Publicação
Multidimensional poverty in Benin : evidence from classic and machine learning analysis
| Resumo: | This dissertation provides new insights on immediate factors affecting multidimensional poverty in Benin. Ordered probit and fractional probit models are compared to the random forest model, and poverty-targeting indicators are derived for the country, using 2018/2019 individual-level cross-sectional data. In most cases, the effects of regressors on the response variable have the same direction of impact in both glass box and black box models, whereas accumulated local effects plots on random forest suggest a highly nonlinear relationship between individual’s welfare condition and the age of household head and inequality, as well as a nonlinear but non-concave relationship with household size and child dependency ratio. While all models corroborate suggesting that education, agroecological zones, financial access, household size, and employment sector are among most important variables associated with welfare condition, only the black box model, through SHAP values, ranked variables with highly nonlinear effects among the most important regressors, as well child dependency ratio. Moreover, the random forest model, by computing more complex interactions between variables, was able to present a broader range of important variables in the top 15. In general, my findings are consistent with most literature on poverty in Africa and Benin, with all models indicating that education is the most important "proximate" determinant of the welfare condition in Benin. The most important poverty-targeting indicators are household size, food diversification, household head without education, households that gather wood for home cooking, and child dependency ratio. |
|---|---|
| Autores principais: | Barbosa, Lágida Kórcia Almeida Coimbra Monteiro |
| Assunto: | Multidimensional Poverty Ordered Probit Fractional Probit Random Forest Explainable Model Techniques Pobreza Multidimensional Probit Ordenado Probit Fracionado Floresta Aleatória Interpretação de Modelos Black Box. |
| Ano: | 2023 |
| País: | Portugal |
| Tipo de documento: | dissertação de mestrado |
| Tipo de acesso: | acesso aberto |
| Instituição associada: | Universidade de Lisboa |
| Idioma: | inglês |
| Origem: | Repositório da Universidade de Lisboa |
| Resumo: | This dissertation provides new insights on immediate factors affecting multidimensional poverty in Benin. Ordered probit and fractional probit models are compared to the random forest model, and poverty-targeting indicators are derived for the country, using 2018/2019 individual-level cross-sectional data. In most cases, the effects of regressors on the response variable have the same direction of impact in both glass box and black box models, whereas accumulated local effects plots on random forest suggest a highly nonlinear relationship between individual’s welfare condition and the age of household head and inequality, as well as a nonlinear but non-concave relationship with household size and child dependency ratio. While all models corroborate suggesting that education, agroecological zones, financial access, household size, and employment sector are among most important variables associated with welfare condition, only the black box model, through SHAP values, ranked variables with highly nonlinear effects among the most important regressors, as well child dependency ratio. Moreover, the random forest model, by computing more complex interactions between variables, was able to present a broader range of important variables in the top 15. In general, my findings are consistent with most literature on poverty in Africa and Benin, with all models indicating that education is the most important "proximate" determinant of the welfare condition in Benin. The most important poverty-targeting indicators are household size, food diversification, household head without education, households that gather wood for home cooking, and child dependency ratio. |
|---|