Transcript Document

Stata in Space:
An example for the econometric
analysis of spatially explicit raster data
--- Daniel Müller --Institute of Agricultural Economics
and Social Sciences
Humboldt University Berlin
Berlin -- August, 12th, 2003
Outline
1. Introduction
2. Spatial data analysis
3. Data preparation
4. The empirical example
5. Econometric estimation
6. Export of results and geovisualization
Introduction
- Socioeconomic data usually exist for (discrete)
social entities, rarely explicitly linked to location
(georeferenced)
- ‘Natural’ data: often continuous (rainfall, slope,
elevation) and georeferenced
- Integration of both data sources can provide
additional insights
- Allows to understand spatial patterns & processes
- Knowing the where can help us infer the why
Spatial data analysis
- Spatial analysis is the analysis of data linked to
location (spatial data)
- Why analysis of spatial data ?
- Variables of interest vary in space
- Location matters!
- Spatial analysis can provide important insights:
- geographical targeting of investments
- diffusion of technologies
- causes and consequences of land-use change
Spatial data analysis
What’s special about spatial data ?
=> Location matters !!!
=> Tobler’s 1st law of geography (1979):
“Everything is related to everything else, but near
things are more related than distant things.”
=> Spatial effects:
- spatial autocorrelation
- spatial heterogeneity
Spatial data analysis
Peculiarities in space: Spatial effects
1. Spatial autocorrelation
- Coincidence of value similarity with locational similarity
- Second dimension adds mathematical complexity
(multiple directions)
2. Spatial heterogeneity
- Each location is unique
- Units of observations not homogeneous across space
- Structural instability over space, e.g. heteroskedasticity
Spatial data analysis
Peculiarities in space: spatial effects [2]
Spatial effects due to:
-
interactions among neighboring agents
data from different sources
different sample designs
varying aggregation rules
“Spatial relationships among observations can result
in unreliable estimates and misguided statistical
inference of the parameters.” (Anselin 1988).
=> Corrections necessary
Spatial data analysis
Geographic Information Systems (GIS):
- Compile, store, manipulate, analyse, visualize
spatial data
- Consist of hardware, software, data and
procedures
- Data models: vector & raster
Spatial data analysis
Raster data model:
- Arrangement of regularly shaped, contiguous cells
- Continuous data layers; fit together edge-to-edge
- Typically consist of square cells
- Each cell represents a location in a raster GIS
- Cells are arranged in layers
- Values of a cell indicate characteristics of that location
- Data is composed of many layers covering the same
geographical area
Spatial data analysis
Raster data model --- file structure:
1
1
2
2
2
3
1
3
2
1
2
3
3
4
2
3
5
5
6
6
6
6
4
4
6
Header:
Contains spatial information!
Spatial data analysis
Raster data model --- land use map:
Spatial data analysis
From data layers to resulting map
data layers
overlays
analyses
output
Data preparation
Importing grids into Stata
ras2dta , files(filelist) [ idcell(varname) nodata(#) dropmiss
xcoord(#) ycoord(#) genxcoor(varname) genycoor(varname)
header(filename) saving(filelist) replace clear ]
→
infile-s grids (filelist) into Stata:
→
-generate-s IDcode for each cell (=observation)
→
reads the information from the header (if present)
→
“
sets missing values to a specified number
→
“
-drop-s unnecessary empty cells
→
“
-generate-s X and Y coordinates
→
“
-save-s the header information in a file
Data preparation
Integration of data layers
1. Import of raster grids (-ras2dta-)
2. Combination of raster layers in Stata (-joinby-,
-merge-) based on spatial identifier (ID-code
of cells)
3. Socioeconomic (survey, census) data can be
joined to grids based on, e.g., administrative
boundaries
Data preparation
Corrections of spatial effects
1. Spatial lag variables with index values for
latitude (Y) and longitude (X)
2. Spatially lagged variables
3. Regular sampling from a grid
=> 1. can be done with -ras2dta-
=> 2. we ignore here
=> 3. is easy in Stata, e.g. with : -spatsam-
Data preparation
spatsam , gap(#) xcoord(varname)
ycoord(varname) [ saving(filename) norestore
noseed replace ]
Basically that‘s:
keep if (xcoord / gap) == int (xcoord / gap) &
(ycoord / gap) == int (ycoord / gap)
Therefore, only every #-th observation in X and
Y direction is kept in the sample.
The empirical example
Land use change in Vietnam
- Land use as an inherently spatial process
- Returns to land use are (spatially) affected by:
- market accessibility (von Thünen)
- land rent (Ricardo)
- Possible factors to consider:
- soil quality, topography, climate, market
locations, population density, technology
- Limited dependent variable problem (-mlogit-)
The empirical example
Data
- Satellite image interpretation:
- land cover => land use (change)
- GIS, maps, point measurements:
- geophysical indicators => topography, soil, climate
- Socioeconomic & policy variables:
- village survey, secondary statistics
=> technology, population, education, market access
- Data integration based on spatial identifier
and (approximated) village areas
Econometric estimation
Observations:
964,000 pixels (50 x 50 m)
Spatial sample: every 5. cell in X & Y direction
Estimation:
35,000 observations
=> Dependent: five land cover classes (1, 2, .., 5)
=> Independent: a) geophysical
b) socioeconomic
c) policy
d) spatial effects
Econometric estimation
1. Estimation of the influence of hypothesized
determinants on land use.
2. What is the probability that a certain pixel falls
into one of the five land-use categories?
=> -mlogit- (reduced form, clustered for villages)
=> -mlogtest, iia-, -fitstat- (Long & Freese)
Then we take the highest predicted probability
as predicted land use.
Export of results
Outputting results from Stata
dta2ras [varlist], xcoord(#) ycoord(#) cellsize(#)
[ header(filename) idcell(varname) nodata(#)
xllcorner(#) yllcorner(#) saving(varlist) replace ]
→ writes header in front of file with the information
from xcoord(#) ycoord(#) cellsize(#) or header();
(optionally) nodata(#) xllcorner(#) yllcorner(#)
→ then the results can be mapped in the GIS
Geovisualization of results
Prediction map
Geovisualization of results
Maximum predicted probabilities
Thank you !
Questions, comments and critique welcome !
____________________________
© Daniel Müller ([email protected])
Institute of Agricultural Economics and Social Sciences
---- Humboldt University Berlin ---Stata ados available for download at:
http://amor.cms.hu-berlin.de/~muelleda