Introduction – Multiview Ensemble Learning for Poverty Mapping
In this project we are looking to apply Probabilistic Machine Learning Techniques to Poverty Mapping. [1] points out that the current techniques and the precision of these are limited. They argue that these are due to the strong assumptions (especially Area Homogeneity) the estimation models require.
We are trying to apply ML techniques (specifically multi view clustering) to the dataset (described below) and to estimate the poverty index of a particular area.
Current Dataset :
We are working on the datasets used by [1]. It consists 2 data sets of Census level (2000 Mexican Census) data and Survey level data of the muncipios in the Mexican states of Chiapas, Oaxaca and Veracruz. It also has a separate synthetic data set for a 2nd experiment. Refer to [1] for more details.
Related work:
[1] focuses on the short comings of simple Projection Based estimator as well as the ELL estimator [2] techniques which are used to make the final estimation model. They argue that the conditions necessary to match Census and Survey Data may not hold in practice. They conduct 2 experiments – one on synthetic data and the other one on 2000 Mexican Census. Their conclusion from the results are that the assumption of Area Homogeneity (AM) has a significant effect on the model parameters and that these assumptions may severely under estimate the variance of error and thus, potentially lead to incorrect results.
References:
[1] “Using Census and Survey Data to Estimate Poverty and Inequality for Small Areas”, Alessandro Tarozzi and Angus Deaton (2008)
[2]“Micro-level estimation of Poverty and Inequality”, Chris Elbers, Jean O. Lanjouw, and Peter Lanjouw (Econometrica, Vol 71, No.1 (Jan 2003), 355-364)
Recent Comments