BRINGING MODELS DOWN TO EARTH - Locally grounded network models for supporting HIV policy planning UW Network Modeling Group - GitHub Pages
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
BRINGING MODELS DOWN TO EARTH Locally grounded network models for supporting HIV policy planning Martina Morris Deven Hamilton, Jeanette Birnbaum Susan Buskin, Roxanne Kerani, Sara Glick, Tom Jaenicke UW Network Modeling Group
Start by acknowledging my collaborators Modelers Epidemiologists UW Network Modeling Group Joint UW/PHSKC Deven Hamilton Sarah Glick Jeanette Birnbaum Roxanne Kerani Advisory Board PHSKC – Amy Bennett, Susan Buskin, Katelyn Gardner Toren WA DOH – Jason Carr, Tom Jaenicke UW – Matt Golden, David Katz Funding: NIAID R21 NME 2018 2
Outline Project overview Model structure … and data sources Demography Transmission system Care continuum and clinical outcomes Epidemic results Preliminary – first set of runs NME 2018 3
Project goals Build locally grounded projection model to support HIV policy Models have traditionally been built at the country level But there is significant variation in HIV prevalence within countries And in the US, prevention happens at the State/Local level Start with the heterosexual epidemic in King County Why? Small, but potential for eradication First step towards a more comprehensive model It’s a challenge… NME 2018 4
Heterosexual cases of HIV Measured with Uncertainty New HIV Diagnoses in King In King County County: 2011-2016 350 6-7% of incidence is 300 attributed to 250 Heterosexual contact 200 Another 15-20% is “No Identified Risk” 150 100 Total range: 6-27% 50 0 2011 2012 2013 2014 2015 2016 In Southeast US As high as 30% Total Het/NIR Heterosexual NME 2018 5
Race and Immigration in KC HIV HIV Prevalence: 2016 Estimated HIV US Born Foreign Born Prevalence/100K White 314.2 93% 2% Black 1001.0 56% 41% Hispanic 434.6 42% 52% We see large racial disparities And profound differences in country of origin by race NME 2018 6
Importance of local MSM epidemic MSM comprise the largest group of HIV diagnoses Several papers have shown evidence that there is substantial transmission across subpopulations Based on phylogenetic clustering of HIV sequence data In the US: Oster et al. (2015) “Of heterosexual women for whom we identified potential transmission partners, 29% were linked to MSM, 21% to heterosexual men... a higher percentage of women in the West (52%) were linked to MSM” NME 2018 7
What this suggests Ongoing transmission in the heterosexual population could be below the reproductive threshold sustained by cross-boundary transmissions? This has implications for targeting prevention policy Target the boundary to have the maximum impact NME 2018 8
9 The modeling framework Dynamic network foundation (statnet) Epidemic model components (EpiModel) NME 2018
Key components of our framework Transmission system Dynamic partnership network Handled by Multilayer: Cohab, Persistent and one-time partnerships statnet Behavior within partnerships Transmission Function of viral load/stage of infection Care continuum Handled by Testing, treatment, viral suppression EpiModel Demography Aging Travel Entry/Exit NME 2018 10
Dynamic network model(s) Partnerships modeled with a STERGM • Formation ERGM • Dissolution ERGM • Estimated from egocentrically sampled data 3 different types of partnerships • Cohabiting • Persistent So three different STERGMs • One-time NME 2018 11
Transmission system components Several processes are overlaid, and interact with the network Within discordant Boundary exposure partnerships: Behavior Foreign FB Travel • Coital Frequency • Condom Use Force of infection Infectivity, by MSMF • Stage Local MSM • CC engagement • Clinical outcome FB: Foreign Born MSMF: Males who have sex with both males and females NME 2018 12
So our population has multiple subgroups Race / Immigration subgroups (5) US and foreign-born Black US and foreign-born Hispanic Other (predominantly White) Sex / Sexual preference subgroups (3) Female (F) Males who have sex with females only (MSF) Males who have sex with males and females (MSMF) And age… NME 2018 13
Lots of other model components Engagement in Care HIV Testing (at sex and race-specific rates, some never test) Treatment with ARVs Adherence, with episodes of drop and return Viral Suppression (some fraction are not full suppressors) Clinical progression after infection 4 stages (acute rise, acute fall, chronic, AIDS) Progression time: 6.4 wks, 3 wks, 10 yrs, 2 yrs Stage-specific viral load (influences infectivity) Demography: Open population model Entry at 18, exit at 45 Age and AIDS specific mortality rates Travel for foreign born (pauses local sexual activity, activates boundary exposure) NME 2018 14
15 Data sources Locally sourced, … when possible NME 2018
Model components: LOCAL DATA NEEDED Model Component Governs: Source Sexual network Partnership NSFG (18-45) formation/dissolution dynamics Behavior within partnerships Coital frequency, condom use, NSFG (18-45) HIV status disclosure Natural history of within-host Viral load, CD4, symptoms and Global Estimates HIV infection infectivity Clinical care cascade Testing, referral, adherence, PHSKC HIV Core Surveillance suppression Demographics Entries and Exits into the King County Census population (pop’n growth, mortality and in/out migration) NME 2018 16
US Data on sexual behavior National Survey of Family Growth (NSFG, 2006-15) Representative national sample with annual waves Age 15-45 Egocentric data on last 3 heterosexual partners Partner attributes (age, race/immig, cohab, duration/once only, ongoing) Behavior within partnerships HIV testing rates Combined sample size: ~40K Reweighted by age, sex, race/immigration group To match King County demographics NME 2018 17
Local data on travel back to home country: collected in public health interviews of new HIV cases Added in 2010 But only for newly diagnosed HIV cases NME 2018 18
Local “Cascade of Care” We use sex/race specific values NME 2018 19
20 Some descriptive statistics Population attributes Partnerships patterns by subgroup NME 2018
KC Demographics by race/imm/sex King County is predominantly white About 15% of the population is Black or Hispanic And half of those are foreign born NME 2018 21
KC Sex Group estimates by Age and Race About 1% of the population are MSMF Based on NSFG, reweighted by age/race/sex to KC 49% 50% 50% 50% 50% 49% 1% 1% 1% NME 2018 22
Partnership Type Prevalence by Age, Sex, and Race Cohab Persistent One Time Race Sex Age NME 2018 23
Age Mixing Cohabiting Persistent One time Ego Age Alter Age Strong age homophily for all types of partnerships NME 2018 24
Partnership Durations: Cohab & Persistent by Race AGE OF ACTIVE TIES Cohab: Persistent: average 10-15 years average 2 – 4 years NME 2018 25
Mixing by Race/Immigration group Cohab Persistent Ego Race/Immigration group Alter Race/Immigration group NME 2018 26
Overlapping partnership networks At any point in time, a person can have none, or some of each partnership type About 1-2% of the population has two or more concurrent partners Sex Cohabiting partners Cohab 1.2 9.6 0.0 Rate of 1 0.1 2.5 0.0 time partners per 100 persons 4.1 1.9 2.2 23.2 15.3 13.1 Persistent partners # Persistent Partners NME 2018 27
Concurrency by sex group Highest overall rates are in 2.5 one of the boundary populations: MSMF ~2% 2.0 About half of this is cross- 1.5 network 1.0 This is just the concurrency with opp sex partners 0.5 45% of the MSMF also have 0.0 M partners during the year F MSF MSMF Any Cross-Network Lowest overall rates are among women: ~0.5% NME 2018 28
Concurrency by race/immigration group Female Male 5 5 4 4 percent 3 3 2 2 1 1 0 0 B BI H HI W B BI H HI W Any CrossNet Any CrossNet Highest rates are for Black, Hispanic and Hispanic immigrant men ~4% Mostly multiple persistent partners for Black men Mostly cross-network for Hispanic immigrants Split equally for US born Hispanics Black women have slightly higher rates among females : ~2% NME 2018 29
Concurrency: By age This is a young person’s game Highest rates for young men: 3-7% But the configuration changes with age too The cross-network fraction rises, as rates of cohabitation rise NME 2018 30
Boundary force of infection Boundary Groups: BI HI MSMF Percent of 2.3% 5.4% 1.5% population Exposure Depart: 0.01 probabilities 2.5 partners/yr Return: 0.25 HIV acquisition F: 2.0e-04 F: 2.0e-05 probabilities * 7.2e-06 M: 1.0e-04 M: 9.9e-06 * The HIV acquisition probabilities are For example for MSMF: MSM prevalence x a function of several components, and condom use (.304) x efficacy (cond.rr=.4) x determine the FOI at the boundary P(transmission | contact) (((.0082*1.09)+(.0031*1.09))/2) x P(contact per week) (2.5/52) NME 2018 31
Much uncertainty about boundary inputs So, we will use these for model calibration At this stage, by just manually trying some values Multiplying the FOI by a factor Later: we have a better plan NME 2018 32
33 ERGM Results NME 2018
Formation models for each network Cohab Pers OT Age -0.87 -0.20 -0.38 Age2 0.02 0.00 0.01 Age Diff -3.22 -2.59 -2.40 Race (main) Black 1.10 1.14 0.42 Black Imm 1.14 1.61 -0.52 Hispanic 3.17 2.00 0.52 Hisp Imm 1.63 1.13 -0.78 Race (matching) Black 3.35 3.21 Black Imm 3.85 2.86 Hispanic 0.01 0.27 Hisp Imm 2.88 2.30 White 3.14 2.17 Concurrency Cross net -5.96 -4.36 Within net NA -2.85 NME 2018 34
Model assessment: Convergence This is what you want to see But we found it requires a very long MCMC interval (1e5) NME 2018 35
Model assessment: Network fidelity We want the dynamic simulations to reproduce the observed network statistics (on average) Degree distributions ERGMS should be Within and between networks able to reproduce By sex, age, race/imm the joint distribution of all of the network statistics Mixing patterns By age, race/imm in each model Partnership durations NME 2018 36
Persistent network: All model stats Good fit to observed… NME 2018 37
Example: Persistent degree by race Good fit to observed, even though the degree terms are not in the model NME 2018 38
Durations by partnership type These also look roughly like the observed stats … But there is more of a story here NME 2018 39
40 Epidemic results Now we can simulate the epidemic, on a network that we know closely represents the observed data NME 2018
First things first Even with a simulation size of 50,000 nodes The smallest groups are
Persistence and equilibrium 1. We can sustain an epidemic 2. And we are “in the ballpark” KC Observed Prevalence: Het only: 0.006 Het+NIR: 0.012 NME 2018 42
Prevalence by race & immigration status Simulation Observed KC prevalence Black Immigrant Here, the rank order matches the observed pattern But the BI prevalence is too high (by an order of magnitude) NME 2018 43
Prevalence by sexual partner group Simulation Here the rank order is correct And the prevalence for Females and MSF are about right Working on estimating the true local prevalence is for MSMF, current estimate is ~30% NME 2018 44
Infections by source Infections from No persistence original local Boundary without heterosexual Infections continual seeds infections across MSMF the boundary And the primary contribution is via MSMF Downstream Infections NME 2018 45
A note on workflow The project is managed on GitHub Code and data repository Organizing issues with Projects We’re keeping a lab book using markdown/html Exploratory data analysis: bookdown Records both the descriptives and our decisions Model results It’s still a bit overwhelming … NME 2018 46
47 What’s next? Model Calibration against Local Phylogenetics NME 2018
From another project here at UW Phylogenetic analyses of local HIV diagnoses White Black Asian • B Clades predominate in Latino the US • Non-B clades predominate in Africa and Asia They form a distinctive cluster here, Non-B predominantly black NME 2018 48
Herbeck & Kerani Phylogenetic analyses of local HIV diagnoses MSM Heterosexual They form a distinctive cluster here, Non-B And predominantly heterosexual NME 2018 49
An empirical foundation for calibration Directly relevant for calibrating our most uncertain parameters – the FOI across the boundaries And separate from HIV incidence and prevalence data So those can be preserved for model validation Deven Hamilton is taking the lead on this project NME 2018 50
You can also read