the previous section, Table 2 summarises the expected revealed interests-incentives scenarios by firm types.

Sampling strategy and data collection
Tanzania's Annual Survey of Industrial Production 2016/17 was used as sampling frame and was cross-referenced with data collected by REPOA for the Global Competitiveness Report and the most recent list of companies registered with the Business Registration and Licencing Authority (BRELA).Firms across 10 manufacturing sectors with more than 20 people engaged, located in six regions in the northeast of the country were selected: Arusha, Dar es Salaam, Iringa, Kilimanjaro, Morogoro and Tanga. 9This sample was then complemented by several smaller firms located in Dar es Salaam (See Appendix A2 for a complete account of the sampling strategy and data collection).
The six regions combined cover about 82 per cent of total manufacturing value added in Tanzania, and hence the selection provides a cost-efficient way to cover a substantial share of large manufacturing companies in the country (NBS, 2016).The choice of manufacturing sectors was motivated by our aim to cover sectors across different technological classifications; low-tech, medium-tech and hightech. 10 A total of 178 large and medium size firms and 38 small firms were reached, resulting in a total sample of 216.Seven of these firms were excluded from the analysis as the missing data share was too large, leaving a sample size of 209 (See Appendix A3 for more detail on data cleaning).In line with our sampling strategy, the largest shares of firms in the sample are from Dar es Salaam (62 %), Arusha (14 %) and What would be an appropriate size for the skills levy?
What would be an appropriate size for the skills levy?Source: The authors. 9 Tanga includes Coast as a newly established region. 10As defined by the United Nations Industrial Development Organization (UNIDO); see https://stat.unido.org/content/learning-center/classification-ofmanufacturing-sectors-by-technological-intensity-%28isic-revision-4%29.
Morogoro (11 %); the three regions that rank highest in terms of total manufacturing value addition (NBS, 2016).

Research results
Details about employees' skills profile and the amount, type and costs of training provided by employers sampled are summarised in subsection 4.1.Sub-sections 4.2.and 4.3 summarise the DCE results, first presenting average results and then differentiating by observed firm characteristics.

The skills gap and training provision
In line with previous reports, English language and IT skills are assessed by employers as being the least adequate across all employee categories (WB 2017;Andreoni, 2018).This is followed by critical thinking and writing skills, and then problem-solving skills.Non-skilled workers are rated the most deficient in all skills categories, while managerial and professional staff skills are rated mostly adequate.In all three employee categories interpersonal, work ethic and job-specific skills are rated the highest; see Fig. 1.
The perceived skills gap varies not only across employee categories but also across the type of firm.Smaller Tanzanian-owned companies are more heavily affected by the perceived skills gap in English and IT.These companies might be unable to attract highly skilled labour or tap into overseas labour markets to satisfy their skills requirements.Interestingly, job-specific skills are particularly highly rated in terms of adequacy for companies with fully automated production processes, possibly indicating the importance of on-the-job training for these firms to operate production technologies.
Fig. 2 supports the conjecture about the varied access of firms to skilled labour due to differences in ownership.While the overall share of foreign workers in the total workforce is small, it is significant for managerial roles and roles requiring education in science, technology, engineering, and mathematics (STEM).A quarter of all STEM positions on average are filled by foreign workers.As expected, the share of foreign workers in all categories is higher for foreign-owned firms or joint ventures, indicating easier access to overseas labour markets.
Asked about where employees gathered relevant work experience before filling their current post, employers indicated that most employees have gained experience within the same company.The relatively higher share of employees recruited internally, 45-55 per cent across employee categories, indicates that trained employees can and are being retained.About 61 per cent of the recently employed VET graduates required additional training and more than 66 per cent of the sampled firms indicated that they provide employees with training.
The training received differs by employee type (see Fig. 3a).Most training is non-certified and is delivered in-house either in the form of shadowing, mentoring and supervised work or in the form of structured but non-certified training programmes.Structured training programmes are used for operative non-skilled workers, while for managerial and professional staff training is mainly received through shadowing, mentoring and supervised work.Interestingly, a substantial share of operative skilled workers receives additional certified training through VETA SCs, the costs of which is covered by the firm.
The objectives of the training provided also differ by employee type.The main purpose for managerial and professional staff is professionalism and soft skills, followed by general productivity-enhancing skills (Fig. 3b).For operative skilled and non-skilled workers, the main purpose is productivity-enhancing skills.For unskilled workers industryspecific knowledge is another important objective.The average duration of training provided ranges between three and four weeks across all employee types.The average costs for training provided are highest for managerial and professional staff. 11The average training costs if covered internally are substantially higher than costs that are covered externally (Fig. 3c).
Overall, the data suggests that firms make substantial investments into further training for their employees.Further, SCs are in much higher demand than LCs for post-employment training, which explains why firms voice dissatisfaction over the need to invest in further certified training.Yet, with only 66 per cent of the sampled firms reporting that they provide training for their employees, these issues affect some but not all firms.

Preference structure
The DCEs reveal latent preference structures for sampled firms with respect to: (1) the design of training, (2) the nature of the skills levy, and (3) the design of incentives for firms and VETA for skills provision.Complementing these results with insights from the questionnaire, we can evaluate how these preference structures vary based on observed characteristics, as hypothesised in section 2. For ease of reading the subsequent figures and tables, Appendix B Table 2 provides a glossary of all variable names and their description.
The ranking exercise identifies a consistent preference structure across firms for length of training (two years is preferred), placements (a greater share of work placements is preferred), and size of the levy (unsurprisingly, the lowest option of 2.5 % is preferred).Results are summarised in Fig. 4. As hypothesised, preferences appear to vary with the observed characteristics.For instance, smaller firms and those that mainly rely on manual production processes prefer the SDL to be based on profits rather than payroll and show a strong preference for subsidising SCs.Larger firms and firms that rely more heavily on semiautomatic and fully automatic production processes instead prefer the levy to be based on payroll and VET centres to retain the profits from these SCs or be rewarded for each student who gets a full-time job.
These preference structures are largely confirmed in the DCEs.Results are summarised in Table 3.For the dummy coded attributes, the Source: The authors.
11 High costs are incurred as managerial and professional staff are often sent abroad for specific and highly technical training.
A. Andreoni et al. size of the coefficient shows the relative preference for a particular level over the omitted level.A clear preference for a smaller skills levy, as expected, is revealed.The skills levy estimates are strongly significant and consistent throughout all three DCEs, which enables us to derive WTP estimates, making the preference structures comparable across all three experiments (Hole, 2007;Lancsar et al., 2017).The WTP coefficient can be interpreted as the amount in percentage of the SDL that the average firm is willing to sacrifice to gain utility (positive coefficients) or avoid disutility (negative coefficients).
Fig. 5 visualises the WTP estimates for all three experiments combined with their respective standard errors.A significant preference for longer training durations with declining marginal utility (maximum utility is reached at 22 months) is revealed.Further, a strong and significant preference for a greater share of work experience over theoretical training and for investment and joint investment in training staff and machinery over providing access to machinery is revealed. 12Results for the remaining attribute are less pronounced and largely insignificant (see Table 3).Overall, there appears to be a small preference for customised curricula, for annual frequency of payments, for the levy to be allocated to VETA and used primarily for sector-specific training, for an upfront reduction of the levy if employing a recent VETA graduate, and for VETA centres to receive a subsidy to provide SCs customised for  individual firms over the alternatives.

Preference structure by differences in firm characteristics
The results from the mixed logit estimation suggest heterogeneous preferences; see the number of significant η j estimates in Table 3.Hence, in a second step, we evaluate potential heterogeneities in preferences across observed characteristics.Based on our hypotheses, we construct different measures that capture dimensions of what we term productive capabilities.Specifically, we construct five indices by means of PCA: (i) size, (ii) technical capabilities, (iii) skills demand, (iv) human resources, and (v) labour intensity (See Appendix A4 for further information about the index construction).We expect HCMP firms to be larger, technically more capable, with a higher level of human resources and less labourintensive production processes (and the reverse for LCLP firms).These indices are interacted with attribute levels of the three DCEs.Interactions are summarised in Fig. 6.Indices are divided into quartiles with the top and bottom 5 per cent cut to avoid distortions due to outliers.The horizontal line divides the space between utility (top) and disutility (bottom).As before, WTP estimates measure the strength of preference of the attribute level relative to the omitted level of the attribute.Results are reported for each DCE separately.The tiles are organised by attribute (column) and by index (row).Where standard errors include the zero line (dotted), WTP estimates are insignificant.Where WTP estimates are relatively invariant across index quartiles, differences by the observed characteristic are insignificant.
According to Fig. 6, technically capable and large firms have similar preference profiles and so have labour intensive firms and firms faced with a shortage of skilled workers.Some preferences are also universal across firms.Firms agree on a stronger focus on field placements over theory as part of the curriculum and show a strong dislike for the proposal to provide access to machinery and equipment for training purposes.The degree of dislike increases with the sophistication of the machinery.Further, smaller and less technically capable firms show indifference over most attribute levels.
Larger and technically capable firms prefer being reimbursed for incurred training costs, while smaller and technically less capable firms prefer an upfront reduction of the levy for employment of a recent VETA graduatewhich suggests that larger and technical capable firms are already engaged in training provision.Large firms and those that rank highly in the human resources index show a strong preference for a customised curriculum co-developed by individual firms and VETA over the alternatives and a preference for most of the SDL to be used for sector specific training.Larger firms also prefer a joint investment in training as part of a sectoral council and VETA centres to provide customised SCs.Less technically capable firms prefer the status quo instead, which is SCs to be offered by VETA centres at a competitive rate.
Lager, more technically capable firms and firms with more labourintensive production processes show a preference for the levy to be paid annually, while firms that rank highly on the human resource index, have a slight preference for monthly payments.The latter is explained by the group of firms that rank highly on the human resource index being dominated by foreign firms.Monthly payments of the SDL as part of payroll are preferred by foreign firms as this aligns with their bookkeeping, while domestic firms prefer annual payments. 13Unsurprisingly, labour-intensive firms strongly prefer the SDL to be based on profits, while more technically capable firms prefer the SDL to be based on payroll instead.
Firms with a strong demand for skills and firms with labour intensive production processes have a strong preference for shorter training courses while firms which rank highly on the human resource index prefer longer training courses.Specifically, for firms ranking highest on the human resource index, the preferred duration is 21 months, while for those with the highest skills needs, the preferred duration is 16 months.Firms with a strong demand for skills signal willingness to invest in training resources (but not as part of a sectoral council), prefer to be reimbursed for training costs incurred and show a preference for customised SCs to be offered by VETA centreswhich suggest that these firms are already investing in training and would welcome some of the training to be provided by VETA.
In a final stage of analysis, we use training provision as a dummy interaction term to exploit heterogeneities in binary characteristics.Fig. 7 summarises the results.Firms that provide training prefer the SDL to be paid annually, the SDL to be allocated to VETA with the majority being used for sector-specific training, have no dominant preference for firm incentives for skills provision (this is expected, since they provide training regardless), 14 and prefer a joint investment in training and teaching staff as part of a sectoral council.Firms that provide training also show a stronger preference for a focus on work placements over theory than those that do not provide training.
The results suggest significant heterogeneity across firms.Firms that Fig. 5. WTP estimates and their standard errors.Source: The authors. 13We interacted the attribute levels with an ownership dummy (foreign, domestic) and returned highly significant results which confirm this conjecture. 14Interestingly and expectedly, firms that do not provide training prefer an upfront reduction of the levy in return for employment of a recent VETA graduate over being reimbursed for training provision.
A. Andreoni et al. provide training welcome the introduction of a sectoral council and the provision of more tailored training courses.These preferences align with firms that have previously been identified as large and highly technically capable (HCMP firms).Such firms also show a preference for the levy to be based on payroll rather than profits and to be reimbursed for training costs rather than an upfront reduction of the levy.Despite the differences in preference structure overall, firms agree that greater emphasis should be placed on field placements as part of standard VET courses,  A. Andreoni et al. that firms should not provide access to their machinery for training purpose and that the duration of courses should not exceed two years.

Designing for differences: An anti-corruption strategy for Tanzania's skills sector
Reforms in the skills sector have been slow in Tanzania and this has made it difficult to improve development outcomes in a key sector of the economy.While all stakeholders recognise the skills sector as a priority, there has been increasing mistrust between the government and the private sector.Allegations of SDL misallocation and double taxation from the private sector have resulted in frustration and have cascaded in chains of rule-breaking behaviours.The high level of the SDL has also acted as a dis-incentive towards formal employment.On the other hand, accusations of large-scale tax avoidance have made it difficult to improve VET-centre outcomes, and several informal practices have emerged at that level along with parallel semi-formal streams of financing.
The general policy response to this low-productivity skills trap has seen the government shifting from a position in which 'vertical enforcement' of the rules (e.g., fighting tax avoidance) has been preferred (indeed, SDL tax collection has increased), to one in which monetary incentives have been introduced in an undifferentiated manner (indeed, incentives have gone to both HCMP and LCLP firms). 15 Our firm-level evidence shows why these efforts have not generated the expected outcomesthe government did not take firms' differences sufficiently into account, hence it failed to re-align the incentives of key stakeholders.
Results from our DCEs, drawing on a stratified sample of more than 200 Tanzanian companies, reveal substantial heterogeneity in preferences across observed firm characteristics.The main differences relate to different dimensions of productive capabilities.
HCMP firms support sectoral councilshence more sector-specific skills development and joint commitment of fundingand show willingness to engage with curriculum design in collaboration with VETA.These firms also put great emphasis on work experience as part of training.Crucially, these firms are less concerned about the size of the levy or with other financial incentives without improvements in outcomes.This means that HCMP firms would be willing to support an incentive package that prioritises the quality of VET outcomes over a reduction in the cost of the scheme.
In contrast, the preferences revealed for LCLP firms appear skewed towards an incentive package that prioritises a reduction in the costs of the VET scheme while remaining indifferent to most other attributes studied.These firms are happy with most of the levy going to VETA and do not have an interest in a sectoral council or sector-specific training provision.
Against this evidence, an effective strategy to unlock failings in the sector and to enforce rule-following behaviour through horizontal  Source: The authors. 15The distinction between 'horizontal' and 'vertical' enforcement is introduced in (Khan et al., 2019).Differently from top-down approaches to policy enforcement, horizontal enforcement leverages peer-pressure and differences in incentives and characteristics across stakeholders.
A. Andreoni et al. enforcement should be based on an incentive package designed around the preferences of HCMP firms (see Table 4).Such a package would not only crowd-in the 'right' type of firmsthat is, high capability firms that are crucial to lift the sector, jointly with VETAit would also crowd-out low capability firms that are mainly interested in a reduction in the SDL.
The opposite strategyan incentive package that prioritises the costs of the VET schemewould be in line with the preferences of LCLP firms but it would not be effective in crowding in the HCMP firms.It must be noted that the incentive package targeting HCMP can also help overcoming some of the search and matching frictions.With HCMP companies being more directly involved in the VET design and deliveryespecially through on-the-job trainingworkers have a chance to signal HCMP firms their level of skills and potential contribution to enhanced productivity.Simply enhancing the quality of VET training, without a more direct involvement of firms would not achieve the same outcomes, and search and matching frictions would remain high.
An incentive package prioritising the quality of VET outcomes has three advantages.First, it is feasibleit will allow the emergence of coalitions of interests across the private and public sector willing to support high-quality VET outcomes.Second, it is effectiveit aligns the interests of the powerful organisations with the desired development outcomes.Finally, it is enforceable and less vulnerable to resource misallocation and tax evasionstakeholders at the sectoral level will have a strong incentive in peer monitoring the flow of resources and in making sure they reward productive arrangements between companies and VET centres.
While selecting an incentive package that prioritises the quality of VET outcomes and is aligned with HCMP firms, the government will need to implement complementary measures to support LCLP firms reaching minimum capability thresholds and crowd-in these firms incrementally via extension services, and training of entrepreneurs.These measures must consider the specific characteristics of the LCLP firms and address the specific challenges that they face.The government must offer such firms a pathway to become more capable and to contribute more to skills development in the country.
These complementary measures for LCLP firms (Table 4, bottom left) should have three objectives: First, VET centres should work to raise awareness of sector-and technology-specific skills and their benefits.This could help low capability firms to realise the potential of joining sectoral council solutions, being involved in the co-design of skills development programmes, and contributing to the SDL.In turn, this would establish a different dialogue between VET centres and firms, particularly around the proliferation of SCs and their formalisation and transparency of VET centres.
Second, VET centres could partner with existing domestic intermediate technology institutions16 to provide manufacturing extension services and improve organisational capabilities and operational design solutions so that skills can be deployed effectively.Less than one quarter of firms in Tanzania have ever interacted with these organisations, another example of the problems faced by LCLP firms (MITI and UNIDO, 2016).
Third, dynamic incentives with conditionalities attached could open pathways for LCLP firms to graduate towards high capability status.This could be achieved using time-limited incentives that reward matched investments and other investments in the development of organisational and technical capabilities.The incentives could link also to the extent to which companies retain employees and are willing to invest in their workforce along the graduation journey.
In sum, there is evidence that an incentive package that combines prioritising the quality of VET outcomes (tailored around the preferences and characteristics of high capability firms) with complementary measures (tailored around the characteristics of low capability firms) that offer pathways to graduation from low to high capability status can deliver an effective strategy for the Tanzanian skills sector.This strategy has the merit of horizontal enforcement mechanisms, realigned incentives for key stakeholders and options for those firms willing to be rule-following.The overall outcomes would be one of incremental improvement in this key sector of Tanzania's economy.

Declaration of competing interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
.Following this augmented sampling strategy, a total of 178 large and medium size firms and 38 small firms were reached, resulting in a total sample of 216.Seven of these firms were excluded from the analysis as the missing data share was too large, leaving a sample size of 209 (See Appendix A3 for more detail on data cleaning and missingness).In line with our sampling strategy, the largest share of firms in the sample are from Dar es Salaam (62 %), Arusha (14 %) and Morogoro (11 %); the three regions that rank highest in terms of total manufacturing value addition (NBS, 2016).
For almost all sectors, the number of medium and large firms in our sample exceeds the number in the pre-selected sample.Some firms have only recently started operations and hence are not included in the 2016/17 survey.However, the great majority of firms not covered in the sample frame have grown over recent years and have since crossed the threshold of 20 or more people engaged.With the coverage of sectors expanded, we were able to reach a more balanced sample in terms of technological classification with 55 % of firms falling into the low-tech category (70 % in the pre-selected sample).
The DCEs and the questionnaire were piloted in January and February 2019 and the DCEs, questionnaires and sampling strategy were finalised in March 2019.Between June and August 2019, questionnaires and DCEs were coded using Open Data Kit17 which enabled data collection using tablets.REPOA recruited 12 research assistants from its own database for enumeration.Enumerators went through a two-day intensive training course in early September 2019.After completion of the training, the survey was piloted again with selected firms in Dar es Salaam that met the sampling criteria but that were not part of the sample.

A3. Data cleaning and transformation
The questionnaire is divided into three parts: A, B and C.These parts were answered by different staff, at different points in time, and were later matched by a unique identifier.The unique ID of the device on which the answers were recorded (held by the enumerator) as well as a timestamp of the entry were also used to match the parts of the survey.This measure was introduced in case errors were made when the unique company identifiers were recorded.Appears twice throughout for different companies.Matched the first entries with ID_35 and the second with ID_60 judging by the time stamp when the data was collected.

GM005
Appears twice in C with entries identical up until question 3.2.Afterwards entries differ.We selected the first entry as the data was more in line with data from question 3.1.
Source: The authors.
Matching left us with 216 companies that responded to all three parts of the questionnaire.Part A and B were complete for these companies.However, Part C suffered from missing observations for several companies.Note that some questions with qualitative responses are missing from Table A3.2.Out of these 216 companies, 4 companies did not respond to any question in Part C and another 3 did not respond to any questions regarding employees in Part C (Section 3).We excluded these seven companies.This leaves a set of 209 companies with complete responses for most variables.

A4. Productive capabilities index construction
Table A4.1 provides a summary of the variables used for index construction via PCA.

Fig. 1 .Fig. 2 .Fig. 3 .
Fig.1.Perceived skills gap by employee category.Note: Average of assessed adequacy of skills: − 1 below required, 0 adequate, 1 above required.A breakdown of the data is provided in Appendix B Table1.Source: The authors.

Fig. 4 .
Fig. 4. Average ranking of attribute levels.Note: Rankings are adjusted to scale, ranging from top (first choice) to bottom (last choice).Source: The authors.

Fig. 6 .
Fig. 6.Changes in WTP estimates with firm characteristics.Note: The horizontal line divides the space between utility and disutility.The error bars are standard errors.Source: The authors.

Fig. 7 .
Fig. 7. WTP estimates by training provision.Note: Only significant interaction effects are depicted.Bars indicate standard errors.Light grey (N) = firms with no training provision; dark grey (Y) = firms with training provision.Source: The authors.
This publication is an output of the SOAS Anti-Corruption Evidence (ACE) research consortium funded by UK aid from the UK Government [IATI Identifier: GB-1-203752, Contract P07073].The views presented are those of the author(s) and do not necessarily reflect the UK government's official policies or the views of SOAS-ACE or other partner organizations.For more information on SOAS-ACE visit www.ace.soas.ac.uk.CRediT authorship contribution statementAntonio Andreoni: Conceptualization, Data curation, Formal analysis, Funding acquisition, Investigation, Methodology, Project administration, Resources, Software, Supervision, Validation, Visualization, Writingoriginal draft, Writingreview & editing.Sophie van Huellen: Conceptualization, Data curation, Formal analysis, Funding acquisition, Investigation, Methodology, Project administration, Resources, Software, Supervision, Validation, Visualization, Writingoriginal draft, Writingreview & editing.Lucas Katera: Conceptualization, Data curation, Funding acquisition, Project administration, Resources, Software, Supervision, Validation, Writingreview & editing.Cornel Jahari: Data curation, Project administration, Resources, Software, Supervision, Validation, Writingreview & editing.

Table 1
DCE attributes and attribute levels.

Table 3
DCE estimation results.

Table 4
A feasible anti-corruption strategy for the skills sector.

Table A2 .1
Firms selected by sampling strategy and firms reached.Notes: For instance, the initial sampling strategy taking the Tanzania's Annual Survey of Industrial Production 2016/17 as sample frame resulted in the selection of 16 large firms and 4 small firms in the Manufacture of beverages sector.The selected firms covered 40 % of all large firms and 22 % all firms in the sector, 61 % of value added by large firms and all firms in the sector, and 55 % and 54 % of total number of employees by all large firms and all firms respectively in the sector.The revised sample frame resulted in the selection of 20 large firms and 1 small firm resulting in 21 firms interviewed that operate in the sector, making up 10.05 % of the total sample.) Medium and large firms are those with 20 or more employees.Small firms are those with fewer than 20 employees.Firms with no information about number of employees are counted as small firms also.d) Industry affiliations have been established by drawing from information provided in Part A question 1.4 with Part C question 2.1 and with reference to ISIC Ref. 4: https://unstats.un.org/unsd/classifications/Econ/Download/In%20Text/ISIC_Rev_4_publication_English.pdf .

Table A3 .1
Matching parts A, B, C of the questionnaire.and hence not matched (A only has GM 001 but the time stamp does not match, hence both are excluded) JM002

Table A3 .2
Missingness of data for part C of questionnaire.

Table A3 .3 provides a full breakdown of the sample by size of firm, sector and technological classification in comparison to the targeted sample. Table A3.3
Firms reached by sector and size.
(continued on next page) A.Andreoni et al.

Table A3 .
3 (continued ) Notes: a) Sectors included in the sampling strategy are highlighted in bold.Industry affiliations have been established by drawing from information provided in Part A question 1.4 with Part C question 2.1 and with reference to ISIC Ref. 4: https://unstats.un.org/unsd/classifications/Econ/Download/In%20Text/ISIC_Rev_4_pu-blication_English.pdf.b) L: low-tech, M: medium-tech, H: high-tech.c) Medium and large firms are those with 20 or more employees.Small firms are those with fewer than 20 employees.Firms with no information about number of employees are counted as small firms also.Source: The authors.

Table A4 .1
Capability indices, their interpretation and input variables for construction.

Table B2 (
continued ) In your opinion, which incentive would make VETA centres most responsive to your company's needs and improve the quality of training?Prof VETA centres retaining the profits from offering SCs at competitive rates Sub/ Substi VETA centres receiving a subsidy for providing SCs customised for individual firms Reward VETA centres receiving a financial reward with each student who gets a full-time job