To the Editors:
Studies show that timely linkage to care for persons with HIV infection is associated with better clinical outcomes and improved length and quality of life.1,2 Despite the improvement in linkage to care among persons newly diagnosed with HIV,3,4 a relatively large proportion of persons living with HIV had never been linked to care, eg, 20% in the United States in 2009 and 12% in New York City (NYC) in 2013.5,6 Given the natural history of HIV infection, the presence of such a large proportion suggests that previous estimates of linkage to care may be inaccurate. The purpose of this analysis was to use predictive modeling to provide more accurate estimates.
The data source was the NYC HIV surveillance registry, and the analysis population included persons newly diagnosed with HIV in NYC between January 1, 2006, and December 31, 2014 (N = 24,076).
“Alive but never linked to care” was defined as a person diagnosed with HIV in NYC who was not known to be dead by December 31, 2015, and had no evidence of ever having been linked to HIV care in NYC or elsewhere in the United States. The data on death and HIV care in NYC were obtained from NYC HIV registry, and the data on HIV care elsewhere in the United States were obtained from the Routine Interstate Duplicate Review (RIDR).7,8 Linkage to care measures included “linkage to care within 3 months of diagnosis” defined as ≥1 CD4/viral load test 8–91 days after diagnosis and “linkage to care within 12 months of diagnosis” defined as ≥1 CD4/viral load test 8–365 days after diagnosis.9
For uncorrected estimates of linkage to care, we included all persons diagnosed with HIV in 2006–2014 in the denominator. For corrected estimates, different methods were used for 2006–2010 and 2011–2014. Among persons diagnosed with HIV in 2006–2010, we excluded those from the denominator who had never been linked to HIV care and were not known to be dead by December 31, 2015. Such a simple method cannot be used for persons diagnosed with HIV in 2011–2014 because of the shorter follow-up period. Instead, all of them were included in the denominator, after those who were linked to care within 12 months of diagnosis were assigned a weight of 1, and those who were not were assigned a weight based on the predictive model described next.
To build the predictive model, we first selected persons who (1) were diagnosed with HIV in 2006–2010, (2) were alive for ≥12 months after being diagnosed, and (3) had no evidence of linkage to care within 12 months of diagnosis. We then partitioned the data set randomly into a training (80%) and testing data set (20%). In the training data set, we fitted logistic regression models to predict “alive but never linked to care” with the candidate variables sex, age at diagnosis, race/ethnicity, transmission risk, diagnosing provider type, and borough of residence at diagnosis, using backward, forward, backward stepwise, and all-subsets selection methods. The backward, forward, and backward stepwise selection methods generated identical models with a c-statistic of 0.706, whereas the all-subsets selection method generated a different model with a c-statistic of 0.690. In the testing data set, we compared the 2 models and selected the model derived from the backward, forward, and backward stepwise selection methods with a higher c-statistic (0.733 vs. 0.727). The final model was used to predict each person's probability of being “alive but never linked to care,” and then each person was assigned a weight equal to 1 minus the probability.
To assess the factors associated with linkage to care within 3 months of diagnosis, corrected estimates were used, and weighted bivariate analyses were conducted to estimate crude prevalence ratios (PRs), and a weighted log-binomial regression model was used to estimate adjusted prevalence ratios (adjPRs).
By December 31, 2015, almost 10 years after diagnosis, approximately 7.5% of persons diagnosed with HIV in 2006 remained alive but never linked to care. Persons diagnosed with HIV in 2007–2010 showed a similar pattern in that after 4 or 5 years of diagnosis; the proportion alive but never linked to care was stable at 8.3% in 2007, 8.6% in 2008, 8.9% in 2009, and 7.9% in 2010 diagnoses (data not shown in figure or table).
Figure 1 shows both uncorrected and correct estimates of linkage to care. The uncorrected estimates show that the proportion of persons linked to care within 3 months of diagnosis increased from 65.7% [95% confidence interval (CI): 64.1% to 67.4%] in 2006 to 75.7% (95% CI: 73.8% to 77.5%) in 2014. The corrected estimates show that the proportion increased from 71.1% (95% CI: 69.5% to 72.7%) in 2006 to 80.0% (95% CI: 78.3% to 81.8%) in 2014.
Table 1 shows crude PRs and adjPRs of factors associated with linkage to care within 3 months of diagnosis, based on corrected estimates. Compared with blacks, persons of other races/ethnicities had better timely linkage (Hispanics: PR = 1.06 and adjPR = 1.03; whites: PR = 1.10 and adjPR = 1.05). Persons aged 13–24 (PR = 0.75 and adjPR = 0.91) and persons with a history of injection drug use (PR = 0.86 and adjPR = 0.92) were the least likely to be linked to care within 3 months of diagnosis compared with other age and transmission risk groups. Persons diagnosed at screening/diagnostic facilities were less likely to have timely linkage than those diagnosed at outpatient or inpatient facilities. The crude PR shows that timely linkage in NYC improved by 13% (PR = 1.13, 95% CI: 1.09 to 1.16) from 2006 to 2014; the adjPR reduces this increase to 4% (adjPR = 1.04, 95% CI: 1.02 to 1.06) and shows no improvements in the last 4 years.
Our analysis shows that among persons diagnosed with HIV in NYC in 2006, 7.5% seemed to have never been linked to care by December 31, 2015. A recent study from Centers for Disease Control and Prevention reported that among persons diagnosed with HIV in the United States in 2006, 10.2% never had a CD4 test by December 31, 2015.10 Such large proportions suggest that previous estimates of new diagnoses may be inaccurate. Over-counting of never-linked-to-care cases is likely caused by duplicate registry records.11,12 The NYC HIV registry is regularly deduplicated through RIDR and manual reviews. However, the deduplication process is limited by its conservative approach; potential matches must have high concordance on key identifiers, otherwise they are called nonduplicates.
Studies report that some patients do not use their health insurance or real name when seeking sexually transmitted disease care or HIV testing because of concerns about confidentiality.13–15 In this case, duplicate records will be entered in the HIV registry if (1) a previously undiagnosed individual is newly diagnosed with HIV with an assumed name at an sexually transmitted disease clinic and later receives care under his/her real name at an HIV care facility, or (2) a previously diagnosed HIV-infected individual uses his/her real name for HIV care and seeks other services elsewhere using an assumed name.
Duplicate records caused by data entry errors or different name spellings can be deduplicated through manual reviews and field investigations, but surveillance may not be able to deduplicate all of them. Duplicate records caused by assumed names are more difficult to deduplicate. Including these duplicate diagnoses in the analysis would have overestimated the denominators. Many previous analyses treated all records in the registry as unique individuals; few tried to exclude the duplicates.11,12,16 In this analysis, we used predictive modeling to remove duplicates among persons newly diagnosed with HIV in NYC, 2006–2014, to report corrected estimates of linkage to care.
Care outcomes such as timely linkage and viral suppression are generally better in NYC than the national average.4,16–22 However, in this analysis we found that even corrected estimates of linkage to care within 3 months of diagnosis in NYC were lower than the national average (80.0% vs. 84.0% in 2014). The discrepancy is likely due to different definitions of timely linkage to care: We define linkage as ≥1 CD4/viral load test 8–91 days after diagnosis, whereas Centers for Disease Control and Prevention defines linkage as ≥1 CD4/viral load test 0–91 days after diagnosis.4 We implemented the 7-day lag after an NYC study showed that many CD4/viral load tests performed within 7 days of diagnosis were part of the diagnostic workup and did not represent a medical care visit.9
Our analysis is limited by how we excluded duplicates from the denominator: not based on direct evidence of duplication, but rather based on the fact that the registry contained no evidence of care or death after diagnosis. Some patients could be falsely excluded because of (1) incomplete laboratory data reporting, (2) missed matches across jurisdictions by RIDR after patients' migration out of NYC, (3) missed matches against death registries, and (4) the possibility of remaining alive for years after diagnosis despite never having received care. However, given the high completeness (>97%) of HIV laboratory data reporting in NYC (unpublished data, NYC Department of Health and Mental Hygiene), the low possibility that persons who left NYC or died were missed by RIDR and death registry matches and had not received one single CD4/viral load test in NYC between diagnosis and migration or death, and the low possibility of remaining alive for years without ever receiving HIV care in NYC in the context of the natural history of HIV and wide availability of HIV care in NYC, we believe that the number of falsely excluded patients is small and would have minimal impact on the estimates.
Duplicates are common in HIV registries and compromise efforts to provide accurate estimates. Manual reviews and field investigations can be used for deduplication, but they are time consuming and tend to be too conservative. As an alternative, mathematical modeling can be used to “clean” the data and produce more accurate estimates of linkage to care and other HIV care outcomes.11,23–28
The authors thank Drs. Kent Sepkowitz, Demetre Daskalakis, Jay Varma, and James Hadler for their review and comments on this article.
1. Kitahata MM, Gange SJ, Abraham AG, et al. Effect of early versus deferred antiretroviral therapy for HIV on survival. N Engl J Med. 2009;360:1815–1826.
2. Lundgren JD, Babiker AG, Gordin F, et al. Initiation of antiretroviral therapy in early asymptomatic HIV infection. N Engl J Med. 2015;373:795–807.
3. Centers for Disease Control and Prevention. Monitoring selected national HIV prevention and care objectives by using HIV surveillance data—United States and 6 U.S. dependent areas—2010. HIV Surveill Supplemental Rep. 2013;18:1–25.
4. 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, 2014. HIV Surveill Supplemental Rep. 2016;21:1–87.
5. Centers for Disease Control and Prevention. CDC Fact Sheet: HIV in the United States: The Stage of Care. Washington, DC: U.S. Department of Health and Human Services; 2012.
6. HIV Epidemiology and Field Services Program. Care and clinical status of persons with HIV in NYC in 2013 as based on HIV surveillance data. 2015. Available at: https://www1.nyc.gov/assets/doh/downloads/pdf/dires/hiv-related-medical-care-2013.pdf
. Accessed September 12, 2016.
7. Hanna DB, Pfeiffer MR, Sackoff JE, et al. Comparing the national death index and the social security Administration's death master file to ascertain death in HIV surveillance. Public Health Rep. 2009;124:850–860.
8. Mitsch A, Tang T, Whitmore S. Accurate Monitoring of HIV in the United States—CDC's Routine Interstate Duplicate Review, 2005–2008. XIX International AIDS Conference, July 22–27, 2012, Washington, DC. Abstract #TUPE268.
9. Sabharwal CJ, Braunstein SL, Robbins RS, et al. Optimizing the use of surveillance data for monitoring the care status of persons recently diagnosed with HIV in NYC. J Acquir Immune Defic Syndr. 2014;65:571–578.
10. Song R, Hall HI, Green TA, et al. Using CD4 data to estimate HIV incidence, prevalence, and percent of undiagnosed infections in the United States. J Acquir Immune Defic Syndr. 2017;74:3–9.
11. Song R, Green T, McKenna M, et al. Using occupancy models to estimate the number of duplicate cases in a data system without unique identifiers. J Data Sci. 2007;5:53–66.
12. Xia Q, Braunstein SL, Wiewel EW, et al. Persons living with HIV in the United States: fewer than we thought. J Acquir Immune Defic Syndr. 2016;72:552–557.
13. Spillane HC, McNulty AM, Wand H, et al. Who declines to give a name at a sexual health service? Sex Transm Infect. 2007;83:160–162.
14. Stephens SC, Cohen SE, Philip SS, et al. Insurance among patients seeking care at a municipal sexually transmitted disease clinic: implications for health care reform in the United States. Sex Transm Dis. 2014;41:227–232.
15. Pearson WS, Cramer R, Tao G, et al. Willingness to use health insurance at a sexually transmitted disease clinic: a survey of patients at 21 US clinics. Am J Public Health. 2016;106:1511–1513.
16. Xia Q, Kersanske LS, Wiewel EW, et al. Proportions of patients with HIV retained in care and virally suppressed in New York City and the United States: higher than we thought. J Acquir Immune Defic Syndr. 2015;68:351–358.
17. Torian LV, Xia Q, Wiewel EW. Retention in care and viral suppression among persons living with HIV/AIDS in New York City, 2006-2010. Am J Public Health. 2014;104:e24–e29.
18. Xia Q, Wiewel EW, Braunstein SL, et al. Comparison of indicators measuring the proportion of human immunodeficiency virus-infected persons with a suppressed viral load. Ann Epidemiol. 2015;25:226–230.
19. Braunstein SL, Robertson MM, Myers J, et al. Using HIV viral load from surveillance to estimate the timing of antiretroviral therapy initiation. J Acquir Immune Defic Syndr. 2016;73:222–227.
20. Wiewel EW, Braunstein SL, Xia Q, et al. Monitoring outcomes for newly diagnosed and prevalent HIV cases using a care continuum created with New York City surveillance data. J Acquir Immune Defic Syndr. 2015;68:217–226.
21. Xia Q, Lazar R, Bernard MA, et al. New York City achieves the UNAIDS 90-90-90 targets for HIV-infected whites but not Latinos/Hispanics and blacks. J Acquir Immune Defic Syndr. 2016;73:e59–e61.
22. Hanna DB, Felsen UR, Ginsberg MS, et al. Increased antiretroviral therapy use and virologic suppression in the Bronx in the context of multiple HIV prevention strategies. AIDS Res Hum Retroviruses. 2016;32:955–963.
23. Xia Q, Wiewel EW, Torian LV. Revisiting the methodology of measuring HIV community viral load. J Acquir Immune Defic Syndr. 2013;63:e82–e84.
24. Dombrowski JC, Buskin SE, Bennett A, et al. Use of multiple data sources and individual case investigation to refine surveillance-based estimates of the HIV care continuum. J Acquir Immune Defic Syndr. 2014;67:323–330.
25. Schwarcz S, Hsu LC, Scheer S. Disparities and trends in viral suppression during a transition to a “Test and Treat” approach to the HIV epidemic, San Francisco, 2008–2012. J Acquir Immune Defic Syndr. 2015;70:529–537.
26. Xia Q, Wiewel EW, Torian LV. Comparison of single-visit and multiple-visit measures of retention in care for HIV monitoring and evaluation. J Acquir Immune Defic Syndr. 2016;71:e59–e62.
27. Xia Q, Sun X, Wiewel EW, et al. HIV prevalence and the prevalence of unsuppressed HIV in New York City, 2010–2014. J Acquir Immune Defic Syndr. 2017;75:143–147.
28. Hall HI, Tang T, Johnson AS, et al. Timing of linkage to care after HIV diagnosis and time to viral suppression. J Acquir Immune Defic Syndr. 2016;72:e57–e60.