Page 250 - Special Topic Session (STS) - Volume 4

STS583 Michael W. et al.
3. Results
One limitation of the results is related to different time periods in which
the data was collected. Whereas the census was conducted in 2011, the
satellite imagery comes from 2015 (built up area) and 2016 (land cover). This
however only renders the results more prudent, since the estimates derived
from or with the support of satellite imagery would subsequently give a more
accurate picture, since it would reflect the correct distribution of the target
population.
5.1. Sampling from a conventional sampling frame stratified, by the
available census variables.
To address the problem of a lack of informativeness in the sampling frame,
we will in a first step add the above described landcover types contained in a
raster image to the sampling frame at the level of the PSU. In this way we can
demonstrate, that already by adding this widely available type of data, we can
substantially improve the estimates. One important perquisite for an
improvement of the estimates through stratification is a sufficiently strong
relationship between the variable(s) of interest and the stratification variable.
In the case of the landcover, we decided to estimate the number of housing
types. The landcover aggregation at the PSU level was done by calculating the
share of crop area (category 4 in Table ..).
To make the most out of the additional information we also used a newly
developed allocation algorithm, which optimizes stratification by
simultaneously creating and allocating strata, such that the overall variance of
the estimate is minimized at a prespecified level for each domain of interest.
This algorithm is implemented in R by using the package sampling strata
(Barcaroli, 2014).
We will first compare a sample drawn in the conventional way from the
census data frame, with a sample making use of the additional stratification.
Results are presented in Table 1 below

239 | I S I W S C 2 0 1 9