摘要：Since 2000 census, the American Community Survey (ACS) publishes poverty rate data based on five-year estimates only. We look at poverty rate estimation in two stages. In part 1, a situation where 5% of the poverty rate data is purposely missing from census tracts is simulated. Several interpolation methods were tried in GIS including Empirical Bayesian Kriging (EBK) and local polynomial interpolation (LPI). It is seen that using the EBK method a mean absolute percent error (MAPE) of 4.1% in the estimation process can be achieved as validated by the 2007-11 five year interval estimates of ACS poverty data. In part 2, the census tract poverty rates from 2000 as well as the ACS five year interval estimates from 2005-09, 2006-10 and 2007-11 were processed by first devising a procedure for unifying the underlying variable census tract geography. Then, poverty data for the time periods were used to create three dimensional poverty rate surfaces using the EBK method. Geographically Weighted Regression method enabled validation of the prediction process with a very low MAPE of 1.5% in comparison to the predicted poverty surface, followed by prediction of poverty rates across census tracts for a “future” period in time.
