Social Vulnerability Modifies the Effects of Geographic Proximity on Engagement in Latent Tuberculosis Infection Care in a United States Safety Net Healthcare Network

Abstract

Background

Latent tuberculosis (TB) infection care often requires engagement with multiple teams in several locations throughout the diagnostic and treatment steps of the TB infection care cascade. The intersecting effects of geographic proximity and social drivers on care cascade retention have not been well examined.

Methods

We conducted a retrospective cohort study of patients with a positive TB infection test between 2018–2019 within a health system in Boston, Massachusetts. The primary outcome was attendance at a TB clinic after a referral was placed. The primary exposure was geographic proximity, as measured by travel time by car. We assessed effect modification of proximity by Social Vulnerability Index (SVI), a composite measure of census tract social drivers.

Results

We identified 1677 patients with positive TB infection tests; 1208 (72%) were referred to a TB clinic, of whom 748 (62%) completed referral. Longer travel times were associated with lower odds of referral completion (furthest vs nearest quartiles: adjusted odds ratio, 0.76 [95% confidence interval, .71–.82]). SVI significantly modified the effects of proximity: Increasing travel time was associated with decreasing probability of clinic attendance for patients in lower-vulnerability census tracts but had minimal effect on clinic attendance among patients in higher vulnerability census tracts.

Conclusions

Additional support is needed for individuals referred to TB clinics that require long travel times to attend. Support should also account for other social drivers affecting care access for those living near TB clinics.

Increased travel time was associated with decreased odds of tuberculosis infection care referral completion. Increased travel time was associated with increased odds of treatment completion. Social Vulnerability Index modified the effects of travel time on referral, treatment, and cascade completion.

An estimated 13 million individuals in the United States (US) have latent tuberculosis (TB) infection (henceforth: “TB infection”), and approximately 9600 people are diagnosed with TB disease annually in the US [1, 2]. Without treatment, 5%–10% of individuals with TB infection progress to TB disease [3]. TB preventive treatment reduces progression (or “reactivation”) risk by 90% [4]. Yet while 80% of TB disease in the US arises from reactivated TB infection [5], only 1 in 10 individuals with TB infection in the US has completed preventive treatment [6]. The TB infection care cascade refers to the steps patients and their care teams need to complete to diagnose and complete treatment for TB infection [7]. Losses from the care cascade represent missed opportunities to treat TB infection and thereby prevent TB disease.

In the US, TB infection care for a patient is often provided in several clinical settings. For example, initial testing may be obtained in a primary care clinic; chest radiographs may then be obtained in a geographically separate radiology suite; and treatment may be prescribed and monitored in specialty TB clinics. Access to and accessibility of these care locations likely affects overall retention in TB infection care. However, prior studies that have examined distance and travel time—proxy measures of accessibility—to TB infection care sites have heterogeneous effects on retention in care [8–11]. A potential driver of this heterogeneity is differences in the prevalence and effects of other social drivers of health on care access, though this hypothesis has largely not been tested. While several proximity-unrelated social drivers (eg, social support, socioeconomic status) have been theorized to affect access to medical care in general [12], the ways in which social drivers of health modify the effects of proximity for TB infection care retention in the US have not been assessed.

Here, we examine the relationship between proximity and the Social Vulnerability Index (SVI) on completion of key TB infection care cascade steps in a single, safety net healthcare network. The SVI, composed from American Community Survey data, is an index measuring socioeconomic status, household characteristics, racial and ethnic minority status, and housing and transportation access [13]. We and others have found that the SVI is independently associated with healthcare delivery outcomes, making it an appealing unified measure of social and environmental factors that promote or inhibit the ability to reach care [14, 15]. Our analysis is informed by the Theory for Access to Healthcare [12] and prior qualitative work in our setting [16], which conceptualized healthcare access as a function both of a health system's accessibility (here represented by proximity) and patients' abilities (here represented by the SVI).

METHODS

Setting and Population

We conducted a retrospective cohort study of patients tested for TB infection at an urban safety net hospital and 9 affiliated community health centers in Boston, Massachusetts. In our setting, most TB infection screening and testing is conducted in outpatient primary care or specialty clinics, which then refer patients with positive tests to a Department of Public Health–funded TB clinic for further evaluation (including chest radiographs) and treatment. Some patients receive treatment in primary care or other subspecialty (eg, human immunodeficiency virus [HIV]) clinics.

Patients, regardless of age, were included if they had received a tuberculin skin test (TST) and/or interferon-gamma release assay (IGRA) with a valid result between 1 January 2018 and 31 December 2019. For patients with positive tests, we performed manual chart abstraction to determine completion of care cascade steps. For this analysis, we excluded patients who had missing clinical records, who were pregnant or had delivered a child within 7 days of the positive test (whom we analyzed separately [17]), who were previously treated for TB disease or TB infection, who were diagnosed with TB disease at the time of the positive test, or who were deemed to have false-positive TB infection tests. We also excluded patients for whom no referral information was available and, because our analysis focused on local care patterns, those who lived >100 km from the clinic where they were tested. We excluded patients with unknown treatment completion from analyses of treatment completion and overall cascade completion.

Outcomes: TB Infection Care Cascade

The primary outcome was completion of referral after referral initiation because this cascade step has been less well studied than others, is a major source of losses from the cascade, and plausibly is most susceptible to proximity-related barriers to care engagement. Secondary outcomes were (1) treatment completion among those who started treatment; and (2) completion of the overall care cascade among all individuals with a positive TB infection test, defined as referral completion (among those referred) plus treatment completion or determination of ineligibility for treatment (per judgment of the treating clinician).

We defined referral initiation as placement of an electronic order for referral to a TB clinic or clinical documentation that a referral was made. We defined referral completion as attending at least 1 visit at a TB clinic. Referral initiation and referral completion needed to be completed within 6 months of the positive test to be considered complete, given recommendations to repeat TB infection testing and radiography if treatment is not started within 6 months for high-risk individuals [18]. We defined treatment initiation as prescription of a Centers for Disease Control and Prevention–approved drug regimen for TB infection or class IV (inactive) TB [19]. We determined treatment completion using documentation in the clinical records.

Exposures

The primary measure of proximity was travel time by car from each patient's primary address to the address of the clinic where the patient received TB-related care (when patients completed referral or received TB infection treatment in a primary care clinic), or to our clinical center's central TB clinic (when patients were not referred and not treated or were referred but did not complete referral). Automobile travel time was calculated as occurring at 12 Pm on a Tuesday and was computed using the georoute HERE API plugin in Stata. Although there are multiple TB clinics in the Boston area and the clinic to which patients were referred was not always apparent from clinical records, we assumed that patients referred within our network were referred to the central TB clinic unless otherwise specified in the health record. This assumption is justified because this is the standard referral pattern for our clinical center and affiliated community health centers and 99% of patients who attended a TB clinic in our cohort attended the central clinic. Travel time was separated into quartiles. Because we did not have information about the mode of transportation patients used to get to and from clinics, we performed sensitivity analyses using 2 alternative measures of proximity: Euclidian distance between addresses and public transit travel time.

Covariates

We selected covariates from TB infection care cascade literature [15, 20, 21], and abstracted them from patients’ electronic health records. Sociodemographic covariates included sex, age, language preference, and US birth. Clinical covariates were defined using International Classification of Diseases, Tenth Revision (ICD-10) codes, and included conditions associated with disease- or treatment-related immunosuppression (HIV, transplant receipt, any cancer, inflammatory bowel disease, and rheumatologic conditions) and “other” important comorbidities (end-stage renal disease, chronic kidney disease, type 1 diabetes, type 2 diabetes, and liver insufficiency/failure) (Supplementary Table 1). We assessed for effect modification of proximity by SVI. In this analysis, we parameterized SVI as state-level quartile rankings, measured at the census tract of each patient's address.

Statistical Analysis

We used summary and descriptive statistics to characterize the cohort and the care cascade. To illustrate the spatial distribution of proximity to clinic, we generated maps of patients' residences, using donut method geomasking to protect privacy [22]. We used mixed-effects logistic regression with robust standard errors to model the odds of cascade step completion, accounting for clustering at the testing clinic level (for analysis of referral completion) or treatment initiation clinic (for analysis of treatment completion). First, we performed univariable analysis and identified covariates that were associated with outcomes at the level of P < .2. We included these variables in multivariable models along with measures of proximity. We used interaction terms to test for effect modification between proximity and SVI; patients with missing SVI data were excluded from this analysis. We illustrated the interaction effects using postestimation margins plots and assessed the global interaction significance using joint Wald tests for each interaction model. We performed mapping with ArcGIS Online (Esri, Redlands, California, USA), and statistical analysis using Stata version 17 (StataCorp, College Station, Texas, USA).

RESULTS

We identified 16 685 tests obtained in the study timeframe, of which 2337 (14%) were positive. Of these, 1677 patient testing encounters (henceforth “patients”) among 1497 individuals met inclusion criteria (Figure 1). Most patients were female (56%), preferred a language other than English (54%), and were non-US-born (69%); the median age was 46.2 years (interquartile range [IQR], 34.4–60.9 years) (Table 1). The median travel time by car from residence to TB clinic was 21.5 minutes (IQR, 14.9–29.0 minutes); 43% of patients lived in a census tract in the highest (most vulnerable) SVI quartile. Supplementary Table 2 summarizes proximity between patients' addresses and the clinic where they attended for TB infection care, or proximity to the health system's central TB clinic for those who did not receive evaluation or treatment. Supplementary Figure 1 shows the spatial distribution of patients and proximity to the central TB clinic or the clinic where they started therapy.

Figure 1.

Consort diagram and care cascade for tests; there were 1677 positive tests among 1497 patients included in the care cascade analysis. Red boxes indicate patients who were lost from the care cascade. Green boxes represent patients who successfully completed the care cascade. Abbreviations: EHR, electronic health record; LTBI, latent tuberculosis infection; MRN, medical record number; neg, negative; pos, positive; TB, tuberculosis.

Table 1.

Cohort Characteristics

Characteristic	N = 1677a
Female sex	947 (56)
Age, y, median (IQR)	46.2 (33.4–60.9)
Age category
<5 y	9 (1)
5–14 y	37 (2)
15–24 y	150 (9)
25–44 y	600 (36)
45–64 y	578 (34)
≥65 y	303 (18)
Language
English	778 (46)
Spanish	176 (11)
Haitian Creole	438 (26)
Otherb	284 (17)
Missing	1 (1)
Born in the United States
No	1158 (69)
Yes	245 (15)
Missing	274 (16)
Testing modality
Tuberculin skin test	295 (18)
Interferon-gamma release assay	1382 (82)
Immunocompromising comorbidities	158 (9)
Other comorbidities	306 (18)
Social Vulnerability Index quartile
First (least vulnerable)	151 (9)
Second	254 (15)
Third	321 (19)
Fourth (most vulnerable)	716 (43)
Missing	235 (14)
Time to testing clinic by car, min, median (IQR) [range]	20.9 (12.0–28.7) [0.5–80.6]
Time to treatment clinic by car, min, median (IQR) [range]	21.5 (14.9–29.0) [0.2–79.7]
Time to testing clinic by public transport, min, median (IQR) [range]	53 (34–70) [2.1–186]
Time to treatment clinic by public transport, min, median (IQR) [range]	53 (38–71) [1.3–198]
Distance to testing clinic, km, median (IQR) [range]	8.6 (3.3–17.7) [0.1–99.5]
Distance to treatment clinic, km, median (IQR) [range]	9.0 (4.9–18.2) [0.1–99.6]

Data are presented as No. (%) unless otherwise indicated.

Abbreviation: IQR, interquartile range.

^aData refer to 1677 testing encounters among 1497 patients. Demographic data reflect patient characteristics at time of testing.

^bOther languages spoken by patients include Acholi, Albania, American Sign Language, Amharic, Arabic, Bengali, Brazilian Portuguese, Burmese, Cape Verdean Creole, Cantonese, Mandarin, Taoisanese, French, Ga, Gujarati, Hindi, Igbo, Krio, Luganda, Mandingo, Nepali, Oromo, Portuguese, Russian, Somali, Swahili, Tagalog, Thai, Tigrinya, Turkish, Ubu, Vietnamese, Wolof, Yoruba, and Urdu.

A total of 1208 of 1677 patients (72%) were referred to our medical center's TB clinic for treatment, and 48 (3%) were not referred but started treatment outside of the medical center's TB clinic (Figure 1). Of the 421 patients who were not referred and did not start treatment, no reason was documented for 379 (90%), 29 patients refused (7%), and other reasons were given for 13 (3%). Of the 1208 referred patients, 748 (62%) attended. Of these 748 patients, 551 (74%) started treatment, 80 (11%) were deemed ineligible for treatment (and were considered to have successfully completed the cascade), and 117 (16%) did not start therapy. Of the 551 patients who started therapy, 313 (57%) completed, 224 (41%) did not complete, and 14 (3%) had unknown completion status. Of the 48 patients who started treatment outside of the central TB clinic, 21 (44%) completed therapy, 11 (23%) did not complete therapy, and 16 (33%) had unknown completion status. In total, 414 of 1647 (25%) patients with known treatment completion status completed the care cascade.

Proximity as a Predictor of Referral, Treatment, and Cascade Completion

We found that intermediate (second and third quartiles) and extended (fourth quartile) driving times were associated with decreased referral completion compared to the shortest times (second quartile vs first quartile: adjusted odds ratio [aOR], 0.85 [95% confidence interval {CI}, .75–.96]; third quartile vs first quartile: aOR, 0.88 [95% CI, .76–1.01]; fourth quartile vs first quartile: aOR, 0.76 [95% CI, .71–.82]) (Figure 2, Supplementary Table 3). We found a U-shaped relationship between drive time and treatment completion: The quartile 2 drive times were associated with decreased treatment completion (second vs first quartile: aOR, 0.84 [95% CI, .74–.94]), while longer-intermediate and extended drive times were associated with increased treatment completion (third vs first quartile: aOR, 1.29 [95% CI, 1.17–1.42]; fourth vs first quartile: aOR, 1.98 [95% CI, 1.77–2.22]) (Figure 2, Supplementary Table 4). There was no significant association between drive times and overall cascade completion. We observed similar associations when we used public transportation and distance as measures of proximity (Supplementary Figure 3, Supplementary Table 5).

Figure 2.

Adjusted odds ratios and 95% confidence intervals for predictors of referral completion, treatment completion, and overall care cascade completion. Solid lines represent an odds of 1. Abbreviations: F, female; IGRA, interferon-gamma release assay; TB, tuberculosis; TST, tuberculin skin test; US, United States.

Other Factors Associated With Referral, Treatment, and Cascade Completion

Several sociodemographic and medical factors were associated with referral, treatment, and cascade completion (Figure 2, Supplementary Tables 3, 4, and 6). We found that female sex and higher age brackets were associated with decreased referral, treatment, and cascade completion. Speaking Spanish or “other” languages was associated with increased completion of these outcomes. Non-US birth was associated with increased referral and cascade completion, but decreased treatment completion. Testing with IGRA was associated with increased completion of all outcomes. While immunocompromising and other important medical conditions were associated with increased referral completion, they were associated with decreased treatment completion. Treatment with a rifamycin-based therapy was associated with increased treatment completion. We did not identify differences in treatment completion when treatment was started in primary care versus TB clinics. Treatment initiation in “other” clinical settings was associated with increased treatment completion, though this finding was limited by a small sample size.

Effect Modification of Proximity by SVI

We found significant effect modification by SVI on the association between drive time and referral completion, treatment completion, and overall cascade completion (joint Wald test P < .0001 for each model). Within the first and second travel time quartiles, probability of completing referral decreased as SVI quartile increased (ie, as vulnerability increased), but there was minimal variation by SVI in the third and fourth travel time quartiles (Figure 3). In contrast, while there was a trend toward decreasing treatment completion as SVI increased within each proximity quartile, this relationship reversed for patients in the third quartile. While overall care cascade varied by SVI within proximity quartiles, there was no consistent pattern to this variation.

Figure 3.

Predicted probability of referral completion, treatment completion, and overall cascade completion, showing effect modification of car travel time by Social Vulnerability Index (SVI). Models are adjusted for sex, age, language preference, birth within or outside the United States, testing with interferon-gamma release assay or tuberculin skin test, presence of immunocompromising condition, and presence of other major comorbidities. The treatment completion models are also adjusted for use of rifamycin-based therapy and whether treatment was started in a tuberculosis clinic.

DISCUSSION

In this analysis of proximity and social factors related to TB infection care cascade completion in a large healthcare network in Boston, we found that longer driving time was associated with decreased odds of TB infection referral completion: The odds of completing referral were 24% lower for patients living furthest compared to closest to the TB clinic. SVI modified this effect, with more pronounced differences by SVI quartile for those who lived closest to the TB clinic. Surprisingly, we found a nonlinear relationship between proximity and treatment completion, with those living at intermediate travel times having lower odds of treatment completion, whereas those living farther away demonstrating nearly twice the odds of completing treatment. Key demographic and medical characteristics, including age, language, US birth, and testing modality, were also associated with referral, treatment, and cascade completion. Our study provides insights into how TB services can be organized and optimized to mitigate referral- and treatment-based losses from TB infection care.

Our findings suggest that proximity modestly facilitates completion of referral for TB infection care, which has implications for how referrals are made and subspecialty TB care is distributed in our setting. Inaccessibility to subspecialty clinics compared to primary care clinics has been cited as a reason to “task shift” TB infection care to primary care clinicians and away from referral TB clinics [23, 24]. One motivation of this strategy is to minimize loss to follow-up that occurs when referrals to centralized TB clinics are required for treatment, because patients likely live closer to—and thus are more likely to attend—their primary care clinics. In Massachusetts, the TB infection Extension for Community Healthcare Outcomes (ECHO) program seeks to educate primary care clinicians about TB infection care, to overcome knowledge gaps that would otherwise prevent such care [25]. Likewise, a pilot program in Massachusetts demonstrated the feasibility of embedding dedicated TB infection services within a community health center [26]. Nationally, the TB Epidemiology Studies Consortium cycle III is examining ways to enhance management of TB infection in primary care settings [27]. Our findings lend support to the contention that more proximate TB infection services enhance establishing TB infection care, though we were not able to compare this effect between primary care versus dedicated TB clinics.

However, we found that patients living at intermediate proximity to TB treatment clinics had lower odds of treatment completion than patients living either very close or very far from clinic. Two prior studies of TB infection care among children in urban US settings (including a prior study in a different healthcare system in Boston) similarly found that patients who lived near clinics had decreased visit attendance and treatment completion compared to those living farther away [11, 28]. This result may reflect a selection bias among those who accessed care; it is possible that among patients living far from treatment clinics, only those who were highly committed and able to complete treatment established care in the first place. This U-shaped relationship may also reflect different transportation modalities that patients living near, intermediate, and far from clinic used, though we were unable to directly measure transportation method. Overall, this finding suggests that once patients were established on treatment, distance was a less salient barrier to treatment completion. In the context of our referral completion findings, these results indicate that public health interventions to reduce travel-related barriers to initial referral intake are needed, but that addressing barriers unrelated to travel for patients who have established care may be high-yield.

We found that SVI significantly modified the effects of proximity on referral completion and treatment completion, indicating that social drivers may impede care access even when proximity is not a barrier to care. This effect was most pronounced in our analysis of referral completion, which showed that the odds of completing a referral decreased as vulnerability increased for those living most proximate to TB clinics, while the effects of SVI were less pronounced among patients living further from clinic. These results may be due to differential access to resources that would enable attendance at a TB clinic—such as access to a car—between urban and rural populations with similar SVI rankings. Meanwhile, the odds of completing treatment decreased as social vulnerability increased within 3 of 4 proximity quartiles, suggesting that social vulnerability imposes a consistent barrier to treatment retention regardless of proximity.

In both the referral and treatment completion analyses, the observed disparities by SVI ranking that are apparent across proximity quartiles illustrate the ways in which the theoretical constructs of “accessibility” (measured in our study as proximity) and “ability” (measured in our study as SVI) operated independently to affect care retention. Several innovative strategies have been developed to reduce accessibility- and ability-related barriers to TB infection care, many targeted toward populations facing significant socioeconomic barriers to care. These include telehealth [11], mobile treatment delivery [29], and building capacity within primary care and community health centers [25, 26, 30, 31]. Our study suggests that multilevel approaches tailored toward mitigating the joint effects of proximity and other social drivers are most likely to be impactful for optimizing overall care engagement.

Additional factors associated with TB infection care engagement point toward structural and behavioral barriers to care engagement for specific populations. Overall, care engagement was lower for older compared to younger patients. A combination of parental motivation and health system support likely contributed to higher rates of referral completion among younger patients. (Notably, while children <5 years old were less likely to complete therapy, this finding may have been influenced by small sample size.) Many adults ≥65 years old started therapy in our study, but treatment completion in this group was significantly lower than in younger groups. Rates of antitubercular therapy toxicity increases with age [32], and high rates of treatment discontinuation in older groups may have been due to toxicity. Few prior care cascade studies have incorporated all age ranges in their analysis, and our study illustrates both successful engagement of young children and the need to improve TB infection care for older adults.

US birth and language preference were heterogeneously associated with completion of cascade steps. We found a strong positive association between US birth and referral completion, followed by a strong negative association between this characteristic and treatment completion. These observations may be due to patients' need for work- or immigration-related documentation after a positive screening test, followed by lack of support to complete treatment after completing the referral. Notably, other large US TB infection care cascade studies have not found differences in treatment initiation or completion by nativity [6, 20], suggesting that immigration-related effects may be different in different localities or among different immigrant populations in the US. Meanwhile, patients speaking languages other than English had generally higher odds of referral completion, treatment completion, and overall care cascade completion compared to English-speaking patients. Non-English-language preference has been associated with higher TB infection testing uptake [33], but few studies have identified differences in engagement in subsequent cascade steps by language. Our results may reflect our health system's ability to support non-English-language speakers. They also indicate a need to better understand social and structural barriers to care for English-speaking patients with TB infection.

Finally, medical factors were associated with differential completion of cascade steps. We found a strong positive relationship between receiving an IGRA (vs TST) and completing referral, treatment, and the overall care cascade, mirroring results from prior research [20, 21, 34]. One potential explanation for this result is that BCG-vaccinated individuals with positive IGRAs may be less skeptical about a TB infection diagnosis than they would be with a positive TST. This hypothesis is supported by qualitative research that found that TST distrust is a barrier to TB infection care [16]. While patients with immunodeficiency or other comorbidities were at increased likelihood of completing referral, they had lower likelihood of treatment completion. It is possible that these patients were motivated and supported to attend initial evaluations following a positive TB infection test, but experienced increased rates of toxicity on treatment.

Strengths of our analysis include use of a dataset with rich sociodemographic and precise geographic information, derived from a medical record that enabled tracking of patients between multiple clinical settings across several health systems. Our study also has limitations. First, data were obtained from a single healthcare system, limiting generalizability, particularly to settings in which median distances and transportation modes to clinics are much different from our setting (eg, rural areas). Second, this was a retrospective study covering the year prior to the coronavirus disease 2019 (COVID-19) pandemic, and follow-up appointments may have been disrupted by the pandemic. Notably, we found similar results when we restricted the analysis to patients tested ≥6 months prior to COVID-19 lockdowns in Massachusetts. Third, our estimated travel times may not match patients' actual travel times. Similarly, our database did not contain information about transportation modes that each patient would have used to reach clinical sites. Fourth, we did not capture data on reason for testing or treatment discontinuation, because these data were often missing or ambiguous in clinical records. Fifth, as an area-based marker of vulnerability, SVI is subject to potential ecological fallacy, whereby census tract population-level vulnerability may not reflect the actual vulnerability of individual patients in the study. Our use of census tracts rather than larger geographic areas likely attenuates but does not fully eliminate the potential for ecological fallacy.

In conclusion, in this retrospective study of patients with TB infection receiving care in a large, safety net healthcare system in Boston, we found that social vulnerability modified the effects of proximity on referral, treatment, and overall cascade completion. We found disparities in referral completion between the most and least socially vulnerable patients who lived close to, but not far from, the TB clinic. Increasing social vulnerability was associated with lower odds of treatment completion across most proximity quartiles, even though proximity exhibited a nonlinear association with treatment completion. Efforts to improve access to and accessibility of TB infection services should consider both location and social drivers that impede patients' ability to reach care. To wit, improving access may not simply involve using or placing TB infection services at locations closest to patients. Our results should inform future implementation research into strategies that could mitigate the joint effects of proximity and social drivers that impede care.

Supplementary Data

Supplementary materials are available at Open Forum Infectious Diseases online. Consisting of data provided by the authors to benefit the reader, the posted materials are not copyedited and are the sole responsibility of the authors, so questions or comments should be addressed to the corresponding author.

Notes

Acknowledgments. We would like to thank the patients and staff of Boston Medical Center and affiliated community health centers.

Author contributions. J. I. C. designed the study, collected and analyzed data, and wrote the manuscript. A. G., D. L., and M. H. collected and analyzed data, and reviewed and critically revised the manuscript. S. B., J. E. H., M. B. B., P. S., L. W., V. S., C. A. T., C. R. H., H. E. J., and K. R. J. guided data analysis and reviewed and critically revised the manuscript.

Ethics approval. This study was approved by the Boston University Medical Campus institutional review board (H-43385).

Disclaimer. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health.

Financial support. This work was supported by the Joel and Barbara Alpert Endowment for Children of the City; the Providence/Boston Center for AIDS Research (grant number P30AI042853); and the Boston University Clinical and Translational Science Institute (grant number 1UL1TR001430). J. I. C. and M. B. B. are supported by the National Institute of Allergy and Infectious Diseases of the National Institutes of Health (award numbers K23AI186597 and K01AI151083, respectively).