How robust are indicator based poverty assessment tools over time? Empirical evidence from Central Sulawesi, Indonesia

Eradicating poverty is one of the most urgent concerns of development policies. Organisations aiming at reducing poverty need simple and stable tools to detect poor households. Using data from Central Sulawesi, Indonesia, this study aims to test first whether two indicators sets for poverty assessment found in 2005 are still capable in predicting absolute poverty and second, if the indicator composition remains robust over time. Data from two household surveys were used: In 2005 we surveyed 264 households in the vicinity of the Lore Lindu National Park in Central Sulawesi to obtain indicators of poverty and to derive the daily per capita consumption expenditures. In total 280 indicators were sampled. Two different multivariate regression models were fit to this data-set. One model (Model 1) included all sampled indicators and the other one (Model 2) contained only easily verifiable indicators as ranked by local staff. Each of the models yielded a different set of 15 indicators that predicted poverty best. In 2007, we conducted an additional survey with the identical questionnaires in the same households. We used the data from 2007 to estimate the poverty status of the households with the indicators derived in 2005. Furthermore, we applied the same regression models again to detect changes in the indicator composition. In Central Sulawesi, almost 20% of the rural population was identified as being very poor in the years 2005 and 2007. Regarding the prediction power of the 2005 indicators we found that the prediction power for 2007 mainly was influenced by the error of over-predicting the poor. When re-estimating the models, the accuracy levels remained similar, but the indicator composition changed.