No Slide Title

Download Report

Transcript No Slide Title

Part 09
Application of Multi-level Models
to Spatial Epidemiology
Term 4, 2005
BIO656--Multilevel Models
1
DATA STRUCTURE & GOALS
• We have geographically indexed dependent
variable and covariates
– Outcome, exposures, demographics, ...
• Want to study the relation between spatiotemporal variation in the dependent variable
and covariates
Term 4, 2005
BIO656--Multilevel Models
2
Term 4, 2005
BIO656--Multilevel Models
3
Term 4, 2005
BIO656--Multilevel Models
4
Term 4, 2005
BIO656--Multilevel Models
5
CRUDE, COUNTY-SPECIFIC RELATIVE RISKS
Rates appear to cluster, with a noticeable
grouping of counties with SMR> 200 in the North
Term 4, 2005
BIO656--Multilevel Models
6
Term 4, 2005
BIO656--Multilevel Models
7
Term 4, 2005
BIO656--Multilevel Models
8
Term 4, 2005
BIO656--Multilevel Models
9
Term 4, 2005
BIO656--Multilevel Models
10
Term 4, 2005
BIO656--Multilevel Models
11
Term 4, 2005
BIO656--Multilevel Models
12
Term 4, 2005
BIO656--Multilevel Models
13
Term 4, 2005
BIO656--Multilevel Models
14
Term 4, 2005
BIO656--Multilevel Models
15
Term 4, 2005
BIO656--Multilevel Models
16
Term 4, 2005
BIO656--Multilevel Models
17
Term 4, 2005
BIO656--Multilevel Models
18
SHRINKAGE
When the population in a region is large
• The statistical uncertainty is relatively small
• High credibility (weight) is given to the direct (MLE)
estimate
• The smoothed rate is close to observed rate
When the population in a region is small
• The statistical uncertainty is relatively large
• Little credibility (weight) is given to the direct (MLE)
estimate
• The smoothed rate is shrunken toward a local
(computed by other nearby regions) or a global target
Term 4, 2005
BIO656--Multilevel Models
19
THE CAR MODEL
Local Smoothing
Crude SMR
Term 4, 2005
Smoothed SMR
BIO656--Multilevel Models
20
Posterior distribution of Relative Risk
for maximum exposure
(Maximum AFF)
Global smoothing
(posterior mean = 3.25)
Term 4, 2005
Local smoothing
(posterior mean = 2.18)
BIO656--Multilevel Models
21
Posterior distribution of Relative Risk
for average exposure
Global smoothing
(posterior mean = 1.08)
Term 4, 2005
Local smoothing
(posterior mean=1.09)
BIO656--Multilevel Models
22
Best
Histogram
Term 4, 2005
BIO656--Multilevel Models
23
Term 4, 2005
BIO656--Multilevel Models
24
Comparison of Ranks
Term 4, 2005
BIO656--Multilevel Models
25
Percentiles and Moments
Term 4, 2005
BIO656--Multilevel Models
26
Highest estimated relative risk regions
Term 4, 2005
BIO656--Multilevel Models
27
Lowest estimated
relative risk regions
Term 4, 2005
BIO656--Multilevel Models
28
DISCUSSION
• It is important to explore sensitivity of the
results to modeling assumptions
– Priors, data models, .....
• For spatially correlated data use of global
smoothing may not be effective
• In the lip cancer study, the sensitivity of
results to choice of prior (global and local
smoothing) suggest presence of spatially
correlated latent factors
Term 4, 2005
BIO656--Multilevel Models
29
Term 4, 2005
BIO656--Multilevel Models
30
Term 4, 2005
BIO656--Multilevel Models
31
Term 4, 2005
BIO656--Multilevel Models
32
Socio-economic and dietary factors of
pellagra deaths in southern US
Shum, Dominici & Marks
•
•
•
1930 data from approximately 800 counties in 9
states in Southern US
Outcome is county-specific observed and expected
number of pellagra deaths
Data set includes county-specific socio-economic
characteristics and dietary factors
• % acres in cotton
• % farms under 20 acres
• dairy cows per capita
• Access to a mental hospital
• % afro-american
• % single women
Term 4, 2005
BIO656--Multilevel Models
33
PELLAGRA
• Disease caused by a deficient diet or failure of the body
to absorb B complex vitamins or an amino acid
• Prevalent in locations where people consume large
quantities of corn
• Characterized by scaly skin sores, diarrhea, mucosal
changes and mental symptoms such as schizophrenialike dementia
• May develop after gastrointestinal diseases or
alcoholism
Term 4, 2005
BIO656--Multilevel Models
34
THE SMR
Term 4, 2005
BIO656--Multilevel Models
35
Computing Expected Deaths
Term 4, 2005
BIO656--Multilevel Models
36
Crude SMR (Observed/Expected) of
Pellagra Deaths in Southern USA in 1930
(Courtesy of Harry Marks)
Term 4, 2005
BIO656--Multilevel Models
37
Analysis Questions & Framework
Use a Conditional Autoregressive (CAR) Model
• To assess which social, economic, behavioral or
dietary factors best explain the spatial distribution of
pellagra in southern US
• Which of the above factors are most important in
explaining the history of pellagra incidence in the US
• To what extent have state laws affected pellagra
incidence
• To adjust and smooth estimated rates 
Term 4, 2005
BIO656--Multilevel Models
38
Crude and smoothed SMRs:
Pellagra Deaths in Southern USA in 1930
Crude SMR
Term 4, 2005
Smoothed SMR
BIO656--Multilevel Models
39
Complex Spatial Relations
and Data Structures
Requiring a hierarchical model
to sort things out
Term 4, 2005
BIO656--Multilevel Models
40
Term 4, 2005
BIO656--Multilevel Models
41
Term 4, 2005
BIO656--Multilevel Models
42
Term 4, 2005
BIO656--Multilevel Models
43
Need a rosetta stone!
Mugglin, A.S., and Carlin, B.P. (1998)
Hierarchical Modeling in Geographic
Information Systems, Population interpolation
over incompatible zones
J. Agricultural, Biolological and Environmental
Statistics, 3: 111-130
Term 4, 2005
BIO656--Multilevel Models
44