Title: Temporal Changes in the Geographical Pattern of Infant Deaths, Des Moines, Iowa 198994
1Temporal Changes in the Geographical Pattern of
Infant Deaths, Des Moines, Iowa 89-94
Aniruddha Rudy Banerjee Department of
Geography, The University of Iowa
2purpose of this study
- To determine significant changes in infant
mortality patterns through time
3Study Area
Des Moines urban area, Iowa
17miles E-W 11miles N-S
N
4first time period
Location of infant births deaths in Des Moines,
Iowa, 1989-92
5second time period
Location of infant births deaths in Des Moines,
Iowa, 1993-94
6Motivation
- Often, as geographers, we study whether the rate
of occurrence of some event changes from one
period to another - These rates may be the relationship between
- infant deaths and infant births
- fire damage of homes in an urban area
- cancer deaths and the population at risk
7Motivation
- These changing relationships are usually mapped
to identify patterns - However
- When mapping these changing patterns, can we
infer true rates from observed rates? - Problem overlooked by geographers
8Goal
- Is there a method that determines change in true
rate patterns from change in observed rates?
9Possible solutions
- One way to observe changing relationship between
patterns of events is to map their location in
space and time - i.e. map the observed rates
10Calculating rates
Location of infant births deaths in Des Moines,
Iowa, 1989-92
0 deaths! 22 births or 0 imr
29 deaths 603 births or 31 imr
4 deaths, 247 births or 17 imr
11How to Map these rates?
- Rates being a continuous variable we may use a
powerful cartographic technique - Isopleth Mapping
- Isopleth mapping converts rates observed at
sample locations to a continuous spatial
distribution using spatial interpolation
techniques - Square grid (0.4 miles) vertices are used for
sample locations
12Calculating rates from a spatial grid
Location of grid points (0.4miles apart)
13Selecting spatial filters
Spatial filters (0.8 miles, 1.2 miles and 1.6
miles)
Adjacent local estimates share information
because of overlaps
14isopleths time period 1
Infant death rates (0.8 miles) in Des Moines,
Iowa, 1989-92
15isopleths time period 2
Infant death rates (0.8miles) in Des Moines,
Iowa, 1993-94
16Any Regional Change?
Statistically, the region as a whole experienced
a significant decline in IMR (drop from 9.2 to
7.9) A general test was done using the following
z-test statistic
8.56
Z-critical is 1.96 at p lt 0.5
where, r1 rate in first period r2 rate in
second period l total number of live births
17detect any change within?
Infant death rates in Des Moines, Iowa, 1989-92
1993-94
18Map the Change
Difference in IMR (93-94) - (89-92)
increase() or decrease(-)
19Problems?
- Yes!
- When mapping changes in these patterns, can we
determine real changes from observed changes? - Unusually high or low rates in local areas may
result from the small number problem
20small number problem
1989-92
1993-94
A large change in imr (from 250 to 0!) with 1
fewer death
0.4 mi
1 deaths,4 births, 250 imr!
0.8 mi
2 deaths,204 births, 9.8 imr!
0 deaths,4 births, 0 imr!
0 deaths,93 births, 0 imr!
A small change in imr (from 9.8 to 0!) with 1
fewer death
21small number problem
1989-92
1993-94
A small change in imr (from 9.5 to 7.5!) with 1
fewer death
4 deaths,419 births, 9.5 imr!
3 deaths,402 births, 7.5 imr!
22variability and the small number problem
Expected change between two time periods i.e. 9.2
- 7.9 1.3 imr
Small numbers
large numbers
1.3
1.3
Large variations
Small variations
23What do we do?
We would like to map real changes in rates
instead of observed changes in rates
24Past research
- Examples
- Choynowski 1959, Berry 1960s -- Probability maps
- Spatial point processes -- Cluster Statistics -
1980s-90s - Bayesian Adjustments Maps-- 1980s-90s
- Monte Carlo Simulation -- 1990s
25Past research
- Few studies done on temporal change of geographic
pattern
26Present research
- To determine real changes in geographic pattern
- We follow a stochastic model that has the form
- ? imrt1 - imrt ?
- To model the error-distribution D(?) a Monte
Carlo simulation is used - the Monte Carlo simulation generates a null
distribution from an expected hypothesis for any
area without the use of parametric distributional
assumptions
27Monte Carlo Simulation vs theoretical
distributions
- In a Monte Carlo simulation
- i.i.d assumptions are not required
- each filter at any grid point generates its own
distribution by appropriate transformation of the
expected or null hypothesis for non-i.i.d
variables, the Jacobian (determinant vector) of
the transformation is theoretically intractable
28simulations from chance the null hypothesis
- To Simulate rate changes a null hypothesis may be
stated - each birth, in a given time period, has the same
chance of dying - in other words what expected changes would occur
if in each period each infant birth had the same
risk of becoming a death?
29Null Hypothesis Simulation
- Time Period 2
- there were 73 deaths and 9241 births
- 73 deaths are selected randomly from the
locations of the 9241 births - made 500 maps
- Time Period 1
- there were 178 deaths and 19,348 births
- 178 deaths are selected randomly from the
locations of the 19,348 births - made 500 maps
30generating simulated differences
- We have
- Each time period has 500 simulated maps
- Now, simulation 1 for time period 1 may differ
with all the 500 simulations in time period 2 - simulation 2 for time period 1 may differ with
all the 500 simulations in time period 2 - . And so on for 500 times
31generating simulated differences
- We have for any area 500500 or 250,000 simulated
(or expected) differences - We have 1084 grid points with observed
differences this makes 250,0001084 or
approximately a quarter billion differences - All these quarter billion calculations are done
by a computer program written in an object
oriented language (C) and uses 30 minutes on a
100 MHz processor speed
32observed versus expectedapplying (constrained)
Monte Carlo simulations
- We compare the observed difference with the
differences generated from the null - In doing so, we measure how many times the
simulated differences for any area (here, spatial
filter) are lower than the observed difference
in IMR
33example of simulated differences
Figure 1 Simulated differences of infant
mortality rates (period 1 period 2) and the
observed difference at one grid point.
34Significance of IMR differences in Des Moines,
Iowa 1989-94
35Significance of IMR differences in Des Moines,
Iowa 1989-94
36conclusions
- In this case, several small areas of Des Moines,
Iowa were shown to have significant declines in
infant mortality rates between 1989-92 1993-94.
No areas showed significant increases in imr. - We have developed and tested a model that
successfully separates change in disease rates in
small areas due to chance from changes that are
NOT due to chance and therefore MUST be real - By using spatially filtered data (as in this case
study), this model can be implemented at any
chosen geographical scale
37future work
- Redefining the null hypothesis
- by redefining the null hypothesis for every local
area a more powerful map can be produced that
estimates the risk of change in infant mortality - to achieve this a bootstrapping method is
applied to achieve unique local acceptable null
hypotheses - bootstrapping is an approach designed to
iteratively reach an acceptable null estimate
38future work
- introduce independent variables to estimate
causal relationship in studying change
f(xi )
? imrt1 - imrt ?
- Where f(xi ) are functions of independent
variables (i 1, 2, 3 ) affecting infant
mortality change, e.g. - Health policy, welfare of neighborhood, education
of mother etc.
39We shall not cease from exploration And the end
of all our exploring Will be to arrive where we
started And know the place for the first
time. (T. S. Eliot)
40References
Chowynowski M. 1959. Maps based upon
probabilities. Journal of American Statistical
Association, 54385-388. Grimson, R.C. and Oden,
N. 1996. Disease clusters in structured
environments. Statistics in Medicine,
15851-871. Langford I., 1994. Using empirical
Bayes estimates in the geographical analysis of
disease mapping. Area 26142-149. Nagarwalla, N.
1996. A Scan Statistic with a variable window.
Statistics in Medicine, 15845-850. Rushton G.
and Lolonis P., 1996. Exploratory spatial
analysis of birth defect rates in an urban
population. Statistics in Medicine, 15717-726.
41This paper was presented at the Annual Meetings
of The Association of American Geographers,
Boston, Massachusetts, March 1998
42Acknowledgements
Dr. Gerard Rushton, Professor, Department of
Geography, University of Iowa Diane S.
Krishnamurti, M.P.H., Birth Defects Registry,
University of Iowa
43for further information contact
Aniruddha Rudy Banerjee (email
aniruddha-banerjee_at_uiowa.edu) for information on
software to run the models used in this research.
The programs are written in the object oriented
C programming language and run on Windows95?
Windows NT? operating systems
44End