| dc.description.abstract | The study describes the effect of digitalization of SMEs on the productivity of the labor force in Bangladesh. It seeks to determine the tools which are most significant. According to the information presented by the World Bank in 2022, in its Enterprise Survey, I have applied both the methods of econometrics and machine-learning implementation to get beyond the relationships and find the causality. I have two portions of our empirical approach. Part A uses fixed-effects OLS, interaction models, quantile regressions, and propensity-score matching (PSM) to estimate the impacts of adopting digital on productivity controlled by firm variation and due to observed selection bias. Part B applies machine-learning models and methods of explainability to tree models, including SHAP values and Partial Dependence plots (PDPs), to determine non-linear effects and threshold effects of a predictive, non-causal setting. The effects of digital strength on labor productivity point to it being, according to fixed-effects estimates, a positive relationship. Regression of the components also indicates that this growth is mainly due to the transactional tools, especially online payment, which is not due to the presence of online entities, such as just having a website. The interactive model shows that companies that have more experienced managers get increased gains, and this outlines the importance of managerial capacity. The quantile regressions determine high amounts of heterogeneity; the productivity increases are non-significant in low-productivity firms and significantly higher in high-productivity firms. To estimate the causal impact of the use of online payment, we use PSM in which we take it as a discrete treatment. The Average Treatment Effect on the Treated (ATT) has the closest approximation of adding 0.40 log points to the labor productivity of the adopters as compared with non-adopters of the same kind. In a range of corresponding approaches, clustered bootstrap tests and doubly robust estimators, this observation is so. The sensitivity tests show that it is resistant to moderate unknown confounding. It was also tried using an instrument-variable design based upon the diffusion of regions, giving a weak first-stage relevant test, and is therefore hard to interpret causally. The results of the machine-learning algorithms are added to the econometric data. Out-of-sample predictive models are acting in a good manner, and SHAP and PDP analysis show good evidence that the non-linearity exists: productivity growth is more rapid when not less than two digital tools are implemented in the company. Overall, we have identified a capacity-based digital dividend - an improvement of productivity in the presence of transactional digitalization and the capacity of managers, and not the online nature of the firms. | en_US |
