Assessing Effective Factors on Poverty Using Bayesian Networks

In this paper, the causes and the way that factors affect poverty are examined using Bayesian networks. Factors are considered include education level, family size, sex, married status and activity status of the head of the household. Applied data are extracted from the urban family’s cost and income survey in statistical center of Iran. The analyses are performed by choosing a random sample of size 500 from these data. From the data, some important variables are chosen. These include an ordinal scale categorized version of the total expenditure of the household (with four categories poor, pseudo-poor, not- poor and pseudo-rich or rich), sex (with two categories male and female), education level (with two categories illiterate and not illiterate), married status (with four categories), activity status (with four categories), kind of accommodation (with three categories) and family size (with six categories). In this paper by applying Bayesian networks’ method, which gives researcher the capability of analyzing the whole variables simultaneously, the most important factors that directly or indirectly affect the poverty of Iranian households are detected. Furthermore, the quality and the quantity of effectiveness of these factors are considered. To clarify the approach we started with an illustrative example. In this example some inferential approaches are also explained. The software HUGIN and its PC and EM algorithm capabilities are used to find the best Bayesian Network. The different approach of Bayesian Network, using the urban family’s cost and income survey of Statistical Center of Iran, is used to analyze the poverty problem. Then, in order to assess the efficient factors on Iranian families’ poverty, the suitable Bayesian Network is found and used in diagnostic analysis. Some reasoning aspects (up-down reasoning and down-up reasoning) are also presented. Some advantages in using Bayesian networks are found. These are: (i) the output is a probability which can be easily interpreted. (ii) Bayesian Networks can be guaranteed to exploit all known features of poverty problem. (iii) they can be used to reason in two different directions (up-down reasoning and down-up reasoning), as mentioned previously. (iv) the utility of the Bayesian networks graph itself is of high importance since the graph is a compact representation of the knowledge surrounding the system. The results of the analysis show that the education level and family size are the most important factors in determining the poverty level of families. In addition, by using ‘down-up’ inference, if we just consider families with poverty status being poor, we found that. ... So, the final result is that for families in absolute poverty level, the most probable cause is the low educational level or the illiterateness of the head of the household.