The second important part explained in the MAER-Net guidelines is literature searching, compilation, and coding. We used the databases of Google Scholar, JSTOR, Ideas RePEc, Econlit, and NBER. The keywords used in the literature search on these databases are “FDI,” “FDI inflow,” “employment,” “job opportunities,” and “labor.” In addition, several phrases are also used as keywords, including “FDI on employment,” “FDI inflow on employment,” “FDI and job opportunities,” and “FDI on labor.”

This study determined the inclusion criteria for selecting the literature reviewed. First, the literature must use at least one of the FDI proxies as the explanatory variable and one of the employment proxies as the dependent variable. Although the unemployment rate is often used as one of the employment proxies, that proxy was hostile (negative). Therefore, we excluded the literature that used unemployment as the dependent variable. For more details, the econometric model must examine the effect of FDI inflow on employment by using the following equation:

Y = α + β₁FDI + β_x Z + ε, (3)

where Y is employment; FDI is FDI inflow; Z is the vector of other explanatory variables used in the model, and ε is the error term.

The second inclusion criterion is that the literature reviewed must report econometric estimation results. According to Stanley and Doucouliagos (2011), to be used as a meta-analysis dataset the literature must provide the results of the regression coefficients. The third inclusion criterion is that it must examine the direct effect of FDI inflow on employment. Therefore, this study excluded the literature that examined the indirect effect of FDI inflow on employment. Lastly, the literature must be written in English so as not to cause errors in understanding due to language problems.

The dataset was collected from February to May 2022. We collected 61 publications with a total of 490 estimates. However, according to Havranek and Irsova (2011), we also attempted to anticipate data outliers. In this context, the estimate is considered an outlier if its t-statistic value exceeds 10. From the 490 estimates, we identified 13 estimates that had more than 10 in t-statistic. Thus, our final number of estimates employed as the dataset is 477. Furthermore, following the recommendation of MAER-Net, the dataset should be accessible to the public. To fulfill this term, our dataset is available in Supplementary material.

According to the recommendations of MAER-Net, the study dataset and moderator variables also need to be described. In fulfilling this term, the descriptive statistics of the datasets are presented in Table 1.

Table 1 shows that the average effect resulting from each study has heterogeneity. The effect of FDI on employment from these studies was positive and negative. The heterogeneity of the magnitude of the effect is higher than the average value. Although Table 1 shows the average effect of FDI on employment, this is not the actual mean effect size. The average effect in Table 1 is obtained by finding the average value of each estimated regression coefficient produced by the study. Meanwhile, the mean effect size is the average effect which magnitude has been weighted to show the original effect of FDI on employment.

However, the mean effect size does not control the possibility of a publication selection bias. In this context, we employed a basic meta-analysis to ascertain the dataset’s mean effect size of FDI’s effect on employment and used MRA with the FAT–PET technique to find the genuine effect (effect beyond bias).

Studies on economic indicators such as supply and demand for labor or the relationships — consumption, investment, imports, exports, and production tend to have an endogeneity bias (Baltagi, 2005). In this context, endogeneity is a condition where explanatory variables are correlated with error terms (Ullah et al., 2018). To anticipate such a bias, Baltagi (2005) suggested an instrumental variable (IV) based analysis or Generalized Method of Moments (GMM). Because of this reason, we identify the literature that employed IV or GMM method as the studies that control endogeneity.

We anticipate an endogeneity bias in the literature in two ways. First, we classify the data into three categories in the FAT–PET analysis. They are the overall sample, ignoring endogeneity and control endogeneity. Thus, differences in the publication bias and genuine effects among samples will be identified. Second, this study includes the SEPcc × NoEndog as one of the explanatory variables in the multiple MRA in order to examine the difference between the overall standard error coefficients and the standard error coefficient from the literature that does not control endogeneity.

We estimated the basic meta-analysis to identify the mean effect size and heterogeneity. We use I² and τ² (tau square) values to detect heterogeneity. If I² exceeds 75%, it indicates great heterogeneity (Higgins and Thompson, 2002). Meanwhile, the value of τ² is the variation among studies or the standard deviation distribution that underlies the mean effect size. Therefore, greater τ² indicates greater heterogeneity.

The value of τ² can be generated by Restricted Maximum Likelihood (REML), Sidik–Jonkman (SJ), Hedges, or Random Effects Empirical Bayes (EB). Meanwhile, I² can be generated from a Fixed Effect Estimator (FEE), REML, Maximum Likelihood, or EB. Therefore, we estimate the basic meta-analysis using REML, FEE, and Random Effects EB (see Table 2).

The FEE in Table 2 assumes that all reported estimates come from the same population as the common mean. Therefore, it relatively produces a smaller mean effect size. REML tends to be more relevant because the estimates come from different populations. In this context, although in some literature REML stands for Restricted Maximum Likelihood, it also stands for Multilevel Random Effect (Stanley and Doucouliagos, 2011). Therefore, FEE is a fixed effect, while REML is a random effect. The estimation of EB in the third column resulted from REML estimation with empirical Bayes iterative procedure.

Furthermore, from the three estimates in Table 2, all mean effect sizes are positive. It shows that based on three estimates, the effect of FDI on employment is positive. However, the value tends to be low. Table 2 does not include the Cochran Q-test results, which the meta-analysis often employs to detect heterogeneity (Stanley and Doucouliagos, 2011). This study estimated the Q-test by regressing the t-value against the precision (1/SEPcc). The value of sum square errors from the regression results is a Q-test distributed as a chi-square with L – 1 degrees of freedom. The resulting Q-test value is 3,624 with a mean sum square error of 7.614 and a probability lower than 0.05. These results indicate that heterogeneity among studies is significant. Because of this reason, there has been heterogeneity among studies regarding the effect of FDI on the host countries’ employment.

The basic meta-analysis procedure presented in Table 2 does not control for a possible publication selection bias and heterogeneity. Meanwhile, in the meta-analysis, examining the publication bias is critical. This bias is caused by a publication selection. In this context, one of the motives of researchers conducting such a selection is when they tend to prioritize reporting significant results. The publication selection bias usually arises when editors only publish studies relevant to a particular topic (Stanley and Doucouliagos, 2011). Simply speaking, the publication bias is a condition where negative results are not published (Cleophas and Zwinderman, 2017). The publication selection bias could distort research findings. In this study, we detected it using a funnel graph and FAT–PET.

We perform the funnel using precision (1/SEPcc) as the y-axis and the partial correlation coefficient as the x-axis (see Fig. 2). Fig. 2 shows that the distribution of the lower-left area’s partial correlation coefficients from the literature tends to be asymmetrical. The studies that report negative results tend to be fewer than those that report positive ones. In other words, some researchers try to report positive results only. However, this funnel plot tends to be subjective (Stanley and Doucouliagos, 2011). Because of that reason, we also performed FAT–PET to correct the publication bias. This study calculates the FAT–PET by referring to Stanley and Doucouliagos (2011), who employed this formula:

Pcc_i = β₀ + β₁SEPcc_ij + ε_ij, (4)

where Pcc_i is the partial correlation coefficient from the i-th study; SEPcc_ij is the standard error from the ith estimate on the jth study; β₀ is the correction of a publication bias, known as the effect beyond bias or genuine effect; β₁ is a publication bias; ε_ij is the error term.

According to Stanley and Doucouliagos (2011), the FAT–PET in the above equation has heteroscedasticity and cannot be estimated by the ordinary least square (OLS) method. The standard error in the above equation comes from the effect size. Because the effect size between studies has different variances, the weighted least squares (WLS) are needed. Therefore, we estimate the FAT–PET with five methods: OLS, Fixed Effect, REML, WLS with 1/SEPcc as weight, and WLS with 1/number of estimates per-study as weight (WLS–WS) (see Table 3).

All estimates from Table 3 have a genuine effect. Compared with the mean effect size of 0.047 in Table 2, the value of the effect beyond bias results from the OLS and WLS estimations are the closest. However, because the FAT–PET analysis contains heteroscedasticity, the WLS results are more relevant (Stanley and Doucouliagos, 2011). From the WLS analysis of the overall sample, we found an upward publication bias and the effect beyond bias. The results of the WLS analysis also show that studies that do not control for endogeneity tend to have a higher upward bias. Meanwhile, a downward bias occurs in studies that control endogeneity.

Studies that ignore endogeneity produced a lower genuine effect than those which control it. Furthermore, according to Doucouliagos (2011), if the genuine effect is lower than 0.07, it indicates a small effect. In other words, the genuine effect of FDI on employment in this study is in a low range. This low genuine effect needs to be studied further by using multiple MRA which can explain heterogeneity with moderator variables to get a more comprehensive picture of the factors that cause it. Several meta-regression studies use the term “modeling heterogeneity”’ to describe multiple MRA procedures.

In addition to showing the possibility of a publication bias, the funnel plot in Fig. 2 also shows that the analyzed studies have heterogeneity. We employed multiple MRA to explain it more comprehensively. The initial step in the multiple MRA is to model heterogeneity as follows:

Pcc_i = β₁ + ∑ β_x Z_xij + β₀SEPcc_ij + ε_ij, (5)

where Pcc_i is the partial correlation from the regression coefficient regarding the effect of FDI on employment from i-th to the number of studies (in this study, there were 61 publications with 477 estimates). SEPcc is the standard error of the Pcc. Z is a vector variable that shows heterogeneity, such as differences in the measurement of FDI and employment, sample country basis, and estimation methods.

Z variables in equation (5) are implemented into moderator variables to explain heterogeneity. We refer to Jarrell and Stanley (1989) in determining Z variables. The moderator variables are created in a dummy variable format based on some categories. Therefore, we set six categories for determining Z variables. They are the type of FDI and employment measurement, data characteristics, FDI receiving countries, estimation method, publications characteristics, and the type of estimation model. In the category of FDI and employment measurement types, we determine eight moderator variables identified from the literature: inward FDI, FDI growth, merger and acquisition FDI, employment, employment growth, unskilled employment, skilled employment, and other employment. Overall, this study identified 6 Z vector groups with 34 moderator variables.

In more detail, following the MAER-Net guidelines on the need to describe variables through descriptive statistics, the definitions of variables are presented in Table 4.

The moderator variables in Table 4 are Z variables identified from the literature. According to the MAER-Net recommendation, the meta-analysis in the economic field needs to simplify a meta-regression model. We employed the Bayesian Model Averaging (BMA) method to fulfill this term. According to Havranek et al. (2017), the BMA method can anticipate the model uncertainty in meta-analysis. In this context, BMA estimates millions of models generated from the sub-sample to determine the model with greater explanatory power (Havranek et al., 2018). BMA adjusted the linear regression model with the uncertainly explanatory variable model to choose the most appropriate one. BMA estimation in this study refers to the procedure described by De Luca and Magnus (2011). The results are presented in Table 5.

Table 5 shows that the BMA identifies six moderator variables that can explain heterogeneity. The estimation of WLS and REML reinforces the results of the BMA as frequentist checks in this study that still includes SEPcc and SEPcc × NoEndog, even though the BMA does not identify these two variables. They were included to re-identify the publication bias from the studies that did not control for endogeneity. When referring to the results of the WLS analysis, there is a reasonably high upward bias from the studies that do not control for endogeneity. These results strengthen the results of the FAT–PET analysis in Table 3. In addition, the analysis results of BMA, WLS, and REML also confirm that the genuine effect of FDI on the host countries’ employment was positive.

We identified eight moderator variables regarding FDI and employment measurement. However, some of these moderators have a relatively shallow size. In addition, some of the other moderator variables have collinearity problems, so the BMA analysis can only use four moderator variables. They are Inward_ FDI, Merger_FDI, Employment, and Employment_Growth. Of them, Merger_FDI got the highest PIP value.

Based on the results of the BMA analysis, we found that the method of measuring FDI and employment that can explain heterogeneity is Merger_FDI (Merger and Acquisition FDI). The negative notation indicated by the t value and the posterior mean of Merger_FDI shows that FDI originating from mergers and acquisitions of foreign companies will have a lower effect on employment. In other words, Merger_FDI reduced the host countries’ employment. WLS also captured the negative effect of Merger_FDI on the latter.

These results strengthen the arguments of Mcdonald et al. (2002), Jenkins (2006), and Akcoraoglu and Acikgoz (2011), who stated that FDI from mergers and acquisitions reduced employment. Our results also support the findings of Jude and Silaghi (2016) that these FDI might reduce employment because the acquired companies tend to be more efficient. FDI from mergers and acquisitions could also reduce employment if foreign companies cut domestic supply in the host country. Such FDI can also attract high technology to replace employment.

Based on the study dataset, we found nine moderator variables from the characteristic data point of view. Most studies used panel data on FDI and employment from databases such as the World Development Index (WDI) and others. Other studies used time series data in a specific country. Only a few studies used cross-sectional data, so we did not include the data as moderator variables to be tested with BMA. Thus, most studies use non-sectoral data (overall), while others use FDI and employment data in specific sectors. Three sectors are identified as the most widely used: manufacturing, service, and other sectors (mining, agriculture, construction, logistics, and others).

In addition, we add Across_Countries and Single_Country as moderator variables in the characteristic data category. Thus, eight moderator variables are tested with BMA: Panel Data, Time Series, Overall (non-sectoral), Manufacturing, Services, Other Sectors, Across Countries, and Single Country. These eight moderator variables get different posterior means notation. For example, Panel Data, Other Sectors, Services, and Single Country have a negative posterior mean. Meanwhile, Time Series, Overall, Manufacturing, and Across Countries have a positive posterior mean. However, of the eight moderator variables, only Other Sectors has a PIP value of more than 0.5. Therefore, only it can explain heterogeneity.

The BMA estimation results show that FDI in Other Sectors tends to have a lower effect on employment. The results of the BMA are reinforced by WLS and REML, which also found a significant adverse effect of the Other Sectors variable on the Pcc. In other words, FDI that is included in the Other Sectors category tends to reduce the level of employment in host countries because FDI entering these sectors is relatively less labor intensive. This stands in contrast to the manufacturing sector, which can absorb more labor. Although the BMA does not identify the manufacturing sector as part of the variable that can explain the Pcc, the posterior mean value of the manufacturing variable is positive. Thus, the type of FDI entry sector determines the increase in employment in host countries.

According to Table 4, from the 477 estimates, 331 came from single-country data, while the rest came from across-country datasets. It indicates that most researchers focused on discussing the effect of FDI on employment at the country level. However, based on the results of the BMA analysis, the moderator variables Across_Country and Single_Country did not have an adequate PIP value. In other words, no empirical evidence exists that using global, regional or country-level data could determine heterogeneity.

This study identifies six moderating variables based on aspects of FDI recipient countries: Asia, Latin America, Europe, Africa, Developing Countries, and Developed Countries. Developing countries cannot be analyzed using BMA because of collinearity. Meanwhile, from five moderator variables analyzed by BMA, Asia, Latin America, and Africa have a positive posterior mean. In contrast, Europe and developing countries have a negative posterior mean. However, only Europe, Africa, and developed countries have a PIP value more significant than 0.5.

Based on the results of the BMA estimation, FDI entering European countries has a lower Pcc. It was confirmed by WLS and REML, which showed that the European moderator variable negatively affected the Pcc. In other words, FDI entering European countries reduced employment levels. One study examining FDI entering European countries was by Hunya and Geishecker (2005). They mentioned that privatized state-owned companies are among the causes of FDI reducing employment rates in European countries. After the takeover of state-owned companies, foreign firms cut off relations with domestic suppliers.

We confirm that FDI entering developing countries also reduced employment. The analysis of WLS and REML strengthens this finding. Therefore, the results of our study indicated that FDI entering developed countries would only increase high-skilled jobs. FDI into developed countries brings more significant technological changes to replace employment. On the other hand, we found that FDI entering African countries has a relatively more significant influence on employment. To judge by their characteristics, most African countries are developing ones.

Therefore, the positive effect of FDI on employment in African countries was not bringing high-technological changes. However, the BMA results related to the association between African moderator variables and the Pcc were not confirmed by WLS and REML, so these results are weak. However, Asiedu and Brepong (2007), who employed an African sample, found that FDI from multinational companies in African countries increases job opportunities. Moreover, Asiedu and Brepong (2007) suggest that African countries attract more FDI through the liberalization of investment regulations.

Our study coded the estimation methods employed by the literature into OLS, other estimation, and control endogeneity estimation. The other estimation method contains publications that used methods other than OLS and control endogeneity, such as instrumental variables and GMM. Several items in the other estimation category include those using least squares dummy variable (LSDV) analysis, Heckman estimation, ARDL, and others.

Unfortunately, of these three moderator variables, only OLS and other estimations can be analyzed by BMA because the moderator variable Control_Endogeneity has collinearity. Finally, of the remaining two moderator variables, only OLS was shown to have a PIP value greater than 0.5. The posterior mean of the OLS moderator variable is positive, indicating that studies using OLS tend to get a higher Pcc. In other words, studies using the OLS method produce a higher coefficient of FDI influence on employment. WLS and REML corroborate the results of this BMA analysis.

Based on Table 4, our study identifies four moderator variables from the aspect of publication characteristics, namely Q1, Q2, Q3–Q4, and Unranked. The most widely found literature came from the Unranked category, followed by Q1, Q3–Q4, and Q2. Q3–Q4 experienced collinearity, so it was not included in the BMA analysis. In this case, the results of the BMA analysis show that Q2 and Unranked have a positive posterior mean value, while Q1 is the opposite. However, none of the moderator variables has a PIP value of more than 0.5. Therefore, we found no empirical evidence that publication characteristics explain heterogeneity.

Each publication estimates a different model. Most literature used several other explanatory variables accompanying FDI in affecting employment. If they also employed output, wages, and technology as other explanatory variables, we identified them as Model 1. This study sets the model estimation type into three moderator variables (see Table 4 for details). However, none of them was identified by BMA. Therefore, we justified that the type of estimation model cannot explain heterogeneity of the effect of FDI on employment.

Our study checks the robustness by excluding estimates from studies not published by the leading journals (indexed by Scopus or Web of Science). Of the 477, only 309 estimates came from them. Furthermore, these estimates were re-analyzed using BMA, WLS, and REML by eliminating all moderator variables in the publication characteristics category. The results are presented in Table 6.

The results of the BMA analysis in Table 6 are slightly different from those in Table 5 because there is a change in the number of datasets. In Table 6, the dataset contains estimates only from the leading journals. The results reveal that six moderator variables have a PIP value more significant than 0.5. They are Manufacturing, Other Sectors, Across Countries, Europe, Africa, and OLS variables. The difference between the BMA results on this robustness checking occurs in Manufacturing, Across Countries, Merger_FDI, and Developed variables. Meanwhile, Other Sectors, Europe, Africa, and OLS variables did not change. Therefore, the four moderating variables are robust in explaining the effect of FDI on employment heterogeneity.

Table 6 indicates that Merger_FDI and the Developed variables could not explain heterogeneity in the leading journals dataset because their PIP value is less than 0.5. However, the posterior means notations for the two moderator variables have not changed, which are negative. On the other hand, in this leading journals dataset, Manufacturing and Across Countries variables get a higher PIP value with positive posterior means notation. Therefore, studies in the leading journals that used FDI in the manufacturing sector and data across countries tend to report higher FDI effects on employment.

This study has several limitations. First, we have not employed the year of a study publication as a moderator variable. Consequently, we cannot justify heterogeneity by this parameter. Second, we only categorize heterogeneity based on the estimation method into three moderator variables: OLS, Other Estimate, and Control Endogeneity. These three moderator variables may be too general because the estimation methods used in the literature tend to vary widely. Lastly, the moderator variable’s estimation results based on the publication type have a potential bias because we identify the types of publications based on journal rankings in 2022. When the study was published, it was possible that the journal had not been indexed by Scopus or had a different Scimago ranking.