HARNESSING HIV and SYPHILIS NETWORKS to REDUCE HIV TRANSMISSION in NORTH CAROLINA Rachael Marie Billock a Dissertation Submitted

HARNESSING HIV AND SYPHILIS NETWORKS TO REDUCE HIV TRANSMISSION IN NORTH CAROLINA

Rachael Marie Billock

A dissertation submitted to the faculty at the University of North Carolina at Chapel Hill in partial fulfillment of the requirements for the degree of Doctor of Philosophy in the Department of Epidemiology in the Gillings School of Global Public Health.

Chapel Hill 2020

Approved by:

Kimberly A. Powers

Jennifer L. Lund

Peter J. Mucha

Brian W. Pence

Erika Samoff

ABSTRACT

Rachael Marie Billock: Harnessing HIV and Syphilis Networks to Reduce HIV Transmission in North Carolina (Under the direction of Kimberly Powers)

Sexual contact networks containing persons diagnosed with HIV and/or syphilis are efficient platforms for delivery of enhanced HIV treatment support to persons living with HIV

(PLWH) and prevention resources to HIV-negative persons. We conducted three inter-related studies among men who have sex with men (MSM) in North Carolina (NC) to assess the potential for network-based interventions to reduce HIV transmission.

We first generated independent and combined HIV and syphilis networks using contact tracing data for all MSM diagnosed with HIV or syphilis 2015-2017 in NC. We identified network communities, or clusters of densely connected nodes, in the combined network and evaluated interconnectivity between the syphilis and HIV networks by community. Heightened interconnectivity was associated with younger median age; higher proportions of persons self- identifying as Black, non-Hispanic; and higher proportions of syphilis cases diagnosed at sexually transmitted disease clinics.

Next, we used NC surveillance data to identify two types of qualifying “network events” among MSM in NC 01/2013-06/2017: being diagnosed with early syphilis or being named as a recent sexual contact of a person diagnosed with HIV or early syphilis. We estimated prevalent and incident viral suppression among previously diagnosed PLWH at and after network events

(52.6% and 35.4%, respectively) and assessed the effect of contact tracing services on incident viral suppression. The six-month cumulative incidence of reported viral suppression was 13.1

(95% CI: 8.9, 17.4) percentage points higher after events among persons reached vs. not

iii reached by these services. Using linked insurance claims data, we also evaluated prevalent and incident pre-exposure prophylaxis (PrEP) use among HIV-negative network members (5.4% and

4.1%, respectively).

Finally, we generated static contact networks with community structure drawn from the combined HIV/syphilis network described above and varied endemic HIV prevalence across network communities. We applied a stochastic transmission model to simulate HIV spread, evaluated community-level HIV incidence rates by endemic community-level HIV prevalence, and modeled the efficiency of community-prevalence-based vs. randomly allocated interventions.

PrEP interventions were most efficient when preferentially deployed to susceptible persons in communities with higher endemic HIV prevalences, while viral suppression intervention efficiency did not vary significantly by intervention allocation scheme.

For the people of North Carolina living with HIV

ACKNOWLEDGMENTS

I must first thank Kim Powers, my dissertation advisor, for her endless support during my time at UNC. Kim has generously given so much of her time and has acted as a true mentor over the past five years. She patiently guided my exploration of many potential research questions in the conceptualization of this project, thoughtfully edited countless manuscript drafts, and consistently posed insightful questions and creative solutions. This dissertation would not exist without her kind and knowledgeable direction, or without her personal support at each stage of the process.

Prior to beginning this dissertation, I worked as a research assistant with Ann Dennis.

Ann introduced me to the field of network analysis, and the skills that I gained under her gentle and expert supervision provided the foundation for these analyses. Ann also supported my development as I learned to work with the surveillance data applied in this dissertation and connected me with collaborators at the NC Division of Public Health (DPH).

I am immensely grateful to everyone at NC DPH, particularly the HIV/STD/Hepatitis

Surveillance Unit, for sharing their skills, knowledge, and resources. Among many others, Erika

Samoff, John Barnhart, Jason Maxwell, and Anna Cope have patiently answered my many questions and modeled constant dedication to the people of North Carolina. Erika has provided an essential voice on my dissertation committee, aiding in the translation of analyses to programmatically meaningful results. She is truly an advocate for people living with HIV, and she inspires me to strive to be the same. Much of the data used in this study were provided by

NC DPH. NC DPH does not take responsibility for the scientific validity or accuracy of methodology, results, statistical analyses, or conclusions presented.

In addition to serving on my committee, Brian Pence is the supervisor of the “Training in

Infectious Disease Epidemiology” NIAID grant (5T32AI070114-12) that has funded my dissertation research and supported my professional development over the past two years. In both roles, Brian has offered straightforward solutions to any potential problems, always sharpening my work and fostering my progression as a scientist. Jennifer Lund provided an introduction to pharmacoepidemiology and invaluable guidance in analyses of insurance claims data as a member of my committee. Jenny’s enthusiasm made a new field more approachable and her many insights refined my analyses. Peter Mucha’s work on community detection in networks inspired much of this dissertation, and his personal support in translating his research was instrumental to its success. His service on my committee allowed me to explore new analytical approaches and to envision the application of these methods to public health.

I would like to thank Valerie Hudock and Jennifer Moore for their help with the administrative side of this program, for always keeping me on track, and for sharing my excitement at each new stage of the MSPH and PhD. I am also very grateful for the support provided by the Cornoni-Huntley Scholarship in the final year of this degree.

I greatly appreciate the support of the staff at the Cecil G. Sheps Center for Health

Services Research at UNC, particularly Abigail Haydon, Roger Akers, Joe Meskey, and Eric

Eyre. Their patient guidance and willingness to support student work was critical to the success of PrEP-related analyses in Aim 2. Database infrastructure used for this project was supported by the Cecil G. Sheps Center for Health Services Research and the CER Strategic Initiative of

UNC’s Clinical and Translational Science Award (UL1TR001111). I would also like to thank Blue

Cross Blue Shield of North Carolina for allowing their data to be applied in this dissertation.

This work would not have been possible without the contribution of the Disease

Intervention Specialists (DIS) who perform contact tracing for persons diagnosed with HIV and syphilis in North Carolina. The often extraordinary efforts of DIS to reach persons diagnosed with HIV and syphilis and their named contacts are the frontline of HIV and syphilis control in

vii

NC, and meticulous data collection throughout this process enables analyses like those performed here. Most of all, I would like to acknowledge the persons living with HIV, and all others, who are represented in this dissertation.

Finally, I would like to thank the many friends that I have made within this program, who helped me believe that I belonged at UNC, and the friends outside of the program who loved, distracted, and uplifted me throughout this process. I want to thank my family, for instilling the joy of science early, and my partner, for his positivity, perspective, and unwavering support. I would not be here without all of you.

viii

TABLE OF CONTENTS

LIST OF TABLES ...... xiv

LIST OF FIGURES ...... xv

LIST OF ABBREVIATIONS ...... xvi

CHAPTER ONE: SPECIFIC AIMS ...... 1

CHAPTER TWO: BACKGROUND ...... 5

A. HIV Epidemiology ...... 5

1. HIV in the Southern United States ...... 5

2. HIV in North Carolina ...... 7

B. Pre-Exposure Prophylaxis and Antiretroviral Therapy ...... 8

1. Pre-Exposure Prophylaxis ...... 9

2. Antiretroviral Therapy ...... 12

C. HIV/Syphilis Syndemic ...... 15

D. Contact Tracing for HIV and STI Control ...... 17

E. Network Epidemiology ...... 20

1. Limitations of Contact Tracing Network Data ...... 22

2. Combined HIV and Syphilis Networks ...... 23

3. Network Systems ...... 24

F. Network Community Structure ...... 25

G. Summary ...... 27

H. Figures ...... 29

CHAPTER THREE: DATA SOURCES AND EXTENDED METHODS ...... 31

A. Overview of Data Sources ...... 31

1. North Carolina Electronic Disease Surveillance System ...... 31

2. Blue Cross Blue Shield of North Carolina ...... 33

B. Linkage of Data Sources ...... 34

C. Aim 1 Extended Methods ...... 35

1. Study Design ...... 35

2. Study Population and Data Extraction ...... 35

3. Network Generation and Description ...... 36

4. Network Overlap and Interconnectivity ...... 38

5. Community Detection...... 39

5. Comparison to Modularity of Random Networks ...... 41

6. Heterogeneity of Syphilis/HIV Network Interconnectivity by Community ...... 41

D. Aim 2 Extended Methods ...... 42

1. Study Design ...... 42

2. Study Population and Data Extraction ...... 43

3. Prevalent and Incident Viral Suppression among Previously Diagnosed PLWH ..... 44

3. Effect of DIS Services on Incident Viral Suppression ...... 45

4. Prevalent and Incident PrEP Use among HIV-Negative Persons ...... 47

5. PrEP Indication among HIV-Negative MSM ...... 48

E. Aim 3 Extended Methods ...... 51

1. Study Design ...... 51

2. Network Generation ...... 51

3. HIV Transmission Model ...... 55

4. Intervention Scenarios ...... 57

5. Figures and Tables ...... 61

CHAPTER FOUR: NETWORK INTERCONNECTIVITY AND COMMUNITY DETECTION IN THE HIV/SYPHILIS SYNDEMIC AMONG MEN WHO HAVE SEX WITH MEN ...... 72

A. Introduction ...... 72

B. Methods ...... 73

1. Study Data ...... 73

2. Independent and Combined Contact Network Generation ...... 74

3. Network Description ...... 75

4. Network Overlap and Interconnectivity ...... 75

5. Community Detection ...... 77

6. Heterogeneity of Syphilis/HIV Interconnectivity by Network Community...... 77

C. Results ...... 78

1. Network Populations ...... 78

2. Independent and Combined Networks Characteristics ...... 79

3. Network Communities ...... 79

D. Discussion ...... 81

E. Figures and Tables ...... 85

CHAPTER FIVE: HIV VIRAL SUPPRESSION AND PRE-EXPOSURE PROPHYLAXIS IN HIV AND SYPHILIS CONTACT TRACING NETWORKS: AN ANALYSIS OF DISEASE SURVEILLANCE AND PRESCRIPTION CLAIMS DATA ...... 90

A. Introduction ...... 90

B. Methods ...... 91

1. Data Sources and Study Population ...... 91

2. Analyses ...... 93

C. Results ...... 95

1. Viral Suppression among Previously Diagnosed PLWH ...... 95

2. PrEP among HIV-Negative Persons ...... 96

D. Discussion ...... 97

E. Figures and Tables ...... 102

CHAPTER SIX: TRANSMISSION DYNAMICS AND CONTROL OF ENDEMIC INFECTION IN NETWORKS WITH STRONG COMMUNITY STRUCTURE ...... 105

A. Introduction ...... 105

B. Methods ...... 106

1. Network Generation ...... 106

2. HIV Transmission Model ...... 107

3. Intervention Scenarios ...... 109

C. Results ...... 112

D. Discussion ...... 113

E. Figures ...... 118

CHAPTER SEVEN: DISCUSSION ...... 121

A. Summary of Findings ...... 121

1. Aim 1 ...... 121

2. Aim 2 ...... 122

2. Aim 3 ...... 123

B. Study Strengths ...... 124

C. Limitations ...... 126

D. Future Directions ...... 129

1. Expanded Populations ...... 129

2. Routine Integration of PrEP Data ...... 130

3. Longitudinal Trajectories of HIV/STI Risk ...... 131

4. Sexual and Injection Drug Use Network Interconnectivity ...... 132

6. Model Impact of Interventions Deployed Through Contact Tracing Networks ...... 133

E. Public Health Implications ...... 134

F. Final Remarks...... 136

xii

APPENDIX: TRUVADA EXCLUSION CRITERIA ...... 138

REFERENCES ...... 141

xiii

LIST OF TABLES

Table 3.1. Match process for linkage between North Carolina (NC) Electronic Disease Surveillance System (EDSS) and insurance claims data...... 70

Table 3.2. Selected partition statistics considered in selection of the final resolution parameter γ for community detection ...... 71

Table 4.1. Individual characteristics by network event status within the combined HIV/syphilis network among men who have sex with men in North Carolina 2015-2017 ...... 88

Table 4.2. Network parameters for the combined HIV/syphilis and independent HIV and syphilis sexual contact networks among men who have sex with men in North Carolina 2015-2017 ...... 89

Table 5.1. Individual characteristics by HIV viral suppression at the time of an HIV/syphilis network event occurring January 2013 – June 2017 in North Carolina among men who have sex with men who were previously diagnosed with HIV ...... 103

Table 5.2. Individual characteristics by pre-exposure prophylaxis (PrEP) use at the time of an HIV/syphilis network event and in the subsequent six months among HIV-negative men who have sex with men in North Carolina, January 2013 – June 2017 ...... 104

Table A.1. International Classification of Diseases, 9th and 10th revisions, Clinical Modification (ICD-9/10-CM) codes and National Drug Code (NDC) package codes applied to identify and exclude FTC/TDF prescribed as treatment for HIV or hepatitis B virus (HBV), or as post-exposure prophylaxis (PEP) against HIV ...... 138

xiv

LIST OF FIGURES

Figure 2.1. HIV pre-exposure prophylaxis (PrEP) care continuum...... 29

Figure 2.2. HIV care continuum ...... 30

Figure 3.1. HIV and early syphilis contact tracing lookback periods ...... 61

Figure 3.2. Data flow for North Carolina (NC) Division of Public Health (DPH) – Cecil G. Sheps Center for Health Services Research at the University of North Carolina at Chapel Hill (UNC-CH) data linkage ...... 62

Figure 3.3. Network interconnectivity vs. overlap ...... 63

Figure 3.4. Example multilayer network ...... 64

Figure 3.5. Causal diagram used to select confounders of the association between contact tracing interview by disease intervention specialists (DIS) and incident viral suppression ...... 65

Figure 3.6. Algorithm for identification of individuals prescribed emtricitabine and tenofovir disoproxil fumarate (FTC/TDF) for pre-exposure prophylaxis (PrEP) against HIV ...... 66

Figure 3.7. Power law and lognormal distribution fits to empirical network data ...... 67

Figure 3.8. Adapted SI transmission model ...... 68

Figure 3.9. Beta distributions for sampling of community-level endemic HIV prevalences ...... 69

Figure 4.1. Largest connected components in the independent and combined HIV/syphilis networks ...... 85

Figure 4.2. Network communities ...... 86

Figure 4.3. Demographic characteristics of network community members ...... 87

Figure 5.1. Populations for analysis of prevalent and incident viral suppression and pre-exposure prophylaxis (PrEP) use in the HIV/syphilis contact tracing networks ...... 102

Figure 6.1. Log-transformed community-level HIV incidence rate by endemic community-level HIV prevalence and overall endemic network HIV prevalence under no intervention ...... 118

Figure 6.2. Overall HIV incidence rate by initial overall endemic HIV prevalence, intervention allocation scheme, and intervention coverage level ...... 119

Figure 6.3. Intervention efficiency by overall endemic HIV prevalence, intervention allocation scheme, and intervention coverage level ...... 120

LIST OF ABBREVIATIONS

aCID Adjusted cumulative incidence difference

ART Antiretroviral therapy

BCBS NC Blue Cross Blue Shield of North Carolina

CAI Condomless anal intercourse

CDC Centers for Disease Control and Prevention

CHAMP Convex Hull of Admissible Modularity Partitions

CIC Correlation information criterion

CTRS Counseling, Testing, and Referral Services

DAG Directed acyclic graph

DIS Disease Intervention Specialist

EHE Ending the HIV Epidemic

FTC/TDF Emtricitabine/Tenofovir disoproxil fumarate

GEE Generalized estimating equation

HBV Hepatitis B virus

HCM* Randomized hierarchical configuration model

HIV Human immunodeficiency virus

HMO Health maintenance organization

ICD-9/10-CM International Classification of Diseases, Ninth and Tenth Revisions Clinical Modification

IDU Injection drug use

IIHI Individually identifiable health information

INSTI Integration inhibitor

IQR Interquartile range

MICE Multiple imputation by chained equations

xvi

MSM Men who have sex with men

NC North Carolina

NC DPH North Carolina Division of Public Health

NC EDSS North Carolina Electronic Disease Surveillance System

NDC National Drug Code

NNRTI Non-nucleoside reverse transcription inhibitor

NRTI Nucleoside reverse transcription inhibitor

PEP Post-exposure prophylaxis

PI Protease inhibitor

PLWH People living with HIV

PnR PrEP-to-need ratio

PrEP Pre-exposure prophylaxis

PWID People who inject drugs

QIC Quasi-likelihood under independence model criterion

SBC State Bridge Counselor

SES Socioeconomic status

STD Sexually transmitted disease

STI Sexually transmitted infection

TAP Treatment as prevention

UNC-CH University of North Carolina at Chapel Hill

US United States

WHO World Health Organization

xvii

CHAPTER ONE: SPECIFIC AIMS

Despite the development of highly effective human immunodeficiency virus (HIV) prevention modalities, HIV incidence is steady or increasing in many populations.1 Available interventions, including antiretroviral therapy (ART) as treatment among PLWH and PrEP for prevention among HIV-negative persons, substantially reduce HIV transmission.2-5 However,

ART and PrEP uptake and adherence are widely suboptimal,6,7 preventing these interventions from reaching their full potentials at the population level. New strategies are needed to increase

ART and PrEP use for improved clinical and prevention outcomes in the modern epidemic era.

Sexual contact networks containing persons diagnosed with HIV and/or syphilis are efficient platforms for delivery of HIV interventions for both epidemiological and practical reasons. HIV and syphilis are syndemic, particularly among MSM, and primarily affect the same populations in the United States (US).8 HIV-uninfected HIV/syphilis network members experience heightened HIV risk,9,10 and among previously diagnosed, unsuppressed PLWH, being newly diagnosed with syphilis or named as a sexual contact of someone diagnosed with

HIV or syphilis suggests potential for ongoing HIV transmission.10 As contact tracing for HIV and syphilis is routine in the US,11 existing public health infrastructure can serve as a conduit to reach HIV-negative and HIV-positive network members with PrEP and ART, respectively.

Network-focused HIV prevention efforts will require improved understanding of the relationship between HIV and syphilis contact networks and up-to-date knowledge of PrEP and viral suppression use and coverage gaps within these networks. Identification of network communities, or clusters of individuals who are densely connected to each other and sparsely

1 connected to the remainder of the network,12 may provide a novel means of identifying populations at greatest risk of HIV acquisition and allocating interventions within the interconnected HIV/syphilis network.13,14 Network community structure impacts disease transmission across networks,13,14 and tailored interventions could harness this effect to maximize the HIV prevention benefits of limited public health resources.

We conducted a detailed investigation of HIV and syphilis contract tracing networks among MSM in NC, a state with a large HIV epidemic that serves as a useful representation of

HIV in the American South.15 We applied social network analysis, descriptive and causal inference epidemiological methods, and network transmission modeling using public health surveillance and insurance claims data to:

(1) Evaluate interconnectivity of HIV and syphilis sexual contact networks among MSM in

NC. Using contact tracing data, we generated independent and combined HIV and syphilis

networks for all MSM diagnosed with HIV or early syphilis, respectively, in North Carolina

between 2015 and 2017. We treated the independent networks as layers and identified

network communities, or clusters of densely connected nodes, in the two-layer network. We

assessed interconnectivity by comparing the mean node degree among syphilis network

members in the syphilis network alone vs. in the combined HIV/syphilis network, both overall

and by community.

Hypotheses: HIV and syphilis sexual contact networks among MSM in NC are

interconnected and network interconnectivity differs by network community.

Justification: The biological underpinnings of the HIV/syphilis syndemic have been

comprehensively described,8 but the syndemic has never been evaluated using a network

approach. Using a simple measure that we devised to assess interconnectivity of the

syphilis network with the HIV network, we aim to identify populations within the syphilis

network that may disproportionately benefit from intensified HIV prevention interventions.

(2) Assess viral suppression and PrEP use among MSM members of the HIV/syphilis

sexual contact network in NC. We used surveillance data to identify two types of qualifying

“network events” occurring 01/2013-06/2017 among MSM in NC: being diagnosed with early

syphilis, or being named as a recent sexual partner of a person diagnosed with HIV or early

syphilis. We estimated prevalent and incident viral suppression among previously diagnosed

PLWH at and after the network event, and we assessed the effect of contact tracing

services on six-month cumulative incidence of viral suppression. Using linked prescription

claims data, we also evaluated prevalent and incident PrEP use among HIV-negative

network members.

Hypotheses: Six-month cumulative incidence of viral suppression is heightened among

viremic, previously diagnosed PLWH interviewed by Disease Intervention Specialists (DIS)

during contact tracing. Prevalent and incident PrEP use are low among HIV-negative

network members, despite high frequency of PrEP indication in this population.

Justification: The NC DPH is strengthening efforts to expand PrEP use and increase viral

suppression under their emergent “Ending the HIV Epidemic” campaign. Identification of

gaps in current intervention coverage and resource needs among known HIV/syphilis

network members may assist NC DPH in development of effective intervention strategies to

reach this critical population.

(3) Model the impact of HIV prevention interventions in modular networks with

differential endemic HIV prevalence across network communities. We generated static

contact networks with community structure drawn from the combined HIV/syphilis network

generated in Aim 1 and varied endemic HIV prevalence within communities. We applied a

simple, stochastic transmission model to simulate HIV spread over these networks. We

evaluated community-level HIV incidence rates by endemic community-level HIV prevalence

and modeled the effect and efficiency of community-prevalence-based vs. randomly

3 allocated ART and PrEP interventions on overall HIV incidence rates at a range of endemic overall network HIV prevalences and intervention coverage levels.

Hypotheses: PrEP interventions may be most efficient when preferentially allocated to persons in high-prevalence vs. low-prevalence communities because each susceptible person has a higher probability of HIV exposure via membership in a discordant dyad in a high-prevalence community. ART interventions may be most efficient when preferentially allocated to persons in low-prevalence vs. high-prevalence communities because viral suppression of PLWH may largely insulate susceptible persons in a low-prevalence community from HIV exposure.

Justification: Prior studies of the impact of network community structure on infection transmission have assessed only newly introduced infections,13,14 a scenario that is not relevant to the current context of endemic HIV. We aim to elucidate the contribution of community structure to endemic infection transmission and test intervention allocation strategies that leverage heterogeneous infection distribution across network communities as a first step toward the design of HIV intervention strategies accounting for community structure.

CHAPTER TWO: BACKGROUND

A. HIV Epidemiology

Despite great advances in HIV treatment and prevention in recent years,16,17 estimated

HIV incidence is steady or increasing in many key populations.1,17 Estimated global HIV incidence has declined dramatically in the era of widespread ART and improved testing and prevention interventions.16,17 However, only very gradual declines in incidence have been estimated in most global regions since 2010, while large increases have been estimated in eastern Europe and central Asia.17 In the US, over one million people are living with HIV and an estimated 40,000 become newly infected each year.18 Estimated HIV incidence remains flat among Black MSM and is rising among Hispanic and Latino MSM, contrary to decreasing estimated incidence among all other demographic groups in the US.1,19 There is a pressing need for novel intervention strategies to build upon current successes in HIV treatment and control, and to extend these successes to the most vulnerable populations.

HIV in the Southern United States

The South is the core of the modern US HIV epidemic.20 Southern states, as well as the

District of Columbia, contain just 38% of the national population but host approximately half of all new HIV diagnoses.20 In 2018, the South experienced an HIV diagnosis rate of 15.7/100,000 population.18 This rate eclipses those observed in other regions: 10.0/100,000 in the Northeast,

9.3/100,000 in the West, and 7.2/100,000 in the Midwest.18 Although HIV incidence is heterogeneous across the region, eight of the ten states and all of the ten metropolitan statistical areas with the highest rates of new HIV diagnoses in the US are found in the South.20

As in other regions of the US, HIV in the South primarily affects disadvantaged populations.20 In

2018, African American and Hispanic and/or Latino populations accounted for 52% and 22% of new diagnoses in the South, respectively.18 Self-identified Black and Hispanic and/or Latino

MSM comprise 48% of all new diagnoses in the South, while Black women comprise 69% of new diagnoses among women.20

Access to health services, including HIV care, primary care, mental health treatment, and addiction services, is a critical component of effective HIV prevention and treatment.

However, southern states have higher average uninsured rates than states in other regions, which can impede regular care.21,22 High uninsured rates in the South are attributed to comparatively depressed average economic conditions,20 lower availability of employer-based health insurance coverage,20 and failure of many southern states to expand Medicaid following the 2010 Patient Protection and Affordable Care Act.21-23 Only seven of sixteen southern states, and the District of Columbia, have expanded Medicaid as of January 2020.23 Medicaid expansion was associated with a significant increase in HIV testing rates for individuals below the expansion income cutoff from 2010-2017, with the largest impacts among older Black and rural testers.24 Failure to expand Medicaid in many southern states may contribute to higher proportions of PLWH who are undiagnosed in the South compared to PLWH in other regions of the country.20 As a consequence of delayed diagnosis, PLWH in the South, particularly non- white PLWH, are more likely to experience HIV-related morbidity.20

High levels of HIV-related stigma, including internalized, perceived, and experienced stigma, may also contribute to heightened HIV incidence and poorer health outcomes among

PLWH in the South.25 Persistent HIV-related stigma in the South is associated with poor ART adherence, increased depressive symptoms, and increased substance use, each of which can complicate effective HIV care. Stigma may also prevent patients from accessing HIV testing and decrease HIV status disclosure to sexual partners, increasing the likelihood of transmission.

Intersectional stigma among PLWH building on race, sex, sexual orientation, gender identity,

6 rurality, substance use, and socioeconomic status can further inhibit formation of the social support systems that are critical for successful HIV care.25

HIV control in the South is complicated by the relative rurality of the epidemic. HIV is a primarily urban epidemic in the US, with newly diagnosed HIV infections in rural and suburban areas comprising only 23%-32% of all infections.26,27 The South contains a disproportionate percentage of both rural and urban populations of PLWH.26,28,29 Approximately two-thirds of

PLWH in rural regions of the US live in the South,28,30 and the HIV incidence rate within rural areas in the southeast is three times the overall US rural incidence rate.29 Barriers to successful

HIV care associated with rurality include increased distance to specialized medical facilities and personnel, a shortage of qualified medical and mental health professionals in rural regions, a lack of personal or public transportation to travel to HIV care and testing, and increased perceived stigma towards PLWH in rural communities.30-32 In particular, HIV-related stigma may be magnified in rural areas because of a perceived lack of privacy in small communities.32,33

HIV in North Carolina

NC hosts a large HIV epidemic with an epidemiology reasonably representative of trends across the American South, allowing it to serve as an effective laboratory for new intervention development. In 2018, approximately 35,000 PLWH resided in NC and 1,218 persons were newly diagnosed with HIV (13.9 diagnoses/100,000 population over 13 years old),15 a slight decline from 2017 (15.1/100,000). NC ranked 13th in rate of new HIV diagnoses and 7th in absolute new HIV diagnoses among all US states and Washington D.C in 2018.18 A majority of newly diagnosed PLWH in NC self-identify as MSM (53%) and 3% self-identify as people who inject drugs (PWID); 2% identify as both MSM and PWID.15 Approximately three-fourths of newly diagnosed PLWH in NC self-identify as Black (63%) or Hispanic/Latino (10%).15

As observed across the South, vulnerable populations are disproportionately affected by

HIV in NC. Adult and adolescent Black and Hispanic/Latino MSM were newly diagnosed at

7 estimated rates of 1,908 and 845 diagnoses/100,000 population in 2018, respectively, compared to 200 diagnoses/100,000 population among White MSM.15 Adult and adolescent

Black women experienced a new diagnosis rate of 15 diagnoses/100,000 population, compared to 1.4 diagnoses/100,000 population among White women in NC.15 HIV in NC also predominately affects young persons. The highest new diagnosis rates are observed among persons ages 20-29 and 46% of all newly diagnosed PLWH in 2018 were <30 years of age at diagnosis.15

Inconsistent access to health care and barriers to care associated with rurality and poverty also contribute to HIV transmission in NC. Like many other southern states, NC has not yet moved to expand Medicaid;25 11% of NC residents remain uninsured and may experience difficulty accessing health care, including preventative services that can reduce HIV incidence.15

NC also hosts the third highest proportion of newly diagnosed PLWH residing in rural areas, following only Mississippi and South Carolina. Nearly 20% of all newly diagnosed PLWH in NC live in rural counties, compared to 9% across the US.29 Persons living in NC census tracts with higher proportions of persons living below the federal poverty line, including many rural areas, also experienced higher rates of HIV and STI diagnoses in 2018.15

B. Pre-Exposure Prophylaxis and Antiretroviral Therapy

Available interventions, including ART as treatment for PLWH2,3 and PrEP for prevention of HIV among HIV-negative persons,4,5 substantially reduce HIV transmission. However, ART6 and PrEP7 uptake and adherence are widely suboptimal, preventing these interventions from reaching their full potentials at the population level. Failures in uptake and adherence for each intervention have been visualized as continuums through which a patient may progress.34,35

Although some patients move quickly through the HIV PrEP or care continuums, many patients stall at one stage, drop backward to an earlier stage, or cycle through continuum stages multiple times.36 The HIV PrEP continuum [Figure 2.1] includes PrEP awareness, uptake, and

8 adherence.35 These three processes are divided into nine discrete stages, including 1) identifying individuals at highest risk for HIV infection; 2) increasing HIV risk awareness; 3) enhancing PrEP awareness; 4) facilitating PrEP access; 5) linking to PrEP care; 6) prescribing

PrEP; 7) initiating PrEP; 8) adhering to PrEP; and 9) retaining patients in PrEP care.35 The HIV care continuum [Figure 2.2] spans 1) HIV infection; 2) HIV diagnosis; 3) HIV care engagement;

4) ART prescription; and 5) viral suppression.34 Brief overviews of PrEP and ART, including challenges that impede regular progression through each continuum and existing interventions to address these challenges, are presented below.

Pre-Exposure Prophylaxis

Daily PrEP for HIV prevention, also known as Truvada, was approved by the FDA in

2012.37 Truvada, oral combination emtricitabine/tenofovir disoproxil fumarate (FTC/TDF), contributes to the HIV prevention arsenal as a biomedical prevention option for HIV-negative persons who are at risk for HIV infection.37 Individual-level HIV risk reduction may be as great as 99% with full PrEP adherence.38 To date, only a handful of confirmed PrEP failures under full adherence have been reported worldwide, and several of these cases acquired drug-resistant strains of HIV not susceptible to Truvada.39-42 PrEP also offers population-level benefits. From

2012 through 2016, increased rates of PrEP uptake in US states were associated with significant declines in new HIV diagnoses.43 Similar results were observed in New South Wales,

Australia following a pilot intervention targeting MSM indicated for PrEP; rapid PrEP roll-out was associated with a 35% decline in HIV diagnoses among all MSM in New South Wales.44

Mathematical models have demonstrated that PrEP may be a cost-effective use of public health resources, in comparison to other prevention interventions, when adherence is high and individuals at greatest risk of HIV acquisition are prioritized for uptake.45-48

In addition to prevention of HIV infection, PrEP use may also increase frequency of sexually transmitted infection (STI) testing,49 reduce anxiety about HIV transmission,50-52 allow

9 for safer conception by serodiscordant couples,50,53 and decrease HIV stigma.50,52 The US

Centers for Disease Control and Prevention (CDC) recommends that persons using PrEP be tested for HIV every three months and for bacterial STIs every three to six months, with testing typically occurring at each prescribing visit.54 The frequent STI testing required for PrEP prescription maintenance may have downstream effects on STI incidence. One modeling study estimated a 42% reduction in gonorrhea incidence and 40% reduction in chlamydia incidence over ten years among US MSM with uptake of PrEP by 40% of MSM and 40% risk compensation.49 Risk compensation is an increase in transmission-associated sexual behaviors, such as declines in condom use or increases in number of condomless sexual partners, because of perceived HIV risk reduction with PrEP.49 In this model, risk compensation of nearly

100% among persons using PrEP was required to fully negate the benefits of increased STI screening frequency and early STI treatment.49

The benefits of PrEP use are partially counterweighted by negative effects at the individual and population levels. Small amounts of risk compensation following PrEP uptake have been observed, particularly among MSM who used condoms inconsistently prior to PrEP initiation.55 Significant increases in STI diagnoses among persons using PrEP have also been identified. However, it is unclear if detected increases in diagnoses can be attributed to increases in STI incidence subsequent to increased risk compensation or to increased coverage and frequency of STI testing.55 Antiretroviral resistance as a result of PrEP use during acute, undiagnosed HIV infection has been highlighted as another potential drawback to widespread

PrEP uptake but remains a chiefly theoretical concern because of strict screening for acute HIV infection during initial PrEP prescription.56-58 PrEP use may also cause serious adverse events including declines in bone mineral density and renal disfunction.59-61 Poor adherence among some recipients is currently the largest barrier to successful PrEP use following uptake, necessitating new intervention strategies to support consistent PrEP adherence.62,63

Although PrEP indications encompass a large population, uptake has been low. HIV- negative, adult MSM who report any male sex partners in the past six months and are not in a monogamous partnership with a recently tested, HIV-negative partner are partially indicated for

PrEP. MSM must also fulfil at least one of the following criteria for indication: 1) report any condomless anal intercourse (CAI) in the past six months; 2) have been diagnosed with or reported a bacterial STI in the past six months; or 3) are in an ongoing sexual relationship with an HIV-positive male partner.54 Other adult men and women who experience substantial risk of

HIV acquisition, including PWID, may also be indicated.54 Adolescents who weigh a minimum of

77 pounds and meet other indication criteria in any category are further eligible to receive

PrEP.54 PrEP uptake has been slower than anticipated; an estimated 1.2 million persons in the

US, including 492,000 MSM, 115,000 PWID, and 624,000 non-MSM/non-PWID are indicated for

PrEP,64 but only an estimated 18% of these persons were prescribed PrEP in 2018.7 In NC, just an estimated 2,500 persons were prescribed PrEP in 2018.65

PrEP is underused by some populations, including persons under age 25,66,67 Black and

Hispanic and/or Latino men and women,68,69 low-income persons,68 women,67 and persons with a high school education or less.68 Uptake has also been disproportionately low in the South and

Midwest.67,70,71 Unfortunately, many of the populations least likely to utilize PrEP experience the greatest HIV risks.71,72 This reality is demonstrated by lower PrEP-to-need ratios (PnR), or ratios of the number of persons prescribed PrEP to the number of new HIV diagnoses in a given population. Persons under age 25 (PnR = 0.9) and residents of the South (PnR = 1.0) display lower PnRs than the US population as a whole (PnR = 1.8).67 PrEP access is also highly variable by geographic location; PrEP providers in NC are largely located in urban areas and approximately two-thirds of NC counties do not have a PrEP provider.73 TelePrEP has also been successfully piloted in Iowa as an alternative option for rural clients to address this barrier to care.74 There is a need for innovative strategies to increase PrEP uptake among the most vulnerable populations.

Pilot programs offering full service PrEP care or PrEP referrals out of select county health departments in NC are underway,75-78 exploring the potential for these existing providers to reach currently underserved populations and address financial barriers to care. Although

Gilead’s copay and medication assistance programs reduce the cost of Truvada for underinsured and uninsured patients, respectively, these programs do not cover the cost of provider care and laboratory tests at prescribing and follow-up visits, leaving a coverage gap that can be addressed by county health departments with established STI testing programming.79 NC counties have taken multiple approaches to PrEP provision, including low- cost direct PrEP prescriptions and follow-up at the Orange County and Cabarrus County Health

Departments,75,76 referrals to free PrEP care out of the Mecklenburg County Health

Department,77 and referrals to internal and external PrEP clinics at the Durham County Health

Department.78 These pilot programs may serve as a foundation for wider health department

PrEP distribution programs across the state. For instance, Florida introduced a program in 2018 offering PrEP through all 67 county health departments in the state and offers sliding scale provider and laboratory costs.80 This approach ensures that PrEP is accessible to all patients, regardless of distance to a conventional PrEP provider or insurance coverage for PrEP-related costs.

Antiretroviral Therapy

Widespread ART has changed clinical outcomes for millions of PLWH by allowing treatment of HIV as a chronic condition and preventing the onset of AIDS.16,17,81 ART has radically improved since the introduction of Zidovudine (AZT) monotherapy in 1987.81 Modern antiretroviral agents belong to one of five drug classes: entry inhibitors, including CCR5 binding and fusion inhibitors, reverse transcription inhibitors, including nucleoside reverse transcription inhibitors (NRTI) and non-nucleoside reverse transcription inhibitors (NNRTI), integration inhibitors (INSTI), transcription inhibitors, and virion assembly and release inhibitors (PI).82

Triple-drug combination therapy is standard of care for all PLWH worldwide and safely suppresses HIV replication below detectable viral loads in most patients.81 Standard treatment in the US is a backbone composed of two NRTIs and a PI/NNRTI/INSTI base.82 This combination may be altered to account for viral drug resistance, interactions with therapies for co-morbidities, adverse events, or medication availability in resource-limited settings.82,83

Immediate ART initiation following diagnosis for all PLWH is recommended by the US

Department of Health and Human Services84 and the World Health Organization (WHO).83

ART also provides transmission prevention benefits; the landmark HPTN 052 trial demonstrated in 2011 that suppressive ART in PLWH prevents sexual HIV transmission to uninfected partners.85 This study, as well as the supporting PARTNER 1 and 2 and Opposites

Attract trials, identified zero transmissions from study participants with fully suppressed viral loads to an enrolled partner.85-88 This evidence sparked the re-conceptualization of ART as

“treatment as prevention” (TAP).89 Modeling studies have shown that a TAP approach to HIV control has the potential to substantially reduce population-level HIV incidence.2 Recently, this paradigm has been rephrased for the public in the “U=U” campaign, designed to communicate that HIV infection in PLWH with undetectable viral loads is untransmissible.90

ART use is increasing rapidly but falls short of coverage goals. Globally, 62% of PLWH were receiving ART and 53% were virally suppressed by the end of 2018.91 The UNAIDS 90-90-

90 goals call for 90% of all PLWH to be diagnosed, 90% of diagnosed PLWH to be receiving

ART, and 90% of those on ART to be virally suppressed. In 2018, 78% of diagnosed PLWH globally were prescribed ART and 86% of those on ART were virally suppressed, demonstrating substantial progress toward the 90-90-90 goals.91 However, some global regions have markedly lower coverage. In 2018, just 33% of PLWH in the Middle East/North Africa were receiving ART, indicating large gaps in diagnoses and care. In Eastern Europe/Central Asia and

Western/Central Africa, 38% and 52% of PLWH, respectively, were receiving ART.91

ART coverage in the US, and specifically in North Carolina, also shows room for improvement. In NC, an estimated 68% of PLWH were retained in care in 2018, a proxy for ART use, and 62% were virally suppressed.15 Viral suppression in NC is similar to that observed across the US and is higher among newly diagnosed PLWH, demonstrating improvements in care linkage and maintenance in recent years.92,93 Viral suppression in NC is lowest among patients ages 18-24 years at diagnosis, Black and Hispanic/Latino patients, and PWID.92 In

2016, 55% of PLWH ages 18-24 years at diagnosis in NC were singly virally suppressed, compared to 76% of PLWH aged ≥55 years at diagnosis.94 Disparities in durable viral suppression were even greater; just 42% of PLWH ages 18-24 years at diagnosis showed durable viral suppression, compared to 68% of PLWH aged ≥55 years.94 A significant increase in viral suppression by age at diagnosis was observed across all race/ethnicity strata.94 Self- identified Hispanic/Latino PLWH were less likely to be singly virally suppressed (60%) compared to Black, non-Hispanic (63%) and White, non-Hispanic (72%) PLWH in NC.94

PLWH engaged in care encounter four possible failure points preventing successful treatment with ART: 1) delay in initiation or failure to initiate ART; 2) discontinuation of ART; 3) poor adherence to ART; and 4) viral resistance to ART.95 Although the HIV Medication

Assistance Program and the Ryan White HIV/AIDS Program offer low-income NC residents financial support for HIV care and medication,96 many patients face additional barriers that can lead to failure to initiate, continue, or adhere to ART. Barriers to successful treatment with ART encompass structural issues, including lack of health insurance, inconsistent housing or transportation, few nearby care facilities, inconvenient clinic hours and appointment availability, and culturally-insensitive medical care, as well as intra-and inter-personal issues, including low socioeconomic status (SES), substance use, mental illness, perceived stigma, and insufficient social support systems97-99 Engagement in HIV care and adherence to ART are also fluid over time, and PLWH may require a range of interventions to support consistent treatment as personal circumstances fluctuate.36

NC offers targeted support for HIV care engagement through State Bridge Counselors

(SBC), or case managers who have also undergone field intervention training.100,101 SBC operate under the Data to Care framework, which leverages HIV surveillance data to support progression through the HIV care continuum.102 SBC receive a list of diagnosed PLWH without a recent HIV care visit documented in NC HIV surveillance data each month. They then attempt to contact these individuals, identify those who are truly not engaged in care, and support their linkage or re-linkage to care. SBC provide individualized services to as many PLWH as possible, including assistance scheduling provider visits, support with financial assistance applications and insurance plan enrollment, referrals for mental health and substance abuse counseling, and aid with housing and transportation, language barriers, and childcare.100,101

C. HIV/Syphilis Syndemic

The interacting HIV and syphilis epidemics among MSM exacerbate the impact of the other to form a syndemic.8,103 In the US, HIV,103-105 syphilis,103-105 and HIV-syphilis coinfections104,106-108 are concentrated among MSM, particularly among young MSM of color.107,108 Nationwide, approximately 53% of MSM newly diagnosed with early syphilis were coinfected with HIV in 2013.107 The high prevalence of co-infections is tied to an increase in syphilis infections among MSM over the last two decades.109 Although syphilis diagnosis rates fell sharply in the late 1990s, spurring calls for syphilis elimination, the epidemic resurged in the

2000s among MSM.109 This resurgence is attributed to reductions in perceptions of HIV risk following the introduction of ART, leading to reductions in condom use.103,109 Serosorting, or selection of sexual partners by shared perceived HIV infection status, also allowed increased opportunities for STI transmission in the absence of condom use.103,109 Syphilis rapidly shifted from primarily affecting heterosexual populations to predominately affecting MSM.110 From 2000 to 2003, the estimated percentage of syphilis diagnoses among MSM in the US increased from

7% to 62%.110 This trend was replicated across high-income nations globally, with syphilis diagnoses among MSM increasing from 27% to 55% from 2000 through 2013.111

The HIV/syphilis syndemic is highly observable in NC. Much like other states in the

American Southeast, NC experiences higher HIV and syphilis incidence than the national average.15,112 NC also showed the highest rate of primary and secondary syphilis diagnosis among MSM of all states in 2015.113 In 2018, NC reported 1,914 new early syphilis diagnoses

(18.4/100,000 population), of which approximately 55% occurred among self-identified MSM.114

Dramatic increases in detected HIV-syphilis coinfection in NC have also been observed in recent years; approximately 45% of persons diagnosed with syphilis were coinfected with HIV in

2014, compared to 6% in 2002.115 HIV/syphilis coinfections in NC are largely concentrated among MSM, and particularly among Black MSM.115 Further, both HIV and syphilis contact tracing investigations yield new HIV diagnoses. Nationally, an estimated 1.3-100 HIV cases must be interviewed to yield one new HIV case among named contacts.116 In NC, 41 syphilis cases must be interviewed to yield one new HIV case, demonstrating substantial undiagnosed

HIV prevalence within the known syphilis contact network.116

Syphilis infection increases both susceptibility to HIV among HIV-negative patients and transmissibility of HIV from coinfected patients.103,117 Syphilis is an ulcerative condition that decreases mucosal integrity in the genital tract. Syphilitic lesions recruit activated CD4+ T cells and dendritic cells. Increased concentrations of these immune cells then heighten direct infection of CD4+ T cells at the site of the lesion and amplify dendritic presentation of virions to

CD4+ T cells in regional lymph nodes for indirect infection.117,118 Syphilis infection in co-infected patients also upregulates HIV gene expression via increased binding of nuclear factor (NF)-κB to the promoter region of the HIV provirus117,118 and induces generation of pro-inflammatory cytokines that trigger proliferation of inflammatory cells, enhancing HIV replication in these cell populations.117,118

A syphilis or HIV diagnosis is a strong indicator of generalized STI risk and offers opportunities for prevention and screening for the other infection. Syphilis infection predicts HIV diagnosis among MSM in longitudinal studies9,119,120 and HIV diagnosis is approximately three times as high among MSM recently diagnosed with syphilis compared to HIV-negative MSM not recently diagnosed with syphilis, even following effective treatment of the syphilis infection.121 It is recommended that all persons diagnosed with HIV or syphilis be screened at diagnosis or entry to care for the other infection and that contacts of both HIV and early syphilis cases be screened for both infections.11 PLWH are also encouraged to test at least annually for syphilis coinfection.104

D. Contact Tracing for HIV and STI control

Contact tracing, also termed partner notification in the STI domain, has long served as an integral component of infectious disease control. Mandatory contact tracing was first implemented with the passage of the Contagious Disease Act of 1864 in Great Britain to combat the spread of venereal disease among military members.122 Contact tracing procedures have since been refined to carefully balance the privacy interests of index cases, individual health interests of named contacts, and public health interests of the state.123 Contact tracing aims to detect previously undiagnosed infection to reduce morbidity and mortality among infected persons and prevent onward transmission of infection via treatment and behavior change.124

First, persons newly diagnosed with the infection of interest are interviewed to elicit recent sexual, social, and/or injection drug use (IDU) contacts, depending on the mechanism(s) of transmission for the relevant infection. Named contacts are then informed of potential exposure to a transmissible infection and provided testing, prevention counseling, and when applicable, access to treatment.124

Contact tracing is widely utilized in both domestic and global settings. Although used in high-income nations for decades, contact tracing for HIV is rapidly gaining acceptance in low-

17 income nations as well.125 In 2016, the WHO issued a new recommendation that voluntary, assisted partner notification services should be offered to all newly diagnosed PLWH.125

Globally, sixty-seven nations practice contact tracing or partner notification and twenty-one nations require some form of mandatory notification. State law in much of the US, including NC, mandates contact tracing and infection screening for contacts of persons diagnosed with HIV and early syphilis.126-128 In NC, DIS are automatically notified when a patient is diagnosed with

HIV or early syphilis and attempt to interview all new cases to elicit information about recent contacts. DIS then attempt to reach these named contacts to provide HIV and syphilis testing and prevention counseling.129

Contact tracing efforts also offer opportunities to intervene at several stages in the HIV

PrEP and care continuums. A systematic review evaluating the effectiveness of HIV contact tracing programs found a range of one to eight sexual or IDU contacts named per index case.130

Although only approximately two-thirds of named contacts were identified with sufficient information to begin an investigation, and only 63% of those persons consented to be tested for

HIV, a mean of 20% of named contacts tested were newly diagnosed with HIV.130 Contact tracing programs for HIV and syphilis effectively identify a population with a high prevalence of undiagnosed infection that may benefit from testing. Of the estimated 1.1 million PLWH in the

US in 2015, approximately 14.5% remained undiagnosed.131 Contact tracing efforts may allow for earlier diagnosis of HIV than would have otherwise been observed in this population, triggering changes in behavior and allowing for treatment with ART to reduce the likelihood of onward transmission.

Syphilis diagnosis and contact tracing for HIV or syphilis may also serve as cues to action within the Health Belief Model and trigger acceptance of PrEP or ART uptake among previously reluctant or unaware individuals. The Health Belief Model was developed in the early

1950s by the US Public Health Service to describe why some preventative measures and screening tests were not widely accepted by the public.132 The authors hypothesized that health-

18 related behaviors were dependent on two primary factors: (1) desire to avoid illness; and (2) belief that a health action effectively prevents or treats illness. The model has been further developed into six subcategories: (1) perceived susceptibility to illness; (2) perceived severity of illness; (3) perceived benefits of a recommended health action; (4) perceived barriers to uptake or maintain the recommended health action; (5) cues to action; and (6) self-efficacy.132 A cue to action is a stimulus needed to trigger a decision-making process to accept a recommended health action and can be internal or external.132 Syphilis diagnosis or being named as a contact of a person diagnosed with HIV or syphilis could serve this function as stimuli for PrEP or ART uptake.

NC and several other jurisdictions have recently expanded HIV prevention services to include referrals for PrEP by DIS during contact tracing133-135 because the behaviors and sexual partnerships of persons diagnosed with syphilis or named as contacts of persons diagnosed with HIV or syphilis place them at substantial HIV risk.14,15 A recent study of a sexual network in

NC found that one-quarter of HIV-negative named contacts of persons diagnosed with HIV or syphilis were diagnosed with HIV within three years.136 Contact tracing also offers an opportunity to intervene on the first five stages of the HIV PrEP continuum [Figure 2.1], allowing for PrEP awareness building and support for PrEP uptake among indicated persons.35 Although several studies are currently assessing the potential for contact tracing services to identify and connect eligible PrEP users to providers, as well as interventions to support PrEP linkage and retention among this population,135,137 the size of the populations indicated for PrEP and currently using

PrEP within the HIV/syphilis contact tracing network are unknown.

Previously diagnosed PLWH identified in the HIV/syphilis contact tracing network may show continued behaviors that allow for ongoing HIV transmission in the absence of viral suppression.10 These PLWH may be prime beneficiaries of intensified care engagement interventions by DIS and SBC. One recent study reported that 20% of newly diagnosed PLWH in NC were subsequently diagnosed with an STI within three years, and that HIV-positive

19 network members often rejoined the same sexual contact network from which they had acquired infection following HIV diagnosis.136 An estimated 70% of HIV transmissions in the US138 (and

77% in NC)139 arise from contact with previously diagnosed PLWH, highlighting the importance of reaching this population with suppressive ART for both clinical and public health benefits.

However, limited resources often force prioritization of persons newly diagnosed with HIV or diagnosed with syphilis, and particularly acute/recent HIV cases and symptomatic early syphilis cases, over non-symptomatic, previously diagnosed PLWH by DIS and SBC. Presentation of viremic previously diagnosed PLWH in the network may be a missed opportunity to identify persons at high risk of HIV transmission and to assist them with care engagement or re- engagement [Figure 2.2].140

E. Network Epidemiology

Modern network analysis is rooted in sociometry, or the study of relationships within small groups of individuals. Sociometry was first applied to evaluate group structure and individuals’ positions within those structures in 1932.141 The field rapidly expanded, now encompassing computational, biological, neural, and social networks and methodologies to analyze these networks.142 In network analyses involving relationships of any form between individuals, individuals are represented as nodes and relevant connections between individuals are documented as edges.143 There are two primary types of network analysis: egocentric and sociometric. Egocentric analyses evaluate the personal networks of a sample of nodes from the population of interest, including just the nodes that are connected to the sample nodes and the edges that connect them. Sociometric analyses evaluate the complete network of nodes and edges in a set population.143

Network analyses are critical components of infectious disease control, particularly public health efforts to minimize the transmission of HIV and STIs. In the 1930s, patients at publicly funded venereal disease clinics were interviewed about sexual relationships to

20 ascertain common sources of infection, allowing for the first visualizations of sexual contact networks.144 Sexual contact network analyses in the mid-1980s then demonstrated connections between early AIDS cases among MSM in New York City, Los Angeles, and San Francisco, supporting the hypothesis that AIDS was caused by an infectious agent transmitted via sexual activity.145,146 HIV and STI sexual contact networks have since been repeatedly characterized in efforts to halt outbreaks and improve targeting of prevention and treatment interventions.142,147,148

Network features, including density, concurrency, and mixing patterns, heavily influence infection transmission within a network.142,147,148 Infection diffuses more quickly over dense networks, or networks with a greater proportion of potential connections between nodes fulfilled, than over sparse networks with fewer edges per node.142 Infection diffusion also increases with concurrency.149-151 Concurrent partnerships overlap in time, where one or more additional partnerships begin before the first terminates.142 Independent of network density, increased concurrency creates greater network connectivity, exposing larger portions of the network to infection.149-151 Sexual networks typically demonstrate fairly assortative mixing, with the exception of dissortative mixing by sex in heterosexual networks.152,153 Assortative mixing, also termed homophily, occurs when individuals partner with those who share an attribute, such as age, race, socioeconomic status, or number of partners.152,153 Homophily may allow for differential infection incidence within network sub-populations. Several studies have illustrated the role of racial homophily, alongside existing racial and ethnic disparities in HIV prevalence and treatment, in perpetuating heightened HIV incidence among Black MSM.154,155

Underlying social and cultural forces drive network features.142 Social norms impact partner selection, including the number of lifetime partnerships and length of partnerships, creating higher or lower levels of density within a network.142 These norms also control the acceptability of concurrent partnerships and the tendency towards homophilous sexual partnerships in a given population.142 Cultural factors, including the digital and physical spaces

21 where individuals meet sexual partners, may enforce or circumvent existing social norms.142

Dating websites and mobile applications can expose individuals to a wider pool of potential sexual partners than they might otherwise encounter and increase opportunities for sexual partnerships, particularly among LGBTQ individuals.156,157 Conversely, physical spaces such as bars and nightclubs tend to host a smaller, more homogenous pool of prospective partners because they cater to specific demographic and geographic populations.158

Network factors are associated with HIV acquisition and transmission risk. Following an

HIV outbreak in a population of PWID in Brooklyn, New York, several network factors were identified as significant predictors of HIV infection.159,160 One significant predictor was membership in a large, connected component, or a group of persons in which each person is connected to at least one additional person within the group. Component membership exposes each person to a potential transmission chain from any infected person in the component.159 A recent study in NC combined demographic characteristics, HIV/STI testing history, and sociosexual network data to develop a risk score algorithm for future indicators of transmission risk behavior among newly diagnosed PLWH.10 A model including network features correctly classified a greater percentage of PLWH than a model including only individual-level data.10

Finally, although concurrency does not increase an individual’s HIV/STI acquisition risk independent of number of partners, it is associated with increased acquisition risk among the concurrent partners161 and increased population-level HIV prevalence.150 Augmentation of traditional individual-based analyses with network-based factors may improve evaluation of HIV acquisition and transmission risk.

Limitations of Contact Tracing Network Data

Contact tracing data has inherent biases and limitations stemming from non-randomly missing data.151,162-165 First, the contact tracing network is identified with selection on infection or contact status.163 Thus, network members represent a distinct, high-risk population and network

22 results are implicitly non-generalizable to the full population from which cases and contacts are drawn. As only persons diagnosed with HIV or early syphilis are queried for contacts, analyses of these data are restricted to egocentric network analysis methods against the sporadically interconnected local networks of each diagnosed person. Analyses may also treat the resulting network as a sparse, non-random sample from the underlying sociometric network.

Contact tracing is unable to reach all newly diagnosed persons and their contacts. In

NC, approximately 20% of persons diagnosed with HIV or early syphilis are not reached by DIS for interview or do not agree to name partners when reached for interview.10 Further, in one meta-analysis of US contact tracing programs with reported data, 33% of all persons named as contacts of persons diagnosed with HIV or early syphilis were not reached by public health personnel.130 Analyses of network populations identified via contact tracing may thus be subject to an additional source of selection bias.163 Contact tracing data are also self-reported and can be plagued by incomplete reporting and poor recall.166-168 Persons may additionally fail to name contacts because of stigma towards MSM/PWID and PLWH, anonymous partnerships, and infidelity,162,163 creating a substantial amount of missing data that may bias network analyses by reducing observed density and connectivity.163,165

Combined HIV and Syphilis Networks

Few analyses have generated combined networks containing persons diagnosed with

HIV or syphilis and their sexual contacts.10,136,169,170 The first combined network study visualized persons diagnosed with HIV or syphilis in Colorado Springs from 1985 through 1999 and described the structure of the resulting network.169 Subsequent studies have added to our understanding of how these networks intersect. One study assessed the combination of HIV and syphilis networks in a single NC county, concluding that contemporaneous HIV and syphilis transmission persists among MSM in NC.170 This study identified two distinct, large network components. One component predominantly included young, Black MSM co-infected with HIV

23 and syphilis, while the second was largely composed of older MSM diagnosed with syphilis and their named contacts, with higher proportions of persons who reported engaging in transactional sex.170 Another recent study generated a combined HIV/syphilis network seeded by two young,

Black MSM diagnosed with acute HIV infection in NC.136 Following HIV diagnosis, 21% of

PLWH in the network were later diagnosed with an incident STI. Conversely, 25% of HIV- negative named contacts of persons diagnosed with HIV or syphilis were newly diagnosed with

HIV within three years.136 Finally, a combined HIV/syphilis network generated within one NC region demonstrated bridging between HIV contact network components by persons diagnosed with early syphilis and their named contacts.10

This literature illuminates a relationship between HIV and syphilis networks among

MSM. However, each of the prior studies evaluating combined HIV/syphilis networks have focused on small sub-populations of interest and have largely evaluated the combined network as a single network, rather than specifically interrogating how the two networks inteconnect.10,136,169,170 A statewide HIV/syphilis contact network has never been characterized, and independent HIV and syphilis networks within a single population have never been compared. There is a need for a more nuanced understanding of how HIV and syphilis networks overlap and diverge if the HIV/syphilis syndemic among MSM is to be effectively addressed.

Network Systems

Many approaches have been proposed to illuminate the complex relationships between networks.171 The simplest approach to comparison of two networks is to assess network overlap, or the presence of shared nodes and edges between two or more networks. The

Jaccard node similarity quantifies overlap between two networks as the number of nodes belonging to both independent networks, or the intersection, divided by the number in the combined network, or the union.172 The Jaccard edge similarity is the number of edges in the intersection of the two networks divided by the number in the union of the two networks. For

24 both measures, a value of 1 represents complete overlap and a value of 0 represents no overlap.172 Overlap may be measured between separate networks or between layers of a network system. Other approaches treat the networks as interacting or interdependent, but distinct, where nodes and edges are not present in multiple networks, but edges may connect nodes in network A to nodes in network B. Information, infection, or activity can pass from one network to another over these edges.171

Interconnected or interconnecting networks have been conceptualized in several ways but are broadly defined as network systems in which nodes in network A may be connected both to other nodes in network A via intra-network edges, as well as to nodes in network B via inter-network edges. Inter-network edges may occur between shared nodes in both networks or between distinct nodes across the two networks.171 One proposed measure of interconnectivity between networks assesses the change in node degree distributions from the independent networks to the combined network in coupled, disjoint network systems.173 These network systems require that each node in network A is connected to a node in network B (coupled), but that no nodes are present in both networks (disjoint). This measure distinguishes between strongly and weakly coupled network systems, where disease or information diffuses over a strongly coupled network system as if it were a single network. However, this measure requires that all networks demonstrate an internal scale, which is violated by scale-free networks, such as most sexual networks, and that the networks be disjoint.173 No measure has been previously proposed to quantify interconnectivity between two overlapping, or non-disjoint, networks with some proportion of shared nodes and edges.

F. Network Community Structure

Network community structure is the presence of clusters of nodes that are densely connected to each other and sparsely connected to the remainder of the network.12 Community structuring has been widely observed across biological, neural, computational, and social

25 networks, including epidemiological contact networks for directly-transmitted infections in human and animal populations.13,174,175 However, few studies have been conducted evaluating community structure in sexual contact networks. In one such analysis, strong network community structure characterized by homogeneity of HIV transmission risk group within communities was identified in a sexual contact network of persons diagnosed with HIV and their named contacts in Cuba.176 Community detection may be applied to identify community structure within multipartite, multiplex, directed or undirected, and weighted or unweighted networks, or most combinations of network types.12,177

One major school of community detection methodologies is modularity optimization.12

This approach identifies an approximation of the strongest partition, or division of a network into communities with the greatest edge density within communities and least edge density between communities. Modularity optimization applies an algorithm, or quality function, that counts the number of edges observed within a community, minus the number of edges that would have been expected to occur within the same group of nodes under a null model with only randomly occurring community structure.12 The quality function is calculated for potential partitions, where the partition with the highest value for the quality function displays the strongest community structure. As modularity optimization is NP-hard, or likely cannot be solved in polynomial time under most algorithms, the strongest partition identified may be one of multiple optimal partitions.12 This heuristic may be particularly relevant for identification of network communities to inform the design of network-focused transmission prevention interventions because it optimizes edge density within communities. As transmissions occur over network edges, most transmissions must occur over dense intracommunity edges and few transmissions may occur over sparse intercommunity edges in networks with strong community structure. Thus, modularity optimization allows for identification of network subgroups or clusters within which most transmissions are expected to occur (all else equal).

Community structure impacts epidemic features, including size, duration, and peak prevalence, independent of other network parameters.13,14 In a network with strong community structure, infection is more likely to spread within communities than between communities because of higher edge density within communities.13 Early outbreaks may die out within the initial community or spread serially through sparse edges between communities. Thus, outbreaks in networks with strong community structure tend to exhibit greater variance in final incidence and duration than those in networks with weak community structure, which typically experience more explosive outbreaks.13 Overall community structure also appears to play a larger role in infection diffusion over networks than the precise internal structure of communities.

This phenomenon has been attributed to the relative density of within-community edges compared to between-community edges.14

The study of infection transmission over networks with community structure was a byproduct of the study of information diffusion over networks with community structure, and thus, all analyses to date have been limited to transmission of newly introduced infections over naïve networks with community structure.14 The effect of community structure on the spread of an endemic infection, such as HIV, through a network population has not been assessed.

Further, the effect of community structure on transmission when infection is unevenly distributed across communities is unknown, as are the corresponding relationships among community infection prevalence, intervention type, intervention allocation approach, and intervention impacts. Identification of network communities may provide a novel means of identifying populations at greatest infection risk and allocating network-based interventions.

G. Summary

The ongoing burden of the HIV epidemic among MSM, particularly among young MSM of color in the American South, highlights the need for improved interventions to reduce HIV incidence. Greater understanding of interconnectivity of HIV and syphilis networks in this

27 population may be valuable in the design of effective ART and PrEP interventions that can be deployed through known sexual contact networks. Insights into how network community structure impacts infection transmission over these networks may also aid in precise delivery of these interventions for maximum prevention benefits. In addition, up-to-date knowledge of PrEP and viral suppression levels within these networks is required to assess resource needs and current intervention coverage gaps. In an era where “Ending the HIV Epidemic” (EHE) campaigns are being mounted across the US,178 findings from this dissertation may be helpful in the design and deployment of EHE-backed interventions against persistent HIV incidence in high-risk populations.

Figure 2.1 HIV pre-exposure prophylaxis (PrEP) care continuum

Adapted from Nunn et al.35

Figure 2.2 HIV care continuum

CHAPTER THREE: DATA SOURCES AND EXTENDED METHODS

A. Overview of Data Sources

North Carolina Electronic Disease Surveillance System (NC EDSS)

NC EDSS is a secure, integrated system for monitoring of all reportable communicable infections in NC, including HIV and syphilis.179 This system is used by NC DPH, all 86 local and multi-county health departments in NC, and eight regional HIV/STI offices. Individuals are assigned unique identifiers upon diagnosis or interaction with NC DPH related to one of the infections included in NC EDSS. This identifier is then used to follow the individual over time and to link reports across disease fields, enabling longitudinal evaluation at the individual level.

Reporting of HIV and early syphilis diagnoses to NC EDSS is mandated by NC law, as is reporting of all CD4 counts, viral loads, and viral sequences performed for PLWH. The rich information managed in NC EDSS allows for study of individuals over time as they experience

HIV and STI diagnoses, are named as contacts, engage in HIV care, and disengage from HIV care.179

As described in Chapter 2, NC law mandates DIS contact tracing for named contacts of

HIV and early syphilis cases. Early syphilis comprises three infection stages – primary, secondary, and early latent – approximately corresponding to the first three, six, and twelve months of infection, respectively.104 DIS attempt to interview all persons newly diagnosed with

HIV and persons diagnosed with early syphilis to identify recent sexual and IDU contacts, as appropriate. Sexual and IDU contacts are elicited for the twelve months prior to DIS interview for persons newly diagnosed with HIV127 and sexual contacts are elicited for the twelve, six, and

31 three months prior to symptom onset for persons diagnosed with early latent, secondary, and primary syphilis, respectively [Figure 3.1].128 DIS then attempt to trace and screen these named contacts, and a subsequent round of contact tracing is initiated when a contact is newly diagnosed with HIV or diagnosed with early syphilis. This process produces exponential discriminative snowball sampling of the underlying contact network.180

SBC assist persons newly diagnosed with HIV with linkage to HIV treatment to expedite

ART initiation.100,101 SBC also identify previously diagnosed PLWH who are not receiving regular

HIV care and are not virally suppressed through frequent record evaluations, provider reports, and DIS report following early syphilis diagnoses or identification of these persons as contacts of persons diagnosed with HIV or early syphilis. SBC then provide support for care linkage or re- linkage. All DIS and SBC interviews are documented in NC EDSS.100,101

Information on persons diagnosed with HIV and early syphilis and their named sexual and injection drug use contacts available in NC EDSS include:

1. HIV and STI diagnosis information, including dates, tests, results, providers, and

locations, as well as self-reported STI diagnoses; and

2. Laboratory and medical information, including HIV viral loads, CD4 counts, and viral

sequences, as well as the associated dates, providers, and locations; and

3. Demographic information, including age, sex, self-identified gender identity, and self-

identified race and ethnicity; and

4. Behavioral information, including self-identification as MSM or PWID, condom use,

and substance use behaviors; and

5. Contact information, including names of contacts, type of contact (sexual, injection

drug use, or social), contact duration (start and end dates or start date to current),

and frequency of sexual or injection-related acts; and

6. Geographic information, including residential and clinic location at diagnosis and/or

DIS/SBC interview and updated residential location over time.179

All data were originally collected for public health surveillance purposes and no further contact was made with study participants on behalf of this study. A Data Use Agreement with the Communicable Disease Branch of the NC DPH governing data collection, use, and publication was executed prior to data extraction from NC EDSS.

Blue Cross Blue Shield of North Carolina (BCBS NC)

BCBS NC data contain insurance claims from insured groups, administrative services only groups, and individual market and Affordable Care Act Exchange groups.181 These data cover approximately 3.6 million NC residents:182 35% of the NC population,183,184 50% of the non-Medicare/Medicaid insured population,182,184,185 and 60% of the privately insured population.186 BCBS NC data are housed at the Cecil G. Sheps Center for Health Services

Research at the University of North Carolina at Chapel Hill (UNC-CH). Available data include:

1. Insurance claims information, including provider visits, inpatient and outpatient care,

prescription medications, charge amounts, dates of service, diagnoses, and

procedures; and

2. Provider information, including specialty and location; and

3. Member information, including age, self-identified gender, and county/ZIP code of

residence; and

4. Plan information, including dates of enrollment in BCBS NC, type of insurance plan,

pharmacy coverage, and relationship to BCBS NC subscriber.181

These data do not include State Employee’s Health Plan subscribers, which is also administered by BCBS NC. A Data Use Agreement with BCBS NC governing data access, use, and publication was executed prior to data linkage and analysis.

B. Linkage of Data Sources

A subset of persons present in NC EDSS (described below in Aim 2 Extended Methods) were linked to BCBS NC insurance claims data for analyses in Aim 2 by an honest broker at the

Cecil G. Sheps Center for Health Services Research at UNC-CH. Individually identifiable health information (IIHI) and dummy identifiers, stripped of all health-related information, were provided by both NC EDSS and BCBS NC for data linkage [Figure 3.2].

Linkages occurred on the Sheps Center’s secure server using an iterative set of three matching algorithms [Table 3.1]. Before matching, names were stripped of all blanks, periods, dashes, commas, apostrophes, and suffixes. A strict deterministic match on first and last name, date of birth, self-identified gender, and ZIP code was applied as the first pass, followed by subsequent passes using subsets of these fields. Matched persons were removed after each pass. Persons in NC EDSS who matched to multiple BCBS NC IDs in any of the three matching passes (multiple matches) were cross-checked against a table of persons known to have multiple BCBS NC IDs and data were condensed across all matched IDs. Multiple matches not previously known to Sheps Center staff were manually resolved and samplings of high confidence matches were reviewed for verification.

Access to IIHI data was restricted to the honest broker who performed matching and data linkage work. IIHI were stored in an encrypted form while at rest and destroyed as soon as linkage was completed. Dummy identifiers were then used to merge health-related information from both sources. The linked dataset containing BCBS NC and NC EDSS data was stored on

Sheps Center secure servers and accessed for analyses via password-protected VPN.

C. Aim 1 Extended Methods

Study Design

Using contact tracing data, we generated independent and combined HIV and syphilis networks for all MSM diagnosed with HIV or early syphilis, respectively, in North Carolina between 2015 and 2017. We treated the independent networks as layers and identified network communities, or groups of densely connected people, in the two-layer network. We assessed interconnectivity of the syphilis network with the HIV network by comparing the mean node degree among syphilis network members in the syphilis network alone vs. the combined

HIV/syphilis network, both overall and by community. All analyses were conducted in R v.

3.5.0187 and MATLAB 9.3 (The MathWorks Inc., Natick, MA).

Study Population and Data Extraction

We defined a network event as receiving an HIV or early syphilis diagnosis between

2015 and 2017, or being named as a sexual contact by someone receiving one of these new diagnoses. Each network event was assigned a network event status: HIV diagnosis, early syphilis diagnosis, HIV contact, or early syphilis contact.

All new HIV and early syphilis diagnoses among persons diagnosed in NC between

2015 and 2017 who self-identified as cisgender MSM aged ≥ 18 years were included as HIV and syphilis diagnosis events, respectively. Adult, cisgender men named as sexual contacts at these diagnosis events were included as experiencing HIV and early syphilis contact events, respectively. Some persons experienced multiple or repeat network events during the study period; all such events were included for network generation. We excluded diagnosis events among newly diagnosed persons not reached for – or not willing to provide information about recent contacts in – DIS interview, as the absence of information about contacts precludes any contribution to network construction. Newly diagnosed persons who consented to interview but

35 named no sexual contacts during the relevant time frame were included as isolates. Events among self-identified transgender persons were excluded because gender identity was not collected for persons diagnosed with early syphilis during the study time frame, preventing analyses specific to this vulnerable population.

Data extracted from NC EDSS for diagnosis and contact events included dummy identifiers, demographic information, geographic information, diagnosis dates, locations, and statuses through December 31, 2018, laboratory confirmed and self-reported STI history through December 31, 2018, contact tracing dates and statuses through December 31, 2018,

HIV viral loads through December 31, 2018 for all PLWH, and self-reported HIV transmission risk factors. Data were extracted using SAS 9.4 (SAS Institute Inc., Cary, NC) from a copy of

NC EDSS data current through December 31, 2018.

Network Generation and Description

In social network analysis, persons are represented as nodes and relevant connections between persons are included as edges.143 We generated an independent HIV network comprising persons diagnosed with HIV and their contacts, with sexual relationships reported during HIV contact tracing between these persons as edges. We generated an analogous independent syphilis network comprising persons diagnosed with syphilis and their contacts, linked by edges denoting sexual relationships reported during syphilis contact tracing. Finally, we generated a combined HIV/syphilis network comprising all persons and edges from both independent networks.

We calculated descriptive statistics for the combined network population and by network event status, summarizing the number and percentage of network events according to age and self-identified race/ethnicity. We also described the diagnostic site for persons diagnosed with early syphilis, categorizing sites as 1) primary care physician’s office or health maintenance organization (HMO) provider; 2) STD clinic, including STD clinics housed within city and county

36 health departments; 3) other clinics; 4) emergency department; and 5) other (inpatient care, correctional facility, blood bank, etc.).

Next, we calculated standard network parameters for the combined and independent networks. In network analysis, a component is a group of two or more nodes, where each node is connected by an edge to at least one other node in the component. An isolate is a node that is not connected by an edge to any other node.143 We identified isolates, calculated the largest component size for each network, and characterized the corresponding component size distributions as the number and percentage of components in each network that were dyads (2 nodes), and small (3-9 nodes), medium (10-29 nodes), and large components (≥ 30 nodes).

One network characteristic of particular interest was the size of the largest connected component in each network. The largest connected component is the largest group of persons within a network within which each person is connected to at least one other person within the group. An increase in the percentage of persons included in the largest connected component from the independent networks to the combined network demonstrates bridging by nodes and edges in each of the independent networks between disconnected components in the other network.

We calculated node degree as the number of nodes to whom each node was connected by an edge143 and estimated the mean and 95% confidence intervals in each network population. Homophily (assortative mixing) by age and self-identified race/ethnicity for pairs of diagnosis nodes and their sexual contacts were characterized for each network. Homophily by age was calculated as the median absolute difference in diagnosis-contact ages at the time of the network event. Homophily by race/ethnicity was evaluated via mean homophilic preference,188 or the proportion of edges within diagnosis-contact pairs with concordant self- identified race/ethnicity classifications (Black, non-Hispanic; White, non-Hispanic; Hispanic;

Asian, non-Hispanic; or other/multiracial). Pairs in which either partner was missing self- identified race/ethnicity were excluded from homophily calculations.

Network Overlap and Interconnectivity

We used a traditional measure of network overlap, the Jaccard node similarity, to quantify direct overlap between the independent HIV and syphilis networks. We calculated this measure as the number of nodes belonging to both independent networks (intersection) divided by the number in the combined network (union).172 We calculated the Jaccard edge similarity analogously as the number of edges in the intersection divided by the number in the union. For both measures, a value of 1 represents complete overlap and a value of 0 represents no overlap.

The Jaccard node and edge similarities do not fully describe the potential for HIV to propagate across the syphilis network, as opportunities for transmission of HIV through the syphilis network are dependent not only on the amount, but also the distribution, of overlap

[Figure 3.3]. If the HIV network overlaps with only a distinct segment of the syphilis network, limited numbers of persons within the syphilis network are directly connected to the HIV network. However, if the HIV and syphilis networks are more widely interwoven, a larger proportion of syphilis network members may be directly connected to the HIV network. To reflect this phenomenon, we devised a simple measure of interconnectivity that accounts for both the proportion of overlapping nodes and the distribution of this overlap.

More specifically, we conceptualized interconnectivity as the relative percentage increase in mean node degree (mean number of persons to whom each person is connected by an edge) following combination of an index network with a second network. To estimate interconnectivity of the syphilis network with the HIV network, we thus compared the mean node degree among syphilis network members in the independent syphilis vs. combined HIV/syphilis networks. If there were no interconnectivity between the syphilis network and the HIV network, we would expect to observe no increase in mean node degree; higher relative increases represent greater interconnectivity. This measure is similar to the approach proposed by

Dickison et al.173 for evaluation of coupled, disjoint network systems, but our simpler measure allows for nodes and edges to be shared between networks. Further, our measure does not require that networks’ degree distributions demonstrate an internal scale, a requirement is violated by scale-free networks, including many sexual contact networks.

Community Detection

Community detection identifies clusters of nodes within a network that are more densely connected to each other than to the remainder of the network.12 We employed community detection to identify clusters of persons within the combined HIV/syphilis network among whom the syphilis network was highly interconnected with the HIV network, suggesting greater potential for exogenous HIV entry into the syphilis network. We applied a generalized Louvain method for community detection, including the independent HIV and syphilis networks as distinct network layers.177 This method maximizes a quality function Q, also termed the modularity of the network, to identify the set of community assignments that produces the highest concentration of edges within network communities. Larger values of Q (maximum possible value = 1) indicate that more edges occur within communities (rather than between communities), while Q = 0 indicates that edge occurrence is uncorrelated with community labels. Community detection analyses were restricted to the largest connected component in the combined network to ensure sufficient network density.

The generalized Louvain method for community detection maximizes a quality function

Q that sums intra-community edge weights within layers, less the edge weight expected under a null model where nodes are linked randomly within layers.177 Nodes in each layer are connected to their corresponding node in each other layer in which they are present under an interlayer coupling parameter ꙍ = 1 [Figure 3.4]. Interlayer coupling with ꙍ = 1 ensures that each node is identified in only one community across all layers. However, not all nodes are required to be present in all layers. The multilayer quality function can be expressed as:

1 푘푖푠푘푗푠 푄 = ∑ {(퐴푖푗푠 − 훾푠 ) 훿푠푟 + 훿푖푗퐶푗푠푟} 훿(푔푖푠, 푔푗푟) 2휇 2푚푠 푖푗푠푟

This function sums the total edge weight observed, less the edges expected within layers in a null model under the set resolution parameter, within communities. The sum is then divided by the total edge weight observed within the network. Adjacencies 퐴푖푗푠 denote edges within a layer s between nodes i and j, and inter-layer couplings 퐶푗푠푟 denote edges between node j in layer s to itself in layer r. The total edge strength for each node j within a single layer is

푘푗푠 = ∑푖 퐴푖푗푠 and the total edge strength across layers is 푐푗푠 = ∑푟 퐶푗푠푟. The multilayer edge strength for node j in layer s is defined as κ푗푠 = 푘푗푠 + 푐푗푠. The total weight of all edges within a layer is 2푚푠 and the total weight of all edges in the network is 2μ = ∑jr κjr. The parameters

훿푠푟, 훿푖푗, and 훿(푔푖푠, 푔푗푟) are Kronecker deltas: 훿푠푟 = 1 if the layers s and r are the same layer and

0 if different layers, 훿푖푗 = 1 if nodes i and j are the same node in different layers and 0 if different nodes, and 훿(푔푖푠, 푔푗푟) = 1 if the community assignments 푔푖푠 and 푔푗푟 of nodes i and j in layers s and r are the same and 0 if they differ.177 The resolution parameter 훾 is adjusted to modify community size.

We applied the Convex Hull of Admissible Modularity Partitions (CHAMP) algorithm to our network189 to identify domains of potentially optimal resolution parameters within the range

[0, 10] for both 훾 and ꙍ. Pairs of (훾, ꙍ) parameter selections within any single domain produce community assignments with similar numbers of communities and modularity (Q). After setting

ꙍ = 1, CHAMP identified a domain over 훾 = [0.2, 1.2] within which all possible parameter assignments should produce similar community assignments.

We then applied a modified version of the iterated generalized Louvain algorithm for community detection within categorical, multilayer networks available at http://netwiki.amath.unc.edu/GenLouvain190 and discussed above, adapting the script to allow each network layer to contain only the nodes present in that layer. Final communities for each

40 tested 훾 were identified as the community assignments for each node that maximized the quality function Q in the highest modularity partition of 1,000 non-deterministic runs.177 We compared identified communities across a subset of the 훾 = [0.2, 1.2] domain with ꙍ = 1 and selected 훾 =

0.25 as the final parameter assignment because it minimized the number of individuals in small communities excluded from interconnectivity analyses. 훾 = 0.2-0.3 produced very similar community assignments and required exclusion of the same number of nodes due to membership in communities containing <30 members of the syphilis network [Table 3.2].

Network community membership was visualized using an adaptation of the code available at http://netwiki.amath.unc.edu/VisComms/VisComms.191

Comparison to Modularity of Random Networks

Large, sparse networks naturally tend to have high modularity, so for comparison we assessed modularity of the largest connected component of the same size and density of random multilayer networks generated from an Erdos-Renyi model. These networks showed modularity (Q) = 0.954 (± SD: 0.0032). Although the absolute difference in modularity is slight, given the small available range between the modularity expected by chance and the maximum value of Q = 1, our empirical network does display significantly higher modularity.

Heterogeneity of Syphilis/HIV Network Interconnectivity by Community

We repeated our analysis of syphilis network interconnectivity with the HIV network in each community to assess heterogeneity of interconnectivity. As the distribution of node degree increases was non-normal within communities, we restricted this analysis to network communities containing ≥30 members of the syphilis network to allow for reliable estimates of the mean percentage increase in node degree under the central limit theorem.192

To characterize potentially distinguishing features of highly interconnected communities that could aid in intervention development and deployment, we applied simple linear regression

41 models to assess the unadjusted associations between community membership characteristics and community-level interconnectivity (relative percentage increase in node degree following combination of the syphilis network with the HIV network). Independent variables included median age of community members at first network event; proportion of community members who self-identified as Black, non-Hispanic; proportion of persons with syphilis diagnoses who received their diagnoses at STD clinics in each community; and proportion of community members who were members of both the HIV and syphilis networks. Each independent variable was assessed in a distinct regression model. With the exception of syphilis diagnosis site, we included all members of each community in the calculation of these community descriptors because understanding communities as a whole may allow for insights into the forces driving overall network structure. Associations of interconnectivity with demographic descriptors (age, race/ethnicity) and syphilis diagnostic site may aid in identifying indicators of heightened interconnectivity that could guide resource allocation within the syphilis network without requiring complex network analyses in real time.

D. Aim 2 Extended Methods

Study Design

We used surveillance data to identify two types of qualifying “network events” occurring between January 2013 and June 2017 among MSM in NC: being diagnosed with early syphilis, or being named as a recent sexual partner of a person diagnosed with HIV or early syphilis. We estimated prevalent and incident viral suppression among persons with a previous HIV diagnosis at the network event, and we assessed the effect of contact tracing services on six- month cumulative incidence of viral suppression. Using linked prescription claims data, we also evaluated prevalent and incident PrEP use among HIV-negative network members. PrEP- related data were cleaned and analyzed in SAS 9.4 (SAS Institute Inc., Cary, NC) using a

42 remote server submit. Viral suppression analyses with NC EDSS data alone were performed in

R v. 3.5.0.187

Study Population and Data Extraction

≥18 years old on the date of the network event; 2) resided in NC at the time of the event; 3) did not self-identify as transgender; and 4) either self-identified as MSM or identified as male and were named as a sexual contact by a self-identifying male partner. Self-identified transgender individuals were excluded because gender identity was unavailable for persons diagnosed with syphilis during the study period, preventing subgroup analyses of transgender women.

We then identified sub-populations for viral suppression and PrEP use analyses. Viral suppression analyses were conducted only among persons with an HIV diagnosis ≥30 days before their network event. Persons diagnosed with HIV at the network event or <30 days prior were excluded because they could not yet have achieved suppression through ART. Incident viral suppression analyses included only previously HIV-diagnosed persons without prevalent viral suppression at the time of their network event.

PrEP use analyses were conducted only among persons with confirmed HIV-negative or unknown HIV status at the time of the network event and included only those persons in NC

EDSS who matched to insurance claims data as described above in “Linkage of Data Sources” and were covered by plans with prescription coverage. We further restricted populations for

43 prevalent PrEP use analyses to those with continuous coverage over at least six months before the network event. Incident PrEP use analyses were restricted to those with continuous coverage over the sixty days prior to the network event and over the six months following the network event.

As most persons without a prior HIV diagnosis are tested for HIV at the time of a network event, persons without an HIV diagnosis on record in NC EDSS before or in the 14 days following the event were presumed HIV-negative. The impact of this assumption was interrogated in a sensitivity analysis limiting PrEP analyses to events among persons who 1) were reached by DIS at the time of the network event, consented to HIV testing, and had a negative result; OR 2) had an unrelated, documented HIV-negative test result in NC EDSS in the three months prior to the network event or ever following the network event, through

December 2019.

Prevalent and Incident Viral Suppression among Previously Diagnosed PLWH

HIV viral load is routinely measured during HIV clinical care visits to evaluate disease progression and treatment success. In NC, HIV viral load measures generated by all HIV care providers are mandatorily reported to NC DPH.127 We identified all viral load measures during the year prior to each network event among previously diagnosed PLWH to assess prevalent viral suppression. Using the measure collected closest to the event, we classified persons with viral load <200 copies/mL as virally suppressed according to the CDC definition for surveillance purposes.193 Persons with no viral load measures available during the year prior to the network event (20.6% of previously diagnosed PLWH without viral load <200 copies/mL) were assumed to be out of care and classified as not virally suppressed. Virally non-suppressed PLWH were subsequently assessed for six-month cumulative incidence of viral suppression, defined as ≥1 viral load measure <200 copies/mL in the six months following the event. In an ancillary

44 analysis, we assessed the twelve-month cumulative incidence of viral suppression, defined analogously.

Effect of DIS Services on Incident Viral Suppression

To assess the impact of current DIS services on viral load outcomes among previously diagnosed PLWH in our population, we estimated the effect of being reached by DIS in association with the network event on the six-month cumulative incidence of viral suppression thereafter. Previously diagnosed PLWH were classified as reached by DIS services if DIS interacted with them in any capacity, including contact tracing interviews for those diagnosed with early syphilis, re-interviews among those named as contacts of persons diagnosed with

HIV, and contact interviews for syphilis screening among those named as contacts of persons diagnosed with early syphilis. We applied linear risk regression to estimate adjusted cumulative incidence differences (aCID) comparing incident viral suppression among those reached vs. not reached by DIS.

A generalized estimating equations (GEE) model with an autoregressive correlation structure was used to account for the presence of multiple observations for some persons.194

Under this condition, standard regression methods would underestimate standard errors because of correlation between repeated observations from a single individual. Repeated observations in our study population are valuable and cannot be discarded because they capture the reality of HIV and early syphilis contact tracing programs in NC – many individuals do experience multiple network events in close succession. GEE addresses this condition by down weighting within-subject observations according to the degree of correlation between residuals and the number of correlated observations for parameter estimation.194 First, a standard linear regression model assuming that all observations are independent is applied.

Residuals are calculated from this model and are used to estimate a working correlation matrix.

Regression coefficients are then refit, correcting for any correlation between observations within

1 individuals by weighting each observation ( ), where x is the number of within-subject 1+(푥−1)∗푅 observations and R is the correlation between residuals. This process is repeated until estimates converge.194

The Quasi-likelihood under Independence model Criterion (QIC) is applied with GEE for model selection.195 The QIC sums the log quasi-likelihood for all observations under the independence model, or the inaccuracy of the fitted model to the observed data, plus a covariance penalty term.195 Confounders were singly assessed to identify optimal functional form via estimation of the QIC of a model with the confounder as the independent variable and the outcome variable as the dependent variable. The QIC was estimated for such models under all relevant correlation structures and the functional form that minimized QIC across all correlation structures was selected for each confounder.195

Plausible confounders of the possible effect of DIS services on reported incident viral suppression were identified following a comprehensive literature review and assessed using a directed acyclic graph (DAG)196 [Figure 3.5]. Network event status, age, and time since HIV diagnosis may affect prioritization of DIS services, while the event date may affect the size of the DIS caseload at that time. SES and perceptions of public health/willingness to engage with health services may create uncontrolled confounding that cannot be addressed with the available data. Confounders were then evaluated to identify optimal functional form, resulting in nominal coding of event year (2013, 2014, 2015, 2016, 2017), network event status (syphilis diagnosis, HIV contact, syphilis contact), and age at network event (18-24, 25-34, 35-44, ≥45 years), and restricted cubic splines to represent years since HIV diagnosis. This adjustment set was included in the linear risk GEE model to estimate an aCID with robust standard errors using the sandwich estimator. GEE models were repeated with twelve-month cumulative incidence of viral suppression as the outcome.

We evaluated several potential correlation structures because the pattern of correlation between multiple observations from a single individual in our study population is not known. If all within-subject observations are equally correlated regardless of other factors, an exchangeable correlation structure is appropriate. However, if correlation of within-subject observations is influenced by the number of network events between observations, an autoregressive correlation structure should more closely approximate the true covariance matrix. The

Correlation Information Criterion (CIC) is the ratio of the robust covariance against the model- based covariance, where a lower value indicates that the estimated covariance matrix more closely approximates the true covariance matrix.197 We selected the most appropriate correlation structure by comparing models with exchangeable and autoregressive working correlation structures for within-subject observations to an independent correlation structure as a reference using the CIC.

We fit an identity-binomial (linear risk) regression model, where i is the subject ID, o is the observation number for a network event, DISio represents DIS services at that event, and

X1, …, Xn represent adjustment variables. Adjustment variable values may or may not differ by observation for a single individual. GEE models were fit and evaluated using the geepack198 R package as: E(Yio) = B0 + B1*DISio + B2X1io + … + BnXnio.

Prevalent and Incident PrEP Use among HIV-Negative Persons

PrEP use, as measured by FTC/TDF prescription claims, was evaluated among confirmed or presumptive HIV-negative persons using an adaptation of a previously described algorithm for PrEP utilization.199 Persons prescribed FTC/TDF, also known as Truvada, for PrEP were identified via exclusion of FTC/TDF prescriptions for HIV or Hepatitis B virus (HBV) treatment or as HIV post-exposure prophylaxis (PEP) [Figure 3.6]. There is substantial overlap between PEP and PrEP users; even if a person had an excluded FTC/TDF prescription claim due to indicators of PEP usage, an earlier or later FTC/TDF prescription claim was allowed to

47 be classified as a PrEP prescription. PrEP prescriptions and exclusions199 due to prior or concomitant prescriptions and/or diagnoses were determined from International Classification of

Diseases, Ninth and Tenth Revisions Clinical Modification (ICD-9/10-CM) codes and National

Drug Codes (NDC) in insurance claims data [Appendix]. Claims records were queried from

January 1, 2010 through 30 days after FTC/TDF prescription as an “all-available” lookback period to identify exclusions. We manually reviewed all records for each event with an FTC/TDF claim excluded due to prior HIV diagnosis, identifying and correcting erroneous ICD-9/10-CM codes for prior HIV diagnosis in two persons who tested HIV-negative at the network event and indicated to DIS that they were currently using PrEP.

We assessed prevalent PrEP use, defined as ≥1 PrEP prescription claim in the six months prior to the network event date, among those with continuous pharmacy insurance coverage during that period. We then assessed incident PrEP use following a network event among all persons with pharmacy insurance coverage over the sixty days prior to the event date and continuously for six months after the event date, excluding persons with documented PrEP use prior to the network event. Six-month cumulative incidence of PrEP use was defined as ≥1

PrEP prescription claim in the six months following the network event. We described prevalent and incident PrEP use by age, self-identified race/ethnicity, HIV/syphilis network event status

(syphilis diagnosis, HIV contact, syphilis contact), year of network event, insurance subscriber

(self, parent, other), and apparent PrEP indication (defined below).

PrEP Indication among HIV-Negative MSM

The CDC recommended indications for PrEP use by MSM include 1) adult man; 2) without acute or established HIV infection; 3) any male sex partners in the past six months; 4) not in a mutually monogamous partnership with a recently tested, HIV negative man; and either

5a) any anal sex without condoms (receptive or insertive) in the past six months; or 5b) a bacterial STI diagnosed or reported in the past six months.54,64 We applied a surveillance based

48 adaptation of the CDC recommended indications among MSM to describe PrEP use by indication for PrEP.

PrEP indication at the time of the network event was defined as: 1) had ≥1 male sex partner in the prior six months documented in NC EDSS (named by a man diagnosed with HIV or syphilis as a sexual contact or named any male sexual contacts during the six months prior to the event date); AND either 2) reported any CAI (reported recent condom use frequency as anything other than “Always” in DIS interview); OR 3) became diagnosed with or reported a bacterial STI, including syphilis, gonorrhea, chlamydia, or chancroid, in the prior six months

(self-reported a bacterial STI diagnosis, had a bacterial STI diagnosis recorded in NC EDSS, or became diagnosed with early syphilis at the relevant network event). Missing values for condom use frequency were treated as lack of consistent condom use and a proxy for CAI. We evaluated the impact of this “worst-case” assumption in a sensitivity analysis by imputing missing condom use values and identifying the PrEP-indicated population after imputation.

As condom use is binary in our analyses, we used multiple imputation by chained equations (MICE) to impute unobserved condom use values conditional on observed demographic, clinical, and risk behavior characteristics.200 Under MICE, a series of regression models are applied to model each variable with missing data conditional upon all other variables available. Each variable is modeled according to its distribution; binary variables such as condom use as defined here are modeled using logistic regression.200 First, observed values for condom use were regressed on all other variables. Missing values were replaced by predictions, or imputations, from this regression model. We repeated this process until the model coefficients converged and retained the final imputations as the variable values. We then repeated the imputation procedure to create 50 imputed datasets.200 All imputation was performed using the mice R package.201

Imputation was restricted to the first two events for each unique person; no persons in our dataset reported always using condoms at the third event, preventing further imputation.

Data were transposed from long to wide to create a single record for each person across both events. We then applied MICE to impute condom use conditional on all available variables, including: age, self-reported race/ethnicity, network event status, county, ZIP code, event date,

DIS services status, prior documented HIV negative status, STD laboratory confirmed diagnoses in the six months prior to the event, self-reported STD diagnoses in the six months prior to the event, documented sexual contact in the six months prior to the event, documented sexual contact with a diagnosed PLWH at the event, and naming case condom use status for contacts. If a person had at least two eligible events, data from the other event, including condom use if non-missing, were included in the imputation model. Condom use information was missing for 2% of persons diagnosed with syphilis, 69% of persons named as HIV contacts, and 91% of persons named as syphilis contacts.

We treated PrEP indication as a stratification variable rather than an inclusion criterion for PrEP analyses due to challenges in reliably assessing PrEP indication with surveillance data. In addition to challenges related to missing data on condom use frequency, the surveillance-based adaptation likely underestimates PrEP indication because we cannot reliably identify the number of male sex partners in the six months prior for persons named as contacts and instead must rely on information provided by naming cases or by these same persons at other network events within the time frame of interest. In addition, sexual contacts are only elicited for the three months prior to symptom onset among persons diagnosed with primary syphilis, preventing complete assessment of partnerships in the six months prior within this population.

E. Aim 3 Extended Methods

Study Design

We generated static contact networks with community structure and other critical network features drawn from the Aim 1 sexual network and varied initial HIV prevalence within communities. Then, we applied a simplified, stochastic transmission model to simulate HIV spread through these networks. We evaluated HIV incidence in each community by endemic community HIV prevalence and modeled the efficiency of community-prevalence-based vs. randomly allocated treatment and prevention interventions on overall HIV incidence within the known network over a range of initial overall network HIV prevalences and intervention coverage levels.

Network Generation

Following other studies of network structure,13,14 we used an empirical network as a starting point and test case for simulations and comparisons. Network parameters were drawn from the largest connected component of the HIV and syphilis sexual contact tracing network among MSM in NC from 2015 through 2017. This network demonstrates strong modularity, or community structure, and was comprehensively characterized in Chapter 4. Empirical network parameters drawn from this network included the network size (N), modularity (Q), number of communities (K), and overall node degree, intracommunity node degree, and community size distributions.

The network modularity quality function Q evaluates edge density within and between communities. If all edges fall randomly between nodes without regard to community, Q = 0, while if all edges fall only within communities, Q = 1. We used a simplification of the generalized-Louvain algorithm175,190 discussed in the Aim 1 Expanded Methods for a single-layer network to assess the number of communities (K), network community size distribution, and

51 modularity (Q) in our empirical network, treating the combined HIV and syphilis network as a single network layer. We also later applied this function to identify network communities within each generated network. With only a single layer, the expression discussed above for Aim 1 analyses simplifies to the Potts generalization of Newman-Girvan modularity:

1 kikj Q = ∑ {(A − γ ) } δ(g , g ) 2μ ij 2μ i j ij

This function sums the total edge weight observed, less the edges expected with the set resolution parameter 훾 under a null model assuming no community structure, within network communities. The sum is then divided by the total edge strength 2μ observed within the network. Adjacency Aij denotes edges between nodes i and j. The total edge strength for each node j is kj = ∑i Aij and the total edge strength in the network, 2μ = ∑j k,. The Kronecker delta

δ(gi, gj) = 1 if the community assignments gi and gj of nodes i and j are the same and 0 if they differ.175 Modularity (Q) was estimated with 훾 = 0.25175 for concordance with community detection analyses of the empirical network conducted in Aim 1 using a publicly available script hosted at http://netwiki.amath.unc.edu/GenLouvain.190

Large, sparse networks like the empirical network assessed here tend to display strong modularity. To evaluate the departure of this network’s structure from expected modularity, we compared the observed modularity to modularity of largest connected components of the same size and density from random Erdos-Renyi networks. These random networks showed modularity (Q) = 0.914 (± SD: 0.0051).

To determine the appropriate distributions for additional network simulation input parameters, we fit the overall node degree, intracommunity node degree, and community size distributions from the empirical network against common distributions for network data, including uniform, Poisson, geometric, lognormal, and power law distributions. Lognormal and power law distributions provided appropriate fits to the data under the Kolmogorov Smirnoff statistic.202 All

52 three distributions were then best fit by the lognormal distribution under the Vuong test [Figure

3.8].203 Parameters (μ and σ) from these distributions were identified as input parameters for network generation.

We used these input parameters with the algorithm presented in Sah et al204 to generate

100 static networks as modular random graphs using the randomized hierarchical configuration model (HCM*), which preserves the empirically observed community size distribution and distributions of node degrees within and between communities, but randomizes specific inter and intracommunity edges.14,204 The algorithm proceeds as follows:

1. Assign N nodes to K communities. Community sizes sk are sampled for each of the K

communities from the lognormal community size distribution such that ∑sk = N. N nodes

are then arbitrarily assigned to communities under the sampled community size

sequence.

2. Assign overall degrees, d(vi), to each node vi. A degree sequence is sampled from the

lognormal overall degree distribution. This sequence must verify the Handshake

Theorem, or the requirement that the sum of degrees is even, and the Erdős-Gallai

Criterion, which ensures that the sequence is able to be graphed. Under the Erdős-

Gallai criterion, the degree sequence is sorted from highest to lowest and then queried

to ensure that there are sufficient degrees remaining among the nodes with a lower

degree to satisfy the degree requirements of any chosen node. No node may be

assigned a degree = 0.

3. Assign an intracommunity degree (number of edges within the assigned

community), dw (vi) to each node vi. The intracommunity degree sequence is sampled

from the lognormal intracommunity degree distribution. Intracommunity degrees must

satisfy four conditions.

1. After sorting the degree sequence and the intracommunity degree sequence

from highest to lowest independently, d(vi) ≥ dw(vi) for all vi.

2. The intracommunity degree sequence must be realizable for each community Ck

under the Handshake Theorem and the Erdos-Gallai criterion.

3. The intracommunity degree sequence attains the expected mean, dw within a set

tolerance εdw = 0.3.

4. Each community achieves the expected mean intracommunity degree, dw, within

a set tolerance εdw = 0.3. This is achieved as long as max[{dw(vi)}vi∈G] ≤ min[sk].

4. Assign an intercommunity degree (number of edges outside the assigned community),

db(vi), to each node vi as db(vi) = d(vi) − dw(vi).

5. Generate intercommunity edges using a modified Havel-Hakimi model and randomize

them. Nodes are sorted by intercommunity degree from highest to lowest and the node

at the top of the list is connected randomly with another node. For an edge to form, the

two nodes may not be previously connected, must belong to different communities, and

may not both have an intracommunity degree = 0. The list is then resorted and a new

edge is formed, repeating until all intercommunity degrees are fulfilled. Edges are then

randomized using double-edge swaps. Two intercommunity edges (a,b) and (c,d) are

randomly chosen, removed, and then replaced by two new edges (a,c) and (b,d). Newly

connected nodes cannot belong to the same community, and nodes are not permitted to

form self-loops or connect to another node multiple times. This process is then repeated

N/2 times to randomize all edges. If the full network is not a single connected component

after this process, one edge within the largest connected component and one edge

within the next largest component are swapped, repeating until the network is a single

connected component.

6. Generate intracommunity edges using the Havel-Hakimi model and randomize them.

Within each community independently, nodes are sorted by intracommunity degree from

highest to lowest. The node with the highest intracommunity degree is connected to the

node with the next highest intracommunity degree in the same community. Nodes are

then resorted and the process repeats until all intracommunity degrees are satisfied.

Edges cannot be formed if nodes are previously connected. Nodes are not permitted to

form self-loops. Edges are then randomized using double-edge swaps as described in

Step 5. If each community is not a single connected component after this process, one

edge within the largest connected component and one edge within the next largest

component are swapped, repeating until each community is a single connected

component.

We adapted publicly available scripts204 to generate networks under this algorithm in

Python 3.7 (Python Software Foundation, https://www.python.org). Network adjacency matrices were then exported to MATLAB 9.3 (The MathWorks Inc., Natick, MA) for community detection via optimization of the quality function described above.175,190 Generated networks replicated critical features of our empirical network: 1) network size (N) = 2,350 persons; 2) modularity (Q)

= 0.96; 3) mean node degree = 2.1; 4) mean community size = 112 members.

HIV Transmission Model

Simple transmission models are widely used to evaluate fundamental network properties and potential intervention strategies.13,14 We applied a highly simplified HIV transmission model to simulate endemic infection spread across a sexual contact network, specifying a modified SI structure with one additional compartment for persons protected from HIV infection and another for persons who are rendered non-infectious via effective treatment [Figure 3.8]. Persons in the model may be “susceptible” to HIV (S), “protected” (P) via PrEP for HIV, infected and

“infectious” (I), or “virally suppressed” (V) from highly adherent ART use. Transmission was modeled within HIV-discordant dyads, or those in which one person is at state I and one person is at state S. Transmission could not occur from persons who were in state V or to persons who were in state P, reflecting an assumption of perfect intervention efficacy, and persons in these

55 states could not revert to the I or S states, respectively. In each daily time step, transmission occurred within each discordant dyad with probability 1 − (1 − 푝)c where 푝 was the per-act transmission probability associated with condomless anal intercourse (CAI) between HIV- discordant partners and c was the act rate per partner per day. We simulated transmission within each discordant dyad in each time step using a draw from a Bernoulli distribution with a probability parameter equal to the daily transmission probability.

Act rate per partner per day and per-act transmission probability were drawn from the literature. We specified a fixed daily act rate per partner of 0.22, based on a large internet survey of US MSM in which the mean annualized number of sex acts with the most recent male sexual partner in the previous year was reported to be 80.6.205 We also specified a fixed per-act transmission probability of 0.0076, representing the midpoint between meta-analysis estimates of 0.0011 and 0.0140 for receptive and insertive anal intercourse, respectively, in the absence of viral suppression, PrEP, and condom use.206,207

We first simulated a series of nine intervention-free scenarios in which we varied the starting overall network prevalence from 0.10 to 0.50 at increments of five percentage points.

For each overall endemic prevalence, we seeded each of the 100 generated networks with that prevalence at the start of each transmission simulation. To achieve this overall endemic prevalence while maximizing variability in prevalence across communities, we assigned unique prevalences between 0% and 100% to each community at time zero using random draws from a beta distribution with a mean equal to the specified overall endemic prevalence and a fixed α parameter of 2 [Figure 3.10]. Persons were then randomly sampled within each community and assigned to be in state I at the start of the simulation to fulfill assigned community-level endemic prevalences. No persons were initially seeded in the P or V states, nor did any persons move to the P or V states over time (PrEP and ART assignment under varying intervention scenarios are described below).

We then simulated HIV transmission within all 100 networks over 1000 days for each overall endemic HIV prevalence, estimating the HIV incidence rate within the full network as incident infections per susceptible person-year. HIV incidence rates across all 100 networks were used to estimate the mean HIV incidence rate and 95% confidence intervals at each overall endemic prevalence. We also assessed the relationship between endemic community- level HIV prevalences and community-level HIV incidence rates at each overall endemic HIV prevalence. Transmission simulations were conducted in R v. 3.5.0187 with the R EpiModel208 package using the Longleaf computing cluster at UNC-CH.

Intervention Scenarios

PrEP) per allocation scheme. Under the random allocation scheme, 10%, 30%, or 50% of eligible persons (i.e., persons who would otherwise start in the “S” or “I” state) were randomly allocated at time 0 to PrEP or ART, depending on the intervention being modeled. Under the preferential allocation schemes, the probability of an eligible person receiving an intervention was linearly related (directly for the high-to-low scheme and inversely for the low-to-high scheme) to the endemic HIV prevalence in that person’s community.

The sample function with the prob argument in R was applied to achieve weighted sampling without replacement for intervention assignment, sampling from persons in state S for

PrEP interventions and from persons in state I for ART interventions. Under high-to-low community-level endemic prevalence intervention allocation schemes, the probability applied was the initial community HIV prevalence for each eligible person, normalized to sum to 1 across all persons in state S for PrEP interventions or across all persons in state I for ART interventions. Under low-to-high intervention allocation schemes, the probability applied was 1 – initial community HIV prevalence, normalized in the same manner. The prob argument samples from the specified population, where each person is selected with the normalized probability given. The sampled person is then removed from the population and all probabilities are re- normalized to sum to 1, with the process repeating until the sample of desired size (in our case, intervention coverage level) is achieved.

Preferential allocation schemes were probabilistic, rather than deterministic, to mirror the practicalities of public health interventions: although some persons or populations within a network may be prioritized for enhanced PrEP or viral suppression support, public health personnel will typically attempt to reach all relevant persons with interventions. Under such an approach, not all persons in prioritized populations will initiate PrEP or achieve viral suppression under any intervention, while some persons not prioritized will start PrEP or become virally suppressed.

Each of the six intervention scenarios (ART at 10%, 30%, and 50% coverage; PrEP at

10%, 30%, and 50% coverage) was modeled according to each of the three distribution schemes (random, low-to-high, high-to-low) within each of the nine overall endemic HIV prevalence scenarios, for a total of 162 intervention scenarios. The HIV incidence rate in each scenario was estimated as the mean number of new HIV infections per susceptible or protected person-year across the 100 networks over the full 1000-day simulation. We estimated the intervention effect as the percentage reduction in the HIV incidence rate under each intervention

58 in comparison to the incidence rate at the same overall endemic HIV prevalence under no intervention.

Comparisons of ART vs. PrEP strategies based on intervention effect may be misleading because the absolute number of persons reached with an intervention at a given percent- coverage level varied across the two interventions, with the extent of the difference depending upon the distance between 0.5 and the overall endemic HIV prevalence. To account for this discrepancy and allow for direct comparison of interventions, we defined a measure of intervention efficiency that adjusted for the number of persons reached by a given intervention.

We estimated overall intervention efficiency for each intervention scenario as the percentage reduction in the HIV incidence rate per person reached with the intervention, compared to the mean incidence rate under the no-intervention model at the same overall endemic network prevalence.

We pursued two distinct objectives: 1) compare the effect and efficiency of ART vs.

PrEP interventions under a set scenario (intervention coverage level and overall endemic HIV prevalence); and 2) compare the effect and efficiency of a given intervention type (ART or PrEP) across the three allocation schemes under a set scenario. Direct comparison of ART and PrEP interventions was conducted to identify optimal allocation of limited intervention resources under each starting scenario, while ART and PrEP interventions were tested across intervention allocation schemes to evaluate if differential allocation by community-level endemic prevalence provided added value in comparison to a random allocation scheme, which is simpler to implement.

59 variation in community-level endemic prevalence under our infection seeding procedure [Figure

3.10]. At low overall endemic prevalences, most communities have low endemic prevalences and very few have high endemic prevalences, creating little variation for a preferential allocation scheme to exploit. At high overall endemic prevalences, community-level endemic HIV prevalences are more normally distributed, and there are both communities with low and high endemic prevalences from which persons can be sampled for intervention under high-to-low or low-to-high allocation schemes.

Figure 3.1 HIV and early syphilis contact tracing lookback periods

Figure 3.2 Data flow for North Carolina (NC) Division of Public Health (DPH) – Cecil G. Sheps Center for Health Services Research at the University of North Carolina at Chapel Hill (UNC- CH) data linkage

Figure 3.3 Network interconnectivity vs. overlap

CAPTION: The top row provides a high-level visualization and the bottom row illustrates this concept in a mock network. Example networks in panels A and B have the same number and proportion of overlapping nodes (3/15), but divergent levels of interconnectivity (12.5% vs. 50%). Members of the HIV network only are shown in blue, members of the syphilis network only are shown in red, and members of both the HIV and syphilis networks are shown in yellow. Nodes within the HIV network that contribute to the observed percentage increase in node degree among syphilis network members upon combination with the HIV network are distinguished with bold borders. A. Network with comparatively low interconnectivity of the syphilis network with the HIV network. Members of both networks are clustered together, bridging distinct HIV and syphilis networks, and a large portion of the syphilis network is relatively insulated from the HIV network. B. Network with comparatively high interconnectivity of the syphilis network with the HIV network. The HIV network is embedded within the syphilis network, directly connecting much of the syphilis network with the HIV network.

Figure 3.4 Example multilayer network

CAPTION: Two independent networks are shown in separate boxes. Nodes that are present in both layers are shown in white and nodes that are distinct to one layer are shown in black. Each node may be connected by an edge to any other node within a given layer (solid lines) and/or to itself between layers (dashed curves).

Figure 3.5 Causal diagram used to select confounders of the association between contact tracing interview by disease intervention

specialists (DIS) and incident viral suppression

CAPTION: Directed acyclic graph assessing the effect of DIS interview during contact tracing on the six month cumulative incidence of reported viral suppression. Age, event date, network event status, and years since HIV diagnosis were identified as potential confounders of the association of interest and are highlighted in red. Factors that may create additional uncontrolled confounding that cannot be addressed with the available data, including socio-economic status and perceptions of public health services, are highlighted in blue.

Figure 3.6. Algorithm for identification of individuals prescribed emtricitabine and tenofovir disoproxil fumarate (FTC/TDF) for pre-exposure prophylaxis (PrEP) against HIV

Figure 3.7 Power law and lognormal distribution fits to empirical network data

6 7

CAPTION: Power law and lognormal distribution fits to the A. observed node degree; B. intra-community node degree; and C. community size distributions from the largest connected component of the combined HIV/syphilis contact network in the Aim 1 network. Empirical data are represented by circles and fitted distributions are represented by lines. The power law fit to the data is shown in red and the lognormal fit is shown in blue.

Figure 3.8 Adapted SI transmission model

CAPTION: Under no intervention, susceptible (“S”) nodes become infected and infectious (“I”) through HIV acquisition, which is a function of contact rates and transmission probabilities within HIV-discordant dyads. In intervention scenarios, susceptible nodes may become fully protected (“P”) against HIV acquisition through sustained and adherent PrEP use and infectious nodes may become virally suppressed (“V”) and fully non-infectious through sustained and adherent treatment with ART.

Figure 3.9 Beta distributions for sampling of community-level endemic HIV prevalences

CAPTION: Beta distributions for sampling of community-level endemic HIV prevalences at each overall endemic HIV prevalence 0.10 to 0.50. Beta distributions were generated with mean = overall endemic HIV prevalence and α = 2.

Table 3.1 Match process for linkage between North Carolina (NC) Electronic Disease Surveillance System (EDSS) and insurance claims data

Unique people Single matches in Multiple matches Match Matching variables in NC EDSS insurance data in insurance data First name, last name, date 1 1000 968 32 of birth, gender, ZIP code First name, last name, date 2 of birth, gender, ZIP code 487 470 17 within +/- 50 miles First name, last name, date 3 301 274 27 of birth, gender Total 1788 1712 76

Table 3.2 Selected partition statistics considered in selection of the final resolution parameter 훾 for community detection

Number of Proportion of communities with ≥30 nodes in Number of members of syphilis Largest eligible 훾 communities network community Q communities 0.45 34 26 144 0.9641 0.919 0.40 33 24 171 0.9658 0.911 0.35 32 24 164 0.9675 0.919 0.30 29 24 191 0.9692 0.965 0.25 26 21 229 0.9711 0.965 0.20 25 20 243 0.9734 0.965

CHAPTER FOUR: NETWORK INTERCONNECTIVITY AND COMMUNITY DETECTION IN THE HIV/SYPHILIS SYNDEMIC AMONG MEN WHO HAVE SEX WITH MEN

A. Introduction

Due to both biological117 and behavioral8 synergies, the interacting HIV and syphilis epidemics each exacerbate the impact of the other to form a syndemic.8 In the US, HIV and syphilis constitute substantial public health burdens, particularly among MSM.104 In the US in

2017, 25,748 and 17,736 MSM were newly diagnosed with HIV209 and syphilis,210 respectively.

Although estimated HIV incidence has decreased among US MSM overall in recent years, HIV incidence has remained steady among Black MSM and is rising among Hispanic/Latino MSM.209

Syphilis diagnoses increased steadily from 2013 through 2017 within all populations in the US, with the largest increases also observed among minority MSM.210 Novel intervention strategies are needed to curtail the ongoing HIV/syphilis syndemic, particularly among MSM of color.

Although numerous studies of HIV and STD networks have been undertaken to guide prevention and treatment efforts for a single infection,148,142 few analyses have generated combined networks containing persons diagnosed with HIV or syphilis and their sexual contacts.

Each of the existing combined-network analyses10,136,169,170 primarily evaluated the union of the two networks as a single network, rather than interrogating how the networks interact. New approaches that assess the relationship between HIV and syphilis contact networks could inform the design of interventions deployed through these networks. For example, identification of populations in which the syphilis network is highly interconnected with the HIV network may be a powerful tool for guiding focused HIV prevention efforts.

We conducted a statewide, network-based evaluation of the HIV/syphilis syndemic among MSM in NC, a setting where the burden of both infections is high.15,113 Using surveillance-based diagnosis and contact tracing data for both infections, we sought to characterize the independent and combined HIV and syphilis networks and identify populations wherein these networks are highly connected. We devised a simple measure to quantify the amount and distribution of overlap across two networks; we termed this measure

“interconnectivity” and posited that interconnectivity between HIV and syphilis contact-tracing networks could allow infection transmission to cross networks. We then identified distinct network communities, or groups of densely connected nodes,12 in which the syphilis network showed comparatively heightened interconnectivity with the HIV network. Detection of these communities may: 1) highlight key populations disproportionately burdened by the syndemic; and 2) guide resource allocation to portions of the syphilis network that experience elevated exposure to the HIV network and may thus disproportionately benefit from HIV prevention efforts.

B. Methods

Study Data

All data were drawn from the NC EDSS, which is operated by the NC Division of Public

Health. NC EDSS captures demographic, medical, and intervention data for all reportable communicable diseases in NC, including HIV and syphilis.179 Contact tracing data in NC EDSS are collected by DIS who interview all persons newly diagnosed with HIV and early syphilis in

NC to identify recent sexual and injection drug use contacts. Sexual contacts are elicited for the twelve months prior to DIS interview for persons newly diagnosed with HIV127 and for the twelve, six, and three months prior to symptom onset for persons diagnosed with early latent, secondary, and primary syphilis, respectively (corresponding to the mean incubation times for

73 these three phases of early syphilis).128 DIS then attempt to trace and screen all named contacts for both HIV and syphilis.

Data were extracted from NC EDSS using SAS 9.4 (SAS Institute Inc., Cary, NC), and all analyses were conducted in R v. 3.5.0187 and MATLAB 9.3 (The MathWorks Inc., Natick,

MA). The Non-Biomedical Institutional Review Board at UNC-CH approved this study (IRB

Number: 18-3110).

Independent and Combined Contact Network Generation

We defined a network event as receiving an HIV or early syphilis diagnosis between

All new HIV and early syphilis diagnoses among persons diagnosed in NC between

We generated an independent HIV network comprising persons diagnosed with HIV and their contacts, with sexual relationships reported between these persons during HIV contact tracing as edges. We generated an analogous independent syphilis network comprising persons diagnosed with syphilis and their contacts, linked by edges representing sexual relationships reported during syphilis contact tracing. Finally, we generated a combined HIV/syphilis network comprising all persons and edges from both independent networks.

Network Description

We calculated standard network parameters, including number of nodes, number of edges, component sizes, node degrees, and homophily143 for the combined and independent networks. Methodological details about these measures are provided in Chapter 3.

Network Overlap and Interconnectivity

We used a traditional measure of network overlap, the Jaccard node similarity, to quantify the direct overlap between the independent HIV and syphilis networks. We calculated this measure as the number of nodes belonging to both independent networks (intersection) divided by the number in the combined network (union).172 We calculated the Jaccard edge similarity analogously as the number of edges in the intersection divided by the number in the

75 union. For both measures, a value of 1 represents complete overlap and a value of 0 represents no overlap.

Dickison et al.173 for evaluation of coupled, disjoint network systems, but our simpler measure allows for nodes and edges to be shared between partially overlapping networks. Further, this measure does not require that networks’ degree distributions demonstrate an internal scale, a requirement is violated by scale-free networks, including many sexual contact networks.

Community Detection

Community detection identifies clusters of nodes within a network that are more densely connected to each other than to the remainder of the network.12 We employed community detection to identify clusters of persons within the combined HIV/syphilis network among whom the syphilis network was highly interconnected with the HIV network, suggesting greater potential for exogenous HIV entry into the syphilis network. We applied a generalized Louvain method for community detection, including the independent HIV and syphilis networks as distinct network layers.177 This method maximizes a quality function Q, also termed the modularity of the network, to identify the set of community assignments that produces the highest concentration of edges within network communities. Larger values of Q (maximum possible value = 1) indicate that more edges occur within communities (rather than between communities), while Q = 0 indicates that edge occurrence is uncorrelated with community labels. The quality function and its application in this study are detailed in Chapter 3. Community detection analyses were restricted to the largest connected component in the combined network to ensure sufficient network density.

Heterogeneity of Syphilis/HIV Network Interconnectivity by Network Community

To characterize potentially distinguishing features of highly interconnected communities that could aid in intervention development and deployment, we applied simple linear regression models to assess the unadjusted associations between community membership characteristics

77 and community-level interconnectivity (relative percentage increase in syphilis network node degree following combination of the syphilis network with the HIV network). Independent variables included median age of community members at first network event; proportion of community members who self-identified as Black, non-Hispanic; proportion of persons with syphilis diagnoses who received their diagnoses at STD clinics in each community; and proportion of community members who were members of both the HIV and syphilis networks.

With the exception of syphilis diagnosis site, we included all members of each community in the calculation of these community descriptors because understanding communities as a whole may allow for insights into the forces driving overall network structure. Associations of interconnectivity with demographic descriptors (age, race/ethnicity) and syphilis diagnostic site may aid in identifying indicators of heightened interconnectivity that could guide resource allocation within the syphilis network without requiring complex network analyses in real time.

C. Results

Network Populations

There were 2,090 and 3,419 HIV and early syphilis diagnoses, respectively, that met our inclusion criteria between 2015 and 2017 in NC [Table 4.1]. These diagnosis events produced

2,439 eligible HIV contact events and 3,545 eligible syphilis contact events, for a total of 11,493 network events among 8,020 unique network members. Most network events were among persons who self-identified as Black, non-Hispanic (60%) or White, non-Hispanic (29%) and the median age at the first eligible network (diagnosis or contact) event during the study period was

28.7 (interquartile range [IQR]: 24.2-36.8) years. Age at first diagnosis or contact event was lower for HIV vs. syphilis network members (median age 27.7 years vs. 29.0 years), and HIV diagnosis and contact events were more likely than syphilis diagnosis and contact events to occur among persons self-identifying as Black, non-Hispanic (62% vs. 58%). Persons

78 diagnosed with early syphilis were most commonly diagnosed by a primary care physician or

HMO provider (36.8%) or at an STD clinic (36.2%).

Independent and Combined Network Characteristics

The largest connected component in the syphilis network comprised 13% of all syphilis network members, whereas the largest connected component in the HIV network comprised 9% of all HIV network members [Table 4.2, Figure 4.1]. In the combined network, edges in each of the independent networks bridged disconnected components across the other independent network, creating a largest connected component containing 2,350 persons (29% of all combined network members). Similar proportions of persons were identified as isolates in the

HIV, syphilis, and combined networks (19.0%, 19.8%, and 19.6%, respectively). The overall mean node degree, including isolates, in the combined HIV/syphilis network (1.49; 95% CI:

1.45, 1.53) was greater than that observed in the independent HIV (1.24; 95% CI: 1.19, 1.28) and syphilis (1.34; 95% CI: 1.30, 1.38) networks, while levels of homophily were similar across all three networks. In the combined HIV/syphilis network, the median difference in age between persons diagnosed with HIV and/or syphilis and their named contacts was 5.3 years and 70% of case-contact pairs shared concordant racial and ethnic self-identities.

We observed minimal direct overlap between the independent networks according to

Jaccard node and edge similarities, with 15% of persons and 6% of edges included in both the

HIV and syphilis networks, including isolates. The mean node degree among members of the syphilis network increased by a relative 26.0% following combination with the HIV network (1.34 vs. 1.69; interconnectivity = 26%).

Network Communities

We identified twenty-six distinct network communities within the largest connected component of the combined network [Figure 4.2] and observed modularity (Q) = 0.97 out of a

79 maximum value of 1.0. Twenty-one communities comprising 2,268 persons (97% of persons in the largest connected component) contained ≥ 30 members of the syphilis network each and were thus evaluated for interconnectivity. The mean relative increase in node degree among members of the syphilis network following combination with the HIV network spanned 15.2% to

72.2% across communities, demonstrating heterogeneity in interconnectivity by community.

Our measure of interconnectivity incorporates overlap between networks, and consequently, we observed an association between the percentage of community members who were members of both the HIV and syphilis networks and interconnectivity. Across communities, an absolute increase of one percentage point in the percentage of community members belonging to both networks, compared to the lowest observed value, was associated with an absolute increase of 2.0 (95% CI: 1.4, 2.6) percentage points in interconnectivity (i.e., in the relative increase in node degree in the combined vs. independent syphilis networks within a community. We also identified an association between syphilis diagnostic site and interconnectivity: a ten-percentage-point absolute increase in the percentage of persons diagnosed with syphilis at STD clinics, compared to the lowest observed value, was associated with an absolute increase of 2.5 (95% CI: -1.7, 6.7) percentage points in interconnectivity.

We observed associations between demographic identifiers and interconnectivity as well. A one-year absolute decrease in the median age of community members, compared to the highest observed value, was associated with an absolute increase of 1.7 (95% CI: -0.4, 3.8) percentage points in interconnectivity. A ten-percentage-point absolute increase in the percentage of persons who self-identified as Black, non-Hispanic, compared to the lowest observed value, was associated with an absolute increase of 3.2 (95% CI: 0.7, 5.7) percentage points in interconnectivity. Age and race/ethnicity within communities were also associated with each other: communities with a younger median age tended to show a higher proportion of members who self-identified as Black, non-Hispanic [Figure 4.3].

D. Discussion

Despite the persistent HIV/syphilis syndemic among MSM and nuanced scientific understanding of the biological interaction between the two infections,117 the relationship between HIV and syphilis contact networks has not been well characterized. Improved understanding of interactions between these networks may allow for development of efficient intervention approaches, including HIV prevention strategies prioritizing sub-populations in which the syphilis network is highly interconnected with the HIV network. In this study, we devised and applied a simple measure of network interconnectivity in a statewide HIV/syphilis syndemic as an initial step toward such an approach.

Our application of community detection is a novel strategy for identifying network subsets with potentially increased risk of infection entry. Network community structures have been widely observed across biological, neural, computational, and social networks, including contact networks in human and animal populations.175,176 Our network showed extremely high modularity, or concentration of edges within network communities, of 0.97. A high modularity of

0.85 was also observed in an analysis of a sexual contact network of PLWH in Cuba under a different community detection algorithm that identified distinct network communities characterized by HIV transmission risk group.176 High modularity is associated with explosive transmission within communities and rare transmission from one community to another.175 The high modularity that we observed in the HIV/syphilis network among NC MSM suggests that specific sub-populations belonging to communities in which the syphilis network is particularly interconnected with the HIV network are potential priorities for interventions to prevent HIV transmission.

Homophily can contribute to the formation of network community structure,211 and in the combined HIV/syphilis sexual contact network among MSM in NC, homophily by age and race/ethnicity produces network communities with demographic commonalities. In our network

81 population, younger median age and greater self-identification as Black, non-Hispanic among community members were associated with heightened interconnectivity of the syphilis network with the HIV network. This finding is congruent with prior studies illustrating the role of racial homophily, alongside existing racial and ethnic disparities in HIV prevalence and treatment, in driving HIV incidence among Black MSM.154,155 As most transmissions within highly modular networks occur within communities,175 community structuring derived from homophily211 may also drive transmission within vulnerable demographic sub-groups. This effect likely contributes to disparities in HIV incidence and HIV/syphilis coinfection among young, Black MSM.115 If enhanced HIV prevention interventions must be prioritized within the syphilis network due to resource constraints, young MSM of color appear to be optimal recipients due to their increased membership in network communities with heightened interconnectivity of the syphilis network with the HIV network. Persons diagnosed with syphilis at STD clinics, many of which are housed in city and county health departments, may also experience substantial exposure to the HIV network. Enhanced HIV prevention services for persons receiving a syphilis diagnosis at these sites and for their partners may be particularly beneficial as well.

Beyond the detection of specific network communities with elevated interconnectivity and potentially heightened risk of HIV exposure among syphilis network members, we observed interconnectivity between the HIV and syphilis networks overall. This interconnectivity, in conjunction with increased coverage of the largest connected component in the combined vs. independent networks, suggests that the syphilis network may be treated as largely interwoven with the HIV network during intervention planning. That is, in the absence of constraints necessitating more focused interventions, HIV prevention interventions that attempt to reach the full combined network have the potential to make a strong impact. For example, NC and other jurisdictions have recently expanded HIV prevention services to include referrals for pre- exposure prophylaxis (PrEP) by DIS during contact tracing for both HIV and syphilis.133

Our analysis is limited to assessment of network interconnectivity and does not attempt to quantify individual HIV risk, as our networks aggregated sexual partnerships over several years and did not capture the time-course of individual infection statuses or the spread of infection within or across networks. Instead, we aimed to evaluate the HIV/syphilis syndemic from a network perspective by identifying populations in which the syphilis network is more likely to be interconnected with the HIV network, irrespective of individual syphilis network members’

HIV statuses. With this exploratory inquiry, we aim to spark future research on network interconnectivity and community detection in public health.

We acknowledge that persons diagnosed with HIV and syphilis and their named contacts do not represent the complete HIV and syphilis transmission networks in NC. Network membership was subject to selection bias because members were identified via self-reported sexual orientation and gender identity, and our networks only included persons receiving infection diagnoses or being named as contacts.163 Further, contact tracing data were limited to sexual contacts, were self-reported, and can suffer from social desirability bias and recall error.163 Our networks were also retrospective and static, ignoring some important complexities of sexual partnerships for analytical tractability. Although these factors limit generalizability beyond the reported network population, our findings are highly germane to the development of

MSM-focused HIV treatment and prevention interventions for deployment within the existing HIV and syphilis contact tracing infrastructure. Future analyses may extend these methods to additional populations, transmission modes, and contact networks to more fully capture the entire transmission network.

Sexual contact networks containing HIV and syphilis are interconnected among MSM in

NC, and the sexual network of known HIV and syphilis cases and contacts offers an efficient platform for deployment of intensified HIV prevention interventions. Limited HIV prevention resources may be most efficiently utilized by leveraging existing contact tracing infrastructure to

83 reach populations in which the syphilis network is highly interconnected with the HIV network. In an era where “Ending the HIV Epidemic” campaigns are being mounted across the US,178 this type of information may be helpful in the design and deployment of interventions against persistent HIV incidence in particularly vulnerable populations.

Figure 4.1 Largest connected components in the independent and combined HIV/syphilis networks

CAPTION: Blue indicates persons diagnosed with HIV 2015-2017, red indicates persons diagnosed with early syphilis 2015-2017, gold indicates persons diagnosed with HIV and early syphilis 2015-2017, and gray indicates persons not diagnosed with HIV or early syphilis in the study period. Shape indicates network event status. Circular nodes were diagnosed with HIV or early syphilis 2015-2017, square nodes were named as sexual contacts of persons diagnosed with HIV or early syphilis, and triangular nodes were diagnosed with HIV or early syphilis and named as contacts. A. Largest connected component in the independent HIV network (N=334), including persons diagnosed with HIV 2015-2017 and their named sexual contacts. B. Largest connected component in the independent syphilis network (N=672), including persons diagnosed with early syphilis 2015-2017 and their named sexual contacts. C. Largest connected component in the combined HIV/syphilis network (N=2,350), including persons diagnosed with HIV or early syphilis and their named sexual contacts 2015-2017.

Figure 4.2 Network communities

CAPTION: A. Network communities in the largest connected component of the HIV/syphilis network. Communities are distinguished by color. B. HIV/syphilis network membership by network community. Each network community containing ≥30 members of the syphilis network in the largest connected component of the two-layer network is represented by a pie chart. Pie chart area is proportional to the community size and colors reflect community members’ network membership. Communities not included in interconnectivity analyses (due to syphilis network membership <30) are shown in gray. Lines connecting communities represent ≥1 edge between members of the two communities. Communities are laid out under the force-directed Fruchterman-Reingold algorithm. Network community membership was visualized using an adaptation of the code available at http://netwiki.amath.unc.edu/VisComms/VisComms.191

Figure 4.3 Demographic characteristics of network community members

CAPTION: Demographic characteristics of the twenty-one network communities containing ≥30 members of the syphilis network in the largest connected component of the combined HIV/syphilis network. The relative percentage increase in mean node degree among members of the syphilis network following combination with the HIV network in each community is indicated by color and community size is represented by point size. The percentage of community members who self-identify as Black, non-Hispanic on the y-axis is compared to the median age of community members at their first included network event on the x-axis.

Table 4.1 Individual characteristics by network event status within the combined HIV/syphilis network among men who have sex with men in North Carolina 2015-2017

Combined Network event status Individual characteristics HIV/syphilis HIV HIV Early syphilis Early syphilis (N, %) network diagnoses contacts diagnoses contacts Age at network event 18-24 3342 (29.1) 741 (35.4) 774 (31.7) 872 (25.5) 955 (26.9) 25-34 4643 (40.4) 822 (39.3) 1047 (42.9) 1319 (38.6) 1455 (41.0) 35-44 1871 (16.3) 251 (12.0) 383 (15.7) 612 (17.9) 625 (17.6) 45+ 1637 (14.2) 276 (13.2) 235 (9.6) 616 (18.0) 510 (14.4) Race/ethnicity White, non-Hispanic 3339 (29.0) 533 (25.5) 601 (24.6) 1082 (31.6) 1123 (31.7) Black, non-Hispanic 6865 (59.7) 1266 (60.6) 1581 (64.8) 2011 (58.8) 2007 (56.6) Hispanic 814 (7.1) 211 (10.1) 126 (5.2) 263 (7.7) 214 (6.0) Asian, non-Hispanic 76 (0.7) 14 (0.7) 7 (0.3) 32 (0.9) 23 (0.6)

Other 121 (1.0) 19 (0.9) 22 (0.9) 31 (0.9) 49 (1.4) 88

Missing 278 (2.4) 47 (2.2) 102 (4.2) 0 (0.0) 129 (3.6)

Syphilis diagnostic site STD clinic ------1090 (36.2) --- Private physician/HMO ------1110 (36.8) --- Other clinic ------405 (13.4) --- Emergency room ------117 (3.9) --- Other ------293 (8.6) --- Total network events 11493 2090 2439 3419 3545 STD; Sexually transmitted disease. HMO; Health maintenance organization

Table 4.2 Network parameters for the combined HIV/syphilis and independent HIV and syphilis sexual contact networks among men who have sex with men in North Carolina 2015-2017

Network parameter Combined network HIV network Syphilis network Network size (nodes) 8020 3928 5283 Number of unique edges 5984 2439 3545 Isolates (N, %) 1572 (19.6) 779 (19.8) 1003 (19.0)

Component size (N, %) Dyad 883 (63.6) 468 (57.0) 642 (61.4) Small (3-9) 473 (34.1) 318 (38.7) 365 (34.9) Medium (10-29) 30 (2.2) 31 (3.8) 30 (2.9) Large (≥ 30) 2 (0.1) 4 (0.5) 8 (0.8) Total 2960 1596 2048 Largest connected component (N, %) 2350 (29.3) 334 (8.5) 672 (12.7)

Node degree (mean, 95% CI) 1.5 (1.5, 1.5) 1.2 (1.2, 1.3) 1.3 (1.3, 1.4) 89 Homophily

Age in years: |age1 – age2| (median, IQR) 5.3 (2.2, 7.8) 4.9 (2.1, 7.4) 5.6 (2.4, 8.1) Race/ethnicity: homophilic preference (mean, 95% CI) 0.7 (0.7, 0.7) 0.7 (0.7, 0.7) 0.7 (0.7, 0.7) Jaccard node similarity 0.15 ------Jaccard edge similarity 0.06 ------

IQR; interquartile range. 95% CI; 95% confidence interval

CHAPTER FIVE: HIV VIRAL SUPPRESSION AND PRE-EXPOSURE PROPHYLAXIS IN HIV AND SYPHILIS CONTACT TRACING NETWORKS: AN ANALYSIS OF DISEASE SURVEILLANCE AND PRESCRIPTION CLAIMS DATA

A. Introduction

Despite the development of highly effective HIV prevention modalities, HIV incidence is steady or increasing in many populations.1 Available interventions, including ART for PLWH and

PrEP among HIV-negative persons, can substantially reduce HIV transmission.2-5 However,

ART and PrEP uptake and adherence are widely suboptimal,6,7 preventing these interventions from reaching their full potential at the population level. New strategies are needed to increase

ART and PrEP use for improved clinical and prevention outcomes.

For both epidemiological and practical reasons, sexual networks containing persons diagnosed with HIV and/or syphilis may be efficient platforms for HIV prevention interventions.

HIV and syphilis are syndemic, particularly among MSM,8 and HIV-uninfected members of these networks experience heightened HIV risk.9,10 Among persons with a previous HIV diagnosis and unsuppressed viremia, being diagnosed with early syphilis or named as a sexual contact of a person diagnosed with HIV or early syphilis suggests potential for ongoing HIV transmission risk.10 Because contact tracing for HIV and syphilis is routine in the US,11 existing public health infrastructure can serve as a conduit for reaching HIV-positive and HIV-negative network members with ART and PrEP, respectively.

Estimates of HIV viral suppression and PrEP use within HIV/syphilis networks are required to identify resource needs and coverage gaps for enhanced intervention design. Using communicable disease surveillance data linked to prescription claims data, we evaluated

prevalent and incident viral suppression and PrEP use in the HIV/syphilis contact tracing network among MSM in NC.

B. Methods

Data Sources and Study Population

We identified our study population within NC EDSS, which is maintained by NC DPH.179

NC EDSS hosts personal, laboratory, medical, and contact tracing data for reportable communicable diseases in NC, including HIV and syphilis.179 Contact tracing data in NC EDSS are collected by DIS who interview persons diagnosed with HIV or early syphilis (primary, secondary, or early latent) to identify recent sexual and injection drug use contacts. DIS elicit contacts during the twelve months before DIS interview for persons newly diagnosed with HIV127 and during the three, six, and twelve months before symptom onset for persons diagnosed with primary, secondary, and early latent syphilis,128 respectively. DIS trace and test named contacts for both infections if possible.

We first identified all persons becoming diagnosed with HIV or early syphilis in NC over the period January 2013 – June 2017 (inclusive) within NC EDSS. We then identified the sexual partners named by these persons during contact tracing. We defined a qualifying HIV/syphilis network event as receiving an early syphilis diagnosis or being named as a sexual contact by a person becoming diagnosed with HIV or early syphilis. Some persons experienced multiple network events during the study time frame; all such events were included in analyses. We further restricted our study population to include only persons who: 1) were ≥18 years old on the date of the network event; 2) resided in NC at the time of the event; 3) did not self-identify as transgender; and 4) either self-identified as MSM or identified as male and were named as a sexual contact by a self-identifying male partner. Self-identified transgender individuals were

excluded because gender identity was unavailable for persons diagnosed with syphilis during the study period, preventing subgroup analyses of transgender women.

For analyses of prevalent and incident PrEP use, persons with confirmed HIV-negative or unknown HIV status at the time of the network event were first linked to insurance claims data from a single large, commercial insurer in NC181 by an honest broker at the Cecil G. Sheps

Center for Health Services Research at UNC-CH [Table 3.1]. We included only those individuals linked who were enrolled in a commercial insurance plan with prescription coverage over the relevant time frames (specified below). As most persons without a prior HIV diagnosis are tested for HIV at the time of a network event, persons without an HIV diagnosis on record in NC

EDSS before or in the 14 days after the event were presumed HIV-negative. The impact of this assumption was interrogated in a sensitivity analysis limiting PrEP analyses to events among persons who 1) were reached by DIS at the time of the network event, consented to HIV testing, and had a negative result; OR 2) had an unrelated, documented HIV-negative test result in NC

EDSS in the three months before or ever following the network event, through December 2019.

Data were extracted from NC EDSS using SAS 9.4 (SAS Institute Inc., Cary, NC) and analyses were conducted in SAS 9.4 and R v. 3.5.0.187 The Non-Biomedical Institutional Review

Board at UNC-CH approved this study (IRB Number: 18-3110).

Analyses

Prevalent and Incident Viral Suppression among Previously Diagnosed PLWH

Effect of DIS Services on Incident Viral Suppression

HIV, and contact interviews for syphilis screening among those named as contacts of persons diagnosed with early syphilis. We applied linear risk regression to estimate an aCID comparing

incident viral suppression among those reached vs. not reached by DIS. A generalized estimating equations (GEE) model with an autoregressive correlation structure was used to account for the presence of multiple observations for some persons194

Plausible confounders of the possible effect of DIS services on incident viral suppression were identified through a comprehensive literature review and assessed using a DAG [Figure

3.6].196 Confounders were then evaluated to identify optimal functional form, resulting in nominal coding of event year (2013, 2014, 2015, 2016, 2017), network event status (syphilis diagnosis,

HIV contact, syphilis contact), and age at network event (18-24, 25-34, 35-44, ≥45 years), and restricted cubic splines to represent years since HIV diagnosis. This adjustment set was included in the linear risk GEE model to estimate an aCID with robust standard errors using the sandwich estimator. GEE models were repeated with twelve-month cumulative incidence of reported viral suppression as the outcome. Further GEE modeling details are supplied in

Chapter 3.

Prevalent and Incident PrEP Use among HIV-Negative Persons

PrEP use, as measured by FTC/TDF prescription claims, was evaluated among HIV- negative persons using an adaptation of the CDC algorithm for PrEP utilization.199 Persons prescribed FTC/TDF, also known as Truvada, for PrEP were identified via exclusion of

FTC/TDF prescriptions for HIV or Hepatitis B virus treatment or as HIV PEP [Figure 3.7]. PrEP prescriptions and exclusions due to prior or concomitant prescriptions and/or diagnoses were determined from ICD-9/10-CM codes and NDC in insurance claims data199 [Appendix]. Claims records were queried from January 1, 2010 through 30 days after FTC/TDF prescription as an

“all-available” lookback period to identify exclusions. We manually reviewed all records for each event with an FTC/TDF claim excluded due to prior HIV diagnosis, identifying and correcting erroneous ICD-9/10-CM codes for prior HIV diagnosis in two persons who tested HIV-negative at the network event and indicated to DIS that they were currently using PrEP.

We assessed prevalent PrEP use, defined as ≥1 PrEP prescription claim in the six months before the network event date, among those with continuous pharmacy insurance coverage during that period. We then assessed incident PrEP use following a network event among all persons with pharmacy insurance coverage over the sixty days before the event date and continuously for six months after the event date, excluding persons with documented PrEP use before the network event. Six-month cumulative incidence of PrEP use was defined as ≥1

PrEP prescription claim in the six months following the network event.

We described prevalent and incident PrEP use by age, self-identified race/ethnicity,

HIV/syphilis network event status (syphilis diagnosis, HIV contact, syphilis contact), year of network event, insurance subscriber (self, parent, other), and apparent PrEP indication. We defined PrEP indication at the time of the network event as: 1) having ≥1 male sex partner in the prior six months documented in NC EDSS; AND either 2) reporting condom use during the contact tracing period as anything other than “Always”; OR 3) becoming diagnosed with or reporting a bacterial STI in the prior six months. PrEP indication assessment is detailed in

Chapter 3.

C. Results

We identified 3,963 eligible early syphilis diagnoses and 8,456 HIV/syphilis contact events under our inclusion criteria [Figure 5.1]. Of these network events, 4,959 occurred among

2,970 previously diagnosed PLWH and 6,975 occurred among 5,245 confirmed or presumptively HIV-negative persons. Previously diagnosed PLWH were older (median age:

31.1 vs. 27.3) and more likely to identify as Black, non-Hispanic (71% vs. 54%) than HIV- negative persons, who were more likely to be reached by DIS following a network event (85% vs. 68%).

Viral Suppression among Previously Diagnosed PLWH

Approximately half (52.6%; N = 2,609/4,959) of all events among previously diagnosed

PLWH occurred among persons who were virally suppressed according to surveillance records

[Table 5.1]. Incident viral suppression in the six months after the event was then documented for

35.4% (831/2,350) of events among previously diagnosed PLWH without prevalent viral suppression. Among events without incident viral suppression, 38% had a documented viral load > 200 copies/mL in the six months following the event and 62% had no documented viral loads during this time frame. After adjustment for event year, network event status, age at network event, and years since HIV diagnosis, the six-month cumulative incidence of viral suppression among persons reached by DIS was 13.1 (95% CI: 8.8, 17.4) absolute percentage points higher than that observed among persons not reached by DIS. Twelve-month cumulative incidence of viral suppression was documented for 49.1% (1,154/2,350) of events among previously diagnosed PLWH without prevalent viral suppression. The aCID comparing twelve- month incident viral suppression among those reached vs. not reached by DIS was 6.7 (95% CI:

2.1, 11.3) absolute percentage points.

PrEP among HIV-Negative Persons

Of the 5,245 unique HIV-negative persons in our dataset, 1,788 (34.1%) were successfully linked to insurance claims data with any coverage in the period January 2010 –

December 2017. Among these persons, 441 HIV/syphilis network events occurred among 372 persons with pharmacy coverage at and during the six months prior to the event for analyses of prevalent PrEP use. No FTC/TDF prescriptions were excluded as treatment for HIV or HBV or as PEP. We identified prevalent PrEP use for 5.4% (24/441) of these network events. Our assessment of incident PrEP use was conducted with 411 network events among 360 persons who were not prevalent PrEP users and had pharmacy coverage during the sixty days prior to the event and continuously for six months after the event. No FTC/TDF prescriptions were excluded as treatment for HIV or HBV, while three prescriptions were excluded as PEP. Six-

month cumulative incidence of PrEP use following an HIV/syphilis network event was 4.1%

(17/411).

Small increases in prevalent and incident PrEP use were observed from 2013 to 2017, but PrEP use remained low throughout this period [Table 5.2]. Persons who self-identified as

White, non-Hispanic and who were diagnosed with early syphilis were overrepresented within the population matched to insurance claims data, particularly among those with prevalent or incident PrEP use. Prevalent PrEP users at the network event were older than those without prevalent PrEP use (median age = 34.4 vs. 28.2 years), while incident PrEP users were younger than those without (24.9 vs. 28.1 years). Most network events among HIV-negative persons (82.7%; N = 5,768) occurred among those indicated for PrEP under our surveillance- based definition, and most incident PrEP use (88%; 15/17) occurred among indicated persons.

Six-month cumulative incidence of PrEP use among indicated persons was 4.5% (15/337).

In a sensitivity analysis restricting PrEP analyses to persons with documented HIV negative status to estimate upper bounds for PrEP use, the size of the initial population for PrEP analyses was reduced from 6,975 to 5,787 events. After matching to insurance claims data,

44/441 (10.0%) persons were excluded from prevalent PrEP use analyses and 41/411 (10.0%) persons were excluded from incident PrEP uptake analyses. The prevalence of PrEP use was

5.8% and the six-month cumulative incidence of PrEP uptake was 4.3% in this restricted population. In a sensitivity analysis imputing condom use for events with missing values, we observed only a slight decrease in PrEP indication (79% vs. 83%).

D. Discussion

Combination ART and PrEP interventions with high levels of uptake and adherence are required to end the HIV epidemic,212,213 and sexual networks with prevalent HIV and syphilis are high-priority populations for such programming. However, current levels of coverage in these networks are not well characterized. To assess resource needs and inform intervention planning

in this population, we evaluated prevalent and incident viral suppression and PrEP use among

MSM diagnosed with early syphilis or named as sexual contacts of persons diagnosed with HIV or early syphilis in NC, finding low levels of all outcomes.

Presentation of viremic, previously diagnosed PLWH in the HIV/syphilis network is a critical opportunity to activate support for care engagement or re-engagement. An estimated

70% of HIV transmissions in the US138 (and 77% in NC)139 arise from contact with previously diagnosed PLWH, highlighting the importance of reaching this population with suppressive ART for clinical and public health benefits. In our population, 31% of all early syphilis diagnoses and

HIV/syphilis contact events occurring among previously diagnosed PLWH were among persons with unsuppressed viremia at the event and for at least six months thereafter. These events represent missed opportunities to leverage a syphilis diagnosis or HIV/syphilis contact event as a springboard to viral suppression.

Although estimated viral suppression incidence following a network event was low, existing public health services offer a potential route for improvement. We observed a significant increase in six-month cumulative incidence of viral suppression among persons reached by DIS in comparison to those who were not reached. DIS services may act as a “cue to action” for

ART use under the Health Belief Model132 by reinforcing potential health, transmission, and legal consequences associated with unsuppressed viremia, as well as a channel to care engagement services. Because the highest-priority responsibilities of DIS are to interview newly diagnosed persons and screen their contacts for infection, only 59% of previously diagnosed PLWH with unsuppressed viremia in this study population were reached by DIS. Support for higher coverage of DIS services among previously diagnosed PLWH may facilitate attainment of 2020

National HIV/AIDS Strategy targets of 90% care engagement and 80% viral suppression among diagnosed PLWH.214

PrEP is indicated for a large population on the basis of HIV exposure risk,54 but uptake has been poor. In the US, only an estimated 18% of the 1.2 million persons indicated for PrEP64

were prescribed PrEP in 2018.7 In NC, just an estimated 2,500 persons were prescribed PrEP in

2018.65 We observed very high levels of PrEP indication in our study population, but little PrEP use. However, prevalent and incident PrEP use trended upward from 2013 to 2017, reflecting increased PrEP awareness, interest, and access in this population since the introduction of

Truvada in 2012.37 Persons with incident PrEP use following a network event were substantially younger than prevalent PrEP users, suggesting that the contact tracing network may be a valuable pathway for reaching at-risk young MSM with PrEP.

NC and several other jurisdictions have recently expanded HIV prevention services to include referrals for PrEP during contact tracing.133 Pilot programs offering PrEP care or referrals out of select county health departments in NC are also underway,215 allowing assessment of these existing providers’ potential to reach underserved populations and address financial barriers to care. Although copay and medication assistance programs can reduce the cost of Truvada, these programs do not include provider care and laboratory tests,79 leaving a coverage gap that can be addressed by existing public health infrastructure.80 County health departments are also efficient PrEP access points for HIV/syphilis network members already engaged with public health personnel. Even if linkage to PrEP care is not immediately achieved, contact tracing and syphilis diagnoses offer opportunities to intervene on earlier stages of the

HIV PrEP continuum,35 including building PrEP awareness and support for PrEP use.

As demonstrated here, linkage of insurance claims and surveillance data offers a novel, useful approach to monitor and support PrEP use among key populations for HIV prevention.

Surveillance data allowed identification of persons at high risk of HIV exposure within their sexual networks and linkage to insurance claims data enabled assessment of PrEP use in this population. Linkage further allowed detection of erroneous ICD-9/10 CM codes for HIV diagnoses in persons receiving FTC/TDF, highlighting the value of triangulating multiple data sources to identify and correct misclassification.

Contact tracing data are subject to social desirability bias and recall error, preventing complete elucidation of the underlying network population.163 In addition, our analyses of DIS services and viral suppression may be subject to uncontrolled confounding by socioeconomic status and perceptions of public health services, which were unmeasured in our data source but could simultaneously affect both care-seeking behavior and being reached by DIS. We also note that some persons without a viral load measure in the year before the network event may have received care in another state; however, this number is likely small because we excluded persons without a NC address at the time of the network event. We were further unable to account for potential out-migration after a network event, potentially leading to underestimation of incident viral suppression. Despite mandatory viral load reporting to NC EDSS implemented in 2013, slow uptake by several providers delayed full statewide reporting until 2016 (Dr. Erika

Samoff, NC DPH, personal communication 2020), creating another potential source of error in viral suppression estimates. Error may also have arisen from our assumption that all persons with unknown HIV status at the time of the network event were HIV-negative. However, HIV testing occurred in association with most network events and would have resulted in a recorded diagnosis if positive. Additionally, a sensitivity analysis requiring a negative HIV test in the three months before, at, or ever following the event showed only small increases in PrEP use estimates.

Insurance claims data provide estimates of medication dispensing but may not accurately reflect medication consumption. We also required only a single PrEP prescription claim for prevalent or incident PrEP use, preventing evaluation of long-term use. Observed

PrEP use is generalizable only to network members with pharmacy coverage under a single private insurer in NC, a population which differs substantially from all HIV-negative network members; however, estimates may also be reasonably generalizable to the remainder of the privately insured population. Although uninsured and publicly insured individuals are critical populations for HIV prevention efforts, PrEP use in these populations could not be assessed

100

with available data. We note, however, that in our population of persons with high HIV acquisition risk and access to PrEP under insurance coverage – a population in which PrEP use might be expected to approach an upper bound – we observed very little PrEP use overall.

In summary, although many HIV-negative members of the HIV/syphilis contact tracing network were indicated for PrEP, prevalent and incident PrEP use were low. Viral suppression among previously HIV-diagnosed persons in the network was also sub-optimal. Incident viral suppression among previously HIV-diagnosed, viremic network members was significantly lower among persons not reached by DIS, suggesting a need for increased support for DIS services in this population. These findings may inform expansion of health department PrEP services and guide efforts to increase viral suppression as jurisdictions across the US join the campaign to end HIV.178

101

Figure 5.1 Populations for analysis of prevalent and incident viral suppression and pre-exposure prophylaxis (PrEP) use in the HIV/syphilis contact tracing networks

CAPTION: HIV/syphilis network events among adult, cisgender, North Carolina (NC)-residing men who have sex with men (MSM), January 2013 – June 2017, were identified from NC surveillance data. Previously diagnosed people living with HIV (PLWH) were evaluated for prevalent and incident viral suppression. HIV-negative persons were matched to insurance claims data and evaluated for prevalent and incident PrEP use. Some HIV-negative persons were eligible for both prevalent and incident PrEP use analyses. *Events among HIV-negative persons, defined as those with no HIV diagnosis on record in NC EDSS at any point before or in the 14 days after the event. +Events among previously diagnosed PLWH, defined as those with an HIV diagnosis on record in NC EDSS ≥30 days prior to the event. #Events among recently diagnosed PLWH, defined as those with an HIV diagnosis on record in NC EDSS <30 days prior to or ≤14 days following the event. ^Population for prevalent PrEP use analysis was restricted to events among persons with continuous pharmacy coverage over the six months prior to the network event. **Population for incident PrEP use analysis was restricted to events among persons with continuous pharmacy coverage over the sixty days prior to and the six months following the network event, excluding persons with prior PrEP use.

102

Table 5.1 Individual characteristics by HIV viral suppression at the time of an HIV/syphilis network event occurring January 2013 – June 2017 in North Carolina among men who have sex with men who were previously diagnosed with HIV.

Not virally suppressed Prevalent viral Individual characteristics suppression* Incident viral No incident viral suppression+ suppression Years since HIV diagnosis 5.5 (2.4, 9.8) 4.2 (0.8, 8.2) 4.8 (2.1, 8.6) (median, IQR) Age at network event 33.2 (27.6, 44.2) 29.8 (25.8, 38.3) 28.9 (24.9, 35.8) (median, IQR) Race/ethnicity (N, %) White, non-Hispanic 717 (27.5) 182 (21.9) 220 (14.5) Black, non-Hispanic 1690 (64.8) 582 (70.0) 1233 (81.2) Hispanic 144 (5.5) 44 (5.3) 41 (2.7) Other 38 (1.5) 12 (1.4) 6 (0.4) Missing 20 (0.8) 11 (1.3) 19 (1.2) Network event Syphilis diagnosis 1165 (44.6) 289 (34.8) 425 (28.0) HIV contact 537 (20.6) 260 (31.3) 555 (36.5) Syphilis contact 907 (34.8) 282 (33.9) 539 (35.5) Reached by DIS# 1951 (74.8) 564 (67.9) 833 (54.8) Year 2013 200 (7.7) 87 (10.5) 242 (15.9) 2014 410 (15.7) 158 (19.0) 297 (19.6) 2015 820 (31.4) 229 (27.6) 428 (28.2) 2016 818 (31.4) 266 (32.0) 376 (24.8) 2017++ 361 (13.8) 91 (10.9) 176 (11.6) Total network events 2609 831 1519 IQR; interquartile range * ≥1 viral load <200 copies/mL in the 12 months prior to the date of the syphilis diagnosis or the date named as an HIV or syphilis contact +≥1 viral load <200 copies/mL in the 6 months following the date of the syphilis diagnosis or the date named as an HIV or syphilis contact #Disease Intervention Specialist; refers to DIS contact after a network event, not the time of the previous HIV diagnosis ++January 1, 2017 – June 31, 2017

103

Table 5.2 Individual characteristics by pre-exposure prophylaxis (PrEP) use at the time of an HIV/syphilis network event and in the subsequent six months among HIV-negative men who have sex with men in North Carolina, January 2013 – June 2017.

All events among Prevalent PrEP use* Incident PrEP uptake+ Individual characteristics HIV-negative Yes No Yes No persons Age at event (median, IQR) 27.3 (23.2, 35.3) 34.4 (26.8, 39.6) 28.2 (23.1, 38.0) 24.9 (20.6, 32.2) 28.1 (23.2, 38.1) Race/ethnicity (N, %) White, non-Hispanic 2411 (34.6) 16 (66.7) 230 (55.2) 9 (52.9) 201 (51.0) Black, non-Hispanic 3783 (54.2) 5 (20.8) 153 (36.7) 5 (29.4) 161 (40.9) Hispanic 487 (7.0) <5# 16 (3.8) <5 14 (3.6) Other 124 (1.8) <5 12 (2.9) <5 14 (3.6) Missing 170 (2.4) <5 6 (1.4) <5 <5 Network event status Syphilis diagnosis 1839 (26.4) 9 (37.5) 123 (29.5) 8 (47.1) 113 (28.7) HIV contact 2235 (32.0) <5 120 (28.8) <5 117 (29.7)

Syphilis contact 2901 (41.6) 11 (45.8) 174 (41.7) 5 (29.4) 164 (41.6) 104 Reached by DIS 5951 (85.3) 23 (95.8) 378 (90.7) 16 (94.1) 357 (90.6)

Year 2013 938 (13.5) <5 43 (10.3) <5 46 (11.7) 2014 1282 (18.4) <5 88 (21.1) <5 94 (23.9) 2015 1968 (28.2) 6 (25.0) 125 (30.0) <5 98 (24.9) 2016 1827 (26.2) 11 (45.9) 101 (24.2) 9 (52.9) 94 (23.9) 2017** 960 (13.8) <5 60 (14.4) 6 (35.3) 62 (15.7) Insurance subscriber Self --- 23 (95.8) 342 (82.0) 13 (76.5) 327 (83.0) Parent --- <5 71 (17.0) <5 62 (15.7) Other --- <5 <5 <5 5 (1.3) Indicated for PrEP 5768 (82.7) 18 (75.0) 337 (80.8) 15 (88.2) 322 (81.7) Total 6975 24 417 17 394 IQR; interquartile range. DIS; Disease Intervention Specialist *Defined as ≥ one insurance claim for emtricitabine and tenofovir disoproxil fumarate (FTC/TDF) prescribed for ≥30 days in the six months prior to the date of the event +Defined as ≥ one insurance claim for emtricitabine and tenofovir disoproxil fumarate (FTC/TDF) prescribed for ≥30 days in the six months following the date of the event #Cell sizes <5 are not presented to prevent deductive identification. **January 1, 2017 – June 31, 2017

CHAPTER SIX: TRANSMISSION DYNAMICS AND CONTROL OF ENDEMIC INFECTION IN NETWORKS WITH STRONG COMMUNITY STRUCTURE

A. Introduction

The increasing availability of infectious disease contact and transmission network data offers new opportunities for development of network-based disease prevention strategies.162,216-

218 Network community structure, or the presence of clusters of individuals densely connected to each other and comparatively sparsely connected to the remainder of the network,12 is one of many factors that influences infection transmission over contact networks.13,14 Newly introduced infection in networks with weak community structure (many connections between communities) typically experience explosive growth, whereas emerging epidemics in networks with strong community structure (few connections between communities) tend to exhibit greater variance in final epidemic size, peak prevalence, and duration.13,14 However, the effect of community structure on endemic infection dynamics has not been assessed.

In networks with strong community structure, most edges occur within network communities.12 Consequently, most opportunities for infection transmission occur within communities. When endemic infection is unevenly distributed across communities in these networks, susceptible persons may thus bear differential infection risks and experience variable benefits from different types of prevention interventions. For instance, susceptible persons in network communities with high infection prevalence are more likely to be exposed to infection than those in low-prevalence communities, and thus prevention interventions that reduce susceptibility of uninfected persons may be most efficient when preferentially deployed to susceptible persons in high-prevalence communities. Conversely, in network communities with

105

low infection prevalence, preferential treatment of infected persons may allow for efficient protection of susceptible members of those communities.

Consideration of network community structure and both overall and community-level endemic infection prevalence may aid in selection of optimal intervention approaches and efficient allocation of limited public health resources. Using HIV transmission as an example, we examined infection incidence and intervention effects in simulated networks with strong community structure and unevenly distributed endemic infection across communities. Within this framework, we estimated: 1) the impact of endemic community-level infection prevalence on community-level infection incidence, and 2) the efficiency of prophylactic and treatment interventions under random intervention allocation vs. preferential allocation according to endemic community-level infection prevalence.

B. Methods

Network Generation

Following other studies of network structure,13,14 we used an empirical network as a starting point and test case for simulations. Network parameters were drawn from the largest connected component of the HIV and syphilis sexual contact tracing network among MSM in NC from 2015 through 2017. This network demonstrates strong modularity, or community structure, and was comprehensively characterized in Chapter 4. Empirical network parameters drawn from this network included the network size (N), modularity (Q), number of communities (K), and overall node degree, intracommunity node degree, and community size distributions.

We used these input parameters with the algorithm presented in Sah et al204 to generate

100 static networks as modular random graphs under a randomized hierarchical configuration model (HCM*), which preserves the empirically observed community size distribution and

106

distributions of node degrees within and between communities, but randomizes specific inter and intracommunity edges.14,204 The algorithm proceeds as follows:

1. Assign N nodes to K communities according to the community size distribution.

2. Assign overall degrees, d(vi), to each node vi according to the node degree distribution.

Assign intracommunity degrees (number of edges within the assigned

community), dw (vi), and intercommunity degrees (number of edges between

communities), db(vi), to each node vi.

3. Generate intercommunity edges using a modified Havel-Hakimi model and randomize

specific placement between nodes across communities.

4. Generate intracommunity edges using the Havel-Hakimi model and randomize specific

placement between nodes within communities.

We adapted publicly available scripts to generate networks204 in Python 3.7 (Python

Software Foundation, https://www.python.org) and identify network communities177,190 in

MATLAB 9.3 (The MathWorks Inc., Natick, MA). Simulated networks replicated critical features of our empirical source network: 1) network size (N) = 2,350 persons; 2) modularity (Q) = 0.96;

3) mean node degree = 2.1; 4) mean community size = 112 members. Additional details on input parameters, network generation, and community detection are provided in Chapter 3.

HIV Transmission Model

Simple transmission models are widely used to evaluate fundamental network properties and potential intervention strategies.13,14 We constructed a highly simplified HIV transmission model to simulate endemic infection spread across a sexual contact network, specifying a modified SI structure with one additional compartment for persons protected from HIV infection and another for persons who are rendered non-infectious via effective treatment [Figure 3.8].

Persons in the model may be “susceptible” to HIV (S), “protected” (P) via PrEP for HIV, infected and “infectious” (I), or “virally suppressed” (V) from highly adherent ART use. Transmission was

107

modeled within HIV-discordant dyads, or dyads in which one person is in state I and one person is in state S. Transmission could not occur from persons who were in state V or to persons who were in state P, reflecting an assumption of perfect intervention efficacy, and persons in these states could not revert to the I or S states, respectively. In each daily time step, transmission occurred within each discordant dyad with probability 1 − (1 − 푝)c where 푝 was the per-act transmission probability associated with condomless anal intercourse (CAI) between HIV- discordant partners and c was the act rate per partner per day. We simulated transmission within each discordant dyad in each time step using a draw from a Bernoulli distribution with a probability parameter equal to the daily transmission probability.

Act rate per partner per day and per-act transmission probability were drawn from the literature. We specified a fixed daily act rate per partner of 0.22, based on a large internet survey of US MSM in which the mean annualized number of sex acts with the most recent male sexual partner in the previous year was reported to be 80.6.205 We specified a fixed per-act transmission probability of 0.0076, representing the midpoint between meta-analysis estimates of 0.0011 and 0.0140 for receptive and insertive anal intercourse, respectively, in the absence of viral suppression, PrEP, and condom use.206,207

We first simulated a series of nine intervention-free scenarios in which we varied the starting overall network prevalence from 0.10 to 0.50 at increments of five percentage points.

For each overall endemic prevalence, we seeded each of the 100 generated networks with that prevalence at the start of each transmission simulation. To achieve this overall endemic prevalence while maximizing variability in prevalence across communities, we assigned unique prevalences between 0% and 100% to each community at time zero using random draws from a beta distribution with a mean equal to the specified overall endemic prevalence and a fixed α parameter of 2 [Figure 3.9]. Persons were then randomly sampled within each community and assigned to be in state I at the start of the simulation to fulfill assigned community-level endemic prevalences. No persons were initially seeded in the P or V states, nor did any persons move to

108

the P or V states over time (PrEP and ART assignment under varying intervention scenarios are described below).

Intervention Scenarios

We simulated two types of interventions: ART, which instantaneously moved persons from state I to state V at simulation start, and PrEP, which instantaneously moved persons from state S to state P at simulation start. For each intervention, we modeled three sets of allocation schemes: one in which interventions were distributed randomly (i.e., without regard to community structure), one in which interventions were distributed preferentially to persons in communities with higher endemic HIV prevalences (high-to-low allocation), and one in which interventions were distributed preferentially to persons in communities with lower endemic HIV prevalences (low-to-high allocation). Within each of these three allocation schemes, we modeled intervention coverage levels of 10%, 30%, and 50% for each intervention, for a total of six intervention scenarios (three for ART, three for PrEP) per allocation scheme. Under the random allocation scheme, 10%, 30%, or 50% of eligible persons (i.e., persons who would otherwise start in the S or I state) were randomly allocated at time 0 to PrEP or ART, depending on the intervention being modeled. Under the preferential allocation schemes, the probability of an eligible person receiving an intervention was linearly related (directly for the high-to-low

109

scheme and inversely for the low-to-high scheme) to the endemic HIV prevalence in that person’s community (see Chapter 3 for details). Preferential allocation schemes were probabilistic, rather than deterministic, to mirror the practicalities of public health interventions: although some persons or populations within a network may be prioritized for enhanced PrEP or viral suppression support, public health personnel will typically attempt to reach all relevant persons with interventions. Under such an approach, not all persons in prioritized populations will initiate PrEP or achieve viral suppression under any intervention, while some persons not prioritized will start PrEP or become virally suppressed.

Each of the six intervention scenarios (ART at 10%, 30%, and 50% coverage; PrEP at

We estimated overall intervention efficiency for each intervention scenario as the percentage reduction in the HIV incidence rate per person reached with the intervention, compared to the

110

mean incidence rate under the no-intervention model at the same overall endemic network prevalence.

We pursued two distinct objectives: 1) compare the effect and efficiency of ART vs.

We hypothesized that the impact of intervention allocation scheme on intervention effect and efficiency would decrease with increasing intervention coverage because differential sampling becomes increasingly similar to random sampling as greater proportions of eligible persons are sampled. Conversely, we hypothesized that the impact of intervention allocation scheme would increase with increasing overall endemic HIV prevalence because of heightened variation in community-level endemic prevalence under our infection seeding procedure. At low overall endemic prevalences, most communities have low endemic prevalences and very few have high endemic prevalences, creating little variation for a preferential allocation scheme to exploit. At high overall endemic prevalences, community-level endemic HIV prevalences are more normally distributed, and there are both communities with low and high endemic prevalences from which persons can be sampled for intervention under high-to-low or low-to- high allocation schemes.

111

C. Results

Community-level HIV incidence rates increased with endemic community-level HIV prevalences under our simplified HIV transmission model [Figure 6.1]. This relationship varied by overall endemic HIV prevalence, with the strongest effect observed at the lowest overall endemic prevalence tested. In intervention-free scenarios, estimated incident HIV infections per susceptible person-year within the full network ranged from 0.19 (95% CI: 0.19-0.20) to 0.64

(95% CI: 0.62, 0.66) at overall endemic HIV prevalences of 0.10 to 0.50, respectively. [Figure

6.2]. Modeled interventions reduced overall HIV incidence rates by 6.7% to 81.1% in comparison to no-intervention incidence rates.

We observed greater absolute and percentage reductions in the HIV incidence rate with

PrEP versus viral suppression interventions for each three-way combination of intervention allocation scheme, coverage level, and overall endemic HIV prevalence. However, greater absolute numbers of persons were required to be reached with PrEP than with viral suppression to achieve the same coverage level at overall endemic HIV prevalences < 0.50. When we compared interventions on the basis of efficiency, which accounted for variation in number of persons reached across scenarios, we found that viral suppression interventions were more efficient than PrEP interventions at low overall endemic HIV prevalences, with this relationship reversing as overall endemic prevalence increased [Figure 6.3]. PrEP became more efficient than ART at lower overall endemic prevalence levels in scenarios with lower (vs. higher) intervention coverage levels.

PrEP interventions showed divergent intervention effects by intervention allocation scheme: prioritization of susceptible persons in communities with highest initial HIV prevalences

(high-to-low allocation) demonstrated the greatest percentage reductions in HIV incidence rate and highest intervention efficiency across all overall endemic prevalences and intervention coverage levels. Intervention effects under high-to-low allocation were followed first by those

112

under random allocation, then by effects under low-to-high allocation. The impact of PrEP intervention allocation scheme increased with overall endemic HIV prevalence, with the greatest differences across allocation schemes observed at the highest tested overall endemic prevalence. This effect was most observable at 10% coverage of susceptible persons with PrEP and was dampened by heightened intervention coverage. In contrast, viral suppression interventions showed no significant differences in intervention effect or efficiency by allocation scheme.

D. Discussion

HIV and other sexually transmitted infections are spread through discrete contacts within sexual contact networks, with infection risk among network members depending on network position and individual characteristics.10,159-161 One network feature known to play a critical role in infection diffusion over infection-naïve networks is community structure, or modularity.13,14

However, despite strong community structure identified in sexual contact tracing networks,176 including in Chapter 4, the role of network community structure in transmission of endemic infections like HIV has not been previously examined. We assessed the impact of network community structure on HIV transmission and HIV prevention intervention efficiency, finding a strong association between endemic community-level infection prevalence and infection incidence rates – as well as clear relationships between intervention type, allocation scheme, and overall endemic infection prevalence – that could potentially be leveraged in the design of prevention interventions.

We observed increasing community-level HIV incidence rates with increasing endemic community-level HIV prevalences across a wide range of overall endemic HIV prevalences.

Strong community structure, as simulated here, allowed for heightened infection incidence within high prevalence vs. low prevalence communities. This effect was most observable at low overall endemic prevalences and decreased as overall endemic prevalence increased. At low

113

overall prevalences, infection is largely contained within a few communities with comparatively higher endemic infection prevalences, and many communities have little or no infection. Most edges within a highly modular network link persons within the same community, and transmission principally occurs over intracommunity edges within the few communities with comparatively high endemic HIV prevalences. As overall endemic network prevalence increases and more communities host substantial numbers of infections, transmission occurs more frequently over the dense intracommunity edges in each community. In addition, opportunities for infection to enter low-prevalence communities via intercommunity edges increase with the number of communities with substantial HIV burden.

Longstanding racial and ethnic disparities in HIV prevalence have been attributed to homophily in sexual partnerships,154 differences in progression through the HIV care continuum,219 potential for stigma-associated transmission risk behaviors among persons of color,220 and variation in frequencies of the CCR5 genotype by racial and ethnic background.221

However, in a recent agent-based model, these four factors accounted for only half of the disparity in HIV prevalence between black and white MSM in the United States.222 Homophily contributes to the formation of network communities with demographic commonalities among members,211 and strong community structure may have played a role in the establishment of historical disparities in HIV prevalence across racial and ethnic groups. However, most network transmission models do not explicitly model community structure.211 Generation of appropriate community structure in addition to local network properties in network transmission models may account for a portion of the previously unexplained disparity in HIV prevalence by race and ethnicity.

PrEP intervention efficiency was greatest under allocation scenarios prioritizing susceptible persons in high prevalence communities, as these persons were more likely than susceptible persons in lower prevalence communities to belong to discordant dyads and experience opportunities for HIV acquisition. This effect was most pronounced at high overall

114

endemic prevalences, likely because these networks displayed greater variation in community- level endemic prevalence under our infection seeding mechanism. At low overall endemic prevalences there was less variation in community-level endemic prevalence, dampening the utility of community-prevalence-based allocation schemes. The benefit of a high-to-low allocation scheme was also most observable at low PrEP coverage levels because most persons reached with PrEP under low coverage scenarios were members of high prevalence communities. Intervention sampling under the high-to-low and low-to-high allocation schemes more closely approximated sampling under the random allocation scheme as coverage increased, mitigating this effect.

We identified no clear efficiency gains from prioritizing PLWH for viral suppression by network community prevalence. We initially hypothesized that viral suppression interventions preferentially allocated to persons in low prevalence communities would be more efficient than viral suppression interventions under other allocation schemes because they could effectively remove endemic infection from some communities, largely protecting susceptible community members from exposure. However, all PLWH in a low prevalence community would need to be sampled for viral suppression intervention coverage to achieve this protection, which was extremely rare under our nondeterministic intervention assignment scheme. Although viral suppression interventions preferentially allocated to infected persons in low prevalence communities reached more persons in discordant dyads than were reached under other allocation schemes, some discordant dyads remained in nearly all communities. Susceptible persons thus still experienced the opportunity for transmission in dyads with persons who did not receive the intervention and in dyads with persons who became infected during the simulation.

This highly simplified analysis is a preliminary exploration of the role of community structure in endemic infection transmission and the potential impact of control efforts leveraging community structure. We generated sparse, static contact networks replicating some critical

115

features of the empirical HIV/syphilis contact tracing network among MSM in NC, and these networks are subject to the limitations governing contact network analysis described in Chapter

4.163 Static networks forced all partnerships to be concurrent and required partnership maintenance over the full study time frame, increasing network connectivity beyond that which would occur in a true contact network. Our empirical source network is also highly modular and results may not be generalizable to contact networks with lower modularity. In addition, we applied a highly simplified transmission model that does not account for medication adherence, maintenance in care, transmissibility by infection stage, condom use, positional preferences, and many other factors influencing HIV transmission.223 As a result of these limitations, estimated HIV incidence rates, intervention effects, and intervention efficiency do not reflect expected transmission within the empirical network from which we drew parameters. Rather, this model explores general trends in infection transmission and intervention utility in modular networks across a wide range of intervention scenarios.

Although this study was highly simplified and was not designed to provide actionable guidance about specific intervention approaches, our findings do provide some broad insights around network structure and intervention allocation that may be instructive from a public health perspective. Specifically, our results suggest that prioritization of susceptible persons in high- prevalence communities for enhanced interventions supporting PrEP uptake and maintenance may prevent more HIV infections per person reached than other PrEP allocation strategies. Viral suppression interventions may be more efficient than PrEP interventions under all allocation strategies in network populations with endemic HIV prevalences <20%. Future research will be useful in clarifying any relationship between network position and ART intervention allocation and elucidating more realistic estimates of intervention effects under various strategies. Many practical factors – such as relative costs of reaching sub-populations of network members with enhanced ART vs. PrEP interventions, heterogeneity in ART or PrEP uptake and adherence following intervention allocation, and feasibility of intervention scaling at high coverage levels –

116

will also be necessary considerations in any future translation of observed relationships to public health action.

We demonstrate a clear effect of endemic community-level infection prevalences on community-level incidence rates and assess the potential for community-based treatment and prevention interventions that incorporate network community structure. Consideration of network community structure in network models for transmission of endemic infections, including HIV, may improve existing models. Further study of network-based prevention interventions accounting for community structure and differential infection prevalences across communities under a more realistic model may inform new approaches to prioritize limited public health resources.

117

Figure 6.1 Log-transformed community-level HIV incidence rate by endemic community-level HIV prevalence and overall endemic network HIV prevalence under no intervention

CAPTION: Log-transformed community-level HIV incidence rate, defined as incident HIV infections per susceptible person-year, by endemic community-level HIV prevalence and overall endemic network HIV prevalence under no intervention. Points represent individual community- level HIV incidence rates in a single simulation. Overall endemic network HIV prevalence is indicated by color and linear trend lines in each color illustrate the relationship between endemic community-level HIV prevalence and community-level HIV incidence rate (log-transformed for visualization purposes).

118

Figure 6.2 Overall HIV incidence rate by initial overall endemic HIV prevalence, intervention allocation scheme, and intervention

coverage level

119

CAPTION: Overall HIV incidence rate, defined as mean incident HIV infections per susceptible or protected person-year, by initial overall endemic HIV prevalence, intervention allocation scheme, and intervention coverage level. The mean HIV incidence rate under a given scenario across 100 network simulations is represented as a point. Color indicates intervention allocation scheme, where individual nodes may be prioritized for intervention by endemic community-level HIV prevalence from highest to lowest prevalence, lowest to highest prevalence, or randomly. Point shape indicates intervention coverage among eligible persons, either protection via pre-exposure prophylaxis (PrEP) among susceptible persons or viral suppression of infected and infectious persons. Linear trend lines highlight the relationship between overall HIV incidence rate and overall endemic HIV prevalence within each intervention coverage level/allocation scheme pair. A. PrEP interventions. B. Viral suppression interventions.

Figure 6.3 Intervention efficiency by overall endemic HIV prevalence, intervention allocation scheme, and intervention coverage level

120 CAPTION: Intervention efficiency, defined as the mean percentage reduction in HIV incidence rate per person reached with the

intervention, by overall endemic HIV prevalence, intervention allocation scheme, and intervention coverage level. The mean

intervention efficiency under a given scenario across 100 simulations is represented as a point. Color indicates intervention allocation scheme, where individual nodes may be prioritized for intervention by endemic community-level HIV prevalence from highest to lowest prevalence, lowest to highest prevalence, or randomly. Point shape indicates intervention type, either protection via pre- exposure prophylaxis (PrEP) among susceptible persons or viral suppression of infected and infectious persons. Locally estimated scatterplot smoothing (LOESS) trend lines highlight the relationship between intervention efficiency and overall endemic HIV prevalence within each intervention coverage level/allocation scheme pair. A. Intervention efficiency at 10% coverage of eligible persons. B. Intervention efficiency at 30% coverage of eligible persons. C. Intervention efficiency at 50% coverage of eligible persons.

CHAPTER SEVEN: DISCUSSION

Existing public health infrastructure for HIV and syphilis contact tracing can serve as a conduit to reach HIV-positive and HIV-negative persons with enhanced ART and PrEP interventions, respectively. However, the relationship between HIV and syphilis contact networks and the potential utility of network-based interventions in this population is poorly understood. This dissertation aimed to analyze existing HIV and syphilis surveillance data in novel ways to assess the need for and potential added value of supplemental ART and PrEP interventions deployed through HIV and syphilis sexual contact tracing networks among MSM.

Innovative strategies are needed to increase ART and PrEP use for improved clinical and prevention outcomes to realize the objectives of “Ending the HIV Epidemic” campaigns nationwide.178

This dissertation pursued three specific aims: 1) evaluate interconnectivity of HIV and syphilis sexual contact networks among MSM in NC; assess viral suppression and PrEP use among MSM members of the HIV/syphilis sexual contact network in NC; and model the impact of treatment and prevention interventions in modular networks with differential endemic infection prevalence across network communities.

A. Summary of Findings

Aim 1

We generated independent and combined HIV and syphilis sexual contact networks among MSM, identified network communities, and assessed HIV and syphilis network interconnectivity both overall and by community. The syphilis network was interconnected with

121

the HIV network, with the magnitude of this interconnectivity varying across network communities. Heightened interconnectivity was associated with younger median age; higher proportions of persons self-identifying as Black, non-Hispanic; and higher proportions of persons diagnosed with syphilis diagnosed at STD clinics.

Strong network community structure is associated with explosive infection transmission within communities.175 The very strong community structure (Q = 0.97) detected in the combined HIV/syphilis network suggests that specific sub-populations within the syphilis network that tend to belong to communities that are particularly interconnected with the HIV network may be potential priorities for enhanced interventions to prevent HIV transmission.

Strong homophily by age and self-identified race/ethnicity leads to the formation of network communities with demographic commonalities. This community structure may then drive transmission within some demographic populations, contributing to the maintenance of longstanding disparities in HIV prevalence222 and HIV/syphilis coinfection107 among young,

Black MSM.

Aim 2

We used NC surveillance data to identify persons diagnosed with early syphilis or named as a recent sexual partner of a person diagnosed with HIV or early syphilis. We estimated prevalent and incident viral suppression among persons with a previous HIV diagnosis at and after the network event, respectively, and we assessed the effect of contact tracing services on six-month cumulative incidence of viral suppression among persons who were not virally suppressed at the time of the event. Prevalent viral suppression among persons previously diagnosed with HIV was 52.6%, and the six-month cumulative incidence of viral suppression was 35.4%. Although incident viral suppression following a network event overall was low, existing public health services offer a promising opportunity for improvement. We

122

observed a significant increase in six-month cumulative incidence of viral suppression among persons reached by DIS for interview in association with the network event in comparison to those who were not reached (aCID = 13.1; 95% CI: 8.9, 17.4 percentage points). Increased support for DIS services among previously-diagnosed PLWH with unsuppressed viremia identified in the HIV/syphilis contact tracing network may facilitate attainment of viral suppression targets in NC.

Using linked prescription claims data, we evaluated prevalent and incident PrEP use among HIV-negative persons diagnosed with early syphilis or named as recent sexual partners of persons diagnosed with HIV or early syphilis. Although most network events among HIV- negative persons occurred among persons indicated for PrEP, few HIV-negative persons showed prevalent (5.4%) or incident (4.1%) PrEP use. Persons with incident PrEP use following a network event were substantially younger than persons with prevalent PrEP use prior to a network event, suggesting that the contact tracing network may provide a useful mechanism to encourage PrEP use among young MSM. Linkage of insurance claims data and surveillance data offered a novel approach to evaluate PrEP use among a key population for HIV prevention, and extension of this process may support efforts to increase PrEP use in NC.

Aim 3

We generated sparse, static contact networks with strong community structure, replicating key features of the largest connected component of the combined HIV/syphilis contact network characterized in Aim 1. We simulated HIV transmission over these networks using a simple transmission model, allowing persons to be susceptible to HIV infection, protected via PrEP, infected and infectious, or virally suppressed via ART. We first evaluated the community-level HIV incidence rate by initial community-level HIV prevalence under no intervention, finding a strong association that was particularly pronounced at low overall endemic HIV prevalences.

123

We also estimated the effect and efficiency of community-prevalence-based vs. randomly allocated viral suppression and PrEP interventions across a range of initial overall network HIV prevalences and intervention coverage levels. Modeled interventions reduced HIV incidence rates across the full network by 6.7% to 81.1% relative to corresponding intervention- free scenarios. Viral suppression interventions were more efficient than PrEP interventions at low overall endemic HIV prevalences, while PrEP interventions were more efficient at high overall endemic HIV prevalences. PrEP interventions were most efficient when preferentially deployed to susceptible persons in communities with higher endemic HIV prevalences, especially at high overall endemic HIV prevalences and low intervention coverage levels. ART intervention efficiency did not vary meaningfully by intervention allocation scheme.

This study was highly simplified and was intended as a preliminary exploration of the role of community structure in endemic infection transmission and the potential impact of infection control efforts leveraging community structure. Further study of network-based prevention interventions accounting for community structure and differential infection prevalences across communities under a more realistic model may inform new approaches to prioritize limited public health resources.

B. Study Strengths

NC’s HIV and STI surveillance data are a valuable resource for research because each person is assigned a unique identifier that is universally applied across time, infections, and contact tracing experiences.179 In contrast, many other states and jurisdictions maintain separate identifiers for each infection and cannot not link contact tracing events to diagnosis events, even within a single infection. NC EDSS enables analyses like those performed in the course of this dissertation by allowing people to experience patterns of interaction with the public health system, rather than distinct events. Many people in NC have repeat interactions with NC DPH services, including HIV and/or syphilis diagnoses, contact tracing for HIV and/or

124

syphilis, and HIV and STI screening and treatment. We leveraged this rich data source in

Chapters 4 and 5 to conduct analyses that would not be possible in most other jurisdictions.

The analysis conducted in Chapter 4 is the first to characterize an integrated statewide

HIV/syphilis sexual contact network and to specifically interrogate how HIV and syphilis sexual contact networks interact. Few analyses have generated integrated networks containing persons diagnosed with HIV and syphilis and their named sexual contacts. All such analyses were restricted to small demographic or geographic sub-populations and primarily evaluated the union of the HIV and syphilis networks as a single network.10,136,169,170 We specifically sought to understand the relationship between the HIV and syphilis sexual contact networks among MSM in NC. Jaccard node and edge similarities were applied to estimate direct overlap between the

HIV and syphilis networks and largest connected component sizes in the independent and combined networks were identified to evaluate bridging between the network populations.

However, these methods could not fully quantify direct connectivity between two networks. We devised a simple new measure of interconnectivity between two partially overlapping networks, drawing from literature on interconnectivity within coupled, disjoint network systems.173 The new measure proposed here evaluates both the amount of overlap between nodes in each network and the distribution of this overlap. This measure may be applied to assess interconnectivity within a full network or within network subsets, such as communities.

We evaluate network community structure in Chapters 4 and 6 to assess the potential for differential HIV risk among HIV-negative persons within the combined HIV/syphilis network population. Although community detection is common in sociology, biology, and computer science,12 it has been applied to only a small number of empirical and simulated sexual contact networks.13,14,176,224 Community detection offers a useful mechanism to group persons into clusters within which most network edges,12 and thus most opportunities for transmission, occur.

As explored in Chapters 4 and 6, community detection could be a novel tool for identification of populations to be prioritized for prevention interventions under limited public health resources.

125

Although network community structure has repeatedly been shown to impact disease spread across a network, all prior studies of this effect have been confined to newly introduced epidemics in a naïve network population.13,14 We extended this work to examine the impact of community structure on the spread of an endemic infection that is unevenly distributed within a network, using HIV as an illustrative example.

HIV laboratory and prescription data are regularly monitored to evaluate HIV care engagement among PLWH under the Data to Care framework.102,225 We extended this framework by linking HIV/STI surveillance data and prescription claims data to assess PrEP use among HIV-negative MSM within the known HIV/syphilis sexual contact network in Chapter 5.

Surveillance data allowed us to identify persons at heightened risk of HIV exposure within their personal sexual networks and linkage to insurance claims data enabled assessment of prevalent and incident PrEP use in this vulnerable population. Routine linkage of prescription claims or prescribing data, including data covering populations not included in this analysis, and surveillance data may offer a novel approach to monitor and support PrEP use among key populations for HIV prevention.

C. Limitations

We chose to apply surveillance data in these analyses because NC EDSS offers the best representation of the HIV epidemic in North Carolina. However, surveillance data are also subject to several important limitations that required the use of strong assumptions in our analyses. NC EDSS captures all reported HIV and STI diagnoses and HIV-related laboratory values in NC, but not all persons seek testing or care at regular intervals. In addition, although viral load reporting to NC EDSS was mandated for all providers in NC in 2013, slow uptake by several providers delayed full statewide reporting until 2016 (Dr. Erika Samoff, NC DPH, personal communication 2020). Our estimates are reliant on these reported viral loads and likely underestimate prevalent and incident viral suppression, with the degree of underestimation

126

dependent on the year of the network event. We included network event year in our GEE model in Aim 2 to adjust for any potential confounding by this underreporting in analyses of the association between DIS interview status and incident viral suppression. Some persons may also have been misclassified in analyses of prevalent and incident viral suppression as having unsuppressed viremia if care was sought outside of NC or if no care was sought but ART adherence was maintained. In Chapter 5, we also assumed that all persons with unknown HIV status at the time of the network event were HIV-negative, interrogating this assumption in sensitivity analyses that produced similar results to those of our main analyses.

We acknowledge that the complete HIV and syphilis transmission networks in NC cannot be fully reconstructed from data on persons diagnosed with HIV and syphilis and the named contacts of these persons. Network membership is subject to selection bias because members are identified via self-reported characteristics and were restricted by diagnosed infection or contact status.163 Persons who are infected with HIV or syphilis but are undiagnosed could not be included in our analyses, and diagnosed persons not reached by DIS for interview were excluded from analyses in Chapter 4 because they could contribute no information for network generation. Further, contact tracing data were limited to only MSM sexual contacts across all analyses, ignoring the contributions of heterosexual and IDU-related contacts to the network.

We were unable to include marginal contacts, or persons for whom there were too little information for DIS to begin contact investigations. Contact data were also self-reported and may suffer from social desirability bias and recall error, leading to underreporting.163

We initially sought to characterize all PrEP use among HIV-negative persons diagnosed with syphilis or named as contacts of persons diagnosed with HIV or syphilis. However, PrEP data could only be obtained from a single private insurer in NC. Approximately one-third of all

HIV-negative network members matched to insurance claims data, but only 441 and 457 network events occurred among persons with continuous pharmacy coverage under BCBS NC over the relevant time frames for prevalent and incident PrEP use analyses, respectively.

127

Among matched persons, 309/1,788 (17%) were not covered by an insurance plan with pharmacy benefits, while the remainder of excluded persons did not have continuous insurance coverage. Estimates of prevalent and incident PrEP use are reasonably generalizable only to network members with pharmacy coverage under a private insurer, and this population differs in meaningful ways from the full population of HIV-negative HIV/syphilis network members.

Although uninsured and publicly-insured persons are now being reached by existing NC public health infrastructure with PrEP referrals and care, we were unable to evaluate PrEP use in these vulnerable populations with the available data.

We restricted all analyses to self-identified adult, cisgender MSM with a documented NC address at the time of their HIV or syphilis diagnosis or contact event. Although these factors limit generalizability beyond this specific population, our findings are highly relevant to the development of interventions deployed through HIV and syphilis contact tracing networks to support ART and PrEP use among MSM. Our networks are also retrospective and static, ignoring several important complexities of sexual partnerships and resulting sexual networks. In

Chapter 4, we aggregated sexual partnerships over a three year period. The resulting network displays an inflated mean node degree compared to the true network at any single point in time because our static network artificially created concurrency of all partnerships over the study time frame. This analysis also classified persons by network membership, rather than infection status, because of variability in diagnosed and undiagnosed infection during the study time frame. Assessment of HIV exposure among HIV-negative network members would require use of a dynamic network that allows formation and dissolution of partnerships and near-omniscient knowledge of infection status at all time points. Instead, we aimed to evaluate the HIV/syphilis syndemic from a network perspective by identifying individuals and populations within the syphilis network that are highly interconnected with the HIV network.

The largest connected component of the combined HIV/syphilis network identified in

Chapter 4 was then applied in Chapter 6 as a starting point for a simple network model. Forced

128

concurrency and maintenance of all relationships for the full duration of the simulation within this network may heighten transmission estimates produced in Chapter 6. This effect is in direct contrast to missing marginal and unreported contacts in the combined HIV/syphilis network. Our generated networks in Chapter 6 thus cannot be considered realistic sexual contact networks. In addition, we applied a highly simplified transmission model that does not account for medication adherence, maintenance in care, variations in transmissibility with infection stage, condom use, positional preference, and many other factors influencing HIV transmission. Estimated HIV incidence rates and intervention effects do not reflect expected transmission among MSM in the known HIV/syphilis sexual contact network in NC. Rather, this model aims to explore general trends in intervention efficiency in modular networks with endemic HIV infection.

D. Future Directions

Expanded Populations

We restricted analyses in Chapters 4 and 5 to adult, self-identified cisgender MSM diagnosed with HIV or early syphilis and their named sexual contacts. However, transgender persons, persons in heterosexual partnerships, and PWID are also critical populations for HIV prevention and contribute to the full transmission network. The analyses conducted in Chapter 4 could be expanded to include additional populations and transmission modes. One prior analysis identified distinct network communities in the Cuban HIV contact tracing network that were characterized by transmission risk group.176 Assessment of transmission risk group within network communities in the combined HIV/syphilis network in NC may provide additional insights into the HIV/syphilis syndemic within each sub-population and bridging between sub- populations.

In addition, all PLWH and persons at risk of acquiring HIV may benefit from increased support for viral suppression and PrEP use, respectively. The analyses conducted in Chapter 5

129

could be extended to all persons diagnosed with HIV or early syphilis and the named contacts of these persons in NC. Stratified analyses evaluating viral suppression and PrEP use prior to and following a network event by sex, self-identified transmission risk group, and self-identified gender identity may be useful for identification of gaps in ART and PrEP coverage within the combined HIV/syphilis network population. Increased understanding of inequities in care utilization within this population may help NC DPH design enhanced intervention strategies to aid populations with comparatively lower incidence of viral suppression or PrEP use following a network event. Further, stratified analyses assessing the effect of DIS interview on incident care utilization may highlight populations for which current DIS services are particularly effective or ineffective and contribute to the design of adapted programming to fit the needs of all populations in NC.

Routine Integration of PrEP Data

Our findings highlight very low prevalent and incident PrEP use among a population of persons at heightened risk of acquiring HIV with access to PrEP under their insurance coverage. Although NC DPH and partnering county health departments across NC have made great strides in increasing PrEP access for NC residents, NC does not currently measure PrEP use in a systematic manner. AIDS Vu provides estimates of PrEP use for each state using a nationally representative prescription database from Gilead linked to an insurance claims database.65 In 2018, AIDS Vu estimated that a minimum of 2,515 people in NC were prescribed

PrEP. This minimum estimate does not include all sources of PrEP prescriptions and excludes persons with FTC/TDF prescriptions that have no or insufficient insurance claims data (28%) to confirm FTC/TDF use for PrEP.65 The AIDS Vu estimate is a useful metric for NC DPH to track statewide PrEP use over time, but additional measures within specific populations of interest may provide more targeted information for intervention design.

130

Following the end of the study period for analyses in Chapter 5, NC EDSS introduced new data fields to record PrEP indication, awareness, and use among persons diagnosed with early syphilis or named as contacts of persons diagnosed with HIV or early syphilis (known

HIV/syphilis network members). Similar fields have also been added to NC Counseling, Testing, and Referral Services (CTRS) HIV test reporting forms. This data may be applied in conjunction with any available insurance claims data and prescribing records to regularly assess PrEP use within this critical population for HIV prevention. Specifically, NC EDSS data could be applied to identify the number of known HIV/syphilis network members within each geographic jurisdiction that are potentially indicated for PrEP. Indicated persons could then be matched to any available insurance claims data, including Medicare and Medicaid data, and PrEP prescribing records from NC county health departments and clinical partners to assess PrEP use in this population. If possible, this data may be supplemented with Gilead prescription assistance data or pharmacy dispensing data. Routine analyses could identify specific demographic and geographic subpopulations within the known HIV/syphilis contact network that require enhanced support for PrEP uptake and maintenance as NC expands PrEP distribution efforts, particularly in rural regions. Analogous analyses could also be conducted among persons testing for HIV captured in NC CTRS data.

Linkage with the data sources listed above may also facilitate additional analyses in this population, including efforts to assess the effect of insurance type (private, public, or uninsured) on PrEP uptake and maintenance. Linkage of prescribing and dispensing data may further allow for evaluation of the gap between prescribed medication and filled medication, allowing insights into a potential break-point in the HIV PrEP continuum.

Longitudinal Trajectories of HIV/STI Risk

Trajectories of HIV/STI-related events, including HIV testing documented in NC

Counseling, Testing, and Referral Services (CTRS) data, HIV and STI diagnoses (syphilis,

131

gonorrhea, or chlamydia) documented in NC EDSS, and being named as a sexual contact of a person diagnosed with HIV or early syphilis, could be analyzed using group-based trajectory models.36,224 Time zero would be defined as the date of the first documented negative HIV test,

STI diagnosis, or HIV/syphilis contact tracing event, depending on the starting population and outcome of interest. Longitudinal patterns of HIV/STI-related events could then be established, with each person in the analytical cohort having an estimated probability of “membership” in latent groups defined by each pattern. After identification of distinct trajectories, potential predictors of membership in consistently or increasingly high priority groups could be evaluated.

Time to important HIV/STI-related outcomes within each starting population could also be compared using survival analyses.

Data on incident HIV infection by trajectory could also be combined with individual demographic, risk behavior, and sexual contact network features to evaluate the likelihood of future HIV diagnosis among persons who test negative for HIV, are diagnosed with a reportable

STI, or are named as a contact of a person diagnosed with HIV or early syphilis. These data could be analyzed using a predictive model or machine learning algorithm to identify persons who are at greatest risk of HIV diagnosis and may experience the greatest benefits from enhanced support for PrEP linkage and adherence following interaction with public health programming.

Sexual and Injection Drug Use Network Interconnectivity

We evaluated interconnectivity between HIV and syphilis sexual contact networks to identify populations within the syphilis network that are more likely to be interconnected with the

HIV network. This type of analysis may be useful for assessment of other infections and network types, particularly for investigation of HIV and HCV. Sexual and IDU contact networks often partially overlap,227 and sexual network data from HIV and syphilis contact tracing investigations may offer insights into IDU contact networks. In 2018, NC DPH identified a putative HIV

132

outbreak among PWID in western NC. Of the 96 persons investigated in association with this outbreak with diagnosis data available, 15% were ever diagnosed with HIV and 44% were ever diagnosed with HCV. A majority of diagnosed PLWH identified in the investigation were coinfected with HCV (64%).228

Network interconnectivity could be applied to evaluate the relationship between sexual and IDU contact networks, rather than separate infection networks. These analyses would require extensive data on shared injecting equipment in a population that can be linked to HIV and syphilis contact tracing data. We would expect to identify some overlapping edges, known as multiplex edges, where some persons are connected via both sexual contact and shared injecting equipment. We would also expect to identify some overlapping nodes, where unique persons appear as IDU contacts of some persons and sexual contacts of other persons.

Network communities in which the sexual and injection drug use networks are highly interconnected may be potential bridges for introduction of HIV from the known HIV/syphilis sexual contact network to the IDU contact network, and HIV negative persons within these communities could be ideal recipients for PrEP. In addition, all persons within these network communities may benefit from HCV testing and subsequent treatment, if necessary.

Model Impact of Interventions Deployed through Contact Tracing Networks

The transmission model applied in Chapter 6 is a simplification of HIV transmission within a network with modularity and other network features drawn from the empirical contact tracing network generated in Chapter 4. However, as discussed previously, this network does not reflect the true underlying transmission network in NC. To model the potential impact of HIV prevention strategies deployed through the known HIV/syphilis contact tracing network with more realism, a different approach would be required. This model would begin with simulation of a large, dynamic sexual contact network using exponential random graph models208 that replicate empirical sexual behaviors and partnering within the MSM population in NC. We would

133

simulate HIV and syphilis transmission over this network and sample from infected persons to fit diagnosis rates in NC. Next, we would identify sexual contacts of these nodes with sampling reflecting contact tracing in NC and reconstruct contact tracing networks. Cumulative partnerships over time in the dynamic networks would be fit against an empirical contact tracing network to ensure that critical network features, including community structure, have been reproduced. Finally, we would simulate the deployment of ART and PrEP interventions to persons diagnosed with HIV or syphilis or named as contacts of persons diagnosed with HIV or syphilis.

A more realistic model would also allow for the introduction of additional features not included in our simple model presented in Chapter 6, including medication adherence, maintenance in care, variations in transmissibility with infection stage, condom use, positional preferences, and biological effects of syphilis infection on HIV acquisition and transmission. This type of interrogation could give credence to or disprove the findings of our model presented in

Chapter 6. It would also allow the intervention strategies explored in Chapter 6 to be tested against simpler strategies using demographic proxies for community-level interconnectivity or

HIV prevalence.

E. Public Health Implications

In an era where “Ending the HIV Epidemic” campaigns are being mounted across the

US,178 novel approaches are required to design and deploy interventions against persistent HIV incidence. Despite this lofty goal, resources for public health efforts are thinly stretched. New mechanisms are needed to identify populations at greatest risk of HIV transmission and acquisition, to reach these populations with HIV treatment and prevention interventions, and to monitor the success of this programming.

134

HIV and syphilis sexual contact networks are interconnected among MSM in NC, and the sexual network of persons diagnosed with HIV and syphilis and their named contacts offers an efficient platform for deployment of intensified HIV prevention interventions. Limited HIV prevention resources may be most efficiently utilized by leveraging existing contact tracing infrastructure to reach persons and populations within the HIV/syphilis network that are likely to acquire HIV. HIV-negative persons in populations within the syphilis network that are highly interconnected with the HIV network, including young MSM of color and persons diagnosed with syphilis at STI clinics, may experience heightened exposure to the HIV network and benefit disproportionately from enhanced PrEP interventions. In addition, we demonstrated a clear effect of community structure in transmission of endemic infection through a highly modular network and identified potential for community-based interventions to exploit this network structure. Further study of network-based interventions accounting for community structure and differential infection risk across communities may offer new avenues to prioritize limited public health resources for HIV prevention and control.

Contact tracing for HIV and syphilis provides a useful mechanism to identify previously diagnosed PLWH who are virally non-suppressed and may benefit from additional support for

HIV care linkage and maintenance. Although the six-month cumulative incidence of incident viral suppression following a network event was low, existing DIS services significantly increased this outcome. However, previously diagnosed PLWH named as contacts of persons diagnosed with HIV currently receive the lowest priority for allocation of limited DIS resources and are often not reached for DIS interview. Increased resources for DIS and SBC could allow these services to reach additional virally non-suppressed, previously diagnosed PWLH and support engagement or re-engagement in HIV treatment at this critical juncture in their care.

Despite high levels of PrEP indication among HIV-negative members of the HIV/syphilis contact tracing network, prevalent and incident PrEP use in this population were low. NC and several other jurisdictions have recently expanded HIV prevention services to include referrals

135

for PrEP by DIS during contact tracing.133 Pilot programs offering full-service PrEP care or PrEP referrals out of select county health departments in NC are also ongoing,75-78,215 allowing assessment of these existing providers’ potential to reach underserved populations and address financial barriers to care. These new programs offer critical pathways to enhance PrEP use among HIV negative HIV/syphilis network members specifically. Linkage of surveillance data with PrEP prescribing, insurance claims, and pharmacy dispensing data, as described previously, will be necessary to monitor the success of these offerings and to identify populations not adequately served by current programming.

F. Final Remarks

Contact tracing is a traditional tool in the public health arsenal because it allows for rapid identification of undiagnosed infection, interrupting onward transmission.124 However, the secondary benefits of the contact tracing process are also incredibly valuable: highlighting uninfected persons who may benefit from prevention interventions and identifying previously diagnosed persons who may require additional support for HIV care utilization. We used NC surveillance data to quantify a direct effect of DIS services for HIV and syphilis on incident viral suppression among virally non-suppressed, previously diagnosed PLWH identified in the

HIV/syphilis network, and to estimate the size of the HIV-negative and HIV-positive populations that may benefit from enhanced PrEP and viral suppression services delivered during contact tracing, respectively. We also assessed HIV and syphilis contact tracing data using community detection to identify populations within the syphilis network that may be ideal recipients of HIV prevention interventions.

Existing HIV treatment and prevention interventions have the potential to end HIV transmission in the US,212,213 but they cannot reach this potential without new intervention strategies that place populations with insufficient ART and PrEP use at the forefront. We must

136

find novel solutions that connect with people where they are and when they are able to engage with the public health system. New intervention strategies bypassing geographic, social, and economic barriers are needed to reduce inequities in HIV incidence and care utilization and reach all persons who may benefit from ART or PrEP. The existing public health infrastructure, including DIS, SBC, and publicly funded STD clinics housed at county and local health departments, offers many potential avenues for enhanced support for ART and PrEP use.

These services already reach critical populations for HIV prevention and treatment, and do so during HIV- and/or syphilis-related events that may act as cues to action for initiation of ART of

PrEP use. Increased support for and extended utilization of these frontline services can only strengthen the effort to end the HIV epidemic in NC.

137

APPENDIX: TRUVADA EXCLUSION CRITERIA

Table A.1 International Classification of Diseases, 9th and 10th revisions, Clinical Modification (ICD-9/10-CM) codes and National Drug Code (NDC) package codes applied to identify and exclude FTC/TDF prescribed as treatment for HIV or hepatitis B virus (HBV), or as post- exposure prophylaxis (PEP) against HIV. All NDC are formatted as 5-4-2 with dashes and leading zeros stripped.

Diagnosis or Code Code Type Prescription HIV diagnosis 042, 79.53, V08 ICD-9-CM HBV diagnosis 070.2, 70.3, 70.52, V02.61 ICD-9-CM HIV diagnosis B20, B97.35, Z21 ICD-10-CM HBV diagnosis B16.0, B16.1, B16.2, B16.9, B17.0, B18.0, B18.1, B19.10, B19.11 ICD-10-CM HIV treatment 50090098000, 70518034100, 53808020801, 15584010101, NDC prescription 61958110101, 50090124800, 70518067700, 50090172400, 3364111, package 50090172300, 59676057530, 70518071400, 53808088701, codes 50090131800, 70518067400, 61958120101, 49702020218, 53808089501, 50090061000, 70518184700, 61958060201, 61958060101, 173066200, 173066300, 49702020548, 49702020318, 49702020413, 50090055000, 53808089301, 70518069100, 50090087400, 49702020613, 49702021120, 49702021326, 49702021248, 53808099001, 49702021718, 87667117, 87667217, 87667317, 87667417, 87663241, 87663341, 87667117, 87667217, 87667317, 87667417, 61958040301, 50090075400, 61958040601, 61958040401, 70518054600, 61958040101, 61958040501, 3196401, 3196501, 3196601, 3196701, 3196801, 49702022118, 49702022248, 59676027801, 59676057001, 59676057201, 59676057101, 49702022517, 70518048200, 56047030, 56047492, 56051030, 597004660, 597004724, 597012330, 597000201, 597000302, 6057143, 6057362, 4024451, 4024515, 55289094712, 74052260, 70518009100, 74395646, 74679922, 74679930, 70518009102, 70518009101, 53808027601, 53808101001, 49702020718, 49702020853, 74333330, 74339930, 53808111901, 74194063, 50090116200, 59676056301, 59676056201, 59676056401, 59676056501, 50090149100, 70518148300, 59676056630, 70518068900, 3362212, 3362412, 3363112, 3363810, 50090158100, 50090100700, 70518072500, 70518068800, 70518046900, 63010001030, 63010002770, 55289047727, 4038140, 50090148600, 49702022318, 49702023508, 49702023755, 49702023308, 49702022418, 6022761, 6047361, 6047761, 6308001, 6360360, 6360361, 50090108501, 50090108500, 70518162101, 70518162102, 70518162100, 76519113006, 68071211306, 50090108502, 70518148700, 50090135501, 49702022813, 53808113001, 50090135500, 49702022713, 49702022613, 527194748, 70518200700, 54040713, 70518151900, 31722059730, 31722059712, 68382069606, 65862068701, 65862068730, 65862068799, 65862068705, 65162006132, 65162006106, 60687042025, 70518108600, 60687036425, 74308228, 74006301, 74006328, 74309301, 74309328, 59676080030, 59676080099, 50090160600, 49702023113, 49702024613, 49702024213 HBV treatment 3161112, 3161212, 3161412, 61958040401, 50090075400, NDC prescription 61958040601, 61958040301, 70518054600, 61958040501, package 68382092101, 68382092116, 68382092177, 70771102009, codes

138

52343014730, 52343014830, 60687021625, 68382092106, 51991089633, 43547043615, 69097042602, 31722083330, 31722083332, 31722083390, 31722083430, 31722083431, 31722083432, 70771101903, 70771102001, 65162044603, 65162044903, 42291026130, 65862084230, 49884054809, 70771101909, 16729038810, 49884054709, 49884054811, 43547043715, 43547043603, 43547043609, 51407006530, 65862084190, 65862084199, 70771102003, 70771102004, 51407006430, 65862084299, 16729038910, 49884054711, 68382092001, 68382092016, 51991089590, 70771101904, 43547043703, 70771101901, 68382092077, 31722083490, 69097042605, 33342036210, 33342036257, 33342036207, 69238154703, 69238154803, 69238154903, 69238152703, 31722053510, 31722053530, 42385090150, 53808112801, 60687036625, 69097053302, 69097053315, 60505466603, 65862081930, 70518100201, 65862081903, 65862081830, 70518215100, 65862082030, 42385090103, 31722053505, 31722053501, 31722053560, 65862042130, 42291080030, 65862081803, 65862082003, 16714082001, 93710456, 70518100200, 64380071404, 33342009607, 68180028806, 65862090030, 35573040230, 70518071500, 70771105309, 42291011530, 70710104903, 70771105303, 69097036202, 68180028809, 68180028833, 65862033530, 66993048230, 70710104909, 65862033510, 93538256, 68180028607, 68180028601, 68180028602, 33342036210, 33342036257, 33342036207, 64330000101, 66993047860, 65862002510, 60429035430, 70518206700, 70518207000, 60505325108, 60505325208, 60505325006, 53808084301, 65862002560, 65862002610, 64380071003, 64380071104, 33342000109, 33342000207, 65862055330, 68180060306, 50742062360, 50742062460, 65862005524, 65862055310, 904658304, 70518163400, 70518044001, 61958050101, 42794000308, 60505394703, 173066200, 173066300, 49702020318, 49702020548, 53808089301, 49702020413, 50090055000, 61958230101, 4035009, 4035730, 4036509, 4036530, 85113301, 85116801, 85435001, 85435101, 85435201, 85129701, 85129702, 85129703, 85131601, 85131602, 85132301, 85132302, 85137001, 85137002, 85137003, 85435301, 85435401, 85435501, 85435601 HIV post- Truvada (61958070401, 61958070501, 63629758102, 50090087000, NDC exposure 63629758101, 70518009700, 70518009702, 61958070101, package prophylaxis 35356007003, 61958070301, 70518009701, 68071211203, codes (PEP) 54868514100, 54569558800, 54569558802, 54569558803, 67296123703, 53808080501, 52959096903, 69189070101, 61919066902, 67296123704) + Raltegravir (6022761, 6047361, 6047761, 6308001, 6360360, 6360361, 50090108501, 50090108500, 70518162101, 70518162102, 70518162100, 76519113006, 68071211306, 50090108502) OR Truvada + Dolutegravir (49702024613, 49702024213, 50090135500, 49702022713, 49702022613, 70518148700, 50090135501, 49702022813, 53808113001, 50090160600, 49702023113); OR Truvada + Darunavir (70518071400, 50090172300, 59676057530, 59676056501, 70518148300, 59676056301, 59676056201, 50090149100, 70518068900, 59676056401, 59676056630, 59676080030, 59676080099) OR Truvada + Ritonavir (55289094712, 74052260, 70518009100, 74395646, 74679922, 74679930, 70518009102, 70518009101, 53808027601, 527194748,

139

53808111901, 50090116200, 74194063, 74339930, 74333330, 70518200700, 54040713, 70518151900, 31722059730, 31722059712, 68382069606, 65862068701, 65862068730, 65862068799, 65862068705, 65162006132, 65162006106, 60687042025, 70518108600, 60687036425, 74308228, 74006301, 74006328, 74309301, 74309328, 61958110101).

FTC/TDF prescription duration <30 days with ≥30 day fill gap between prescriptions and concomitant Raltegravir, Dolutegravir, Darunavir, or Ritonavir. One additional claim was treated as PEP with FTC/TDF prescribed for 10 days but no concomitant 3rd medication. HIV pre- Truvada (61958070401, 61958070501, 63629758102, 50090087000, NDC exposure 63629758101, 70518009700, 70518009702, 61958070101, package prophylaxis 35356007003, 61958070301, 70518009701, 68071211203, codes (PrEP) 54868514100, 54569558800, 54569558802, 54569558803, 67296123703, 53808080501, 52959096903, 69189070101, 61919066902, 67296123704).

Prescription duration ≥30 days or <30 days with <30 day fill gap between prescriptions, totaling ≥30 days. No concomitant Raltegravir, Dolutegravir, Darunavir, or Ritonavir.

140

REFERENCES

1. Singh S, Song R, Johnson AS, et al. HIV incidence, HIV prevalence, and undiagnosed HIV infections in men who have sex with men, United States. Ann Intern Med. 2018;168:685- 694.

2. Eaton JW, Johnson LF, Salomon JA, et al. HIV treatment as prevention: Systematic comparison of mathematical models of the potential impact of antiretroviral therapy on HIV incidence in South Africa. PLoS Med. 2012;9(7):e1001245.

3. Cohen MS, Chen YQ, McCauley M, et al. Prevention of HIV-1 Infection with early antiretroviral therapy. N Engl J Med. 2011;365:493-505.

4. Jenness SM, Goodreau SM, Rosenberg E, et al. Impact of the Centers for Disease Control's HIV preexposure prophylaxis guidelines for men who have sex with men in the United States. J Infect Dis. 2016;214(12):1800-1807.

5. Grant RM, Lama JR, Anderson PL, et al. Preexposure chemoprophylaxis for HIV prevention in men who have sex with men. N Engl J Med. 2010;363:2587-2599.

6. Beer L, Mattson CL, Bradley H, et al. Understanding cross-sectional racial, ethnic, and gender disparities in antiretroviral use and viral suppression among HIV patients in the United States. Medicine. 2016;95(13):e3171.

7. Harris NS, Johnson AS, Huang YA, et al. Vital Signs: Status of human immunodeficiency virus testing, viral suppression, and HIV preexposure prophylaxis – United States, 2013- 2018. Morb Mortal Wkly Rep. 2019;68(48):1117-1123.

8. O'Leary D. The syndemic of AIDS and STDs among MSM. Linacre Q. 2014;81(1):12–37.

9. Girometti N, Guiterrez A, Nwokolo N, et al. High HIV incidence in men who have sex with men following an early syphilis diagnosis: Is there room for pre-exposure prophylaxis as a prevention strategy? Sex Transm Infect. 2017;93(5):320-322.

10. Pasquale D. Epidemiological analysis of sociosexual HIV networks in central North Carolina [dissertation]. Chapel Hill, NC: University of North Carolina at Chapel Hill. 2018.

11. Dooley SW. Recommendations for partner services programs for HIV infection, syphilis, gonorrhea, and chlamydial Infection. Morb Mortal Wkly Rep. 2008;57(RR09);1-63.

12. Fortunato S. Community detection in graphs. Physics Reports. 2010;486(3-5):75-174.

13. Salathe M and Jones JH. Dynamics and control of diseases in networks with community structure. PLoS Comput Biol. 2010;6(4):e1000736.

14. Stegehuis C, van der Hofstad R, van Leeuwaarden JSH. Epidemic spreading on complex networks with community structures. Sci Rep. 2016;6:29748.

15. North Carolina HIV/STD/Hepatitis Surveillance Unit. 2018 North Carolina HIV Surveillance Report. https://epi.dph.ncdhhs.gov/cd/stds/figures/hiv18rpt_12172019.pdf. Accessed January 30, 2020.

141

16. Fenton KA and Valdiserri RO. Twenty-five years of HIV/AIDS - United States, 1981–2006. Morb Mortal Wkly Rep. 2006;55(21):585–589.

17. UNAIDS. Global AIDS Update. http://www.unaids.org/sites/default/files/media_asset/global-AIDS-update-2016_en.pdf. Accessed January 20, 2019.

18. Centers for Disease Control and Prevention. HIV Surveillance Report, 2018; 30. https://www.cdc.gov/hiv/pdf/library/reports/surveillance/cdc-hiv-surveillance-report-2018- vol-30.pdf. Accessed January 30, 2020.

19. Centers for Disease Control and Prevention. Estimated HIV incidence and prevalence in the United States, 2010–2015. HIV Surveillance Supplemental Report. 2018;23(1).

20. Centers for Disease Control and Prevention. HIV in the Southern United States. https://www.cdc.gov/hiv/pdf/policies/cdc-hiv-in-the-south-issue-brief.pdf. Accessed January 20, 2019.

21. Snider JT, Juday T, Romley JA. Nearly 60,000 uninsured and low-income people with HIV/AIDS live in states that are not expanding Medicaid. Health Aff. 2014;33(3):386–393.

22. Kahn, JG, Haile, B, Kates, J, Chang, S. Health and federal budgetary effects of increasing access to antiretroviral medications for HIV by expanding Medicaid. Am J Public Health. 2001;91(9):1464–1473.

23. The Kaiser Commission on Medicaid and the Uninsured. Key Facts about the Uninsured Population. http://files.kff.org/attachment/fact-sheet-key-facts-about-the-uninsured- population. Accessed January 20, 2019.

24. Gai Y, Marthinsen J. Medicaid expansion, HIV testing, and HIV-related risk behaviors in the United States, 2010-2017. Am J Pub Health. 2019;109(10):1404-1412.

25. Ingram L, Stafford C, Deming M, et al. A systematic mixed studies review of the intersections of social–ecological factors and HIV stigma in people living with HIV in the U.S. South. J Assoc Nurses AIDS Care. 2019;30(3):330-343.

26. Weis KE, Liese AD, Hussey J, et al. Associations of rural residence with timing of HIV diagnosis and stage of disease at diagnosis, South Carolina 2001-2005. J Rural Health. 2010;26(2):105-112.

27. Weissman S, Duffus WA, Vyavaharkar M, et al. Defining the rural HIV epidemic: Correlations of 3 definitions-South Carolina, 2005-2011. J Rural Health. 2013;30(3):275- 283.

28. Reif S, Golin CE, Smith SR. Barriers to accessing HIV/AIDS care in North Carolina: Rural and urban differences. AIDS Care. 2005;17(5):558-65.

29. Hall HI, Li J, and Mckenna MT. HIV in predominantly rural areas of the United States. J Rural Health. 2005;21(3):245-253.

142

30. Reif S, Whetten K, Wilson E. HIV/AIDS epidemic in the South reaches crisis proportions in last decade. Duke Center for Health Policy and Inequalities. 2012.

31. Heckman TG, Somlai AM, Peters J, et al. Barriers to care among persons living with HIV/AIDS in urban and rural areas. AIDS Care. 1998;10(3):365-375.

32. Ohl ME and Perencevich E. Frequency of human immunodeficiency virus (HIV) testing in urban vs. rural areas of the United States: Results from a nationally-representative sample. BMC Public Health. 2011;11(1):681.

33. Gonzalez A, Miller CT, Solomon SE, et al. Size matters: Community size, HIV stigma, & gender differences. AIDS Behav. 2008;13(6):1205-1212.

34. Bradley H, Hall H, Wolitski R, et al. Vital signs: HIV diagnosis, care and treatment among persons living with HIV – United States, 2011. Morb Mort Wkly Rep. 2014;63(47):1113- 1117.

35. Nunn A, Brinkley-Rubinstein L, Oldenburg C, et al. Defining the HIV pre-exposure prophylaxis care continuum. AIDS. 2017;31(5):731-734.

36. Powers KA, Samoff E, Weaver MA, et al. Longitudinal trajectories of HIV care retention in North Carolina. J Acquir Immune Defic Syndr. 2017;74(Suppl 2):S88-S95.

37. Food and Drug Administration. FDA approves first drug for reducing the risk of sexually acquired HIV infection. http://wayback.archive- it.org/7993/20170112032741/http://www.fda.gov/NewsEvents/Newsroom/PressAnnounce ments/ucm312210.htm. Accessed January 20, 2019.

38. Anderson PL, Glidden DV, Liu A, et al. Emtricitabine-tenofovir concentrations and pre- exposure prophylaxis efficacy in men who have sex with men. Science Trans Med. 2012;4(151):S122-129.

39. Knox D, Anderson P, Harrigan P, et al. Multidrug-resistant HIV-1 infection despite pre- exposure prophylaxis. N Engl J Med. 2017;376:501-502.

40. Hoornenborg E and de Bree G. Acute infection with a wild-type HIV-1 virus in PrEP user with high TDF levels [abstract]. Presented at 2016 HIV Research for Prevention, Chicago, USA, October 17-21, 2016.

41. Grossman H, Anderson P, Grant R, et al. Newly acquired HIV-1 infection with multi-drug resistant (MDR) HIV-1 in a patient on TDF/FTC-based PrEP [abstract]. Presented at 2016 HIV Research for Prevention, Chicago, USA, October 17-21, 2016.

42. Thaden J, Gandhi M, Kuncze K, et al. Seroconversion on PrEP: A protocol for untangling adherence vs resistance failure [abstract]. Presented at 2018 Conference on Retroviruses and Opportunistic Infections, Boston, USA, March 4-7, 2018.

43. Sullivan PS, Smith SK, Mera-Giler R, et al. The impact of pre-exposure prophylaxis with TDF/FTC on HIV diagnoses, 2012-2016, United States [abstract]. Presented at 2018 International AIDS Conference, Amsterdam, the Netherlands, July 23-27, 2018.

143

44. Grulich A, Guy RJ, Amin J, et al. Rapid reduction in HIV diagnoses after targeted PrEP implementation in NSW, Australia [abstract]. Presented at 2018 Conference on Retroviruses and Opportunistic Infections, Boston, USA, March 4-7, 2018.

45. Shen M, Xiao Y, Rong L, et al. The cost-effectiveness of oral HIV pre-exposure prophylaxis and early antiretroviral therapy in the presence of drug resistance among men who have sex with men in San Francisco. BMC Med. 2018;16:58.

46. Leech AA, Burgess JF, Sullivan M, et al. Cost-effectiveness of pre-exposure prophylaxis for HIV prevention for conception in the United States. AIDS. 2018;32(18):2787-2798.

47. Cambiano V, Miners A, Dunn D, et al. Cost-effectiveness of pre-exposure prophylaxis for HIV prevention in men who have sex with men in the UK: A modelling study and health economic evaluation. Lancet Infect Dis. 2018;18(1):85-94.

48. Gomez GB, Borquez A, Case KK, et al. The cost and impact of scaling up pre-exposure prophylaxis for HIV prevention: A systematic review of cost-effectiveness modelling studies. PLoS Med. 2013;10(3):e1001401.

49. Jenness SM, Weiss KM, Goodreau SM, et al. Incidence of gonorrhea and chlamydia following human immunodeficiency virus preexposure prophylaxis among men who have sex with men: A modeling study. Clin Infect Dis. 2017;65(5):712–718.

50. Grant RM and Koester KA. What people want from sex and preexposure prophylaxis. Curr Opin HIV AIDS. 2016;11(1):3–9.

51. Brooks RA, Kaplan RL, Lieber E, et al. Motivators, concerns, and barriers to adoption of preexposure prophylaxis for HIV prevention among gay and bisexual men in HIV- serodiscordant male relationships. AIDS Care. 2011;23(9):1136–1145.

52. Storholm ED, Volk JE, Marcus JL, et al. Risk perception, sexual behaviors, and PrEP adherence among substance-using men who have sex with men: A qualitative study. Prev Sci. 2017;18(6):737–747.

53. Matthews LT, Heffon R, Mugo NR, et al. High medication adherence during periconception periods among HIV-1–uninfected women participating in a clinical trial of antiretroviral pre- exposure prophylaxis. J Acquir Immune Defic Syndr. 2014;67(1):91–97.

54. US Public Health Service. Preexposure prophylaxis for the prevention of HIV infection in the United States—2017 update: A clinical practice guideline. https://www.cdc.gov/hiv/pdf/risk/prep/cdc-hiv-prep-guidelines-2017.pdf. Accessed February 20, 2019.

55. Traeger M, Schroeder S, Wright E, et al. Effects of pre-exposure prophylaxis for the prevention of human immunodeficiency virus infection on sexual risk behavior in men who have sex with men: A systematic review and meta-analysis. Clin Infect Dis. 2018;67(5):676-686.

56. Strobos J, Hauschild BC, Miller V. Safety considerations in the prevention of transmission of HIV by pre-exposure prophylaxis (or "PrEP"). Ann Forum Collab HIV Res. 2011;13(5):1- 11.

144

57. Supervie V, Garcia-Lerma JG, Heneine W, et al. HIV, transmitted drug resistance, and the paradox of preexposure prophylaxis. Proc Natl Acad Sci USA. 2010;107(27):12381– 12386.

58. Hurt CB, Eron JJ, Cohen MS. Pre-exposure prophylaxis and antiretroviral resistance: HIV prevention at a cost? Clin Infect Dis. 2011;53(12):1265–1270.

59. Peterson L, Taylor D, Roddy R, et al. Tenofovir disoproxil fumarate for prevention of HIV infection in women: A phase 2, double-blind, randomized, placebo-controlled trial. PLoS Clin Trials. 2007;2(5):e27.

60. Liu AY, Vittinghoff E, Sellmeyer DE, et al. Bone mineral density in HIV-negative men participating in a tenofovir pre-exposure prophylaxis randomized clinical trial in San Francisco. PLoS ONE. 2011;6(8):e23688.

61. Mulligan K, Glidden DV, Anderson PL, et al. Effects of emtricitabine/tenofovir on bone mineral density in HIV-negative persons in a randomized, double-blind, placebo-controlled trial. Clin Infect Dis. 2015;61(4):572-580.

62. Marcus JL, Buisker T, Horvath T, et al. Helping our patients take HIV pre-exposure prophylaxis (PrEP): a systematic review of adherence interventions. HIV Med. 2014;15(7):385–395.

63. Sidebottom D, Ekstrom A, Stromdahl S. A systematic review of adherence to oral pre- exposure prophylaxis for HIV – How can we improve uptake and adherence? BMC Infect Dis. 2018;18:581.

64. Smith DK, Van Handel M, Wolitski RJ, et al. Vital signs: estimated percentages and numbers of adults with indications for preexposure prophylaxis to prevent HIV acquisition — United States, 2015. Morb Mortal Wkly Rep. 2015;64(46):1291-1295.

65. AIDSVu. Local data: North Carolina. https://aidsvu.org/local-data/united- states/south/north-carolina/. Accessed October 6, 2019.

66. Bush S, Ng L, Magnuson D, et al. Significant uptake of Truvada for pre-exposure prophylaxis (PrEP) utilization in the US in late 2014 – 1Q2015 [abstract]. Presented at 10th International Conference on HIV Treatment and Prevention Adherence, Miami, USA, June 28-30, 2015.

67. Siegler AJ, Mouhanna F, Giler RM, et al. The prevalence of pre-exposure prophylaxis use and the pre-exposure prophylaxis–to-need ratio in the fourth quarter of 2017, United States. Ann Epidemiol. 2018;28(12):841-849.

68. Hoots BE, Finlayson T, Nerlander L, et al. Willingness to take, use of, and indications for pre-exposure prophylaxis among men who have sex with men—20 US cities, 2014. Clin Infect Dis. 2016;63(5):672-677.

69. Bush S, Magnuson M, Rawlings T, et al. Racial characteristics of FTC/TDF for pre- exposure prophylaxis users in the US [abstract]. Presented at ASM Microbe 2016, Boston, USA, June 16-20, 2016.

145

70. US Census Bureau. United States population growth by region. https://www.census.gov/popclock/data_tables.php?component=growth. Accessed January 16th, 2018.

71. Hess KL, Hu X, Lansky A, et al. Lifetime risk of a diagnosis of HIV infection in the United States. Ann Epidemiol. 2017;27(4):238–243.

72. Denning P and Dinenno E. Communities in crisis: Is there a generalized HIV epidemic in impoverished urban areas of the United States [abstract]? Presented at 27th International AIDS Conference, Vienna, Austria, July 18-23, 2010.

73. North Carolina AIDS Training and Education Center. North Carolina PrEP providers. http://www.med.unc.edu/ncaidstraining/files/PrEPProvidersforDownload.pdf. Accessed January 16, 2019.

74. Hoth AB, Shafer C, Dillon DB, et al. Iowa telePrEP: A public-health-partnered telehealth model for HIV pre-exposure prophylaxis (PrEP) delivery in a rural state. Sex Transm Dis. 2019;46(8):507-512.

75. Cabarrus Health Alliance. PrEP: Pre-Exposure prophylaxis to prevent HIV. https://www.cabarrushealth.org/375/PrEP-for-HIV-Prevention. Accessed January 28, 2019.

76. Orange County Health Department. PrEP – Pre-exposure prophylaxis for high-risk individuals. https://www.orangecountync.gov/1988/PrEP. Accessed January 28, 2019.

77. Mecklenburg County Public Health. PrEP. https://www.mecknc.gov/HealthDepartment/CDControl/HIVSTDSurveillance/Pages/PrEP.a spx. Accessed January 28, 2019.

78. Durham County Department of Public Health. PrEP DCODPH. http://dcopublichealth.org/services/std-hiv-testing/prep-dcodph. Accessed January 29, 2019.

79. Gilead. How can you get help paying for Truvada for PrEP? https://www.truvada.com/how- to-get-truvada-for-prep/truvada-cost. Accessed February 20, 2019.

80. Florida Department of Health Communications. On world AIDS day, Florida Health continues statewide fight against HIV/AIDS. http://www.floridahealth.gov/newsroom/2017/12/120117-world-aids-day.html. Accessed January 29, 2019.

81. Dieffenbach CW and Fauci AS. Thirty years of HIV and AIDS: Future challenges and opportunities. Ann Int Med. 2011;154(11):766-771.

82. Gunthard H, Saag M, Benson C, et al. Antiretroviral drugs for treatment and prevention of HIV infection in adults: 2016 recommendations of the International Antiviral Society – USA Panel. JAMA. 2016;316(2):191-210.

83. World Health Organization. Consolidated guidelines on the use of antiretroviral drugs for treating and preventing HIV infection: 2016 recommendations for a public health approach.

146

http://apps.who.int/iris/bitstream/handle/10665/208825/97892?sequence=1. Accessed January 17, 2019.

84. Panel on Antiretroviral Guidelines for Adults and Adolescents. Guidelines for the use of antiretroviral agents in HIV-1-infected adults and adolescents. https:// aidsinfo.nih.gov/contentfiles/lvguidelines/adultandadolescentgl.pdf. Accessed December 22, 2019.

85. Cohen MS, Chen YQ, McCauley M, et al. Antiretroviral therapy for the prevention of HIV-1 transmission. N Engl J Med. 2016;375:830-839.

86. Bavinton B, Pinto A, Phanuphak N, et al. Viral suppression and HIV transmission in serodiscordant male couples: An international, prospective, observational, cohort study. Lancet HIV. 2018;5(8):e438-e447.

87. Rodger A, Cambiano V, Brunn T, et al. Sexual activity without condoms and risk of HIV transmission in serodifferent couples when the HIV-positive partner is using suppressive antiretroviral therapy. JAMA. 2016;316(2):171-181.

88. Rodger A, Cambiano V, Brunn T, et al. Risk of HIV transmission through condomless sex in MSM couples with suppressive ART: The PARTNER2 study extended results in gay men [abstract]. Presented at 2018 International AIDS Conference, Amsterdam, the Netherlands, July 23-27, 2018.

89. Cohen M. Treatment for HIV prevention, one couple at a time. Lancet HIV. 2018;5(8):e408-e409.

90. The Lancet HIV. U=U taking off in 2017. Lancet HIV. 2017;4:e475.

91. UNAIDS. Global HIV & AIDS statistics — 2019 fact sheet. https://www.unaids.org/sites/default/files/media_asset/UNAIDS_FactSheet_en.pdf. Accessed January 30, 2020.

92. HIV/STD/Hepatitis Surveillance Unit. HIV care outcomes in North Carolina, 2017. https://epi.publichealth.nc.gov/cd/stds/figures/factsheet_HIV_care_outcomes_2017_rev2.p df. Accessed January 20, 2019.

93. Yehia BR, Rebeiro P, Althoff KN, et al. The impact of age on retention in care and viral suppression. J Acquir Immune Defic Syndr. 2015;68(4):413-419.

94. Bullo M, Samoff E, Dzialowy-Adams N, et al. HIV viral suppression assessed by two metrics in North Carolina during 2016 [abstract]. Presented at 2018 Conference on Retroviruses and Opportunistic Infections, Boston, USA, March 4-7, 2018.

95. Gardner E, McLees M, Steiner J, et al. The spectrum of engagement in HIV care and its relevance to test-and-treat strategies for prevention of HIV infection. Clin Infect Dis. 2011;52:793-800.

96. North Carolina Division of Health and Human Services. HIV medication assistance program. https://epi.publichealth.nc.gov/cd/hiv/hmap.html. Accessed January 17, 2019.

147

97. Napravnik S, Eron J, McKaig R, et al. factors associated with fewer visits for HIV primary care at a tertiary care center in the southeastern U.S. AIDS Care. 2006;18(Suppl 1):S45- S50.

98. Giordano T. Retention in HIV care: What the clinician needs to know. Top Antivir Med. 2012;19:12-16.

99. Rajabiun S, Mallinson R, McCoy K, et al. Getting me back on track: The role of outreach interventions in engaging and retaining people living with HIV/AIDS in medical care. AIDS Patient Care STDS. 2007;21(Suppl 1):S20-S29.

100. Sena A, Donovan J, Swygard H, et al. North Carolina HIV bridge counselor program: Outcomes from a statewide intervention to link and reengage HIV-infected persons in care in the South. J Acquir Immune Defic Syndr. 2017;76(1):e7-e14.

101. Swygard H, Sena A, Mobley V, et al. Implementation of the North Carolina HIV bridge counseling program to facilitate linkage and reengagement in care for individuals infected with HIV/AIDS. N C Med J. 2018;79(4):210-217.

102. Centers for Disease Control and Prevention. Data to care: Using HIV surveillance data to support the HIV care continuum. https://effectiveinterventions.cdc.gov/docs/default- source/data-to-care-d2c/pdf-of-important-considerations.pdf. Accessed January 28, 2019.

103. Zetola N and Klausner J. Syphilis and HIV infection: An update. Clin Infect Dis. 2007;44(9):1222–1228.

104. Centers for Disease Control and Prevention. 2015 sexually transmitted diseases treatment guidelines. https://www.cdc.gov/std/tg2015/hiv.htm. Accessed January 16th, 2018.

105. Purcell DW, Johnson CH, Lansky A, et al. Estimating the population size of men who have sex with men in the united states to obtain HIV and syphilis rates. Open AIDS J. 2012;6(1):98–107.

106. Centers for Disease Control and Prevention. Sexually transmitted disease surveillance, 2015. http://www.cdc.gov.libproxy.lib.unc.edu/std/stats15/std-surveillance-2015-print.pdf. Accessed November 7, 2017.

107. Su JR, Zaidi AA, Torrone EA, et al. HIV co-infection among individuals with primary and secondary syphilis – U.S., 2013 [abstract]. Presented at 2015 Conference on Retroviruses and Opportunistic Infections, Seattle, USA, February 23-26, 2015.

108. Rottingen JA, Cameron WD, Garnett GP. A systematic review of the epidemiologic interactions between classic sexually transmitted diseases and HIV. Sex Transm Dis. 2001;28(10):579–597.

109. Solomon MM and Mayer KH. Evolution of the syphilis epidemic among men who have sex with men. Sex Health. 2016;12(2):96-102.

110. Heffelfinger JD, Swint EB, Berman SM, et al. Trends in primary and secondary syphilis among men who have sex with men in the United States. Am J Public Health. 2007;97(6):107-1083.

148

111. Roberts C and Klausner J. Global challenges in human immunodeficiency virus and syphilis co-infection among men who have sex with men. Expert Rev Anti Ther. 2016;14(11):1037-1046.

112. Centers for Disease Control and Prevention. 2016 sexually transmitted disease surveillance: syphilis. https://www.cdc.gov/std/stats16/Syphilis.htm. Accessed January 17th, 2018.

113. De Voux A, Kidd S, Grey JA, et al. State-specific rates of primary and secondary syphilis among men who have sex with men — United States, 2015. Morbid Mortal Weekly Rep. 2017;66(13):349-354.

114. North Carolina HIV/STD/Hepatitis Surveillance Unit. 2018 North Carolina STD surveillance report. https://epi.dph.ncdhhs.gov/cd/stds/figures/std18rpt_02102020.pdf. Accessed January 30, 2020.

115. North Carolina HIV/STD Surveillance Unit. Syphilis and HIV in North Carolina: Reported syphilis and HIV case data, 2014. https://epi.publichealth.nc.gov/cd/stds/figures/factsheet_syphilis_and_hiv_coinfection_201 4r3.pdf. Accessed January 17th, 2018.

116. Samoff E, Cope A, Maxwell J, et al. The number of interviews needed to yield new syphilis and human immunodeficiency virus cases among partners of people diagnosed with syphilis, North Carolina, 2015. Sex Transm Dis. 2017;44(8):451-456.

117. Galvin SR and Cohen MS. The role of sexually transmitted diseases in HIV transmission. Nat Rev Microbiol. 2004;2(1):33-42.

118. Mayer KH and Venkatesh KK. Interactions of HIV, other sexually transmitted diseases, and genital tract inflammation facilitating local pathogen transmission and acquisition. Am J Reprod Immunol. 2011;65(3):308–316.

119. Solomon MM, Mayer KH, Glidden DV, et al. Syphilis predicts HIV incidence among men and transgender women who have sex with men in a preexposure prophylaxis trial. Clin Infect Dis. 2014;59(7):1020–1026.

120. Pathela P, Braunstein SL, Blank S, et al. The high risk of an HIV diagnosis following a diagnosis of syphilis: A population-level analysis of New York city men. Clin Infect Dis. 2015;61(2):281–287.

121. Katz DA, Dombrowski JC, Bell TR, et al. HIV incidence among men who have sex with men following diagnosis with sexually transmitted infections. Sex Transm Dis. 2017;43(4):249-254.

122. Cowan FM, French R, Johnson AM. The role and effectiveness of partner notification in STD control: A review. Sex Transm Infect. 1996;72(4):247–252.

123. Laar AK, DeBruin DA, Craddock S. Partner notification in the context of HIV: An interest- analysis. AIDS Res Ther. 2015;12(1):

149

124. European Centre for Disease Prevention and Control. Public health benefits of partner notification for sexually transmitted infections and HIV. https://ecdc.europa.eu/sites/portal/files/media/en/publications/Publications/Partner- notification-for-HIV-STI-June-2013.pdf. Accessed January 20, 2019.

125. World Health Organization. Guidelines on HIV self-testing and partner notification: Supplement to consolidated guidelines on HIV testing services. https://www.who.int/hiv/pub/vct/Annex16.pdf. Accessed January 18, 2019.

126. Golden MR, Hogben M, Handsfield HH, et al. Partner notification for HIV and STD in the United States: Low coverage for gonorrhea, chlamydial infection, and HIV. Sex Transm Dis. 2003;30(6):490-496.

127. North Carolina Division of Health and Human Services. Rule 10A NCAC 41A .0202.

128. North Carolina Division of Health and Human Services. Rule 10A NCAC 41A .0204.

129. Centers for Disease Control and Prevention. Description of the North Carolina data to care initiative (use of HIV surveillance data for HIV care engagement and re-engagement activities). https://effectiveinterventions.cdc.gov/docs/default-source/data-to-care-d2c/nc- data-to-care_sf.pdf?sfvrsn=0. Accessed January 20, 2019.

130. Hogben M, McNally T, McPheeters M, et al. The effectiveness of HIV partner counseling and referral services in increasing identification of HIV-positive individuals. Am J Prev Med. 2007;33(2):S89–S100.

131. Centers for Disease Control and Prevention. Monitoring selected national HIV prevention and care objectives by using HIV surveillance data – United States and 6 dependent areas, 2016. HIV Surveillance Supplemental Report. 2018;23(4).

132. Jan NK and Becker MH. The health belief model: A decade later. Health Educ Behav. 1984;11(1):1-47.

133. Katz DA, Dombrowski JC, Barry M, et al. STD partner services to monitor and promote HIV pre-exposure prophylaxis use among men who have sex with men. J Acquir Immune Defic Syndr. 2019;80(5):533-541.

134. Truss K. Partner services PrEP delivery model in Tennessee. Presentation given at: University of Washington Public Health Capacity Building Center: HIV Technical Cooperation Group. http://depts.washington.edu/hivtcg/presentations/uploads/62/partner_services_prep_delive ry_model_in_tennessee.pdf. Accessed January 18th, 2018.

135. Katz DA, Barry M, Dombrowski JC, et al. Integrating PrEP referrals Into STD partner services increases PrEP use among MSM [abstract]. Presented at 2018 Conference on Retroviruses and Opportunistic Infections, Boston, USA, March 4-7, 2018.

136. Hurt CB, Beagle S, Leone PA, et al. Investigating a sexual network of black men who have sex with men: Implications for transmission and prevention of HIV infection in the United States. J Acquir Immune Defic Syndr. 2012;61(4):515-521.

150

137. Schneider JA. Partner services pre-exposure prophylaxis (PS-PrEP). ClinicalTrials.gov. Bethesda (MD): National Library of Medicine (US). Identifier NCT02749955. https://clinicaltrials.gov/ct2/show/NCT02749955. Accessed January 20, 2019.

138. Skarbinski J, Rosenberg E, Paz-Bailey G, et al. Human immunodeficiency virus transmission at each step of the care continuum in the United States. JAMA Int Med. 2015;175(4):588-596.

139. Cope AB, Powers KA, Kuruc JD, et al. Ongoing HIV transmission and the HIV care continuum in North Carolina. PLoS ONE. 2015;10(6):e0127950.

140. Golden MR, Katz DA, Dombrowski JC. Modernizing field services for HIV and sexually transmitted infections in the United States. Sex Transm Dis. 2017;44(10):599-607.

141. Moreno JL and Jennings HH. Who shall survive: A new approach to the problem of human interrelations. Washington, DC: Nervous and Mental Disease Pub. Co; 1934.

142. Doherty IA, Padian NS, Marlow C, et al. Determinants and consequences of sexual networks as they affect the spread of sexually transmitted infections. J Infect Dis. 2005;191(S1):S42–S54.

143. Wasserman S and Faust K. Social network analysis: Methods and applications. New York: Cambridge University Press; 1994.

144. Parran T and Vonderlehr RA. Plain words about venereal disease. New York: Reynal & Hitchcock; 1941.

145. Auerbach DM, Darrow WW, Jaffe HW, et al. Cluster of cases of the acquired immune deficiency syndrome: patients linked by sexual contact. Am J Med. 1984;76(3):487-492.

146. Klovdahl AS. Social networks and the spread of infectious diseases: The AIDS example. Soc Sci Med. 1985;21(11):1203-1216.

147. Amirkhanian YA. Social networks, sexual networks and HIV risk in men who have sex with men. Curr HIV/AIDS Rep. 2014;11(1):81-92.

148. Luke DA and Harris JK. Network analysis in public health: History, methods, and applications. Ann Rev Public Health. 2007;28:69-93.

149. Morris M and Kretzschmar M. Concurrent partnerships and the spread of HIV. AIDS. 1997;11:641-648.

150. Watts C and May R. The influence of concurrent partnerships on the dynamics of HIV/AIDS. Math Biosci. 1992;108:89-104.

151. Kretzschmar M. Sexual network structure and sexually transmitted disease prevention: A modeling perspective. Sex Transm Dis. 2000;27:627-635.

152. Liljeros F, Edling CR, Nunes Amaral LA. Sexual networks: Implications for the transmission of sexually transmitted infections. Microbes Infect. 2003;5:189-196.

151

153. Garnett GP, Hughes JP, Anderson RM, et al. Sexual mixing patterns of patients attending sexually transmitted diseases clinics. Sex Transm Dis. 1996;23:248-257.

154. Mustanski B, Morgan E, D’Aquila R, et al. Individual and network factors associated with racial disparities in HIV among men who have sex with men. J Acquir Immune Defic Syndr. 2019;80(1):24-30.

155. Morris M, Kurth AE, Hamilton DT, et al. Concurrent partnerships and HIV prevalence disparities by race: Linking science and public health practice. Am J Public Health. 2009;99(6):1023-1031.

156. Hobbs M, Owen S, Gerber L. Liquid love? Dating apps, sex, relationships and the digital transformation of intimacy. J Sociol. 2017;53(2):271-284.

157. Xu J, Yu H, Tang W, et al. The effect of using geosocial networking apps on the HIV incidence rate among men who have sex with men: Eighteen-month prospective cohort study in Shenyang, China. J Med Internet Res. 2018;20(12):e11303.

158. McPherson M, Smith-Lovin L, Cook JM. Birds of a feather: Homophily in social networks. Ann Rev Sociol. 2001;27:415-444.

159. Friedman SR, Neaigus A, Jose B, et al. Sociometric risk networks and risk for HIV infection. Am J Public Health. 1997;87(8):1289-1296.

160. Kottiri BJ, Friedman SR, Neaigus A, et al. Risk networks and racial/ethnic differences in the prevalence of HIV infection among injection drug users. J Acquir Immune Defic Syndr. 1999;30(1):95-104.

161. Morris M. Concurrent partnerships and syphilis persistence: New thoughts on an old puzzle. Sex Transm Dis. 2001;28:504-507.

162. Eames K, Bansal S, Frost S, et al. Six challenges in measuring contact networks for use in modelling. Epidemics. 2015;10:72-77.

163. Ghani AC, Donnelly CA, Garnett GP. Sampling biases and missing data in explorations of sexual partner networks for the spread of sexually transmitted diseases. Stat Med. 1998;17(18):2079-2097.

164. Smith JA and Moody J. Structural effects of network sampling coverage I: Nodes missing at random. Soc Networks 2014; 35(4).

165. Smith JA, Moody J, Morgan JH. Network sampling coverage II: The effect of non-random missing data on network measurement. Soc Networks. 2017;48:78-99.

166. Brewer DD, Garrett SB, Kulasingam S. Forgetting as a cause of incomplete reporting of sexual and drug injection partners. Sex Transm Dis. 1999;26(3):166-176.

167. Brewer DD and Garrett SB. Evaluation of interviewing techniques to enhance recall of sexual and drug injection partners. Sex Transm Dis. 2001;28(11):666-677.

152

168. Bell DC, Belli-McQueen B, Haider A. Partner naming and forgetting: Recall of network members. Soc Networks. 2007;29(2):279-299.

169. Potterat JJ, Phillips-Plummer L, Muth SQ, et al. Risk network structure in the early epidemic phase of HIV transmission in Colorado Springs. Sex Transm Infect. 2002;78:i159-i163.

170. Doherty I, Pasquale D, Liebman A, et al. P1-S5.39 Triple jeopardy? Sexual networks of syphilis & HIV infection among heterosexuals, MSM, and bisexual men in North Carolina. Sex Transm Infect. 2011;87:A192-A193.

171. Kivela M, Arenas A, Barthelemy M, et al. Multilayer networks. J Complex Networks. 2014;2(3):203-271.

172. Levandowsky M and Winter D. Distance between sets. Nature. 1971;234:34-35.

173. Dickison M, Havlin S, Stanley HE. Epidemics on interconnected networks. Phys Rev. 2012;E85:066109.

174. Guclu H, Read J, Vukotich C, et al. Social contact networks and mixing among students in K-12 schools in Pittsburg, PA. PLoS One. 2016;11(3):e0151139.

175. Sah P, Leu S, Cross P, et al. Unraveling the disease consequences and mechanisms of modular structure in animal social networks. Proc Nat Acad Sci. 2017;114(16):4165-4170.

176. Clémençon S, De Arazoza H, Rossi F, et al. A statistical network analysis of the HIV/AIDS epidemics in Cuba. Soc Netw Anal Min. 2015;5:58.

177. Mucha PJ, Richardson T, Macon K, et al. Community structure in time-dependent, multiscale, and multiplex networks. Science. 2010;328(5980):876.

178. Centers for Disease Control and Prevention. Ending the HIV epidemic: A plan for America. https://www.cdc.gov/endhiv/index.html. Accessed August 20, 2019.

179. NC Health and Human Services. North Carolina Electronic Disease Surveillance System (NC EDSS). http://epi.publichealth.nc.gov/cd/about/ncedss.html. Accessed January 25, 2019.

180. Etikan I, Alkassim R, Abubakar S. Comparison of snowball sampling and sequential sampling techniques. Biom Biostat Int J. 2016;3(1):00055.

181. The Cecil G. Sheps Center for Health Services Research. The Carolina Cost and Quality Initiative. http://www.shepscenter.unc.edu/data/bcbsnc-claims-data-ccqi/. Accessed January 25, 2019.

182. Blue Cross and Blue Shield of North Carolina. Company information. http://www.bcbsnc.com/content/corporate/company-info.htm. Accessed January 25, 2019.

183. United States Census Bureau. QuickFacts: North Carolina. https://www.census.gov/quickfacts/NC. Accessed January 25, 2019.

153

184. North Carolina Department of Health and Human Services. Medicaid and health choice enrollment reports. https://dma.ncdhhs.gov/documents/reports/enrollment- reports/medicaid-and-health-choice-enrollment-reports. Accessed January 25, 2019.

185. Centers for Medicare & Medicaid Services. Medicaid enrollment dashboard. https://www.cms.gov/Research-Statistics-Data-and-Systems/Statistics-Trends-and- Reports/Dashboard/Medicare-Enrollment/Enrollment%20Dashboard.html. Accessed January 25, 2019.

186. North Carolina Institute of Medicine. Claims data to improve health in North Carolina: A report from the NCIOM task force on all-payer claims database. https://www.nciom.org/apcd/. Accessed January 25, 2019.

187. R Core Team (2013). R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria. http://www.R-project.org/.

188. Fu F, Nowak MA, Christakis NA, et al. The evolution of homophily. Sci Rep. 2012;2:845.

189. Weir WH, Emmons S, Gibson R, et al. Post-processing partitions to identify domains of modularity optimization. Algorithms. 2017;10(3):93.

190. Jeub LGS, Bazzi M, Jutla IS, et al. A generalized Louvain method for community detection implemented in MATLAB. 2011-2017. http://netwiki.amath.unc.edu/GenLouvain. Accessed February 21, 2020.

191. Traud A, Frost C, Mucha P, et al. Visualization of communities in networks. Chaos. 2009;19(4):041104.

192. Stevens JP. Applied multivariate statistics for the social sciences (5th ed.). New York, NY: Routledge; 2012.

193. Division of HIV/AIDS Prevention. Selected national HIV prevention and care outcomes in the United States. Centers for Disease Control and Prevention. https://www.cdc.gov/hiv/pdf/library/factsheets/cdc-hiv-national-hiv-care-outcomes.pdf. Accessed March 10, 2019.

194. Hanley JA, Negassa A, Edwardes MD, et al. Statistical analysis of correlated data using generalized estimating equations: An orientation. Am J Epidemiol. 2003;157(4):364-375.

195. Pan W. Akaike’s information criterion in generalized estimating equations. Biometrics. 2001;57(1):120-125.

196. Greenland S, Pearl J, Robins JM. Causal Diagrams for Epidemiologic Research. Epidemiology 1999; 10(1): 37-48.

197. Hin LY and Wang YG. Working-correlation-structure identification in generalized estimating equations. Stat Med. 2009;28(4):642-658.

198. Hojsgaard S, Halekoh U, Yan J. The R package geepack for generalized estimating equations. J Stat Soft. 2006;15(2):1-11.

154

199. Wu H, Mendoza MCCB, Huany YA, et al. Uptake of HIV preexposure prophylaxis among commercially insured persons—United States, 2010–2014. Clin Infect Dis. 2017;64(2):144-149.

200. Azur NJ, Stuart EA, Frangakis C, et al. Multiple imputation by chained equations: What is it and how does it work? Int J Methods Psychiatr Res. 2011;20(1):40-49.

201. van Buuren S and Groothuis-Oudshoorn K. mice: Multivariate imputation by chained equations in R. J Stat Softw. 2011;45(3):1-67.

202. Massey FJ. The Kolmogorov-Smirnov test for goodness of fit. J Am Stat Assoc. 1951;46(253):68-78.

203. Vuong QH. Likelihood ratio tests for model selection and non-nested hypotheses. Econometrica. 1989;57(2):307-333.

204. Sah P, Singh LO, Clauset A, et al. Exploring community structure in biological networks with random graphs. BMC Bioinformatics. 2014;15:200.

205. Wall KM, Stephenson R, Sullivan PS. Frequency of sexual activity with most recent male partner among young, internet-using men who have sex with men in the United States. J Homosex. 2013;60(10):1520–1538.

206. Baggaley RF, White RG, Boily MC. HIV transmission risk through anal intercourse: Systematic review, meta-analysis and implications for HIV prevention. Int J Epidemiol. 2010;39(4):1048–1063.

207. Jin F, Jansson J, Law M, et al. Per-contact probability of HIV transmission in homosexual men in Sydney in the era of HAART. AIDS. 2010;24(6):907-13.

208. Jenness SM, Goodreau SM, Morris M. EpiModel: An R package for mathematical modeling of infectious disease over networks. J Stat Softw. 2018;84:8.

209. Centers for Disease Control and Prevention. HIV in the United States and dependent areas. https://www.cdc.gov/hiv/pdf/statistics/overview/cdc-hiv-us-ataglance.pdf. Accessed September 22, 2019.

210. Centers for Disease Control and Prevention. Sexually transmitted disease surveillance 2017. https://www.cdc.gov/std/stats17/syphilis.htm. Updated July 24, 2018. Accessed September 22, 2019.

211. Krivitsky PN, Handcock MS, Raftery AE, et al. Representing degree distributions, clustering, and homophily in social networks with latent cluster random effects models. Soc Networks. 2009;31(3):204-213.

212. Kurth AE, Celum C, Baeten JM, et al. Combination HIV prevention: Significance, challenges, and opportunities. Curr HIV/AIDS Rep. 2011; 8(1): 62-72.

213. Rozhnova G, Heijne J, Bezemer D, et al. Elimination prospects of the Dutch HIV epidemic among men who have sex with men in the era of preexposure prophylaxis. AIDS 2018; 32(17): 2615-2623.

155

214. Office of National AIDS Policy. National HIV/AIDS strategy for the United States: Updated to 2020. https://files.hiv.gov/s3fs-public/nhas-update.pdf. Accessed February 12, 2020.

215. North Carolina DHHS Communicable Disease Branch. HIV in North Carolina, 2017. https://epi.dph.ncdhhs.gov/cd/stds/figures/factsheet_HIV_2017_rev11272018.pdf. Accessed February 12, 2020.

216. Vasylyeva TI, Friedman SR, Paraskevis D, et al. Integrating molecular epidemiology and social network analysis to study infectious diseases: Towards a socio-molecular era for public health. Infect Genet Evol. 2016;46:248-255.

217. Vazquez-Prokopec GM, Bisanzio D, Stoddard ST, et al. Using GPS technology to quantify human mobility, dynamic contacts and infectious disease dynamics in a resource-poor urban environment. PLoS ONE. 2013;8(4):e58802.

218. Barrat A, Cattuto C, Tozzi AE, et al. Measuring contact patterns with wearable sensors: methods, data characteristics and applications to data-driven simulations of infectious diseases. Clin Microbiol Infect. 2014;20(1):10-16.

219. Rosenberg ES, Millett GA, Sullivan PS, et al. Understanding the HIV disparities between black and white men who have sex with men in the USA using the HIV care continuum: A modeling study. Lancet HIV. 2014;1:e112–18.

220. White D, Rosenberg ES, Cooper HL, et al. Racial differences in the validity of self-reported drug use among men who have sex with men in Atlanta, GA. Drug Alcohol Depend. 2014;138:146–153.

221. Marmor M, Sheppard HW, Donnell D, et al. Homozygous and heterozygous CCR5- Delta32 genotypes are associated with resistance to HIV infection. J Acquir Immune Defic Syndr. 2001;27:472–81.

222. Goodreau SM, Rosenberg ES, Jennes SM, et al. Sources of racial disparities in HIV prevalence in men who have sex with men in Atlanta, GA, USA: a modelling study. Lancet HIV. 2017;4:e311-e320.

223. Miller WC, Rosenberg NE, Rutstein SE, et al. The role of acute and early HIV infection in the sexual transmission of HIV. Curr Opin HIV AIDS. 2010;5(4):277-282.

224. Villandre L, Stephens DA, Labbe A, et al. Assessment of overlap of phylogenetic transmission clusters and communities in simple sexual contact networks: applications to HIV-1. PLoS One. 2016;11(2):e0148459.

225. Evans D, Van Gorder D, Morin SF, et al. Acceptance of the use of HIV surveillance data for care engagement: National and local community perspectives. Immune Defic Syndr. 2015;69(1):S31-S36.

226. Franklin JM, Shrank WH, Pakes J, et al. Group-based trajectory models: A new approach to classifying and predicting long-term medication adherence. Med Care. 2013;51:789- 796.

156

227. Rudolph AE, Crawford ND, Latkin C, et al. Multiplex relationships and HIV: Implications for network-based interventions. AIDS Behav. 2016;21:1219-1277.

228. Samoff E, Mobley V, Hudgins M, et al. HIV outbreak control with effective access to care and harm reduction in North Carolina (2017–2018). Am J Pub Health. 2020;110(3).

157