Human Capital Spillovers and Plant Productivity in ASEAN: Evidence from the Plant Panel Data

doi:10.21203/rs.3.rs-3548224/v1

Download PDF

Research Article

Human Capital Spillovers and Plant Productivity in ASEAN: Evidence from the Plant Panel Data

https://doi.org/10.21203/rs.3.rs-3548224/v1

This work is licensed under a CC BY 4.0 License

Journal Publication

published 01 Apr, 2024

Read the published version in Humanities and Social Sciences Letters →

Version 1

posted

You are reading this latest preprint version

This paper examines whether the local human capital spillovers can affect plant productivity in three ASEAN countries using panel data from World Bank Enterprise Survey between 2009 and 2015. The spillovers exist if plants can produce more in the region with the abundant number of skilled workers given their inputs. This study focuses on the proportion of nonproduction workers outside the plant as the proxy of skilled workers. I use augmented Cobb-Douglas production function to estimate this effect. Results suggest that human capital spillovers do matter for plant productivity. I find that a 1 percent increase in the proportion of nonproduction workers within a region will enhance the productivity of plants in that region by roughly 7.1–8.9 percent. Moreover, the productivity rises are found to be higher for low-tech plants. The economic value of such spillovers effect indicates that per plant value-added will increase by US$ 330,000 or correspondingly, an increase of US$ 866.75 million for the whole manufacturing industry.

JEL codes: D24, J24, L25, L60

Human Capital Spillover

Plant Productivity

Manufacturing Sector

ASEAN

Economists have long identified that knowledge spillovers in the form of human capital are crucial to determine the long-run economic growth (Lucas, 1988). The accumulation of human capital can be acquired through schooling or learning-by-doing. In the context of the latter, Marshall (1920) emphasised that social interaction among workers in the same location could generate learning opportunity to implement productivity-enhancing ideas which increase firm productivity. The quality of the pool of workers will determine the success rate of such collaboration. For example, if firms located in an area with more educated workers, it is easier for them to hire workers who are able to use firm’s asset in a very efficient way. To achieve this goal, a way that those workers can do is to learn new knowledge and information from other skilled workers within the location. A close regional proximity will make easier for them to interact and share those things because it increases the possibility through face-to-face interaction. However, as elaborated by Rosenthal and Strange (2008), the impact of such activity is inversely correlated with spatial proximity. In other words, the exchange of ideas will sharply attenuate when the distance among skilled workers increases. This fact implies plants location does matters because the availability of more educated workers within the area is an important element for them to acquire new knowledge and as a result, affect their productivity.

In this paper, I examine the presence of local human capital spillovers on plant productivity for manufacturing industries in three ASEAN countries. The story is straightforward and easy to see. Plants are expected to produce larger output if they are located in particular region with high level of human capital for any given input rather than similar plants located in an area where the human capital level is low. To test this premise, I specifically take a more direct approach by estimating augmented Cobb-Douglas production function at plant level using World Bank Enterprise Survey for the year of 2009 and 2015.

I measure the degree of human capital within the region as the proportion of nonproduction workers among all workers in the region. However, previous empirical works mainly use education attainment (i.e., the fraction of workers with a college degree or above) to capture regional human capital level. I use such indirect indicator such as the share of nonproduction workers to capture labour skill-intensity because of lack of comprehensive data on worker’s education in the survey. This measure, in turn, will not be problematic as Berman et al. (1994) argue that

Both conceptually and empirically, the production/nonproduction worker distinction closely mirrors the distinction between blue- and white-collar occupations… The blue-collar/white-collar classification, in turn, closely reflects an educational classification of high school/college (pp. 371–372).

This statement implies that the use of nonproduction workers as the degree of labour skills is acceptable because it reflects the separation between blue and white-collar workers. In turn, blue and white-collar workers categories mirror the educational distinction. In other words, nonproduction workers tend to be more educated than production workers, and as a consequence, the former type is more likely to absorb new ideas and knowledge.

Our data constitute a balanced sample of 194 plants combined for Indonesia, Philippines, and Vietnam with complete information about workforce characteristics and plant performance. The dataset only covers for two years, 2009 and 2015. This rich dataset enables us to deal with unobserved characteristics of plants. A key econometric issue is the possibility of unobserved factors may affect plant productivity and the proportion of nonproduction workers. More productive plants, for instance, may choose to locate and produce in the region with a high percentage of skilled workers. Taking advantage of longitudinal data, I can tackle some of the relevant endogeneity problems. I can control permanent unnoticed elements of the plant by adding plant and time-fixed effect into the regression equations. Plant fixed effect will absorb any permanent industry and location heterogeneity as both features do not change over time. Time fixed effect, contrarily, will absorb any time-varying factors that influence productivity and the share of nonproduction workers within the region.

Based on Ordinary Least Square (OLS) estimates, I found that if the portion of nonproduction workers in any particular area increases by one percentage point, there will be around 7.1 until 8.9 percent rise in plant productivity. These figures are remarkably consistent, even after controlling for plant and year-fixed effect. However, there is still a potential problem that the overlooked time-varying region characteristics will affect the findings. Thus, I provide additional sections of evidence to examine the validity of the results before reaching the concluding remarks.

Although I control for plant and time-fixed effect to absorb unobserved permanent shocks, this approach only overcomes the omitted variable bias but not the reverse causality bias. For example, a region with better amenities attract skilled workers to move to that area, and at the same time also raise the productivity of existing plants in the area. A common method to address such problem is implementing the instrumental variable (IV) approach. In this paper, I use the lagged number of higher education institutions in the region. The existence of universities and colleges may determine the changes in the aggregate skill level of workers. The findings show that the regional human capital spillovers are still relevant using this method although the instrument is not as strong as expected.

Another plausible source of bias exists if the quality of workers is unobserved. Thus, I construct an alternative measure of human capital spillovers using average yearly wage in the region, excluding the plant itself. The explanation is that plants encourage to pay skilled workers with higher wage bill because they tend to be more productive than unskilled workers. Furthermore, this different indicator can be used to assess whether the results are still consistent with the main indicator of human capital. The findings suggest that spillovers effect are positively and significantly associated with higher plant productivity.

I provide several other specification tests to check the reliability of our core results. I begin with the use of Total Factor Productivity (TFP) estimation approach, assuming that factors of production are correlated with unobservable shocks to productivity. Using this technique, the results show that human capital spillovers are still in line with the core findings, despite with smaller impact on plant productivity.

I also test whether the magnitude of spillovers maybe not equal between high-tech and low-tech industries. Spillovers are expected to be stronger for plants in high-tech industries because the presence of highly skilled workers is more valuable in these sectors. However, my findings suggest the other way around for ASEAN countries. Plants in low-tech industries receive the largest benefit from the availability of skilled workers provided by both other low-tech and high-tech industries in the region.

In the last section of this paper, I also carry out some sensitivity analysis including the estimation with a translog production function, as well as implementing regressions with different sample based on specific categories such as plant size and multi-plant status. In all these exercises, I find that the human capital spillovers substantially affect plant productivity. Therefore, my results are robust under different identification tests.

Overall, I find a supportive evidence of the presence of human capital spillovers in ASEAN manufacturing sector. In other words, the proportion of nonproduction workers in any given region has a meaningful effect on plant productivity. Moreover, the contribution of human capital spillovers seems significant. It implies that the stock of human capital grows considerably over time in ASEAN countries. The most robust estimates in this paper specify that human capital spillovers account for an average of 1.5 percent increase in plant productivity per year from 2009 to 2015.

Empirical literature indicates that human capital spillovers are an important aspect of determining plant productivity, especially in the manufacturing sector (Moretti, 2004b; Liu, 2013; Chang et al., 2016). In spite of its significance for policy implications, there is a little systematic attempt to examine the size of spillovers in developing countries. All studies I mention before focused on developed countries. Hence, this paper will fill the gap by estimating the magnitude of human capital spillovers in three ASEAN countries and all of them are developing economies.

Although the empirical approach adopted in this paper has quite similar with the work done by Moretti (2004b), my study has a significant contribution to the literature about human capital externalities which differs in three ways. First, to the best of my knowledge, this study is the first attempt to employ panel-plant level data for multiple countries. Previous studies have emphasised the magnitude of local skilled workers on plant productivity for a given country. Another difference comes from the use of the dataset. Instead of constructing matched employer-employee data as Moretti did, I internally calculate regional human capital level using the data from Enterprise Surveys. Such computation is relevant and reliable since the survey is sufficient to exploit the overall variation in plant condition within regions. Lastly, the new additional piece is that in this paper, I employ a different instrument to eliminate endogeneity issue, compared to the previous studies. The proportion of skilled workers at the regional level is instrumented with the lagged number of higher education institutions.

The remainder of the paper is organised as follows. Section 2 describes the related theoretical and empirical literature on human capital spillovers. The augmented production function used is explained in section 3. This section also presents the estimation issues that I need to address. Section 4 discusses data source and variables I employ in this paper. The estimated effect of human capital spillovers on plant productivity is described in section 5. This section also provides sensitivity exercises of the main specification. Finally, section 6 is the concluding remark.

The theoretical literature about human capital spillovers and its relationship to firm productivity was firstly discussed by Marshall (1920), who argued that the localisation of skilled workers in a narrow area be the key element of productivity improvement. This view is then supported by Kuznets (1962) who elaborates that

Creative effort flourishes in a dense intellectual atmosphere… The possibility of more intensive intellectual contact… afforded by greater numbers may be an important factor in stepping up the rate of additions to new knowledge (pp 328–329).

Both views focused on the average quality of the local workforce. However, it is not the sole determinant of economic productivity but also depends on the role of skill diversity and cultures. Jacobs (1969) highlights that the interactions among people with different knowledge and perspectives within a city tend to generate new ideas and innovations. These arguments are then formalised with learning model by Lucas (1988) who proposes that the benefit associated with urban areas comes from the firm acquiring ideas from their citizens where they operate. Other types of learning model have been richly presented in the literature (Jovanovic and Rob, 1989; Glaeser, 1999). On the other hand, Acemoglu (1996) builds a model where the aggregate human capital can affect plant productivity in the absence of learning mechanism. In his model, the human capital of workers and physical investment of firms are complementary. For instance, if workers are expected to increase their educational attainment in the future, firms will decide to invest in more advanced machinery and equipment now to match with the expected worker’s skill requirement and thus raise productivity. More recently, Duranton and Puga (2004) analyse that spatial concentration and structure of local workforce can enhance firm productivity through three mechanisms — sharing, matching, and learning. The first type includes sharing narrower specialisation and risk by establishing labour pooling. The matching process focuses on improving the quality of the match between workers and employers. Finally, as discussed earlier, learning mechanism is based on knowledge accumulation and diffusion.¹

Several papers have examined the empirical extent of local human capital spillover either in the region or firm level. The first wave of studies focused on its impact on individual wage based on Mincerian model. Rauch (1993) is the first and most influential work to exploit the human capital variation across cities to identify its effect on wage. Using the 1980 US Census of Population, he reported that one year increase in average years of schooling in any given city would raise the average hourly wage by 3–5 percent in 1980. Nevertheless, the main limitation of Rauch’s paper is failing to deal with the endogeneity problem associated with the city-level human capital. To overcome such issue, Acemoglu and Angrist (2000) use the state variation of compulsory education turn out and child labour laws as instruments for average schooling across cities. They also address the endogeneity concern in individual schooling by instrumenting it with a quarter of birth. They only find little evidence of such spillovers in practice. In a similar manner, Moretti (2004a) implemented two instrumental variables at the city level, which are age structure and land grant institutions. He also uses the share of college graduates as the measure of city human capital level. He highlighted that one percentage point increase in the proportion of college graduates in particular city increases average wage by 0.6–1.2 percent in 1980. The main difference in findings between Acemoglu and Angrist (2000) and Moretti (2004) lies on the instrument they use. The latter uses the share of college graduates which reflects the upper bound estimate of human capital spillovers whereas the former employs compulsory schooling laws that mostly affect lower education.

However, another strand of literature in human capital spillovers has shifted away from wage model to firm productivity model by estimating production function. De la Fuente and Domenech (2001) found that there is a positive relationship between growth rate of average years of schooling and productivity using cross-country dataset. However, employing cross-country data is not particularly useful as the human capital level is possibly correlated with unobserved shocks which can affect productivity. Hence, the subsequent studies tried to explore exogenous variation at the city or regional level to tackle this issue. Adam and Jaffe (1996) investigate the effect of transfer knowledge composition within and across

firms on plant productivity for chemical industry in the US over the period 1974–1988. They use the number of R&D performed in labs instead of human capital stock as the explanatory variable. Research activities can be conducted informally at plants, formally at parent firm, or by other plants in the same city or industry. They document that R&D spillovers are essential both within and across plants. More interestingly, the magnitude of spillovers diminished as geographical distance between labs and plants increases. They suggest that such condition reflects the rising communication cost. Similarly, growth and the spatial distribution of scientists is the main aspect to determine the birth and location of biotechnology industry in the US (Zucker et al., 1998). A number of outstanding scientists in cities will influence the location decision of firms in such sector.

Both studies described above emphasise on high-tech industries. Moretti (2004b) then systematically try to estimate the impact of regional human capital spillovers on plant productivity for all US manufacturing industries between 1982 and 1992. Moretti indicated that the impact of inter-industry human capital spillovers is quite low, only responsible for an average of 0.1% increase in firm output per year during the period. Estimated spillovers tend to come from high-tech industries, and it is virtually zero for low-tech industries. More importantly, the scale of spillovers between plants within city depends on their interaction level. Spillovers are found to be significant for plants that regularly interact each other.

Recent studies in human capital spillovers using productivity model particularly follow Moretti’s (2004b) study. The only difference essentially comes from the country sample and instruments they apply when controlling endogeneity problem. More specifically in Asia, Liu (2014) and Chang et al. (2016) report that high level of city human capital is positively associated with larger plant productivity for manufacturing sector in China and Taiwan respectively while Park (2016) does not find the similar result for Korean manufacturing plants. In the context of instrument they apply for IV approach, the study in China uses the number of the book collection in public libraries whereas the latter work proposes the exogenous change in Korea’s educational reform. In sum, the findings of existing empirical evidence in local human capital spillovers, both based on wage and productivity approach, are found to be mixed.

The hypothesis of this paper is whether plants located in the region with a higher quality of human capital will generate more output with any given input. Thus, production function will be estimated to evaluate the effect of human capital spillovers on plant productivity. However, such strategy is not straightforward to identify. Unobservable factors can affect productivity of the plants and are correlated with the overall human capital across regions. More productive plants may decide to produce in areas with ample amount of highly skilled workers for any reasons which are not associated with human capital spillovers.

I begin by explaining how production function can fit to capture such externalities. Then, in the remainder of this section, I describe the panel data model choices, endogeneity issues in detail and the appropriate econometric approach to deal with such concerns.

3.1 Model Specifications

First, consider the following standard Cobb-Douglas production function:

$${Y}_{ijrt}={A}_{ijrt}{LNP}_{ijrt}^{\alpha }{LP}_{ijrt}^{\beta }{K}_{ijrt}^{\delta }$$

Where Y_ijrt indicates the value added of plant i, belonging to industry j, in region r, and at time t; LNP_ijrt and LP_ijrt denote the number of nonproduction and production workers in the plant respectively; and K_ijrt is capital stock. If I assume that plant total factor productivity, A_ijrt, reflects the regional human capital level, then A_ijrt can be expressed as follows:

$${\text{ln}A}_{ijrt}= \gamma {HS}_{rt}+\theta {X}_{ijrt}^{{\prime }}+{\epsilon }_{i}{+\epsilon }_{j}{+\epsilon }_{r}{+\epsilon }_{t}{+\epsilon }_{rt}{+\epsilon }_{ijrt}$$

HS_rt is the share of nonproduction workers in particular region r outside the firm and coefficient of interest γ represents the external effect of human capital on productivity. In Eq. (2) above, I emphasise my analysis on the human capital spillovers within a region. I do not distinguish the spillovers within and between industries. Hence, estimates of γ can be interpreted as the upper bound on the total spillover magnitude. Statistically, the main challenge from Eq. 2 is how to establish the coefficient γ consistent and not correlated with time-varying shocks, which are embodied in ε_rt and ε_ijrt. If γ = 0, the model goes back to standard production function without spillovers. ε_i, ε_j, and ε_r are time-invariant unobserved productivity shocks at the plant, industry, and region respectively. Similarly, specific time-varying factors and region-specific time-varying shocks are represented by ε_t and ε_rt. However, I do not include any industry-specific time-varying shocks that are shared by all plants in the same industries because it will cause perfect multicollinearity with permanent heterogeneity at the industry level (ε_j). I also add control for individual plant characteristics, X_ijrt, such years of operation and single/multi-plant category.

Equation (1) can be rewritten into natural log and plugged back into$\text{ln}{A}_{ijrt}$, then the production function becomes:

$${\text{ }\text{ln}y}_{ijcrt}=\delta +\alpha \text{ln}{LNP}_{ijrt}+\beta \text{ln}{LP}_{ijrt}+\delta \text{ln}{K}_{{ijrt}}+\gamma {HS}_{rt}+\theta {X}_{ijrt}^{{\prime }}+{f}_{i}{+f}_{t}+{\epsilon }_{rt}+{\epsilon }_{ijrt}$$

It is important to notice that c indexes the country where the region r is located. All the regressions are estimated using clustered standard errors on plants.

The main advantage of using panel dataset is that I can control the permanent and time-varying factors that affect productivity and overall share of nonproduction workers. Plant-fixed effect f_i absorbs any unobserved time-invariant variations at the plant, regional, and industry level (ε_i, ε_j, and ε_r). In a similar way, year-fixed effect f_t deal with the time-varying heterogeneity attributable to plant productivity. Nevertheless, I leave the time-varying, region-specific factors ε_rt as it is. If I add region-time fixed effect, f_rt, as a covariate, it will cause perfect multicollinearity as the human capital spillovers variable HS_rt is at the region-year level. As a result, Eq. (3) cannot be estimated properly. When I can certainly control plant specific characteristic, I am not able to do the same thing for regional level due to data limitations across these countries. Hence, to overcome this issue, I provide the alternative measure that varies over time within a region. One can expect that the regional human capital level will have a positive impact on productivity and hence the coefficient γ is predicted to be positive. However, unobserved factors can affect either region-human capital or plant productivity. This will be the sources of identification threat that I need to tackle.

Key hurdle from Eq. 3 is the endogeneity issue. It arises if the time-varying factors at regional level ε_rt are positively correlated with the share of college graduates within the region H_rt. When such problem is not solved, the magnitude of coefficient γ will be biased upwards.

The common solution to this concern is employing the two-stage least square estimation technique. When using this method, selecting the proper instrument will be very critical. It has to be correlated with the share of nonproduction workers (relevant condition), but no relevance to productivity shocks of individual plant (exogeneity condition). In addition to that, the instrument has to vary at the regional-year level to extract the portion of the region-specific change in the share of nonproduction workers. Then, I propose the logarithm of total higher education institutions for the instrument. Using this instrument, which arguably satisfies these two conditions, I use two-stage least square to investigate the causal effect and the extent of human capital spillovers. More formally, the first stage equation is described as follows:

$${HS}_{rt}=\tau \text{ln}{HEI}_{r,t-1}+\sigma \text{ln}{LNP}_{ijrt}+\rho \text{ln}{LP}_{ijrt}+\pi \text{ln}{K}_{{ijrt}}+\vartheta {X}_{ijrt}^{{\prime }}+{f}_{i}{+f}_{t}+{\epsilon }_{rt}$$

Here, HS_rt denotes the share of nonproduction workers, as defined earlier in this section. HEI_r,t−1 is the number of universities and colleges in the previous year. In other words, I instrument the employment share of nonproduction workers based on the assertion that higher education institutions will exogenously have a stronger association with the number of highly-skilled workers. The second stage regression of IV approach uses the predicted value of the share of nonproduction workers in the first stage, $\widehat{{HS}_{rt}}$, and estimates with the following equation.

$${\text{ }\text{ln}y}_{ijcrt}=\delta +\alpha \text{ln}{LNP}_{ijrt}+\beta \text{ln}{LP}_{ijrt}+\delta \text{ln}{K}_{{ijrt}}+\gamma \widehat{{HS}_{rt}}+\theta {X}_{ijrt}^{{\prime }}+{f}_{i}{+f}_{t}+{\epsilon }_{rt}+{\epsilon }_{ijrt}$$

Again, the coefficient of interest is γ. If my instrument is valid, it will tackle the positive biased attributed with simple OLS estimates and γ will further reveal the regional human capital spillovers.

To check the validity requirements of the instrument, I need to refer back to those two conditions. The relevant condition can be tested by the looking at the significance of the estimated result of the first stage in the two-stage least square analysis. As to exogeneity condition, an instrument will be invalid if unobservable productivity shocks have positive impacts toward plant productivity and tend to increase the presence of universities in the region. Nevertheless, because my instrument using the information for the previous year, it is unlikely to be caused by unobservable productivity shocks in the current year. Hence, it is assumed that the exogeneity condition is satisfied in the analysis.

The next subsection will describe the detail of econometric concern to obtain unbiased estimates. However, before I begin the discussion on estimation issues, I will briefly describe the panel data models which are regularly used in the empirical papers.

3.2 Linear Static Panel Data Models

As a starting point, I utilise Linear Static Panel Data estimation techniques (i.e., Pooled OLS, Random Effects and Fixed Effects) to investigate the presence of human capital spillovers as in baseline Eq. (3). To begin the estimation process, it will be estimated using Pooled Ordinary Least Squares and Random Effect, with cluster-robust standard errors being used to account for the presence of heteroscedasticity. Clustering is applied at the firm level which allows for serial correlation of the error term (both the unobserved and the idiosyncratic errors) over time, which assumes that the individual plant unobserved heterogeneity carries over from one year to another.

A typical example of a linear static panel data model will be denoted as follows:

$${y}_{it}= {\beta }_{0}+ {\beta }_{1}{x}_{it}+ {u}_{i}+ {\epsilon }_{it}$$

where i = 1,…,N and t = 1,….,T. The main characteristic in Eq. (6) above is that coefficient β₁ and β₂ do not vary across individual and over time, which do not allow for individual heterogeneity. This criterion leads to the concept of Pooled Least Squares model. As a result, it is assumed that there is no contemporaneous correlation of error terms with the independent variables x_it.² Thus, the estimators are consistent and can be estimated using standard Ordinary Least Square technique. However, in case of panel dataset, this restrictive assumption is likely to be violated since it is quite unrealistic. Heterogeneity across individuals tend to happen as the unobserved characteristics that are not captured in the regressors will be included in error terms and leads to incorrect standard errors.

An alternate model is then proposed and known as Random Effect model. The main rationale behind the Random Effects model is that the variations across individuals are assumed to be random and not correlated with regressors. If we change Eq. (6) a little bit where the individual specific effects ${u}_{i}$ are assumed to be uncorrelated with the regressor x_it;

$${y}_{it}= {\beta }_{0}+ {\beta }_{1}{x}_{it}+ {u}_{i}+ {\epsilon }_{it};i=1,\dots ,N;t=1,\dots ,T.$$

$${\epsilon }_{it}∼N\left(0,{\sigma }_{\epsilon }^{2}\right) and {u}_{i}∼N(0, {\sigma }_{u}^{2})$$

Once both the Pooled Ordinary Least Squares and Random Effect techniques for the baseline model are estimated, we employ the Breusch and Pagan (1980) Lagrangian Multiplier (LM) test for random unobserved individual effects. If the p-value of the Breusch and Pagan LM is less than 5 percent significant level, the null hypothesis of homoscedastic errors are rejected, then the Random effects model is selected over the Pooled Ordinary Least Squares estimator. By convention in related empirical work, before one can select the random effect estimator as the appropriate choice of model, it must be evaluated against the fixed effects model.

I then proceed to estimate the model using the Fixed Effects Estimator which transforms the data through a demeaning procedure which removes the unobserved heterogeneity from the model and is then estimated using the standard Ordinary Least Square estimation technique. In what follows, I provide a brief theoretical presentation of the fixed effects model as it is most frequently used between two models, due to the restrictive assumption in Random Effect model that the unobservable factors are not random and uncorrelated with regressors, which does not reflect the data structure of most dataset.

Suppose that we have the following generic static panel data linear regression model stated in Eq. (8) where a_i and u_it are the unobserved time-invariant plant-fixed effects and the idiosyncratic error term respectively, which includes all other time-varying unobservable.

$${y}_{it}={\alpha }_{i}+{\beta }^{{\prime }}{x}_{it}+{u}_{it}$$

In the context of Ordinary Least Squares, if the unobserved fixed time-invariant effect f_i is correlated with the regressor x_it, usually either omitted variables or simultaneity bias, the Ordinary Least Squares Estimator will be biased as the strict exogeneity assumption is then violated. One way of overcoming this issue is the inclusion of dummy variables for each cross-section, where for every observation of the cross-sectional variable one dummy is assigned 1 and for all other cross-sectional variables 0 is assigned. This method is known as the Least Squares Dummy Variable model or LSDV model. It should be noted that the LSDV is theoretically the same as the fixed effects estimator, but a major weakness is the issue of degree of freedom losses. In light of this, LSDV model is not preferred.

Then, the more preferred option is what known as time demeaning (within) transformation process. Now, each variable in the Eq. (3) is averaged to obtain the following expression in Eq. (9) below:

$${\stackrel{-}{ y}}_{it}={\stackrel{-}{\alpha }}_{i}+{\beta ́}_{1}{\stackrel{-}{x}}_{i}+{\stackrel{-}{u}}_{i}$$

where ${\stackrel{-}{y}}_{it}$, ${\stackrel{-}{x}}_{i}$, ${\stackrel{-}{\alpha }}_{i}$ and $\stackrel{-}{u}$ are time averages of the dependent variable, independent variable, the time-invariant unobserved factor and idiosyncratic error term. The core thing of this process is that it subtracts the Eq. 9 from Eq. 8 to obtain fixed effect estimator.

$$\left({y}_{it}- {\stackrel{-}{y}}_{it}\right)=\left({\alpha }_{i}-{\stackrel{-}{\alpha }}_{i}\right)+{\beta }^{{\prime }}\left({x}_{it}-{\stackrel{-}{x}}_{i}\right)+\left({u}_{it}- {\stackrel{-}{u}}_{i}\right)$$

In so doing, we obtain a transformed equation below with time demeaned variables:

$${\ddot{y}}_{it}={\beta {\prime }}_{1}{\ddot{x}}_{it}+{\ddot{u}}_{it}$$

Where ${\stackrel{-}{\alpha }}_{i}= \frac{1}{N}\sum _{t=1}^{N}{\alpha }_{it}$ and ${\stackrel{-}{X}}_{i}= \frac{1}{N}\sum _{t=1}^{N}{X}_{it}$. From Eq. 11 above, it should be evident that the time invariant fixed effect which is constant over time has been excluded (${{\stackrel{-}{\alpha }}_{i}=\alpha }_{i}$). Given the removal of the unobserved time invariant fixed effects Eq. 11 can then be estimated using Ordinary Least Squares.³

To select between the Fixed and Random effect models, the Hausman Test is employed. It is used to verify the orthogonality condition of the random effect and the regressors (Greene, 2003). However, the standard Hausman Test does not function on non-i.i.d (non-independent and identically distributed errors) as we assumed from the onset in this paper. Hence a robust version of the Hausman Test will be used in this article.

3.3 Estimation Issues

Econometric problems described in this section are related to the potential unobservable heterogeneity that may happen across plants and over time under Fixed Effect model. As stated before, the main assumption of suitable identification after controlling plant characteristic is that there is no correlation between the fraction of nonproduction workers in region-observable variation H_rt and the time-varying regional unobserved factor ε_rt. If this assumption is violated, the coefficient of interest γ will be inconsistent. There are two sources of threats to identification regarding such relationship. The first problem comes from the existence of reverse causality. It is a condition where plants with better management practice, for instance, choose to locate their production in regions with a high proportion of nonproduction workers because they can attract them with a higher wage. Therefore, the local density of such employees increases and provide an overestimated coefficient of regional human capital. This problem is unlikely a major issue here since the inclusion of plant fixed effect to the model, which also captures both permanent industry and region-specific heterogeneity. Secondly, the variation in local amenities that are not included in H_rt but correlated with regional human capital level and plant productivity at the same time will also pose an omitted variable problem. The exclusion of such feature will be captured in region error term ε_rt. An obvious example of this is local infrastructure such as highways and airports. Both transport means will certainly raise productivity for all plants in any region and at the same time attract workers to move to that area. If this happens, the estimates of γ will lead a positive bias. However, there are no plants which move to another region in the dataset between two subsequent years. I can include plant fixed effect f_i to eliminate the heterogeneity across regions. Additionally, I also employ an instrumental variable approach to adjust with this endogeneity. More explicitly, I use the one-year lagged number of higher education institutions as the instrument.

My estimates may also be invalid if any positive relationship between plant productivity and human capital spillovers are found to be caused by time-varying factors other than the proportion of nonproduction workers in the region. Obviously, if I ignore this possibility, the estimates tend to exhibit an upward bias. Thus, the inclusion of year-fixed effect is essential to be able to avoid the failure to recognise these unobservable factors.

Another potential econometric problem is the correlation between firm-level human capital and random error term. The latter may contain plant-specific unobserved factors that affect productivity. Such potential issue will lead to inconsistent estimates. To deal with this, I add time-varying plant-specific control variables such as the longevity of plant operation and multi-plant status.

To provide consistent estimates of the regional human capital variable, I also employ an alternative measure to examine the regional human capital endowment. The measure involves log of average monthly real wage within the regions. This indicator can capture the local quality of workers besides schooling. Plants located in areas with abundant nonproduction workers will pay higher wage because these employees are anticipated to be more productive and hence increase plant output. Therefore, I safely assume that nonproduction workers, who are more likely to be skilled, are paid more than production workers.

A final concern is related to theory when input factors such as labour and capital cannot be treated as exogenous. Endogenous inputs are the only issue to the extent that it will result in the biased estimates of coefficient γ. Therefore, I assume that this problem is solved after controlling for plant fixed effect. To test the reliability of my results to this assumption, I directly estimate Total Factor Productivity (TFP) as the function of H_rt using Levinsohn and Petrin (2003) procedure. Results are substantially consistent with the main model, suggesting that labour and capital are fully exogenous in this case.

In sum, it is evident that many plausible sources of endogeneity have been accounted for while I cannot wholly dismiss the probability that some of the estimated effects reflect the regional productivity shocks. I also carry out a set of specification tests to prove that the positive impact on plant productivity in this study is solely due to human capital spillovers.

The data used in this study is taken from World Bank Enterprise Surveys for 3 ASEAN countries, namely Indonesia, Philippines, and Vietnam for the period 2009 and 2015. I exclude other ASEAN countries due to differences in year surveyed since World Bank just began to survey them at the beginning of the 2000s. I only focus on plants in the manufacturing sector to be consistent with previous studies. It contains approximately 700 manufacturing plants for each year within a country, but some plants leave and join the survey in 2015 and thus creates unbalanced panel format. However, in this study, I only include plants that appear in both years to construct a balanced panel dataset. Finally, there are 198 plants from all countries combined in my dataset. Although the number of plants seems small, Enterprise Surveys can sufficiently represent the whole manufacturing sector within the country because it applies stratified random sampling.⁴ The dataset also covers detailed information about individual plants such as net value of capital, locations, industrial classification, cost of production, sales, and among others.

The dataset also shows the number of employees, both for production and nonproduction employees, and average schooling years. However, the information on the latter is restricted to production workers only. That is the reason I do not use schooling as the variable of interest. Hence, I only use the proportion of white-collar (nonproduction) workers as the proxy of college graduates share. Enterprise Survey defines nonproduction worker as “…workers employed in sales (including driver-salespersons), sales delivery (highway truck drivers and their helpers), janitorial and guard services, advertising, credit, collection, installation and servicing of own products, clerical and routine office functions, executive, purchasing, financing, legal, personnel (including cafeteria, medical, etc.)” [World Bank Group, 2017]. In addition to that, I also construct an alternative measure of regional human capital to check the validity of core results by employing log of average yearly wage outside the plant within a region, calculated as yearly labour cost divided by total workers.

I use value-added as the indicator of plant output. Value added is defined as the sales less cost of production, which includes the cost of electricity and materials. Labour will be measured and separated using the number of production and nonproduction employees. I use such indicator instead of a number of work hours because of data constraint. Net book value of machinery, vehicles, and equipment is assigned as a proxy for plant capital stock. We also include the total cost of materials as input factors, but the value of each variable is reported in local currency. Thus, I convert these local currencies into US Dollar using the official exchange rate from World Development Indicator (WDI-WB) database. ⁵

I also obtain the data for the number of higher education institutions in one-year lag as the instrument from Statistical Yearbook of each country. These documents contain the information on the regional distribution of total universities and colleges. The regional classification between Statistical Yearbook and Enterprise Survey is the same. Therefore, I calculate the number of higher education institutions from the latter, and then I assign those figures to each sample plants in the dataset without any further adjustment.

In the analysis, I will emphasise on manufacturing sectors to be consistent with the value-added terms and thus the productivity. The industrial classification is based on four-digit level manufacturing industries. I regroup them into 17 sectors at two digit level to capture more plants within industries when examining the spillovers effect based on the sectoral analysis. Nevertheless, the number of plants for each industry are not perfectly distributed because some industries have a larger contribution to the national economy than the others. A consequence of it is that large plants are more likely to be oversampled. The Same methodology also applies in the context of plant location. Not all regions are included in the survey. Nine provinces are selected for Indonesia, five districts for the Philippines, and only four regions picked in Vietnam. There will be a probability that more skilled workers prefer to work in one of those regions.⁶

It is also important to bear in mind for the caveat of my results presented here due to data availability. The most matched panel dataset for ASEAN countries is only for the period 2009 and 2015. Plants included in the survey are the only small part of plant population in each country and therefore not essentially representative to describe the whole effect in ASEAN. One critical consequence, though admitted by the World Bank, is the oversampling of large plants due to their bigger output in the economy. To control this, I reweight the model by estimating the separate regressions based on three broad categories of plant size, which are small, medium, and size in the robustness analysis (see Section 4.2.4 for more details).

Table 1 provides summary statistics for main variables used in this paper. Columns in Table 1 contain mean and standard deviation for each year and country.⁷ One can verify that both sales and value-added increased significantly between 2009 and 2015. The average sales of each plant increased greatly from around 7 million U.S. Dollars in 2009 to 78 million U.S Dollars in 2015, which this figure increased nearly 11 times in six years. During this period, the average value-added also rose rapidly by more than 600%. On the other hand, capital stock is the variable that slightly fell during the term of interest. Each plant’s net value of capital, on average, declined approximately 6%, from a base of 1.5 million U.S Dollars in 2009. As the production inputs, the cost spent to buy fuel (energy) and materials also both went up at a similar pace. Cost of materials escalated roughly more than 346%, and there was 485% increase in energy expenses respectively. As a consequence, energy cost is as important as the cost of material.

Table 1

Means of Summary Statistics
	All Countries		Indonesia		Vietnam		Philippines
	2009	2015	2009	2015	2009	2015	2009	2015
Sales (*1000)	7658.1	77980.1	5988.6	74532.3	2749.4	5073.6	13914.7	131938.6
Value Added (*1000)	3876.0	40005.9	2690.5	55590.6	590.9	1458.3	8310.2	34832.9
Capital (*1000)	1542.8	1447.7	1780.5	1014.8	478.9	678.4	1770.0	2788.1
Materials (*1000)	2815.1	12574.8	2411.5	16080.0	1946.1	2615.5	4163.3	12214.0
Energy (*1000)	159.1	932.0	145.5	1480.5	30.8	68.7	268.8	282.0
Number of Production Workers	145.8	398.7	109.5	180.7	92.4	80.9	109.5	87.4
Number of Nonproduction Workers	35.9	124.0	41.4	54.0	13.0	12.8	41.4	32.5
Nonproduction Workers Share in Region	0.200	0.213	0.193	0.207	0.128	0.130	0.258	0.280
Average Yearly Wage in Region	2234.3	7099.4	1478.1	2852.8	1699.6	3560.1	4051.3	18271.3
Age of Plants	17.3	22.5	19.7	24.5	10.1	15.2	17.5	23.2
Proportion of Plants Belong to Multi-unit Firm	0.15	0.13	0.22	0.09	0.09	0.06	0.07	0.26
Number of Plants	194		105		35		54
Note: All monetary values are reported in US Dollar. Nonproduction worker share and average yearly wage are calculated outside the plant.

Moreover, the average proportion of white-collar workers in regions within three ASEAN countries does not differ significantly between two periods. In particular, there was only one percentage point increase in the fraction of nonproduction workers within a region outside the plant in the space of six years. This can be inferred by a slight increase in total nonproduction workers, from on average of 36 to 41 people employed, while there is a small decline by around nine nonproduction workers during the period of interest. On the other hand, the average age of individual plants slightly increased while the fraction of multi-unit plants achieved a minor decline during the period. Overall, summary statistics show that plant level characteristics increased significantly whereas the regional human capital level remained stable between these two years.

Now, the general pattern can be drawn for each country as well. Productivity, input factors except for the capital stock, and aggregate human capital increased considerably during the period when plants in the Philippines have the most significant rise in every aspect. For example, we can take a look at the value-added as our productivity measure. It is interesting that per plant value-added in Indonesia is approximately 2.5 times larger than in Vietnam, whereas the Indonesia’s and Vietnam’s value-added is just 10 and 5 percent of each plant’s value-added in the Philippines respectively over a six-year period. This fact implies that there are huge variations in productivity across plants within the country.

This condition also applies to worker endowments as well. Although the number of skilled workers is the highest for plants in Indonesia, we cannot say that manufacturing plants in Indonesia have the best worker’s skill intensity if not compared to total production workers. Moreover, plants in the Philippines are having the highest employment share of nonproduction workers in plants where every three production workers consist of one nonproduction workers (3:1). On the other hand, in Indonesia, one skilled worker is equivalent to four unskilled workers (4:1) while the comparison between production and nonproduction workers in Vietnamese plants is around seven to one (7:1). As a result, the share of highly-skilled workers within region among countries is also the largest for Philippines manufacturing plants.

This disaggregated means also reveal the similar feature for years of operation and multi-unit status as plant-specific controls. Filipino plants have been operating longer, nearly 20 years, than the similar plants in Indonesia and Vietnam. Similarly, 17 percent of total manufacturing plants in the Philippines has more than one establishment, compared to 15 and 7 percent for Indonesian and Vietnamese plants respectively. From the summary statistics, I can conclude that the human capital spillovers may significantly have a positive correlation with productivity, as can be revealed by the fact figures for plants in the Philippines. Such prediction is consistent with the findings from section 5.

At first, I perform some tests to pick the best panel data models for the analysis. I then continue to estimate the production function in Eq. (3) and use the logarithm of value-added as a dependent variable. I also employ instrumental variable estimation as estimated using Eqs. (4) and (5) to eliminate endogeneity issue. Then, a series of robustness checks are carried out to strengthen my findings. It is critical to note that I do not separate estimation results for each country since I take them as an integrated analysis.

5.1 Basic Results

5.1.1 Ordinary Least Square Estimation

I start the analysis by choosing the best static panel data models to be used in this paper. Thus, I estimate the baseline Eq. (3) under different models, and the results are reported in Table 2 below. To choose between Pooled Least Squares and Random Effect model, we can look at the statistical significance of Breusch-Pagan Test. The p-value of this test is 0.2382, more than 5 percent level of significance. It means that the null hypothesis cannot be rejected and thus the Pooled Least Squares model is more appropriate. However, we may assume that there is unobservable heterogeneity across plants in the dataset that has an impact on productivity. To validate this argument, I need to examine whether I can employ Fixed Effect model using Hausman Test. This test is used to choose between Random Effect and Fixed Effect model. According to p-value of Hausman Test, I can reject the null hypothesis. This verdict suggests that the variation across plants is correlated with the independent variables and then Fixed Effect estimation is selected over Random Effect. Therefore, throughout this paper, I resume the analysis using Fixed Effect model.

Table 2

Estimates of Human Capital Spillovers - Model Choices
	POLS	RE	FE
Nonproduction worker share in region	4.061** (1.700)	4.380*** (1.684)	8.822*** (3.027)
ln (production workers)	0.566*** (0.116)	0.575*** (0.115)	0.334 (0.214)
ln (nonproduction workers)	0.731*** (0.149)	0.712*** (0.123)	0.289 (0.216)
ln (capital)	0.191** (0.075)	0.183*** (0.047)	0.053 (0.066)
Plant specific controls	+	+	+
R²	0.53	0.53	0.43
Breusch-Pagan Test		0.2382
Hausman Test		0.0012
Note: All specifications include plant characteristics such as years of operation and multi-unit status but without dummy year variable. The sample size is 388 observations. * p < 0.01, p < 0.05, * p < 0.1.

Table 3 reports the regression results under different specifications of Eq. (3). I begin by documenting the relationship between share of nonproduction workers in the region and plant productivity. I also include a dummy equal to one if plant belongs to multi-establishment unit and logarithm of plant longevity to control for plant-specific characteristics that affect productivity.

I start with a model in column (1), where the only factor endowments and plant-fixed effect included. It purges estimates of the permanent plant, industry, and region unobserved heterogeneity. Identification of the human capital spillovers comes from changes in plant productivity and nonproduction worker share between 2009 and 2015. The coefficient on nonproduction worker share in column (1) is 8.857.

In column (2), I modify the model specification by adding additional plant controls such as plant age and multi-plant status and retaining the plant-fixed effect. Based on such estimation, the estimated effect of spillovers is 8.822, quite similar to the figure in column (1). This implies that plant characteristics do not completely affect the impact of aggregate human capital at the regional level.

Similarly, I change the specification a little bit by removing additional plant controls but providing the year-fixed effect as well. The spillovers impact becomes smaller to 7.320 as documented in column (3). This figure indicates that time-invariant unobserved characteristics and plant-specific controls may confound the external human capital faced by the plants and lead to underestimating the effect of human capital spillovers. Bloom et al. (2013) pointed out that spillovers will result in two contradictive effects on firm performance: a positive effect from knowledge spillovers and a negative impact from product market rivalries. If we do not control for plant-fixed effect specifically, this leads to downward biased estimates due to plant competition within regions. Hence, after controlling for this, the effect becomes larger.

Table 3

Fixed Effect Estimates of Human Capital Spillovers
	(1)	(2)	(3)	(4)
Nonproduction worker share in region	8.857*** (3.364)	8.822*** (3.393)	7.320** (3.331)	7.152** (3.450)
ln (production workers)	0.345 (0.220)	0.334 (0..220)	0.297 (0.221)	0.302 (0..220)
ln (nonproduction workers)	0.281 (0.231)	0.289 (0.231)	0.300 (0.232)	0.295 (0.233)
ln (capital)	0.047 (0.089)	0.053 (0.090)	0.080 (0.090)	0.079 (0.090)
Plant specific controls		+		+
Plant-fixed effect	+	+	+	+
Year-fixed effect			+	+
R²	0.44	0.43	0.47	0.48
Note: Standard errors are adjusted for clustering on the plant. Specification (2) and (4) include plant characteristics such as years of operation and multi-unit status. The sample size is 388 observations. * p < 0.01, p < 0.05, * p < 0.1.

According to the most robust specification as in column (4), one percent increase in the share of nonproduction workers is associated with productivity increase equal to 7.1 percent. To interpret this figure, consider that the average yearly increase of nonproduction workers between 2009 through 2015 is about 0.2 percentage points. Therefore, an increase in nonproduction workers of 0.2 percentage points is estimated to raise value-added by 1.42 percent annually. For the average plant in three ASEAN countries, this figure is up to about $55,000 per year.

One may wonder why the coefficients of all input variables are not statistically different from zero despite having positive signs. It implies that factor endowments may suffer from endogeneity issue. It means that there may be a correlation between the inputs and the unobservable productivity shocks. However, because I focus on the analysis of the human capital spillovers variable, such condition is not a big concern. Regardless, I still try to correct such problem using TFP specification, which is presented in Section 5.2.1, to prove that we do not need to worry about this.

To sum up, the evidence in Table 3 shows strong evidence that human capital level within region leads to significant increase in plant productivity. As explained earlier in the previous paragraphs, one percent growth in nonproduction worker share of a given plant in any particular region, there will be a corresponding rise of 71.1 until 8.8 percent in the value-added of the plant over a five-year period. If these estimated figures are annualised, the effects are almost three times as big as one measured by Moretti (2004b). A primary reason of this is that the growth effect is expected to be stronger for countries in the catching up process of economic development stage (Benhabib and Spiegel, 1994). A common situation in developing countries is that their level of human capital is still low whereas human capital in the advanced economies has reached a high level. The opportunity to improve such condition is more opened for countries in the “take-off” phase. As a result, the effect on economic performance which is measured using plant productivity in this paper will be larger than for those identified as developed economies.

5.1.2 Instrumental Variable Estimation

Although some unobserved factors have been particularly considered in section 5.1.1, there is still some potential endogeneity bias in the estimators. The inclusion of plant and year-fixed effect only deal with omitted variable bias, but not the reverse causality problem. More productive plants, for example, are more likely to locate in the region with more educated labour force and as a result to hire more nonproduction workers. Therefore, I try to solve this type of issue using instrumental variable approach. I suggest using the number of higher education institutions within regions at one year lag as the variable of instrument following Moretti (2004a). However, in this case, I will restrict my sample for plants located in Indonesia and Philippines only due to unavailable information about it for Vietnam. ⁸

The presence of higher education institutions will be a valid instrument if it is correlated with the regional human capital level and uncorrelated with the error term. I can test for the former requirement, but not for the latter if I have only one instrument. They are often thought to be a suitable instrument for the local human capital level because they are the best place to produce highly-skilled workers. I use the lagged number of universities and colleges, so they will not affect current productivity. However, it still could be the case that higher education institutions improve plant productivity beyond the effect that they have on creating more skilled population. For example, higher education institutions may also bring external dollars into the local area through grants to conduct research. Such an inflow of dollars from outside the area might improve productivity for local plants because plants may use the outcomes of research for their own purpose through public access or private contracts. Thus, the presence of a higher education institution may be a questionable instrument.

From the econometric point of view, Moulton (1990) describes when micro unit observations like firm, includes the aggregate variables, the standard errors will be biased downward if there is a correlation in the error term of individual units that share the same value of aggregate variables. In this case, I assign the logarithm of total higher education institutions in the regions to each plant based on its location. Hence, test statistics for this approach is sensitive to the type of clustered standard errors used. Clustering standard errors at the regional level will lower first-stage F-statistics, which has tendency view instrument as weak, but at the same time, it will lower Hansen-J statistics, increasing the likelihood to interpret the model as correctly specified. Since theory does not offer firm guidance about this, in this paper standard errors will be clustered at the regional level because the spillovers variable also varies over time within the region.⁹

Table 4

Estimates of Human Capital Spillovers – IV Estimates
	(1)	(2)
First Stage:
ln (Number of Universities)	0.064*** (0.024)	0.060** (0.026)
F statistic	6.73	4.87
Second Stage:
Nonproduction worker share in region	59.131*** (17.719)	77.528*** (25.418)
ln (production workers)	0.577* (0.307)	0.734** (0.363)
ln (nonproduction workers)	0.252 (0.265)	0.257 (0.322)
ln (capital)	0.031 (0.076)	0.011 (0.090)
Plant specific control		+
Note: Figures in parentheses are clustered standard errors at the regional level. All regressions include plant-fixed effect. Specification (2) include plant characteristics such as years of operation and multi-unit status. The sample size is 318 observations. * p < 0.01, p < 0.05, * p < 0.1.

The two-stage least square estimation results are highlighted in Table 4. From the first-stage regressions in both models, we can see clear evidence of a positive correlation between a number of universities and the employment share of nonproduction workers. The first-stage coefficients are between 0.060 and 0.064. To help in interpreting with the first-stage regression, consider the instrument has a mean (in logarithmic value) of 5.470. The number of higher education institutions accounts for 0.044–0.047 percentage-point increase in the aggregate nonproduction workers share outside the plant or about 1.5–1.7 percent of the typical increase in skilled workers over a six-year period. In addition to that, the coefficients are statistically significant at 1 percent level. This result implies that the proposed instrumental variable satisfies the relevant condition. However, the F-statistics in the first stage regressions is reasonably small, less than 10. According to Stock and Yogo (2005), I cannot reject the null hypothesis of the weak instrument at 5 percent significance level. This indicates that the instrument I use cannot predict endogenous variable very well.

The second stage estimation specifies that when the aggregate human capital within the region (as explained by the instrumental variable) rises by 1 percent, plant productivity under model (1) and (2) will increase by 59.1 and 77.5 percent over the five-year time span. If we compare the coefficients between Ordinary Least Square and Instrumental Variable model, the magnitude of the former is way lower than the latter. Thus, one may underestimate the spillovers effect of highly skilled workers when I do not deal with endogeneity problem. Findings under two-stage least square estimation also show us that our estimates of human capital spillovers are robust and can be stated as causal inference. To conclude, based on the results using instrumental variable approach, external human capital faced by plants in the region can still be determined to have a positive effect on plant productivity.

5.2 Robustness Checks

5.2.1 Estimation using TFP

In this subsection, I use TFP as the dependent variable for an indicator of plant productivity. Such measure is useful to check whether the factor inputs are correlated with unobserved productivity shocks. If the spillover effect based on TFP is contradictory with the findings in Section 5.1, the previous results may be regarded as unstable ones. When using TFP estimation, I argue that a correlation may exist between factors of production and productivity shocks. If plants receive positive productivity shocks (e.g., new technology on manufacturing process), plants will encourage to increase their inputs because the expected future output will rise as well. Hence, input endowments become endogenous in this case.

If this problem is not addressed properly, results based on OLS technique may be biased. Thus, Olley and Pakes (1996) suggest a method to such issue by using investment decision of the plants to capture the correlation between unobserved productivity shocks and factor inputs. They argued that changes in investment level is a useful instrument to reflect the change in unobservable productivity shocks and these are found to be monotonically increasing.

However, the main drawback of this method is that, in the data, level of investment in some plants is usually zero, since plants may not invest every year. Hence, Levinsohn and Petrin (2003) point out this limitation and provide more appropriate instrument using intermediate input such as material or electricity cost. Given the data that investment value for many plants is unavailable, I adopt the Levinsohn and Petrin (2003) method (L&P TFP) as the plant productivity indices. In this case, I use the logarithm of material cost as the proxy to capture the unobserved productivity shocks at the plant level. ¹⁰

Estimation results in Table 4 show that the estimated human capital spillovers effects vary from 1.987 to 3.857. As we can see in the third row as the most robust TFP specification, one percent increase in the proportion of nonproduction workers will lead to 2.4% increase in the TFP of the plant. If we compare the result between Tables 2 and 5, we can see that although the coefficients in the former are much larger than in the latter, there are no difference in statistical significance among them. This outcome indicates that human capital level in the region has a positive effect on plant productivity under different productivity measure.

Table 5

**Estimates of Human Capital Spillovers Based On TFP**
	(1)	(2)	(3)
Nonproduction worker share in region	3.857** (2.020)	1.987** (0.970)	2.404* (1.470)
Plant-fixed effect	+		+
Time-fixed effect		+	+
Note: Standard errors are adjusted for clustering on the plant. The dependent variable is Levinsohn and Petrin (2003)’s TFP (L&P TFP). The proxy used is the logarithm of material cost. All specifications include plant characteristics such as years of operation and multi-unit status. The sample size is 388 observations. * p < 0.01, p < 0.05, * p < 0.1.

5.2.2 Alternative Measures to Region-Human Capital

Now, I check the robustness of estimated spillovers effect if I use an alternative measure of region-human capital. Following Liu (2013), the new indicator is the logarithm of average yearly wage within a region and excluding the plant itself. The identification mainly comes from the variations over time within the region. The underlying of assumption is that the wage is fairly a good proxy to determine the level of human capital endowment under conventional wisdom. Firms pay their workers equal to their marginal product, and the marginal product is positively correlated with human capital of workers. Based on this theory, if the proportion of nonproduction workers is high in any particular region, one may expect that the productivity in that area will be large as well.

However, firms may not pay their workers their marginal product and wages is determined by the profitability of the firms. To deal with this problem, all regressions include the plant-fixed effect, and as a result, the spillovers effect are identified by within-firm variations in the wage bill. Therefore, as long as firms pay proportionally nonproduction workers higher than production workers and there is no huge variation in profitability across years, this proxy is acceptable.

Table 6

Estimates of Human Capital Spillovers: Alternative Measure of Region-Human Capital
	(1)	(2)	(3)	(4)
Log of Average Yearly Wage in the Region	0.381** (0.174)	0.393** (0.183)	0.020 (0.243)	0.034 (0.246)
ln (production workers)	0.316 (0.227)	0.311 (0.228)	0.274 (0.227)	0.280 (0.226)
ln (nonproduction workers)	0.307 (0.240)	0.307 (0.244)	0.288 (0.246)	0.281 (0.247)
ln (capital)	0.060 (0.097)	0.063 (0.097)	0.087 (0.097)	0.085 (0.097)
Plant specific controls		+		+
Plant-fixed effect	+	+	+	+
Year-fixed effect			+	+
R²	0.52	0.51	0.52	0.51
Note: Standard errors are adjusted for clustering on the plant. Specification (2) and (4) includes plant characteristics such as years of operation and multi-unit status. The sample size is 385 observations. * p < 0.01, p < 0.05, * p < 0.1.

I entirely duplicate the regressions of Table 2 using this new measure. The spillovers estimates are both positive and statistically significant. If I take column (1) as an example, one percent increase in wage outside the plant within a region is positively associated with plant productivity by 0.38%. Additionally, it is also possible to combine both plant-specific characteristics and control for time-varying shocks by adding year-fixed effect into the model. As we can see in column (3), surprisingly, the estimated spillover effect still has a positive sign but statistically insignificant. Furthermore, the magnitude is far lower than two previous rows. Despite such imperfection, this measure is partially a good proxy for region-human capital.

The potential explanation for this finding is that the variation of plant profitability over the five-year period seems high and regional wage only slightly change during the period. This condition is often called as a sticky-wage theory. This theory states that the pay of employed workers tends to respond more slowly to the changes in the firm or economic performance. To illustrate this concept more apparent, I provide the simple correlation between average yearly wage level within the region, outside the plant, and plant value-added (both are in logarithmic values).

We can see that the regional wage levels are not positively attributed to higher plant productivity. Within the region, many productive plants are more likely to pay their workers at the similar rate as less productive plants do. As a consequence, it cannot be inferred whether the regional wage rate can reflect the aggregate level human capital in the area. From the perspective of employer’s interest, Solow (1979), firms will offer wage stability regardless of the output level to minimise their cost, especially in the short-run. Furthermore, if even there is a downturn in output, firms will not cut wage rate because it will have an adverse effect on worker’s morale and hence reduce production. So, it is safer for plants to offer stable wage because it will protect production continuity in the crisis period when their output falls.

5.2.3 Human Capital Spillovers in High Tech Industries

Now we turn our analysis depending on the separation between high and low-tech industries, since the effect of spillovers may not be equal across industries. In general, human capital spillovers within high-tech industries are expected to be larger than low-tech industries because the importance of human capital in the production process is more substantial for the former type of industries. The underlying argument is that plants in high-tech industries, which face rapid change in technological progress, are more sensitive to human capital spillovers than plants in low-tech sectors. To verify this prediction, I adopt a strategy that separating each plant within a region into low-tech and high-tech industries based on OECD classification.¹¹ I then go onto calculating the share of nonproduction workers faced by each plant from high-tech and low-tech industries separately.

In the next step, I test whether the human capital externalities are more important in high-tech than low-tech industries. More importantly, I also examine whether the spillovers from the human capital in high-tech industries of the region as a whole are more important for high-tech plants than human capital from low-tech industries of the region and vice versa. Orlando (2004) claims that the size of spillovers between firms depend on the sectorial proximity, where spillovers among firms in similar industries are usually larger than among firms with different type of industries. Thus, for instance, we expect that medical industry will benefit more from white-collar workers in chemical industry than from food industry. If results are not consistent with these expectations, the spillovers relationship between high-tech and low-tech industries emerge new pattern.

Table 7

Plants and Workers Distribution, by Industrial Technological Intensity
	Number of Plants			Number of Nonproduction Workers
	Indonesia	Vietnam	Philippines	Indonesia	Vietnam	Philippines
Low-Tech	87	32	30	45.6	13.4	20.7
High-Tech	18	3	24	55.6	8.1	57.3
Note: These figures above are averaged between 2009 and 2015.

Table 7 above shows the details on the distribution of plants based on technological intensity, along with the average workers employed by each type of industry. We can see clearly that high-tech sectors hire more skilled than the low-plants, although the number of plants in high-tech industries is far lower than their counterparts. However, this pattern does not occur in Vietnam. Each country indicates the similar feature which is consistent with our analysis of summary statistics. The manufacturing sector in the Philippines is the country which has the most proportion of plants in high-tech industries. 40% of total plants in the Philippines are classified as high-tech plants. This coincides with the fact that on average, the number of nonproduction workers per plant hired in the Philippines is 57 workers, the biggest among other countries. The reverse condition does happen for Vietnam, manufacturing sector. From this figures, we can expect that the size of spillovers will be higher for high-tech rather than low-tech industries.

For the estimation method, I use the dummy variable for each industry type interacted with human capital outside the plant for both high-tech and low-tech industries. The estimation results are presented in Table 8 with the logarithm of value-added as the dependent variable. The first row focuses on human capital spillovers from high-tech industries on high-tech plants and low-tech plants. In a similar fashion, coefficients in the second row indicate the magnitude of human capital spillovers from low-tech industries toward both high-tech plants and low-tech plants. All entries in each panel are from same regressions. In row 1 column (1), for example, is the coefficient on nonproduction workers share in high-tech industries interacted with dummy variable equal to one if the relevant plant is high-tech. Similarly, in row 1 column (2), is the coefficient on white-collar share in high-tech industries interacted with dummy variable equal to one if the relevant plant is low-tech.

Table 8

Human Capital Spillovers between High and Low-tech Industries
	High-tech Plants (1)	Low-tech Plants (2)
Nonproduction worker share of high-tech industries in region	0.301 (2.584)	2.834** (1.153)
Nonproduction worker share of low-tech industries in region	8.530 (7.295)	12.564** (6.508)
Plant-fixed effect	+	+
Time-fixed effect	+	+
Note: Standard errors are adjusted for clustering on the plant. All specifications include plant characteristics such as years of operation and multi-unit status. The sample size is 388 observations. * p < 0.01, p < 0.05, * p < 0.1.

Results in Table 8 show that low-tech plants reap the benefit from more skilled workers in both high-tech and low-plants, with effect from low-tech industries (12.564) are approximately four times greater than high-tech industries (2.834) while there is no significant impact of such type of workers for high-tech plants. The implication for low-tech plants is higher from high-tech industries although the effect is larger from low-tech industries. These findings show that the presence of highly skilled workers will generate spillovers only for low-tech plants. The conclusion drawn from this subsection somehow is not in line with our expectations of human capital spillovers, (i.e., the spillovers between relevant industries have a larger effect and exhibit greater magnitude in high-tech industries). We can infer that the composition of the manufacturing sector in ASEAN countries is dominated by low-tech industries such as food and textile industry. The importance of such low-skilled industries is enormous for the economy in those countries. Thus, the presence of skilled workers within the region will mainly spill over to them since the proportion of plants in high-tech industries is not substantial and less influential to have an impact on other plants, even for other high-tech plants.

5.2.4 Other Tests for Robustness Checks

I conclude this subsection by performing a series of alternative specifications to review the results reported in Table 9. Firstly, I use the logarithm of sales instead of value-added as the dependent variable, and the result is shown in the first row.¹² The coefficient decreases to 5.212, compared to the value of 7.152 in Table 3 of row (4).

I also divide the plants into size groups of 5–19 employees, 20–99 employees, and equal or more than 100 employees. Then, I run separate regressions to estimate human capital spillovers within these categories. The results, which are documented in row (2)-(4), show that no clear pattern arises. Spillovers only occur for small and large plants only while there is insignificant and small effect for medium plants. Large plants can generate higher value-added from the presence of skilled workers than small plants. One percent increase in the proportion of nonproduction workers will induce plants to produce 14.3% higher value-added for small plants whereas this figure is nearly 19.2% for plants with 100 workers or more.

Table 9

Other Robustness Checks
Specifications	Coefficient on white-collar worker share outside plant in the region
(1) Sales	5.212** (2.650)
(2) Small Plants (5–19 workers)	14.255* (8.377)
(3) Medium Plants (20–99 workers)	0.567 (4.950)
(4) Large Plants (workers + 100)	19.209*** (6.796)
(5) Single-unit Plants	8.285** (3.588)
(6) Multi-unit Plants (7) Translog Production Function	63.245* (0.068) 7.063 (3.502)
Note: (2) The number of observations is 89; (3) The number of observations is 159; (4) The number of observations is 140; (5) The number of observations is 333; (6) The number of observations is 55. Multi-unit status is excluded for estimating specifications in row (5) and (6). All regressions include plant and year-fixed effect. Standard errors are adjusted for clustering on the plant. * p < 0.01, p < 0.05, * p < 0.1.

The next thing I do is to separate the plants based on by their multi-unit status. Single-plant firms should be more responsive to changes in their local conditions than firms that have plants in several locations (Henderson, 2003). According to this argument, multi-plant firms more heavily rely on their internal-firm connections and hence these firms are more protected from the local external environment than single-plant firms. Thus, I expect that the magnitude of human capital spillovers is stronger for single-unit plants than multi-establishment plants. Row (3)-(4) of Table 8 report the corresponding results for single-unit and multi-unit plants. The findings are surprisingly not consistent with our prediction. The impact of human capital spillovers for multi-unit firms is 63.245 and only 8.460 for single-unit firms. This condition may happen when multi-establishment plants apply dual human capital networking system, internal and external firm. These firms take advantage of being located in dispersed location to absorb the regional spillovers better, which complement their ability to exploit internal-firm human capital.

As a last specification check, I perform a more general form of production function, which is a translog specification. I then estimate the Eq. (3) including the squared of each log input and the interaction term between each log input. The coefficient of interest is somewhat sizeable compared to the main specifications in Table 3. One percentage increase in nonproduction workers within the region will increase value-added of the plants by approximately 7 to 9 percent under Cobb-Douglas functional form whereas the impact of nonproduction workers on plant’s value added is expected to be around seven percentage point using translog production function. This finding implies that the results are robust with regard to the different type of production function.

In this paper, I investigate the existence of regional human capital spillovers in ASEAN. In principle, I test whether plants located in the region with high level of human capital can produce more output with a similar amount of input. I also provide an instrumental variable approach to solve endogeneity issue caused by the correlation between regional human capital level and plant productivity. Additionally, I also perform some robustness checks to test the validity of findings under alternative measure or specifications.

The results clearly reveal that after controlling for plant-specific characteristics, unobservable plant as well as time-varying heterogeneity, while also excluding the sample plant itself, the percentage of nonproduction workers in the region is increased by 1 percent, and approximately increases plant productivity by around 7.15 percent.

To overcome endogeneity problem originated from unobservable time-varying factor in the regional human capital and plant productivity, I select the number of higher education institutions in the previous year as an instrument to facilitate two-stage estimation. The results also show that the causal effect of human capital spillovers obviously happens and significant since we may underestimate the impact when using OLS estimation. However, this instrument is considered as a weak instrument.

As a check of the main specification, I also examine the spillover effect using Levinsohn and Petrin (2003) TFP as the measure of plant productivity. There is no statistical difference with the Ordinary Least Square and TFP estimation, although the magnitude of spillovers obtained is substantially lower when using the latter. This suggests that the factor endowments are exogenous and uncorrelated with regional human capital spillovers.

In addition to that, I also employ average yearly wage within the region as another independent variable to measure the aggregate level of human capital in the area. My results suggest that the effect of human capital spillovers are still associated with higher productivity, although it cannot confirm the spillover effects when time-varying unobservable factors exist.

When drawing our attention to spillovers effect across industries within a region, I separate the human capital in any given region into low-tech and high-tech industries to examine the impact of spillovers from each type of industry. The results show that low-tech and high-tech industries only spill over their benefit of their highly skilled workers to plants in low-tech industries only, while the effect from the former is found to be stronger. This kind of situation reflects that industries with highly skilled workers are far less influential in affecting other plant's productivity in developing countries, particularly in ASEAN.

In conclusion, I present some microeconomic evidence that human capital spillovers is robust and significantly benefit productivity in ASEAN manufacturing industry. The monetary benefit of this spillovers effect is large, which yield productivity gain with rising per plant value-added by US$ 55,000 and a total gain of US$ 866.75 million over the next six years.¹³ There could be several explanations of such relatively large magnitude of regional human capital spillovers in ASEAN. Rosenthal and Strange (2008) explain that the benefit of human capital spillovers is inversely related to distance and depends on the proximity to highly skilled workers. Thus it is more likely that plants in ASEAN are concentrated in the clustered area (e.g., industrial district), and the reason behind it is to have closer access to skilled workers. When this does exist, workers will have more opportunity to interact more with the workers outside their plants as they work in close collaboration and therefore facilitates the flow of ideas. Then, such new knowledge will be translated into larger plant productivity.

I can draw two policy implications from our study. First, there is clear evidence that firms are benefiting from by locating in areas with an abundant amount of highly-skilled workers. Policies that want to reduce the regional inequality of human capital level need to consider this fact carefully and alter the incentives firms have to locate in the regions with a higher concentration of human capital density. The same explanation also applies to the labour force. Skilled workers are more likely to move to areas where more industrious jobs are accessible and again trying to adjust the incentive arrangement in this respect is vital (although challenging) aspect of these policies. Therefore, incentives driven from policies will determine how firms and workers are willing to move to areas with lower quality of human capital to spread the impact on plant productivity around the country as a whole. Second, while workers’ skill is closely associated with workers’ education, encouraging the local population to pursue more education is an effective way in the context of regional development. It will raise the stock of skilled labour, and at the same time, it increases the supply of better human capital which helps plants to hire skilled workers at a lower wage. Based on my results, this aim can be achieved by encouraging either government or private sector to establish new universities or colleges. Also, policies which encourage plants to increase their skill intensity is a useful instrument for regional development.

I would like to end this paper by providing some limitations in this work regarding any interpretation of my findings. First, this article does not observe the effect of spillovers separately for each industry with different characteristics such as technology intensity. When technological progress is not similar across industries, some industries may benefit more from the human capital level of industries with similar characteristics. Moreover, since plants in service sector are not included here, I cannot investigate whether the spillover effect coming from service sector can affect manufacturing productivity. Second, this work does not decompose the spillovers within and between industries as I emphasise on the regional boundaries. When the connection between two or more industries is strong, the effect of human capital is expected to be large as well. For example, if plants in industry A and industry B have vertical linkage, the level of human capital in industry A will have a greater impact on industry B and vice versa.

As the existence of human capital spillovers has established, the progressive agenda is to determine the exact mechanism of which spillovers arise. One of the possible sources is knowledge exchange through foreign-technology licensing. When plants acquire foreign technology, it is assumed that only skilled workers can operate it most effectively. From my dataset, I find that around 22% of plants in ASEAN manufacturing sector acquire such foreign technology for both periods. Hence, an investigation into the relationship between foreign technology and human capital spillover effect will be an interesting topic for future research.

Funding No funding required from other parties

Conflict of Interest The authors declare no competing interests.

Ethical approval The author never required any ethics approval for the study.

Informed consent None; the research never involved humans or animals.

Author contribution The author has been personally and actively involved in substantial work leading to the paper and will take public responsibility for its content. The author is responsible for any errors.

Data Availability Statement The data underlying this article are available in https://www.enterprisesurveys.org/en/data

Acemoglu, D., 1996. A Microfoundation for Social Increasing Return in Human Capital Accumulation. Quarterly Journal of Economics, 111(3), pp. 779-804.
Acemoglu, D. & Angrist, J., 2000. How Large are Human Capital Externalities? Evidence from Compulsory Schooling Laws. NBER Macroeconomic Annual, Volume 15, pp. 9-59.
Adams, J. D. & Jaffe, A. B., 1996. Bounding the Effect of R&D: An Investigation Using Matched Establishment-Firm Data. The RAND Journal of Economics, 27(4), pp. 700-721.
Benhabib, J. & Spiegel, M., 1994. The Role of Human Capital in Economic Development: Evidence from Aggregate Cross-country Data. Journal of Monetary Economics, Volume 34, pp. 143-173.
Berman, E., Bound, J. & Griliches, Z., 1994. Changes in the Demand for Skilled Labor Within U.S. Manufacturing: Evidence from the Annual Survey of Manufactures. Quarterly Journal of Economics, 109(2), pp. 367-397.
Breusch, T. & Pagan, A., 1980. The Lagrange Multiplier Test and Its Applications to Model Specification in Econometrics. Review of Economic Studies, 47(1), pp. 239-253.
Chang, C.-F., Wang, P. & Liu, J.-T., 2016. Knowledge Spillover, Human Capital, and Productivity. Journal of Macroeconomics, Volume 47, pp. 214-232.
De la Fuente, A. & Domenech, R., 2001. Schooling Data, Technological Diffusion, and the Neoclassical Model. American Economic Review, 91(2), pp. 323-327.
Duranton, G. & Puga, D., 2004. Micro-Foundations of Urban Agglomeration Economies. Handbook of Regional and Urban Economics, Volume 4, pp. 2063-2117.
Glaeser, E. L., 1999. Learning in Cities. Journal of Urban Economics, 46(2), pp. 254-277.
Greene, W. H., 2003. Econometric Analysis. 5th ed. New Jersey: Upper Saddle River.
Henderson, V., 2003. Marshall's Scale Economies. Journal of Urban Economics, Volume 53, pp. 1-28.
Jovanovic, B. & Bob, R., 1989. The Growth and Diffusion of Knowledge. Review of Economic Studies, 56(4), pp. 569-582.
Kuznets, S., 1962. Population Change and Aggregate Output. In: Demographic and Economic Change in Developed Countries. Princeton: Princeton University Press, pp. 324-340.
Levinsohn, J. & Petrin, A., 2003. Estimating Production Function Using Inputs to Control for Unobservables. Review of Economic Studies, 70(2), pp. 317-341.
Liu, Z., 2014. Human Capital Externalities in Cities: Evidence from Chinese Manufacturing Firms. Journal of Economic Geography, Volume 14, pp. 621-649.
Lucas, R. E., 1988. On the Mechanics of Economic Development. Journal of Monetary and Banking, 22(1), pp. 3-42.
Marshall, A., 1920. Principles of Economics. 8th ed. London: McMillan.
Moretti, E., 2004b. Worker's Education, Spillovers, and Productivity: Evidence from Plant-level Production Function. The American Economic Review, 94(3), pp. 656-690.
Moretti, E., 2004. Estimating the Social Return of Higher Education: Evidence from Longitudinal and Repeated Cross-sectional Data. Journal of Econometrics, Volume 121, pp. 175-212.
Moulton, R., 1990. An Illustration of a Pitfall in Estimating the Effect of Aggregate Variables on Micro Unit. The Review of Economics and Statistics, 72(2), pp. 334-338.
Olley, S. & Pakes, A., 1996. The Dynamics of Productivity in the Telecommunications Equipment Industry. Econometrica, 64(6), pp. 1263-1297.
Orlando, M., 2004. Measuring Spillovers from Industrial R&D: On the Importance of Geographic and Technological Proximity. RAND Journal of Economics, 35(4), pp. 777-786.
Park, W., 2016. How Large Are Human Capital Spillover?: Evidence from Korea. KDI Journal of Economic Policy, 38(4), pp. 77-94.
Rauch, J. E., 1993. Productivity Gains from Geographic Concentration of Human Capital: Evidence of the Cities. Journal of Urban Economics, 34(1), pp. 380-400.
Rosenthal, S. S. & Strange, W. C., 2008. The Attenuation of Human Capital Spillovers. Journal of Urban Economics, Volume 64, pp. 373-389.
Solow, R. M., 1979. Another Possible Source of Wage Stickiness. Journal of Macroeconomics, 1(1), pp. 79-82.
Stock, J. & Yogo, M., 2005. Testing for Weak Instrument in Linear IV Regression. In: Identification and Inference for Econometric Models. New York: Cambridge University Press, pp. 80-108.
World Bank Group, 2017. Survey Methodology for Enterprise Surveys. [Online]
Available at: http://www.enterprisesurveys.org/methodology [Accessed 15 July 2017].
Zucker, L. G., Darby, M. R. & Brewer, M. B., 1998. Intellectual of Human Capital and the Birth of US Biotechnology Enterprises. The American Economic Review, 88(1), pp. 290-306.

It is critical to notice that in this paper, I do not test hypothesis in any of these theoretical aspects. Thus, testing them is well beyond the scope of this study.
E(
${x}_{it}$
,
${u}_{i}$
) = 0 and E(
${x}_{it}$
,
${\epsilon }_{it}$
) = 0 for t = 1,....,T
I do not describe the fixed-effect estimation using first difference transformation as it leads to same results with within transformation if there Is only two period in the dataset.
Enterprise Survey use stratified random sampling methodology to collect the key variables from plants (See Data Appendix for detailed explanation). The strata used are based on plant size, sector, and region.
2009: US$1 = IDR 10389.94; US$1 = PHP 47.68; US$1 = VND 17065.08
2015: US$1 = IDR 13389.41; US$1 = PHP 45.50; US$1 = VND 21697.57
The fraction of nonproduction workers for each region is presented in Data Appendix
Summary statistics for each country in details are presented in Data Appendix
The regional distribution of higher education institutions is only available for Philippines and Indonesia. As a consequence, I limit the sample to plants located in both countries only.
I do not apply clustered standard errors at regional level when using OLS estimation because I do not match between two different data sources.
In my dataset, nearly 70% of my observations have zero investment value.
The definition of high-tech industries is calculated based on R&D intensities. They consist of high-tech and medium high-tech industries. These include chemicals and pharmaceuticals; aircraft and spacecraft; ordinary machinery; transportation equipment; electrical machinery; electronics and telecommunications; instruments and office machinery; motor vehicles and trailers.
I use log of sales because my dataset does not provide information on output.
The number of plants in each year is 2580 (2009) and 2673 (2015). With an average of 2627 plants, the pecuniary benefit for manufacturing sector is estimated around US$ 866.75 million.

No competing interests reported.

DataAppendix.docx

Download PDF

Journal Publication

published 01 Apr, 2024

Read the published version in Humanities and Social Sciences Letters →

Version 1

posted

You are reading this latest preprint version

Human Capital Spillovers and Plant Productivity in ASEAN: Evidence from the Plant Panel Data

Status:

Journal Publication

Version 1

Abstract

Figures

1 Introduction

2 Literature Review

3 Empirical Framework

3.1 Model Specifications

3.2 Linear Static Panel Data Models

3.3 Estimation Issues

4 Data and Variables

5 Result

5.1 Basic Results

5.1.1 Ordinary Least Square Estimation

5.1.2 Instrumental Variable Estimation

5.2 Robustness Checks

5.2.1 Estimation using TFP

5.2.2 Alternative Measures to Region-Human Capital

5.2.3 Human Capital Spillovers in High Tech Industries

5.2.4 Other Tests for Robustness Checks

6 Conclusion

Declarations

References

Footnotes

Additional Declarations

Supplementary Files

Status:

Journal Publication

Version 1