You are on page 1of 13

PROBLEMS IN SPATIAL DATA ANALYSIS

(3RD PRESENTATION)

Presented by:
TauFiQ (10-8705-601-8)

Master Degree of Civil and Environment Engineering


山口 大学  2010
Objective:

To understand ways in which properties of spatial


data raise important issues for the statistical
analysis of such data.

Problems in Spatial Data


Analysis
Discussion Materials :
2.3.1 Conceptual Models & Inference Frameworks for Spatial Data

2.3.2 Modelling Spatial Variation (Next Meeting)

Problems in
Spatial
Spatial
Data
Data
Analysis
Analysis
2.3 Problems in Spatial Data Analysis
2.3.1 Conceptual models and inference frameworks for
spatial data

The classical inference model assumes that data are the


outcome of some well defined experiment. This experiment can
be replicated as many times as necessary.

In this classical context, Inference with spatial data, is concerned


with making statements and assessing the evidence regarding
properties of the underlying process responsible for the data.

Problems in
Spatial
Spatial
Data
Data
Analysis
Analysis
The hypothetical universe of realisation or super-population is
often viewed with considerable skepticism as a model for
real spatial data analysis.

It is sometimes argued that the classical perspective is


justifiable when the processes responsible for the
observations appear to contain inherent random components
and where other instances of the same process can be
indicated.

Problems in
Spatial
Spatial
Data
Data
Analysis
Analysis
Another model for spatial data which has some features in
common with the classical inference model is where the data
represent sample observations from a given surface. Henley
(1981) describes that the spatial surface is continuously varying
but fixed, all variation is associated with the actual form of the real
surface.

The model which Henley uses as the model for non parametric
approaches to geo-statistical site interpolation and mapping, reject
assumptions of replication and stationarity. This   is contrasts
with the earlier models for kriging where the realised surface is but
one of many form a super-population of surfaces (Matheron,
1971).

Problems in
Spatial
Spatial
Data
Data
Analysis
Analysis
The difference between these two analogues of the classical
inference model lies in the existence of a universe of possible
surfaces. Difficulties for classical inference arise if we adopt a
deterministic perspective and the data represent an exhaustive
survey.

The classical inference model still seems tenable if:


a. there is random error associated with measurement of surface
properties
b. the surface is sampled
c. the values at the sites are themselves the result of a sample

Problems in
Spatial
Spatial
Data
Data
Analysis
Analysis
Both ways of relating spatial data analysis to the framework provided by
classical statistical inference assume that data arise as if from a
controlled experimental situation. This assumption does carry with it
some implicit assumptions, in particular the axiom of correct
specification. Leamer (1978) specifies the axiom as follows
a. The set of explanatory variables that are thought to determine the
response variable must be: unique, complete, small in number,
observable
b. Other determinations of the response variable must have a probability
distribution with at most a few unknown parameters
c. All unknown parameters must be constant

Problems in
Spatial
Spatial
Data
Data
Analysis
Analysis
If there is no experiment defining or specifying the model that should be fit
to the data, probabilities are not well defined. As a consequence, classical
inference procedures that depend on evaluating probabilities are not on firm
ground.

Leamer defines three groups of data analysts in terms of their response to


this situation: believers (those who report results as if they were the outcome
of a controlled experiment), agnostics (those who do not deal with standard
errors and seek only descriptive summaries of data and discount all results
until tried out on another set of data), and pragmatists (who think that tehe
agnostics go too far and that standard errors shoul be reported)

Leamer (1978) emphasis two other areas of difference with classical


inference procedures based on Fisher & Neyman-Pearson inference:
a. The purpose of statistical inference in the context of control experimentation
is to estimate unknown parameters and test hypothesis
b. Classical judgments are made solely with respect to the current set of data

Problems in
Spatial
Spatial
Data
Data
Analysis
Analysis
The brackets after each description indicate the type of failure of
The axiom of correct specification as follows (Leamer, 1978):
a. Hypothesis testing
Several possible regression models justified by theoretical
considerations and the analyst test to see which one is best
supported by the evidence
b. Interpretive searching
One registration model is chosen to describe the data and the analyst
tries to make the model fit the data better, perhaps by imposing
constraints on the parameters
c. Simplification searches
A regression model chosen under (a) or (b) may be very complicated
and the analyst tries to reduce the complexity of the model whilst
retaining adequacy of fit

Problems in
Spatial
Spatial
Data
Data
Analysis
Analysis
Continue …
d. Proxy searches
A regression model is chosen but it is known that variables can
be measured in many different ways
e. Data selection searches
Fitting a model with different subsets of data or different
transformations of the data and selecting the result that
appears best
f. Post-data model construction
A purely inductive search to try and account as well as possible
for observed variation in a response variable, often involves
searching for additional variable s to improve the level of
explanation

Problems in
Spatial
Spatial
Data
Data
Analysis
Analysis
Diaconis (1985) reviews alternatives to the usual classical
theories of inference in the context of explanatory data
analysis, where the purpose of data analysis on non-
experimental data is to detect structure and pattern.

The choice of inference framework will arise again later and it


is at least arguable that the devising of appropriate inference
frameworks is one of the most fundamental issues in spatial
analysis.

Problems in
Spatial
Spatial
Data
Data
Analysis
Analysis
THANK YOU

You might also like