QUANTIFYING SPATIAL RELATIONSHIPS BETWEEN LANDSCAPE PATTERNS
LINKED TO ANTHROPOGENIC DISTURBANCES AND BURULI ULCER DISEASE
By
Lindsay P. Campbell

A THESIS
Submitted to
Michigan State University
In partial fulfillment of the requirements
for the degree of
MASTER OF SCIENCE
Geography
2011

ABSTRACT
QUANTIFYING SPATIAL RELATIONSHIPS BETWEEN LANDSCAPE PATTERNS
LINKED TO ANTHROPOGENIC DISTURBANCES AND BURULI ULCER DISEASE
By
Lindsay P. Campbell
Anthropogenic ecosystem disturbances play an important role in the distribution of
emerging and re-emerging infectious diseases. Advances in GIS, remote sensing technologies,
and spatial statistical methods facilitate the observation and quantification of anthropogenic
landscape disturbances, providing the tools necessary to link these disturbances to disease
emergence. Investigations into disturbances linked to environmental bacterial infections are an
underrepresented research area. One such disease is Buruli ulcer disease (BU), caused by the
environmental pathogen Mycobacterium ulcerans. The ecological drivers behind pathogen
proliferation and transmission to humans are currently unknown. The main objective of this
study included using a spatial landscape ecological approach to determine whether land cover
patches indicative of anthropogenic landscape disturbances surrounded villages with higher BU
rates. Landscape-level results supported study hypotheses that more fragmented landscapes, with
more uniform land cover patch shapes, lying within areas more likely to collect water surround
villages with higher BU rates, but results were not consistent. Class-level results, analyzing
forest, wetland, and a mixed agriculture/forest class, suggested that more aggregated patches
with more complex shapes surround villages with higher BU rates. Incorporation of a spatial
random effects component accounted for spatial autocorrelation and provided a spatial structure
from which to predict BU rates at unsampled locations.

Copyright by
Lindsay P. Campbell
2011

Dedicated to my wonderful and supportive husband, David, and to my parents and sister
who have always believed in me.

iv

ACKNOWLEDGEMENTS

Foremost, I would like to acknowledge my advisor, Dr. Jiaguo Qi, for his support and
guidance throughout my studies. The opportunities he provided to me proved life-changing and
laid the foundation to further my academic career.
I would also like to acknowledge the time and effort of my committee members. Dr.
Andrew Finley introduced me to the spatial analytical methods central to the formation of this
thesis. Dr. M. Eric Benbow introduced me to fieldwork and cultivated all aspects of my research
skills. Dr. Richard Merritt offered sage advice and provided excellent opportunities to attend
academic conferences where I presented our work.
In addition to my committee members, Dr. Pamela Small acted as a valuable mentor and
provided opportunities for international travel. Jenni van Ravensway became a wonderful friend
and was an excellent resource throughout this experience. I had the opportunity to meet and to
work with several dedicated researchers and graduate students from the U.S, Benin, and Ghana,
including Dr. Lance Waller, Dr. Heather Williamson, Dr. Mollie McIntosh, Ryan Kimbirauskas,
Dr. Christian Johnson, Dr. Ghishlain Sopoh, Charles Yeboah, and Charles Quaye. I would also
like to acknowledge fellow graduate students Chuan Qin, Siam Lawawirowong, Tanita Suepa,
Zhang Feng, and Mark DeVisser.
Finally, I would like to express my sincere appreciation for my husband, family, and
friends, particularly Cheryl Lyons, who supported me throughout my studies.

v

TABLE OF CONTENTS
LIST OF TABLES ....................................................................................................................... viii
LIST OF EQUATIONS ............................................................................................................... xiv
INTRODUCTION .......................................................................................................................... 1
Chapter 1 LITERATURE REVIEW ............................................................................................... 4
1.1 Buruli ulcer disease..........................................................................................................4
1.2 Symptoms and Treatment ................................................................................................5
1.3 Prevalence and Economic Hardship ................................................................................8
1.4 Mycobacterium ulcerans ................................................................................................10
1.5 Hypothesized Modes of Transmission...........................................................................11
1.6 Risk Factors ...................................................................................................................14
1.7 Landscape Epidemiology/Ecology, GIS, and Remote Sensing .....................................16
1.8 Spatial Autocorrelation ..................................................................................................21
1.9 Past BU Landscape Studies ...........................................................................................22
1.10 Objectives ....................................................................................................................27
1.11 Hypotheses ...................................................................................................................28
Chapter 2 MATERIALS AND METHODS ................................................................................. 30
2.1 Study Area .....................................................................................................................30
2.2 BU Case Data ................................................................................................................32
2.3 Satellite imagery and derivatives ...................................................................................33
2.4 Statistical Modeling .......................................................................................................39
2.5 Landscape-level Configuration Analysis .......................................................................39
2.6 Landscape-level Configuration Analysis Predictor Variables .......................................42
2.7 Class-level Configuration Analysis ...............................................................................48
2.8 Class-level Configuration Analysis Predictor Variables ...............................................49
2.9 Spatial GLM ..................................................................................................................52
2.10 Model Verification.......................................................................................................54
2.11 Risk Surface .................................................................................................................55
Chapter 3 RESULTS..................................................................................................................... 58
3.1 Non-Spatial vs. Spatial Model Results ..........................................................................58
3.2 Model Output Interpretation ..........................................................................................60
3.3 Landscape-level Configuration Analysis Results ..........................................................61
3.4 Class-level Composition and Configuration Analysis ...................................................64
3.5 Model Verification.........................................................................................................71
3.6 Predictive Surface Model ..............................................................................................75
Chapter 4 DISCUSSION AND LIMITATIONS .......................................................................... 77
4.1 Discussion ......................................................................................................................77
4.2 Limitations .....................................................................................................................86
4.3 Conclusions and Future Directions ................................................................................88
vi

APPENDICES .............................................................................................................................. 90
Appendix A ............................................................................................................... 91
Appendix B ............................................................................................................. 112
Appendix C ............................................................................................................. 126
Appendix D ............................................................................................................. 149
LITERATURE CITED ............................................................................................................... 154

vii

LIST OF TABLES
Table 2-1. Land use and cover classification categories............................................................... 35
Table 2-2.Initial landscape metrics calculations. .......................................................................... 41
Table 2-3. Class-level initial landscape metric candidates ........................................................... 48
Table 3-1. Non-spatial and spatial GLM DIC values. .................................................................. 58
Table 3-2. Non-spatial vs. spatial binomial GLM variable significance. ..................................... 59
Table 3-3. 1_2k number of patches spatial binomial GLM model summary ............................... 61
Table 3-4. Gelman diagnostic 1_2k patch richness density model ............................................... 61
Table 3-5. 1_6k number of patches spatial binomial GLM model summary ............................... 62
Table 3-6. Gelman diagnostic 1_6k numbers of patches model ................................................... 62
Table 3-7. 2k numbers of patches spatial binomial GLM model summary.................................. 63
Table 3-8. Gelman diagnostic 2k number of patches model ........................................................ 63
Table 3-9. 2k patch richness density spatial binomial GLM summary ........................................ 64
Table 3-10. Gelman diagnostic 2k patch richness density model ................................................. 64
Table 3-11. 800m wetland class spatial model results .................................................................. 65
Table 3-12. 800m wetland class convergence diagnostic ............................................................. 65
Table 3-13. 800m agriculture/forest class spatial model results ................................................... 65
Table 3-14. 800m agriculture/forest class convergence diagnostic .............................................. 66
Table 3-15. 1_2k forest spatial model results ............................................................................... 66
Table 3-16. 1_2k forest class convergence diagnostic.................................................................. 67
Table 3-17. 1_2k agriculture/forest spatial model results ............................................................. 67
Table 3-18. 1_2k agriculture/forest class convergence diagnostic ............................................... 68
Table 3-19.1_6k agriculture/forest spatial model results .............................................................. 69
viii

Table 3-20. 1_6k agriculture/forest class convergence diagnostic ............................................... 69
Table 3-21. 2k forest class spatial model results .......................................................................... 70
Table 3-22. 2k forest class model convergence diagnostic........................................................... 70
Table 3-23. 2k agriculture/forest class spatial model results ........................................................ 71
Table 3-24. 2k agriculture/forest model convergence diagnostic ................................................. 71
Table 3-25. Root mean square errors for all spatial binomial GLM training models ................... 72
Table A-1. Landscape-level 400m correlation matrix .................................................................. 92
Table A-2. Landscape-level 500m correlation matrix .................................................................. 93
Table A-3. Landscape-level 600m correlation matrix .................................................................. 94
Table A-4. Landscape-level 700m correlation matrix .................................................................. 95
Table A-5. Landscape-level 800m correlation matrix .................................................................. 96
Table A-6. Landscape-level 900m correlation matrix .................................................................. 97
Table A-7. Landscape-level 1k correlation matrix ....................................................................... 98
Table A-8. Landscape-level 1_1k correlation matrix ................................................................... 99
Table A-9. Landscape-level 1_2k correlation matrix ................................................................. 100
Table A-10. Landscape-level 1_3k correlation matrix ............................................................... 101
Table A-11. Landscape-level 1_5k correlation matrix ............................................................... 102
Table A-12. Landscape-level 1_6k correlation matrix ............................................................... 103
Table A-13. Landscape-level 1_7k correlation matrix ............................................................... 104
Table A-14. Landscape-level 1_8k correlation matrix ............................................................... 105
Table A-15. Landscape-level 1_9k correlation matrix ............................................................... 106
Table A-16. Landscape-level 2k correlation matrix ................................................................... 107
Table A-17. 800m forest class correlation matrix ...................................................................... 108

ix

Table A-18. 800m wetland class correlation matrix ................................................................... 108
Table A-19. 800m agriculture/forest class correlation matrix .................................................... 108
Table A-20. 1_2k forest class correlation matrix........................................................................ 109
Table A-21. 1_2k wetland class correlation matrix .................................................................... 109
Table A-22. 1_2k agriculture/forest class correlation matrix ..................................................... 109
Table A-23. 1_6k forest class correlation matrix........................................................................ 110
Table A-24. 1_6k wetland class correlation matrix .................................................................... 110
Table A-25. 1_6k agriculture/forest class correlation matrix ..................................................... 110
Table A-26. 2k forest class correlation matrix............................................................................ 111
Table A-27. 2k wetland class correlation matrix ........................................................................ 111
Table A-28. 2k agriculture/forest class correlation matrix ......................................................... 111
Table C-1. Landscape-level non-spatial binomial GLMs ........................................................... 127
Table C-2. Landscape-level non-spatial binomial GLMs ........................................................... 129
Table C-3. Landscape-level non-spatial binomial GLMs ........................................................... 131
Table C-4. Landscape-level non-spatial binomial GLMs ........................................................... 133
Table C-5. Landscape-level non-spatial binomial GLMs ........................................................... 135
Table C-6. 800m forest class non-spatial binomial GLMs ......................................................... 137
Table C-7. 800m wetland class non-spatial binomial GLMs ..................................................... 138
Table C-8. 800m agriculture/forest class non-spatial binomial GLMs....................................... 139
Table C-9. 1_2k forest class non-spatial binomial GLMs .......................................................... 140
Table C-10. 1_2k wetland class non-spatial binomial GLMs..................................................... 141
Table C-11. 1_2k agriculture/forest class non-spatial binomial GLMs ...................................... 142
Table C-12. 1_6k forest class non-spatial binomial GLMs ........................................................ 143

x

Table C-13. 1_6k wetland class non-spatial binomial GLMs..................................................... 144
Table C-14. 1_6k agriculture/forest class non-spatial binomial GLMs ...................................... 145
Table C-15. 2k forest class non-spatial binomial GLMs ............................................................ 146
Table C-16. 2k wetland class non-spatial binomial GLMs......................................................... 147
Table C-17. 2k agriculture/forest class non-spatial binomial GLMs .......................................... 148
Table D-1. 800m forest class spatial binomial GLM .................................................................. 150
Table D-2. 800m wetland class spatial binomial GLM .............................................................. 150
Table D-3. 800m agriculture/forest class spatial binomial GLM ............................................... 150
Table D-4. 1_2k forest class spatial binomial GLM ................................................................... 151
Table D-5. 1_2k wetland class spatial binomial GLM ............................................................... 151
Table D-6. 1_2k agriculture/forest class spatial binomial GLM ................................................ 151
Table D-7. 1_6k forest class spatial binomial GLM ................................................................... 152
Table D-8. 1_6k wetland class spatial binomial GLM ............................................................... 152
Table D-9. 1_6k agriculture/forest spatial binomial GLM ......................................................... 152
Table D-10. 2k forest class spatial binomial GLM ..................................................................... 153
Table D-11. 2k wetland class spatial binomial GLM ................................................................. 153
Table D-12. 2k agriculture/forest class spatial binomial GLM .................................................. 153

xi

LIST OF FIGURES
Figure 1-1. BU stages. (Images Courtesy of Dr. K. Asiedu and Dr. .............................................. 6
Figure 1-2. BU Complications (Images Courtesy of the ............................................................... 7
Figure 2-1. Togo and Benin administrative districts included in study area ................................ 31
Figure 2-2. Benin 2004 and 2005 BU case data subset ................................................................ 33
Figure 2-3. Land use and land cover classification....................................................................... 36
Figure 2-4. Concentric buffers example ....................................................................................... 37
Figure 2-5. Wetness index map. ................................................................................................... 38
Figure 2-6. Methods and approaches used for final model generation ......................................... 40
Figure 2-7. Shape index mean surface .......................................................................................... 44
Figure 2-8. Number of patches surface ......................................................................................... 45
Figure 2-9.Wetness index average surface ................................................................................... 46
Figure 2-10. Patch richness density surface .................................................................................. 47
Figure 2-11. Percent land cover adjacency surface 2k wetland class ........................................... 50
Figure 2-12. 1_2k agriculture/forest class landscape shape.......................................................... 51
Figure 2-13. New locations and observed locations used to create risk surface........................... 56
Figure 3-1. 1_2k wetland class observed rates, 90% data ............................................................ 73
Figure 3-2. 1_2k wetland class fitted model, 90% data ................................................................ 74
Figure 3-3. 1_2k wetland class spatial random effects ................................................................. 75
Figure 3-4. Predictive BU rates surface across the study domain ................................................ 76
Figure 4-1. Anthropogenic disturbance within wetland land cover patch .................................... 81
Figure B-1. Landscape-level “best model” variable scatterplots ................................................ 113
Figure B-2. 800m “best model” variable scatterplot .................................................................. 114

xii

Figure B-3. 800m wetland “best model” variable scatterplot ..................................................... 115
Figure B-4. 800m agriculture/forest class “best model” variable scatterplot ............................. 116
Figure B-5. 1_2k forest class “best model” variable scatterplot ................................................. 117
Figure B-6. 1_2k wetland class “best model” variable scatterplot ............................................. 118
Figure B-7. 1_2k agriculture/forest class “best model” variable scatterplot .............................. 119
Figure B-8. 1_6k forest class “best model” variable scatterplot ................................................. 120
Figure B-9. 1_6k wetland class “best model” variable scatterplot ............................................. 121
Figure B-10. 1_6k agriculture/forest class “best model” variable scatterplot ............................ 122
Figure B-11. 2k forest class “best model” variable scatterplot ................................................... 123
Figure B-12. 2k wetland class “best model” variable scatterplot ............................................... 124
Figure B-13. 2k agriculture/forest class “best model” variable scatterplot ................................ 125

xiii

LIST OF EQUATIONS
Equation 2-1 .................................................................................................................................. 38
Equation 2-2 .................................................................................................................................. 42
Equation 2-3 .................................................................................................................................. 43
Equation 2-4 .................................................................................................................................. 44
Equation 2-5 .................................................................................................................................. 47
Equation 2-6 .................................................................................................................................. 49
Equation 2-7 .................................................................................................................................. 51
Equation 2-8 .................................................................................................................................. 55

xiv

INTRODUCTION
Anthropogenic ecosystem disturbances, for example, land use change, human movement,
encroachment and wildlife translocation, rapid transport, and climate change play an important
role in the distribution of emerging and re-emerging infectious diseases (Wilcox and Gubler,
2005; Patz et al., 2000; Foley et al., 2005; Patz and Confalonieri, 2005; Confalonieri, 2005).
These activities have the potential to disrupt natural community assemblages in ecosystems,
impacting predator/prey relationships, thereby resulting in an imbalance of population control
mechanisms that may have prevented disease pathogens from infecting human populations prior
to disturbance (Wilcox and Gubler 2005; Morse 1995).
Land use and land cover (LULC) change at multiple scales are one anthropogenic
disturbance linked to disease incidence. Examples include: 1) changes in dry-season irrigation
practices at a local scale that created new mosquito-breeding habitats leading to elevated Ross
River virus cases in Australia; 2) deforestation of hillsides at a regional scale contributing to
nutrient run-off that proliferated macroalgae growth suitable to bacteria responsible for ciguatera
fish poisoning; and 3) rainforest degradation at a global scale resulting in new transmission
opportunities for simian retroviruses such as HIV/AIDS between human and non-human
primates (Cook et al., 2004).
Anthropogenic activities with major impacts on LULC are land degradation, including
agriculture intensification and water projects, urbanization, and deforestation (Patz et al., 2008).
These activities contribute to habitat fragmentation that can disrupt vector breeding sites and
reservoir distributions while creating habitats suitable to ecological edge effects that provide
opportunities for niche invasions that promote disease emergence (Patz and Confalonieri, 2005).
Further, these activities generate new pathways for humans to interact with environments
1

undisturbed previously, resulting in reduced proximities to potential vectors, reservoirs, and
isolated pathogens (Patz and Confalonieri, 2005; Morse, 1995; Epstein, 2002; Pongsiri et al.,
2009).
Landscapes impacted most by widespread anthropogenic ecosystem disturbance are those
situated within the tropics, which have experienced the largest ecosystem transformations in
history within the last 60 years (Clark et al., 1990). Agriculture intensification is the major driver
behind tropical deforestation, and as population growth rates continue to rise, agricultural land is
expected to increase by approximately 23% by the year 2050 to meet food and fuel demands
(Patz et al., 2008; Hansen et al., 2008). As these regions continue to transition from natural to
managed ecosystems, new habitat opportunities suitable to a variety of reservoirs, vectors, and
pathogens will continue to surface; therefore, identifying landscape patterns favorable to disease
emergence will be a critical first step in human disease prevention.
Advances in GIS and remote sensing technologies, along with spatial statistical methods,
facilitate the observation and quantification of anthropogenic landscape disturbances, providing
the tools necessary to study location-specific landscape characteristics that might be responsible
for disease incidence while enabling spatial modeling capabilities to link these characteristics
and to predict disease risk across a landscape (Kitron, 1998). While current research focuses
largely on wildlife and vector-borne zoonotic disease emergence, exploring linkages between
anthropogenically-disturbed landscapes and human bacterial infections is an underrepresented
area of research, although recent results suggest that these disturbances play an important role in
spatial distributions of environmental bacteria that pose a human health risk (Goldberg et al.,
2008). Quantifying landscape patterns related to bacterial disease emergence is central to the

2

prediction of present and future public health risk because often the ecological drivers behind
these diseases are poorly understood.
One example is Buruli ulcer (BU) disease, caused by the environmental pathogen
Mycobacterium ulcerans. Although the ecological drivers behind MU growth remain a mystery,
dramatic increases in BU cases since the 1980s (Merritt et al., 2005) have been linked
empirically and anecdotally to anthropogenic landscape changes; for example, deforestation,
habitat fragmentation, aquatic ecosystem disturbances from dam construction and agriculture
irrigation, farming practices, and mining activities (Merritt et al., 2010). Although BU is not
transferred between persons, the mode or modes of transmission is not determined, and no
vaccine exists (Wansbrough-Jones and Phillips, 2006). Therefore, identifying landscape patterns
linked to BU incidence will provide a powerful tool for surveillance and prevention while
affording opportunities to learn more about the disease system.

3

Chapter 1 LITERATURE REVIEW

1.1 Buruli ulcer disease
Mycobacterium ulcerans (MU), the causative agent of BU, is an environmental pathogen
(Williamson et al., 2008) and the second most common mycobacterium infection in humans after
tuberculosis and leprosy (Wansbrough-Jones and Phillips, 2006; Walsh et al. 2008; WHO 2000).
BU presents as a necrotizing skin condition that often causes mobility loss and permanent
disabilities in patients due to the potential for ulcers to cover large areas of the body. The disease
is endemic in over 32 countries worldwide, but occurs predominantly in tropical areas of subSaharan Africa and also in more temperate regions; specifically, the Melbourne area of Victoria,
Australia (WHO, 2007; Walsh et al., 2008). The World Health Organization (WHO) established
the Global Buruli Ulcer Initiative in 1998 with the goals of raising disease awareness, improving
treatment access, strengthening surveillance, and providing priority research into disease
diagnosis, treatment and prevention (WHO, 2007). BU affects all age groups, but children under
the age of 15 are most at risk for the disease (Merritt et al., 2005).
MacCallum and colleagues identified MU as the causal agent of BU in Victoria, Australia
in 1948 (MacCallum et al., 1948), although Sir Albert Cook in Africa in 1897 and Kleinshmidt
in northeast Congo during the 1920s identified likely BU cases (Johnson et al., 2005). BU has
had numerous names worldwide, each representing the region from which cases occurred. In
Australia, it was known as Bairnsdale ulcer after the town in which the identification of MU took
place and Searls’ ulcer after a physician practicing in Bairnsdale during the same era; the
infection is named Kumusi in Papua New Guinea and Kakerifu in Zaire (Meyers, 2007). The
name Buruli is linked to the former Buruli district in Uganda where a large number of cases
4

surfaced between 1950 and 1970 (Clancey et al., 1964), and it is the name adopted by the WHO
to promote disease awareness (Meyers, 2007).

1.2 Symptoms and Treatment
Three BU disease stages exist: non-ulcerative or pre-ulcerative disease, ulcerative
disease, and healing or scarring (Figure 1-1, WHO, 2000). Pre-ulcerative symptoms include
painless, raised papules that are usually < 1cm in diameter; painless, mobile nodules beneath the
skin; and firm plaques (WHO 2000). Edema and swelling can also occur, but subsides once an
ulcer develops (Dobos et al., 1999). An incubation period of approximately three months exists
before ulcer eruption, but varies by individual (Horsbough and Meyers, 1997) with as little as
two weeks to as long as one year being reported by individuals in contact with endemic regions
(Veitch et al., 1997).
The ulcerative disease stage develops when pre-ulcerative symptoms erupt into necrotic
skin lesions. Ulcer edges are deeply undermined and necrosis often extends beyond the visible
infection (Johnson et al., 1999; Hayman J, 1993). Ulcers are usually painless and continue to
enlarge through necrosis of the underlying epidermis surrounding the initial lesion (Huygen et
al., 2009). Healing is a slow process that may occur spontaneously, but scarring of affected
tissues leads to debilitating conditions. Four methods of laboratory diagnosis exist, including
direct smear, culture, Polymerse Chain Reaction (PCR), and histopathology (Walsh et al., 2008).

5

Figure 1-1. BU stages. (Images Courtesy of Dr. K. Asiedu and Dr.
A. Chauty <http://www.who.int/buruli/photos/en/index.html>). For
interpretation of the references to color in this and all other figures,
the reader is referred to the electronic version of this thesis.
Approximately 80% of ulcers present on the limbs or extremities (Walsh et al., 2008),
and no significant difference exists between male and female infection rates (Debacker et al.,
2004b). A country-wide case study in Ghana found that women and girls were more likely than
boys to develop lesions on their arms, but ulcerations on the legs were still dominant in both
sexes (Amofah et al., 2002).
Recent treatment advances determined that a combination of Streptomycin-Rifampin
antibiotics alone had a high success rate for persons presenting with ulcers < 15 cm in diameter
(Chauty et al., 2007), but larger ulcers must be surgically excised along with a large area of
surrounding, healthy tissue followed by skin grafting to help prevent disease recurrence, and
severe cases require amputation of ulcerated limbs (Sizaire et al., 2006). Although BU eventually
subsides on its own, it is not without devastating, long-term negative effects, and disease
recurrence rates range between 6.1% and 47% in some endemic regions (Debacker et al., 2005;
Amofah et al., 2002).
Although study results suggested that the M. bovis bacilli Calmette-Guérin, or BCG,
vaccine used commonly to protect against tuberculosis provided partial protection against BU for
a six month time period (Portaels et al., 2002), a definitive vaccine does not exist (Huygen et al.,
2009). However, the BCG vaccine may prevent more severe infections from developing in
6

patients (Portaels et al., 2002; Noeske et al., 2004) and may decrease disease duration (Amofah
et al., 1993). While these results are promising, contrasting results also exist, suggesting that the
BCG vaccination provides no BU protection and may increase infection risk (Nackers et al.,
2006). Persons infected with the human immunodeficiency virus (HIV) are not at a higher risk
for contracting BU (Raghunathan et al., 2005), but may experience a more severe infection once
inoculated (Johnson et al., 1999).
BU complications include mobility loss from surgical treatment and/or scarring of
affected tissues. Without adequate physical therapy, patients often experience contracture, or the
inability to straighten or flex muscles in affected areas (Hayman 1993). In severe cases, the
infection moves into the bone causing osteomyelitis (Johnson et al, 1999). In these cases, the
bone develops a “moth-eaten” appearance and severe disability may result (WHO, 2007).
Reactive osteitis, or contiguous osteitis, is another BU complication that results from soft tissue
destruction above the bone and also creates the potential for mobility loss in patients (WHO,
2007). A recent study in Ghana found that at least 27% of patients experienced a reduced range
of motion following treatment and the study suggested that better follow-up care should be a
priority to help limit disability from the disease (Figure 1-2, Schunk et al., 2009).

Figure 1-2. BU Complications (Images Courtesy of the
National Buruli ulcer Control Programme, Benin, Professor
H. Assé, and Dr. K. Asiedu
<http://www.who.int/buruli/photos/en/index.html)
7

1.3 Prevalence and Economic Hardship
BU cases increased steadily and expanded geographically within the past few decades
(Merritt et al., 2005). Although elevated disease awareness contributed to higher diagnosis rates,
researchers believe an increase in the exposed population also took place (Debacker et al.,
2004a). The WHO estimated approximately 24,000 cases were recorded between 1978 and 2006
in Côte d’Ivoire; approximately 7,000 cases were recorded between 1989 and 2006 in Benin; and
more than 11,000 cases were recorded since 1993 in Ghana (WHO 2007). Underreporting is a
common problem when estimating BU case numbers because many patients do not have access
to health care facilities, are misdiagnosed, or live in countries in which BU is not a notifiable
disease (WHO 2007).
BU cases increased dramatically in the Melbourne area of Victoria, Australia with only 4
cases reported in 1999 but 61 cases reported in 2006 (Quek et al., 2007) with an outbreak of 83
cases occurring in the city of Point Lonsdale on the Bellarine Peninsula in 2002 (Johnson et al.,
2007). Between 1992 and 1995, an outbreak occurred in the city of Cowes located on Phillips
Island off the mainland of Australia near Melbourne (Veitch, 1997). Although BU cases were
recorded since the 1930s in Melbourne suburbs, previous BU cases on Phillips Island had not
been recorded, suggesting that the pathogen expanded its geographic range. BU is also endemic
to the Daintree Forest region of Northern Queensland along the eastern coast of Australia,
though fewer cases are reported each year than in the Victoria region (Jenkin et al., 2002). Early
detection and access to health care, along with fewer overall cases, has prevented BU from
becoming a more serious public health problem in Australia, but persons living in developing
countries face greater challenges when seeking treatment, and the majority of cases received at
medical facilities are in advanced disease stages (WHO, 2000).

8

Persons presenting with BU often require extensive surgery and hospital stays, creating
economic challenges for families of patients (Aujoulat et al., 2003); for example, patient hospital
stays average approximately three months in Ghana (Stienstra et al., 2002). The WHO reported
in Ghana that the cost of treating a pre-ulcerative nodule in a patient is comparable to 16% of the
average total family income for the work-year; the cost of treating a patient requiring amputation
is equal to 89% of the average total family income for the work-year; and the average cost of
treating a patient in 1994-1996 exceeded the per capita government spending on health care
(WHO, 2007). Compounding the problem is income loss due to incapacitation from the disease
or from caring for a family member with BU, and families face additional challenges involving
prolonged care for members permanently disabled. Fear of surgery, long distances to treatment
facilities, and socioeconomic factors contribute to the rural poor being affected most negatively
by BU (Raghunathan et al., 2005).
Beyond prohibitive treatment costs and access to medical facilities, a stigma associated
with BU exists in many regions due to a belief that the infection is a result of witchcraft. A case
controlled study in Ghana determined that 59% of study respondents felt that witchcraft was
involved in disease development (Stienstra et al., 2002). Persons without BU reported avoiding
patients suffering from the disease, and patients reported hiding symptoms from community
members. Many respondents attributed the disease to a lack of hygiene by patients, and patients
reported a fear of the “evil eye” and did want community members staring at wounds or
watching dressing changes for fear that these activities would influence healing negatively.
Importantly, many respondents did not believe that BU was a “hospital disease” because the
infection was the result of a curse, and therefore, traditional healing methods would be the best
treatment option.

9

1.4 Mycobacterium ulcerans
MU is a slow-growing, mycolactone-producing environmental mycobacterium (Merritt et
al., 2005). During the pre-ulcerative stage, acid-fast bacilli localize to fatty tissues under the skin
where microcolony formation takes place (Dobos et al., 1999). MU is unusual because of its
extracellular activity and mycolactone secretion, a lipid toxin responsible for fat and
subcutaneous tissue necrosis (George et al., 1999) The mycolactone secretes an
immunosuppressive property, indicated by the lack of a granulomatous inflammatory response in
early lesion stages, allowing the disease to progress with little or no pain in the patient (Johnson
et al., 1999). Lack of a host immune response leaves the infection undetected by burulin skin
tests that become positive only after the bacterium ceases to reproduce and healing begins
(WHO, 2007).
Additional attributes set MU apart from related mycobacteria. Most notably, MU prefers
lower temperatures and has a narrower temperature range than its counterparts, preferring
temperatures between 28-34°C with an optimal growth range between 30-33°C in a laboratory
setting (Merritt et al., 2010). Although isolates from various regions show differences in
temperature tolerance, for example, isolates from Africa are more thermo-tolerant than those
from more temperate regions, survival at higher temperatures takes place without reproduction
(Edyanni and Portaels, 2007). Further, MU is intolerant to ultra-violet (UV) light because it lacks
pigments found in related mycobacterium, indicating that MU may be protected from sunlight in
its natural environment (Stinear et al., 2007).
Recent technological advances enabled researchers to detect MU DNA in a variety of
environmental samples ranging from soils to invertebrates (for a complete review see Merritt et
al., 2010) using PCR probes based on detection of IS2404 (Ross et al., 1997). Although IS2404

10

was believed to be specific to M. ulcerans, it was later determined that other mycobacteria also
tested IS2404 positive; specifically, M. marinum, M. liflandji, and M. pseudoshottsii (Stragier et
al., 2007), but the advent of real-time PCR provided the tools necessary to differentiate MU from
other mycobacteria testing IS2404 positive (Fyfe et al., 2007).
A significant contribution involving a pure MU culture from an environmental sample
took place recently in Benin (Portaels et al, 2008). The culture was obtained from a water strider,
an aquatic insect (Hemiptera: Gerridae, Gerris sp.). This discovery made a substantial
contribution to BU research because it was the first of its kind and confirmed MU existence in
aquatic habitats (Merritt et al., 2010).

1.5 Hypothesized Modes of Transmission
The mode or modes of BU transmission are currently unknown. The occurrence of ulcer
presentation at the site of previous skin trauma in some patients suggests a direct mode of
transmission from the environment (Meyers et al., 1974). Ulcer presentation in a patient in Togo,
Africa was reported at the site of a wound from a large splinter of wood (Meyers et al., 1996),
and cleaning and covering cuts and abrasions is recommended in Victoria, Australia to help
prevent inoculation with the disease (Johnson et al., 2007).
Aerosol transmission is another mechanism proposed for MU inoculation based on the
ability of Mycobacteria intracellulare to concentrate in air bubbles, suggesting that mycobacteria
may be propelled into the air and dispersed by wind or waves upon gas bubble bursts occurring
in water bodies in the natural environment (Hayman, 1991). Further speculation ensued when a
study in Australia revealed that BU patients did not report having contact with environmental
water other than the ocean before presenting with the disease, but a need exists for additional
research to determine the plausibility of this hypothesis (Johnson et al., 2005).
11

Researchers found MU DNA in biofilm of aquatic vegetation in Ghana, and subsequent
testing determined that MU growth rates increased substantially after two green algae types were
entered into the medium in a laboratory setting (Marsollier et al., 2004b). Persons may come into
direct contact with contaminated water and plants while bathing, farming or conducting domestic
duties, or contaminated insects may act as mechanical vectors of the disease.
Past studies focused on the possibility of aquatic insects as primary BU vectors. In 1999,
researchers investigated five aquatic invertebrate species found on the roots of various plants in
Benin and Ghana and found they were MU positive, but no evidence surfaced of these insects
passing the pathogen to humans, even though they were known to sometimes bite villagers
(Portaels et al., 1999). Further research determined that MU could survive and reproduce within
the salivary glands of biting aquatic bugs (Marsollier et al., 2003), and scientists observed ulcer
presentation in mice bitten by infected aquatic bugs in the laboratory, although insect species
native to West Africa were not used in these studies (Marsollier, 2002). Additional research
demonstrated that aquatic snails may act as passive hosts, harboring MU within their bodies after
eating contaminated aquatic vegetation, and that MU may be transferred to the salivary glands of
biting insects after feeding on infected snails (Marsollier et al. 2004a).
Contrasting results from a field study involving 27 water body sites near BU positive and
BU negative villages in Ghana determined that biting Hermiptera were not likely BU vectors
after finding no positive correlation between Hemiptera abundance and endemic sties; MU
positivity rates and endemic sites; MU positivity rates among Hemiptera versus other aquatic
insects within endemic sites; or overall MU positivity and vector abundance (Benbow et al.,
2008). However, a recent field study conducted in Cameroon (Marion et al., 2010) supported
previous laboratory results, finding MU within biting hemiptera salivary glands, and contrary to

12

field results from Ghana, found insect density counts ten times higher in BU endemic study sites
than in BU non-endemic sites.
Mosi et al. (2008) investigated MU colonization within the salivary glands of African
giant water bugs (Belostomatidae) through MU-infected mosquito larvae ingestion in the
laboratory. Results determined that MU preferred to accumulate on the exoskeleton as opposed
to the guts or salivary glands, but persistent colonization without replication could occur, and
therefore, these insects were capable of supporting the trophic transfer of MU within the
environment.
Researchers in Australia investigated the role of mosquitoes in potential BU transmission
after discovering that areas with BU outbreaks also had a large mosquito population (Johnson et
al., 2007). Although mosquitoes tested MU positive, it was inconclusive whether they were
harboring MU within their bodies or whether they had acquired the pathogen externally from the
environment. A recent study conducted in the laboratory determined that mosquito larvae fed
MU-contaminated materials accumulated the bacterium in their mouths and midgets over four
instars, although when fed closely-related M. marinum, bacterium was not detected (Tobias et
al., 2009). Additional research identified a correlation over a seven year time period between BU
case notifications and Ross River virus case notifications, a mosquito-transmitted disease in
Australia, although the authors conceded that suitable environmental conditions promoting both
diseases may be driving the correlation (Johnson and Lavender, 2009).
Several investigations into potential MU reservoirs exist. Mammals and marsupials in
Australia, including koalas, alpaca, a domestic cat, potaroos, two horses, black rats, and brushtail
and ringtail possums tested MU positive, but these infections were limited within the geographic
region of Australia (Mitchell et al., 1984; Elsner et al., 2008; van Zyl et al., 2009). A recent

13

study in Australia found that 38% of captured ringtail possums and 24% of captured brushtail
possums had lesions testing MU positive or feces testing MU positive in a BU endemic region,
suggesting that these possums contribute to maintenance of the bacterium in the environment
(Fyfe et al., 2010). An additional study in Benin collected 326 rodents and 222 shrews around
water bodies and in houses during the dry season and during the wet season in high BU endemic
and low BU endemic areas, but MU was not identified in any of the collection samples (Durnez
et al., 2010). Finally, some researchers believe that migratory birds aid in disseminating MU
between wetlands, but further research is needed to determine the potential of this hypothesis
(Eddyani et al., 2004).

1.6 Risk Factors
The greatest risk factor for contracting BU is contact with an endemic area (Johnson et
al., 2007), although certain behavioral patterns may increase risk. A matched case-control study
in three endemic districts of Ghana confirmed researchers’ beliefs that BU is an
environmentally-acquired infection associated with rivers and streams (Raghunathan et al.,
2005). The results also indicated that wading in a river or stream was a risk factor, that using
certain soap products and wearing long trousers was protective against BU, and that wearing
clothing that covered the upper body while farming was also protective against the disease. In a
separate study conducted in Benin, wearing clothing during farming activities was not identified
as a protective factor against BU infection, contact with stagnant water increased BU risk, while
contact with flowing water and using soap decreased BU risk (Nackers et al., 2007). Aiga et al.
(2004) found that swimming in rivers and water use from rivers was a BU risk factor, but water
use from piped sources was not a protective factor in Ghana, while Marston et al. (1995) found
that swimming in rivers was not a BU risk factor Côte d’Ivoire.
14

In Cameroon, the use of bed nets was a strongly-associated protective factor against BU,
but bed nets were not found to be protective in Ghana (Pouillot et al., 2007). Applying insect
repellant during outdoor activities reduced BU risk in Australia (Quek et al., 2007), although
mosquito coil use was a behavioral risk factor in Cameroon (Pouillot et al., 2007). An
association between BU-infected patients and mosquito bites on the arms and legs exists in
Australia, but a direct link between insect bites and inoculation was not established (Quek et al.,
2007).
BU cases are known to occur surrounding slow-moving or stagnant water bodies or
wetlands, and flooding is cited as a risk factor for the disease (CDC, 2005). The first recorded
BU cases in Bairnsdale, Australia occurred two to three years after a major flooding event during
Christmas of 1935 (Hayman, 1991). In 1978, Bairnsdale had its wettest year in recorded history,
and in 1980 BU cases emerged. In 1962 and 1964 severe flooding took place in Uganda, and
approximately two to three years later, BU cases surfaced (Dobos et al., 1999). The first reported
BU cases in Togo occurred in two children living near two separate rivers that experienced
seasonal flooding (Meyers et al., 1996), and in Papua New Guinea, cases occur near the Sepik
and Kumusi Rivers where a spike in cases emerged following the flooding and widespread
destruction from the Mount Lamington volcanic eruption in 1951 (WHO, 2000).
Empirical and anecdotal links exist between landscape alterations, including
deforestation, agriculture irrigation, mining activities, and artificial dam creation and increased
BU risk (Hayman 1991; Brou et al., 2009; Duker et al., 2004; Wagner et al., 2008a; Wagner et
al., 2008b; WHO, 2000; Merritt et al., 2005). Hayman (1991) postulated that closed rainforests
contain MU, and disruption of the rainforest results in MU-contaminated runoff water, creating
more concentrated MU quantities in water bodies, resulting in a bacteria bloom under favorable

15

environmental conditions. Deforestation contributes to increased flooding and erosion, creating
ideal conditions for nutrient introduction into water bodies that contribute to higher water
temperatures, decreased oxygen levels, and an increase in turbidity, all of which create a habitat
that may be suitable to MU (Merritt et al., 2005). Duker et al. (2004) found a relationship
between the spatial distribution of BU cases and distance to gold mining sites and exposure to
arsenic-contaminated soils in Ghana, and Wagner et al. (2008b) found that villages experiencing
high BU rates were situated within low-lying areas at risk for flooding and surrounded by an
agriculture matrix, suggesting that deforestation likely took place.
Several associations exist between artificial water body creation and river damming and
increased BU risk. Students on the campus of the University of Ibadan in Nigeria began
contracting BU after the creation of a dam on a small river running through campus in order to
create an artificial lake (Oluwasanmi, 1975). BU cases also occurred following the building of
large dams for agriculture irrigation north of Brisbane, Australia between 1957 and 1958 and in
1962 (Abrahams, 1964), and a recent study in Côte d’Ivoire found a relationship between disease
rates and patients’ proximities to dams and irrigated agricultures (Brou et al., 2009). In Liberia,
cases emerged after the installation of a swamp rice field that replaced an upland rice field and
also in areas where the creation of dams expanded wetland areas (WHO, 2000).
Although little is known regarding MU growth and subsequent transmission, linkages
between human activities and BU emergence suggest that identification of landscapes disturbed
by anthropogenic activities may provide information needed to characterize risk while
contributing to a better understanding of potential ecological drivers behind disease emergence.

1.7 Landscape Epidemiology/Ecology, GIS, and Remote Sensing

16

Landscape epidemiology is a multidisciplinary field focused on abiotic and biotic
environmental conditions and their relationships to pathogen and vector ecology with an
emphasis on mapping and predicting potential disease distributions based on ecological
conditions (Pavlovsky, 1966; Ostfeld, 2005). Landscape epidemiology is nested partially within
the field of landscape ecology, or the study of spatial and temporal landscape patterns and their
interactions with ecological processes at multiple scales (Turner et al., 2001). Observing
landscape heterogeneity and its influences on ecological systems provides a substantial
advantage when working with environmental data, leading to a better understanding of
underlying mechanisms contributing ecosystem functions (Pickett and Cadenasso, 1995).
Emphases on spatial and temporal patterns are unique to landscape ecology in the broader
ecology discipline because spatial homogeneity across ecological systems was the standard
assumption until recently (Turner and Gardner, 1991; Wagner and Fortin, 2005). In general,
landscape ecological studies observe larger spatial extents than those used in more traditional
ecological studies with an understanding based on scale theory that finer-scale processes can act
as mechanisms driving landscape dynamics within broader-scale patterns and that these broaderscale patterns can act as constraints limiting finer-scale processes (Turner et al., 2001). Emphasis
on neighborhood context is important under this framework and acknowledgement that
environmental heterogeneity may occur at any spatial scale differentiates landscape ecology
approaches further from traditional ecology methods (Wagner and Fortin, 2005).
Landscape epidemiology studies incorporate landscape ecology approaches regularly
(Kitron, 1998), from identifying vector habitat distributions based on landscape composition to
determining where anthrax risk from Bacillus anthracis exists in the United States using climate
and soil variables (Beck et al., 2000; Malone, 2005; Coetzee et al., 2000; Blackburn et al., 2007).

17

Several of these studies utilize geographic information systems (GIS), remote sensing
technologies, and spatial analyses to identify where disease emergence is likely to take place.
GIS and remote sensing technologies provide the foundation from which to collect and
process data to use in landscape ecological studies. GIS are unique in their ability to combine
information from multiple sources with a shared location across a large geographic region. GIS
are techniques to input, store, retrieve, manipulate, analyze, and output data that have spatial
attributes associated with them (Otsfeld et al., 2005). Common outputs include maps displaying
features of interest, but the real power of GIS is that it provides a platform from which to
organize a multitude of data for statistical and pattern analyses, for example natural
environmental data, such as climate, LULC, and elevation data, along with anthropogenic
attributes, such as demographic, road network, and socio-economic data.
GIS facilitates vector and raster data formats (Clarke 2003). A vector data format consists
of points, lines and polygons to which information about specific attributes are stored, for
example address locations, road networks, or administrative boundaries, while a raster format
consists of a grid of equally-sized cells. Cell sizes are referred to as the spatial resolution of the
grid; for example, a 30m resolution grid consists of cells measuring 30m on each side,
representing an area equal to 900 square meters total. Each cell contains information such as
elevation, temperature, or land cover classification values, and the possibility exists for
calculations between grids when cells correspond to the same geographic locations.
Satellite sensors that monitor the Earth from space generate remotely-sensed satellite
imagery (Jensen and Hodgson, 2004). The majority of remotely-sensed satellite products are
available in raster format. A variety of products exist, each measuring reflectance from the
Earth’s surface across the electromagnetic spectrum (EMS) (Turner et al., 2003). These images

18

are useful particularly when observing land cover patterns, water turbidity, or when measuring
elevation.
Quantitative approaches characterizing landscape patterns often utilize landscape metric
calculations derived from raster images that are classified into different land cover categories
(McGarigal and Marks, 1995). Landscape metrics quantify landscape patterns based on
composition, or the presence of specific landscape components in a geographic area, and
landscape configuration, or the spatial arrangement of landscape components in an area (Turner,
2005; McGarigal and Marks, 1995). Outputs from these metrics are then interpreted and
analyzed statistically to gain a better understanding of patterns that may be influencing
ecological processes in a geographic area.
Of particular interest are landscape metrics that identify landscape disturbances linked to
anthropogenic activities. A disturbance is defined as “any relatively discrete event in time that
disrupts ecosystem, community, or population structure and changes resources, substrate
availability, or the physical environment” (Pickett and White, 1985). These disturbances include
habitat fragmentation, or a disruption in landscape connectivity (With et al., 1997) resulting from
large land cover expanses transforming into small patch sizes that become isolated by land cover
types alternative to the natural vegetation (Wilcove et al., 1986), and agriculture intensification,
which takes place when cultivated land is altered to produce high-yield crops through the use of
fertilizers and pesticides, increased irrigation methods, or mechanization (Matson et al., 1997).
Observing land cover patch sizes and shape patterns using landscape metrics, along with
landscape composition, can provide insight into human activities in a region; for example,
Krummel et al. (1987) found that deciduous forest patches with more uniform shapes, such as

19

squares or rectangles, indicated managed forest cover and forest cover adjacent to agricultural
land using a fractal dimension index.
A study conducted by Iverson (1988) assessing land cover changes over a160 year
period in Illinois used several landscape metrics, including number of patches, mean patch area,
perimeter-to-area ratio, and fractal dimension indices by land cover type, to characterize
anthropogenic influence on the landscape. Results determined that a relationship existed between
high perimeter-to-area ratios and the presence of road networks; natural wetlands had highlycomplex shapes because they were not confined to regularly-shaped boundaries imposed by
human activities; surface mines, quarries, and gravel pits had highly-regular shapes because of
human management; and forest cover patches often had regular shapes because of adjacency to
agriculture land cover.
A landscape index quantifying the number of landscape patches was used to characterize
habitat fragmentation in a study area in Iowa (Narumalani et al., 2004), and a study observing
forest clearing patterns in Oregon used a landscape metric that employed a weighted patch edge
calculation to compare forest patches in privately-owned land versus in publically-owned land
between 1972 and 1988 (Spies et al., 1994). Results determined that privately-owned land had a
smaller percentage of interior forest patch habitat compared to publically-owned land, suggesting
a more fragmented forest landscape.
Landscape epidemiology studies use landscape metrics frequently to relate landscape
patterns to disease incidence. Graham et al. (2004) determined that mean shape index values at
the landscape-level within concentric polygons surrounding disease incidence locations were
important when investigating human alveolar echinococossis in China. Brownstein et al. (2005)
quantified a link between forest fragmentation and tick densities and numbers of infected ticks

20

responsible for Lyme disease transmission to humans in Connecticut using observations of patch
size and patch isolation. Results identified a positive relationship between fragmentation and
entomologic risk, but a negative relationship between fragmentation and human disease
incidence.
A study investigating and predicting schistosomiasis-harboring mollusk densities in two
mountainous regions in China used an integrated approach, quantifying landscape patterns and
then employing Bayesian spatial modeling methods (Yang et al., 2008). Researchers analyzed
mean shape index values and Shannon’s evenness index values along with a variety of
environmental variables. Generation of non-spatial and spatial model results emphasized the
importance of using spatial statistical approaches when investigating natural phenomena. Final
results determined a positive relationship existed between mean shape index values and mollusk
densities, indicating that mollusks clustered in irregularly-shaped land cover patches such as
irrigation ditches. Using spatial statistical methods to predict mollusk densities took advantage of
information provided by landscape metric calculations while accounting for spatial
autocorrelation in the data, an important factor that can introduce bias and that many researchers
ignore when working with environmental data.

1.8 Spatial Autocorrelation
Tobler’s first law of geography states that everything is related to everything else, but
near things are more related than distant things (Tobler, 1970), or are spatially autocorrelated.
Ignoring positive spatial autocorrelation, a common phenomenon when working with ecological
data, can produce falsely-precise model estimates, the most common of which is to overestimate
regression coefficient significance leading to Type I errors by rejecting null hypotheses when
they are true (Legendre and Fortin, 1989; Wagner and Fortin, 2005; Lennon, 2000). These errors
21

lead not only to inaccurate parameter estimates contributing to erroneous model prediction, but
promote a misunderstanding of processes driving ecological patterns (Lennon, 2000).
Spatial autocorrelation in environmental data can occur for a variety of reasons, including
processes inherent to the organism under investigation or because of spatial dependency of the
organism to spatially-structured environmental drivers (de Knegt et al., 2010) Further, the
processes driving spatial dependence may occur at different scales than which derivation of data
observations and variables took place (Keitt, 2002). Omission of important covariates will bias

 estimations, and if these covariates have a spatial structure, spatial autocorrelation in model
residuals will likely result (Waller and Gotway, 2004). Failure to account for spatially
autocorrelated model residuals violates the assumption of independently and identically
distributed errors, leading to inflated degrees of freedom, resulting in the overestimation of
coefficient significance mentioned above (de Knegt et al., 2010; Legendre and Fortin, 1989;
Wagner and Fortin, 2005; Lennon, 2000; Keitt, 2002; Waller and Gotway, 2004).
Spatial modeling approaches, for example the one employed by Yang et al. (2008),
mitigate spatial dependency problems by accounting for spatial autocorrelation in the modeling
process. Several spatial modeling approaches exist, each of which was designed to incorporate
different data formats and to achieve different goals. For a review of spatial statistical models
and methods see Schabenberger and Gotway (2005).

1.9 Past BU Landscape Studies
GIS, remote sensing technologies, and statistical analyses were used in several BU
studies investigating environmental features related to disease incidence, but to our knowledge
studies quantifying landscape patterns using landscape metrics do not exist, and, with the

22

exception of Duker et al. (2006) and Wagner et al. (2008a, 2008b), these studies failed to
account for spatial autocorrelation among observations, variables, and model residuals.
A study investigating the relationship between arsenic-rich floodplains and BU incidence
in the Amasie West District of Ghana analyzed several human and environmental variables using
GIS and satellite imagery (Duker et al., 2004). BU incidence per settlement and settlement
population were mapped along with stream sediment sample locations and corresponding arsenic
concentration values. Farmlands lying below a specified elevation threshold were identified,
along with proximities to arsenic-enriched streams. A land use and land cover map generated
from Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) imagery
determined that 21% of total farmlands in the study area lie within arsenic-enriched areas.
Although the results revealed a correlation existed between BU cases and arsenic-enriched, lowlying farmlands, researchers did not determine whether the flooded lowlands or arsenic levels
contributed more to BU risk.
Duker et al. (2006) conducted another analysis in the Amansie West district of Ghana
targeting spatial patterns of BU incidence in the district. Stream sediment samples from 1992 and
37 ground water samples from 2003 were related to single year BU incidence reports obtained
from a 1999 national case search, along with BU average incidence from 2000-2003 for each
settlement within the district using a conditional autoregressive (CAR) statistical model. This
approach assumed that covariates from neighboring districts influenced BU prevalence in the
target district, and districts were defined as regions bordering each other or those that were
within a fixed 15km distance from one another. Results suggested that a relationship existed
between mean arsenic levels in soil and BU spatial distributions and distances to gold mining
sites and BU spatial distributions.

23

The RESPOND Health Mapping Activity Ivory Coast study investigated BU occurrence
in the Ivory Coast (Côte d’Ivoire) using remote sensing and GIS to identify areas with a high BU
risk (RESPOND 2006). Landsat satellite imagery along with Moderate Resolution Imaging
Spectroradiometer (MODIS) data were used to classify vegetation and wetland conditions. A
digital elevation model using Shuttle Radar Topography Mission (SRTM) data was used to
search for low-lying areas prone to water accumulation. The results of the study indicated
definite vegetation trends in the area of interest, but researchers were unable to obtain the
necessary epidemiological data to complete their analysis.
A more recent country-wide study by Brou et al. (2008) investigated landscape variables
related to BU incidence in Côte d’Ivoire. Forest surface area in hectares, irrigated rice
production, banana production, and dam surface areas were mapped and quantified using GIS in
preparation for a multivariate statistical analysis. Logistic regression results suggested that closer
proximities to irrigated rice fields and to dams increased BU risk, but spatial autocorrelation in
the observed data, the environmental covariates, or the model residuals were not considered.
A study investigating the role of landscape features related to BU rates at a country-wide
scale in Benin used satellite imagery and GIS to prepare data for statistical analyses (Wagner et
al., 2008a). BU case data from 2004 and 2005 were mapped to corresponding village locations.
Concentric buffers with radii ranging from 100m to 50 km from village centers were generated
to analyze land cover composition, elevation, and potential wetness patterns, and distances from
village centers to the nearest river were quantified. A negative binomial statistical analysis
including random district effects determined that villages lying at low elevations within drainage
basins with variable wetness patterns that are surrounded by forest cover have a high BU risk. A
spatial scan statistic identified spatial disease clusters with higher than expected BU rates near

24

the Zou and Ouémé Rivers located toward the eastern side of the country and along the Couffo
River located toward the western side of the country. Lower than expected BU rates occurred
near the coastline and toward the northern region of Benin.
A second study by Wagner et al. (2008b) used LULC variables to construct a probability
model related to BU infection. Latitude, longitude, distance to nearest river, elevation, total
population and seven LULC class percentages were analyzed statistically using a multilevel
Bayesian logistic regression model with district-level random effects. This modeling approach
used data derived at different scales and the district-level random effects mitigated correlation in
residuals due to unknown covariates at the district-level. Model residual plots revealed that
spatial autocorrelation was no longer present in the data set. The results indicated that the
probability of BU incidence increased with percentage surrounding agricultural land cover, and
the probability of BU incidence decreased as the percentage of surrounding urban land use
increased at large scales. A set of “best” models predicted BU rates from the mean posterior
distribution of each model at validation sites in Benin and also in Ghana. The models predicted
BU positive sites more accurately than BU negative sites.
A study investigating BU incidence in Victoria, Australia from 1980 to 2009 performed a
network analysis to determine whether BU cases occurred randomly before the analysis of
climate, landscape, and socioeconomic data (van Ravensway et al., IN PREP). Network analysis
results determined that BU cases did not occur randomly; therefore, administrative data
equivalent to U.S. census tract-level data demarcated town boundaries from which derivation of
several environmental and sociological data variables took place. Statistical analyses used
multilevel Poisson regression models with random effects for endemic locations. Two model sets
were produced, one investigating an approximate lag time between human infection and

25

symptom presentation, with the assumption that mosquitoes play a direct role in transmission and
that infection occurred during high mosquito abundance periods, and referred to as the mosquito
abundance prediction (MAP) model. The second model set investigated landscape, occupational,
and climate (LOC) variables related to BU incidence in the study region.
The MAP models used climate variables as a proxy for high mosquito abundance periods,
wetland land cover and low elevations represented mosquito habitats, and outdoor occupation
variables represented proportion of humans at risk within each town. MAP model results
suggested that towns with high percentages of water and low elevations with increased
precipitation and minimum temperatures nine months prior to BU symptom presentation with
large proportions of the population working in agriculture, mining, or parks and gardens
professions were at a higher BU risk, but direct links to mosquitoes as vectors are still under
investigation. The LOC model results indicated that towns at lower elevations with high
percentages of forest and agriculture land cover, along with decreased maximum temperatures 14
months prior to symptom presentation, followed by increased precipitation 11 months prior to
symptom presentation are at a higher BU risk. This study was the first of its kind to incorporate
climate, social, and landscape variables into its analyses and the first to investigate a specific
vector habitat in relation to BU incidence.
While these studies investigated important environmental features related to BU risk,
landscape components focused on compositional features, rather than exploring landscape
configurations. Further, many of these studies investigated relationships between specific
anthropogenic activities, for example proximity to dams or percentage of surrounding
agriculture, while insight into overall landscape disturbance was not gained.

26

Additionally, although these studies observed variables spatial in nature, with the
exception of Duker et al. (2006) and Wagner et al. (2008a, 2008b), quantified relationships
failed to account for potential spatial autocorrelation in model residuals, and while several
studies included random effects to help account for unobserved variables contributing to BU
prevalence, no studies incorporated spatial random effects that identify whether a spatial
structure exists for these unobserved covariates. Incorporating spatial random effects into the
modeling process while accounting for spatial autocorrelation among model residuals provides a
powerful tool from which to predict BU risk at new locations and to identify potential covariates
through observations of spatial characteristics related to unknown covariates.

1.10 Objectives
This study builds upon knowledge gain from previous studies in Benin by relating BU
occurrence to landscape patterns (Wagner et al., 2008a; Wagner et al., 2008b). Previous research
in this region related specific land cover composition to BU risk across large spatial extents. The
primary objective of this research is to observe landscape configurations in addition to landscape
composition related to BU rates at a medium spatial extent, ranging from 400m to 2k from
village centers. Observing closer proximities allowed quantification of habitats traversed
regularly by village residents, while maintaining study areas large enough to quantify landscape
patterns.
The specific objectives of this thesis were to
1) determine whether landscape patch configurations, quantified using landscape metrics
and used as indicators of potential anthropogenic-related landscape disturbances, are related to
BU rates in Benin, West Africa,

27

2) whether specific land cover configurations specific to individual land cover
composition types are related to BU rates,
3) whether compound topographic index values using 30m resolution data were
consistent with previous research outcomes implicating a relationship between stagnant water
and BU rates.
4) identify whether a spatial structure exists for underlying processes contributing to BU
rates and finally
5) create a BU risk surface based on identified contributing landscape characteristics and
underlying spatial processes in the study area.
Further, the goal of this study is not to identify specific ecological processes contributing
to MU growth and BU transmission in Benin, but to determine whether human disturbance
indicators as characterized by landscape patterns are linked to and can be used to identify high
BU risk landscapes.

1.11 Hypotheses
Previous research suggesting that a relationship exists between anthropogenic landscape
disturbances and BU risk (Merritt et al., 2005; Hayman 1991; Brou et al. 2009; Wagner et al.,
2008; Duker et al., 2004; WHO, 2000) provided the basis for the first set of hypotheses. The first
hypothesis is that disturbed landscapes characterized by uniformly-shaped land cover patches
and low aggregation levels, representative of potential anthropogenic landscape disturbances,
surround villages with higher BU rates. The second hypothesis is that a positive relationship
exists between BU rates and land cover patch configurations representative of potential
anthropogenic landscape disturbances corresponding to specific land cover composition types;
for example, uniformly-shaped agriculture or forest patches. Previous research findings
28

demonstrating that BU cases occur near slow-moving or stagnant bodies of water or in low-lying
areas within drainage basins (CDC, 2005; WHO, 2007; Wagner et al., 2008) provided the basis
for the final hypothesis that landscapes with higher average compound topographic, or wetness
index, values surround villages with higher BU rates.
A strong inference approach, involving the exploration of multiple, biologically plausible
hypotheses followed by identification of the model that provides the best approximation of the
data, will provide the basis from which inference may be drawn to predict disease risk across the
study domain (Plowright et al., 2008).

29

Chapter 2 MATERIALS AND METHODS
2.1 Study Area
The southern regions of Benin and Togo, West Africa comprise the study area for this
investigation. The study area falls within a 185 x 185 km area that is equal to one Landsat
satellite image extent located in WRS path 192 row 55, with latitudes ranging from
approximately 8.167 N to 6.300 N and longitudes ranging from approximately 0.842 E to 2.482
E. The study area encompasses all or parts of several administrative districts, including
Atlantique, Mono, Plateau, Couffo, Zou, and Collines located in Benin and Centre, Plateaux, and
Maritime in Togo. Four major rivers flow through the study area, including the Couffu, Oueme
and Zou Rivers in Benin, and the Mono River which delineates the border between Togo and
Benin in the southern regions of the countries (Figure 2-1).
Buruli ulcer is endemic in Benin and Togo, but limitations exist regarding data
corresponding to disease incidence in Togo. Therefore, this study used BU case data and
corresponding environmental data from Benin to predict BU risk across the landscape, into
regions within the boundary of Togo.
The Republic of Benin is a coastal country located in West Africa along the Gulf of
Guinea, bounded by the countries of Togo, Burkina Faso, Niger, and Nigeria. Benin is a former
French colony that gained independence on August 1, 1960. The country encompasses a total
surface area of 112,622 sq km, with 110,622 sq km of land and 2,000 sq km of water (CIA). The
2010 estimated population is 9,056,010 people, and the average life expectancy at birth is 62.3
years for women and 60.1 years for men (United Nations Statistical Division). French is the
official language in Benin, but several groups speak local, indigenous languages, including the
Fon, Yorouba, Boun, Bariba, Somba, Aizo, Mina, and Dendi (United Nations Benin, 2010).
30

Figure 2-1. Togo and Benin administrative districts included in study area
Subsistence farming is the main form of agriculture production, but for-profit agriculture
products include cotton, corn, cassava, yams, beans, palm oil, peanuts, cashews and livestock
(CIA). Benin’s landscape consists primarily of woodland and savannas, followed by cultivated
woodland and shrub savanna, then fallow fields mixed with cultivated fields; semi-deciduous
31

forests are present in the southern region of the country and occupied an estimated 1-2% of the
land surface in 2007 (USAID, 2007). Deforestation and environmental degradation is an ongoing
problem in Benin, and the landscape continues to change rapidly. Food-crop and cotton
cultivation expanded by 265% and 79% between 1986 and 1997, while firewood extraction and
charcoal production contributed to a 30,000 ha per year deforestation rate, along with additional
deforestation taking place to construct fish corrals (USAID, 2007).
Southern Benin falls within an a sub-equatorial climate zone, often referred to as the
Dahomey Gap, with bimodal seasonal distributions; two rainy seasons extend from April to July
and from September to November, while two dry seasons extend from December to March and
in the month of August each year (Dossou and Gléhouenou-Dossou, 2007).

2.2 BU Case Data
The Programme National de Lutte contre la Lèpre et l’ulcère de Buruli (PNLLUB)
provided a subset of Benin 2004 and 2005 BU positive and BU negative villages to use in this
analysis. Processing of these data included three levels. Initially, volunteers and trained
personnel reported cases at the village level under the supervision of health care workers from
nearby facilities, and then a trained nurse reported cases at the regional level using BU02
standardized forms, which were then summed quarterly at the national level (Sopah et al. 2007).
A total of 292 villages, 183 positive and 109 negative, from the data subset were located
within the study area (Figure 2-2); 558 cases occurred total, ranging between 1 case per village
to 29 cases per village. A village was identified as positive if at least one case occurred in 2004
or 2005. The data set consisted of BU case counts, population counts, and latitude and longitude
coordinates of villages. No correlation between population and BU incidence existed.

32

Figure 2-2. Benin 2004 and 2005 BU case data subset
Wagner et al. (2008a) identified disease clusters with higher than expected and lower
than expected BU rates located across southern Benin using the same data set. A primary cluster
comprised of higher than expected BU rates was located across the Atlantic, Zou, and Couffo
districts and falls within the study area for this analysis.

2.3 Satellite imagery and derivatives

33

Land use and land cover data was derived from December 13, 2000 Landsat ETM+
imagery obtained from the University of Maryland Global Land Cover Facility
(http://glcf.umiacs.umd.edu/index.shtml). The imagery was geometrically rectified and projected
(UTM Zone 31N) before performing an unsupervised classification (Lillesand and Kiefer, 2000)
using an ISODATA algorithm with100 initial classes on the principal axis, set to approximate
true color for 10 iterations or 0.95 convergence in Erdas Imagine software program. The initial
100 classes were then reduced to 22 classes, which were then reduced further to 10 classes,
following Anderson’s Level I classification scheme, with the exception of a mixed class and the
agriculture class (Table 2-1; Figure 2-3; Anderson, 1976). The final classification consisted of
three mixed agriculture classes because a dominant agriculture class did not emerge from the
classification algorithm and a mixed class that included pixels classified with greater than two
categories. Execution of a 5x5 statistical majority filter helped to eliminate classification noise
(Mather, 2004).
Lack of ground truth data points from this region and during this time period limited the
ability to assign spectral signatures to specific land cover categories; therefore, visual
interpretation methods were used to delineate land cover classes and aggregated land cover
classes were intended to help mitigate classification error.

34

Table 2-1. Land use and cover classification categories
Land Use and Land Cover Classes
Class
Number

Class Name

Class Description

1

Urban

2
3
4
5

Not Present
Rangeland
Forest

Mixed Urban or Built-up Land, Other Urban or Built-up
Land, Transportation
Agriculture: see classes 9, 10, and 11
Shrub and Brush Rangeland: Shrub and Shrubland
Riparian Forest, Evergreen Forest, Forest

Water

Streams and Canals, Lakes, Reservoirs, Bays and Estuaries

Wetland and Forested and Non-forested Wetland:
Water/Wetland, Wetland/Shrub, Wetland/Riparian
Vegetation
7
Barren
Sand/Salt
Evergreen/Ag/Shrub, Shrub/Ag/Wetland,
8
Mixed
Wetland/Shrub/Urban, Shrub/Ag/Urban
9
Agriculture 1
Shrub/Agriculture
10
Agriculture 2
Agriculture/Forest
11
Agriculture 3
Agriculture/Urban
Class categories correspond to Level I classification scheme except for class 8 (Mixed)
and classes 9-11 (Mixed Agriculture) (Anderson et al., 1976)
6

Wetland

Quantifying landscape patterns within concentric polygons is a common approach in
landscape analyses (see Condeso and Meentemeyer, 2007; Graham et al., 2004; Danson et al.,
2004; Brownstein et al., 2005; Wagner et al., 2008a, Wagner et al., 2008b). Concentric polygons
with radii set at 100m intervals, from 400m to 2k, from village centroids acted as buffers from
which LULC classification data was obtained for pattern analysis (Figure 2-4).

35

Figure 2-3. Land use and land cover classification.
The LULC within each buffer was clipped and reformatted for compatibility within
FRAGSTATS software package. A total of 292 individual buffers representing each village
location were created at each distance interval, with 16 distance intervals total.

36

Figure 2-4. Concentric buffers example
Compound topographic index, or wetness index, derivation was completed using 30m
resolution ASTER Global Digital Elevation Model data (https://wist.echo.nasa.gov/wistbin/api/ims.cgi?mode=MAINSRCH&JS=1). DEM sinks were filled, and elevation slope was
calculated in degrees using ArcGIS software program. Flow direction was modeled using a D-8

37

algorthim (O’Callaghan and Mark, 1984) from which flow accumulation and the wetness index
were calculated using the following equation (Equation 2-1; Beven and Kirby, 1979):
Equation 2-1

wi  ln( A / tan  )

s

where

As is the catchment area at a point, or flow accumulation, and tan  is the local slope

gradient (Figure 2-5; Schmidt and Persson, 2003). Wetness index averages within each buffer
were calculated using the Zonal Statistics function within Hawth’s Tools in ArcGIS.

Figure 2-5. Wetness index map.
38

2.4 Statistical Modeling
Model sets were created at the landscape level, characterizing landscape configuration for
all classes across the study area, and at the class level, characterizing landscape configuration for
specific classes across the study area.
Following the integrated approach of Yang et al. (2008), a landscape analysis approach
and a hierarchical Bayesian spatial modeling approach employing Markov Chain Monte Carlo
(MCMC) simulation were used to identify important covariates and to predict the spatial
distribution of BU risk across the study domain. Figure 2-6 outlines the methods and approaches
used to generate the final predictive surface.

2.5 Landscape-level Configuration Analysis
A variety of landscape-level metrics were calculated within FRAGSTATS software
package using the clipped LULC data. These calculations employed an 8-neighbor rule that
included neighboring cells with flat adjacencies and those with diagonal adjacencies to the focal
cell (Gergel and Turner, 2002). The ability to characterize potential human disturbances to the
landscape expressed through relatively high fragmentation values, potential for ecological edge
effects, and potential biodiversity loss provided the basis for landscape metric selection (Table
2-2).

39

Figure 2-6. Methods and approaches used for final model generation
40

Table 2-2.Initial landscape metrics calculations.
Landscape-level Initial Landscape Metrics
Abreviation
Description
Standardized aggregation
LSI
index based on total edge
Standardized land cover
Shape Index Mean
SHAPE_MN patch edge complexity
measurement
Fractal Dimension
Land cover patch edge
FRAC_MN
Index Mean
complexity measurement
Perimeter-to-area
Land cover shape
PARA_MN
Ratio Mean
complexity measurement
Percent Land Cover
Land cover type
PLANDJ
Adjacency
aggregation measurement
Patch Richness
Standardized landcover
PRD
Density
diversity measurement
Shannon's Diversity
Land cover diversity
SHDI
Index
measurement
Measures number of
Number of Patches NP
individual patches present
in area
Metric
Landscape Shape
Index

Interest
Fragmentation
Altered land
cover shapes
Altered land
cover shapes
Altered land
cover shapes
Fragmentation
Heterogeneity
Heterogeneity
Fragmentation

Redundancy among landscape metrics is a common problem (Riitters et al., 1995);
therefore, a Spearman’s rank order correlation matrix identified mulitcolinearity between
metrics, and only those metrics that exhibited an r < 0.60 with one another across all distance
intervals were selected for candidate regression models. Correlation matrices generated for this
study are available in Appendix A.
Metrics were divided into groups representing shape complexity, potential fragmentation,
and land cover diversity. Candidate model generation used non-spatial binomial generalized
linear regression models (GLMs) with BU rates as the response variable and landscape metric
values and average wetness index values as independent variables. BU rates were derived by
dividing BU counts by the total population at each location. The binomial GLM model

41

distribution is y | p ~ binomial (n , p ) where p is the success probability and links to the

i

i

i

i

i

regression covariate x through a logit transformation (Equation 2-2; SAS/STAT(R)):

i

Equation 2-2

p
log it ( p )  log( i )     x
i
i
1 p
i
Nested models were generated at 100m distance intervals from 400m to 2k; data was not
combined across distances. High leverage points were identified using Cook’s distance measure,
and data locations with a value > 1.0 that impacted sign direction and model results substantially
were removed from the data set (Cook, 1977). Fourteen high leverage points were removed from
the landscape-level analysis, and removal of high leverage points at the class-level occurred on
an individual model basis (Appendix B). Final candidate models consisted of data points that
were never high leverage points at any distance interval between 400m and 2k. Observations of
Akaike Information Criterion (AIC) values (Akaike, 1974) determined the “best model” to use in
the final analysis. Two “best” model sets were created at the landscape-level because metrics
representing diversity and fragmentation were correlated with one another, but both model sets
demonstrated explanatory potential. Complete non-spatial candidate model results are available
in Appendix C.

2.6 Landscape-level Configuration Analysis Predictor Variables
Characterization of land cover shape complexity used the shape index mean. Selection of
the shape index mean over the fractal dimension index mean occurred because the fractal
dimension index mean has the potential to introduce bias when working with small areas, as was
42

the case when working with buffers with shorter radii from village centers (Turner, 2005). Shape
index mean values characterize land cover patch shape complexity using an equation that adjusts
for a square standard, thereby eliminating problems introduced from changing perimeter-to-area
ratios (McGarigal and Marks, 1995). The shape index mean is calculated in FRAGSTATS
software package as follows (Equation 2-3):
Equation 2-3

p
ij
SHAPE 
min p
ij
where p  perimeter of patch ij in terms of number of cell surfaces and p  minimum

ij

perimeter of patch

ij

ij in terms of number of cell surfaces. Values closer to 1.0 indicate more

uniformly-shaped land cover patches with complexity increasing as values increase. A plot of the
shape index mean surface at the landscape level plotted areas with more uniform patch shapes in
blue and more complex shapes in red (Figure 2-7).

43

Figure 2-7. Shape index mean surface
Number of patches is a measurement of potential habitat fragmentation. Number of
patches is calculated within FRAGSTATS as follows (Equation 2-4):
Equation 2-4

NP  N
where NP is equal to the number of patches in a landscape and N is equal to the number of
patches in the landscape (McGarigal and Marks, 1995). High numbers of patches correspond to
greater fragmentation potential, while lower numbers of patches indicate a more aggregated
landscape. No background or border cells were included in the calculation, and patch attributes,
for example total area or land cover type, were not described in this metric (Figure 2-8).

44

Figure 2-8. Number of patches surface
Average wetness index value was the third landscape-level predictor variable. As
outlined above, a wetness index identifies areas in a landscape where water is likely to
accumulate during a precipitation event (See Equation 1; Bevin and Kirby, 1979). High wetness
index values indicate greater accumulation potential, while low wetness index values indicate a
lower accumulation potential. A plot of wetness index values across the study region highlights
areas of high accumulation potential in blue and areas with low accumulation potential in red
(Figure 2-9).

45

Figure 2-9.Wetness index average surface
Limitations of mean approaches include the possibility of unusually high or low values
within a buffer distance affecting the mean value disproportionately, although using a large
number of data locations and the elimination of high leverage points was intended to help
mitigate this problem.
The patch richness density metric quantifies the number of different land cover types in a
landscape and then standardizes this measurement in order to facilitate comparison across
landscapes (McGarigal and Marks, 1995). Patch richness density was calculated in
FRAGSTATS as follows (Equation 2-5):

46

Equation 2-5

PRD 
where

m

present and

m
(10, 000)(100)
A

number of patch types present in the landscape, excluding the landscape border if

A  total landscape area m 2 (McGarigal and Marks, 1995). Higher values represent

more diverse landscapes, while lower values represent more homogeneous landscapes. Regions
with higher numbers of land cover types were plotted in red and those with lower numbers of
land cover types were plotted in blue in Figure 2-10.

Figure 2-10. Patch richness density surface

47

2.7 Class-level Configuration Analysis
Additional model generation at the class-level included predictor variables characterizing
land cover patch configurations corresponding to specific land cover types. Shape index mean
values and an additional variable group characterizing class-level fragmentation were calculated
in FRAGSTATS (Table 2-3).
Table 2-3. Class-level initial landscape metric candidates
Class-level Initial Landscape Metrics
Abbreviation
Description
Standardized land cover
Shape Index
SHAPE_MN patch edge complexity
Mean
measurement
Landscape Shape
Standardized aggregation
LSI
Index
index based on total edge
Percent Land
Land cover type
PLANDJ
Cover Adjacency
aggregation measurement
Measures number of
Number of
NP
individual patches present
Patches
in area
Clumpiness
Land cover type
CLUMPY
Index
aggregation measurement
Aggregation
Land cover type
AI
Index
aggregation measurement
Metric

Interest
Altered land cover
shapes
Fragmentation
Fragmentation
Fragmentation
Fragmentation
Fragmentation

As outlined in the landscape-level analysis, a Spearman’s rank order correlation matrix
identified potential multicolinearity between variables (Appendix A), and non-spatial binomial
GLMs were generated for land cover types using BU rates as the response variable and
landscape metric values corresponding to specific land cover classes as independent variables.
Three land cover classes were explored: agriculture/forest, forest, and wetland because these land
cover classes represented land cover types or uses noted in the literature as having anecdotal or
empirical links to BU incidence in past studies.

48

Data sets changed as distance increased from village centers because a buffer at a 400m
distance did not necessarily contain forest patches, while a buffer at 2k might contain several
forest patches. Therefore, models were produced individually for each land cover type within
buffers at 800m, 1_2k, 1_6k, and 2k distances. Identification of high leverage points occurred on
an individual model basis using Cook’s distance measure (Appendix B); therefore, a “best”
model for each distance interval was not identified using lowest AIC values, but was determined
by variable significance values.

2.8 Class-level Configuration Analysis Predictor Variables
Shape index mean acted as a predictor variable to quantify relationships between land
cover patch shape complexity and BU rates. Calculation of this metric was identical to the
landscape-level shape index mean (Equation 3) with the exception of quantifying individual
classes rather than all classes encompassing the landscape across the study region.
Percent land cover adjacency measured land cover class aggregation (Equation 2-6)
Equation 2-6


 g
ij

PLADJ  
m
  g

ik
 k 1
where

g

ij 

the number of like adjacencies (joins) between pixels of patch type (class)

on a double-count method, and
types (classes)









i

g

ik 

i based

the number of adjacencies (joins) between pixels of patch

and k are based on a double-count method (McGarigal and Marks, 1995).

49

High PLADJ values represent a more aggregated land cover class, while low PLADJ values
represent a more fragmented land cover class. An example of PLADJ for the 2k wetland class
demonstrates high aggregation values in the northern study area plotted in red, and low
aggregation in the southeastern corner of the study area plotted in blue (Figure 2-11).

Figure 2-11. Percent land cover adjacency surface 2k wetland class
The landscape shape index (LSI) is another land cover patch aggregation measurement.
LSI was calculated within FRAGSTATS using the following equation (Equation 2-7):

50

Equation 2-7

e
i
LSI 
min e
i
where

e  total length of edge (or perimeter) of class i in terms of number of cell surfaces,
i

includes all landscape boundary and background edge segments involving class i , and

min e  minimum total length of edge (or perimeter) of class i in terms of number of cell
i
surfaces (Figure 2-12; McGarigal and Marks, 1995).

Figure 2-12. 1_2k agriculture/forest class landscape shape index surface

51

2.9 Spatial GLM
The binomial GLMs used to identify significant BU rate predictor variables assumed
model residuals were independent and identically distributed across the study domain (Keitt et
al., 2002). While this approach is adequate in the absence of spatial autocorrelation in model
residuals, this assumption was unrealistic given the spatial structure of the observations and the
heterogeneity of the environmental covariates (Boyd et al., 2005).
In order to mitigate this problem, a hierarchical modeling approach was used to adapt the
non-spatial GLM with a correlation matrix based on the spatial structure of the model residuals,
adding a spatial random effects component that accounted for missing, spatially-structured
covariates (Waller and Gotway, 2004).
Our interest was in modeling the number of cases Y ( s) at n observed locations. Here,

s denotes the geographic location, assuming s  D  2 .
predictors

A set of

p spatially referenced

x( s) were calculated at each location. Given the population at each location, N ( s) ,

we assumed Y ( s) followed a binomial distribution. For the i -th location

(Y (s ) |  (s ) ~ Binomial ( N ( s ), p( ( s ))) where p( ( s )) is the success probability at s
i
i
i
i
i
i
and  ( s )  x( s )  w( s ) . A logit link function

i

i

i

p( (s ))  exp( ( s )) / (1  exp( ( s ))) was assumed for this model.
i
i
i
The process specification for w( s) is a mean 0 Gaussian Process with covariance
function, C ( s , s ) , denoted GP(0, C ( s , s )) . In application, we specify

1 2

1 2

C ( s , s )   2 p( s , s ;  ) where p(; ) is a correlation function and  is the spatial decay
1 2
1 2
52

parameter. The exponential spatial correlation was assumed for p() . Prior distributions on the
remaining parameters complete the hierarchical model. Customarily, the regression effect

 is

assigned a multivariate Gaussian prior, (i.e.  ~ N (  ,   ) , while the latent variance



component  2 is assigned IG(, ) priors. The process correlation parameter,  , was assigned an
informative prior (e.g., uniform over a finite range) based upon the underlying spatial domain.
Model parameter distributions were estimated using Markov Chain Monte Carlo
(MCMC) methods employing an adaptive Metropolis (AM) algorithm with a 43% acceptance
rate. MCMC employs an iterative sampling process, the goal of which is to converge to a
stationary state representing the target distribution, or in the Bayesian approach, the true
distribution of the model parameters, from which inference may be drawn (Brooks and Gelman,
1998).
The AM algorithm differs from a traditional Metropolis algorithm because accumulated
information from all previous chains are incorporated into each sampling iteration as opposed to
only the prior chain’s information contributing to the calculation (Brooks and Gelman, 1998).
This sampling method speeds convergence, or the point at which equilibrium between parameter
mean and variance commences.
Starting values were obtained from non-spatial models and posterior inference was based
on 3 chains at 500,000 post burn-in iterations measured at 1_2k, 1_6k, and 2k for landscape-level
models, and 3 chains at 50,000 to 100,000 iterations measured at 800m, 1_2k, 1_6k, and 2k for
the class-level models, depending on the individual model. Burn-in occurred at 10,000 iterations
and thinning took place at every 100 samples. All model generation was completed using the

53

spGLM function in spBayes R package, and summaries are available in Appendix D of this
document.
The Gelman-Rubin diagnostic measure assessed model convergence within 5,000
MCMC sample iterations. This diagnostic observed within chain and between chain variance
with values closer to 1.0 indicating that convergence likely took place (Brooks and Gelman,
1998).
Comparison of deviance information criterion (DIC) values from non-spatial models and
from spatial models determined whether the spatial models achieved a better fit. A similarity
exists between DIC values and AIC values; a lower DIC value indicates a better model fit, but
DIC value are used with Bayesian hierarchical models because they can be calculated from
MCMC samples directly (Spiegelhalter et al., 2002).

2.10 Model Verification
A random sample consisting of 10% of village locations was withheld from each model
data set for verification purposes. Spatial GLMs were calculated at the landscape-level at 1_2k,
1_6k, and 2k distance intervals using 90% of the data set as training locations, and at the classlevel at 800m, 1_2k, 1_6k, and 2k distance intervals using 90% of the data set as training
locations. Training model results predicted BU rates at the remaining 10% verification data set
locations using the spPredict function in spBayes R. Predicted values were compared to actual
values at these locations, and a root mean squared error (RMSE) determined the model that
achieved the best fit (Figure 2-8).

54

Equation 2-8

m
ˆ
RMSE   ( y ( s )  y ( s ))2 / m
i 
i 
i 1
where

ˆ
y ( s ) are the estimated rates, y ( s ) are the actual rates, and m is equal to the number of



data points (Bolstad, 2005). The model exhibiting the lowest RMSE value was chosen as the
“best” model from which to predict to new, unsampled data locations.

2.11 Risk Surface
A BU risk surface was created employing Bayesian, or model-based, kriging methods
using samples drawn from the mean posterior distribution of the predictor variables from the best
spatial binomial GLM to predict BU rates at new, unsampled locations. Incorporation of the
random spatial effects component accounted for unknown predictor variables and spatial
autocorrelation, improving prediction accuracy.
Kriging is a geostatistical interpolation method that uses data from observed locations to
predict to new, unsampled locations (Schabenberger and Gotway, 2005). Kriging is a
probabilistic approach; therefore, it has standard errors associated with model predictions,
enabling quantification of uncertainty related to model outputs (Waller and Gotway, 2004).
Several kriging approaches exist. A Bayesian kriging approach differs from traditional kriging
approaches because model parameters are treated as random variables rather than estimated
(Helbert et al., 2009), and while traditional kriging methods ignore uncertainty introduced when
estimating the covariance structure, Bayesian kriging incorporates parameter uncertainty into
model predictions (Moyeed and Papritz, 2002).

55

A total of 1030 new locations generated at 5km intervals within the boundary of the study
area represent locations to which BU rates were predicted (Figure 2-13). Gaps in new location
points represent areas where derivation of predictor variables could not occur because a lack of
data existed corresponding to the “best” model covariates and distance interval.

Figure 2-13. New locations and observed locations used to create risk surface

56

Circular buffers with radii equal to the best-fitting spatial model were created
surrounding each new location, and a polygon created at approximately 5k inside the study area
boundary prevented the introduction of uncertainty due to potential edge effects (Haase, 1995).
Derivation of predictor variables followed methods outlined in section 2.4, and a surface based
on the mean posterior predictive distribution and associated uncertainty was created across the
entire study area using the spPredict function in spBayes R.
A population of 100,000 persons was assumed at each location to calculate meaningful
rates, although these locations were not necessarily occupied by villages or residents. These
location points were generated to characterize risk associated with landscape variables across the
study region, but do not suggest that BU transmission is taking place. Unreliable and incomplete
population data and villages with questionable coordinate information prompted pseudo-location
generation.

57

Chapter 3 RESULTS
3.1 Non-Spatial vs. Spatial Model Results
Spatial model results determined that several non-spatial models overestimated the
significance of one or more predictor variables, indicating that spatial autocorrelation was
present in model residuals. Lower DIC values corresponding to the spatial GLMs compared to
the non-spatial GLMs demonstrated that the spatial models achieved a better fit in every
circumstance (Table 3-1).
Table 3-1. Non-spatial and spatial GLM DIC values.

Scale

Model
1_2k
1_2k

Landscapelevel

1_6k
1_6k
2k

Class-level

2k
800 Forest
800 Wetland
800 Ag/Forest
1_2k Forest
1_2k Wetland
1_2k Ag/Forest
1_6k Forest
1_6k Wetland
1_6k Ag/Forest
2k Forest
2k Wetland
2k Ag/Forest

Variables
WIAVG + SHAPE_MN +
PRD
WIAVG + SHAPE_MN +
NP
WIAVG + SHAPE_MN +
PRD
WIAVG + SHAPE_MN +
NP
WIAVG + SHAPE_MN +
PRD
WIAVG + SHAPE_MN +
NP
SHAPE_MN + PLADJ
SHAPE_MN + PLADJ
SHAPE_MN + LSI
SHAPE_MN + LSI
SHAPE_MN + LSI
SHAPE_MN + LSI
SHAPE_MN
SHAPE_MN + LSI
SHAPE_MN + LSI
PLADJ
PLADJ
SHAPE_MN + LSI
58

Post
Burn-in

DIC Value
NonSpatial
Spatial

500,000

7542.948

6774.835

500,000

7558.832

6775.580

500,000

7526.302

6775.539

500,000

7558.955

6777.154

500,000

7533.240

6776.804

500,000
100,000
50,000
50,000
100,000
100,000
100,000
50,000
100,000
50,000
50,000
50,000
50,000

7539.006
1927.261
3918.085
2000.561
3661.611
5335.240
3901.120
3668.271
5937.132
4344.966
4394.660
6558.870
4712.282

6777.211
1780.473
3483.809
1859.042
3289.170
4781.834
3644.391
3257.107
5218.884
4071.081
3883.277
5780.372
4316.881

Table 3-2 provides a comparison of non-spatial and spatial model variable significance
results, with full model results available in Appendices C and D.
Table 3-2. Non-spatial vs. spatial binomial GLM variable significance.
Model

Landscapelevel

Class-Level

Variable
Wetness index average
1_2k LS
Shape Index Mean
Number of Patches
Wetness index average
1_2k LS
Shape Index Mean
Patch Richness Density
Wetness index average
1_6k LS
Shape Index Mean
Number of Patches
Wetness index average
1_6k LS
Shape Index Mean
Patch Richness Density
Wetness index average
2k LS
Shape Index Mean
Number of Patches
Wetness index average
2k LS
Shape Index Mean
Patch Richness Density
Shape Index Mean
800m Forest
Percent Land Cover
Adjacency
800m Wetland
Shape Index Mean
Shape Index Mean
800
Agriculture/Forest Landscape Shape Index
Shape Index Mean
1_2k Forest
Landscape Shape Index
Shape Index Mean
1_2k Wetland
Landscape Shape Index
Shape Index Mean
1_2k
Agriculture/Forest Landscape Shape Index
1_6k Forest
Shape Index Mean
Shape Index Mean
1_6k Wetland
Landscape Shape Index
Shape Index Mean
1_6k
Agriculture/Forest Landscape Shape Index
Percent Land Cover
2k Forest
Adjacency
59

Significance
Non-Spatial
Spatial
Model
Model
Y
N
Y
N
Y
N
Y
N
Y
N
Y
Y
Y
N
Y
N
Y
Y
Y
N
Y
N
Y
N
Y
Y
Y
Y
Y
N
Y
Y
Y
N
Y
N
Y
N
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y

N
Y
Y
Y
Y
N
N
N
Y
Y
N
N
N
Y
Y

Y

Y

Class-Level
Model
2k Wetland
2k
Agriculture/Forest

Variable
Percent Land Cover
Adjacency
Shape Index Mean
Landscape Shape Index

Significance
Non-Spatial
Model

Spatial
Model

Y
Y

N
Y

Y

Y

3.2 Model Output Interpretation
Model results contain two components. The first component consists of a table reporting
variable significance for each model under the Bayesian credible interval approach. Similarities
exist between credible interval interpretations used in Bayesian statistical analyses and
confidence interval interpretations used in frequentist statistical approaches, although exact
meanings vary between the two approaches (Bland and Altman, 1998). The Bayesian approach
determines that a 95% probability exists that a value lies within a distribution, while the
frequentist approach assumes that a population value is fixed and then constructs a 95%
confidence interval around the fixed value.
Under a Bayesian approach, when parameter value directions do not change between the
lower 2.5% and the upper 97.5% credible tails, the variable may be interpreted as significant at a
95% confidence level; for example, in this study, patch richness density maintained a positive
direction in both tails at a 1_2k distance (Table 3-3), indicating that BU rates increased as patch
richness density increased. The output tables also include the effect range which is the distance at
which the spatial correlation drops to 0.05 (i.e., -log(0.05)/phi).
The second model output component consists of a Gelman-Rubin diagnostic table that
outlines the likelihood that convergence took place within 5,000 MCMC sample iterations.
Values close to 1.0 indicate that convergence likely took place.
60

3.3 Landscape-level Configuration Analysis Results
The following tables and figures outline spatial GLM landscape-level results. No models
maintained significance in all variables under the spatial modeling framework, but four out of six
models maintained significance in at least one variable. The first model to maintain a significant
variable was the 1_2k patch richness density model mentioned above. Results indicated that as
patch richness density increased, BU rates increased, although wetness index average and shape
index mean values were no longer significant under the spatial modeling framework (Table 3-3).
Table 3-3. 1_2k number of patches spatial binomial GLM model summary
Summary 1_2k Landscape-level 3 Chains at 500,000
Parameters
50.0%
2.5% 97.5%
Intercept
-12.94 -14.991 -7.244
Wetness Index Average
0.151
-0.035 0.223
Shape Index Mean
0.913
-2.765 2.487
Patch Richness Density
2.397
1.458
2.89
sigma.sq
3.285
3.097 3.922
Phi
24.388
13.647 29.623
effective range 3/phi
0.123
0.101 0.224
max intersite distance
1.254
The Gelman-Rubin diagnostic measure determined that convergence likely took place
with values presenting close to 1.0 in all parameters (Table 3-4).
Table 3-4. Gelman diagnostic 1_2k patch richness density model
Gelman-Rubin Diagnostic 1_2k Landscape-level
Point
97.5
est.
quantile
Intercept
1.13
1.39
Wetness Index Average
1.00
1.01
Shape Index Mean
1.12
1.34
Patch Richness Density
1.01
1.03
sigma.sq
1.00
1.00
phi
1.03
1.03
Multivariate psrf = 1.08
61

The second model that maintained significance in one variable was at a 1_6k distance
(Table 3-5). Results indicated that a relationship exists between high BU rates and a higher
number of patches surrounding villages within a 1_6k distance. Wetness index average and
shape index mean values were no longer significant.
Table 3-5. 1_6k number of patches spatial binomial GLM model summary
Summary 1_6k Landscape-level 3 Chains at 500,000
Parameters
50.0%
2.5% 97.5%
Intercept
-11.107 -12.728 -9.653
Wetness Index Average
0.077
-0.004
0.241
Shape Index Mean
1.586
-0.645
2.052
Number of Patches
0.007
0.004
0.009
sigma.sq
3.41
2.698
4.21
Phi
21.639 21.517 46.297
effective range 3/phi
0.139
0.067
0.139
max intersite distance
1.254
The Gelman-Rubin diagnostic calculation determined that convergence likely took place
in all model parameters (Table 3-6).
Table 3-6. Gelman diagnostic 1_6k numbers of patches model
Gelman-Rubin Diagnostic 1_6k Landscape-level
Point
97.5
est.
quantile
Intercept
1.05
1.15
Wetness Index Average
1.01
1.03
Shape Index Mean
1.06
1.16
Number of Patches
1.01
1.02
sigma.sq
1.00
1.01
Phi
1.02
1.02
Multivariate psrf = 1.03
Results from the spatial model including wetness index average, shape index means, and
numbers of patches at a 2k distance maintained significance in two variables (Table 3-7). Results

62

suggested that higher wetness index averages and more uniformly-shaped land cover patches
surrounded villages with higher BU rates within a 2k distance of village centers.
Table 3-7. 2k numbers of patches spatial binomial GLM model summary
Summary 2k Landscape-level 3 Chains at 500,000
Parameters
50.0%
2.5% 97.5%
Intercept
-9.084 -10.252 -7.187
Wetness Index Average
0.254
0.213 0.268
Shape Index Mean
-1.021
-1.996 -0.095
Number of Patches
0.000
-0.001 0.006
sigma.sq
3.411
3.109 3.430
Phi
24.873
17.339 29.020
effective range 3/phi
0.121
0.104 0.174
max intersite distance
1.254
A Gelman-Rubin diagnostic calculation determined that convergence likely took place in
all model parameters (Table 3-8).
Table 3-8. Gelman diagnostic 2k number of patches model
Gelman-Rubin Diagnostic 2k Landscape-level
Point
97.5
est.
quantile
Intercept
1.09
1.27
Wetness Index Average
1.00
1.01
Shape Index Mean
1.10
1.30
Number of Patches
1.01
1.04
sigma.sq
1.00
1.01
phi
1.03
1.03
Multivariate psrf = 1.07
The final model to maintain significance in at least one variable also occurred at a 2k
distance (Table 3-9). This model included wetness index averages, shape index means, and patch
richness density values. Results suggested that a positive relationship exists between higher BU
rates and higher wetness index values within a 2km distance of village centers, but shape index
mean values were not significant.
63

Table 3-9. 2k patch richness density spatial binomial GLM summary
Summary 2k Landscape-level 3 Chains at 500,000
Parameters
50.0%
2.5% 97.5%
Intercept
-8.986 -9.983 -7.754
Wetness Index Average
0.126 0.111 0.283
Shape Index Mean
-0.942 -1.411 0.711
Patch Richness Density
-0.273 -1.831 2.954
sigma.sq
3.251 3.010 3.886
Phi
22.260 21.599 26.227
effective range 3/phi
0.135 0.115 0.139
max intersite distance
1.254
The Gelman-Rubin diagnostic measure determined that convergence likely took place in
all parameters with the exception of the shape index mean, which had a value of 2.10 (Table
3-10).
Table 3-10. Gelman diagnostic 2k patch richness density model
Gelman-Rubin Diagnostic 2k Landscape-level
Point
97.5
est.
quantile
Intercept
1.38
2.10
Wetness Index Average
1.00
1.00
Shape Index Mean
1.38
2.10
Patch Richness Density
1.00
1.01
sigma.sq
1.02
1.06
phi
1.05
1.08
Multivariate psrf = 1.26

3.4 Class-level Composition and Configuration Analysis
Seven out of twelve class-level candidate models maintained significance in all model
variables under the spatial modeling framework, the results of which are outlined in following
tables.

64

Results from the 800m wetland class model (Table 3-11) indicated that a relationship
exists between high BU rates and wetland patches with high patch shape complexity.
Table 3-11. 800m wetland class spatial model results
Summary 800m Wetland 3 Chains at 50,000
Parameters
50.0%
2.5% 97.5%
Intercept
-10.022 -10.095 -8.918
Shape Index Mean
0.986
0.722 1.096
sigma.sq
3.503
3.158 3.869
Phi
15.224
12.006 21.354
effective range 3/phi
0.197
0.141 0.251
max intersite distance = 1.260
The Gelman-Rubin diagnostic (Table 3-12) indicated that convergence likely took place
in all model parameters with values presenting close to 1.0.
Table 3-12. 800m wetland class convergence diagnostic
Gelman-Rubin Diagnostic 800m Wetland
Point
est.
97.5 quantile
Intercept
1.02
1.07
Shape Index Mean
1.01
1.02
sigma.sq
1.00
1.01
Phi
1.02
1.02
Multivariate psrf = 1.03
Results summarizing the 800m agriculture/forest class model (Table 3-13)
suggested that a relationship exists between mixed agriculture/forest land cover patches with
more complex shapes that are highly aggregated with one another and higher BU rates.

65

Table 3-13. 800m agriculture/forest class spatial model results
Summary 800m Agriculture/Forest 3 Chains at 50,000
Parameters
Intercept
Shape Index Mean
Landscape Shape Index
sigma.sq
Phi
effective range 3/phi

50.0%
-8.868
2.319
-0.549
1.345
283.921
0.011

2.5%
-9.844
1.774
-0.776
1.142
234.639
0.006

97.5%
-8.245
2.500
-0.192
2.309
551.114
0.013

The Gelman-Rubin diagnostic (Table 3-14) indicated that convergence likely took place
with all values presenting close to 1.0.
Table 3-14. 800m agriculture/forest class convergence diagnostic
Gelman-Rubin Diagnostic 800m Agriculture/Forest
Point
est.
97.5 quantile
Intercept
1.01
1.01
Shape Index Mean
1.01
1.02
Landscape Shape Index
1.00
1.00
sigma.sq
1.00
1.00
phi
1.00
1.00
Multivariate psrf = 1.00
The 1_2k forest model (Table 3-15) indicated that more complex forest shapes and more
aggregated forest patches surround villages with higher BU rates within a 1_2k distance from
village centers.

66

Table 3-15. 1_2k forest spatial model results
Summary 1_2k Forest 3 Chains at 100,000
Parameters
50.0%
2.5%
97.5%
Intercept
11.411 15.723
-8.792
Shape Index Mean
3.439 1.076
6.52
Landscape Shape Index
-0.31
-0.45
-0.07
sigma.sq
4.392 4.304
4.449
Phi
75.14 18.076 189.105
effective range 3/phi
0.04 0.017
0.191
Max intersite distance = 1.260
Gelman-Rubin diagnostic results indicated that convergence likely took place in all
parameters with the exception of the shape index mean, which had a value of 2.15 (Table 3-16).
Table 3-16. 1_2k forest class convergence diagnostic
Gelman-Rubin Diagnostic 1_2k Forest
Point
est.
97.5 quantile
Intercept
1.46
2.34
Shape Index Mean
1.41
2.15
Percent Land Cover Adjacency
1.03
1.11
sigma.sq
1.00
1.01
Phi
1.02
1.08
Multivariate psrf = 1.28
The 1_2k mixed class agriculture/forest model indicated that a relationship exists
between villages with high BU rates and agriculture/forest land cover patches with complex
shapes that are highly aggregated within a 1_2k distance from village centers (Table 3-17).

67

Table 3-17. 1_2k agriculture/forest spatial model results
Summary 1_2k Agriculture/Forest 3 Chains at 100,000
Parameters
50.0%
2.5%
97.5%
Intercept
-9.096 -10.403
-6.957
Shape Index Mean
2.240
0.555
3.582
Landscape Shape Index
-0.255
-0.424
-0.242
sigma.sq
1.547
1.112
1.733
Phi
449.736 319.780 592.320
effective range 3/phi
0.007
0.005
0.009
max intersite distance = 1.230
The Gelman-Rubin diagnostic measure (Table 3-18) indicated that convergence likely
took place with all values presenting close to 1.0.
Table 3-18. 1_2k agriculture/forest class convergence diagnostic
Gelman-Rubin Diagnostic 1_2k Agriculture/Forest
Point
est.
97.5 quantile
Intercept
1.04
1.14
Shape Index Mean
1.05
1.15
Landscape Shape Index
1.00
1.01
sigma.sq
1.00
1.00
Phi
1.00
1.00
Multivariate psrf = 1.04
Results corresponding to the 1_6k agriculture/forest model (Table 3-19) suggested that a
relationship exists between higher BU rates and more complexly-shaped agriculture/forest land
cover patches that are aggregated with one another within a 1_6k distance from village centers.

68

Table 3-19.1_6k agriculture/forest spatial model results
Summary 1_6k Agriculture/Forest 3 Chains at 50,000
Parameters
50.0%
2.5% 97.5%
Intercept
10.739 11.111 10.168
Shape Index Mean
3.635
3.004
3.741
Landscape Shape Index
-0.248 -0.319 -0.208
sigma.sq
1.921
1.334
2.199
Phi
16.864 10.402 33.195
effective range 3/phi
0.178
0.093
0.292
max intersite distance = 1.224
The Gelman-Rubin diagnostic model indicated that convergence likely took place with
all parameter values presenting close to 1.0 (Table 3-20).
Table 3-20. 1_6k agriculture/forest class convergence diagnostic
Gelman-Rubin Diagnostic 1_6k Agriculture/Forest
Point
est.
97.5 quantile
Intercept
1.10
1.31
Shape Index Mean
1.09
1.27
Landscape Shape Index 1.01
1.02
sigma.sq
1.01
1.02
Phi
1.01
1.02
Multivariate psrf = 1.08
The 2k forest model indicated that a relationship exists between higher BU rates and
higher percentages of adjacent forest land cover patches within a 2k distance of village centers
(Table 3-21).

69

Table 3-21. 2k forest class spatial model results
Summary 2k Forest 3 Chains at 50,000
Parameters
50.0%
2.5% 97.5%
Intercept
-10.811 -11.266 -9.358
Percent Land Cover Adjacency
0.024
0.014 0.029
sigma.sq
5.164
4.708 5.165
Phi
18.343 15.803 23.018
effective range 3/phi
0.164
0.131 0.190
max intersite distance = 1.174
The Gelman-Rubin diagnostic model suggested that convergence likely took place in all
model parameters (Table 3-22).
Table 3-22. 2k forest class model convergence diagnostic
Gelman-Rubin Diagnostic 2k Forest
Point
est.
97.5 quantile
Intercept
1.04
1.06
Percent Land Cover Adjacency
1.04
1.08
sigma.sq
1.00
1.00
Phi
1.04
1.04
Multivariate psrf = 1.01
Results corresponding to the 2k agriculture/forest model indicated that a relationship
exists between high BU rates and more complexly-shaped agriculture/forest land cover patches
that are aggregated with one another within 2k distances of village centers (Table 3-23).

70

Table 3-23. 2k agriculture/forest class spatial model results
Summary 2k Agriculture/Forest 3 Chains at 50,000
Parameters
50.0%
2.5%
97.5%
Intercept
-9.434 -11.430
-8.954
Shape Index Mean
2.448
1.700
4.771
Landscape Shape Index
-0.253
-0.492
-0.120
sigma.sq
2.232
1.873
2.707
Phi
366.493 334.177 489.995
effective range 3/phi
0.008
0.006
0.009
max intersite distance = 1.256
The Gelman-Rubin diagnostic indicated that convergence likely took place, although the
shape index mean value exhibited a higher value at 1.67 (Table 3-24).
Table 3-24. 2k agriculture/forest model convergence diagnostic
Gelman-Rubin Diagnostic 2k Agriculture/Forest
Point
est.
97.5 quantile
Intercept
1.19
1.61
Shape Index Mean
1.21
1.67
Landscape Shape Index
1.06
1.21
sigma.sq
1.01
1.03
Phi
1.00
1.00
Multivariate psrf = 1.08
1.15

3.5 Model Verification
Withheld data equal to 10% of each significant model’s data set acted as locations for
model verification. Predicted values were compared to observed values and calculation of the
RMSE for each model determined the “best” model to use to predict to new locations.
Table 3-25 outlines the RMSE results for the landscape-level and class-level models with
models with the lowest RMSE values from each model set highlighted in bold text.

71

Table 3-25. Root mean square error results
Root Mean Square Error (RMSE)
Scale

Model

Landscapelevel

Class-level

1_2k
1_2k
1_6k
1_6k
2k
2k
800m forest
800m wetland
800m
agriculture/forest
1_2k forest
1_2k wetland
1_2k
agriculture/forest
1_6k forest
1_6k wetland
1_6k
agriculture/forest
2k forest
2k wetland
2k agriculture/forest

Variables
WI+SHAPE_MN+NP
WI+SHAPE_MN+PRD
WI+SHAPE_MN+NP
WI+SHAPE_MN+PRD
WI+SHAPE_MN+NP
WI+SHAPE_MN+PRD
SHAPE_MN+PLADJ
SHAPE_MN

RMSE
333.780
347.565
311.843
294.507
274.337
291.361
559.395
396.001

Post-BurnIn
Iterations
500,000
500,000
500,000
500,000
500,000
500,000
100,000
50,000

SHAPE_MN+LSI
SHAPE_MN+LSI
SHAPE_MN+LSI

552.751
332.261
160.557

50,000
100,000
100,000

SHAPE_MN+LSI
SHAPE_MN
SHAPE_MN+LSI

311.324
485.752
251.310

100,000
50,000
100,000

SHAPE_MN+LSI
PLADJ
PLADJ
SHAPE_MN+LSI

361.176
213.547
420.077
231.294

50,000
50,000
50,000
50,000

RMSE values determined that the landscape-level 2k model including BU rates as the
response variable and wetness index averages, shape index means, and number of patches as
independent variables was the best-fitting landscape-level model with an RMSE of 274.337. The
best-fitting class-level model, also the overall best-fitting model, was the 1_2k wetland model
that included BU rates as the response variable and shape index means and landscape shape
index values as independent variables with an RMSE of 160.557. Fitted rates are similar to
observed rates across Benin (Figure 3-1, Figure 3-2).

72

Figure 3-1. 1_2k wetland class observed rates, 90% data

73

Figure 3-2. 1_2k wetland class fitted model, 90% data
The random spatial effects component contributed a large amount to the overall model
fit; areas plotted in red represent higher spatial random effects, while areas plotted in blue
represent lower spatial random effects (Figure 3-3). Areas exhibiting high spatial random effects
correspond to areas with higher predicted rates, demonstrating their presence in the overall
model fit.

74

Figure 3-3. 1_2k wetland class spatial random effects

3.6 Predictive Surface Model
Predicted BU rates at new locations across the study region identified areas at risk for BU
cases based on landscape configuration metrics and unknown, spatially-structured covariates
characterized by random spatial effects generated from the 1_2k wetland spatial binomial GLM
(Figure 3-4). Although this surface represents predictive BU rates, this model does not account
for population absences in a region or for any socio-behavioral activities leading to transmission;
therefore, although areas might be identified as high-risk areas, it does not necessarily mean that
transmission will take place.

75

Figure 3-4. Predictive BU rates surface across the study domain
The highest predicted rates occurred within the boundary of Benin along the southern
portions of the Oueme and Couffu Rivers. These predictions were consistent with observed rates
in these areas. Low risk areas consisted of the northern region and areas along the southern coast
in Benin. Moderate risk was identified across large portions of the study area located within the
boundary of Togo and along the Oueme River, north of high risk areas.

76

Chapter 4 DISCUSSION AND LIMITATIONS
4.1 Discussion
This study was a first attempt to quantify relationships between land cover patch patterns
indicative of anthropogenic disturbances and BU incidence in Benin. The methods used and the
approach taken to quantify these attributes were novel to BU research and highlighted the
importance of using spatial statistical methods when investigating environmental phenomena.
The first set of results demonstrated that using non-spatial statistical methods when
investigating positively, spatially-autocorrelated ecological variables can lead to overestimations
in variable significance. Non-spatial binomial GLMs inflated the influence of predictor variables
for all landscape-level models and for several class-level models. Although non-spatial models
exhibited significance in all covariates, these methods failed to account for spatial
autocorrelation among model residuals, leading to false assumptions regarding their
contributions to BU rates. Spatial binomial GLMs mitigated these effects, while identifying
important variables related to BU rates at observed locations. The addition of a random spatial
effects component partitioned unobserved covariates contributing to BU rates while identifying a
spatial structure associated with these covariates, promoting more accurate predictions at new,
unsampled locations.
Model sets constructed at a landscape-level and at a class-level enabled an in depth
analysis of land cover patch configuration relationships to BU rates within the study region.
Several model results determined that land cover patch configurations related to high BU rates at
the landscape-level supported study hypotheses that more fragmented landscapes with more
uniformly-shaped land cover patches, lying within areas more likely to accumulate water during
precipitation events surround villages with higher BU rates, although no models supported all
77

three scenarios simultaneously, and no variables emerged as dominant indicators of BU risk
across all distance intervals. These results were not surprising considering the complexity of
factors involved in determining variable selection and when deciding the scale at which to
measure natural environmental variables.
Although consistency in variable significance lacked across distances at the landscape
level, significance direction in individual variables did not change between models. Further, the
best-fitting model in the landscape-level set supported two components of the study hypotheses,
and although not the primary focus of this study, additional model results provided several
opportunities for a glimpse into potential ecological connections between land cover patch
configurations and BU rates.
Notable results included a positive relationship between patch richness density
surrounding villages with high BU rates at a 1_2k distance. Variable significance occurred at a
1_2k distance only, but the outcome suggested that a relationship exists between more diverse
landscapes closer to village centers and higher BU rates. A positive relationship between
numbers of patches at a 1_6k distance and BU rates supported the hypothesis that more
fragmented landscapes, potentially disturbed by anthropogenic activities, surround villages with
higher disease rates. Although variable significance took place at a 1_6k distance only, these
results warrant further investigation at larger distance intervals to determine whether this
phenomenon is indicative of BU risk across broader geographic regions.
Models generated at a 2k distance produced differing results, but both models indicated
that a relationship exists between higher average wetness index values and higher BU rates,
supporting the hypothesis that areas more likely to accumulate water during a precipitation event
surround villages with higher BU rates. Interestingly, average wetness index value variable

78

significance did not occur at distances closer to village centers, suggesting that a sufficient
distance may be needed to encounter rivers and tributaries of a magnitude large enough to create
local flooding events on a seasonal basis or during extreme weather events.
Two models created at 2k distances from village centers exhibited different results related
to shape index mean values. Low shape index mean values were significant at a 2km distance
when paired with numbers of patches and wetness index averages as model covariates, although
shape index mean values were not significant at a 2km distance when paired with patch richness
density and wetness index averages. One contributing factor to this difference was that shape
index mean convergence likely did not take place in the model including patch richness density
values; therefore, this model was not representative of the shape index mean distribution at a 2k
distance from village centers.
Results from the 2k model including wetness index averages, shape index means, and
numbers of patches had the lowest RMSE value of all models produced at the landscape-level
and supported the hypothesis that more uniformly-shaped land cover patches, a characteristic of
anthropogenically-disturbed landscapes, located in areas more likely to accumulate water during
a precipitation event surround villages with higher BU rates. Further investigation into these
factors at greater distances from village centers could reveal important trends in larger-scale
processes constraining finer-scale phenomena related to BU ecology in the study region.
Class-level models exhibited several consistent results at various distances and across
individual class types, fulfilling the second objective of this study to determine whether a
relationship exists between land cover patch configurations indicative of potential anthropogenic
disturbances corresponding to specific land cover types and high BU rates. While several
significant model outcomes occurred, these results did not support the study hypotheses.

79

Significant forest, wetland, and mixed agriculture/forest patch shape index mean values
suggested that more natural or undisturbed patch shapes corresponding to these classes surround
villages with higher BU rates, and consistent model outcomes occurred across all distance
intervals.
While results corresponding to patch shape complexity did not support the study
hypothesis, these results were particularly interesting when observing the mixed
agriculture/forest class because the composition of this class included some level of
anthropogenic disturbance inherently; therefore, model outcomes provided insight into potential
disturbance patterns. Natural patch shape complexities corresponding to this land cover class
suggested characteristics representative of natural forest cover because vegetation planted by
humans could not likely produce this pattern, indicating that agriculture plots likely intrude into
forest patches undisturbed previously. The agriculture/forest class models maintained
significance at all distance intervals, signifying that a need exists for further investigation into
agriculture planting practices within forested areas and relationships to BU emergence.
Additionally, individual forest, wetland or mixed agriculture/forest class aggregation
values related to higher BU rates did not support the study hypothesis that more fragmented land
cover corresponding to these classes surrounds villages with higher BU rates. Significant model
outcomes reported lower landscape shape index values and higher percent land cover adjacency
values consistently, suggesting that more aggregated land cover patches corresponding to these
classes surround villages with higher BU rates across all distance intervals.
Several potential reasons may explain class-level model outcomes. One possibility is that
anthropogenic disturbance patterns related to the individual classes investigated in this study may
not follow patterns recognized previously. Studies demonstrating uniformly-shaped land cover

80

patches as those created by anthropogenic activities took place largely in European countries and
within the United States where access to heavy farming machinery was readily available and
long-standing property boundaries may demarcate land cover patches more abruptly.
While disturbance patterns may not follow those quantified in western studies, a more
likely explanation has to do with the scale at which the quantification of the study area took
place. Ikonos 4m resolution satellite imagery from January 2009 revealed several anthropogenic
landscape disturbances demarcated by uniform shapes across a 10k x 10k BU endemic area in
the mid-western portion of the study region. Of particular interest were partitioned rice paddy
plots within natural wetland boundaries (Figure 4-1). These areas appeared undisturbed until
observed at a finer resolution, suggesting that 30m resolution Landsat imagery could not identify
within patch anthropogenic disturbance revealed at a finer resolution, and this factor may have
influenced class-level results.

Figure 4-1. Anthropogenic disturbance within wetland land cover patch

81

A third possible explanation may have to do with MU pathogen abundance in the
environment. To date, unknown ecological factors drive MU abundance, but landscapes
experiencing early disturbance stages may experience elevated pathogen levels. As residents
begin to intrude into habitats undisturbed previously to plant agriculture or to collect fuel wood,
their activities likely follow natural land cover boundaries before significant landscape
alterations take place. Future quantitative PCR applications have the potential to provide
knowledge related to pathogen abundance in the environment, the results of which could provide
critical information linking MU presence and abundance to specific land cover types and uses
during multiple succession stages (Merritt et al., 2005).
Finally, landscape-level model results differed from class-level model results and
supported the study hypotheses, suggesting that important land cover categories may not have
been included at the class-level in this study. Alternatively, while uniformly-shaped patches
corresponding to specific classes did not exhibit significance, when measurement of these patch
shapes took place in conjunction with measurements from the entire landscape, uniformlyshaped patches emerged as an important predictor variable; therefore, the importance of this
phenomenon may be unrelated to specific class types.
The “best” model with the lowest RMSE value from both model sets corresponded to the
1_2k wetland class model with BU rates as the response variable and shape index mean values
and landscape shape index values as the independent variables. Interestingly, the independent
variables did not maintain significance in the spatial GLM; therefore, the spatial random effects
component contributed to the majority of the model fit. These results confirmed that a spatial
structure exists for processes driving BU rates in Benin, although at the end of this study, these
covariates remain unknown.

82

Although measuring the contribution of land cover composition alone was not the
primary objective behind class-level model construction, the fact that the “best” model
corresponded to the wetland class while land cover configuration variables corresponding to this
class and at this distance were not significant is an important development. These results suggest
that wetland land cover composition within relatively close distances from village centers may
play an important role in the distribution of BU rates in Benin because additional models at 1_2k
distances, even those maintaining significant variables, exhibited higher RMSE values. These
results supported previous studies implicating close proximities to slow-moving or stagnant
water bodies as a BU risk factor; consequently, a need exists for further investigation into
wetland characteristics at multiple scales near villages experiencing high BU rates.
The creation of a continuous risk surface map across the study region fulfilled the final
study objective. Although “best” model predictor variables alone could not predict BU rates
effectively, the incorporation of the random spatial effects component enabled accurate
predictions at new, unsampled locations while accounting for uncertainty across the study
domain. To our knowledge, this study was the first to produce a continuous risk surface based on
environmental covariates and the only study to account for and to utilize unknown, spatial
covariates driving disease incidence in model predictions.
Predicted rates derived from the mean posterior predictive distribution of the “best”
model produced a BU risk surface map from which to identify high risk regions across the study
domain (Figure 3-4). Predicted rates in Benin followed a similar pattern to the observed case data
with higher rates along the Oueme River in the east and along the Couffu River in the western
section of the country. A large floodplain exists in the eastern portion of the country where the
Oueme and Zou Rivers join before draining into the Gulf of Guinea. Land cover consists of

83

mixed agriculture, forest, and shrubland with wetlands situated in the south where the river flows
into Lake Nokoué before emptying into the Bight of Benin. Several wetlands extend from the
Couffu River with one group situated near the town of Tandji, located within a high predicted
rate area, around which BU cases occur regularly. Land cover within this region consists
primarily of mixed agriculture/shrub, forest, mixed agriculture/forest, and wetlands.
Low prediction rates along the southern coast coincided with few observed cases in the
region, where brackish waters may impact environmental conditions suitable to the MU
pathogen, and toward the northern portion of the study region where higher elevations separate
the two endemic regions and farther north where climate conditions begin to change to a dryer,
less humid environment.
Predicted rates in the study area within the boundary of Togo exhibited moderate values
with less variability than rates predicted within the boundary of Benin. This may be due to
increased distance from observed locations. As distance from known values exceeds the effective
range determined by the spatial structure of the observed data set, predicted rates have a
tendency to move toward a mean predictive value, and although this phenomenon may have
impacted predicted rates within Togo, the results were still interesting and relevant.
Southern Togo shares a similar climate to southern Benin, falling within the Dahomey
Gap. Swampy floodplains reside along the southern portion of the Mono River beginning at an
approximate latitude of 6.9333 N, stretching to the southern coast, while an additional wetland
consisting of herb swamp, swampy forest, and grassy floodplain lies across the border with
Benin (Hughes and Hughes, 1992). The lagoon system in Togo does not experience tidal flow
except during periods of high rainfall, and therefore, is considered a freshwater system.

84

Most notable were moderate-to-high rates predicted along the Mono River west of the
Tandji foci in Benin at the border between the two countries. Although located within a wetland
system, few reported BU cases exist in this region following construction of the Nangbeto dam
in 1987 (R. Christian Johnson, personal communication, March 24, 2009). One hypothesis is that
a reduction in cases occurred because of controlled fluctuations in water levels, reducing
seasonal flooding impacts in the region. While the environmental conditions may be comparable
to those identified as high risk environments in Benin, the model could not account for
anthropogenic interference with river flow and identified this area as having a moderate-to-high
BU risk.
The region south of the Nangbeto dam may prove to be an important surveillance area.
Although the dam construction reduced seasonal flooding risk in the Mono River region,
unusually high rainfall contributed to tragedy when the opening of a sluice gate relieved water
pressure in the reservoir behind the dam, causing the river to burst its banks, wiping away houses
and farms on November 2, 2010 (Ghana News Link, 2010). While troubling, these flooding
events may provide a unique control from which to observe whether BU cases emerge, providing
an opportunity to gain a better understanding of the role in which flooding contributes to BU
incidence.
While multiple, unknown, environmental variables must interact for the bacterium to
flourish and transmission to humans likely requires specific socio-behavioral factors
unaccounted for in this model, a risk surface incorporating the spatial structure of these unknown
variables provided an ideal tool from which to identify high BU risk regions despite little
knowledge regarding environmental factors contributing to the disease. Further, BU disease does
not observe administrative boundaries; therefore, the creation of a continuous surface

85

transcending artificial borders eliminated areal data set constraints, the benefits of which were
demonstrated in predicted rates across Togo where little case data exists.
Although a temptation exists to observe this risk surface as absolute, the natural
environment is not a static phenomenon, nor is BU incidence. This risk surface represents one
snapshot in time, based on land cover configurations derived from one satellite image during one
season and from the spatial structure of factors contributing to BU incidence in 2004 and 2005.
However, observing where BU transmission could likely take place if persons encountered
similar environments provides a first step toward surveillance and prevention, while creating a
framework from which to target future environmental sampling and research efforts.

4.2 Limitations
Several limitations existed in this study, some of which related directly to data used in the
analysis. LULC data derived from 2000 satellite imagery was used to analyze 2004 and 2005 BU
case data. Challenges existed in acquiring suitable satellite imagery from the study region
because of the presence of cloud cover in a large proportion of the imagery; therefore, this study
assumed that the LULC within the study region did not change substantially within a 3-4 year
time period. Suitable ground truth data were not available to validate the LULC classification. A
10 year gap between image acquisition and limited ground truth data existed; therefore, this
study relied on visual interpretation methods along with aggregated land cover classes to create
the land cover classification.
Characterization of land cover configurations took place at one scale using 30m
resolution data. Important organism responses may be linked to these variables, but could occur
at scales differing from those at which variable measurement took place. A 30m resolution may
not be the appropriate resolution to characterize processes driving MU growth and BU
86

emergence; therefore, a need exists for additional investigation into land cover configuration
patterns related to BU rates at different spatial resolutions.
Although active surveillance by government health officials identified BU cases within
the study region, underreporting may have impacted study results. Further, identification of BU
negative villages corresponded to a 2004 and 2005 time period; these villages may have had
positive cases in previous or subsequent years. Additionally, this study assumed that BU
transmission occurred near the village of residence. Travel between regions and migration was
not known and therefore, was not incorporated into this study.
Environmental variables with a seasonal or a highly temporal nature were not included in
this study; for example, precipitation, temperature, or relative humidity factors. Six weather
stations with incomplete inventories exist within the study region, making data interpolation
across the study domain questionable. Beyond data quality issues, unknown lag times between
suitable environmental conditions and pathogen proliferation, inoculation and ulcer presentation,
and ulcer presentation and hospital treatment make linking BU cases to specific environmental
events challenging. Coarser resolution climate data were considered and an exploratory analysis
of several BioClim 30 year average variables did not reveal significant relationships between
these variables and BU rates. Likely, the data was too coarse both spatially and temporally to
identify important relationships. A need exists for further investigation into the relationship
between climate variables and BU cases in West Africa.
Finally, scaling of model coordinates to correspond to a one-to-one surface area may
have introduced uncertainty, although the location of the study area close to the equator reduced
the impact of this phenomenon.

87

4.3 Conclusions and Future Directions
The role of anthropogenic ecosystem disturbances in the emergence of environmental
bacterial infections is poorly understood. This study was a first attempt to link land cover
configurations representative of anthropogenic disturbances to the environmental bacterial
infection Buruli ulcer disease. Although mixed results did not suggest a definite trend toward
positive linkages between land cover patch configurations representative of anthropogenic
disturbances and BU rates, study results identified several significant variables, warranting future
investigations into these factors at different scales.
Beyond the novel exploratory analysis outlined above, a major contribution of this study
included the incorporation of a modeling component that partitioned the spatial structure of
missing variables, providing a structure from which to predict BU rates to new locations without
strong knowledge of environmental factors contributing to disease distribution. The resulting
continuous BU risk surface is the first of its kind and marks the potential to develop and to target
surveillance efforts. The ability to predict potential risk adequately to locations where little data
availability exists provided a first step toward prevention, while creating a tool from which a
more systematic and controlled site selection process may be used to target future environmental
sampling research.
Future directions include model refinement and comparisons of model outcomes
generated from this study with additional modeling approaches, for example those using an
extreme value link function that accounts for data sets where the number of zeros is vastly
greater than the number of ones. Acquisition of climate variables corresponding to endemic
regions may reveal valuable linkages between weather events and landscape attributes related to
disease incidence, and future studies into wetland land cover using higher resolution satellite

88

imagery in conjunction with environmental samples could identify relevant links between
anthropogenic wetland disturbances and BU disease emergence.
Finally and most importantly, continued acquisition of accurate, and georeferenced case
data along with georeferenced pathogen data will provide the best opportunity for robust
empirical studies of linkages between ecological factors, anthropogenic activities, and BU
transmission. Combining these data with the continued efforts of multidisciplinary research
teams, international governments, and aid agencies will provide the tools necessary to one day
understand the mystery behind Buruli ulcer disease.

89

APPENDICES

90

Appendix A
Correlation Matrices

91

Table A-1. Landscape-level 400m correlation matrix

BU
RATE
BU RATE
NP
LSI
SHAPE MN
FRAC MN
PARA MN
PLADJ
PRD
SHDI
WIAVGG

1.000
0.048
0.056
0.138
0.137
0.028
-0.050
0.008
0.033
0.042

NP
0.048
1.000
0.906
0.073
0.000
0.514
0.910
0.642
0.814
0.072

400m Spearman Rank Order Correlation Matrix
SHAPE FRAC
PARA
LSI
PLADJ
MN
MN
MN
0.056
0.906
1.000
0.263
0.272
0.302
0.992
0.554
0.870
0.048

0.138
-0.073
0.263
1.000
0.899
-0.278
-0.255
-0.105
0.181
-0.071

0.137
0.000
0.272
0.899
1.000
-0.318
-0.263
-0.022
0.205
-0.034

0.028
0.514
0.302
-0.278
-0.318
1.000
-0.317
0.345
0.169
-0.007

92

-0.050
-0.910
-0.992
-0.255
-0.263
-0.317
1.000
-0.548
-0.868
-0.047

PRD

SHDI

WIAVGG

0.008
0.642
0.554
0.105
0.022
0.345
0.548
1.000
0.660
0.096

0.033
0.814
0.870
0.181
0.205
0.169
0.868
0.660
1.000
0.131

0.042
0.072
0.048
-0.071
-0.034
-0.007
-0.047
0.096
0.131
1.000

Table A-2. Landscape-level 500m correlation matrix

BU RATE
NP
LSI
SHAPE MN
FRAC MN
PARA MN
PLADJ
PRD
SHDI
WIAVGG

BU
RATE
1.000
0.005
0.035
0.123
0.142
0.099
0.028
0.001
0.043
0.068

NP
0.005
1.000
0.900
0.036
0.015
0.463
0.902
0.669
0.791
0.100

500m Spearman Rank Order Correlation Matrix
SHAPE FRAC
PARA
LSI
PLADJ
MN
MN
MN
0.035
0.123
0.142
-0.099
-0.028
0.900
-0.036
0.015
0.463
-0.902
1.000
0.321
0.306
0.229
-0.991
0.321
1.000
0.913
-0.395
-0.322
0.306
0.913
1.000
-0.448
-0.309
0.229
-0.395
-0.448
1.000
-0.229
0.991
-0.322
-0.309
-0.229
1.000
0.573
-0.045
0.007
0.287
-0.571
0.853
0.277
0.289
0.067
-0.853
0.068
-0.054
0.024
-0.022
-0.071

93

PRD

SHDI

WIAVGG

0.001
0.669
0.573
-0.045
0.007
0.287
-0.571
1.000
0.688
0.152

0.043
0.791
0.853
0.277
0.289
0.067
-0.853
0.688
1.000
0.139

0.068
0.100
0.068
-0.054
0.024
-0.022
-0.071
0.152
0.139
1.000

Table A-3. Landscape-level 600m correlation matrix

BU RATE
NP
LSI
SHAPE MN
FRAC MN
PARA MN
PLADJ
PRD
SHDI
WIAVGG

BU
RATE
1.000
0.014
0.029
0.057
0.032
0.087
0.022
0.014
0.057
0.089

NP
0.014
1.000
0.908
0.081
0.010
0.393
0.910
0.613
0.773
0.078

600m Spearman Rank Order Correlation Matrix
SHAPE FRAC PARA
LSI
PLADJ
MN
MN
MN
0.029
0.057
0.032 -0.087
-0.022
0.908
-0.081 -0.010
0.393
-0.910
1.000
0.255
0.258
0.168
-0.991
0.255
1.000
0.903 -0.452
-0.255
0.258
0.903
1.000 -0.479
-0.257
0.168
-0.452 -0.479
1.000
-0.180
0.991
-0.255 -0.257 -0.180
1.000
0.512
-0.085
0.027
0.195
-0.518
0.831
0.229
0.245
0.015
-0.837
0.069
-0.011
0.006 -0.021
-0.072

94

PRD

SHDI WIAVGG

0.014
0.613
0.512
0.085
0.027
0.195
0.518
1.000
0.629
0.131

0.057
0.773
0.831
0.229
0.245
0.015
0.837
0.629
1.000
0.143

0.089
0.078
0.069
-0.011
0.006
-0.021
-0.072
0.131
0.143
1.000

Table A-4. Landscape-level 700m correlation matrix

BU RATE
NP
LSI
SHAPE MN
FRAC MN
PARA MN
PLADJ
PRD
SHDI
WIAVGG

BU
RATE
1.000
0.016
0.022
0.088
0.044
-0.057
-0.013
0.060
0.065
0.111

700m Spearman Rank Order Correlation Matrix
SHAPE FRAC PARA
NP
LSI
PLADJ
MN
MN
MN
0.016 0.022
0.088
0.044
0.057
-0.013
1.000 0.897
-0.049
0.021
0.440
-0.902
0.897 1.000
0.316
0.268
0.201
-0.991
0.049 0.316
1.000
0.896
0.466
-0.307
0.021 0.268
0.896
1.000
0.507
-0.261
0.440 0.201
-0.466
0.507
1.000
-0.209
0.902 0.991
-0.307
0.261
0.209
1.000
0.572 0.469
-0.096
0.034
0.209
-0.478
0.745 0.806
0.270
0.257
0.039
-0.814
0.073 0.044
-0.042
0.004
0.047
-0.045

95

PRD

SHDI

WIAVGG

0.060
0.572
0.469
0.096
0.034
0.209
0.478
1.000
0.621
0.185

0.065
0.745
0.806
0.270
0.257
0.039
0.814
0.621
1.000
0.137

0.111
0.073
0.044
-0.042
0.004
-0.047
-0.045
0.185
0.137
1.000

Table A-5. Landscape-level 800m correlation matrix

BU RATE
NP
LSI
SHAPE MN
FRAC MN
PARA MN
PLADJ
PRD
SHDI
WIAVGG

BU
RATE
1.000
0.009
0.024
0.118
0.073
0.142
0.013
0.052
0.074
0.131

NP
0.009
1.000
0.910
0.110
0.073
0.472
0.914
0.556
0.741
0.091

800m Spearman Rank Order Correlation Matrix
SHAPE FRAC PARA
LSI
PLADJ
MN
MN
MN
0.024
0.118
0.073
0.142
-0.013
0.910
-0.110
0.073
0.472
-0.914
1.000
0.232
0.202
0.270
-0.989
0.232
1.000
0.909
0.511
-0.225
0.202
0.909
1.000
0.598
-0.196
0.270
-0.511 -0.598
1.000
-0.280
0.989
-0.225 -0.196 -0.280
1.000
0.447
-0.126 -0.054
0.231
-0.459
0.782
0.236
0.234
0.094
-0.794
0.045
-0.024
0.019 -0.012
-0.051

96

PRD

SHDI

WIAVGG

0.052
0.556
0.447
0.126
0.054
0.231
0.459
1.000
0.621
0.206

0.074
0.741
0.782
0.236
0.234
0.094
0.794
0.621
1.000
0.161

0.131
0.091
0.045
-0.024
0.019
-0.012
-0.051
0.206
0.161
1.000

Table A-6. Landscape-level 900m correlation matrix

BU RATE
NP
LSI
SHAPE MN
FRAC MN
PARA MN
PLADJ
PRD
SHDI
WIAVGG

BU
RATE
1.000
-0.013
0.025
0.122
0.071
-0.137
-0.013
0.070
0.076
0.135

NP
0.013
1.000
0.901
0.108
0.037
0.421
0.907
0.545
0.732
0.076

900m Spearman Rank Order Correlation Matrix
SHAPE FRAC PARA
LSI
PLADJ
MN
MN
MN
0.025
0.122
0.071
-0.137
-0.013
0.901
-0.108
0.037
0.421
-0.907
1.000
0.249
0.239
0.200
-0.987
0.249
1.000
0.889
-0.484
-0.238
0.239
0.889
1.000
-0.558
-0.231
0.200
-0.484
0.558
1.000
-0.217
0.987
-0.238
0.231
-0.217
1.000
0.438
-0.130
0.057
0.200
-0.450
0.767
0.221
0.262
0.011
-0.783
0.050
-0.002
0.009
-0.034
-0.057

97

PRD

SHDI

0.070
0.545
0.438
0.130
0.057
0.200
0.450
1.000
0.611
0.199

0.076
0.732
0.767
0.221
0.262
0.011
-0.783
0.611
1.000
0.169

WIAVGG
0.135
0.076
0.050
-0.002
0.009
-0.034
-0.057
0.199
0.169
1.000

Table A-7. Landscape-level 1k correlation matrix

BU RATE
NP
LSI
SHAPE MN
FRAC MN
PARA MN
PLADJ
PRD
SHDI
WIAVGG

BU
RATE
1.000
0.005
0.030
0.112
0.080
0.083
0.016
0.083
0.084
0.133

NP
0.005
1.000
0.906
0.084
0.004
0.431
0.911
0.541
0.706
0.061

1k Spearman Rank Order Correlation Matrix
SHAPE
FRAC
PARA
LSI
PLADJ
MN
MN
MN
0.030
0.112
0.080
-0.083
-0.016
0.906
-0.084
0.004
0.431
-0.911
1.000
0.257
0.278
0.228
-0.986
0.257
1.000
0.910
-0.497
-0.249
0.278
0.910
1.000
-0.565
-0.271
0.228
-0.497
-0.565
1.000
-0.244
0.986
-0.249
-0.271
-0.244
1.000
0.445
-0.087
0.010
0.201
-0.461
0.751
0.266
0.324
0.000
-0.767
0.048
0.036
0.059
-0.033
-0.055

98

PRD
0.083
0.541
0.445
0.087
0.010
0.201
0.461
1.000
0.597
0.236

SHDI
0.084
0.706
0.751
0.266
0.324
0.000
0.767
0.597
1.000
0.165

WIAVGG
0.133
0.061
0.048
0.036
0.059
-0.033
-0.055
0.236
0.165
1.000

Table A-8. Landscape-level 1_1k correlation matrix

BU RATE
NP
LSI
SHAPE MN
FRAC MN
PARA MN
PLADJ
PRD
SHDI
WIAVGG

BU
RATE
1.000
0.000
0.038
0.120
0.100
0.143
0.024
0.127
0.095
0.133

NP
0.000
1.000
0.906
-0.088
0.043
0.438
-0.913
0.543
0.698
0.051

1_1k Spearman Rank Order Correlation Matrix
SHAPE FRAC PARA
LSI
PLADJ
MN
MN
MN
0.038
0.120
0.100
-0.143
-0.024
0.906
-0.088
0.043
0.438
-0.913
1.000
0.254
0.313
0.234
-0.985
0.254
1.000
0.907
-0.468
-0.244
0.313
0.907
1.000
-0.478
-0.303
0.234
-0.468
-0.478
1.000
-0.249
-0.985
-0.244
-0.303
-0.249
1.000
0.474
-0.033
0.083
0.158
-0.491
0.744
0.282
0.382
0.006
-0.762
0.039
0.039
0.052
-0.058
-0.049

99

PRD

SHDI

WIAVGG

0.127
0.543
0.474
-0.033
0.083
0.158
-0.491
1.000
0.616
0.282

0.095
0.698
0.744
0.282
0.382
0.006
-0.762
0.616
1.000
0.174

0.133
0.051
0.039
0.039
0.052
-0.058
-0.049
0.282
0.174
1.000

Table A-9. Landscape-level 1_2k correlation matrix

BU RATE
NP
LSI
SHAPE MN
FRAC MN
PARA MN
PLADJ
PRD
SHDI
WIAVGG

BU
RATE
1.000
0.006
0.044
0.094
0.029
0.072
0.029
0.137
0.093
0.141

NP
0.006
1.000
0.908
-0.090
0.007
0.453
-0.914
0.532
0.684
0.039

1_2k Spearman Rank Order Correlation Matrix
SHAPE FRAC PARA
LSI
PLADJ
MN
MN
MN
0.044
0.094
0.029 -0.072
-0.029
0.908
-0.090
0.007
0.453
-0.914
1.000
0.252
0.280
0.248
-0.985
0.252
1.000
0.895 -0.502
-0.245
0.280
0.895
1.000 -0.550
-0.278
0.248
-0.502 -0.550
1.000
-0.260
-0.985
-0.245 -0.278 -0.260
1.000
0.473
0.004
0.073
0.188
-0.490
0.734
0.296
0.344
0.038
-0.750
0.036
0.058
0.066 -0.077
-0.047

100

PRD

SHDI

WIAVGG

0.137
0.532
0.473
0.004
0.073
0.188
-0.490
1.000
0.607
0.302

0.093
0.684
0.734
0.296
0.344
0.038
-0.750
0.607
1.000
0.178

0.141
0.039
0.036
0.058
0.066
-0.077
-0.047
0.302
0.178
1.000

Table A-10. Landscape-level 1_3k correlation matrix

BU RATE
NP
LSI
SHAPE MN
FRAC MN
PARA MN
PLADJ
PRD
SHDI
WIAVGG

BU
RATE
1.000
0.020
0.053
0.126
0.075
0.062
0.038
0.144
0.094
0.145

NP
0.020
1.000
0.913
-0.088
0.014
0.462
-0.923
0.509
0.686
0.020

1_3k Spearman Rank Order Correlation Matrix
SHAPE FRAC PARA
LSI
PLADJ
MN
MN
MN
0.053
0.126
0.075 -0.062
-0.038
0.913
-0.088
0.014
0.462
-0.923
1.000
0.237
0.277
0.268
-0.985
0.237
1.000
0.907 -0.478
-0.227
0.277
0.907
1.000 -0.520
-0.267
0.268
-0.478 -0.520
1.000
-0.288
-0.985
-0.227 -0.267 -0.288
1.000
0.457
0.022
0.111
0.143
-0.472
0.728
0.287
0.360
0.049
-0.744
0.018
0.038
0.046 -0.076
-0.029

101

PRD

SHDI

WIAVGG

0.144
0.509
0.457
0.022
0.111
0.143
-0.472
1.000
0.603
0.338

0.094
0.686
0.728
0.287
0.360
0.049
-0.744
0.603
1.000
0.172

0.145
0.020
0.018
0.038
0.046
-0.076
-0.029
0.338
0.172
1.000

Table A-11. Landscape-level 1_5k correlation matrix

RATE
NP
LSI
SHAPE MN
FRAC MN
PARA MN
PLADJ
PRD
SHDI
WIAVGG

BU
RATE
1.000
0.029
0.055
0.110
0.057
0.069
0.040
0.142
0.093
0.146

NP
0.029
1.000
0.915
-0.107
0.013
0.526
-0.925
0.490
0.665
0.011

1_5k Spearman Rank Order Correlation Matrix
SHAPE FRAC PARA
LSI
PLADJ
MN
MN
MN
0.055
0.110
0.057 -0.069
-0.040
0.915
-0.107
0.013
0.526
-0.925
1.000
0.215
0.267
0.356
-0.985
0.215
1.000
0.880 -0.442
-0.200
0.267
0.880
1.000 -0.508
-0.252
0.356
-0.442 -0.508
1.000
-0.377
-0.985
-0.200 -0.252 -0.377
1.000
0.431
0.030
0.122
0.182
-0.446
0.710
0.280
0.345
0.136
-0.725
0.010
0.073
0.086 -0.065
-0.022

102

PRD

SHDI

WIAVGG

0.142
0.490
0.431
0.030
0.122
0.182
-0.446
1.000
0.588
0.341

0.093
0.665
0.710
0.280
0.345
0.136
-0.725
0.588
1.000
0.175

0.146
0.011
0.010
0.073
0.086
-0.065
-0.022
0.341
0.175
1.000

Table A-12. Landscape-level 1_6k correlation matrix

BU
RATE

NP

1_6k Spearman Rank Order Correlation Matrix
SHAPE FRAC PARA
LSI
PLADJ
MN
MN
MN

PRD

SHDI

WIAVGG

BU RATE

1.000

0.028

0.055

0.079

0.030

0.030

-0.039

0.118

0.089

0.146

NP

0.028

1.000

0.915

-0.057

0.089

0.508

-0.925

0.476

0.671

0.011

LSI

0.055

0.915

1.000

0.257

0.323

0.343

-0.987

0.409

0.706

0.011

SHAPE MN

0.079

0.057

0.257

1.000

0.875

0.390

-0.245

0.040

0.303

0.067

FRAC MN

0.030

0.089

0.323

0.875

1.000

0.458

-0.313

0.139

0.386

0.068

PARA MN

0.030

0.508

0.343

-0.390

-0.458

1.000

-0.363

0.186

0.144

-0.082

PLADJ

0.039

0.925

0.987

-0.245

-0.313

0.363

1.000

0.426

0.718

-0.020

PRD

0.118

0.476

0.409

0.040

0.139

0.186

-0.426

1.000

0.595

0.317

SHDI

0.089

0.671

0.706

0.303

0.386

0.144

-0.718

0.595

1.000

0.182

WIAVGG

0.146

0.011

0.011

0.067

0.068

0.082

-0.020

0.317

0.182

1.000

103

Table A-13. Landscape-level 1_7k correlation matrix

RATE
RATE
NP
LSI
SHAPE MN
FRAC MN
PARA MN
PLADJ
PRD
SHDI
WIAVGG

1.000
0.028
0.054
0.097
0.081
0.042
0.039
0.112
0.090
0.146

NP
0.028
1.000
0.914
-0.086
0.042
0.497
-0.924
0.451
0.666
0.006

1_7k Spearman Rank Order Correlation Matrix
SHAPE FRAC PARA
LSI
PLADJ
MN
MN
MN
0.054
0.097
0.081
-0.042
-0.039
0.914
-0.086
0.042
0.497
-0.924
1.000
0.232
0.289
0.331
-0.988
0.232
1.000
0.877
-0.413
-0.219
0.289
0.877
1.000
-0.484
-0.275
0.331
-0.413 -0.484
1.000
-0.350
-0.988
-0.219 -0.275
-0.350
1.000
0.399
0.055
0.138
0.159
-0.416
0.702
0.282
0.360
0.140
-0.714
0.009
0.065
0.066
-0.086
-0.018

104

PRD

SHDI

WIAVGG

0.112
0.451
0.399
0.055
0.138
0.159
-0.416
1.000
0.592
0.337

0.090
0.666
0.702
0.282
0.360
0.140
-0.714
0.592
1.000
0.190

0.146
0.006
0.009
0.065
0.066
-0.086
-0.018
0.337
0.190
1.000

Table A-14. Landscape-level 1_8k correlation matrix

RATE
RATE
NP
LSI
SHAPE MN
FRAC MN
PARA MN
PLADJ
PRD
SHDI
WIAVGG

1.000
0.035
0.058
0.089
0.054
0.040
0.045
0.118
0.093
0.148

NP
0.035
1.000
0.914
-0.102
0.023
0.520
-0.922
0.462
0.658
-0.004

1_8k Spearman Rank Order Correlation Matrix
SHAPE
FRAC
PARA
LSI
PLADJ
MN
MN
MN
0.058
0.089
0.054
-0.040
-0.045
0.914
-0.102
0.023
0.520
-0.922
1.000
0.218
0.271
0.357
-0.988
0.218
1.000
0.889
-0.444
-0.208
0.271
0.889
1.000
-0.512
-0.259
0.357
-0.444
-0.512
1.000
-0.373
-0.988
-0.208
-0.259
-0.373
1.000
0.404
0.050
0.127
0.168
-0.418
0.698
0.289
0.376
0.125
-0.710
0.005
0.058
0.068
-0.093
-0.012

105

PRD

SHDI

0.118
0.462
0.404
0.050
0.127
0.168
-0.418
1.000
0.586
0.340

0.093
0.658
0.698
0.289
0.376
0.125
-0.710
0.586
1.000
0.193

WIAVGG
0.148
-0.004
0.005
0.058
0.068
-0.093
-0.012
0.340
0.193
1.000

Table A-15. Landscape-level 1_9k correlation matrix

RATE
RATE
NP
LSI
SHAPE MN
FRAC MN
PARA MN
PLADJ
PRD
SHDI
WIAVGG

1.000
0.034
0.058
0.088
0.064
0.028
0.044
0.129
0.091
0.150

NP
0.034
1.000
0.912
-0.083
0.068
0.526
-0.920
0.453
0.642
-0.021

1_9k Spearman Rank Order Correlation Matrix
SHAPE
FRAC
PARA
LSI
PLADJ
MN
MN
MN
0.058
0.088
0.064
-0.028
-0.044
0.912
-0.083
0.068
0.526
-0.920
1.000
0.241
0.320
0.355
-0.989
0.241
1.000
0.885
-0.423
-0.229
0.320
0.885
1.000
-0.467
-0.307
0.355
-0.423
-0.467
1.000
-0.369
-0.989
-0.229
-0.307
-0.369
1.000
0.403
0.094
0.171
0.191
-0.418
0.689
0.310
0.398
0.156
-0.703
-0.005
0.082
0.064
-0.106
-0.002

106

PRD

SHDI

0.129
0.453
0.403
0.094
0.171
0.191
-0.418
1.000
0.594
0.344

0.091
0.642
0.689
0.310
0.398
0.156
-0.703
0.594
1.000
0.197

WIAVGG
0.150
-0.021
-0.005
0.082
0.064
-0.106
-0.002
0.344
0.197
1.000

Table A-16. Landscape-level 2k correlation matrix

BU RATE
NP
LSI
SHAPE MN
FRAC MN
PARA MN
PLADJ
PRD
SHDI
WIAVGG

BU
RATE
1.000
0.033
0.055
0.091
0.056
0.042
0.042
0.121
0.089
0.149

NP
0.033
1.000
0.913
-0.092
0.088
0.555
-0.920
0.447
0.636
-0.029

2k Spearman Rank Order Correlation Matrix
SHAPE
FRAC
PARA
LSI
PLADJ
MN
MN
MN
0.055
0.091
0.056
-0.042
-0.042
0.913
-0.092
0.088
0.555
-0.920
1.000
0.226
0.330
0.394
-0.990
0.226
1.000
0.866
-0.390
-0.214
0.330
0.866
1.000
-0.422
-0.317
0.394
-0.390
-0.422
1.000
-0.406
-0.990
-0.214
-0.317
-0.406
1.000
0.388
0.099
0.191
0.173
-0.402
0.681
0.309
0.410
0.176
-0.694
-0.017
0.089
0.056
-0.087
0.009

107

PRD

SHDI

0.121
0.447
0.388
0.099
0.191
0.173
-0.402
1.000
0.580
0.344

0.089
0.636
0.681
0.309
0.410
0.176
-0.694
0.580
1.000
0.198

WIAVGG
0.149
-0.029
-0.017
0.089
0.056
-0.087
0.009
0.344
0.198
1.000

Table A-17. 800m forest class correlation matrix
800m Forest Class Spearman Rank Order Correlation Matrix
SHAPE
NP
LSI
CLUMPY PLADJ
MN
NP
1.000
0.853
-0.003
-0.463
0.019
LSI
0.853
1.000
0.444
-0.335
0.257
SHAPE MN
-0.003
0.444
1.000
0.247
0.606
CLUMPY
-0.463
-0.335
0.247
1.000
0.709
PLADJ
0.019
0.257
0.606
0.709
1.000
AI
-0.442
-0.305
0.271
0.998
0.727

AI
-0.442
-0.305
0.271
0.998
0.727
1.000

Table A-18. 800m wetland class correlation matrix
800m Wetland Class Spearman Rank Order Correlation Matrix
SHAPE
NP
LSI
CLUMPY PLADJ
MN
NP
1.000
0.772
-0.238
-0.445
0.036
LSI
0.772
1.000
0.347
-0.247
0.373
SHAPE MN
-0.238
0.347
1.000
0.450
0.699
CLUMPY
-0.445
-0.247
0.450
1.000
0.719
PLADJ
0.036
0.373
0.699
0.719
1.000
AI
-0.354
-0.112
0.514
0.971
0.804

AI
-0.354
-0.112
0.514
0.971
0.804
1.000

Table A-19. 800m agriculture/forest class correlation matrix
800m Agriculture/Forest Class Spearman Rank Order Correlation Matrix
SHAPE
NP
LSI
CLUMPY PLADJ
AI
MN
NP
1.000
0.880
0.075
-0.426
0.188
-0.379
LSI
0.880
1.000
0.472
-0.170
0.525
-0.110
SHAPE MN
0.075
0.472
1.000
0.456
0.851
0.482
CLUMPY
-0.426
-0.170
0.456
1.000
0.614
0.992
PLADJ
0.188
0.525
0.851
0.614
1.000
0.648
AI
-0.379
-0.110
0.482
0.992
0.648
1.000

108

Table A-20. 1_2k forest class correlation matrix
1_2k Forest Class Spearman Rank Order Correlation Matrix
SHAPE
NP
LSI
CLUMPY
PLADJ
MN
NP
1.000
0.946
0.182
-0.441
0.232
LSI
0.946
1.000
0.433
-0.396
0.328
SHAPE MN
0.182
0.433
1.000
0.253
0.660
CLUMPY
-0.441
-0.396
0.253
1.000
0.630
PLADJ
0.232
0.328
0.660
0.630
1.000
AI
-0.409
-0.363
0.275
0.998
0.650

AI
-0.409
-0.363
0.275
0.998
0.650
1.000

Table A-21. 1_2k wetland class correlation matrix
1_2k Wetland Class Spearman Rank Order Correlation Matrix
SHAPE
NP
LSI
CLUMPY
PLADJ
MN
NP
1.000
0.842
-0.095
-0.287
0.267
LSI
0.842
1.000
0.374
-0.122
0.496
SHAPE MN
-0.095
0.374
1.000
0.413
0.623
CLUMPY
-0.287
-0.122
0.413
1.000
0.705
PLADJ
0.267
0.496
0.623
0.705
1.000
AI
-0.166
0.026
0.465
0.967
0.786

AI
-0.166
0.026
0.465
0.967
0.786
1.000

Table A-22. 1_2k agriculture/forest class correlation matrix
1_2k Agriculture/Forest Class Spearman Rank Order Correlation Matrix
SHAPE
NP
LSI
CLUMPY
PLADJ
AI
MN
NP
1.000
0.955
0.322
-0.433
0.374
-0.395
LSI
0.955
1.000
0.535
-0.364
0.490
-0.323
SHAPE MN
0.322
0.535
1.000
0.160
0.728
0.184
CLUMPY
-0.433
-0.364
0.160
1.000
0.479
0.997
PLADJ
0.374
0.490
0.728
0.479
1.000
0.505
AI
-0.395
-0.323
0.184
0.997
0.505
1.000

109

Table A-23. 1_6k forest class correlation matrix

NP
LSI
SHAPE
MN
CLUMPY
PLADJ
AI

1_6k Forest Class Spearman Rank Order Correlation Matrix
SHAPE
NP
LSI
CLUMPY
PLADJ
MN
1.000
0.954
0.158
-0.196
0.311
0.954
1.000
0.364
-0.178
0.351
0.158
-0.196
0.311
-0.160

0.364
-0.178
0.351
-0.142

1.000
0.390
0.604
0.407

0.390
1.000
0.779
0.998

0.604
0.779
1.000
0.799

AI
-0.160
-0.142
0.407
0.998
0.799
1.000

Table A-24. 1_6k wetland class correlation matrix
1_6k Wetland Class Spearman Rank Order Correlation Matrix
SHAPE
NP
LSI
CLUMPY
PLADJ
MN
NP
1.000
0.901
0.137
-0.289
0.348
LSI
0.901
1.000
0.471
-0.177
0.507
SHAPE MN
0.137
0.471
1.000
0.346
0.666
CLUMPY
-0.289
-0.177
0.346
1.000
0.635
PLADJ
0.348
0.507
0.666
0.635
1.000
AI
-0.159
-0.023
0.395
0.963
0.717

AI
-0.159
-0.023
0.395
0.963
0.717
1.000

Table A-25. 1_6k agriculture/forest class correlation matrix
1_6k Agriculture/Forest Class Spearman Rank Order Correlation Matrix
SHAPE
NP
LSI
CLUMPY
PLADJ
MN
NP
1.000
0.956
0.415
-0.321
0.389
LSI
0.956
1.000
0.600
-0.238
0.492
SHAPE MN
0.415
0.600
1.000
0.298
0.799
CLUMPY
-0.321
-0.238
0.298
1.000
0.537
PLADJ
0.389
0.492
0.799
0.537
1.000
AI
-0.284
-0.198
0.325
0.997
0.557

110

AI
-0.284
-0.198
0.325
0.997
0.557
1.000

Table A-26. 2k forest class correlation matrix

NP
LSI
SHAPE MN
CLUMPY
PLADJ
AI

2k Forest Class Spearman Rank Order Correlation Matrix
SHAPE
NP
LSI
CLUMPY
PLADJ
MN
1.000
0.945
0.143
-0.226
0.449
0.945
1.000
0.342
-0.268
0.432
0.143
0.342
1.000
0.225
0.514
-0.226
-0.268
0.225
1.000
0.617
0.449
0.432
0.514
0.617
1.000
-0.186
-0.228
0.239
0.998
0.639

AI
-0.186
-0.228
0.239
0.998
0.639
1.000

Table A-27. 2k wetland class correlation matrix

NP
LSI
SHAPE MN
CLUMPY
PLADJ
AI

2k Wetland Class Spearman Rank Order Correlation Matrix
SHAPE
NP
LSI
CLUMPY
PLADJ
MN
1.000
0.923
0.170
-0.255
0.344
0.923
1.000
0.457
-0.185
0.452
0.170
0.457
1.000
0.306
0.643
-0.255
-0.185
0.306
1.000
0.655
0.344
0.452
0.643
0.655
1.000
-0.143
-0.053
0.352
0.968
0.718

AI
-0.143
-0.053
0.352
0.968
0.718
1.000

Table A-28. 2k agriculture/forest class correlation matrix
2k Agriculture/Forest Class Spearman Rank Order Correlation Matrix
SHAPE
NP
LSI
CLUMPY
PLADJ
MN
NP
1.000
0.969
0.425
-0.165
0.457
LSI
0.969
1.000
0.575
-0.114
0.515
SHAPE MN
0.425
0.575
1.000
0.405
0.765
CLUMPY
-0.165
-0.114
0.405
1.000
0.638
PLADJ
0.457
0.515
0.765
0.638
1.000
AI
-0.134
-0.079
0.429
0.997
0.653

111

AI
-0.134
-0.079
0.429
0.997
0.653
1.000

Appendix B
Non-Spatial Binomial Model Scatterplots

112

Figure B-1.

Landscape-level “best model” variable scatterplots

113

Figure B-2. 800m

“best model” variable scatterplot

114

Figure B-3. 800m wetland “best model” variable scatterplot

115

Figure B-4. 800m agriculture/forest class “best model” variable scatterplot

116

Figure B-5. 1_2k forest class “best model” variable scatterplot

117

Figure B-6. 1_2k wetland class “best model” variable scatterplot

118

Figure B-7. 1_2k agriculture/forest class “best model” variable scatterplot

119

Figure B-8. 1_6k forest class “best model” variable scatterplot

120

Figure B-9. 1_6k wetland class “best model” variable scatterplot

121

Figure B-10. 1_6k agriculture/forest class “best model” variable scatterplot

122

Figure B-11. 2k forest class “best model” variable scatterplot

123

Figure B-12. 2k wetland class “best model” variable scatterplot

124

Figure B-13. 2k agriculture/forest class “best model” variable scatterplot

125

Appendix C
Non-Spatial Binomial Models

126

Table C-1. Landscape-level non-spatial binomial GLMs
INTERCEPT
Buffer
Distance
400m
500m
600m
700m
800m
900m
1k
1_1k
1_2k
1_3k
1_4k
1_5k
1_6k
1_7k
1_8k
1_9k
2k

Est.
-10.178
-10.900
-9.931
-10.739
-12.630
-13.317
-12.374
-12.125
-10.845
-12.529
-11.218
-11.316
-10.841
-9.654
-11.802
-9.983
-11.231

Std.
Error
0.593
0.653
0.663
0.711
0.770
0.823
0.818
0.930
0.933
0.973
1.025
1.128
1.214
1.210
1.248
1.300
1.347

z-value
-17.163
-16.694
-14.971
-15.111
-16.409
-16.182
-15.120
-13.040
-11.628
-12.880
-10.949
-10.032
-8.927
-7.982
-9.454
-7.680
-8.339

Non-Spatial Landscape-level Models
WIAVGG

Pr(>|z|)
< 2e-16
< 2e-16
< 2e-16
< 2e-16
< 2e-16
< 2e-16
< 2e-16
< 2e-16
< 2e-16
< 2e-16
< 2e-16
< 2e-16
< 2e-16
0.000
< 2e-16
0.000
< 2e-16

Est.
0.080
0.097
0.108
0.121
0.127
0.131
0.131
0.141
0.151
0.150
0.161
0.163
0.168
0.176
0.171
0.182
0.181

Std.
Error
0.013
0.013
0.014
0.014
0.014
0.014
0.015
0.015
0.015
0.015
0.015
0.016
0.016
0.017
0.017
0.017
0.018

127

z-value
6.062
7.261
7.948
8.810
9.267
9.311
8.817
9.529
10.079
9.904
10.371
10.137
10.219
10.381
10.029
10.469
10.346

Pr(>|z|)
0.000
0.000
0.000
< 2e-16
< 2e-16
< 2e-16
< 2e-16
< 2e-16
< 2e-16
< 2e-16
< 2e-16
< 2e-16
< 2e-16
< 2e-16
< 2e-16
< 2e-16
< 2e-16

SHAPE_MN

Est.
1.327
1.904
1.083
1.603
3.061
3.597
2.797
2.522
1.422
2.724
1.616
1.704
1.303
0.324
2.014
0.538
1.520

Std.
Error
0.461
0.501
0.539
0.564
0.593
0.631
0.634
0.728
0.730
0.761
0.802
0.891
0.965
0.966
0.990
1.040
1.071

z-value
2.882
3.796
2.010
2.844
5.164
5.700
4.410
3.466
1.949
3.577
2.015
1.911
1.350
0.336
2.034
0.517
1.419

Pr(>|z|)
0.004
0.000
0.045
0.004
0.000
0.000
0.000
0.001
0.051
0.000
0.044
0.056
0.177
0.737
0.042
0.605
0.156

Non-Spatial Landscape-level Models
AIC
LSI
Buffer
Distance
400m
500m
600m
700m
800m
900m
1k
1_1k
1_2k
1_3k
1_4k
1_5k
1_6k
1_7k
1_8k
1_9k
2k

Est.
0.202
0.110
0.116
0.109
0.070
0.049
0.078
0.088
0.108
0.093
0.104
0.093
0.093
0.097
0.084
0.089
0.079

Std.
Error
0.085
0.073
0.065
0.057
0.051
0.046
0.043
0.041
0.038
0.037
0.034
0.033
0.032
0.030
0.029
0.028
0.027

z-value
2.371
1.501
1.796
1.900
1.394
1.049
1.828
2.138
2.812
2.535
3.026
2.822
2.928
3.209
2.879
3.179
2.941

Pr(>|z|)
0.018
0.133
0.073
0.057
0.163
0.294
0.068
0.033
0.005
0.011
0.002
0.005
0.003
0.001
0.004
0.001
0.003

1872.800
1856.100
1853.400
1833.800
1805.200
1798.100
1809.300
1811.000
1812.500
1796.300
1801.600
1802.800
1806.100
1808.700
1803.500
1806.700
1803.000

128

Table C-2. Landscape-level non-spatial binomial GLMs

INTERCEPT
Buffer
Distance
400m
500m
600m
700m
800m
900m
1k
1_1k
1_2k
1_3k
1_4k
1_5k
1_6k
1_7k
1_8k
1_9k
2k

Est.
-10.564
-11.158
-10.284
-11.232
-13.003
-13.612
-12.879
-12.735
-11.555
-13.263
-12.048
-12.166
-11.820
-10.790
-12.940
-11.270
-12.440

Std.
Error
0.612
0.673
0.667
0.728
0.801
0.853
0.855
0.952
0.954
0.999
1.050
1.138
1.211
1.204
1.256
1.282
1.342

z-value
-17.271
-16.574
-15.424
-15.427
-16.234
-15.957
-15.070
-13.384
-12.117
-13.282
-11.478
-10.688
-9.759
-8.963
-10.307
-8.791
-9.273

Non-Spatial Landscape-level Models
WIAVGG

Pr(>|z|)
< 2e-16
< 2e-16
< 2e-16
< 2e-16
< 2e-16
< 2e-16
< 2e-16
< 2e-16
< 2e-16
< 2e-16
< 2e-16
< 2e-16
<2e-16
< 2e-16
< 2e-16
< 2e-16
< 2e-16

Est.
0.080
0.096
0.108
0.119
0.126
0.130
0.129
0.140
0.150
0.149
0.159
0.161
0.166
0.174
0.169
0.180
0.179

Std.
Error
0.013
0.013
0.014
0.014
0.014
0.014
0.015
0.015
0.015
0.015
0.016
0.016
0.016
0.017
0.017
0.017
0.018

129

z-value
5.996
7.211
7.904
8.641
9.146
9.236
8.678
9.430
9.997
9.811
10.282
9.994
10.077
10.226
9.899
10.378
10.232

Pr(>|z|)
0.000
0.000
0.000
8.641
9.146
9.236
8.678
9.430
9.997
9.811
10.282
9.994
10.077
10.226
9.899
10.378
10.232

SHAPE_MN

Est.
1.758
2.194
1.434
2.032
3.396
3.855
3.244
3.093
2.144
3.409
2.433
2.535
2.247
1.392
3.035
1.716
2.617

Std.
Error
0.441
0.484
0.498
0.535
0.583
0.622
0.631
0.708
0.710
0.746
0.787
0.863
0.923
0.920
0.954
0.980
1.023

z-value
3.989
4.531
2.879
3.796
5.824
6.197
5.143
4.371
3.018
4.570
3.090
2.936
2.433
1.513
3.182
1.751
2.557

Pr(>|z|)
0.000
0.000
0.004
0.000
0.000
0.000
0.000
0.000
0.003
0.000
0.002
0.003
0.015
0.130
0.001
0.080
0.011

Non-Spatial Landscape-level Models
NP
Buffer
Distance
400m
500m
600m
700m
800m
900m
1k
1_1k
1_2k
1_3k
1_4k
1_5k
1_6k
1_7k
1_8k
1_9k
2k

Est.
0.023
0.010
0.010
0.010
0.005
0.003
0.005
0.004
0.004
0.004
0.003
0.003
0.003
0.003
0.002
0.002
0.002

Std.
Error
0.008
0.006
0.005
0.004
0.003
0.003
0.002
0.002
0.002
0.001
0.001
0.001
0.001
0.001
0.001
0.001
0.001

z-value
2.959
1.619
2.250
2.693
1.733
1.300
2.234
2.260
2.464
2.665
2.777
2.548
2.605
3.001
3.039
2.992
2.861

Pr(>|z|)
0.003
0.105
0.024
0.007
0.083
0.194
0.026
0.024
0.014
0.008
0.005
0.011
0.009
0.003
0.002
0.003
0.004

AIC

1869.800
1855.700
1851.700
1830.200
1804.100
1797.500
1807.700
1810.500
1814.500
1795.700
1803.100
1804.400
1808.000
1810.100
1802.700
1807.900
1803.500

130

Table C-3. Landscape-level non-spatial binomial GLMs

INTERCEPT
Buffer
Distance
Est.
400m
-7.749
500m
-9.313
600m
-7.821
700m
-8.578
800m
-11.026
900m
-12.082
1k
-10.212
1_1k
-9.434
1_2k
-7.307
1_3k
-9.222
1_4k
-7.270
1_5k
-7.540
1_6k
-6.904
1_7k
-5.251
1_8k
-7.762
1_9k
-5.450
2k
-7.038

Std.
Error
1.143
1.237
1.363
1.394
1.382
1.437
1.439
1.619
1.627
1.683
1.716
1.844
1.963
1.993
2.035
2.134
2.146

Non-Spatial Landscape-level Models
WIAVGG

z-value Pr(>|z|)
-6.780
0.000
-7.529
0.000
-5.738
0.000
-6.152
0.000
-7.978
0.000
-8.406 < 2e-16
-7.095
0.000
-5.828
0.000
-4.492
0.000
-5.479
0.000
-4.236
0.000
-4.089
0.000
-3.517
0.000
-2.635
0.008
-3.813
0.000
-2.554
0.011
-3.279
0.001

Est.
0.080
0.097
0.109
0.121
0.127
0.131
0.131
0.141
0.151
0.150
0.161
0.163
0.168
0.176
0.171
0.182
0.181

Std.
Error
0.013
0.013
0.014
0.014
0.014
0.014
0.015
0.015
0.015
0.015
0.015
0.016
0.016
0.017
0.017
0.017
0.018

131

z-value
6.061
7.272
7.959
8.818
9.269
9.314
8.822
9.536
10.087
9.909
10.374
10.134
10.214
10.376
10.026
10.464
10.340

Pr(>|z|)
0.000
0.000
0.000
< 2e-16
< 2e-16
< 2e-16
< 2e-16
< 2e-16
< 2e-16
< 2e-16
< 2e-16
< 2e-16
< 2e-16
< 2e-16
< 2e-16
< 2e-16
< 2e-16

SHAPE_MN
Std.
Est. Error z-value
1.313 0.461 2.847
1.892 0.503 3.759
1.052 0.541 1.944
1.597 0.564 2.830
3.058 0.593 5.158
3.596 0.631 5.696
2.800 0.634 4.416
2.521 0.728 3.463
1.427 0.730 1.955
2.731 0.761 3.588
1.632 0.801 2.038
1.728 0.889 1.944
1.346 0.962 1.400
0.371 0.962 0.386
2.051 0.988 2.076
0.589 1.036 0.569
1.567 1.067 1.468

Pr(>|z|)
0.004
0.000
0.052
0.005
0.000
0.000
0.000
0.001
0.051
0.000
0.042
0.052
0.161
0.699
0.038
0.570
0.142

Non-Spatial Landscape-level Models
PLADJ
Buffer
Distance
400m
500m
600m
700m
800m
900m
1k
1_1k
1_2k
1_3k
1_4k
1_5k
1_6k
1_7k
1_8k
1_9k
2k

Est.
-0.023
-0.015
-0.020
-0.021
-0.016
-0.012
-0.021
-0.026
-0.035
-0.033
-0.039
-0.038
-0.039
-0.044
-0.040
-0.045
-0.042

Std.
Error
0.009
0.010
0.011
0.011
0.011
0.011
0.012
0.012
0.013
0.013
0.013
0.014
0.014
0.014
0.014
0.015
0.015

z-value
-2.484
-1.510
-1.899
-1.913
-1.412
-1.057
-1.806
-2.134
-2.757
-2.483
-2.944
-2.766
-2.830
-3.119
-2.787
-3.102
-2.852

Pr(>|z|)
0.013
0.131
0.058
0.056
0.158
0.290
0.071
0.033
0.006
0.013
0.003
0.006
0.005
0.002
0.005
0.002
0.004

AIC

1872.300
1856.000
1853.000
1833.800
1805.100
1798.000
1809.400
1811.000
1812.800
1796.600
1802.000
1803.100
1806.700
1809.300
1804.100
1807.200
1803.500

132

Table C-4. Landscape-level non-spatial binomial GLMs
INTERCEPT
Buffer
Distance
400m
500m
600m
700m
800m
900m
1k
1_1k
1_2k
1_3k
1_4k
1_5k
1_6k
1_7k
1_8k
1_9k
2k

Est.
-10.124
-10.818
-9.764
-10.619
-12.503
-13.228
-12.197
-11.894
-10.612
-12.381
-10.970
-11.042
-10.634
-9.443
-11.497
-9.576
-10.797

Std.
Error
0.586
0.653
0.676
0.716
0.783
0.844
0.833
0.954
0.953
0.986
1.033
1.137
1.232
1.227
1.267
1.327
1.370

z-value
-17.269
-16.575
-14.441
-14.822
-15.974
-15.670
-14.643
-12.461
-11.132
-12.562
-10.618
-9.707
-8.630
-7.697
-9.073
-7.219
-7.879

Non-Spatial Landscape-level Models
WIAVGG

Pr(>|z|)
< 2e-16
< 2e-16
< 2e-16
< 2e-16
< 2e-16
< 2e-16
< 2e-16
< 2e-16
< 2e-16
< 2e-16
<2e-16
< 2e-16
< 2e-16
0.000
< 2e-16
0.000
0.000

Est.
0.081
0.094
0.103
0.114
0.121
0.124
0.123
0.131
0.140
0.141
0.150
0.153
0.156
0.161
0.157
0.166
0.166

Std.
Error
0.014
0.014
0.014
0.014
0.014
0.015
0.015
0.015
0.015
0.016
0.016
0.017
0.017
0.017
0.018
0.018
0.018

133

z-value
5.963
6.897
7.479
8.186
8.621
8.551
8.076
8.560
9.030
9.016
9.387
9.234
9.202
9.230
8.979
9.238
9.143

Pr(>|z|)
0.000
0.000
0.000
0.000
< 2e-16
< 2e-16
0.000
< 2e-16
< 2e-16
< 2e-16
<2e-16
< 2e-16
< 2e-16
< 2e-16
< 2e-16
< 2e-16
< 2e-16

SHAPE_MN

Est.
1.570
1.936
1.002
1.542
2.953
3.457
2.627
2.316
1.267
2.654
1.494
1.541
1.192
0.203
1.815
0.252
1.206

Std.
Error
0.449
0.495
0.535
0.555
0.600
0.643
0.646
0.742
0.743
0.768
0.809
0.897
0.972
0.970
1.001
1.054
1.086

z-value
3.499
3.914
1.874
2.778
4.925
5.377
4.067
3.120
1.705
3.455
1.848
1.719
1.226
0.209
1.814
0.239
1.110

Pr(>|z|)
0.000
0.000
0.061
0.005
0.000
0.000
0.000
0.002
0.088
0.001
0.065
0.086
0.220
0.834
0.070
0.811
0.267

Non-Spatial Landscape-level Models
SHDI
Buffer
Distance
400m
500m
600m
700m
800m
900m
1k
1_1k
1_2k
1_3k
1_4k
1_5k
1_6k
1_7k
1_8k
1_9k
2k

Est.
0.105
0.197
0.333
0.368
0.309
0.329
0.421
0.488
0.538
0.456
0.514
0.502
0.533
0.605
0.548
0.632
0.589

Std.
Error
0.135
0.136
0.140
0.142
0.143
0.145
0.147
0.150
0.150
0.152
0.152
0.155
0.157
0.159
0.164
0.167
0.169

z-value
0.776
1.450
2.381
2.594
2.161
2.271
2.858
3.249
3.590
3.003
3.389
3.250
3.399
3.801
3.343
3.779
3.496

Pr(>|z|)
0.437
0.147
0.017
0.009
0.031
0.023
0.004
0.001
0.000
0.003
0.001
0.001
0.001
0.000
0.001
0.000
0.000

AIC

1877.900
1856.200
1850.900
1830.600
1802.400
1793.900
1804.300
1804.800
1807.300
1793.600
1799.100
1800.100
1802.900
1804.200
1800.500
1802.200
1799.200

134

Table C-5. Landscape-level non-spatial binomial GLMs

INTERCEPT
Buffer
Distance
400m
500m
600m
700m
800m
900m
1k
1_1k
1_2k
1_3k
1_4k
1_5k
1_6k
1_7k
1_8k
1_9k
2k

Est.
-9.975
-10.624
-10.024
-10.919
-12.617
-13.495
-13.211
-13.431
-12.599
-13.995
-12.449
-13.048
-12.760
-11.471
-13.687
-11.825
-12.979

Std.
Error
0.595
0.650
0.663
0.734
0.801
0.890
0.865
0.985
0.988
1.009
1.051
1.162
1.238
1.224
1.272
1.287
1.349

z-value
-16.762
-16.353
-15.109
-14.885
-15.748
-15.161
-15.280
-13.629
-12.748
-13.871
-11.848
-11.227
-10.307
-9.371
-10.759
-9.188
-9.619

Non-Spatial Landscape-level Models
WIAVGG

Pr(>|z|)
< 2e-16
< 2e-16
< 2e-16
< 2e-16
< 2e-16
< 2e-16
< 2e-16
< 2e-16
< 2e-16
< 2e-16
< 2e-16
< 2e-16
< 2e-16
< 2e-16
< 2e-16
< 2e-16
< 2e-16

Est.
0.084
0.100
0.108
0.120
0.128
0.129
0.117
0.119
0.120
0.116
0.134
0.122
0.121
0.129
0.118
0.137
0.136

Std.
Error
0.013
0.013
0.014
0.014
0.014
0.015
0.016
0.016
0.017
0.017
0.017
0.019
0.019
0.020
0.020
0.020
0.021

135

z-value Pr(>|z|)
6.358
0.000
7.517
0.000
7.859
0.000
8.579 < 2e-16
8.958 < 2e-16
8.731 < 2e-16
7.355
0.000
7.388
0.000
7.275
0.000
6.756
0.000
7.727
0.000
6.561
0.000
6.415
0.000
6.542
0.000
5.911
0.000
6.858
0.000
6.615
0.000

SHAPE_MN

Est.
1.663
2.091
1.453
1.982
3.252
3.812
3.326
3.317
2.486
3.584
2.457
2.720
2.343
1.379
2.984
1.663
2.606

Std.
Error
0.421
0.462
0.480
0.516
0.569
0.624
0.625
0.717
0.721
0.743
0.779
0.869
0.933
0.926
0.958
0.977
1.020

z-value
3.953
4.522
3.026
3.837
5.715
6.105
5.323
4.624
3.450
4.824
3.157
3.130
2.510
1.489
3.113
1.703
2.555

Pr(>|z|)
0.000
0.000
0.002
0.000
0.000
0.000
0.000
0.000
0.001
0.000
0.002
0.002
0.012
0.137
0.002
0.089
0.011

Non-Spatial Landscape-level Models
PRD
Buffer
Distance
400m
500m
600m
700m
800m
900m
1k
1_1k
1_2k
1_3k
1_4k
1_5k
1_6k
1_7k
1_8k
1_9k
2k

Est.
-0.032
-0.062
-0.019
0.013
-0.012
0.059
0.424
0.706
1.109
1.244
1.181
1.853
2.464
2.608
3.252
3.019
3.082

Std.
Error
0.026
0.039
0.056
0.080
0.099
0.128
0.151
0.190
0.225
0.272
0.322
0.389
0.456
0.525
0.595
0.644
0.715

z-value
-1.247
-1.587
-0.337
0.160
-0.120
0.458
2.801
3.714
4.926
4.573
3.669
4.763
5.409
4.966
5.471
4.688
4.313

Pr(>|z|)
0.212
0.113
0.736
0.873
0.905
0.647
0.005
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000

AIC

1876.900
1855.800
1856.600
1837.400
1807.100
1799.000
1804.700
1801.500
1795.800
1781.500
1797.100
1787.800
1784.900
1793.900
1781.700
1794.900
1793.200

136

Table C-6. 800m forest class non-spatial binomial GLMs

Est.
-11.736

INTERCEPT
Std.
Error
z-value
0.733 -16.010

Est.
-10.078

INTERCEPT
Std.
Error
z-value
0.489 -20.612

Est.
-11.644

INTERCEPT
Std.
Error
z-value
0.721 -16.157

Est.
10.803

800m Forest Class
High Leverage Points: 1,3,5,6,31,34,64,76
SHAPE_MN
Std.
Pr(>|z|)
Est.
Error z-value
Pr(>|z|)
Est.
< 2e-16 2.157
0.384
5.626
0.000
2.302

Pr(>|z|)
Est.
< 2e-16 2.687

SHAPE_MN
Std.
Error z-value
0.355
7.565

Pr(>|z|)
Est.
< 2e-16 2.137

SHAPE_MN
Std.
Error z-value
0.388
5.501

INTERCEPT
Std.
Error
z-value

Pr(>|z|)

SHAPE_MN
Std.
Error z-value

0.562

< 2e-16 1.533

-19.228

Est.

0.470

3.261

137

CLUMPY
Std.
Error
z-value Pr(>|z|)
0.806
2.857
0.004

AIC
421.46

LSI
Pr(>|z|)
0.000

Est.
-0.199

Std.
Error
0.123

Est.
0.022

Std.
Error
0.008

z-value Pr(>|z|)
-1.613
0.107

AIC
427.66

z-value Pr(>|z|)
2.746
0.006

AIC
422.14

PLADJ
Std.
Error
z-value Pr(>|z|)

AIC

AI
Pr(>|z|)
0.000

Pr(>|z|)

Est.

0.001

0.025

0.007

3.831

0.000

413.67

Table C-7. 800m wetland class non-spatial binomial GLMs

Est.
-9.524

800m Wetland Class
High Leverage Points: 9,36,47,52,71,112,114,123,128,152
INTERCEPT
SHAPE_MN
CLUMPY
Std.
Std.
Std.
Error z-value Pr(>|z|)
Est.
Error
z-value Pr(>|z|)
Est.
Error
z-value Pr(>|z|)
0.551 -17.302 < 2e-16 0.885
0.185
4.800
0.000 1.204
0.671
1.793
0.073

AIC
1049.1

Est.
-8.382

INTERCEPT
Std.
Error z-value Pr(>|z|)
Est.
0.258 -32.467 < 2e-16 1.158

SHAPE_MN
Std.
Error
z-value Pr(>|z|)
0.177
6.557
0.000

z-value Pr(>|z|)
-2.997
0.003

AIC
1044.5

Est.
-9.211

INTERCEPT
Std.
Error z-value Pr(>|z|)
Est.
0.533 -17.279 < 2e-16 0.908

SHAPE_MN
Std.
Error
z-value Pr(>|z|)
0.187
4.843
0.000

Est.
-8.866

INTERCEPT
Std.
Error z-value Pr(>|z|)
Est.
0.336 -26.359 < 2e-16 0.894

SHAPE_MN
Std.
Error
z-value Pr(>|z|)
0.202
4.435
0.000

138

LSI
Est.
-0.213

Std.
Error
0.071
AI

Est.
0.008

Std.
Error
0.007

z-value Pr(>|z|)
1.171
0.241

AIC
1052.3

Est.
0.004

PLADJ
Std.
Error
z-value Pr(>|z|)
0.005
0.914
0.361

AIC
1053

Table C-8. 800m agriculture/forest class non-spatial binomial GLMs

Est.
-8.889

800m Agriculture/Forest Class
High Leverage Points: 19,59,62
INTERCEPT
SHAPE_MN
Std.
Std.
Error z-value Pr(>|z|)
Est.
Error z-value Pr(>|z|)
Est.
0.490 -18.133 < 2e-16 1.456
0.378
3.856
0.000 -0.127

CLUMPY
Std.
Error z-value Pr(>|z|)
0.381
-0.333
0.739

AIC
460.12

Est.
-7.750

INTERCEPT
Std.
Error z-value Pr(>|z|)
Est.
0.453 -17.096 < 2e-16 1.826

SHAPE_MN
Std.
Error z-value Pr(>|z|)
0.320
5.703
0.000

Est.
-0.760

LSI
Std.
Error z-value Pr(>|z|)
0.133
-5.717
0.000

AIC
423.04

Est.
-8.713

INTERCEPT
Std.
Error z-value Pr(>|z|)
Est.
0.519 -16.790 < 2e-16 1.568

SHAPE_MN
Std.
Error z-value Pr(>|z|)
0.393
3.992
0.000

Est.
-0.005

AI
Std.
Error z-value Pr(>|z|)
0.006
-0.892
0.372

AIC
459.5

Est.
-8.740

INTERCEPT
Std.
Error z-value Pr(>|z|)
Est.
0.421 -20.752 < 2e-16 2.143

SHAPE_MN
Std.
Error z-value Pr(>|z|)
0.397
5.403
0.000

Est.
-0.016

PLADJ
Std.
Error z-value Pr(>|z|)
0.005
-3.164
0.002

AIC
451.87

139

Table C-9. 1_2k forest class non-spatial binomial GLMs

Est.
-9.734

INTERCEPT
Std.
Error z-value
0.720 -13.527

Est.
-10.022

INTERCEPT
Std.
Error z-value
0.595 -16.836

Est.
-9.586

INTERCEPT
Std.
Error z-value
0.706 -13.569

Est.
-9.413

INTERCEPT
Std.
Error z-value
0.668 -14.082

Pr(>|z|)
< 2e-16

Pr(>|z|)
< 2e-16

Pr(>|z|)
< 2e-16

Pr(>|z|)
< 2e-16

1_2k Forest Class
High Leverage Points: 6,26,45,102
SHAPE_MN
Std.
Est.
Error z-value Pr(>|z|)
Est.
0.787 0.544
1.447
0.148
2.139
High Leverage Points: 26,45,102
SHAPE_MN
Std.
Est.
Error z-value Pr(>|z|)
Est.
3.258 0.506
6.442
0.000
-0.452
High Leverage Points: 6,26,45,102
SHAPE_MN
Std.
Est.
Error z-value Pr(>|z|)
Est.
0.798 0.551
1.448
0.148
0.019
High Leverage Points: 6,26,45,85,102
SHAPE_MN
Std.
Est.
Error z-value Pr(>|z|)
Est.
2.126 0.655
3.248
0.001
-0.004

140

CLUMPY
Std.
Error z-value Pr(>|z|)
0.662
3.230
0.001

AIC
796.43

LSI
Std.
Error z-value Pr(>|z|)
0.070
-6.480
0.000

AIC
773.27

AI
Std.
Error
0.007

z-value Pr(>|z|)
2.937
0.003

AIC
798.3

PLADJ
Std.
Error z-value Pr(>|z|)
0.005
-0.851
0.395

AIC
777.71

Table C-10. 1_2k wetland class non-spatial binomial GLMs

Est.
-8.485

INTERCEPT
Std.
Error z-value Pr(>|z|)
0.429 -19.772 < 2e-16

Est.
-7.386

INTERCEPT
Std.
Error z-value Pr(>|z|)
0.236 -31.359 < 2e-16

Est.
-7.532

INTERCEPT
Std.
Error z-value Pr(>|z|)
0.298 -25.245 < 2e-16

Est.
0.498

Est.
0.852

Est.
0.761

1_2k Wetland Class
SHAPE_MN
Std.
Error z-value Pr(>|z|)
0.180
2.775
0.006
SHAPE_MN
Std.
Error z-value Pr(>|z|)
0.159
5.351
0.000
SHAPE_MN
Std.
Error z-value Pr(>|z|)
0.181
4.204
0.000

141

Est.
0.642

CLUMPY
Std.
Error z-value Pr(>|z|)
0.514
1.251
0.211

AIC
1240.8

Est.
-0.311

LSI
Std.
Error z-value Pr(>|z|)
0.044
-7.040
0.000

AIC
1190.4

Est.
-0.011

PLADJ
Std.
Error z-value Pr(>|z|)
0.004
-2.692
0.007

AIC
1236.1

Table C-11. 1_2k agriculture/forest class non-spatial binomial GLMs

INTERCEPT
Std.
Error
z-value
0.757 -15.450

Pr(>|z|)
< 2e-16

INTERCEPT
Std.
Error
z-value
0.555 -16.576

Pr(>|z|)
< 2e-16

INTERCEPT
Std.
Est. Error
z-value
-11.561 0.735 -15.720

Pr(>|z|)
< 2e-16

INTERCEPT
Std.
Error
z-value
0.593 -15.425

Pr(>|z|)
<2e-16

Est.
-11.687

Est.
-9.196

Est.
-9.146

1_2k Agriculture/Forest Class
High Leverage Points: 21,27,75,83
SHAPE_MN
CLUMPY
Std.
Std.
Est. Error z-value Pr(>|z|)
Est. Error
z-value Pr(>|z|)
1.305 0.458
2.847
0.004
3.929 0.638
6.163
0.000
High Leverage Points: 21,73,83
SHAPE_MN
LSI
Std.
Std.
Est. Error z-value Pr(>|z|)
Est. Error
z-value Pr(>|z|)
2.651 0.457
5.806
0.000
-0.374 0.055
-6.848
0.000
High Leverage Points: 21,27,75,83
SHAPE_MN
AI
Std.
Std.
Est. Error z-value Pr(>|z|)
Est. Error z-value Pr(>|z|)
1.208 0.460
2.626
0.009
0.039 0.006
6.186
0.000
High Leverage Points: 21,27,40,75,83
SHAPE_MN
PLADJ
Std.
Std.
Est. Error z-value Pr(>|z|)
Est. Error
z-value Pr(>|z|)
1.564 0.573
2.732
0.006
0.006 0.004
1.445
0.149

142

AIC
672.61

AIC
691.12

AIC
675.41

AIC
688.36

Table C-12. 1_6k forest class non-spatial binomial GLMs

Est.
-9.228

INTERCEPT
Std.
Error z-value Pr(>|z|)
0.580 -15.919 < 2e-16

1_6k Forest Class
High Leverage Points: 29,36,50,97,101
SHAPE_MN
Std.
Est.
Error z-value Pr(>|z|)
Est.
1.335
0.499
2.677
0.007
0.428

CLUMPY
Std.
Error z-value Pr(>|z|)
0.369
1.159
0.247

AIC
949.23

Est.
-9.161

INTERCEPT
Std.
Error z-value Pr(>|z|)
0.557 -16.436 < 2e-16

SHAPE_MN
Std.
Error z-value Pr(>|z|)
0.443
3.840
0.000

LSI
Std.
Error z-value Pr(>|z|)
0.046
-1.478
0.140

AIC
948.4

Est.
-9.220

INTERCEPT
Std.
Error z-value Pr(>|z|)
0.579 -15.930 < 2e-16

Est.
0.004

AI
Std.
Error z-value Pr(>|z|)
0.004
1.081
0.280

AIC
949.42

Est.
0.003

PLADJ
Std.
Error z-value Pr(>|z|)
0.003
0.990
0.322

AIC
949.62

Est.
-9.142

Est.
-9.142

INTERCEPT
Std.
Error z-value Pr(>|z|)
0.583 -15.684 < 2e-16
INTERCEPT
Std.
Error z-value Pr(>|z|)
0.583 -15.684 < 2e-16

Est.
1.702

Est.
1.348

SHAPE_MN
Std.
Error z-value Pr(>|z|)
0.500
2.695
0.007

Est.
1.351

SHAPE_MN
Std.
Error z-value Pr(>|z|)
0.513
2.633
0.008

Est.
1.351

SHAPE_MN
Std.
Error z-value Pr(>|z|)
0.513
2.633
0.008

143

Est.
-0.068

AIC
949.62

Table C-13. 1_6k wetland class non-spatial binomial GLMs

Est.
-7.784

INTERCEPT
Std.
Error z-value Pr(>|z|)
0.331 -23.524 < 2e-16

Est.
-7.697

INTERCEPT
Std.
Error z-value Pr(>|z|)
0.208 -36.957 < 2e-16

Est.
-7.421

INTERCEPT
Std.
Error z-value Pr(>|z|)
0.392 -18.955 < 2e-16

Est.
-7.004

INTERCEPT
Std.
zError value
Pr(>|z|)
0.266 -26.288 < 2e-16

Est.
0.463

1_6k Wetland Class
SHAPE_MN
Std.
Error z-value Pr(>|z|)
0.160
2.889
0.004

Est.
0.681

SHAPE_MN
Std.
Error z-value
0.151
4.500

Est.
0.505

SHAPE_MN
Std.
Error z-value
0.161
3.142

Est.
0.738

SHAPE_MN
Std.
Error z-value
0.157
4.689

144

Pr(>|z|)
0.000

Pr(>|z|)
0.002

Pr(>|z|)
0.000

Est.
-0.218

CLUMPY
Std.
Error z-value
0.351 -0.620

Pr(>|z|)
AIC
0.535 1412.8

Est.
-0.140

LSI
Std.
Error z-value
0.029 -4.852

Pr(>|z|)
AIC
0.000 1388.9

Est.
-0.007

AI
Std.
Error z-value Pr(>|z|)
AIC
0.005
-1.556
0.120 1410.9

Est.
-0.018

PLADJ
Std.
Error z-value
0.004 -5.008

Pr(>|z|)
AIC
0.000 1392.3

Table C-14. 1_6k agriculture/forest class non-spatial binomial GLMs

Est.
-8.669

INTERCEPT
Std.
Error z-value
0.690 -12.563

Est.
-9.128

INTERCEPT
Std.
Error z-value Pr(>|z|)
0.581 -15.716 < 2e-16

Est.
-8.601

INTERCEPT
Std.
Error z-value
0.680 -12.640

Est.
-8.769

INTERCEPT
Std.
Error z-value
0.690 -12.711

Pr(>|z|)
<2e-16

Pr(>|z|)
<2e-16

Pr(>|z|)
<2e-16

1_6k Agriculture/Forest Class
High Leverage Points: 24,32,68,85,94,101
SHAPE_MN
Std.
Est.
Error z-value Pr(>|z|)
Est.
0.854
0.532
1.605
0.108
0.878
High Leverage Points: 24,32,85,94,101
SHAPE_MN
Std.
Est.
Error z-value Pr(>|z|)
Est.
2.681
0.511
5.244
0.000 -0.328
High Leverage Points: 24,32,68,85,94,101
SHAPE_MN
Std.
Est.
Error z-value Pr(>|z|)
Est.
0.842
0.537
1.567
0.117
0.008
High Leverage Points: 24,32,47,68,85,94,98,101
SHAPE_MN
Std.
Est.
Error z-value Pr(>|z|)
Est.
1.590
0.752
2.114
0.035 -0.001

145

CLUMPY
Std.
Error z-value Pr(>|z|)
0.495
1.775
0.076

AIC
821.61

LSI
Std.
Error z-value Pr(>|z|)
0.044
-7.478
0.000

AIC
787.29

AI
Std.
Error z-value Pr(>|z|)
0.005
1.642
0.101

AIC
822.1

PLADJ
Std.
Error z-value Pr(>|z|)
0.006
-0.170
0.865

AIC
766.23

Table C-15. 2k forest class non-spatial binomial GLMs

Est.
-9.379

INTERCEPT
Std.
Error z-value Pr(>|z|)
0.714 -13.140 < 2e-16

Est.
-8.017

INTERCEPT
Std.
Error z-value
0.530 -15.124

Est.
-9.300

INTERCEPT
Std.
Error z-value Pr(>|z|)
0.704 -13.210 < 2e-16

Est.
-8.464

INTERCEPT
Std.
Error z-value Pr(>|z|)
0.597 -14.187 < 2e-16

Pr(>|z|)
<2e-16

2k Forest Class
High Leverage Points: 1,2,33,57,73,109
SHAPE_MN
Std.
Est.
Error z-value Pr(>|z|)
Est.
0.022
0.456
0.049
0.961
2.508

CLUMPY
Std.
Error z-value Pr(>|z|)
0.666
3.766
0.000

AIC
1062.4

Est.
0.672

SHAPE_MN
Std.
Error z-value Pr(>|z|)
0.428
1.572
0.116

Est.
-0.015

LSI
Std.
Error z-value Pr(>|z|)
0.037
-0.414
0.679

AIC
1080.6

Est.
0.007

SHAPE_MN
Std.
Error z-value Pr(>|z|)
0.459
0.015
0.988

Est.
0.024

AI
Std.
Error z-value Pr(>|z|)
0.007
3.691
0.000

AIC
1063.2

Est.
-0.407

SHAPE_MN
Std.
Error z-value Pr(>|z|)
0.504
-0.808
0.419

146

Est.
0.023

PLADJ
Std.
Error
0.005

z-value Pr(>|z|)
4.686
0.000

AIC
1054.7

Table C-16. 2k wetland class non-spatial binomial GLMs

Est.
-5.895

INTERCEPT
Std.
Error z-value Pr(>|z|)
0.476 -12.394 <2e-16

2k Wetland Class
High Leverage Points: 2,3,39,54,75,96,203
SHAPE_MN
Std.
Est.
Error z-value Pr(>|z|)
Est.
-0.693
0.270
-2.565
0.010 -0.674

CLUMPY
Std.
Error z-value Pr(>|z|)
0.493
-1.367
0.172

AIC
1552.5

Est.
-6.406

INTERCEPT
Std.
Error z-value Pr(>|z|)
0.340 -18.821 <2e-16

SHAPE_MN
Std.
Error z-value Pr(>|z|)
0.274
-2.126
0.034

LSI
Std.
Error z-value Pr(>|z|)
0.023
-1.734
0.083

AIC
1551.2

Est.
-5.838

INTERCEPT
Std.
Error z-value Pr(>|z|)
0.459 -12.731 <2e-16

Est.
-5.837

Est.
-0.583

Est.
-0.039

SHAPE_MN
Std.
Est.
Error z-value Pr(>|z|)
Est.
-0.662
0.271
-2.442
0.015 -0.008
High Leverage Points: 2,3,39,54,75,96,177,203
INTERCEPT
SHAPE_MN
Std.
Std.
Error z-value Pr(>|z|)
Est.
Error z-value Pr(>|z|)
Est.
0.376 -15.512 <2e-16 -0.635
0.302
-2.102
0.036 -0.009

147

AI
Std.
Error
0.005

z-value Pr(>|z|)
-1.640
0.101

AIC
1551.7

PLADJ
Std.
Error
z-value Pr(>|z|)
0.004
-2.000
0.046

AIC
1537.1

Table C-17. 2k agriculture/forest class non-spatial binomial GLMs

Est.
-9.431

Est.
-9.382

Est.
-9.350

Est.
-7.833

INTERCEPT
Std.
Error z-value Pr(>|z|)
0.600 -15.570 < 2e-16

2k Agriculture/Forest Class
High Leverage Points: 25,34,52,93,104
SHAPE_MN
Std.
Est.
Error z-value Pr(>|z|)
Est.
0.607
0.466
1.304
0.192
2.247
High Leverage Points: 25,34,93,104
SHAPE_MN
Std.
Est.
Error z-value Pr(>|z|)
Est.
2.633
0.430
6.128
0.000 -0.227
High Leverage Points: 25,34,52,93,104
SHAPE_MN
Std.
Est.
Error z-value Pr(>|z|)
Est.
0.550
0.469
1.175
0.200
0.022

INTERCEPT
Std.
Error z-value Pr(>|z|)
0.614 -12.750 < 2e-16

SHAPE_MN
Std.
Error z-value Pr(>|z|)
0.624
-1.182
0.237

INTERCEPT
Std.
Error z-value Pr(>|z|)
0.608 -15.520 < 2e-16
INTERCEPT
Std.
Error z-value Pr(>|z|)
0.496 -18.911 < 2e-16

Est.
-0.738

148

Est.
0.026

CLUMPY
Std.
Error z-value Pr(>|z|)
0.466
4.821
0.000

AIC
900.78

LSI
Std.
Error z-value Pr(>|z|)
0.033
-6.858
0.000

AIC
918.08

AI
Std.
Error z-value Pr(>|z|)
0.005
4.765
0.000

AIC
901.54

PLADJ
Std.
Error z-value Pr(>|z|)
0.005
5.244
0.000

AIC
897.88

Appendix D
Spatial Binomial Models

149

Table D-1. 800m forest class spatial binomial GLM
Summary 800m Forest 3 Chains at 100,000
Parameters
50.0%
2.5%
97.5%
Intercept
-8.621 11.906
-8.192
Shape Index Mean
0.703 -1.772
1.903
Percent Land Cover
Adjacency
0.034
0.009
0.045
sigma.sq
2.628
1.853
4.305
Phi
350.850 26.887 387.372
effective range 3/phi
0.009
0.008
0.185
max intersite distance = 1.174

Table D-2. 800m wetland class spatial binomial GLM
Summary 800m Wetland 3 Chains at 50,000
Parameters
50.0%
2.5%
Intercept
10.022 10.095
Shape Index Mean
0.986
0.722
sigma.sq
3.503
3.158
Phi
15.224 12.006
effective range 3/phi
0.197
0.141
max intersite distance = 1.260

97.5%
-8.918
1.096
3.869
21.354
0.251

Table D-3. 800m agriculture/forest class spatial binomial GLM
Summary 800m Agriculture/Forest 3 Chains at 50,000
Parameters
50.0%
2.5%
97.5%
Intercept
-8.868
-9.844
-8.245
Shape Index Mean
2.319
1.774
2.500
Landscape Shape Index
-0.549
-0.776
-0.192
sigma.sq
1.345
1.142
2.309
Phi
283.921 234.639 551.114
effective range 3/phi
0.011
0.006
0.013

150

Table D-4. 1_2k forest class spatial binomial GLM
Summary 1_2k Forest 3 Chains at 100,000
Parameters
50.0%
2.5%
97.5%
Intercept
11.411 15.723
-8.792
Shape Index Mean
3.439
1.076
6.52
Landscape Shape Index
-0.31
-0.45
-0.07
sigma.sq
4.392
4.304
4.449
Phi
75.14 18.076 189.105
effective range 3/phi
0.04
0.017
0.191
Max intersite distance = 1.260

Table D-5. 1_2k wetland class spatial binomial GLM
Summary 1_2k Wetland 3 Chains at 100,000
Parameters
50.0%
2.5%
Intercept
-7.448 -8.688
Shape Index Mean
0.368 -0.401
Landscape Shape Index
-0.184 -0.295
sigma.sq
2.495
2.194
Phi
22.312 16.076
effective range 3/phi
0.134
0.09
max intersite distance = 1.277

97.5%
-7.447
1.16
0.011
2.897
33.448
0.188

Table D-6. 1_2k agriculture/forest class spatial binomial GLM
Summary 1_2k Agriculture/Forest 3 Chains at 100,000
Parameters
50.0%
2.5%
97.5%
Intercept
-9.096 -10.403
-6.957
Shape Index Mean
2.240
0.555
3.582
Landscape Shape Index
-0.255
-0.424
-0.242
sigma.sq
1.547
1.112
1.733
Phi
449.736 319.780 592.320
effective range 3/phi
0.007
0.005
0.009
max intersite distance = 1.230

151

Table D-7. 1_6k forest class spatial binomial GLM
Summary 1_6k Forest 3 Chains at 50,000
Parameters
50.0%
2.5%
Intercept
-9.490 -9.759
Shape Index Mean
-0.111 -2.042
sigma.sq
6.166
4.735
Phi
9.049
6.944
effective range 3/phi
0.332
0.229
max intersite distance = 1.260

97.5%
-6.676
0.629
6.917
13.169
0.434

Table D-8. 1_6k wetland class spatial binomial GLM
Summary 1_6k Wetland 3 Chains at 100,000
Parameters
50.0%
2.5%
Intercept
-9.375 -9.485
Shape Index Mean
0.870 -0.060
Landscape Shape Index
0.051 -0.028
sigma.sq
3.969
3.623
Phi
17.402 11.701
effective range 3/phi
0.172
0.144
max intersite distance = 1.250

97.5%
-8.770
1.127
0.080
6.301
20.866
0.259

Table D-9. 1_6k agriculture/forest spatial binomial GLM
Summary 1_6k Agriculture/Forest 3 Chains at 50,000
Parameters
50.0%
2.5% 97.5%
Intercept
10.739 11.111 10.168
Shape Index Mean
3.635
3.004
3.741
Landscape Shape Index
-0.248 -0.319 -0.208
sigma.sq
1.921
1.334
2.199
Phi
16.864 10.402 33.195
effective range 3/phi
0.178
0.093
0.292
max intersite distance = 1.224

152

Table D-10. 2k forest class spatial binomial GLM
Summary 2k Forest 3 Chains at 50,000
Parameters
50.0%
2.5%
Intercept
10.811 11.266
Percent Land Cover
Adjacency
0.024
0.014
sigma.sq
5.164
4.708
Phi
18.343 15.803
effective range 3/phi
0.164
0.131
max intersite distance = 1.174

97.5%
-9.358
0.029
5.165
23.018
0.190

Table D-11. 2k wetland class spatial binomial GLM
Summary 2k Wetland 3 Chains at 50,000
Parameters
50.0%
2.5%
Intercept
-7.522 -8.621
Percent Land Cover
Adjacency
-0.004 -0.020
sigma.sq
3.376
2.859
Phi
25.083 21.627
effective range 3/phi
0.120
0.049
max intersite distance =

97.5%
-6.267
0.002
4.107
63.926
0.139

Table D-12. 2k agriculture/forest class spatial binomial GLM
Summary 2k Agriculture/Forest 3 Chains at 50,000
Parameters
50.0%
2.5%
97.5%
Intercept
-9.434 -11.430
-8.954
Shape Index Mean
2.448
1.700
4.771
Landscape Shape Index
-0.253
-0.492
-0.120
sigma.sq
2.232
1.873
2.707
Phi
366.493 334.177 489.995
effective range 3/phi
0.008
0.006
0.009
max intersite distance = 1.256

153

LITERATURE CITED

154

Aiga H, Amano T, Cairncross S, Domako JA, Nanas O, and Colemen S (2004). Assessing
Water-Related Risk Factors for Buruli Ulcer: A Case-Control Study in Ghana. American
Journal of Tropical Medicine and Hygiene, 71(4), 387-392.
Abrahams EW, Tonge JI (1964). Mycobacterium ulcerans infection in Queensland. Medical
Journal of Australia i: 334.s
Akaike H. A New Look at the Statistical Model Identification (1974). IEEE Transactions on
Automatic Control, AC-19(6).
Amofah G, Bonsu F, Tetteh C, Okrah J, Asamoa K, Asiedu K, Addy J (2002). Buruli Ulcer in
Ghana: Results of a National Case Search. Emerging Infectious Diseases.
Amofah GK, Sagoe-Moses C, Adjei-Acquah C, Frimpong EH (1993). Epidemiology of Buruli
ulcer in Amansie West district, Ghana. Transactions of the Royal Society of Tropical
Medicine and Hygiene, 87, 644-645.
Anderson JR, Hardy EE, Roach JT, and Witmer RE (1976). A Land Use and Land Cover
Classification System for Use with Remote Sensor Data. Geological Survey Professional
Paper 964.
Aujoulat I, Johnson C, Zinsou C, Guedenon A, Portaels F (2003). Psychosocial aspects of health
seeking behaviours of patients with Buruli ulcer in southern Benin. Tropical Medicine
and International Health 8, 750-759.
Beck LR, Lobitz BM, Wood BL (2000). Remote Sensing and Human Health: New Sensors and
New Opportunities. Emerging Infectious Diseases, 6(3).
Benbow ME, Williamson H, Kimbirauskas R, McIntosh MD, Kolar R, Quaye C, Akpabey F,
Boakye D, Small P, and Merritt RW (2008). Aquatic Invertebrates as Unlikely Vectors of
Buruli Ulcer Disease. Emerging Infectious Diseases, 14(8) <www.cdc.gov/eid>
Beven KJ, and Kirkby MJ (1979). A physically based, variable contributing area model of basin
hydrology. Hydrological Sciences, Bulletin 24(1).
Blackburn JK, McNyset KM, Curtis A, Hugh-Jones ME(2007). Modeling the Geographic
Distribution of Bacillus anthracis, the Causative Agent of Anthrax Disease, for the
Contiguous United States using Predictive Ecologic Niche Modeling. American Journal
of Tropical Medicine and Hygiene, 77(6):1103-1110.
Bland JM, Altman DG (1998). Bayesians and Frequentists. Statistics Notes. British Medical
Journal, 317.
Bolstad P. GIS Fundamentals: A First Text on Geographic Information Systems, Second Ed.
2005. Eider Press, White Bear Lake, MN 55110.

155

Boyd HA, Flanders WD, Addiss DG, and Waller LA (2005). Residual Spatial Correlation
Between Geographically Referenced Observations A Bayesian Hierarchical Modeling
Approach. Epidemiology,16(4).
Brooks SP, and Gelman A (1998). General Methods for monitoring Convergence of Iterative
Simulations. Journal of Computational and Graphical Statistics, 7(4), 434-455.
Brou T, Broutin H, Elguero E, Asse H, Guegan J (2008). Landscape Diversity Related to Buruli
Ulcer Disease in Côte d’Ivoire. PLOS Neglected Tropical Diseases, 2(7), e271.
Brownstein JS, Skelly DK, Holford TR, Fish D (2005). Forest fragmentation predicts local scale
heterogeneity of Lyme disease risk. Oecologia, 146, 469-475.
Condeso TE and Meentemeyer RK (2007). Effects of landscape heterogeneity on the emerging
forest disease sudden oak death. Journal of Ecology, 95, 364-375.
Confalonieri UEC (2005). The Millennium Assessment: Tropical Ecosystems and Infectious
Diseases. EcoHealth 2, 231-233.
Centers for Disease Control (2005). Burul Ulcer. National Center for Immunization and
Respiratory Disease: Division of Bacterial Diseases. Last Accessed Online 22 Nov 2010
<http://www.cdc.gov/ncidod/dbmd/diseaseinfo/buruliculer_t.htm>
Central Intelligence Agency. The World Factbook: Africa: Benin. Last Accessed Online 22 Nov
2010 <https://www.cia.gov/library/publications/the-world-factbook/geos/bn.html>
Chauty A, Ardant MF, Adeye A, Euverte H, Guedenon A, Johnson C, Aubry J, Nuermberger E,
Grosset J (2007). Promising Clinical Efficacy of Streptomycin-Rifampin Combination for
Treatment of Buruli Ulcer (Mycobacterium ulcerans Disease). Antimicrobial Agents and
Chemotherapy, 4029-4035.
Clancey JK, Dodge OG, Lunn HF (1962). Study of a mycobacterium causing skin ulceration in
Uganda. Ann. Soc. Belge Med. Trop, 4, 585-590.
Clark WC, Kates RW, et al. editors. The Earth as transformed by Human Action. Cambridge:
Cambridge University Press; 1990.
Clarke KC Ed. Getting Started with Geographic Information Systems. Fourth Edition. Chapter 3:
Maps as Numbers, pp 67 – 70. Prentice Hall Series in Geographic Information Systems.
Pearson Education Inc.Upper Saddle River, New Jersey.
Coetzee M, Craig M, le Sueur D (2000). Distribution of African Malaria Mosquitoes Belonging
to the Anopheles gambiae Complex. Parasitology Today, 16(2).
Cook A, Jardine, A, Weinstein P (2004). Using Human Disease Outbreaks as a Guide to
Multilevel Ecosystem Interventions. Environmental Health Perspectives, 112(11).
156

Cook RD (1977). Detection of Influential Observations in Linear Regression. Technometrics,
19(1).
Danson FM, Craig PS, Man W, Shi D, and Giraudoux P (2004). Landscape Dynamics and Risk
Modeling of Human Alveolar Echinococcosis. Photogrammetric Engineering & Remote
Sensing, 70(3), 359-366.
de Knegt HJ, van Langevelde F, Coughenour MB, Skidmore AK, de Boer WF, Heitkōnig IMA,
Knox NM, Slotow R, van der Waal C, and Prins HHT (2010). Spatial autocorrelation and
the scaling of species-environment relationships. Ecology, 91(8), 2455-2465.
Debacker M, Aguiar J, Steunou C, Zinsou C, Meyers WM, Guedenon A, Scott J, Dramaix M,
Portaels F (2004a). Mycobacterium ulcerans Disease (Buruli Ulcer) in Rural Hospital,
Southern Benin, 1997-2001. Emerging Infectious Diseases, 10(8), <www.cdc.gov/eid>
Debacker M, Aguiar J, Steunou C, Zinsou C, Meyers WM, Scott JT, Dramaix M, Portaels F
(2004b). Mycobacterium ulcerans disease: role of age and gender in incidence and
morbidity. Tropical Medicine and International Health, 9, 1297-1304.
Debacker M, Zinsou C, Aguiar J, Meyers WM, Portaels F (2003). First case of Mycobacterium
ulcerans disease (Buruli ulcer) following a human bite. Clinical Infectious Disease.
1;36(5).
Debacker M, Aguiar J, Steunou C, Zinsou C, Meyers WM, Portaels F (2005). Buruli Ulcer
Recurrence, Benin. Emerging Infectious Diseases 11(4).
Dobos, Karen M, Quinn, Frederick D, Ashford, David A, Horsburgh, Robert C, King, C Harold.
Emergence of a Unique Group of Necrotizing Mycobacterial Diseases (1999). Emerging
Infectious Diseases, 5(3).
Dossou KM and Gléhouenou-Dossou (2007). The vulnerability to climate change of Cotonou
(Benin): the rise in sea level. Environment and Urbanization 19(65).
Duker AA, Carranza EJM, Hale M. (2004). Spatial dependency of Buruli ulcer prevalence on
arsenic-enriched domains in Amansie West District, Ghana: implications for arsenic
mediation in Mycobacterium ulcerans infection. International Journal of Health
Geographics,3(19).
Duker AA, Stein A, Hale M. A statistical model for spatial patterns of Buruli ulcer in the
Amansie West district, Ghana (2006). International Journal of Applied Earth
Observation and Geoinformation, 8, 126-136.
Durnez L, Suykerbuyk P, Nicolas V, Barrière, Verheyen E, Johnson CR, Leirs H, Portaels F
(2010). Terrestrial Small Mammals as Reservoirs of Mycobacterium ulcerans in Benin.
Applied and Environmental Microbiology, July, 4574-4577.
157

Eddyani M, Portaels F (2007). Survival of Mycobacterium ulcerans at 37°C. European Society
of Clinical Microbiology and Infectious Diseases, CMI 13, 1023-1035.
Eddyani M, Ofori-Adjei D, Teugels G, De Weirdt D, Boakye D, Meyers WM, Portaels F (2004).
Potential role for fish in transmission of Mycobacterium ulcerans disease (Buruli ulcer):
an environmental study. Applied Environmental Microbiology,70(9), 5679-81.
Elsner L, Wayne J, O’Brien CR, McCowan C, Malik R, Hayman JA, Globan M, Lavender CJ,
Fyfe JA (2008). Localized Mycobacterium ulcerans infection in a cat in Australia.
Journal of Feline Med Surg, 10, 407-12.
Epstein PR. Biodiversity, Climate Change, and Emerging Infectious Diseases. Chapter 4. pg 27.
Conservation Medicine: Ecological Health in Practice. Ed. Aguirre AA, Ostfeld RS,
Tabor GM, House C, Pearl MC. Oxford University Press. New York, New York. 2002.
Foley JA, DeFries R, Asner GP, Barfoard C, Bonan G, Carpenter SR, Chapin FS, Coe MT, Daily
GC, Gibbs HK, Helkowski JH, Holloway T, Howard EA, Kucharik CJ, Monfreda C, Patz
JA, Prentice IC, Ramankutty N, Snyder PK (2005). Review: Global Consequences of
Land Use. Science. 22 July (309).
Fyfe JA, Lavender CJ, Johnson PD, Globan M, Sievers A, Azuolas J and Stinear TP (2007).
Development and application of two multiplex real-time PCR assays for the detection of
Mycobacterium ulcerans in clinical and environmental samples. Applied Environmental
Microbiology, 73, 4733–4740.
Fyfe JAM, Lavender CJ, Handasyde KA, Legione AR, O’Brien CR, Stinear TP, Pidot SJ,
Seemann T, enbow ME, Wallace JR, McCowan C, Johnson PDR (2010). A Major Role
for Mammals in the Ecology of Mycobacterium ulcerans. PLoS Neglected Tropical
Diseases, 4(8), e791.
George, KM, Chatterjee, D, Gunawardana, G. (1999). Mycolactone: a polyketide toxin from
Mycobacterium ulcerans required for virulence. Science, 283, 854-857.
Gergel S. E., Turner M.G. Ed. (2002). Learning Landscape Ecology: A Practical Guide to
Concepts and Techniques. Springer, New York, NY.
Godwin E. (2010 Nov 2). Floods waters from Dam carry away Traditional Ruler. Ghana News
Link. Accessed Online 22 Nov 2010. <http://www.ghananewslink.com>
Goldberg TL, T.R. Gillespie, I.B. Rwego, E.L. Estoff., and C.A. Chapman (2008). Forest
fragmentation as cause of bacterial trasmission among nonhuman primates, humans, and
livestock, Uganda. Emerging Infectious Diseases. 14(9),1375-1382.

158

Graham AJ, Danson FM, Giraudoux P, Craig PS (2004). Ecological epidemiology: landscape
metrics and human alveolar echinococossis. Acta Tropica, 91, 267-278.
Haase, Peter (1995). Spatial pattern analysis in ecology based on Ripley’s K-function:
Introduction and methods of edge correction. Journal of Vegegation Science, 6, 575-582.
Hansen MC, Stehman SV, Potapov PV, Loveland TR, Townshend JRG, DeFries RS, Pittman
KW, Arunarwatil B, Stolle F, Steininger MK, Carroll M, DiMiceli C (2008). Humid
tropical forest clearing from 2000 to 2005 quantified by using multiptemporal and
mutiresolution remotely sensed data. PNAS, 105(27), 9941.
Hayman J (1991). Postulated Epidemiology of Mycobacterium ulcerans Infection. International
Journal of Epidemiology, 20(4).
Helbert C, Dupuy D, Carraro L (2009). Assessment of uncertainty in computer experiments from
Universal to Bayesian Kriging. Applied Stochastic Models in Business and Industry,
25(2), 99-113.
Horsburgh Jr CR and Meyers WM. Buruli ulcer. In: Horsburgh Jr CR and Nelson AM (ed).
Pathology of emerging infections. Washington DC: American Society for Microbiology.
P 119-126.
Hughes RH and Hughes, JS. A Directory of African Wetlands. IUCN, Gland, Switzerland &
Cambridge, UK/UNEP, Nairobi, Kenya, WCMC, Cambridge, UK.
Huygen K, Adjei O, Affolabi D, Bretzel G, Demangel C, Fleischer B, Johnson RC, Pedrosa J,
Phanzu DM, Phillips RO, Pluschke G, Siegmund V, Singh M, van der Werf TS,
Wansbrough-Jones M, and Portaels F (2009). Buruli ulcer disease: prospects for a
vaccine. Medical Microbiology and Immunology, 198, 6-77.
Iverson LR (1988). Land-use changes in Illinois, USA: The influence of landscape attributes on
current and historic land use. Landscape Ecology, 2(1), 45-61.
Jenkin JA, Smith M, Fairley M, Johnson PDR (2002). Acute, oedematous Mycobacterium
ulcerans infection in a farmer from far north Queensland. Medical Journal of Australia,
176,180-182.
Jensen J.R. Introductory digital image processing: a remote sensing perspective, 2nd ed. 1996.
Simon & Schuster, Upper Saddle River, NJ.
Johnson PDR, Stinear TP, and Hayman JA (1999). Mycobacterium ulcerans -- a mini-review.
Journal of Medical Microbiology, 48, 511-13.
Johnson PDR, Stinear T, Small PLC, Plushke G, Merritt RW, Portaels F, Huygen K, Hayman
JA, Asiedu K (2005). Buruli Ulcer (M. ulcerans Infection): New Insights, New Hope for
Disease Control. PLoS Med, 2(4), e108.

159

Johnson PDR, Azuolas J, Lavender CJ, Wishart E, Stinear TP, Hayman JA, Brown L, Jenkin
GA, and Fyfe JAM (2007). Mycobacterium ulcerans in Mosquitoes Captured during
Outbreak of Buruli Ulcer, Southeastern Australia. Emerging Infectious Diseases, 13(11).
<www.cdc.gov/eid>
Johnson PDR, Lavender CJ (2009). Correlation between Buruli Ulcer and Bector-borne
Notifiable Diseases, Victoria, Australia. Emerging Infectious Diseases, 15(4).
<www.cdc.gov/eid>
Keitt TH, Bjørnstad ON, Dixon PM, Citron-Pousty S (2002). Accounting for spatial pattern
when modeling organism-environment interactions. Ecography, 25, 616-625.
Kitron U (1998). Landscape Ecology and Epidemiology of Vector-Borne Diseases: Tools for
Spatial Analysis. Journal of Medical Entomology, 35(4), 435-445.
Krummel JR, Gardner RH, Sugihara G, O’Neill RV, Coleman PR (1987). Landscape Patterns in
a Disturbed Environment. Oikos, 48(3), 321-324.
Legendre P and Fortin MJ. Spatial pattern and ecological analysis. Vegetation, 80, 107-138.
Lennon JL (2000). Red-shifts and red herrings in geographical ecology. Ecography, 23(1), 101113.
Lillesand and Kiefer. Remote Sensing and Image Interpretation. 4th ed. New York: John Wiley &
Sons.
MacCallum P, Tolhurst JC, Buckle G, Sissons HA (1948). A new mycobaterial infection in man.
Journal of Pathologic Bacteriology,60, 93-122.
Malone JB, Nieto P, Tadesse A (2006). Biology-based mapping of vector-borne parasites by
geographic information systems and remote sensing. Parassitologia,48, 77-79.
Marion E, Eyangoh S, Yeramian E, Doannio J, Landier J, Aubry J, Fontanet A, Rogier C,
Cassisa V, Cottin J, Marot A, Eveillard M, Kamdem Y, Legras P, Deshayes C, SaintAndré J, and Marsollier L (2010). Seasonal and Regional Dynamics of M. ulcerans
Transmission in Environmental Context: Deciphering the Role of Water Bugs as Hosts
and Vectors. PLoS Neglected Tropical Diseases. 4(7): e731.
Marsollier L, Robert R, Aubry J, Saint Andre J, Kouakou H, Legras P, Manceau A, Mahaza C,
and Carbonnelle B (2002). Aquatic Insects as a Vector for mycobacterium ulcerans.
Applied and Environmental Microbiology, 68(9), 4623-4628.
Marsollier L, Aubry J, Saint-Andre P, Robert R, Legras P, Manceau AL, Bourdon S, Audrain C,
and Carbonnelle B (2003). Ecology and transmission of Mycobacterium ulcerans.
Pathologie Biologie 51,490-495

160

Marsollier L, Severin T, Bubry J, Merritt RW, Saint-Andre JP, Legras P, Manceau A, Chauty A,
Carbonnelle B, and Cole ST. (2004a). Aquatic snails, passive hosts of Mycobacterium
ulcerans. Applied and Enviornmental Microbiology, 70, 6296-6298.
Marsollier L, Stinear T, Aubry J, Saint Andre JP, Robert R, Legras P, Manceau A, Audrain C,
Bourdon S, Kouakou H, and Carbonnelle B (2004b). Aquatic Plants Stimulate the
Growth of and Biofilm Formation by Mycobacterium ulerans in Axenic Culture and
Harbor These Bacteria in the Environment. Applied Environmental Microbiology, 70,
1097-1103.
Marston BJ, Diallo MO, Horsburgh CR Jr, Diomande I, Saki MZ, Kanga JM, Patrice G, Lipman
HB, Ostroff SM, Good RC (1995). Emergence of Buruli ulcer disease in the Daloa region
of Côte d’Ivoire. American Journal of Tropical Medicine and Hygiene, 52(3), 219-24.
Mather PM (2004). Computer processing of remotely sensed images: an introduction 3rd
Edition. John Wiley & Sons Ltd. West Sussex, England.
Matson PA, Parton WJ, Power AG, and Swift MJ. Agricultural Intensification and Ecosystem
Properties (1997). Science 277, 504.
McGarigal K, Marks B. FRAGSTATS: spatial pattern analysis program for quantifying
landscape structure. Gen. Tech. Rep. PNW-GTR-351. Portland, OR: U.S. Department of
Agriculture, Forest Service, Pacific Northwest Research Station. 122 p.
Merrit RW, Benbow ME, Small PLC (2005). Unraveling an emerging disease associated with
disturbed aquatic environments: the case of Buruli ulcer. Reivew. Frontiers in Ecology
and the Environment, 3(6), 323-331.
Merritt RW, Walker ED, Small PLC, Wallace JR, Johnson PDR, Benbow ME, Boakye DA
(2010). Buruli Ulcer Disease: A Systematic Review of Ecology and Transmission. PLoS
Neglected Tropical Diseases, 4(12), e911.
Meyers, Derek H (2007). Letters: Mycobacterium ulcerans infection: an eponymous ulcer.
Medical Journal of Australia, 187(1), 63.
Meyers WM, Shelly WM, Connor DH, Meyers EK. Human Mycobacterium ulcerans infections
developing at sites of trauma to skin (1974). American Journal of Tropical Medicine and
Hygiene, 23, 919.
Meyers WM, Tignokpa N, Priuli GB, and Portaels F. Mycobacterium ulcerans infection (Buruli
ulcer): first reported patients in Togo (1996). British Journal of Dermatology, 134, 11161121.
Mitchell, PJ, Jerrett, IV, Slee, KJ (1984). Skin ulcers caused by Mycobacterium ulcerans in
koalas near Bairnsdale, Australia. Pathology 16, 256-260.

161

Morse S (1995). Factors in the Emergence of Infectious Diseases. Emerging Infectious Diseases,
1(1).
Mosi L, Williamson H, Wallace JR, Merritt RW, Small PLC (2008). Persistent Association of
Mycobacterium ulcerans with West African Predaceous Insects of the Family
Belostomatidae. Applied Envrionmental Microbiology, 74(22) 7036-7042.
Moyeed RA and Papritz A (2002). An Empirical Comparison of Kriging Methods for Nonlinear
Spatial Point Prediction. Mathematical Geology, 34(4).
Nackers F, Johnson RC, Glynn JR, Zinsou C, Tonglet R, and Portaels F (2007). Environmental
and Health-related Risk Factors for Mycobacterium ulcerans Disease (Buruli Ulcer) in
Benin. American Journal of Tropical Medicine and Hygiene. 77(5), 834-836.
Narumalani S, Mishra DR, and Rothwell RG (2004). Change detection and landscape metrics for
inferring anthropogenic processes in the greater EFMO area. Remote Sensing of the
Environment, 91(3-4), 478-489.
Noeske J, Kuaban C, Rondini S, Sorlin P, Ciaffi L, Mbuagbaw J, Portaels F, and Pluschke G
(2004). Buruli Ulcer Disease in Cameroon Rediscovered. American Journal of Tropical
Medicine and Hygiene, 70(5), 520-526.
O’Callaghan JF and Mark DM (1984). The Extraction of Drainage Networks from Digital
Elevation Data. Computer Vision, Graphics, and Image Processing. 28, 323-344.
Oluwasanmi JO, Solanke TF, Olurin EO, Itayemi SO, Alabi GO, and Lucas AO (1976).
Mycobacterium Ulcerans (Buruli) Skin Ulceration in Nigeria. The American Journal of
Tropical Medicine and Hygiene, 25(1).
Ostfeld RS, Glass GE, Keesing F (2005). Spatial epidemiology: an emerging (or re-emerging)
discipline. TRENDS in Ecology and Evolution, 20(6).
Patz JA and Confalonieri UEC, coord. eds. Human Health: Ecosystem Regulation of Infectious
Diseases. In Hassan RM, Scholes R, and Ash N Ecosystems and Human Well-being:
Current State and Trends: Findings of the Condition and Trends Working Group. Pp
393-415. Millenium Ecosystem Assessment (2005). Island Press, Washington D.C.,
USA.
Patz JA, Graczyk TK, Geller N, Vittor AY (2000). Effects of environmental change on emerging
parasitic diseases. International Journal for Parasitology, 30, 1395-1405.
Patz JA, Olson SH, Uejio CK, Gibbs HK (2008). Disease Emergence from Global Climate and
Land Use Change. Medical Clinics of North America, 92, 1473-1491.
Pavlovsky, EN (1966). The natural nidality of transmissible disease. University of Illinois Press,
Urbana.

162

Pickett STA and Cadenasso ML (1995). Landscape Ecology: Spatial Heterogeneity in Ecological
Systems. Science, 269(5222), 331-334.
Pickett STA and White PS, eds. The Ecology of Natural Disturbance and Patch Dynamics.
NewYork: Academic. 472 pp. 1985.
Plowright RK, Sokolow SH, Gorman ME, Daszak P, Foley JE (2008). Causal inference in
disease ecology: investigating ecological drivers of disease emergence. Frontiers in
Ecology and the Environment, 6(8), 420-429.
Pongsiri MJ, Roman J, Ezenwa VO, Goldberg TL, Koren HS, Newbold SC, Ostfeld RS,
Pattanayak SK, and Salkeld DJ (2009). Biodiversity Loss Affects Global Disease
Ecology. BioScience, 59(11).
Pouillot R, Matias G, Wondje CM, Portaels F, Valin N, Ngos F, Njikap A, Marsollier L,
Fontanet A, Eyangoh S (2007). Risk Factors for Buruli Ulcer: A Case Control Study in
Cameroon. PLOS Neglected Tropical Diseases, 1(3), e101.
Portaels F, Elsen P, Guimaraes-Peres A, Fonteyne P-A, Meyers WM (1999). Insects in the
transmission of Mycobacterium ulcerans infection. Lancet, 353(9157), 986.
Portaels F, Aguiar J, Debacker M, Steunou C, Zinsou C, Guédénon A, Meyers WM (2002).
Prophylactic effects of Mycobacterium bovis BCG vaccination against osteomyelitis in
children with Mycobacterium ulcerans disease (Buruli Ulcer). Clininical Diagnostics and
Laboratory Immunology, 9, 1389–1391.
Portaels F, Meyers WM, Ablordey An, Castro AG, Chemlal K, de Rijk P, Elsen P, Fissette K,
Fraga AG, Lee R, Mahrous E, Small PLC, Stragier P, Torrado E, Aerde AV, Silva MT,
Pedrosa J (2008). First Cultivation and Characterization of Mycobacterium ulcerans from
the Environment. PLOS Neglected Tropical Diseases, 2(3), e178.
Quek TY, Athan E, Henry MJ, Pasco JA, Redden-Hoare J, Hughes A, Johnson PDR (2007). Risk
factors for Mycobacterium ulcerans infection, southeastern Australia. Emerging
Infectious Diseases, 13(11), 1661–6.
Raghunathan, PL, Whitney, EA, Asamoa, S, Stienstra, Y, Taylor Jr, TH, Amofah GK, OforiAdjei, David, Dobos, Karen, Guarner, Jeannette, Martin, Stacey, Pathak, Sonal, Klutse,
Erasmus, Etuaful, Samuel, van der Fraaf, Winette TA, van der Werf, Tjip S, King, CH
(2005). Risk Factors for Buruli ulcer disease (Mycobacterium ulcerans infection): Results
from a Case-Control study in Ghana. Clinical Infectious Diseases, 40(10), 1445-53.
Riitters KH, O’Neill RV, Hunsaker CT, Wickham JD, Yankee DH, Timmins SP, Jones KB, and
Jackson BL (1995). A factor analysis of landscape pattern and structure metrics.
Landscape Ecology, 10(1) 23-39.
RESPOND Health Mapping Activity Ivory Coast. Document No: S-18106-DLR-I1.00P1_Health_Mapping. Verson 1.00. ESA/ESRIN contract No: 18106 05/09/06.
163

Ross BC, Honson PD, Oppedisano F, Marino L, Sievers A, Stinear T, Hayman JA, Veitch MG,
Robins-Browne RM (1997). Detection of Mycobacterium ulcerans in environmental
samples during an outbreak of ulcerative disease. Applied Environmental Microbiology,
63(10), 4135-4138.
SAS/STAT(R) 9.2 User’s Guide, Second Edition.
http://support.sas.com/documentation/cdl/en/statug/63033/HTML/default/viewer.htm#/do
cumentation/cdl/en/statug/63033/HTML/default/statug_mcmc_sect045.htm Accessed 27
June 2010.
Schabenberger O, and Gotway CA. Statistical Methods for Spatial Data Analysis. 2005.
Chapman & Hall/CRC, Boca Raton, FL.
Schmidt F and Persson A. Comparison of DEM Data Capture and Topographic Wetness Indices
(2003). Precision Agriculture, 4, 179-192.
Schunk M, Thompson W, Klutse E, Nitschke Jorg, Opare-Asamoah K, Thompson R,
Fleischmann E, Siegmund V, Herbinger Kh, Adjei O, Fleischer B, Loscher T, and Bretzel
G (2009). Outcome of Patients with Buruli Ulcer after Surgical Treatment with or
without Antimycobacterial Treatment in Ghana. American Journal of Tropical Medicine
and Hygiene, 81(1), 75-81.
Sizaire V, Nackers F, Comte E, Poratels F (2006). Mycobacterium ulcerans infection: control,
diagnosis, and treatment. The Lancet Infectious Diseases, 6(5), 288-296.
Sopah GE, Johnson RC, Chauty A, Dossou AD, Aguiar J, Salmon O, Portaels F, Asiedu K
(2007). Buruli ulcer surveillance, Benin, 2003-2005. Emerging Infectious Diseases,
13(9), 1374-1376.
Spiegelhalter DJ, Best NG, Carlin BP, and van der Linde A (2002). Bayesian measures of model
complexity and fit. Journal of the Royal Statistical Society: Series B (Statistical
Methodology), 64(4).
Spies TA, Ripple WJ, Bradshaw GA. Dynamics and Pattern of a Managed Coniferous Forest
Landsape in Oregon (1994). Ecological Applications, 4(3), 555-568.
Stinear TP, Seemann T, Pidot S, Wafa F, Yeysset G, Garnier T, Meurice G, Simon D, Bouchier
C, Ma L, Tichit M, Porter JL, Ryan J, Johnson PDR, Davies JK, Jenkin GA, Small PLC,
Jones LM, Tekaia F, Laval F, Daffé, Parkhill J, Cole ST (2007). Reductive evolution and
niche adaptation inferred from the genome of Mycobacterium ulcerans, the causative
agent of Buruli ulcer. Genome Research, 17(2), 192-200.
Stienstra Y, van der Graaf WTA, Asamoa K, and van der Werf TS. Beliefs and Attitudes Toward
Buruli Ulcer in Ghana (2002). American Journal of Tropical Medicine and Hygiene,
67(2), 207-213.
164

Stragier P, Ablordey A, Durnez L, Portaels F (2007). VNTR analysis differentiates
Mycobacterium ulcerans and IS2404 positive mycobacteria. Systematic and Applied
Microbiology, 30(7), 252-530.
Tobias NJ, Seemann T, Pidot S, Porter JL, Marsollier L, Marion E, Letournel F, Zakir T,
Azuolas J, Wallace JR, Hong H, Davies JK, Howden BP, Hohnson PDR, Jenkin GA, and
Stinear TP (2009). Mycolactone Gene Expression is Controlled by Strong SigA-Like
Promoters with Utility in Studies of Mycobacterium ulcerans and Buruli Ulcer. PLoS
Neglected Tropical Diseases, 3(11).
Tobler WR (1970). A Computer Movie Simulating Urban Growth in the Detroit Region.
Economic Geography, 46, Supplement: Proceedings. International Geographical Union.
Commission on Quantitative Methods, 234-240.
Turner MG and Gardner RH. Quantitative Methods in Landscape Ecology: An Introduction.
Chapter 1. Quantitative Methods in Landscape Ecology. pp 3-14. Springer-Verlag New
York Inc. New York, New York. USA. 1991.
Turner MG, Gardner RH, O’Neill RV. Landscape Ecology In Theory and Practice: Pattern and
Process. Springer-Verlag New Yrok, Inc. 2001.
Turner MG (2005). Landscape Ecology: What is the State of the Science? Annual Review of
Ecology and Evolutionary Systems, 36, 319-44.
Turner W., Spector S., Gardiner N., Fladeland M., Sterling E., and Steininger M (2003). Remote
sensing for biodiversity science and conservation. TRENDS in Ecology and Evolution,
18(6).
U.S. Agency for International Development (USAID), Bureau for Africa, Office of Sustainable
Development (AFR/SD). 118/119 Biodiversity and Tropical Forest Assessment for
Benin. EPIZ IQC: EPP-1-00-03-00014-00, Task Order 02. Accessed online 20 Oct 2010
<http://www.encapafrica.org/documents/biofor/Benin%20118%20119%20FINAL%20O
ctober%202007.pdf>
United Nations, Permanent Mission of the Republic of Benin to the United Nations. Country
Facts. Access online 20 Oct 2010 < http://www.un.int/wcm/content/site/benin/>
United Nations Statistical Division. World Statistics Pocketbook. Benin. Last Accessed Online
22 Nov 2010 < http://data.un.org/Search.aspx?q=benin>
Veitch MGK, Johnson PDR, Flood PE, Leslie DE, Street AC, and Hayman JA (1997). A large
localized outbreak of Mycobacterium ulcerans infection on a temperate southern
Australian island. Epidemiology and Infection, 119(3), 313-318.

165

van Ravensway J, Tsonis AA, Benbow ME, Campbell LP, Johnson PDR, Qi J. Regional-scale
Climate, Landscape, and Social Risk Factors Identified as Potential Drivers of the
Network of Buruli Ulcer Cases in Victoria, Australia. (IN PREP, 2010).
van Zyl A, Daniel J, Wayne J, McCowan C, Malik R, Jelfs P, Lavender CJ, and Fyfe JA (2010).
Mycobacterium ulcerans infections in two horses in south-eastern Australia. Australian
Veterinary Journal, 88,(3),101-106.
Yang K, Wang XH, Yang GJ, Wu XH, Qi YL, Li HJ, and Zhou XN (2008). An integrated
approach to identify distribution of Oncomelania hupensis, the intermediate host of
Schistosoma japonicum, in a mountainous region in China. International Journal for
Parasitology, 38(8-9), 1007-1016.
Wagner HH and Fortin MJ (2005). Spatial analysis of landscapes: concepts and statistics.
Ecology, 86(8), 1975-1987.
Wagner T, Benbow ME, Brenden TO, Qi J, Johnson RC (2008a). Buruli ulcer disease prevalence
in Benin, West Africa: associations with land use/cover and the identification of disease
clusters. International Journal of Health Geographics, 7, 25.
Wagner, Tyler, Benbow, M Eric, Burns, Meghan, Johnson, R Christian, Merritt, Richard W, Qi,
Jiaguo, Small, Pamela LC (2008b). A Landscape-based Model for Prediction
Mycobacterium ulcerans Infection (Buruli Ulcer Disease) Presence in Benin, West
Africa. EcoHealth, 5(1), 69-79.
Walsh DS, Portaels F, Meyers WM (2008). Buruli ulcer (Mycobacterium ulcerans infection).
Transactions of the Royal Society of Tropical Medicine and Hygiene,102(10), 969-78.
Waller LA and Gotway CA. Applied Spatial Statistics for Public Health Data. 2004. John Wiley
& Sons, Inc. Hoboken, NJ.
Wansbrough-Jones M, Phillips R (2006). Buruli ulcer: emerging from obscurity. Lancet,
367(9525), 1849-1858.
Wilcox BA, Gubler DJ (2005). Disease Ecology and the Global Emergence of Zoonotic
Pathogens. Environmental Health and Preventive , 10(5), 263-272.
Williamson HR, Benbow ME, Nguyen KD, Beachboard DC, Kimbirauskas RK, McIntosh MD,
Quaye C, Ampadu EO, Boakye D, Merritt RW, Small PLC (2008). Distribution of
Mycobacterium ulcerans in Buruli Ulcer Endemic and Non-Endemic Aquatic Sites in
Ghana. PLOS Neglected Tropical Diseases, 2(3), e205.
Wilcove DS, McLellan CH, Dobson AP (1986). Habitat fragmentation in the temperate zone.
Conservation Biology, ed. ME Soul´e, pp. 237–56. Sunderland, MA: Sinauer

166

With KA, Gardner RH, Turner MG (1997). Landscape connectivity and population distributions
in heterogeneous environments. Oikos, 78(1),151–69
World Health Organization (2000). Buruli ulcer-Mycobacterium ulcerans infection. Asiedu K,
Scherpbier R, Raviglione M., editors. Geneva: The Organization.
World Health Organization (2007). Fact Sheet: Buruli ulcer disease (Mycobacterium ulcerans
infection). <http://www.who.int/mediacentre/factsheets/fs199/en/print.html>
World Health Organization (2009). Weekly epidemiological record, 6(84), 41-48.
<http://www.who.int/wer>

167