icon-folder.gif   Conference Reports for NATAP  
 
  11th International Workshop
on HIV and Aging
30 September - 2 October 2020
Virtual
Back grey_arrow_rt.gif
 
 
Latent Class Analysis--Unique Predictor
of Hospitalization, Death in Older HIV+

 
 
  11th International Workshop on HIV & Aging Virtual Meeting, September 30 to October 2, 2020
 
By Mark Mascolini for NATAP and Virology Education
 
Latent class analysis, a multilayered, data-driven clinical classification tool, predicted hospital admission and death in 6 groups of older people with HIV identified by a composite score reflecting health conditions present before a given date [1]. Researchers from the Marcus Institute for Aging Research and collaborators at 3 New England centers argued that the method is "ideal" for assigning risk categories to older people with HIV.
 
According to the UCLA Institute for Digital Research & Education, latent class analysis "is a statistical method for identifying unmeasured class membership among subjects using categorical and/or continuous observed variables" [2]. The method can identify previously unobservable subgroups in a population [3].
 
Because patterns of mortality and healthcare use are complex in older groups with HIV, the Marcus Institute collaborators proposed that latent class analysis may offer a more precise method to predict risk of hospital admission or death. They used 2014 Medicare/Medicaid data on chronic conditions and beneficiaries to identify US residents diagnosed with HIV infection before January 2014. Demographics and date of death came from the Medicare/Medicaid Master Beneficiary Summary File. The researchers used 2014 Medicare Part A claims to tabulate hospital admissions. Reviewing published literature on older people with HIV, the researchers outlined 8 distinct health domains that can be scored to reflect the number of conditions a person had before January 2014:
 
-- Mental health index scored from 0 to 6 (depending on number of individual conditions)
-- Noncardiovascular chronic condition index scored from 0 to 11
-- Substance use index scored from 0 to 4
-- Cardiovascular condition index scored from 0 to 9
-- Sensory deficits index scored from 0 to 4
-- Musculoskeletal deficits index scored from 0 to 10
-- Pulmonary deficits index scored from 0 to 2
-- Learning disabilities and neurologic deficits index scored from 0 to 9
 
The investigators used latent class analysis adjusted for age, sex, and race to identify best model fits. They used multinomial logistic regression to predict latent class membership by demographics. Cox proportional hazards regression explored associations between latent classes and death in 1 year of follow-up. And zero-inflated Poisson regression assessed associations between latent class and hospital admission.
 
Among 17,666 HIV-positive people, the researchers identified 6 latent classes:
 
-- Class 1, Healthy, included 10,534 people with average index scores below 1.0 for all 8 indices listed above.
-- Class 2, Mostly healthy, low comorbidity, included 2220 people.
-- Class 3, Low mental health burden with moderate comorbidity, included 1783 people. -- Class 4, Moderate mental health burden and substance use disorder and low comorbidity, included 1178 people.
-- Class 5, Low mental health burden with high comorbidity, included 1428 people.
-- Class 6, High mental health burden with substance use disorder and high comorbidity, included 523 people.
 
Average age was significantly older in Class 2 (64.48 years) and Class 5 (64.41) than in the other classes (range 49.21 to 57.39) (P < 0.001). Classes 1 and 4 had significantly higher proportions of men (75% and 77%) than the other classes (range 60% to 65%) (P < 0.001). Proportions of people who died were 2% in Class 1, 6% in Class 2, 8% in Class 3, 5% in Class 4, 10% in Class 5, and 10% in Class 6 (P < 0.001). Respective average hospital admission counts were 0.18, 0.66, 0.95, 1.25, 0.96, and 3.26 (P < 0.001).
 
Cox regression analysis determined that, compared with people in Class 1, those in all other classes had an independently higher risk of dying in 2014, as indicated by the following hazard ratios (and 95% confidence intervals):
 
-- Class 2: 2.84 (2.28 to 3.54)
-- Class 3: 3.98 (3.21 to 4.95)
-- Class 4: 2.53 (1.89 to 3.41)
-- Class 5: 4.78 (3.83 to 5.96)
-- Class 6: 5.25 (3.82 to 7.21)
 
People over 50 years old had an independently higher risk of dying in 2014, and women had an independently lower risk of dying.
 
Compared with Class 1, all other classes had an independently higher expected change in 1-year hospital admission rate, about 2-fold higher for Classes 2 through 5 and about 4-fold higher for Class 6.
 
The researchers underlined three other findings of their study: (1) Mental health and substance use, which helped define Classes 4 and 6, raised the risk of hospital admission 2.3- and 4.0-fold in those classes compared with Class 1. (2) A heavy comorbidity toll, usually cardiovascular disease, was linked to a 4-fold higher risk of 1-year mortality in Class 3 and a 4.7-fold higher risk in Class 5, compared with Class 1. (3) The oldest Classes, 2 and 5, had similar hospital admission rates, but Class 5 had a much higher 1-year death risk than Class 2.
 
The data-driven nature of latent class analysis, the researchers proposed, makes it the "an ideal method by which to assign older people living with HIV to risk categories for the purpose of directing resources and providers best suited to tailor care focused on reducing hospitalizations and mortality."
 
References
1. Olivieri-Mui B, McCarthy E, Shi S, Montano M, Wilson I, Kim D. Categorizing risk for older persons living with HIV: a latent class analysis of comorbidities. 11th International Workshop on HIV & Aging Virtual Meeting, September 30 to October 2, 2020. Abstract 7.
2. UCLA Institute for Digital Research & Education. Statistical consulting. Latent class analysis in MPLUS. https://stats.idre.ucla.edu/mplus/seminars/lca/ 3. PennState College of Health and Human Development. The Methodology Center. Latent class modeling. https://www.methodology.psu.edu/ra/lca/