Risk Factors Affecting Poverty in Nepal: Statistical Modeling Approach

Abstract

Poverty is one of the main problems of developing countries, like Nepal and its reduction is a central issue. The identification of its determinants to reduce the monetary poverty is one of the key issues. According to previous studies, log-binomial regression model (LBRM) is a good option to logistic regression model (LRM) for common outcomes, mostly used in the analysis of clinical and epidemiological data. However, the use of LBRM and the comparison with LRM for data on poverty has not been discussed yet. The objectives of this study are to identify the important risk factors, to compare the LRM and LBRM in identifying the risk factors and estimating their effects on poverty in Nepal, and to assess the stability of the model through bootstrapping method. The data used for the analysis is the cross-sectional household level data (n = 5988) of Nepal Living Standard Survey 2010/11. All the data required for this study are not available in the provided household level data file of 5,988 households but are available in the individual level data file of 28,670 individuals. The individual level data are converted into household level data in order to generate the data on a number of variables, and merged into the main data file. With the support of rigorous review of literature and the availability of the variables in the dataset, seven possible independent variables have been considered for both the LRM and LBRM. They are: sex of household head (female / male), literacy status of household head (illiterate / literate), status of remittance recipient of household (no / yes), status of land ownership (no / yes), household with access to nearest market center (poor / better), number of children under 15 years (more than two / at most two), and number of literate members of working age population (WAP) (none / at least one). The response variable is household poverty (poor / non-poor). Implementing the stepwise forward and backward selection procedure with all these seven variables for the development of each final multiple regression model, only six variables except sex of household head has come out statistically significant at 5% level of significance. The LRM has yielded the odds ratio (OR) and LBRM has yielded risk ratio (RR) with 95% confidence interval estimate (CIE) for each covariate. Diagnostics of the model, the goodness of fit test, a risk assessment based on the presence of variables, and the stability of each model has been carried out. The classification and discrimination of the LRM has been also assessed. LRM and LBRM have been compared with respect to different criteria such as selection of covariates, effect size and its precision. The model's good fit test using and test of model's diagnostics criteria has also been compared. Further, the comparisons have also been made in risk assessment on the bais of factors present in the model, stability of the model and convergence failure problem. The effect size in terms of OR and in RR of six factors in each final model namely illiterate household head (OR: 2.20, 95% CIE: 1.86 – 2.61, p < 0.001; RR: 1.68, 95% CIE: 1.49 – 1.89, p < 0.001), remittance non recipient household (OR: 1.90, 95% CIE: 1.64 – 2.20, p < 0.001; RR: 1.45, 95% CIE: 1.33 – 1.59, p < 0.001), household with no land holdings (OR: 1.53, 95% CIE: 1.31 – 1.78, p < 0.001; RR: 1.22, 95% CIE: 1.11 – 1.34, p < 0.001), household with poor access to market center (OR: 1.77, 95% CIE: 1.52 – 2.07, p < 0.001; RR: 1.51, 95% CIE: 1.34 – 1.69, p < 0.001), household having > 2 children aged under 15 (OR: 4.69, 95% CIE: 4.06 – 5.42, p < 0.001; RR: 2.96, 95% CIE: 2.66 – 3.28, p < 0.001) and household not having literate members of WAP (OR: 1.29, 95% CIE: 1.07 – 1.56, p < 0.001; RR: 1.16, 95% CIE: 1.05 – 1.29, p < 0.001) are significantly associated with the likelihood of poverty. For each covariate, the OR is overestimated than that of RR. There is narrower 95% CIE of RR than that of OR for each covariate. It shows that RR is more precise than OR. Greater elevation in risk in LRM compared to LBRM varies from 13% to 173%. In each model, there is no convergence issues have been countered, where both the models are equally stable as assessed by bootstrapping procedure. Almost all variables are repeated 100% times among 1000 times repetition. The visual assessments of diagnostics of each model are reasonably satisfactory. There is considerable acceptable discrimination of LRM (AUC: 0.78) and model correct classification values of 67.15%. The good fit of the model is satisfied by LRM [ with 8 d.f.= 6.05, p = 0.53] but not satisfied by LBRM [ with 8 d.f.= 28.60, p = 0.0004]. Since the LRM satisfied the majority of requirements of model performance instead of some limitations, this model seems to be better than the LBRM for this data set. Nevertheless, the LBRM is an option for the LRM since it has better accuracy and avoids overestimating effect size. The findings of this study are expected to be useful for researchers and policy makers in the relevant field.