Sustainable Energy Management in Vietnamese Firms: Evidence from Logit Regression and Random Forest

1. Introduction

Over the last few decades, firms have been adapting their strategies to become more environmentally friendly by minimizing resource use and reducing their environmental impact (Valero-Gil et al., 2017). This is partly due to the increasing pressure that firms face from consumers and to stricter regulatory measures taken by governments (Erdiaw-Kwasi et al., 2024). The literature provides evidence that such measures not only benefit the environment but also confer firms with a comparative advantage and financial gains (Kotkova Striteska et al., 2024; Vučković and Čučković, 2024). For instance, adopting sustainable energy management (SEM) measures reduces operational costs and improves production efficiency (Erdiaw-Kwasi et al., 2024). However, implementing greener measures remains one of the biggest challenges firms face, especially in developing countries (Luthra et al., 2015).

Identifying the factors that drive firms to take sustainability measures is crucial for accelerating the transition to a low-carbon economy. By understanding these drivers, policymakers can design more targeted and effective regulations and incentives to promote sustainable practices. This, in turn, is key to achieving broader climate goals.

Investigating the determinants of greenness measures in Vietnam is especially important, as environmental degradation has accelerated over the last few decades, beginning with the Doi Moi economic reforms in 1986 (Ngo, 2023). Vietnam ranks last among 180 countries in the 2024 Environmental Performance Index, a comprehensive, data-driven ranking that analyzes and scores countries’ environmental policies and sustainability performance worldwide (Block et al., 2024). As such, it is imperative that all actors in the country take sustainability measures at the individual and institutional levels.

Therefore, in this paper, we focus on the energy management aspect of sustainability measures, which include improving existing technologies and procedures or developing new, energy-saving technologies (Asif et al., 2025). Long-standing government policies that indirectly subsidize energy costs through state-owned companies have made Vietnam one of the world’s most energy-intensive economies. This has led to inefficient energy use in businesses’ production processes (Le, 2019). Growth in industry and construction is driving the rapid rise in energy demand in Vietnam (Le, 2019).

Against this backdrop, this paper investigates the determinants of SEM practices of Vietnamese firms by addressing the following research questions: 1) How do firm-level characteristics influence the adoption of SEM practices? 2) How do firms’ innovation capabilities affect SEM decisions? 3) Are firms with external linkages more likely to adopt SEM practices? 4) Are foreign-owned firms more likely to implement SEM measures only if they invest in R&D?

To address these questions, we developed a logit model and a machine-learning-based random forest model.

The remainder of the paper is organized as follows. The next section presents the relevant literature and the theoretical arguments underpinning this analysis. Section 3 describes the data, defines the variables, and explains the methodology used in this study. After that, the empirical results are presented and discussed. The last section concludes with policy recommendations and suggestions for future research.

2. Literature Review, Theoretical Background, and Our Contribution

The independent variables of our model can be classified into three groups: firm characteristics and structural factors, innovation capabilities and knowledge assets, and external linkages. Moreover, we include a control variable and an interaction term. In the following three subsections, we present past empirical findings and the theoretical background for each variable, and we establish the corresponding hypotheses. In the last subsection, we outline our contributions to the literature.

2.1. Firm Characteristics and Structural Factors

This group of variables comprises the firm’s size, the sector in which it operates, its age, and its manager’s sector experience.

Applying sustainable energy management measures is bound to be costly for firms. This makes it less likely that small or medium-sized enterprises take such measures, as they tend to lack the necessary resources (Aragón-Correa et al., 2008). Larger firms also tend to accumulate more knowledge (Nguyen & Vu, 2024). Empirically, Uhlaner et al. (2014) and Veselova & Sidorenko (2022) found a positive association between firm size and environmental management practices. Similarly, Horbach et al. (2023) find a positive relationship between firm size and greenness. Interestingly, Erdiaw-Kwasie et al. (2024) find that firm size negatively affects sustainable energy practices.

The sector in which a firm operates can also be a determinant of whether it takes sustainability measures. According to Demirel et al. (2019), stringent regulations serve as a key driver of innovation, pushing industries—especially those in heavily regulated sectors—to develop sustainable practices and green business models. Truong (2024) states that manufacturing companies are primarily to blame for Vietnam’s severe environmental issues. Thus, environmental management standards certificates became compulsory in Vietnam for polluting industries in the manufacturing sector (Fadly, 2020). Moreover, manufacturing firms may be more willing to adopt sustainable energy management measures because they are more heavily affected by high energy prices than other sectors (Asif et al., 2025). Indeed, Kotkova Striteska et al. (2024) find that firms in the manufacturing sector are more likely to implement energy consumption management measures. Similarly, Vučković and Čučković (2024) find that manufacturing firms are more likely than service firms to implement capital-intensive green measures.

Hypothesis 1:

Large firms and firms operating in the manufacturing sector are more likely to adopt energy management measures to reduce emissions, waste, or pollution.

The effect of a firm’s age—or, in other words, its experience—on its likelihood of taking sustainable energy measures may go either way. On the one hand, over time, firms are more likely to accumulate the capabilities and resources needed to take such decisions (Horbach et al., 2023). On the other hand, it may lead to structural inertia, that is, to a strong tendency for firms to resist change and maintain their current structure (Hannan and Freeman, 1984). There is also no empirical consensus. While Horbach et al. (2023) find a positive effect of firms’ age on their greenness, Vučković and Čučković (2024) find a negative relationship between firms’ age and the adoption of green measures. Others, such as Hoogendorn et al. (2015) or Younis & Sundarakani (2020) fail to find a significant relationship between firms’ age and environmental and green supply practices.

Hambrick and Mason’s (1984) Upper Echelons Theory posits that a part of an organization’s results and decisions can be attributed to the prior experience of its managers. Their abilities and experience in the sector influence their firms’ corporate social responsibility actions (Cortes et al., 2025). More experience may allow managers to build stronger industry-specific knowledge and to access the information, resources, and partnerships needed to implement sustainable energy management measures. On the other hand, Hambrick and Fukutomi (1991) argue that long-serving managers may resist change due to an increasing dedication to their original strategic vision. Indeed, Datta et al. (2003) find a negative correlation between CEOs’ tenure and their openness to change. Older managers are generally more risk-averse and thus less likely to take measures with more uncertain outcomes (Barker & Mueller, 2002). The above arguments suggest that the relationship between managers` experience and their likelihood of taking sustainable energy management decisions may not be linear.

Hypothesis 2:

There is a non-linear relationship between the manager’s experience in the sector and the likelihood of the firm taking SEM decisions.

2.2. Innovation Capabilities and Knowledge Assets

This group of variables includes whether the firm invests in R&D, whether it has a foreign technology license, and whether it holds an international quality certification.

According to the dynamic capabilities theory by Teece (2007), a firm maintains a competitive advantage by developing three key abilities: first, to sense shifts in the market; second, to act on those shifts; and third, to continuously reorganize its resources and strengths to adapt. Developing these capabilities may also help firms to be leaders in sustainability measures. Investment in R&D is part of the learning process that forms the micro-foundations of dynamic capabilities (Teece, 2007).

Cohen and Levinthal (1990) posit that a firm’s innovative potential is driven by its absorptive capacity, which is the capability to recognize valuable external information, assimilate it, and exploit it for commercial applications. The firm’s prior related knowledge largely determines this capacity. This potential can manifest itself in areas such as sustainability and energy management. Cohen and Levinthal (1990) show that firms’ R&D expenditures contribute to this absorptive capacity.

Based on these two theories, we could expect an indirect positive effect of R&D expenditure on sustainability and greenness measures. Indeed, Kotkova Striteska et al. (2024) find a positive impact of firms’ R&D expenditure on energy consumption monitoring and proactive environmental strategies.

Internationally recognized quality management standards such as ISO 9000 and ISO 14000 facilitate the implementation of environmental measures by providing the necessary information (Horbach et al., 2023). This should be more relevant for the implementation of more capital-intensive greening measures (Vučković and Čučković, 2024). Being in possession of such internationally recognized certificates was found to be positively correlated with green practices (Asif et al., 2025; Horbach et al., 2023).

Using better technologies facilitates the implementation of energy management systems (Asif et al., 2025). In developing countries such as Vietnam, access to foreign technology enables the use of better and more efficient tools (Vučković and Čučković, 2024). The use of such technology, in turn, increases the odds of implementing environmental measures (Vučković and Čučković, 2024).

Considering the arguments and findings above, we formulate Hypothesis 3.

Hypothesis 3:

Firms that invest in R&D, use foreign licensed technologies, and hold internationally recognized quality certificates are more likely to implement sustainable energy management measures.

2.3. External Linkages

The last group of variables captures external pressures and influences: whether a foreign private entity owns the firm and whether it exports.

Due to financial constraints, local firms often struggle to obtain access to environmental technologies (Christmann and Taylor, 2001). Foreign ownership may facilitate access to better technologies (Vučković and Čučković, 2024). Due to greater concern for their reputation, foreign firms typically adhere to more responsible business conduct (Hanousek et al., 2019). Thus, it is to be expected that foreign-owned firms are more likely to take environmental actions. This conjecture is supported by empirical evidence. Arora (2023) finds a positive link between foreign ownership and sustainability measures, while Veselova and Sidorenko (2022) and Vučković and Čučković (2024) found a connection between foreign ownership and engagement in environmental practices and greenness measures, respectively.

In developing countries like Vietnam, environmental regulations tend to be less stringent, and consumer awareness is lower (Ngo, 2023). Therefore, firms exporting to more developed countries – with stricter environmental regulations – may have stronger incentives to become greener to gain a competitive advantage (Constantini and Mazzanti, 2012). A firm’s environmental strategy is often based on a cost-benefit analysis that weighs the perceived economic benefits of complying with domestic standards against those of major foreign markets (Christmann and Taylor, 2001). Therefore, firms that export their production to markets with more stringent environmental standards are generally more likely to adopt greener practices. This link is empirically supported by Horbach et al. (2023) and Vučković and Čučković (2024), studies that find that exporters are more likely to implement environmental actions.

Based on the above theoretical arguments and empirical findings, we hypothesize that:

Hypothesis 4:

Foreign-owned and export-oriented firms are more likely to implement sustainable energy management measures.

An interaction term is included in a model to assess whether the effect of one variable on an outcome is moderated by another. In this specific case, we aim to check if the effect of foreign ownership is contingent on whether the enterprise invests in R&D. That is, the possibility of access to better technologies brought about by foreign ownership can only be realized if the firm invests in R&D. Otherwise, the firm might not have the internal capacity to utilize the advanced knowledge.

Hypothesis 5:

The positive effect of private foreign ownership is contingent on the firm investing in R&D.

Firms’ productivity and profitability are adversely impacted by extreme weather conditions and other global challenges (Vučković and Čučković, 2024). As Horbach et al. (2023) find that firms affected by extreme weather events are more likely to implement greenness actions, we also control for the effect of such events in our analysis.

2.4. Our Contribution

Previous studies have examined how various firm characteristics and structural or external factors affect firms’ likelihood of implementing green measures. However, to the best of our knowledge, no other study has examined the relationship between managers’ experience and the likelihood of adopting green measures. Based on the theoretical arguments above, we test whether a non-linear relationship is present between managers’ experience and the probability of taking SEM measures (Hypothesis 2).

Moreover, we investigate whether the potential positive effect of private foreign ownership is contingent on the firm investing in R&D (Hypothesis 5).

Methodologically, to the best of our knowledge, this is the first study to analyze the adoption of SEM in enterprises combining a traditional econometric model with a machine learning algorithm. This approach allows us to uncover nonlinearities and dependencies in the data.

3. Data, Variables, and Methodology

In this section, we first present the source of the survey data we use and provide details on how the survey was conducted. We then describe the dependent and independent variables of our model and, finally, briefly outline the statistical methodology used.

3.1. Data

This study uses the World Bank Enterprise Surveys (WBES)1, a survey conducted with business owners and top managers. They provide information on several topics in the business environment, including infrastructure, performance, corruption, financial accessibility, and innovation. The Vietnamese survey comprises 1,028 observations. WBES employed stratified random sampling across industries, company size (small, medium, and large), and geographic areas (five subnational divisions). The survey was conducted between April and November 2023. Sampling weights were employed to ensure representativeness.

3.2. Dependent and Independent Variables

Table 1: Variables used in the analysis.

The dependent variable and all the independent variables, except for Firm Age and Manager’s Experience, are binary variables.

As the survey was conducted in 2023, to obtain the firm’s age, the establishment year reported by the firm in the survey was subtracted from 2023. The theoretical arguments and empirical evidence presented above do not suggest the direction of the relationship between a firm’s age and the likelihood of taking SEM measures. Since we cannot rule out the possibility of a non-linear relationship, such as a U- or inverse-U-shaped one, the firm’s age was included in the logit model both in level and in quadratic form.

For the sector dummy variables, the omitted category includes the retail, wholesale trade, and services sectors.

To test the hypothesized non-linear relationship between the manager’s experience and the dependent variable (Hypothesis 2), we include the variable in its linear and quadratic levels.

The Foreign Private Ownership dummy variable takes value 1 for firms whose largest ownership type among the three (foreign private, domestic private, state, or government) is foreign private.

Finally, the Exporter variable takes value 1 if the firm exports its products directly or indirectly.

3.3. Methodology

To analyze the determinants of sustainable energy management decisions taken by Vietnamese enterprises, a logit regression model was estimated, an appropriate approach given the nature of the dependent variable: it takes a value of 1 if at least one SEM was taken in the last three years, or 0 if no such measure took place. As explained in the previous section, the independent variables are grouped into three categories (firm characteristics and structural factors, innovation capabilities and knowledge assets, and external linkages), and therefore, the results are presented by variable group. The model also includes a control variable to capture whether the firm was affected by extreme weather, and an interaction variable between foreign ownership and R&D investment (both dichotomous variables).

To complement our analysis, a Random Forest (RF) algorithm was employed (Breiman, 2001). This machine learning algorithm constructs several decision trees during training and, afterward, chooses the class that appears most frequently across the predictions of the individual trees to determine the final classification output.

RF’s main advantage is its ability to capture potential non-linear relationships between the dependent and independent variables, as well as interaction effects between the independent variables, without requiring the researcher to establish them a priori. This is particularly important in our context because parametric models, such as the logit model, rely on pre-specified functional forms. Consequently, RF might uncover information that the logit model cannot detect. Furthermore, the variable importance metrics generated by RF supplement the information obtained from the logit model. Such metrics quantify each variable’s relative contribution to the model’s predictive power, allowing us to identify the most influential factors in our context (Hastie et al., 2009).

The variable importance metric used in our analysis is the mean decrease in accuracy (MDA) index. This index measures a variable’s importance using the model’s built-in test data, the Out-of-Bag (OOB) sample. For each tree in the forest, the method compares the model’s prediction error on the OOB data before and after randomly shuffling the variable values. The variable’s importance is the average of these error increases across all trees, indicating how much the model’s performance relies on that variable (Han et al., 2016).

The logit regression and the random forest algorithm were run in R, while data cleaning and variable creation were performed in Stata.

4. Results and Discussion

Table 2: Descriptive statistics.

Table 2 shows that 17.8% of the sampled Vietnamese enterprises have adopted sustainable energy management decisions in the three years preceding the survey. Small and medium-sized enterprises dominate the sample; only 6.7% of the firms are large.

These results show relatively low levels of innovation capabilities among Vietnamese firms. Only 4.5% of the firms in the sample invested in R&D, and 5.4% held a foreign technology license. In the same vein, Vietnamese enterprises’ foreign linkages are weak. While less than 10% of firms are exporters, the share of private foreign ownership is below 5%.

As mentioned previously, we first run a logit model. Table 3 presents the regression output and the odds ratios for the independent variables.

The results only provide partial evidence in favor of Hypothesis 1. Neither manufacturing firms nor construction companies are more likely to take SEM measures than enterprises in the retail and services sectors. Larger firms are more likely to engage in SEM practices: the odds of a large firm taking sustainable energy measures are 117.7% higher than the odds for a small/medium firm.

The statistical significance of both Manager’s Experience and Manager’s Experience Squared provides evidence for Hypothesis 2. With a negative sign for the linear term and a positive sign for the quadratic term, the results suggest that the negative effect of an additional year of the manager’s experience on the likelihood of adopting SEM measures is smaller the more years of experience s/he has. This suggests a U-shaped relationship: first, the probability of adopting SEM measures decreases with experience, then it reaches a turning point, and finally it goes up. Nonetheless, the increasing part of the curve has little practical value, as fewer than 5% of managers exceed the turning point, estimated at around 29 years. That means that for the vast majority of managers, the relationship between experience and the adoption of SEM measures is negative.

In line with the findings of Hoogendorn et al. (2015) and Younis & Sundarakani (2020), we fail to find a significant relationship between the firm’s age and its likelihood of adopting SEM measures.

Hypothesis 3, which states that firms with stronger innovation capabilities are more likely to implement SEM, is partially supported by our findings. While no link was found between holding a foreign technology license and SEM implementation, investing in R&D and holding an international quality certificate both have a statistically significant positive relationship with SEM. Nonetheless, the association with international quality certification should not be interpreted causally since firms that are already ‘green’ are most likely to seek certifications, as suggested by Horbach et al. (2023).

As for the effect of external linkages, export firms are more likely to engage in SEM. This finding aligns with Horbach et al. (2023) and Vučković and Čučković (2024). On the other hand, contrary to our expectations, foreign-owned enterprises are not more likely to engage in SEM practices.

As discussed in section 3.3, to complement our logit regression analysis, a Random Forest algorithm was employed. We used the Ranger package in R because our dataset was imbalanced, with only 17.8% of firms adopting SEM measures. By maximizing the Youden Index –a metric that considers how well the model classifies firms as SEM-measures adopters or not adopters– the ranger package identified an optimal threshold of 0.27 (instead of using the default 0.5 threshold, which is biased towards the majority class), meaning that the model classifies a firm as adopting SEM measures if the predicted probability is 27% or higher. This optimal threshold improves the model’s ability to correctly identify firms that adopt SEM measures.

Figure 1 shows the relative influence of each variable on the model’s predictive accuracy. The RF model identifies holding an International Quality Certification as the most influential factor in the adoption of SEM measures, which is consistent with the highly significant relationship found by the logit model. This consistency highlights the importance of established quality standards in promoting SEM adoption. Moreover, RF assigns high importance to the manager’s experience, consistent with the logit model results. Its high importance in the RF model might be due to its nonlinear relationship with other variables.

Nonetheless, there are some discrepancies between the RF and the logit model findings, which can be due to the RF’s ability to capture nonlinearities in the data. A key difference is RF’s high importance assigned to the Manufacturing Sector variable, which was not statistically significant in the logit model. This suggests that the relationship between operating in the manufacturing sector and SEM adoption is not linear but conditional and interactive.

The fourth most important variable, according to the RF model, is foreign private ownership. In the logit model, it did not have a direct effect on SEM adoption, but it did when interacted with the R&D spending variable. This supports the idea that the relationship between foreign private ownership and SEM adoption is complex and interactive with other variables, a fact better captured by the RF.

While the size of the enterprise was only significant at the 10% level in the logit model, RF ranks it among the five most relevant variables. Again, this might be due to its effect being nonlinear and conditional on other determinants. Regarding the firm’s age, the RF ranked it sixth in importance, ahead of R&D spending and exporting, but it was not statistically significant in the logit model. One possible explanation is, again, its potential interaction with other variables. Finally, the variable R&D spending ranks seventh in importance. Including it, the cumulative importance reaches 94%, as shown in Table 4 below. We can consider the remaining variables to have low weight in determining whether an enterprise takes SEM measures.

Table 5 below shows the most important performance metrics we can use to compare the models.

The comparison of performance metrics shows that RF is more accurate than the logit model. While the logit model correctly identifies 89.5% of the firms that do not adopt SEM measures (Specificity), the Sensitivity (% of correctly identified enterprises that adopt these measures) is as low as 47.8%. This means that more than half of the firms that take SEM measures are wrongly classified. The random forest performance is much better; it correctly classifies adopters in 86.1% of cases while maintaining a very high Specificity (84.6%). The AUC (area under the curve) shows that RF’s overall ability to discriminate between the two groups is higher than that of the logit model (0.927 vs. 0.786). There is also a clear difference in the F1-score (the harmonic mean of precision and sensitivity) in favor of the RF.

Taken together, the results from the logit regression and the random forest model offer complementary insights into the determinants of SEM adoption among Vietnamese firms. While the logit model identifies statistically significant relationships between the likelihood of taking SEM measures and the explanatory variables R&D investment, international quality certification, and exporting, the random forest model highlights other nonlinear relationships. Both models underline the importance of international quality certification, managerial experience, and R&D spending. However, the random forest model reveals additional important predictors, including manufacturing sector affiliation, firm size, and foreign ownership, which were either marginally significant or insignificant in the logit model. This suggests that these factors may exert influence in conditional or non-linear ways not captured by traditional regressions. Moreover, the statistically significant interaction between foreign ownership and R&D investment found in the logit model is emphasized by the random forest’s ranking of foreign ownership as a key predictor. Overall, the machine learning approach enhances the robustness of the analysis and improves predictive accuracy, thereby offering a richer understanding of the drivers behind SEM decisions.

5. Concluding Remarks

This study provides empirical evidence on the determinants of SEM adoption among Vietnamese firms, using both a logit regression model and an RF algorithm. The findings reveal interactions between the internal capabilities, managerial characteristics, and external pressures.

The logit model shows that firm-level factors, including size, R&D investment, international quality certification, and export orientation, significantly influence the likelihood of SEM adoption. Managerial experience exhibits a negative non-linear relationship with SEM adoption. Notably, foreign ownership alone does not predict SEM adoption unless coupled with R&D investment, highlighting the importance of internal absorptive capacity.

The RF model complements and extends the regression analysis by uncovering nonlinear relationships and interaction effects. It identifies international certification, managerial experience, and manufacturing-sector affiliation as key predictors, even though these variables are not statistically significant in the logit model. This difference underscores the importance of using machine learning techniques alongside regression models to capture conditional and nonlinear effects.

From a policy perspective, the results suggest that support for R&D activities and international certification processes could significantly enhance SEM adoption. As other firm-level characteristics also play an important role, it is essential to design such policies in a targeted manner. For instance, initiatives could include providing grants or tax incentives for R&D investments, particularly for small and medium-sized enterprises. Moreover, the Vietnamese government could subsidize the ISO certification process to encourage firms to adopt international standards. Since foreign ownership per se does not guarantee SEM adoption, to ensure that technological spillovers result in sustainable practices, foreign investment should be combined with programs that enhance local absorptive capacity. Moreover, capacity-building programs focused on managerial development could help overcome resistance to change and promote more sustainable practices.

A manager’s experience is the sum of diverse factors, such as the diversity of sectors and firms worked at, macroeconomic conditions, and financial and other firm-level characteristics, among others. Unfortunately, our dataset does not allow us to move beyond a measure of years. Future research could delve further into the effect of managerial experience diversity on the adoption of green measures. Future research could also explore the specific mechanisms behind the non-linear and interactive relationships.