Regular HIV care is critical for the achievement and maintenance of virologic suppression. Failure to adhere to antiretroviral regimens can lead to resistance, viral breakthrough, clinical progression, and increased likelihood of HIV transmission.1,2 The HIV “cascade of care” is a commonly used framework to quantify the proportion of patients engaged at each stage of HIV medical care, from initial diagnosis to virologic suppression.3 This framework has several limitations. Notably, patient engagement is presumed to follow a linear trajectory through the various stages, failing to capture the clinical reality that individuals may move back and forth along the continuum or choose to minimize their engagement with care for various reasons, including being otherwise well.4,5 A model is therefore required that integrates HIV clinical outcomes and frequency of follow-up in a manner that characterizes transitions through various states of care experienced by individuals throughout the course of their illness and differentiates patterns of engagement yielding poor clinical outcomes from those where patients may be clinically well despite infrequent contact with health care providers. Such an approach would facilitate identification of factors associated with suboptimal patterns of engagement and inform the design of interventions targeted to specific groups at risk of poor outcomes. The objectives of this analysis were to describe patterns of engagement in HIV care over time among individuals receiving combination antiretroviral therapy (cART) and to determine factors associated with transitions among different states of HIV care.
The Canadian Observational Cohort (CANOC) is a collaboration of 8 cohorts from British Columbia (BC), Quebec, and Ontario.6 To be eligible, patients must be HIV-positive Canadian residents older than 18 years who initiated their first antiretroviral regimen composed of at least 3 agents after January 1, 2000 and have at least 1 measurement of HIV plasma viral load (VL) and CD4 cell count within 6 months before initiating cART. Demographic, laboratory, and clinical data of eligible patients were formatted in a standardized fashion at each study site, stripped of identifying information, and compiled at the Project Data Centre in Vancouver, BC. For most Ontario and Quebec participants, start and stop dates of ARV regimens were abstracted from patient charts. For participants from Maple Leaf Medical Clinic in Ontario and participants from BC, start and stop dates were generated using actual prescription records from electronic medical and administrative data, respectively. All participating cohorts have received approval from their institutional ethics boards to contribute non-nominal patient-specific data to CANOC. To be included in this analysis, participants must have at least 1 full year of follow-up after initiation of cART.
States of Care
We developed a system to classify states of HIV care using routine data, such as laboratory tests and medication records. Each year of follow-up after the initial 12-month period following cART initiation was classified into 1 of 5 states of HIV care: (1) HIV care following guidelines (suppressed VL, CD4 >200 cells per cubic millimeter, no gaps in cART >3 months, and no gaps in CD4 or VL measurement >6 months), (2) successful care with decreased frequency of follow-up (as above except no gaps in CD4 or VL measurement >12 months), (3) suboptimal care (unsuppressed VL, CD4 <200 cells per cubic millimeter on 2 consecutive visits, 1 or more gaps in cART >3 months, or 1 or more gaps in VL or CD4 measurement >12 months), (4) loss to follow-up (LTF, no contact for 18 months), and (5) death. These states differentiate individuals who infrequently access care but are otherwise well (state 2) from those who have poor laboratory or clinical outcomes (states 3, 4, and 5).
The frequencies of VL and CD4 count measurement were used as surrogates of access to HIV care,7 with laboratory test results used to classify the clinical status of the patient. A threshold of 6 months between CD4 and VL measurements was chosen as a significant departure from the recommendations to monitor patients every 3 or 4 months.8 Virologic suppression was defined as per United States Department of Health and Human Services guidelines,9 where nonsustained increases in VL were not considered evidence of unsuppressed VL. A value of 200 cells per cubic millimeter for CD4 counts was chosen because the CDC definition for stage 3 HIV infection (AIDS) was based on this threshold10 and because of its association with disease progression11 and increased mortality.12,13 A gap in treatment was defined to be 3 months or more to ensure administrative phenomena resulting from non-uniform prescription refill data were not counted as gaps, and because results from structured treatment interruption trials showed gaps in treatment of 3 months or less, it had no significant impact on future clinical outcomes.14
Each year of follow-up after cART initiation was classified according to the predominant state, that is, for each participant, the state in which he or she spent the greatest time during that year. We classified states of care for each 1 year timeframe rather than for each CD4 or VL measurement to eliminate the dependency of the state classification on time because this would have violated assumptions of multi-state models. The 1 year duration was selected to allow sufficient time for observation of gaps in access to care. States were censored in the last year of follow-up when partial years occurred because of the administrative end of the study period as there was insufficient information to determine the predominant state. Patients who died or were lost to follow-up were classified as such regardless of length of follow-up in the last year.
Fixed covariates were used to examine associations with transitions among care states of demographic and clinical characteristics, including age, sex, ethnicity, HIV risk factors [men having sex with men, people who have injected drugs (PWID)], province, hepatitis C infection, CD4 count at cART initiation, initial third agents, AIDS defining illness at baseline, and calendar year of cART initiation. For the purposes of modeling, missing data were included using indicator variables for unknown race, risk factor, and hepatitis C status.
We compared characteristics among patients included and excluded from the analysis and among patients with different patterns of engagement in care with Wilcoxon rank sum tests for continuous variables and chi-square or Fisher exact tests, as appropriate for categorical variables.
We used multi-state time-homogeneous Markov modeling to identify factors associated with transitioning among care states after the first year of cART using the msm package in R.15–17 These models assume that the sojourn time, that is, the time spent in a state on a single occasion, follows an exponential distribution and does not depend on previous states. Because of small numbers of events, we collapsed LTF and death into a single, absorbing state. We assumed that transitions between the successful decreased follow-up and LTF/death states went through the suboptimal state as too few individuals transitioned directly from successful decreased follow-up to LTF/death. The transition matrix then consisted of 8 possible transitions: state 1 to each state 2, 3, and 4/5; state 2 to states 1 and 3; state 3 to each state 1, 2, and 4/5; as shown in Figure S1, Supplemental Digital Content, http://links.lww.com/QAI/A852. For different values of key covariates, we estimated the mean sojourn times and the probabilities of moving to different states. Age, sex, ethnicity, and PWID were considered for inclusion in the multivariable model based on a priori knowledge.4,18–20 As multi-state models estimate parameters for each transition by each covariate level, inclusion of additional covariates was dependent on model precision and fit. Goodness-of-fit model was assessed by comparing observed and expected numbers of transitions among states using a modified, Pearson-type χ2 test.17,21
We conducted 2 sensitivity analyses. In the first, we included only participants from BC, to determine if the population-based data available for that province influenced the results. In the second, we omitted the criterion requiring a CD4 cell count above 200 cells per cubic millimeter. This analysis was conducted to acknowledge that some patients who initiate cART with low CD4 counts may not achieve this level of immune recovery despite good adherence to cART and regular follow-up with their HIV care provider.22
As of December 2012, data were available for 9694 CANOC participants. We excluded participants who died (n = 236) or were lost to follow-up (n = 237) within the first year of cART initiation, active participants who had less than 1 year of follow-up (n = 697), and participants with a VL <200 copies per millimeter on or before the date of cART initiation as possible non-antiretroviral naïve participants (n = 714). Of the 7810 included CANOC participants, 81% were male and the median age was 40 years (interquartile range: 33–46 years) (Table 1).
The median duration of follow-up was 5.0 years (interquartile range: 2.9–8.1 years). The number and rate of transitions between states are shown in Table S1, Supplemental Digital Content, http://links.lww.com/QAI/A852. Following the first year of cART therapy, 52% of the patients transitioned to guidelines care, 31% to suboptimal, 1.5% to successful decreased engagement, and 3.8% were LTF or died. Of the 2494 patients who transitioned to suboptimal care in the second year of cART, 76% had unsuppressed VL, and 22% had discordant virologic and immunologic responses with suppressed VL and CD4 <200 cells per cubic millimeter. Participants transitioning to guidelines care in the second year of cART were more likely to be male participants, men having sex with men, and have higher baseline CD4 counts than those transitioning to suboptimal care in that year (Table 2).
The estimated sojourn times were 5.17 years [95% (CI): 4.92 to 5.43], 0.72 years (95% CI: 0.66 to 0.78), and 2.72 years (95% CI: 2.61 to 2.83) for the guidelines, successful decreased follow-up, and suboptimal care states, respectively. The estimated probabilities of transitioning among states, according to the care state occupied in the second year of cART, are shown in Table 3. Although most participants who follow guidelines care during the second year of cART are likely to be in guidelines care in the years following, approximately one-fifth of those who are in the successful decreased follow-up state in the second year of cART are likely to be in suboptimal care in subsequent years. The probabilities of being in suboptimal care 1, 2, 5, and 10 years after the second year of cART are 70%, 51%, 27%, and 17%, respectively.
Estimated sojourn times and probabilities of moving among states from univariate models are presented in Table 4. Women spent a mean of 4.04 years in guidelines care with an estimated 48% probability of transitioning to suboptimal care. Men, however, spent a mean of 5.43 years in guidelines care with 42% probability of transitioning to suboptimal care. Similar results were observed between Indigenous and non-Indigenous participants (Table 4). Moreover, Indigenous participants spent 3.93 years on average in suboptimal care compared to 2.70 years for non-Indigenous participants. Similar results were observed for people who have injected drugs compared to those who have not injected drugs.
The multivariable multi-state model is presented in Table 5. Among patients in guidelines care, people who have injected drugs were more likely to transition at any given time to suboptimal care than those who have not injected drugs [hazard ratio (HR) = 1.87, 95% CI: 1.59 to 2.21], those who were older were less likely to transition to suboptimal care (HR = 0.86 per 10 years, 95% CI: 0.78 to 0.95) or successful decrease follow-up care (HR = 0.74 per 10 years, 95% CI: 0.64 to 0.85) than younger individuals, and male patients were less likely than female patients (HR = 0.78, 95% CI: 0.64 to 0.94) to transition to suboptimal care from guidelines care. Among patients in suboptimal care, male patients were more likely to transition at any given time to guidelines care than were female patients (HR = 1.29, 95% CI: 1.14 to 1.46), older patients were more likely than younger individuals to transition to guidelines care (HR = 1.08 per 10 years, 95% CI: 1.03 to 1.14), people who have injected drugs were less likely than people who have not injected drugs to transition to guidelines care (HR = 0.67, 95% CI: 0.60 to 0.75), and people of Indigenous ethnicity were less likely to transition to guidelines care from suboptimal care (HR = 0.70, 95% CI: 0.56 to 0.89) than individuals not of Indigenous ethnicity.
In a sensitivity analysis where guidelines state did not require CD4 cell count >200 cells per cubic millimeter, the clinical inference from the multivariable model remained similar to that of the main model, with the exception that Indigenous individuals were no longer less likely to transition from suboptimal to guidelines care (see Table 2, Supplemental Digital Content, http://links.lww.com/QAI/A852).
In this study of antiretroviral-naive HIV-positive individuals initiating cART since 2000, younger age, female gender, PWID, and Indigenous ethnicity were associated with increased probabilities of transitioning from care meeting guidelines to suboptimal care and reduced likelihood of transitioning from suboptimal care to care meeting guidelines. Participants who started in care meeting guidelines tended to be in this care state in subsequent years while participants who started in suboptimal care gradually transitioned to care meeting guidelines over time. One-fifth of participants starting in successful decreased follow-up were projected to be in suboptimal care 1, 2, 5, and 10 years later, suggesting that interventions to re-engage individuals at high risk of transitioning to suboptimal care are required.
To our knowledge, this is the first study to use multi-state models to describe associations of patient characteristics with bidirectional transitions among states of HIV care engagement over time, as recently suggested by Powers and Miller.23 There have been a number of studies examining associations between patient characteristics and the HIV “cascade of care” and retention.4,18–20 Our findings were similar to those of Burchell et al,18 who found that younger individuals, women, Indigenous people, and people who have injected drugs had lower continuous engagement in care, ART use, and VL suppression. A similar study of HIV-positive individuals in care in North America between 2000 and 2008 also reported that younger individuals and people who have injected drugs were at greater risk of incomplete care; however, women were more likely to be engaged in complete care.4
Our work has several important clinical implications. Notably, our findings that women and people who have injected drugs were more likely to transition to suboptimal care following the first year of antiretroviral therapy highlight subgroups of vulnerable individuals for whom interventions to maximize successful engagement in care are warranted. Although further research is required to characterize the reasons these subgroups of individuals transition to suboptimal care shortly after initiating antiretroviral therapy, possible interventions could include enhanced integration of addiction management into HIV primary care, peer patient navigators, and assistance with costs associated with child care and transportation to appointments.
Using the multi-state model framework, we were also able to estimate the time spent in each care state and the probabilities of transitioning among states according to patient characteristics. These estimates are useful for the design of interventions to enhance engagement in care. Further strengths of the multi-state framework include flexibility in the definitions of both states and trajectories of transitions among states.24 For example, patients can be assumed to transition through suboptimal care before being lost to follow-up, as per physician experience, or state definitions can exclude CD4 criterion, as in our sensitivity analysis. Multi-state models also allow for censoring of the state itself, when the patient is known to be alive but there is insufficient information available to classify the individuals' care state, as was done in our analysis.
There are some limitations to multi-state models. First, a modest number of associations could be evaluated in a multivariable model, as coefficients for each covariate were estimated for each transition, and the power of multi-state models is related to the number of observed transitions. Second, the “memoryless” property that the future state depends only on the current state and time17 does not allow the incorporation of previous, clinically informative CD4 cell counts in the model. Third, we assumed time homogeneity in our models. To ensure that the definitions of states did not depend on time, we classified participants within each year of follow-up, according to the state they were in for the greatest period during that year. This “predominant” care state was chosen over the first or last state held within the year because it was felt to be more representative of the clinical care received by the individual. Future work could explore the use of hidden continuous-time Markov models to address possible misclassification of the states.24 Despite these limitations, we believe the approach of examining transitions among states is useful. Multi-state models have been used in disease progression modeling of HIV.25–27
Strengths of our study include its large sample size, universal health care setting, homogeneity of participants with regard to era of antiretroviral therapy exposure and diversity of participants regarding geographic area, HIV risk factor, age, and gender. A limitation of our study was the large amount of missing data on ethnicity and HIV risk factor and lack of data on socioeconomic status, adherence, antiretroviral drug coverage, and injection drug use over time. Furthermore, we were unable to distinguish CD4 and VL measurements ordered as part of routine clinical care from those ordered because of hospitalizations, emergent care, or unscheduled visits with patients' regular physicians. Consequently, gaps in routine, structured care may be underestimated. We were also unable to distinguish whether decreases in the frequency of follow-up were because of the patient's or the physician's initiative. There were important differences among provinces regarding data collection and participant enrolment. The BC cohort is population-based, including all individuals prescribed cART since 2000. The cohorts from Ontario and Quebec are clinic-based and are not necessarily representative of all HIV-positive individuals who initiated cART since 2000 in these provinces. Transfers between clinics are captured in 2 cohorts, BC and the Ontario HIV Treatment Network Cohort Study. Transfers from other cohorts are not captured or for individuals transferring out of province. LTF may therefore be overestimated. Deaths in BC were ascertained through monthly linkage with the BC Vital Statistics Database, whereas deaths in Ontario were passively reported to individual clinics and may be undercounted. However, a sensitivity analysis that included only participants from BC yielded similar results (data not shown). Generalizability of our findings is limited to patients receiving care in a universal health care setting and to patients who remained engaged in care for at least 1 year. Finally, because our cohort included only HIV-positive individuals who have initiated cART, we were not able to address factors earlier in the “treatment cascade” associated with HIV testing and initial engagement in HIV care.
In conclusion, we have developed a flexible framework and model that characterizes patient transitions among states of HIV clinical care. Our work provides evidence that patterns of engagement and adherence are often established within the first year of cART. We also found that incomplete adherence to monitoring guidelines can be meaningfully separated into that occurring within the context of successful therapy and that occurring within the context of failing therapy, with very different management implications. These findings are relevant for policy, planning, and recommendations for care and highlight the unique information that can be gleaned from using multi-state models to evaluate engagement in HIV care.
The authors would like to thank all the participants for allowing their information to be a part of the CANOC collaboration.
The CANOC Collaborative Research Centre includes the following: CANOC Nominated Principal Investigator: Robert Hogg (British Columbia Centre for Excellence in HIV/AIDS, Simon Fraser University); Principal Investigators: Ann N. Burchell [St. Michael's Hospital, University of Toronto], Curtis Cooper (University of Ottawa, OCS), Deborah Kelly (Memorial University of Newfoundland), Marina Klein (Montreal Chest Institute Immunodeficiency Service Cohort, McGill University), Mona Loutfy (University of Toronto, Maple Leaf Medical Clinic, OCS), Nima Machouf (Clinique Medicale l'Actuel, Université de Montréal), Julio Montaner (British Columbia Centre for Excellence in HIV/AIDS, University of British Columbia), Janet Raboud (University of Toronto, University Health Network, OCS), Christos Tsoukas (McGill University), Stephen Sanche (University of Saskatchewan), Alexander Wong (University of Saskatchewan),Tony Antoniou (St. Michael's Hospital, University of Toronto, Institute for Clinical Evaluative Sciences), Ahmed Bayoumi (St. Michael's Hospital, University of Toronto), Mark Hull (British Columbia Centre for Excellence in HIV/AIDS), and Bohdan Nosyk (British Columbia Centre for Excellence in HIV/AIDS, Simon Fraser University); Co-investigators: Angela Cescon (Northern Ontario School of Medicine), Michelle Cotterchio (Cancer Care Ontario, University of Toronto), Charlie Goldsmith (Simon Fraser University), Silvia Guillemi (British Columbia Centre for Excellence in HIV/AIDS, University of British Columbia), P. Richard Harrigan (British Columbia Centre for Excellence in HIV/AIDS, University of British Columbia), Marianne Harris (St. Paul's Hospital), Sean Hosein (CATIE), Sharon Johnston (Bruyère Research Institute, University of Ottawa), Claire Kendall (Bruyère Research Institute, University of Ottawa), Clare Liddy (Bruyère Research Institute, University of Ottawa),Viviane Lima (British Columbia Centre for Excellence in HIV/AIDS, University of British Columbia), David Marsh (Northern Ontario School of Medicine), David Moore (British Columbia Centre for Excellence in HIV/AIDS, University of British Columbia), Alexis Palmer (British Columbia Centre for Excellence in HIV/AIDS, Simon Fraser University), Sophie Patterson (British Columbia Centre for Excellence in HIV/AIDS, Simon Fraser University), Peter Phillips (British Columbia Centre for Excellence in HIV/AIDS, University of British Columbia), Anita Rachlis (University of Toronto, OCS), Sean B. Rourke (University of Toronto, OCS), Hasina Samji (British Columbia Centre for Excellence in HIV/AIDS), Marek Smieja (McMaster University), Benoit Trottier (Clinique Medicale l'Actuel, Université de Montréal), Mark Wainberg (McGill University, Lady Davis Institute for Medical Research), and Sharon Walmsley (University Health Network, University of Toronto); Collaborators: Chris Archibald (Public Health Agency of Canada Centre for Communicable Diseases and Infection Control), Ken Clement (Canadian Aboriginal AIDS Network), Monique Doolittle-Romas (Canadian AIDS Society), Laurie Edmiston (Canadian Treatment Action Council), Sandra Gardner (OHTN, University of Toronto, OCS), Brian Huskins (Canadian Treatment Action Council), Jerry Lawless (University of Waterloo), Douglas Lee (University Health Network, University of Toronto, ICES), Renee Masching (Canadian Aboriginal AIDS Network), Stephen Tattle (Canadian Working Group on HIV & Rehabilitation), and Alireza Zahirieh (Sunnybrook Health Sciences Centre); Analysts and Staff: Claire Allen (Regina General Hospital), Stryker Calvez (SHARE), Guillaume Colley (British Columbia Centre for Excellence in HIV/AIDS), Jason Chia (British Columbia Centre for Excellence in HIV/AIDS), Daniel Corsi (The Ottawa Hospital Immunodeficiency Clinic, Ottawa Hospital Research Institute), Louise Gilbert (Immune Deficiency Treatment Centre), Nada Gataric (British Columbia Centre for Excellence in HIV/AIDS), Alia Leslie (British Columbia Centre for Excellence in HIV/AIDS), Lucia Light (OHTN), David Mackie (The Ottawa Hospital), Costa Pexos (McGill University), Susan Shurgold (British Columbia Centre for Excellence in HIV/AIDS), Leah Szadkowski (University Health Network), Chrissi Galanakis (Clinique Médicale L'Actuel), Benita Yip (British Columbia Centre for Excellence in HIV/AIDS), Jaime Younger (University Health Network), and Julia Zhu (British Columbia Centre for Excellence in HIV/AIDS).
1. Hogg RS, Heath K, Bangsberg D, et al. Intermittent use of triple-combination therapy is predictive of mortality at baseline and after 1 year of follow-up. AIDS. 2002;16:1051–1058.
2. Bangsberg DR, Perry S, Charlebois ED, et al. Non-adherence to highly active antiretroviral therapy predicts progression to AIDS. AIDS. 2001;15:1181–1183.
3. Cohen SM, Van Handel MM, Branson BM, et al. Center for disease Control and prevention (CDC). Vital signs: HIV prevention through care and treatment—United States. MMWR Morb Mortal Wkly Rep. 2011;60:1618–1623.
4. Rebeiro P, Althoff KN, Buchacz K, et al. Retention among North American HIV–infected persons in clinical care, 2000–2008. J Acquir Immune Defic Syndr. 2013;62:356–362.
5. Gill MJ, Krentz HB. Unappreciated epidemiology: the churn effect in a regional HIV care programme. Int J STD AIDS. 2009;20:540–544.
6. Palmer A, Klein M, Raboud J, et al. Cohort Profile: the Canadian observational cohort collaboration. Int J Epidemiol. 2011;40:25–32.
7. Mugavero MJ, Davila JA, Nevin CR, et al. From access to engagement: measuring retention in outpatient HIV clinical care. AIDS Patient Care STDS. 2010;24:607–613.
8. Antiretroviral (ARV) treatment of adult HIV infection. Therapeutic Guidelines Committee, British Columbia Centre for Excellence (BC CfE), British Columbia, Canada; 2013. Available at: http://www.cfenet.ubc.ca/therapeutic-guidelines/adult
. Accessed February, 2013.
9. Panel on Antiretroviral Guidelines for Adults and Adolescents. Guidelines for the Use of Antiretroviral Agents in HIV-1-Infected Adults and Adolescents. Department of Health and Human Services; 2013. Available at: https://aidsinfo.nih.gov/contentfiles/adultandadolescentgl003371.pdf
. Accessed July, 2013.
10. Selik RM, Mokotoff ED, Branson B, et al. Revised Surveillance Case definition for HIV infection–United States. MMWR Recomm Rep. 2014;63:1–10.
11. Smurzynski M, Wu K, Benson CA, et al. Relationship between CD4+ T-cell counts/HIV-1 RNA plasma viral load and AIDS-defining events among persons followed in the ACTG longitudinal linked randomized trials study. J Acquir Immune Defic Syndr. 2010;55:117–127.
12. Loutfy MR, Genebat M, Moore D, et al. A CD4+ cell count <200 cells per cubic millimeter at 2 years after initiation of combination antiretroviral therapy is associated with increased mortality in HIV-infected individuals with viral suppression. J Acquir Immune Defic Syndr. 2010;55:451–459.
13. Engsig FN, Zangerle R, Katsarou O, et al. Long-term mortality in HIV-positive individuals virally suppressed for >3 years with incomplete CD4 recovery. Clin Infect Dis. 2014;58:1312–1321.
14. Walmsley SL, Thorne A, Loutfy MR, et al. A prospective randomized controlled trial of structured treatment interruption in HIV-infected patients failing highly active antiretroviral therapy (Canadian HIV Trials Network Study 164). J Acquir Immune Defic Syndr. 2007;45:418–425.
15. Cox D, Miller H. The Theory of Stochastic Processes. 1st ed. London, UK: Chapman and Hall; 1965.
16. Kalbfleisch JD, Lawless JF. The analysis of panel data under a Markov assumption. JASA. 1985;80:863–871.
17. Jackson CH. Multi-state models for panel data: the msm package for R. J Stat Softw. 2011;38:1–28.
18. Burchell AN, Gardner S, Light L, et al. Engagement in HIV care among persons enrolled in a clinical HIV cohort in Ontario, Canada, 2001-2011. J Acquir Immune Defic Syndr. 2015;70:e10–e19.
19. Fleishman JA, Yehia BR, Moore RD, et al. Establishment, retention, and loss to follow-up in outpatient HIV care. J Acquir Immune Defic Syndr. 2012;60:249–259.
20. Yehia BR, Fleishman JA, Metlay JP, et al. Comparing different measures of retention in outpatient HIV care. AIDS. 2012;26:1131–1139.
21. Aguirre-Hernandez R, Farewell VT. A Pearson-type goodness of fit test for stationary and time-continuous Markov regression models. Stat Med. 2002;21:1899–1911.
22. Kelley CF, Kitchen CM, Hunt PW, et al. Incomplete peripheral CD4+ cell count restoration in HIV-infected patients receiving long-term antiretroviral treatment. Clin Infect Dis. 2009;48:787–794.
23. Powers KA, Miller WC. Building on the HIV cascade: a complementary “HIV States and Transitions” framework for describing HIV diagnosis, care and treatment at the population level. J Acquir Immune Defic Syndr. 2015;69:341–347.
24. Hougaard P. Multi-state models: a review. Lifetime Data Anal. 1999;5:239–264.
25. Gentleman RC, Lawless JF, Lindsey JC, et al. Multi-state Markov models for analyzing incomplete disease history data with illustrations for HIV disease. Stat Med. 1994;13:805–821.
26. Ndumbi P, Gillis J, Raboud JM, et al. Clinical impact of altered T-cell homeostasis in treated HIV patients enrolled in a large observational cohort. AIDS. 2013;27:2863–2872.
27. Kousignian I, Abgrall S, Duval X, et al. Modeling the time course of CD4 T-lymphocyte counts according to level of virologic rebound in HIV-1-infected patients on highly active antiretroviral therapy. J Aquir Immune Defic Syndr. 2003;35:50–57.