Analyzing the determinants of poverty

Analysis of correlates of poverty can be carried using multivariate income and consumption regressions. In these regressions, the logarithm of consumption or income (possibly divided by the poverty line) is typically used as the left-hand variable. Right-hand explanatory variables span a large array of possible poverty correlates, such as education of different household members, number of income earners, employment characteristics, household composition and size, and geographic location. Special care must be taken when including variables that themselves are likely a function of income or consumption availability, for example access to basic services. The regressions will only return results for the degree of association or correlation and not for causal relationships.

It is important to note at this stage that numerous correlates or determinants of poverty are not quantifiable. For some other variables, one might only be able to use a proxy, which might not fully reflect the underlying dimensions. The method used here is only able to take into account those dimensions which are quantifiable or for which a proxy is available. It is also important to note that the various coefficients obtained from a regression will have different degrees of significance.

These multivariate regressions will estimate the partial correlation coefficient between income or consumption per capita and the included ‘explanatory’ variables while holding all other impacts constant. For example, the results could tell us how strongly an additional year of education for the household head or his spouse is associated with a change in income or consumption per capita while holding gender, employment, age, location, and all other possible influences constant. They can tell us, then, much more than the simple relative poverty risks discussed in the previous section since a high relative poverty risk of a specific population group could in fact be attributable to individual characteristics such as education rather than to a group characteristic.

The table below shows an example of such a regression in Cote d’Ivoire. It indicates that education plays a different role in urban and rural areas (where it does not seem to significantly influence consumption), as do different types of assets. In rural areas, infrastructure has substantial predictive power—households located in villages that are nearer to both paved roads and public markets are better off, as are households living in areas with higher wage levels. The results present further questions about the quality of education in rural areas and to the importance of rural infrastructure in helping families grow out of poverty, that could be addressed in putting together a poverty reduction strategy.

Several variations of these multivariate income regressions can be used to examine the correlates of the income of the poor. Poverty analysis focuses on correlates of income and expenditure at the lower end of the distribution rather than the correlates at the top end. One can then perform different regressions for each quintile, or quartile, of the population. Whether these regressions can be conducted will in part depend on the sample size of the survey. Alternatively, the regression can examine structural differences in parameter estimates for different income or expenditure groups.

When multiple cross-sectional surveys are available, the same regression can be repeated for different years to see how the association of certain correlates with income or consumption varies over time. Variations over time will be reflected in changes in coefficients or parameters. The results of repeated cross-section regressions can also be used to decompose changes in poverty between changes in household characteristics, and changes in the returns to (or impact of) these characteristics (e.g., Wodon 2000). Another possibility is to use parameters from the regression model obtained for year one in order to predict household income or consumption in year two, and to compare this prediction with the prediction obtained using the regression estimates for year two applied to the data for year two. The differences in the predictions with the two models can then be analyzed, and one can test whether changes in income between years is due to changes in structural conditions or changes in the behavior of households between the two years.

Apart from income and consumption regressions, several other types of multivariate regressions can provide additional insights into the determinants of poverty. These can in particular be applied to other dimensions of poverty, such as child nutrition, mortality, morbidity, literacy or other measures of capabilities. The techniques are also sometimes applied to understand the determinants of employment and labor income and to estimate the returns to education. They can also be used to better understand agricultural production patterns by estimating agricultural production functions (which relate production with information on type of crops grown per area, harvest, inputs into agricultural production, and input and output prices).