53
Outline Disease Ecology: What do we want to do? Point Processes: Chagas Disease in Peru Spatial Regression: Raccoon Rabies MODULE 5: Spatial Statistics in Epidemiology and Public Health Lecture 9: Disease Ecology Jon Wakefield and Lance Waller 1 / 54

MODULE 5: Spatial Statistics in Epidemiology and Public ...faculty.washington.edu/jonno/SISMIDmaterial/2020_SISMID_5_9.pdfI Spatio-temporal data, mathematical models, genetic sequences,

  • Upload
    others

  • View
    0

  • Download
    0

Embed Size (px)

Citation preview

Page 1: MODULE 5: Spatial Statistics in Epidemiology and Public ...faculty.washington.edu/jonno/SISMIDmaterial/2020_SISMID_5_9.pdfI Spatio-temporal data, mathematical models, genetic sequences,

OutlineDisease Ecology: What do we want to do?

Point Processes: Chagas Disease in PeruSpatial Regression: Raccoon Rabies

MODULE 5: Spatial Statistics in Epidemiologyand Public Health

Lecture 9: Disease Ecology

Jon Wakefield and Lance Waller

1 / 54

Page 2: MODULE 5: Spatial Statistics in Epidemiology and Public ...faculty.washington.edu/jonno/SISMIDmaterial/2020_SISMID_5_9.pdfI Spatio-temporal data, mathematical models, genetic sequences,

OutlineDisease Ecology: What do we want to do?

Point Processes: Chagas Disease in PeruSpatial Regression: Raccoon Rabies

Disease Ecology: What do we want to do?Pattern and ProcessGaps and Bridges: Ecology and Statistics

Point Processes: Chagas Disease in PeruCluster detectionSpatial relative risk

Spatial Regression: Raccoon RabiesStatistical estimation of landscape barriersWomblingSpatially varying coefficients

2 / 54

Page 3: MODULE 5: Spatial Statistics in Epidemiology and Public ...faculty.washington.edu/jonno/SISMIDmaterial/2020_SISMID_5_9.pdfI Spatio-temporal data, mathematical models, genetic sequences,

OutlineDisease Ecology: What do we want to do?

Point Processes: Chagas Disease in PeruSpatial Regression: Raccoon Rabies

Pattern and ProcessGaps and Bridges: Ecology and Statistics

Disease Ecology

I Interactions between virus, host, landscape.

I Landscape epidemiology (Pavlovsky, 1967), landscape ecology(Manel et al. 2003, TrEE), spatial epidemiology (Osfeld et al.2005, TrEE), landscape genetics (host and virus) (Biek et al.2006, Science), conservation medicine (Aguirre et al. 2002).

I People, animals, diseases, ecology, environment!

I Spatio-temporal data, mathematical models, geneticsequences, missing data, GIS!

3 / 54

Page 4: MODULE 5: Spatial Statistics in Epidemiology and Public ...faculty.washington.edu/jonno/SISMIDmaterial/2020_SISMID_5_9.pdfI Spatio-temporal data, mathematical models, genetic sequences,

OutlineDisease Ecology: What do we want to do?

Point Processes: Chagas Disease in PeruSpatial Regression: Raccoon Rabies

Pattern and ProcessGaps and Bridges: Ecology and Statistics

Epizoology and Epidemiology

I Most emergent infectious diseases have animal reservoir(WNV, Ebola, Avian influenza, Monkeypox, SARS, HIV/SIV).

I History of animal/human disease (Torrey and Yolken, 2005,Beasts of the Earth).

I Interesting intersection of modelers, ecologists, statisticians,medical geographers, ecological geneticists, public healthresearchers, epidemiologists.

I One Health.

4 / 54

Page 5: MODULE 5: Spatial Statistics in Epidemiology and Public ...faculty.washington.edu/jonno/SISMIDmaterial/2020_SISMID_5_9.pdfI Spatio-temporal data, mathematical models, genetic sequences,

OutlineDisease Ecology: What do we want to do?

Point Processes: Chagas Disease in PeruSpatial Regression: Raccoon Rabies

Pattern and ProcessGaps and Bridges: Ecology and Statistics

The “big picture”

Map oftruecasesin hosts

Map ofdiag’dcasesin hosts

Map ofrep’dcasesin hosts

Reportingfilter

Diagnosisfilter

Surveillance of Host Disease

Landscapeand

Climate

Hostdistribution

Vectordistribution

Pathogendistribution

Modeling Disease Ecology

Ecology Public Health

5 / 54

Page 6: MODULE 5: Spatial Statistics in Epidemiology and Public ...faculty.washington.edu/jonno/SISMIDmaterial/2020_SISMID_5_9.pdfI Spatio-temporal data, mathematical models, genetic sequences,

OutlineDisease Ecology: What do we want to do?

Point Processes: Chagas Disease in PeruSpatial Regression: Raccoon Rabies

Pattern and ProcessGaps and Bridges: Ecology and Statistics

Not a new idea (Koch, 2005, Cartographies of Disease)

6 / 54

Page 7: MODULE 5: Spatial Statistics in Epidemiology and Public ...faculty.washington.edu/jonno/SISMIDmaterial/2020_SISMID_5_9.pdfI Spatio-temporal data, mathematical models, genetic sequences,

OutlineDisease Ecology: What do we want to do?

Point Processes: Chagas Disease in PeruSpatial Regression: Raccoon Rabies

Pattern and ProcessGaps and Bridges: Ecology and Statistics

Pattern and Process

I Our ultimate goal is understanding the ecological processesdriving the patterns we see in our observations.

I When linking process (model or reality) to pattern (data),typically:I Ecology focus: Process to pattern

I Emphasis on mathematical model, link to available data

I Statistics: Pattern to processI Collected data, hypothesis test or analytic (e.g., regression)

model.

7 / 54

Page 8: MODULE 5: Spatial Statistics in Epidemiology and Public ...faculty.washington.edu/jonno/SISMIDmaterial/2020_SISMID_5_9.pdfI Spatio-temporal data, mathematical models, genetic sequences,

OutlineDisease Ecology: What do we want to do?

Point Processes: Chagas Disease in PeruSpatial Regression: Raccoon Rabies

Pattern and ProcessGaps and Bridges: Ecology and Statistics

Ultimately futile exercise?

I Process may not yield unique pattern (e.g., chaos,stochasticity).

I Pattern may not reveal unique process without additionalinformation (e.g., spatial point patterns, Bartlett (1964)).

I But the real question is, “Can we learn more than we alreadyknow?”

I If not, what additional data do we need?

8 / 54

Page 9: MODULE 5: Spatial Statistics in Epidemiology and Public ...faculty.washington.edu/jonno/SISMIDmaterial/2020_SISMID_5_9.pdfI Spatio-temporal data, mathematical models, genetic sequences,

OutlineDisease Ecology: What do we want to do?

Point Processes: Chagas Disease in PeruSpatial Regression: Raccoon Rabies

Pattern and ProcessGaps and Bridges: Ecology and Statistics

The whirling vortex

1

Questions you

want to answer

Data you need

to answer the

questions.

Questions you

can answer with

the data you have.

Data you can

get.

How close?

9 / 54

Page 10: MODULE 5: Spatial Statistics in Epidemiology and Public ...faculty.washington.edu/jonno/SISMIDmaterial/2020_SISMID_5_9.pdfI Spatio-temporal data, mathematical models, genetic sequences,

OutlineDisease Ecology: What do we want to do?

Point Processes: Chagas Disease in PeruSpatial Regression: Raccoon Rabies

Pattern and ProcessGaps and Bridges: Ecology and Statistics

Exactly wrong or approximately correct?

I John Tukey noted an approximate answer to the right questionis better than a precise answer to the wrong question.

I Particularly important here...if available data redefine ouranswerable questions, we may be changing course withoutrealizing it!

I Let’s look at how modelers and statisticians address thesequestions...

10 / 54

Page 11: MODULE 5: Spatial Statistics in Epidemiology and Public ...faculty.washington.edu/jonno/SISMIDmaterial/2020_SISMID_5_9.pdfI Spatio-temporal data, mathematical models, genetic sequences,

OutlineDisease Ecology: What do we want to do?

Point Processes: Chagas Disease in PeruSpatial Regression: Raccoon Rabies

Pattern and ProcessGaps and Bridges: Ecology and Statistics

Gaps and Bridges

Conceptual gapI Mathematical modelers

I Build assessments using families of models and derivingproperties.

I Data used to calibrate models.I Using process (model) to understand pattern (data).

I StatisticiansI Build inference from probability model defining observations.I Data define a likelihood function.I Using pattern (data) to understand process (model).

11 / 54

Page 12: MODULE 5: Spatial Statistics in Epidemiology and Public ...faculty.washington.edu/jonno/SISMIDmaterial/2020_SISMID_5_9.pdfI Spatio-temporal data, mathematical models, genetic sequences,

OutlineDisease Ecology: What do we want to do?

Point Processes: Chagas Disease in PeruSpatial Regression: Raccoon Rabies

Pattern and ProcessGaps and Bridges: Ecology and Statistics

Gaps and Bridges

Training gap

I In current modes of training, mathematical modelers oftentake one (or fewer) courses in statistics.

I Statisticians often take one (or fewer) courses inmathematical modeling.

I Furthermore, the importance of one area is seldom stressed inthe other.

I Few working at the intersection of the two but there is a lot ofinteresting work to be done!

To see how this might work, consider the following...

12 / 54

Page 13: MODULE 5: Spatial Statistics in Epidemiology and Public ...faculty.washington.edu/jonno/SISMIDmaterial/2020_SISMID_5_9.pdfI Spatio-temporal data, mathematical models, genetic sequences,

OutlineDisease Ecology: What do we want to do?

Point Processes: Chagas Disease in PeruSpatial Regression: Raccoon Rabies

Pattern and ProcessGaps and Bridges: Ecology and Statistics

How statistics might help...(Ecology, 2010)

1

True

model

Observed

data

Observed

associations

Proposed

model

Generated

data

Induced

associations

Goodness of fit

Regressions,

correlations, etc.

Parameter estimation

13 / 54

Page 14: MODULE 5: Spatial Statistics in Epidemiology and Public ...faculty.washington.edu/jonno/SISMIDmaterial/2020_SISMID_5_9.pdfI Spatio-temporal data, mathematical models, genetic sequences,

OutlineDisease Ecology: What do we want to do?

Point Processes: Chagas Disease in PeruSpatial Regression: Raccoon Rabies

Cluster detectionSpatial relative risk

Point Processes in Disease Ecology: Chagas disease in Peru

I Joint work with Michael Levy (Fogarty International Center,NIH)

I Chagas disease: Vector borne disease (infection with T. cruzi).

I Vector (in southern Peru): Triatoma infestans.

I Study area: Guadalupe, Peru (peri-urban).

I Fields surrounding rocky hilltops with houses.

I GPS all household locations.

I Spraying campaign, identify house locations, houses withvectors (“infested”), and houses with infected vectors(“infected”).

14 / 54

Page 15: MODULE 5: Spatial Statistics in Epidemiology and Public ...faculty.washington.edu/jonno/SISMIDmaterial/2020_SISMID_5_9.pdfI Spatio-temporal data, mathematical models, genetic sequences,

OutlineDisease Ecology: What do we want to do?

Point Processes: Chagas Disease in PeruSpatial Regression: Raccoon Rabies

Cluster detectionSpatial relative risk

How to find a cluster?

I Consider two approaches: scan statistic and intensityestimators.

I Spatial scan statistic:I Define set of potential clusters (elements of scanning window).I Assign “score” to each potential cluster.I Find “most likely cluster” (MLC) as potential cluster with

extreme score.I Evaluate significance of most likely cluster via Monte Carlo

simulation.I Compare observed “score” of MLC to distribution of scores

MLCs (regardless of location) under random assignment.I SaTScan software (www.satscan.org).

15 / 54

Page 16: MODULE 5: Spatial Statistics in Epidemiology and Public ...faculty.washington.edu/jonno/SISMIDmaterial/2020_SISMID_5_9.pdfI Spatio-temporal data, mathematical models, genetic sequences,

OutlineDisease Ecology: What do we want to do?

Point Processes: Chagas Disease in PeruSpatial Regression: Raccoon Rabies

Cluster detectionSpatial relative risk

SaTScan, Infested among households, Most likely cluster(p=0.002)

−71.592 −71.591 −71.590 −71.589 −71.588 −71.587

−16.441

−16.440

−16.439

−16.438

−16.437

−16.436

−16.435

Longitude

Latitude

SaTScan, Most Likely Cluster, Infested, p−value = 0.002

16 / 54

Page 17: MODULE 5: Spatial Statistics in Epidemiology and Public ...faculty.washington.edu/jonno/SISMIDmaterial/2020_SISMID_5_9.pdfI Spatio-temporal data, mathematical models, genetic sequences,

OutlineDisease Ecology: What do we want to do?

Point Processes: Chagas Disease in PeruSpatial Regression: Raccoon Rabies

Cluster detectionSpatial relative risk

SaTScan, Infected among infested, Most likely cluster(p=0.181)

−71.592 −71.591 −71.590 −71.589 −71.588 −71.587

−16.441

−16.440

−16.439

−16.438

−16.437

−16.436

−16.435

Longitude

Latitude

OOOOO OOOO

OOO

O

OOOOOOOO

O OO

O

O

OOOOOOOOOOOOO OOOO

O

OOOOO

OOO

OOOO

OO

OO OOO

O

OOOO

OOOO

SaTScan, Most Likely Cluster, Infected, p−value = 0.181

17 / 54

Page 18: MODULE 5: Spatial Statistics in Epidemiology and Public ...faculty.washington.edu/jonno/SISMIDmaterial/2020_SISMID_5_9.pdfI Spatio-temporal data, mathematical models, genetic sequences,

OutlineDisease Ecology: What do we want to do?

Point Processes: Chagas Disease in PeruSpatial Regression: Raccoon Rabies

Cluster detectionSpatial relative risk

Chagas SaTScan conclusions

I Statistically significant cluster of infested households amongall households.

I No statistically significant cluster of infected householdsamong infested households.

I Note circular most likely cluster may include gaps (top of hill).

I What about non-circular clusters?

18 / 54

Page 19: MODULE 5: Spatial Statistics in Epidemiology and Public ...faculty.washington.edu/jonno/SISMIDmaterial/2020_SISMID_5_9.pdfI Spatio-temporal data, mathematical models, genetic sequences,

OutlineDisease Ecology: What do we want to do?

Point Processes: Chagas Disease in PeruSpatial Regression: Raccoon Rabies

Cluster detectionSpatial relative risk

Kernel intensity estimates, infested vs. all households

−71.592 −71.590 −71.588

−16.441

−16.439

−16.437

−16.435

Longitude

Latitude

Infested sites

−71.592 −71.590 −71.588

−16.441

−16.439

−16.437

−16.435

Longitude

Latitude

Non−infested sites

19 / 54

Page 20: MODULE 5: Spatial Statistics in Epidemiology and Public ...faculty.washington.edu/jonno/SISMIDmaterial/2020_SISMID_5_9.pdfI Spatio-temporal data, mathematical models, genetic sequences,

OutlineDisease Ecology: What do we want to do?

Point Processes: Chagas Disease in PeruSpatial Regression: Raccoon Rabies

Cluster detectionSpatial relative risk

Ratio of kernel intensity estimates, infested vs. allhouseholds

−71.592 −71.591 −71.590 −71.589 −71.588 −71.587

−16.441

−16.440

−16.439

−16.438

−16.437

−16.436

−16.435

Longitude

Latitude

Log relative risk surface: Infested

|||||||||||||||||||||||||||||||||||||||||||||||||||||||||| ||||||||||||||||||| |||||||||||||||||| ||||||||||||||||||| |||||||||||||||||||| ||||||||||||||||||||| |||||||||||||||||||||| ||||||||||||||||||||||| |||||||||||||||||||||||| ||||||||||||||||||||||||| |||||||||||||||||||||||||||| ||||||||||||||||||||||||||||| |||||||||||||||||||||||||||||||| ||||||||||||||||||||||||||||||||||| |||||||||||||||||||||||||||||||||||| |||||||||||||||||||||||||||||||||| ||||||||||||||||||||||||| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||

−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−

20 / 54

Page 21: MODULE 5: Spatial Statistics in Epidemiology and Public ...faculty.washington.edu/jonno/SISMIDmaterial/2020_SISMID_5_9.pdfI Spatio-temporal data, mathematical models, genetic sequences,

OutlineDisease Ecology: What do we want to do?

Point Processes: Chagas Disease in PeruSpatial Regression: Raccoon Rabies

Cluster detectionSpatial relative risk

Kernel intensity estimates, infected vs. infested households

−71.592 −71.590 −71.588

−16.441

−16.439

−16.437

−16.435

Longitude

Latitude

O OOOOOOOO

OOOO

OOOOOOOOOOOO

O

OOOOOOOOOOOOOOOOOOOOO

OOOOO

OOOO

OOOOOOO

OOOOOOOOO

Infected sites

−71.592 −71.590 −71.588

−16.441

−16.439

−16.437

−16.435

Longitude

Latitude

Infested not infected sites

21 / 54

Page 22: MODULE 5: Spatial Statistics in Epidemiology and Public ...faculty.washington.edu/jonno/SISMIDmaterial/2020_SISMID_5_9.pdfI Spatio-temporal data, mathematical models, genetic sequences,

OutlineDisease Ecology: What do we want to do?

Point Processes: Chagas Disease in PeruSpatial Regression: Raccoon Rabies

Cluster detectionSpatial relative risk

Ratio of kernel intensity estimates, infected vs. infested

−71.592 −71.591 −71.590 −71.589 −71.588 −71.587

−16.441

−16.440

−16.439

−16.438

−16.437

−16.436

−16.435

Longitude

Latitude

OOOOO OOOO

OOO

O

OOOOOOOO

O OO

O

O

OOOOOOOOOOOOO OOOO

O

OOOOO

OOO

OOOO

OO

OO OOO

O

OOOO

OOOO

Log relative risk surface: Infected

||||||||||||||||||||||| ||||||||||||||||||||||||||||||||||||||

||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||

−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−

−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−

22 / 54

Page 23: MODULE 5: Spatial Statistics in Epidemiology and Public ...faculty.washington.edu/jonno/SISMIDmaterial/2020_SISMID_5_9.pdfI Spatio-temporal data, mathematical models, genetic sequences,

OutlineDisease Ecology: What do we want to do?

Point Processes: Chagas Disease in PeruSpatial Regression: Raccoon Rabies

Cluster detectionSpatial relative risk

Cluster conclusions

I Relative risk surface adds more geographical precision topatterns initially revealed by SaTScan.

I Large risk of infestation in the south.

I Within this some pockets of increased risk of infection.

I Area of lower risk missed by circular scan statistic, due to itsirregular shape.

I Identifies areas for future studies.

23 / 54

Page 24: MODULE 5: Spatial Statistics in Epidemiology and Public ...faculty.washington.edu/jonno/SISMIDmaterial/2020_SISMID_5_9.pdfI Spatio-temporal data, mathematical models, genetic sequences,

OutlineDisease Ecology: What do we want to do?

Point Processes: Chagas Disease in PeruSpatial Regression: Raccoon Rabies

Cluster detectionSpatial relative risk

Chagas conclusions

I Significant cluster of infested households, but no clusters ofinfected households (circular clusters).

I Relative risk surface also suggests area of low risk (bothinfestation and infection) in northeast.

I K functions suggest significant clustering of infected but notinfested households.

I Taken together, results reveal different aspects of theunderlying process.

I A single cluster does not define clustering, nor does clusteringimply a single cluster.

24 / 54

Page 25: MODULE 5: Spatial Statistics in Epidemiology and Public ...faculty.washington.edu/jonno/SISMIDmaterial/2020_SISMID_5_9.pdfI Spatio-temporal data, mathematical models, genetic sequences,

OutlineDisease Ecology: What do we want to do?

Point Processes: Chagas Disease in PeruSpatial Regression: Raccoon Rabies

Cluster detectionSpatial relative risk

Questions?

25 / 54

Page 26: MODULE 5: Spatial Statistics in Epidemiology and Public ...faculty.washington.edu/jonno/SISMIDmaterial/2020_SISMID_5_9.pdfI Spatio-temporal data, mathematical models, genetic sequences,

OutlineDisease Ecology: What do we want to do?

Point Processes: Chagas Disease in PeruSpatial Regression: Raccoon Rabies

Cluster detectionSpatial relative risk

Chagas conclusions

I Infestation: pockets of higher and lower relative risk, but levelof clustering not different between cases and controls.

I Infection: More clustered at small distances than infestation,but resulting clusters are smaller and more diffuse.

I Scale of clustering different between infestation and infection,and larger than typical range of individual vectors.

I Scale of clustering useful in targeted surveillance for humancases (Levy et al., 2007, PLoS NTD).

26 / 54

Page 27: MODULE 5: Spatial Statistics in Epidemiology and Public ...faculty.washington.edu/jonno/SISMIDmaterial/2020_SISMID_5_9.pdfI Spatio-temporal data, mathematical models, genetic sequences,

OutlineDisease Ecology: What do we want to do?

Point Processes: Chagas Disease in PeruSpatial Regression: Raccoon Rabies

Cluster detectionSpatial relative risk

Questions?

27 / 54

Page 28: MODULE 5: Spatial Statistics in Epidemiology and Public ...faculty.washington.edu/jonno/SISMIDmaterial/2020_SISMID_5_9.pdfI Spatio-temporal data, mathematical models, genetic sequences,

OutlineDisease Ecology: What do we want to do?

Point Processes: Chagas Disease in PeruSpatial Regression: Raccoon Rabies

Statistical estimation of landscape barriersWomblingSpatially varying coefficients

Raccoon rabies

28 / 54

Page 29: MODULE 5: Spatial Statistics in Epidemiology and Public ...faculty.washington.edu/jonno/SISMIDmaterial/2020_SISMID_5_9.pdfI Spatio-temporal data, mathematical models, genetic sequences,

OutlineDisease Ecology: What do we want to do?

Point Processes: Chagas Disease in PeruSpatial Regression: Raccoon Rabies

Statistical estimation of landscape barriersWomblingSpatially varying coefficients

What is rabies?

I Virus in family of Lyssa (“frenzy”) virus.

I Behavioral impact on host.

I Reportable disease.

I Various strains associated with primary host (bat, dog, coyote,fox, skunk, and raccoon).

I Host cross-over, typically transmitted via bite/scratch.

I Most human infection from bat strains.

29 / 54

Page 30: MODULE 5: Spatial Statistics in Epidemiology and Public ...faculty.washington.edu/jonno/SISMIDmaterial/2020_SISMID_5_9.pdfI Spatio-temporal data, mathematical models, genetic sequences,

OutlineDisease Ecology: What do we want to do?

Point Processes: Chagas Disease in PeruSpatial Regression: Raccoon Rabies

Statistical estimation of landscape barriersWomblingSpatially varying coefficients

Raccoon rabies

I Endemic in Florida and South Georgia.

I Translocation of rabid animal(s) to VA/WV border circa 1977.

I Wave-like spread since.

I Connecticut first appearance 1991-1996.

I Ohio 2005.

I Joint work with Leslie Real’s lab in Population Biology,Evolution, and Ecology (David Smith, Colin Russell, RomanBiek, Scott Duke-Sylvester).

30 / 54

Page 31: MODULE 5: Spatial Statistics in Epidemiology and Public ...faculty.washington.edu/jonno/SISMIDmaterial/2020_SISMID_5_9.pdfI Spatio-temporal data, mathematical models, genetic sequences,

OutlineDisease Ecology: What do we want to do?

Point Processes: Chagas Disease in PeruSpatial Regression: Raccoon Rabies

Statistical estimation of landscape barriersWomblingSpatially varying coefficients

Barrier estimation: What do we want?

I Goal: Measure effect of landscape features, (e.g., mountainsand rivers) on the speed of raccoon rabies diffusion.

I Elevation, river or road presence significantly related toraccoon rabies counts (Recuenco et al. 2007) andtransmission time (Russell et al. 2004).

I Landscape features may serve as either barriers or gateways tothe spread of infectious disease.

I Find and visualize barriers: Do they align with certainlandscape features?

31 / 54

Page 32: MODULE 5: Spatial Statistics in Epidemiology and Public ...faculty.washington.edu/jonno/SISMIDmaterial/2020_SISMID_5_9.pdfI Spatio-temporal data, mathematical models, genetic sequences,

OutlineDisease Ecology: What do we want to do?

Point Processes: Chagas Disease in PeruSpatial Regression: Raccoon Rabies

Statistical estimation of landscape barriersWomblingSpatially varying coefficients

Data: What do we have?

I Time in months to first reported raccoon rabies case in 428contiguous counties in the Eastern US (CDC).I 0 for origin county: Pendleton, WV.

I Mean elevation by county (USGS - Geographic NamesInformation System).

I Indicator for major river presence in county (ESRI data and ageographic information system (GIS)).

I Population density by county (US Census and ESRI).

I Distance between origin county and all counties.

32 / 54

Page 33: MODULE 5: Spatial Statistics in Epidemiology and Public ...faculty.washington.edu/jonno/SISMIDmaterial/2020_SISMID_5_9.pdfI Spatio-temporal data, mathematical models, genetic sequences,

OutlineDisease Ecology: What do we want to do?

Point Processes: Chagas Disease in PeruSpatial Regression: Raccoon Rabies

Statistical estimation of landscape barriersWomblingSpatially varying coefficients

Data

33 / 54

Page 34: MODULE 5: Spatial Statistics in Epidemiology and Public ...faculty.washington.edu/jonno/SISMIDmaterial/2020_SISMID_5_9.pdfI Spatio-temporal data, mathematical models, genetic sequences,

OutlineDisease Ecology: What do we want to do?

Point Processes: Chagas Disease in PeruSpatial Regression: Raccoon Rabies

Statistical estimation of landscape barriersWomblingSpatially varying coefficients

Wombling

I Joint work with David Wheeler (Wheeler and Waller, JABES,2008).

I Wombling: determine boundaries on a map by finding wherelocal spread (change) is slower than elsewhere (Womble, 1951Science).

I William H. Womble a bit of an elusive figure...

34 / 54

Page 35: MODULE 5: Spatial Statistics in Epidemiology and Public ...faculty.washington.edu/jonno/SISMIDmaterial/2020_SISMID_5_9.pdfI Spatio-temporal data, mathematical models, genetic sequences,

OutlineDisease Ecology: What do we want to do?

Point Processes: Chagas Disease in PeruSpatial Regression: Raccoon Rabies

Statistical estimation of landscape barriersWomblingSpatially varying coefficients

William H. Womble (?)

35 / 54

Page 36: MODULE 5: Spatial Statistics in Epidemiology and Public ...faculty.washington.edu/jonno/SISMIDmaterial/2020_SISMID_5_9.pdfI Spatio-temporal data, mathematical models, genetic sequences,

OutlineDisease Ecology: What do we want to do?

Point Processes: Chagas Disease in PeruSpatial Regression: Raccoon Rabies

Statistical estimation of landscape barriersWomblingSpatially varying coefficients

Google search: W.H. Womble Professor Robert Stencel

36 / 54

Page 37: MODULE 5: Spatial Statistics in Epidemiology and Public ...faculty.washington.edu/jonno/SISMIDmaterial/2020_SISMID_5_9.pdfI Spatio-temporal data, mathematical models, genetic sequences,

OutlineDisease Ecology: What do we want to do?

Point Processes: Chagas Disease in PeruSpatial Regression: Raccoon Rabies

Statistical estimation of landscape barriersWomblingSpatially varying coefficients

Which leads to...

37 / 54

Page 38: MODULE 5: Spatial Statistics in Epidemiology and Public ...faculty.washington.edu/jonno/SISMIDmaterial/2020_SISMID_5_9.pdfI Spatio-temporal data, mathematical models, genetic sequences,

OutlineDisease Ecology: What do we want to do?

Point Processes: Chagas Disease in PeruSpatial Regression: Raccoon Rabies

Statistical estimation of landscape barriersWomblingSpatially varying coefficients

Are you weady to womble?

I Consider a set of potential boundaries and decide if each is a“real” boundary or not.

I Many algorithmic approaches both deterministic and “fuzzy”.

I Adopt a Bayesian hierarchical model for wombling (Lu andCarlin 2005).

I Bayesian approach provides a direct estimate of theprobability that a line segment between two adjacent areas isa barrier (fuzzy boundary) in contrast to algorithmic versionsbased on thresholds, etc.

38 / 54

Page 39: MODULE 5: Spatial Statistics in Epidemiology and Public ...faculty.washington.edu/jonno/SISMIDmaterial/2020_SISMID_5_9.pdfI Spatio-temporal data, mathematical models, genetic sequences,

OutlineDisease Ecology: What do we want to do?

Point Processes: Chagas Disease in PeruSpatial Regression: Raccoon Rabies

Statistical estimation of landscape barriersWomblingSpatially varying coefficients

Bayesian areal wombling

I Model time to first reported raccoon rabies case Yi :

Yi |µi , τ ∼ N(µi , 1/τ)

whereµi = α + φi

is the expected value of time to first case per county.

I Spatial random effects follow a conditionally autoregressive(CAR) prior φ ∼ CAR(η) with a mean random effectdetermined by its neighboring values.

39 / 54

Page 40: MODULE 5: Spatial Statistics in Epidemiology and Public ...faculty.washington.edu/jonno/SISMIDmaterial/2020_SISMID_5_9.pdfI Spatio-temporal data, mathematical models, genetic sequences,

OutlineDisease Ecology: What do we want to do?

Point Processes: Chagas Disease in PeruSpatial Regression: Raccoon Rabies

Statistical estimation of landscape barriersWomblingSpatially varying coefficients

Bayesian areal wombling

I Boundary likelihood value (BLV) assigned to each potentialboundary (here, edge between two counties), based ondifference in expected (modeled) time to first appearance.

∆ij = |µi − µj |

I Use MCMC to draw sample from posterior [∆ij |y ] based ondraws from posteriors [µi |y ] and [µj |y ].

I This assigns a posterior probability for each edge, then displayedges with with p(∆ij > c|y) for some threshold probability c .

40 / 54

Page 41: MODULE 5: Spatial Statistics in Epidemiology and Public ...faculty.washington.edu/jonno/SISMIDmaterial/2020_SISMID_5_9.pdfI Spatio-temporal data, mathematical models, genetic sequences,

OutlineDisease Ecology: What do we want to do?

Point Processes: Chagas Disease in PeruSpatial Regression: Raccoon Rabies

Statistical estimation of landscape barriersWomblingSpatially varying coefficients

Wombling boundaries: p(∆ij > c |y)

41 / 54

Page 42: MODULE 5: Spatial Statistics in Epidemiology and Public ...faculty.washington.edu/jonno/SISMIDmaterial/2020_SISMID_5_9.pdfI Spatio-temporal data, mathematical models, genetic sequences,

OutlineDisease Ecology: What do we want to do?

Point Processes: Chagas Disease in PeruSpatial Regression: Raccoon Rabies

Statistical estimation of landscape barriersWomblingSpatially varying coefficients

Linking to local covariates

I Bayesian areal wombling provides estimates of barriers butdoes not allow direct inference regarding the impact ofparticular landscape barriers on the evidence for barriers.

I We could expand our fixed effect α to X ′β to include localcovariates (e.g., elevation, boundary based on a river).

I However, what it if the effect of elevation or presence of rivervaries from place to place?

I Russell et al. (2003, PNAS) suggest that river effect dependson direction of movement of the wave (perpendicular? Slower.Parallel? Faster.)

42 / 54

Page 43: MODULE 5: Spatial Statistics in Epidemiology and Public ...faculty.washington.edu/jonno/SISMIDmaterial/2020_SISMID_5_9.pdfI Spatio-temporal data, mathematical models, genetic sequences,

OutlineDisease Ecology: What do we want to do?

Point Processes: Chagas Disease in PeruSpatial Regression: Raccoon Rabies

Statistical estimation of landscape barriersWomblingSpatially varying coefficients

Spatially varying coefficients

I We consider a spatially varying coefficient model with CARpriors on the covariate effects β, i.e.,

Yi |µi , τ ∼ N(µi , 1/τ)

whereµi = X ′iβi + φi

I Spatial priors on elements of βi .

I More specifically, assign a multivariate CAR prior on the set ofβ (Banerjee et al. 2004).

43 / 54

Page 44: MODULE 5: Spatial Statistics in Epidemiology and Public ...faculty.washington.edu/jonno/SISMIDmaterial/2020_SISMID_5_9.pdfI Spatio-temporal data, mathematical models, genetic sequences,

OutlineDisease Ecology: What do we want to do?

Point Processes: Chagas Disease in PeruSpatial Regression: Raccoon Rabies

Statistical estimation of landscape barriersWomblingSpatially varying coefficients

MultiCAR details

I βi = (βi1, βi2, . . . , βip)′

I βi |(β(−i),1,β(−i),2, . . . ,β(−i),p) ∼ N(βi ,Ω/mi )where

βi = (βi1, βi2, . . . , βip)′

andβi1 =

∑k∈κi

βk1/mi

where κi = neighbor set for region i , and |κi | = mi .

I Ω ∼ Inverse-Wishart(ν, 0.02 · Ip×p).

44 / 54

Page 45: MODULE 5: Spatial Statistics in Epidemiology and Public ...faculty.washington.edu/jonno/SISMIDmaterial/2020_SISMID_5_9.pdfI Spatio-temporal data, mathematical models, genetic sequences,

OutlineDisease Ecology: What do we want to do?

Point Processes: Chagas Disease in PeruSpatial Regression: Raccoon Rabies

Statistical estimation of landscape barriersWomblingSpatially varying coefficients

Including covariates

I Include effects of (mean) elevation, presence of a major river,and the natural log of the (human) population density.

I Best fitting (via DIC) model includes spatial variation in allthree (and intercept).

E [Yi ] = βi1 +βi2(mean elev) +βi3(river) +β4i (log(pop dens))

45 / 54

Page 46: MODULE 5: Spatial Statistics in Epidemiology and Public ...faculty.washington.edu/jonno/SISMIDmaterial/2020_SISMID_5_9.pdfI Spatio-temporal data, mathematical models, genetic sequences,

OutlineDisease Ecology: What do we want to do?

Point Processes: Chagas Disease in PeruSpatial Regression: Raccoon Rabies

Statistical estimation of landscape barriersWomblingSpatially varying coefficients

β1: int, β2: elev, β3: river, β4: log pop

46 / 54

Page 47: MODULE 5: Spatial Statistics in Epidemiology and Public ...faculty.washington.edu/jonno/SISMIDmaterial/2020_SISMID_5_9.pdfI Spatio-temporal data, mathematical models, genetic sequences,

OutlineDisease Ecology: What do we want to do?

Point Processes: Chagas Disease in PeruSpatial Regression: Raccoon Rabies

Statistical estimation of landscape barriersWomblingSpatially varying coefficients

Findings/interpretations

I Map of posterior mean (MU): shows the overall wave orspread.

I Random intercept reveals local adjustments.

I River effect indicates increases in time until first appearanceacross Potomac and Susquehanna Rivers, decreases time forHudson River and others.

I Elevation is not difference in elevation so not directlyinforming on elevation gradients as barriers, simply elevationimpact on time until appearance.

47 / 54

Page 48: MODULE 5: Spatial Statistics in Epidemiology and Public ...faculty.washington.edu/jonno/SISMIDmaterial/2020_SISMID_5_9.pdfI Spatio-temporal data, mathematical models, genetic sequences,

OutlineDisease Ecology: What do we want to do?

Point Processes: Chagas Disease in PeruSpatial Regression: Raccoon Rabies

Statistical estimation of landscape barriersWomblingSpatially varying coefficients

SVC wombled boundaries: p(∆ij > c |y)

48 / 54

Page 49: MODULE 5: Spatial Statistics in Epidemiology and Public ...faculty.washington.edu/jonno/SISMIDmaterial/2020_SISMID_5_9.pdfI Spatio-temporal data, mathematical models, genetic sequences,

OutlineDisease Ecology: What do we want to do?

Point Processes: Chagas Disease in PeruSpatial Regression: Raccoon Rabies

Statistical estimation of landscape barriersWomblingSpatially varying coefficients

Including covariates → better wombling?

49 / 54

Page 50: MODULE 5: Spatial Statistics in Epidemiology and Public ...faculty.washington.edu/jonno/SISMIDmaterial/2020_SISMID_5_9.pdfI Spatio-temporal data, mathematical models, genetic sequences,

OutlineDisease Ecology: What do we want to do?

Point Processes: Chagas Disease in PeruSpatial Regression: Raccoon Rabies

Statistical estimation of landscape barriersWomblingSpatially varying coefficients

Raccoon rabies: What do we have so far?

I A Monte Carlo approach to assess whether observed dataseems consistent with our model.

I Spatial analysis of patterns of appearance, inference regardinglocal barriers to geographic spread.

I But where do the data come from? How accurate are they?

50 / 54

Page 51: MODULE 5: Spatial Statistics in Epidemiology and Public ...faculty.washington.edu/jonno/SISMIDmaterial/2020_SISMID_5_9.pdfI Spatio-temporal data, mathematical models, genetic sequences,

OutlineDisease Ecology: What do we want to do?

Point Processes: Chagas Disease in PeruSpatial Regression: Raccoon Rabies

Statistical estimation of landscape barriersWomblingSpatially varying coefficients

Overall Conclusions

I Much to be done to link mathematical models to statisticalideas.

I Disease ecology offers great setting for exploration.

I Models of transmission, interaction, observation.

I Mathematical models can inform statistics, statistics caninform models.

I Room to move past “ad-hockery”.

I Linking landscape features in a more meaningful (inferential)and spatial way.

51 / 54

Page 52: MODULE 5: Spatial Statistics in Epidemiology and Public ...faculty.washington.edu/jonno/SISMIDmaterial/2020_SISMID_5_9.pdfI Spatio-temporal data, mathematical models, genetic sequences,

OutlineDisease Ecology: What do we want to do?

Point Processes: Chagas Disease in PeruSpatial Regression: Raccoon Rabies

Statistical estimation of landscape barriersWomblingSpatially varying coefficients

Closing

“... Nature’s dice are always loaded, ... in her heaps and rubbishare concealed sure and useful results.”Ralph Waldo Emerson, Nature.

52 / 54

Page 53: MODULE 5: Spatial Statistics in Epidemiology and Public ...faculty.washington.edu/jonno/SISMIDmaterial/2020_SISMID_5_9.pdfI Spatio-temporal data, mathematical models, genetic sequences,

OutlineDisease Ecology: What do we want to do?

Point Processes: Chagas Disease in PeruSpatial Regression: Raccoon Rabies

Statistical estimation of landscape barriersWomblingSpatially varying coefficients

References

I Smith et al. (2002) Predicting the spatial dynamics of rabiesepidemics on heterogeneous landscapes. PNAS 99,3668-3672.

I Waller et al. (2003) Monte Carlo assessments of fit forecological simulation models. Eco Mod 164, 49-63.

I Waller (2010) Bridging gaps between statistical andmathematical modeling in ecology. Ecology 91, 3500-3502.

I Wheeler and Waller (2008) Mountains, valleys, and rivers:The transmission of raccoon rabies over a heterogeneouslandscape. JABES 13, 388-406.

53 / 54