Basic subsistence needs and overall health among human immunodeficiency virus-infected homeless and unstably housed women.

Abstract

Some gender differences in the progression of human immunodeficiency virus (HIV) infection have been attributed to delayed treatment among women and the social context of poverty. Recent economic difficulties have led to multiple service cuts, highlighting the need to identify factors with the most influence on health in order to prioritize scarce resources. The aim of this study was to empirically rank factors that longitudinally impact the health status of HIV-infected homeless and unstably housed women. Study participants were recruited between 2002 and 2008 from community-based venues in San Francisco, California, and followed over time; marginal structural models and targeted variable importance were used to rank factors by their influence. In adjusted analysis, the factor with the strongest effect on overall mental health was unmet subsistence needs (i.e., food, hygiene, and shelter needs), followed by poor adherence to antiretroviral therapy, not having a close friend, and the use of crack cocaine. Factors with the strongest effects on physical health and gynecologic symptoms followed similar patterns. Within this population, an inability to meet basic subsistence needs has at least as much of an effect on overall health as adherence to antiretroviral therapy, suggesting that advances in HIV medicine will not fully benefit indigent women until their subsistence needs are met.

Full text

Page 1

American Journal of Epidemiology
ª The Author 2011. Published by Oxford University Press on behalf of the Johns Hopkins Bloomberg School of Public Health.
This is an Open Access article distributed under the terms of the Creative Commons Attribution Non-Commercial
License (http://creativecommons.org/licenses/by-nc/2.5), which permits unrestricted non-commercial use,
distribution, and reproduction in any medium, provided the original work is properly cited.
Vol. 174, No. 5
DOI: 10.1093/aje/kwr209
Advance Access publication:
July 11, 2011
Original Contribution
Basic Subsistence Needs and Overall Health Among Human Immunodeficiency
Virus-infected Homeless and Unstably Housed Women
Elise D. Riley*, Kelly Moore, James L. Sorensen, Jacqueline P. Tulsky, David R. Bangsberg, and
Torsten B. Neilands
* Correspondence to Dr. Elise Riley, University of California at San Francisco, UCSF Box 0874, San Francisco, CA 94143-0874
(e-mail: eriley@epi-center.ucsf.edu).
Initially submitted August 16, 2010; accepted for publication April 8, 2011.
Some gender differences in the progression of human immunodeficiency virus (HIV) infection have been attrib-
uted to delayed treatment among women and the social context of poverty. Recent economic difficulties have led to
multiple service cuts, highlighting the need to identify factors with the most influence on health in order to prioritize
scarce resources. The aim of this study was to empirically rank factors that longitudinally impact the health status of
HIV-infected homeless and unstably housed women. Study participants were recruited between 2002 and 2008
from community-based venues in San Francisco, California, and followed over time; marginal structural models
and targeted variable importance were used to rank factors by their influence. In adjusted analysis, the factor with
the strongest effect on overall mental health was unmet subsistence needs (i.e., food, hygiene, and shelter needs),
followed by poor adherence to antiretroviral therapy, not having a close friend, and the use of crack cocaine.
Factors with the strongest effects on physical health and gynecologic symptoms followed similar patterns. Within
this population, an inability to meet basic subsistence needs has at least as much of an effect on overall health as
adherence to antiretroviral therapy, suggesting that advances in HIV medicine will not fully benefit indigent women
until their subsistence needs are met.
health status; HIV; homeless persons; women
Abbreviations: AIDS, acquired immunodeficiency syndrome; HIV, human immunodeficiency virus; OR, odds ratio; tVIM, targeted
variable importance.
Editor’s note: An invited commentary on this article ap-
pears on page 523, and the authors’ response is published
on page 526.
The National Health and Nutrition Examination Survey
(NHANES) indicates that women and low-income persons
are at increased risk of reporting poor health-related quality
of life (1). Homeless persons, in particular, often suffer from
serious mental and physical health problems that are attrib-
uted to numerous health risks, as well as low social support
and a lack of social resources (2), particularly among
women (3). It has been suggested that the face of homeless-
ness is changing in some geographic areas from adult male
alcoholics to an increasingly diverse population with com-
plexmedicalillnesses,includingpsychiatricconditions,phys-
ical disability, consequences of drug and alcohol abuse, and
violence (4). As part of this changing social landscape, fe-
male gender is now one of the strongest predictors of poor
health among homeless persons (5, 6).
Human immunodeficiency virus (HIV) infection intro-
duces another layer of complexity for vulnerable persons.
HIVis increasingly characterized as a chronic condition that
can be managed through adherence to a healthy lifestyle,
complex drug regimens, and treatment monitoring; how-
ever, a person’s housing status can be a significant determi-
nant of an individual’s ability to meet these requirements
and achieve better health status (7). In addition, the effects
of complex and interconnected risks facing homeless per-
sons may be modified by gender. For instance, some gender
515 Am J Epidemiol. 2011;174(5):515–522

Page 2

differences in the progression of HIV have been attributed to
delayed and fragmented treatment among women (8, 9) and
the social context of poverty (10). This social context high-
lights issues such as drug use, which is an established risk
factor for both HIV infection and homelessness among
women (11, 12).
Improved antiretroviral medications have led to an era in
which HIV has become a more manageable chronic illness,
with increased attention to conditions beyond disease pro-
gression, such as the development of non-acquired immu-
nodeficiency syndrome (AIDS)-defining cancers (13). The
effects of exposures that change over time and influence one
another, such as drug use and social support structures, now
have the opportunity to influence a longer disease course.
Few studies have considered the varying levels of these
exposures over time among unstably housed persons. Addi-
tionally, few studies regarding the impact of competing
needs on the health status of HIV-infected persons have
considered data collected exclusively from unstably housed
persons, and few have recruited participants from noninsti-
tutional settings, which systematically excludes those out-
side the health-care system.
A better understanding of the numerous and changing
health risks experienced by HIV-infected, unstably housed
persons has the potential to improve the delivery of health
and social services; however, funding for programs serving
vulnerable populations, such as public health, mental health,
and drug treatment programs, has been cut drastically (14–
17). Understanding that each factor is important but recog-
nizing program limitations, we aimed first in the current
study to determine the extent to which changing risks influ-
enced the health status of HIV-positive, unstably housed
women over time and then empirically to rank risk expo-
sures by their level of influence.
MATERIALS AND METHODS
Participants
From July 2002 to September 2008, a mobile outreach
team recruited a probability sample of homeless and unsta-
bly housed biologic women from San Francisco homeless
shelters, free-food programs, and a random sample of low-
income hotels in 3 neighborhoods selected with probability
proportional to the number of residents (18, 19). In addition,
informational flyers were posted in these venues. All indi-
viduals who interfaced with outreach staff, obtained study
information at these venues, or were referred by venue staff
were eligible to participate in screening activities. The only
inclusion criterion was HIV-positive serostatus. Study par-
ticipants were part of ‘‘Shelter, Health, and Drug Outcomes
among Women’’ (SHADOW), an ongoing cohort study that
assesses the impact of competing risks on the health and
well-being of poor and unstably housed women.
Procedures
Participants were asked to visit a community-based field
site every3 months. There,theyprovided blood for CD4 cell
count and viral load assessments and completed a confiden-
tial interviewer-administered questionnaire to assess the
presence and extent of factors that might influence health
status. All study procedures were conducted with the ap-
proval of the Committee on Human Research at the Univer-
sity of California at San Francisco.
Outcome variables: health status
The Medical Outcome Study’s Short Form 36 (SF-36)
(20) is a reliable and valid measure of health in this popu-
lation (21) and was used to assess overall health; the Mental
Composite Score and the Physical Composite Score ranged
from 0 to 100, where higher scores indicated better health.
In addition, given a more recent focus on gynecologic health
that has been made relevant by improved HIV therapies
(22), the presence of gynecologic symptoms was assessed.
A binary variablewas based on a panel of gynecologic symp-
toms and conditions developed by Wenzel et al. (23) and
included abnormal vaginal discharge, severe pelvic pain,
burning during urination, blood in the urine, skipped period,
abnormally heavy periods, and new warts, sores, or lumps
on the genitals. Thus, although better health was indicated
by a unit increase in the Mental Composite Score and the
Physical Composite Score, it was indicated by a decrease
in (the absence of) gynecologic symptoms; that is, better
health was indicated by higher linear estimates for the
Mental Composite Score and the Physical Composite Score
but lower odds ratios for gynecologic symptoms.
Main effects: competing health risks
Variables considered as having potential influence on
health status have been previously established as important
factors in predicting health. These factors included socio-
demographic variables; unmet subsistence needs (difficulty
gaining access to a bathroom, place to wash, clothing, food,
or a place to sleep, regardless of where the respondent slept)
(24); social support (having at least 1 close friend/confidant)
and instrumental support (having someone who would lend
the respondent money or give her a place to sleep) (25);
heavy alcohol use (>1 drink/day) (26); any use of heroin,
crack cocaine, or methamphetamine; symptoms of with-
drawal from heroin, crack cocaine, methamphetamine, and
alcohol detected in the Diagnostic Interview Schedule (27);
exchange of sex for money or drugs (28); use of and adher-
ence to antiretroviral therapy; and chronic health conditions
experienced during the study (asthma, diabetes, heart dis-
ease, high blood pressure, and/or emphysema). The CD4
cell count and viral load are intermediate variables on the
causal path between study variables and overall health sta-
tus. Thus, in accordance with recommendations regarding
overadjustment bias (29, 30), these variables were not con-
sidered potential candidates for the treatment model.
Health services were not adjusted in the current study, as
service use is a consequence of poor health and not a risk
factor.
Statistical methods
Targeted variable importance (tVIM) was used to analyze
data. tVIM assesses the effects of large numbers of variables
516Riley et al.
Am J Epidemiol. 2011;174(5):515–522

Page 3

with unknown or diverse correlation structures; it more ac-
curately assesses effects when compared with techniques
that rely on parametric regression models (31–34). tVIM
estimates the effect of 1 variable at a time, which tailors
the estimation approach toward the specific effect of inter-
est, thereby providing more accurate effects and assessment
of uncertainty. This approach is important for the analyses
described herein, because different types and a broad spec-
trum of variables were analyzed for the current study
and, thus, a single multivariate regression model approach
would be impossible. tVIM involves 2 steps: the first step
to estimate the population effect for each variable of in-
terest by using a marginal structural model approach to
estimate the parameters, and the second step to rank sig-
nificant risk factors by the strength of the effect on the
outcome.
First,marginalstructuralmodelswereusedtoestimatethe
population effect (35) of each risk factor, adjusting for con-
founding (36) via the general formula logit(P(Ya(t)¼ 1) ¼
m(a(t ? 1)jb) ¼ b0þ b1a(t ? 1) for binary variables and
E(Ya(t)) ¼ m(a(t ? 1)ja) ¼ a0þ a1a(t ? 1) for continuous
variables(Figure1).Confounderswereassessedfor eachrisk
factor separately to obtain estimates of the adjusted effects
on health status. Also, variables were considered potential
confounders in all models for which they were not being
considered the primary effect. Standard errors were calcu-
lated through bootstrap methods, which do not treat weight-
ing as implicitly fixed but as stochastic (37). In addition, the
following variables were considered potential confounders
for all analyses: 1) to account for secular trends, the year
of study participation was considered; 2) to account for
potential study effects, the number of study participation
months was considered; 3) in recognition that age is an
established predictor of health status, age was considered;
and 4) in consideration of the potential cumulative effects
of drug use, an individual’s drug use history was consid-
ered a potential confounder in order to focus on the impact
of current drug use only. To ensure that the exposure pre-
ceded the outcome (an assumption of risk and the statisti-
cal models that estimate it), the effect of a change in the
level of the exposure on the outcome in the current quarter
was assessed (i.e., a 1-unit change for the Mental Composite
Score and the Physical Composite Score or a change in
odds ratios for gynecologic symptoms), adjusting for all
other significant confounding variables from the prior
quarters.
Next, tVIM techniques were applied such that the risk
factor-specific effects obtained by the marginal structural
models were ranked by P value, which was appropriate
for the current study because the exposure variables (i.e.,
risk factors) had different units of measure. Because the
population and sample size were consistent between
models, ranking variables based on the P value was a stan-
dardized approach to ranking effect estimates (i.e., signal-
to-noise ratio). Thus, ranking is not from the most negative
to the most positive effect; rather, it is from thevariable with
the largest population-level effect on the outcome to that
with the smallest. With this approach, variables with the
strongest individual influence will not necessarily have the
largest population-level effect if their prevalence is low. If
all covariates represent all confounding, then the tVIM es-
timate can be interpreted as the average causal effect of that
exposure on the outcome. If this assumption is untenable,
then the tVIM estimate can be interpreted as a summary
measure of the importance of the variable after controlling
for covariates.
Because participants were recruited into the study at dif-
ferent time points, a staggered entry approach was used
where follow-up time began at the time of study entry for
each individual. Because individuals belonged to different
comparison groups as the risk factors changed, estimates
yielded by each model represented the overall average effect
A0
A1
Y2
L1
L0
t1
Y1
t2
t3
Figure 1.
housed women living in San Francisco, California, 2002–2008. A0and A1represent measurements of a single exposure variable at 2 consecutive
timepoints (t1and t2). L0and L1denote measured confounders that are associated with both A0and A1and, therefore, are included in the treatment
model. Y1and Y2represent the health status at 2 consecutive timepoints (t2and t3). It is important to note that marginal structural models account
for changes in confounding over time and account for the fact that the outcome can influence the exposure.
Causal diagram for the effect of competing social needs on the health status of human immunodeficiency virus-positive, unstably
Basic Subsistence Needs and Overall Health517
Am J Epidemiol. 2011;174(5):515–522

Page 4

of each risk factor, adjusting for measured confounders. To
account for missing values, we used a forward imputation
method in which the most recent previously observed value
for that participant was analyzed, which is a common ap-
proach in analyses that include marginal structural modeling
(38–40). If no previous value was observed, the median
population value was used for continuous variables, and
categorical variables were imputed at the category with
the highest proportion of observed values. This approach
was chosen because it does not rely on modeling assump-
tions to predict missing information and ensures that
original data are preserved.
RESULTS
Among the 249 female adults who were screened for
study participation, 133 were confirmed as HIV infected
and, thus, eligible to participate. A 3% refusal rate resulted
in a cohort of 129 HIV-infected women. Loss to follow-up
was approximately 2% per year, and the mortality rate was
approximately 0.5% per year; the median follow-up time
was 10 months per person. Outcome variables did not con-
tain any missing values, and missing data for exposure vari-
ables did not exceed 6.5%.
The median age of study participants was 44 years, 42%
were African American, 29% Caucasian, 5% Latina, and
14% ‘‘other’’ (Table 1). At baseline, 33% of respondents
reported the use of crack cocaine, and 20% reported sleep-
ing on the street or in a homeless shelter during the past
3 months. Unmet subsistence needs were reported by 25%
of the population. Having no sources of instrumental sup-
port was reported by 20% of study participants; any in-
surance coverage was reported by 97% of participants
(89% Medicaid; 15% Veterans; 10% private). At baseline,
the median viral load was 1,105 copies/mL, the median
CD4 cell count was 385 cells/lL, and 57% of participants
reported at least 1 chronic health condition. Highlighting
the importance of change over time, 52% of participants
took antiretroviral therapy at baseline, while 71% took anti-
retroviral therapy at some point during the study; 31% of
those on antiretroviral therapy at baseline reported ?90%
adherence, while 72% were >90% adherent at some point
in the study; 3% experienced cocaine withdrawal, and 1%
experienced alcohol withdrawal at baseline, while 11% and
20% experienced cocaine and alcohol withdrawal, respec-
tively, at some point in the study.
Exploratory factor analysis of the 5-item subsistence
needs variable measured at baseline was performed with
Mplus, version 6, software (Muthen & Muthen, Los An-
geles, California). A single dominant factor emerged (ei-
genvalue ¼ 4.68) with all other eigenvalues being negligible
in size (?0.25). The fit of a 1-factor model to the data was
excellent, as indicated by the v2test of exact fit (v25df¼
3.64; P ¼ 0.60). Factor loadings for all items exceeded 0.85,
demonstrating strong factor-item relations for all 5 items.
Reliability for this factor’s items was excellent (Cronbach’s
a ¼ 0.90). These items were, therefore, considered to com-
prise a subsistence needs scale whose composite score was
used in the analyses described below.
Table 1.
of HIV-infected Homeless and Unstably Housed Women (n ¼ 129)
Living in San Francisco, California, 2002–2008
Baseline Characteristics and Behaviors (Past 3 Months)
Median%
Socioeconomic
Age, years 44
Graduated from high school40
Race
African American52
Caucasian29
Latina5
Other14
Minor children living at home7
Overnight jail/prison stay7
Slept on the street or in a homeless
shelter
20
Employed5
Any income from SSI or SSDI80
Monthly income, $812
Drug and alcohol use (past 3 months)
Crack cocaine 33
Heroin 15
Methamphetamine1
Cocaine withdrawal3
Heavy alcohol use (>1 drink/day) 14
Alcohol withdrawal 1
Subsistence needs and social support
(past 3 months)
Difficulty meeting basic subsistence
needsa
25
At least one close friend/confidant69
No sources of instrumental support20
Health
CD4 cell count, cells/lL385
Viral load, copies/mL 1,105
Taking antiretroviral therapy52
>90% adherence (among those taking
ART)
31
Any chronic health conditionb
57
Physical health composite scorec
41/100
Mental health composite scorec
42/100
Any gynecologic symptomsd
43
Abbreviations: ART, antiretroviral therapy; HIV, human immunode-
ficiency virus; SSDI, Social Security Disability Insurance; SSI, Sup-
plemental Security Income.
aAccess to a bathroom, sufficient food, clothing, or a place to
sleep.
bAsthma, diabetes, heart disease, high blood pressure, or emphy-
sema.
cGeneral population median ¼ 50/100.
dAbnormal vaginal discharge, severe pelvic pain, burning during
urination/blood in the urine, skipped period, abnormally heavy pe-
riods, or new warts, sores, or lumps on the genitals.
518Riley et al.
Am J Epidemiol. 2011;174(5):515–522

Page 5

Overall mental health
ThepopulationmedianfortheMentalCompositeScorewas
42(ofapossible100;medianforthegeneralpopulation¼ 50).
With adjustment for all significant study confounders, unmet
subsistence needs was the most important explanatoryvariable
(i.e.,hadthe largesteffect) amongthestudy’s estimatedeffects
on the overall mental health of women in this population
(Table 2). On average, women reporting unmet subsistence
needs had Mental Composite Scores that were 5.37% lower
than those who did not (P ¼ 0.000027), after adjustment
for all other significant study confounders. In separate
models and, in order of the population effect on the Men-
tal Composite Score, ?90% antiretroviral therapy adher-
ence and having a close friend/confidant were the second
and third most important explanatory variables for overall
mental health, corresponding to a 5.07% (P ¼ 0.0006) and a
3.20% (P ¼ 0.0014) higher Mental Composite Score, re-
spectively; subsequently ranked variables were crack use
with a score of ?4.55% (P ¼ 0.0018), any drug use with a
score of ?4.15% (P ¼ 0.0018), sleeping on the street with
a score of ?2.92% (P ¼ 0.0036), experiencing cocaine
withdrawal with a score of ?3.02% (P ¼ 0.011), heavy
alcohol use with a score of ?2.78% (P ¼ 0.032), and having
no sources of instrumental support with a score of ?2.26%
(P ¼ 0.040).
Overall physical health
The population median for the Physical Composite Score
was 41 (of a possible 100; median for the general popula-
tion ¼ 50). With adjustment for all significant study con-
founders, crack use was the most important explanatory
variable among the study’s estimated effects on the overall
physical health of women in this population (Table 3). On
average, women reporting crack use had Physical Compos-
ite Scores that were 3.62% lower (P ¼ 0.0052) than those
who did not, after adjustment for all other significant study
confounders. In separate models, and in order of their ad-
justed population effects on physical health status scores,
any drug use and unmet subsistence needs also had signifi-
cant negative impact on overall physical health, with ?3.1%
(P ¼ 0.013) and ?2.9% (P ¼ 0.016) scores, respectively.
Gynecologic symptoms and conditions
At least 1 gynecologic symptom or condition was re-
ported by 43% of this population. With adjustment for all
Table 2.
HIV-positive Homeless and Unstably Housed Women Living in San Francisco, California, 2002–2008 (n ¼ 129)b
Ranked Influence of a Unit Increase in Competing Needs During the Past 3 Months on the Overall Mental Health (Score, 0–100)aof
Main Effect
Crude Population
Effect
Crude 95% CI
Adjusted Population
Effect
Adjusted 95% CI
tVIM
Rank
P Value
Unmet subsistence needs?8.23
7.16
?11.31, ?5.05
4.12, 10.07
?5.37
5.07
?7.34, ?2.35
2.12, 7.76
10.000027
?90% ART adherence
Has a close friend/confidant
2 0.0006
3.97 1.70, 6.083.20 1.05, 4.9830.0014
Crack use?7.58
?7.43
?10.06
?9.32
?4.57
?4.49
?10.47, ?4.58
?10.3, ?4.43
?14.35, ?5.76
?12.98, ?5.69
?7.43, ?1.66
?6.65, ?2.19
?4.55
?4.15
?2.92
?3.02
?2.78
?2.26
?7.2, ?1.69
?6.81, ?1.6
?4.47, ?0.49
?4.95, ?0.34
?5.44, ?0.38
?4.49, ?0.15
4 0.0018
Any drug use 50.0018
Slept on the street6 0.0036
Cocaine withdrawal70.011
Heavy alcohol use80.032
No reported sources of instrumental support90.040
Abbreviations: ART, antiretroviral therapy; CI, confidence interval; HIV, human immunodeficiency virus; SF-36, Short Form 36; tVIM, targeted
variable importance.
aSF-36 mental health composite score in which a higher score indicates better health.
bEach row represents a separate model in which a single main linear effect is estimated, adjusting for the potential confounding of all other
significant study variables.
Table 3.
HIV-positive Homeless and Unstably Housed Women Living in San Francisco, California, 2002–2008 (n ¼ 129)b
Ranked Influence of a Unit Increase in Competing Needs During the Past 3 Months on the Overall Physical Health (Score, 0–100)aof
Main Effect
Crude Population
Effect
Crude 95% CI
Adjusted Population
Effect
Adjusted 95% CI
tVIM
Rank
P Value
Crack use?3.89
?3.59
?3.02
?6.63, ?1.08
?6.15, ?0.88
?5.31, ?0.8
?3.62
?3.1
?2.91
?5.89, ?0.94
?5.5, ?0.6
?5.33, ?0.61
10.052
Any drug use20.013
Unmet subsistence needs30.016
Abbreviations: CI, confidence interval; HIV, human immunodeficiency virus; SF-36, Short Form 36; tVIM, targeted variable importance.
aSF-36 physical health composite score in which a higher score indicates better health.
bEach row represents a separate model in which a single main linear effect is estimated, adjusting for the potential confounding of all other
significant study variables.
Basic Subsistence Needs and Overall Health519
Am J Epidemiol. 2011;174(5):515–522