*5.2. Regression Results and Analysis*

In this section, static and dynamic panel data models are estimated for the industry panel data available from 2005 to 2016. The data contains 15 industry groups listed in Table 1 and all of them are included in the analysis. Thus, the data includes the entire population of the industry groups. The estimated models are pooled OLS, fixed effects (FE), and random effects (RE) models from the static panel estimators, while difference GMM and system GMM are presented as dynamic estimators. Three different model specifications are used in the estimation. In the first model, industry group production is the dependent variable, while energy, labor, and capital are explanatory variables. In this model, the returns in relation to the scale of production are calculated based on the sum of the coefficients for the three input variables. In the second model, manufacturing labor productivity is specified as employment (labor), capital (fixed assets), value of energy, and time trend (technology) as the explanatory variables. In this model, the coefficients measure the scale effect of the explanatory variables on labor productivity of the industry groups and labor represents the scale effect. In the third model, the manufacturing sector's labor productivity is explained by measuring energy and capital intensities respectively. In all the three model specifications, a trend is included to capture a shift in the labor productivity function or rate of technological change. All variables (with the exception of trend) are transformed into logarithmic form so that the coefficients are interpreted as input elasticities.

Accordingly, Table 4 shows the results of the pooled OLS for the three model specifications. In the first model, labor, capital, energy, and technology are found to be statistically significant and positive. These are among the key factors used for explaining the manufacturing industry's production growth. The elasticity of the output with respect to capital is higher than the corresponding figures for labor and energy in these industries. The returns in relation to the scale of the production process are 1.06 implying increasing returns in relation to scale in this specification coinciding with predictions in the literature [1,67]. In the second model, labor is significant and positive at the one percent significance

level. However, we do not interpret the coefficient of labor and, instead, based on Equation (5d), we find the value of α by adding one to the estimated coefficient in our model, which is zero. Then α, in our case, will be positive, indicating the positive effect of labor on productivity in the manufacturing industries. This can be attributed to the increasing returns in relation to the scale of production and the type of existing industries, which are dominated by labor-intensive industries. In this model, capital is significant and positive for labor productivity, which is a boost for the industry groups. These results are in accordance with Otalu and Anderu and Velucchi Viviani [7,43]. Energy use also positively effects productivity in line with other empirical studies [40–45]. In the third model, capital and energy intensities are significant and positive and help explain labor productivity in the manufacturing industries in line with other studies [46,50,51]. Our results confirm that labor productivity is high and more elastic for energy intensity than for capital in the Ethiopian manufacturing industries. The models show that adjusted R2 is high and the probability of F-statistics is significant, confirming the appropriateness of the model's specifications (see Table 4).


Notes: \*\*\*, \*\*, \* denote the statistical significance levels at 1%, 5%, and 10%, respectively. \*Model 1: Output is the dependent variable. \*Model 2: Labor productivity is the dependent variable (scale effect). \*Model 3: Labor productivity is the dependent variable (input intensity effect).

It should be noted that the pooled OLS model ignores industry effects that may generate biased results. However, it serves well to establish the model's specifications. Table 5 presents the static panel data model estimation results. In this section, only the second and third models are estimated using fixed effects (FE) and random effects (RE) estimation methods. The fixed effects model allows the industry effects and inputs to be correlated, while the random effects model assumes that these are not correlated. The fixed effects model is consistent and unbiased regardless of the correlated effects, but the random effects model is valid and efficient. In this case, since the industry groups are made up of the population of industries, the fixed effects model is a better choice. For a comparison, we estimate the models using both estimation methods.

In the fixed effects model, labor is statistically significant and is a positive factor in explaining the variations in manufacturing productivity in Ethiopia. This is expected based on theoretical predictions as more labor employment induces labor productivity. The fixed effects model's estimation results confirm that energy, capital, and technology positively effects labor productivity, and all of them are statistically significant at the one percent level of significance. The input intensity model based on the fixed effects estimation shows that capital intensity and energy intensity are statistically significant factors for explaining labor productivity in the Ethiopian manufacturing industries. However, in this case, productivity is more elastic in relation to capital intensity than energy intensity. In the random effects model, energy, capital, and technology are positive and statistically significant in explaining the industry groups' labor productivity, while the coefficient for labor is negative, but, based on Equation (5d), α is found by adding one to the coefficient, which gives us a positive coefficient with a value of 0.45. For the intensity model, the random effects estimation approach confirms the significance of energy and capital intensities positively effecting labor productivity. Like the fixed effects model's results, productivity is less elastic in relation to energy intensity than it is to capital intensity. In all the models, the coefficients for trends are positive and significant, implying a positive shift in labor productivity because of technological changes in Ethiopian manufacturing industries during the study period.

The models give different results for some of the explanatory variables, so we cannot take into account the results of all the models. Instead, we must select a model that explains the data using different tests and base the analysis on the optimal model's specifications. To choose between pooled and random effects models, we used the Breusch and Pagan lagrange multiplier (LM) tests with the null hypothesis that pooled OLS is an appropriate model that explains the data better relative to the random effects model. The Hausman test compares the random effects model with the fixed effects model and the null hypothesis for the Hausman test shows that the random effects model is not appropriate for representing the data relative to the fixed effects model. Accordingly, in both cases, the *p*-value of chi2 and chibar2 forces us to reject the null hypothesis. Therefore, the fixed effects model is preferred to the pooled OLS model and the fixed effects model is preferred to the random effects model to represent our data. To control for the heteroscedasticity problem, standard errors reported in all the models are robust.


**Table 5.** Static panel estimation results for Models 2 and 3.

Notes: \*\*\*, \*\*, \* denote statistical significance levels at the 1%, 5%, and 10%, respectively. \*Model 2: Labor productivity as the dependent variable (scale effect). \*Model 3: Labor productivity as the dependent variable (intensity effect).

Table 6 gives the dynamic panel model's estimation results for both the difference GMM and system GMM models. Unlike static panel models, these models include the lag of the dependent variable as an explanatory variable in addition to the other variables. In the dynamic models, problems of heteroscedasticity and autocorrelation are considered. In both the scale effects (Model 2) and the input intensity models (Model 3), lagged labor productivity is found to be significant and positive in explaining changes in the manufacturing industry's labor productivity in Ethiopia. This shows that the previous year's productivity increases current productivity, which, in our case, is labor productivity. An increase in employment for the industry groups has a positive and significant effect which is attributed to increasing returns to scale and the labor-intensive nature of manufacturing industries in both the cases. In both the difference GMM and system GMM models, energy induces labor productivity. Comparing our results with those from developing countries suggests that our results are in line with those from some sub-Saharan African countries, such as those reported by Kebede

et al. and Akinlo [12,18]. However, the effect of energy on productivity for some African countries shows that it is not as important as labor and capital [10,39], signifying the mixed empirical results of the relationship between energy and growth as one major reason for undertaking this specific study. The empirical validation in our case is at the industry level and not at the aggregate national level and this is one of the contributions of this study to the existing literature, as it is what makes this study different from the existing studies. Unlike other studies, the consistency of our results is empirically confirmed using different model specifications and alternative estimation strategies. In addition to the role of energy in productivity, the effects of labor, capital, and technological change on manufacturing productivity are also empirically validated in Ethiopia. This provides crucial policy input for the country's industrial policy.


**Table 6.** Dynamic panel estimation results for Models 2 and 3.

Notes: \*\*\*, \*\*, \* denote the statistical significance levels at 1%, 5%, and 10% levels respectively. \*Model 2: Labor productivity as the dependent variable (scale effect) \*Model 3: Labor productivity as the dependent variable (input intensity effect).

In the input intensity model (Model 3), the elasticity productivity for energy intensity is higher than capital intensity, while the opposite is the case for the system GMM model. Capital is positive and significant in all the models for increasing labor productivity. The coefficient of the time trend has a positive sign in all the models, indicating technological progress with an expected positive effect on the productivity of the industries (see Table 6).

Table 7 discusses the results of the system GMM dynamic estimator, including dummies for trends. Our results show that, in both the models, energy magnitude and energy intensity are statistically significant and positive factors in increasing labor productivity in the manufacturing industry groups; this finding coincides with other findings in the literature [34–36]. Besides, the magnitude of capital and capital intensity are positive factors for labor productivity. In both the models, time dummies are positive throughout. The results show that there is no cyclical effect and, instead, labor productivity increases in both cases over time, which can be attributed to technical changes, increasing labor productivity.


**Table 7.** System GMM dynamic panel with time dummies for Models 2 and 3.

Notes: \*\*\*, \*\*, \* denote the statistical significance levels at 1%, 5%, and 10% levels respectively. \*Model 2: Labor productivity as the dependent variable (scale effect). \*Model 3: Labor productivity as the dependent variable (input intensity effect).

One major objective of this study was to ascertain whether an empirical relationship existed between energy and labor productivity in Ethiopian industries, along with investigating whether it positively affected productivity or limited it. The results of all the models confirm that the energy-related parameter is significant and positive, showing that an increase in energy consumption enhances labor productivity in Ethiopian manufacturing industry groups. This result coincides with other empirical studies [15–17,46]. However, in Ethiopia, agriculture was previously a major source of livelihood for the population. Agriculture was a dominant sector in terms of the employment share up until recently, when traditional services emerged to dominate the economy [21,23]. The share of manufacturing in Ethiopian GDP was very low, indicating output and premature deindustrialization [20,21,24]. This requires serious engagement for identifying and prioritizing the major explanatory factors for the manufacturing industry. Furthermore, manufacturing is more energy intensive relative to other sectors and the interdependence between energy and industries is a crucial tool for sustainable economic development [6,32]. Accordingly, empirically identifying the role of energy in the manufacturing productivity of Ethiopia can contribute to industrial policy input. The labor input is significant and positive in the scale effects model (Model 2). This means that an increase in labor employment will increase labor productivity due to increasing returns and the labor-intensive nature of the industries [7,30]. Finally, we reported the diagnostic tests for serial correlation and heteroscedasticity. The AR (2) test validated the model, free from the serial correlation problem. The number of instruments used were less than the groups in both the dynamic panel estimation approaches.

#### **6. Conclusions and Policy Implications**

This study investigated the effect of energy on manufacturing labor productivity in Ethiopia using panel data for manufacturing industry groups. Fifteen industries were included in the study covering 12 years of data from 2005 to 2016. The number of industry groups and the period was determined by data availability. Data were obtained from the Central Statistical Authority (CSA) in Ethiopia. We used both descriptive and econometric approaches for examining the empirical relationships between the

variables of interest conditional on some other variables and characteristics. This study had two specific objectives: examining the existence of an empirical relationship between energy and labor productivity in the manufacturing industry and estimating the elasticity effect of energy on labor productivity.

Three models were estimated. The first model is a conventional production function with labor, capital, and energy as the explanatory variables along with a time trend to proxy for capturing technological change. The second model measures the scale effect of energy with the control variables labor, capital, and technology. The third model measures the intensity effect of energy and capital on labor productivity in Ethiopian manufacturing industries. Accordingly, static and dynamic panel data models were estimated—pooled OLS, fixed effects, and random effects static panel estimators, along with difference and system GMM dynamic panel models.

The data for industrial group production showed that the overall trends in production were steady and constant over the study period, except for the food and beverage industry, which rapidly increased (industry code 1). On average, the energy use trend increased in the food and beverage industry (industry code 1) as well as the textile industry (industry code 3). The share of production across the 15 industry groups was dominated by the food and beverage industry (industry code 1), followed by the non-metallic mineral products industry (industry code 10). The non-metallic mineral products industry was found to be more energy intensive than the others.

In the first model, the manufacturing production function was estimated with labor, capital, and energy as the inputs in the production process. The time trend was included to capture technological change. In this model, energy, capital, and labor were statically significant and positive in augmenting manufacturing production in Ethiopia; this result is similar to that of other empirical studies [15–17]. Technology was also significant and a positive factor in industrial growth in Ethiopia. In this model, the sample average returns in relation to the scale of production were 1.07, implying increasing returns in relation to the scale of the manufacturing industries. Labor and capital were statistically significant in all the models at the one percent level of significance.

Across the models, some variables had different significance levels, which led us to select an appropriate model that fit the data best. Both static and dynamic model estimation methods were considered, and we got different estimated coefficient results. For the static models, limitations in considering endogeneity, omitted variable bias, autocorrelation, and heteroscedasticity led to the dynamic panel model estimator being selected over the static panel estimator. The system GMM estimator was chosen over the difference GMM model based on the diagnostic tests and to overcome the limitations of missing observations in the difference GMM model.

In all the models, an increase in employment induced labor productivity due to increasing returns to scale and the labor-intensive nature of the industries. Energy positively explains labor productivity in manufacturing industries in Ethiopia. This means an increase in the use of energy-enhanced labor productivity in the industry groups. Capital intensity use gave a boost to labor productivity, which is consistent with theoretical predictions. In addition, a system GMM model was estimated, including time dummies for the scale effect and input intensity models. In both the cases, labor productivity increased over time, signifying the positive effect of technical change on manufacturing labor productivity in Ethiopia. Across the different approaches used, the role of energy use and energy intensity was consistently significant and positive in explaining labor productivity changes in Ethiopian manufacturing industries.

This study showed that energy induces labor productivity in the manufacturing industry groups in Ethiopia, showing that the efficient use of energy increases industrial growth. It also empirically identified labor and capital as essential determinant factors of productivity in the manufacturing industries in Ethiopia, complemented by technological change effects. This indicates a need to organize resources in a way that boosts the growth of the industries. Energy and capital should also be efficiently used, as the results show that productivity is elastic in relation to a change in energy and capital input intensities in the manufacturing industries in Ethiopia.

A review of the existing literature showed that the role of energy in productivity is controversial across countries [10,17,44,57]. This study adds to the literature by empirically validating the positive role of energy in productivity, applying different model specifications and estimation methods to Ethiopia's manufacturing industries. This implies that industrial policies in Ethiopia should focus on the efficient use of energy along with labor, capital, and technical changes to overcome the premature deindustrialization pattern over time. Research on the energy efficiency and energy productivity of the manufacturing industries in Ethiopia is expected to provide additional policy inputs. This type of research can be extended to cross-country analyses in developing countries, using the manufacturing industry as a case study.

**Author Contributions:** Conceptualization, S.G.K. and A.H.; methodology, S.G.K. and A.H.; Data collection, S.G.K.; Formal analysis, S.G.K.; writing—original draft preparation, S.G.K. and A.H.; writing—review and editing, S.G.K. and A.H. All authors have read and agreed to the published version of the manuscript.

**Funding:** This research received no external funding.

**Acknowledgments:** The authors gratefully acknowledge the comments and suggestions by the editor of the journal and three anonymous referees on an earlier version of this manuscript.

**Conflicts of Interest:** The authors declare no conflict of interest.
