Heterogeneous patient preferences for modern antiretroviral therapy: results of a discrete choice experiment

Abstract

Objective:

Limited data describe patient preferences for the growing number of antiretroviral therapies (ART). We quantified preferences for key characteristics of modern ART deemed relevant to shared decision-making.

Methods:

A discrete choice experiment (DCE) survey elicited preferences for ART characteristics, including dosing (frequency and number of pills), administration characteristics (pill size and meal requirement), most-bothersome side effect (from diarrhea, sleep disturbance, headaches, dizziness/difficulty thinking, depression, or jaundice), and most-bothersome long-term effect (from increased risk of heart attacks, bone fractures, renal dysfunction, hypercholesterolemia, or hyperglycemia). Between March and August 2017, the DCE was fielded to 403 treatment-experienced persons living with HIV (PLWH), enrolled from two infectious diseases clinics in the Southern United States and a national online panel. Participants completed 16 choice tasks, each comparing 3 treatment options. Preferences were analyzed using mixed and scale-adjusted latent class (SALC) logit models.

Results:

Most participants were male (68%), older (IQR: 42–58 years), and had substantial treatment experience (IQR: 7–21 years). In mixed logit analyses, all attributes were associated with preferences. Side- and long-term effects were most important, with evidence of substantial preference heterogeneity. SALC analysis identified 5 preference classes. For classes 1 (40%), 2 (24%) and 3 (21%) side effects were most important, followed by long-term effects. For class 4 (10%) dosing was most important. Class 5 (4%) was largely indifferent to ART characteristics.

Conclusion:

Overall, treatment-experienced PLWH valued minimizing side effects and long-term toxicities over dosing and administration characteristics. Preferences varied widely, highlighting the need to elicit individual patient preferences in models of shared antiretroviral decision-making.

Précis:

On average, persons living with HIV value minimizing side effects and long-term toxicities of antiretrovirals more than other regimen characteristics, but individual preferences vary widely.

INTRODUCTION

The advent of multiclass, combination antiretroviral therapy (ART) in 1996 transformed the clinical care of persons living with HIV (PLWH). Despite their inconvenient dosing schedules, high cost, strict adherence requirements, and significant short-term and long-term toxicities, these medications effected dramatic declines in HIV-associated morbidity and mortality. Subsequent advances in modern ART have included pharmacologic boosting, new mechanistic classes, widespread use of single-tablet regimens (STRs), improvements in toxicity profiles, and increasing numbers of generic formulations, including generic STRs. By October 2018, the US Department of Health and Human Service Guidelines for the Use of Antiretroviral Agents in Adults (1) recommended six options for the initial therapy for most persons with HIV and another 22 for certain clinical situations.

The plethora of effective treatment options, each with unique, patient-relevant characteristics and trade-offs, affords greater opportunity to engage patients in shared antiretroviral decision-making. This approach, where clinicians and patients collaborate in making clinical plans, is a key component to delivering person-centered care (2–5). Patients who are more activated (i.e., view themselves as active managers of their health and health care) have been shown to have better outcomes and care experiences, and a key measure of patient activation is their involvement in decision making (6). A recent Cochrane review found that patients exposed to decision aids used for shared decision-making felt more knowledgeable of risks and benefits, better informed, and clearer about their values (7). As new HIV therapies emerge and generic formulations become available, patients and providers faced with making treatment decisions must navigate a complex array of ART characteristics, such as pill burden, food requirements, side effects, long-term toxicities, and out-of-pocket costs, each of which may influence patient satisfaction, adherence, and treatment efficacy. Prescribers, who may be unwittingly influenced by external factors (8, 9), may incorrectly gauge how a given patient values each of these characteristics and what tradeoffs the patient is willing to make.

Discrete choice experiments (DCEs), grounded in the economic theory of utility maximization, are specifically designed to provide information about individuals’ preferences for varying characteristics of multi-attribute products. DCEs are used increasingly to understand patient perspectives and to design patient-centered interventions. Although DCEs have been used in various contexts related to HIV, including testing (10–15), prevention (16–19), service delivery (20–22), and industry-sponsored studies of treatment preferences in select populations (23–26), to our knowledge DCEs have not yet been used to assess modern ART preferences in the United States. As a first step toward the development and implementation of new shared antiretroviral decision-making models, we sought to quantify patients‘ preferences for key characteristics of modern ART using a DCE. The characteristics evaluated in this study were identified using extensive qualitative work (27) and describe features of regimens considered particularly relevant to patient participation in shared antiretroviral decision-making. An innovative approach to the design of the DCE allowed for the consideration of a broad range of potentially preference-relevant side and long-term effects while ensuring relevance to each individual participant.

METHODS

Study sample

Between March and August 2017, 263 adult PLWH from Infectious Diseases clinics at Duke University and the University of South Carolina were enrolled in the study and completed interviewer-assisted, in-person surveys. Participants were recruited using flyers and invitation cards; referrals from providers, patients, and members of a community advisory board; and recruitment of patients following clinic appointments. In addition, 140 PLWH who were members of the KnowledgePanel (28), an invitation-only survey panel comprised of more than 50,000 adults nationally representative of the U.S. population, completed an online version of the survey. GfK, which maintained the KnowledgePanel until 2018, invited eligible members and referred them to the online survey.

Ethical approval

Study activities were approved by the Institutional Review Boards of Duke University Health System (Protocol #Pro00069562) and the University of South Carolina (Protocol #Pro00051940). Written and oral informed consent were obtained from participants at Duke University and the University of South Carolina, respectively. Online respondents completed the survey after reviewing a consent script.

Surveys

Antiretroviral treatment preferences were assessed using a DCE. A supplemental survey assessed sociodemographic characteristics and treatment experience. Surveys with clinic participants were fielded on iPads; online participants accessed surveys via a web browser.

Discrete choice experiment

The design, administration, and analysis of the DCE followed the guidelines for DCE applications in healthcare (29); details on the DCE development and analysis are available in Appendix 1 in Supplemental Materials. The selection of attributes was guided by previously published results of in-depth interviews and ranking exercises with HIV clinic providers and patients (27).

The objective of the DCE was to present survey respondents with hypothetical ART regimens, built on the characteristics of regimens actually available at the time of the survey. We measured preferences over the full range of potential attributes and attribute levels (values taken on by a given attribute) by presenting respondents with systematically designed ART options and observing their stated preferences. The level range of each attribute (Table 1) was determined by the characteristics of the 21 recommended and alternative ART regimens endorsed in the DHHS Panel on Antiretroviral Guidelines for Adults and Adolescents with rating strengths of strong or moderate at the time of the survey development (30). These guidelines were updated in October 2017 (31). The attributes and levels employed in our study were the following:

Table 1.

Variation in treatment characteristics across ART regimens and their operationalization as attribute levels in the discrete choice experiment

Verbal description of attribute levels	Characteristics of existing regimens		N or range
Attribute 1: Dosing	Number of pills (per dose)
	One2		8 regimens
You take 1 pill once daily	Two2		10 regimens
You take 2 pills once daily	Three		5 regimens
You take 3 pills once daily
You take 1 pill twice daily	Dosing frequency
	Once daily		19 regimens
	Twice daily		2 regimens
Attribute 2: Administration
The pills are small but you must take them with a meal of at least 400 kcal	Meal requirement		16 regimens
The pills are large (about 1 inch) but you can take them with or without a meal
The pills are small and you can take them with or without a meal	Pill size		6 to 23 mm
Attribute 3: Side effect
Individually most bothersome side effect, selected from:	Range for probability of side effect	Low	High
You have 2–3 loose bowel movements on most days of the week	Moderate diarrhea	0	0.2
You have no diarrhea
You have trouble sleeping or unusual dreams most nights of the week	Moderate sleeping problems	0	0.161
You have no problems sleeping
You have headaches on most days of the week	Moderate headaches	0	0.07
You have no headaches
You often feel dizzy or have trouble concentrating after taking your medicine	Moderate dizziness	0	0.28
You have no dizziness or trouble concentrating
You feel depressed most days of the week and some people consider suicide	Moderate depression	0	0.09
You have no depression
Your skin and the white part of your eyes turn yellow	Jaundice	0	0.65
You have no jaundice
Attribute 4: Long-term effect
Individually most bothersome long-term effect, selected from:	Probability of long-term effect	Low	High
Over 5 years ...	Over 5 years …
1 out of 100 patients will have a heart attack (lower risk)	Risk of heart attack	0	0.02
2 out of 100 patients will have a heart attack (higher risk)
1 out of 100 patients will have a fracture due to weakened bones (lower risk)	Risk of fracture due to weakend bones	0	0.03
3 out of 100 patients will have a fracture due to weakened bones (higher risk)
1 out of 100 patients will develop new or worse kidney problems (lower risk)	Risk of new or worse kidney problems	0	0.03
3 out of 100 patients will develop new or worse kidney problems (higher risk)
10 out of 100 patients will have high cholesterol (lower risk)	Risk of high cholesterol	0.1	0.25
25 out of 100 patients will have high cholesterol (higher risk)
7 out of 100 patients will develop high blood sugar (lower risk)	Risk of high blood sugar	0.1	0.15
15 out of 100 patients will develop high blood sugar (higher risk)

Notes to Table 1.

¹The DCE also included visual descriptions for all attribute levels (see e.g. Figure 1)

²Two regimens are counted twice; they required intake of 1 pill twice a day plus 1 pill once a day

Source: Author compilation

Attribute 1 – Dosing.

Dosing was implemented in the form of a 4-level “compound” attribute that covered both the dosing frequency, i.e., how often pills are taken, and the number of pills taken. The levels included once- and twice-daily intake, with the number of pills for once-daily intake ranging from one to three. These levels were based on 19 of 21 recommended and alternative regimens allowing for once-daily dosing, with either 1 (6 regimens), 2 (8 regimens), or 3 (5 regimens) pills, and two regimens requiring twice-daily intake (1 pill twice a day plus 1 pill once a day).

Attribute 2 – Administration characteristics.

Administration characteristics included pill size and meal requirements. In the DCE, these were included as a 3-level attribute that allowed for separate estimation of the effects of (a) large pills or (b) a meal requirement, relative to (c) small pills without meal requirement. Pill size was characterized as either large (approximately 1 inch) or small, while food requirements were described as needing to take the pills with a meal of at least 400 kcal, versus taking them with or without a meal. Pill sizes among existing ART regimens ranged from 6–23 mm; 14 of 21 regimens had a meal requirement.

Attribute 3 – Side effects.

Package labeling and clinical trials results describe widely varying probabilities of diverse, potentially preference-relevant, side effects. To allow for individually relevant choice scenarios, we first asked respondents which of six moderate side effects would be most bothersome to them: diarrhea, sleep problems, headaches, dizziness/difficulty thinking, depression, or jaundice. Side effects were presented as moderate (Grade 2), using the Division of AIDS Table for Grading the Severity of Adult and Pediatric Adverse Events Version 2.0 as a reference (32). Next, in the DCE, a binary side effect attribute indicated whether the respondent’s most bothersome moderate side effect was present or absent.

Attribute 4 – Long-term effects.

Long-term effects were described by a single binary attribute describing a plausible increase in the risk of a long-term complication versus no increased risk. As with side effects, respondents saw only the increased risk of complication which they identified as their greatest concern. Participants were asked to choose the scenario they would find most bothersome over a 5-year period: increased risk of heart attacks, bone fractures, kidney problems, high cholesterol, or high blood sugar. The risks for these adverse outcomes varied across conditions and reflected the plausible range of complication rates across regimens; risks were communicated both numerically and visually.

Survey administration

Participants initially ranked dosing and administration options and selected their most bothersome side and long-term effects (see Appendix 2 in Supplemental Materials). These data were used to populate a respondent-specific comprehension task with clearly dominant (preferred levels for all attributes) and dominated (worse levels for all attributes) alternatives, followed by 16 DCE choice tasks. In each choice task, participants were asked to select their most preferred option from three treatment options presented. The individualized integration of side and long-term effects into the choice tasks, and a sample choice task, are illustrated in Appendix 2 in Supplemental Materials.

Experimental design

The experimental design of a DCE refers to the combination of choice tasks that allows for the independent estimation of the influence of each treatment characteristic on preferences. Ngene software (ChoiceMetrics 2017) version 1.12b was used to select an experimental design that minimized the D-error for a mixed logit model. Two constraints were imposed in the selection of choice tasks: (1) to reduce task complexity, either the side effect attribute or the long-term effect attribute was held constant within any given task; and (2) tasks that included more than one attribute with overlap or one or more clearly dominant or dominated treatment options were excluded. The final design consisted of 96 tasks. Participants were randomized across 6 blocks with 16 tasks each. The order of choice tasks in a block was randomized across participants. The order of treatment alternatives was randomized within each choice task.

Statistical analysis

The DCE data are comprised of 16 stated choices per participant, each indicating a preference among 3 varying treatment options. To estimate mean (average) preferences, DCE data were first analyzed using a mixed logistic (logit) model with correlated, normally distributed random coefficients, using Stata version 15 (StataCorp 2018). Models were estimated for the entire study sample and separately for each cohort (two clinic cohorts and the online sample). To facilitate cross-cohort comparisons, parameter estimates were rescaled within models to range from a minimum of −5 for the least preferred attribute level to +5 for the most preferred attribute level. Attribute importance was described by each attribute’s range of parameter estimates as a share of the total range across attributes. To compare preference heterogeneity across attributes, a coefficient of variation was calculated for each attribute as the sum of the standard deviations divided by the sum of the absolute values of the mean parameter estimates across all levels of an attribute.

To model systematic variation in preferences across respondents, a scale-adjusted latent class (SALC) model was estimated in Latent Gold Choice version 5.0 (Statistical Innovations Inc. 2018). Compared to other statistical approaches, SALC models allow for the joint measurement of structural preference heterogeneity (systematic variation in preferences across respondents) and random variation in response certainty or consistency (normal response variation within an individual). Ignoring scale heterogeneity may bias parameter estimates and may affect conclusions about the amount of systematic preference heterogeneity (i.e., the number of latent classes) among participants. The Bayesian Information Criterion (BIC) was used to compare model fit for models with 1 to 6 preference classes and 0 to 3 continuous, normally distributed latent scale factors, and identify the model that yielded the best fit for the data. To evaluate whether preferences varied systematically across study cohorts, cohort was included in the SALC model as a covariate. A sensitivity analysis also included the individually selected most bothersome side and long-term effects as covariates.

Statistical power

Statistical power in DCEs varies with sample size, the number of choice tasks, the number of alternatives per task, and the number of attributes and levels, among other characteristics. Applying an empirical power-test formula (33) conservatively to the DCE design employed in this study, suggests that the combined sample (N=403) allowed us to estimate the utility difference between the most and least-preferred regimens with a standard deviation of 0.52 times the mean, which, owing to the small numbers of attributes (4) and levels (4) and the comparatively large number of choice tasks (16), is substantially better than the precision of ‘the average’ DCE study.

RESULTS

Characteristics of study participants

Table 2 details demographic and clinical characteristics of study participants, highlighting differences between the two clinics as well as between clinic and online panel participants. Compared with Duke participants, South Carolina participants were younger, more likely to be female, had lower education, and had been on ART for a shorter period of time. Online panel participants were older, more educated, more likely to be non-minority males, had been infected with HIV and on ART for a longer period of time, were more likely to have had AIDS, had better immunologic and virologic outcomes at the time of the survey, and were more likely to have broadly disclosed their HIV infection.

Table 2.

Characteristics of study participants (N=403)

		All participants		Clinic participants					Online participants
				Duke		South Carolina
		(N=403)		(N=132)		(N=131)			(N=140)
		n or mean	% or (sd)	n or mean	% or (sd)	n or mean	% or (sd)	1	n or mean	% or (sd)	2
Demographics	Age in years	49.6	12.2	48.6	11.6	44.9	11.7	*	55.1	11.0	***
Gender	Female	128	31.8	45	34.1	69	52.7	**	14	10.0	***
	Male	274	68.0	87	65.9	61	46.6		126	90.0
	Other	1	0.2	0	0.0	1	0.8		0	0.0
Education	Less than high school	39	9.7	11	8.3	21	16.3	*	7	5.0	***
	High school	129	32.2	47	35.6	62	48.1		20	14.3
	Some college	89	22.2	22	16.7	16	12.4		51	36.4
	Bachelor’s degree or higher	133	33.2	46	34.8	25	19.4		62	44.3
	Missing	11	2.7	6	4.5	5	3.9		0	0.0
Race / ethnicity	White, non-Hispanic	138	34.3	29	22.1	22	16.8		87	62.1	***
	Black, non-Hispanic	220	54.7	96	73.3	104	79.4		20	14.3
	Other, non-Hispanic	7	1.7	1	0.8	3	2.3		3	2.1
	2+ Races, non-Hispanic	4	1.0	1	0.8	0	0.0		3	2.1
	Hispanic	33	8.2	4	3.1	2	1.5		27	19.3
HIV infection	Time since diagnosis (years)	16.4	9.2	15	8.7	13.4	8.7		20	8.8	***
	Time on ARVs (years)	13.9	8.6	13	8.2	11.0	8.2	*	17	8.1	***
Ever had AIDS?	AIDS	102	25.3	26	19.7	15	11.5		61	43.6	***
	HIV only	295	73.2	102	77.3	115	87.8		78	55.7
	Not sure / refused	6	1.5	4	3.0	1	0.8		1	0.7
Recent viral load	< 200 or not detected	344	85.4	106	80.3	109	83.2		129	92.1	**
	Between 200 and 1000	24	6.0	8	6.1	9	6.9		7	5.0
	> 1,000	14	3.5	6	4.5	4	3.1		4	2.9
	Not sure / refused	21	5.2	12	9.1	9	6.9		0	0.0
Recent CD4 count	> 500	180	44.7	61	46.2	43	32.8		76	54.3	***
	Between 200 and 500	112	27.8	28	21.2	39	29.8		45	32.1
	< 200	43	10.7	12	9.1	20	15.3		11	7.9
	Not sure / refused	68	16.9	31	23.5	29	22.1		8	5.7
How many people	No one	30	7.4	10	7.6	13	9.9		7	5.0	**
know that you are	A few	192	47.6	70	53.0	70	53.4		52	37.1
HIV-positive?	Most	102	25.3	30	22.7	28	21.4		44	31.4
	All	77	19.1	22	16.7	20	15.3		35	25.0
	Not sure / refused	2	0.5	22	16.7	22	16.8		2	1.4
Preferred	Doctor only	12	3.9	2	1.5	10	7.6
decision-maker	Self only	207	66.6	108	81.8	99	75.6
in ARV decisions	Shared	92	29.6	44	33.3	48	36.6

Notes to Table 2.

¹significance of differences between the two clinic cohorts

²significance of differences between the online and combined clinic cohorts.

^{*, **, ***}indicate statistical significance at the 0.05, 0.01, and 0.001 levels, respectively, from t-tests for continuous variables and chi square / fisher’s exact tests for categorical variables.

For continuous variables the numbers in parentheses represent standard deviations (sd). Observations with missing data (up to n=12) are excluded from the respective analyses.

Heterogeneous preferences for selected ART characteristics

Participants were asked to indicate which moderate side effect and which increased risk of long-term complication would bother them most, and to rank selected dosing and administration characteristics. Side effects that participants considered to be most bothersome were diarrhea (35%), depression (20%), sleep problems (18%), dizziness (11%), jaundice (8%), and headaches (7%). Increases in risks of long-term complication that were considered most bothersome were heart attacks (43%), kidney problems (29%), high blood sugar (12%), high cholesterol (12%), and fractures (5%). Fifty-two percent viewed one pill twice daily to be more bothersome than taking a regimen of three pills once daily. Similarly, 63% considered a meal requirement with the medication to be more bothersome than taking a large pill.

Figures 1a and 1b are Marimekko charts that highlight preference heterogeneity across two dimensions. The area of each cell is proportional to the share of participants indicating these characteristics to be most bothersome. Out of 30 possible combinations of most bothersome side effects and increases in long-term risks, the combination of diarrhea and an increased risk of heart attacks was selected most frequently (14%), followed by diarrhea and an increased risk of kidney problems (11%). When contrasting diverse dosing and administration characteristics, the combination of twice-daily dosing and a meal requirement was most bothersome to 36% of participants.

Figure 1.

Heterogenous preferences for selected ART characteristics (N=403)

Heterogeneous preferences in the discrete choice experiment

Table 3 shows the results of mixed logit models of 403 participants’ choices, totaling 6,432 comparisons (16 choice tasks per participant * 403 participants – 16 skipped tasks = 6,432 choices). The relative importance of the attribute levels is described by the estimated coefficients. Across the three cohorts, the rank order of the estimated coefficients for the 4 most important attribute levels was identical, and the mean parameter estimates, rescaled to range from −5 for the least preferred attribute level to +5 for the most preferred attribute level, were similar: (1) no side effect (all +5.0); (2) low long-term risk (range +3.02 to +4.36), (3) single tablet regimen (range +2.05 to +2.35), and (4) small pill without meal requirement (range +0.89 to +1.05). Preference heterogeneity for each attribute level is described by the standard deviation around the mean estimate for the corresponding parameter. Across the four models, all standard deviations but one were significantly different from zero (p<0.001; not shown).

Table 3.

Preferences for characteristics of HIV antiretroviral therapy: results from mixed logit analyses of data from a discrete choice experiment (N=403).

			All participants			Duke participants			South Carolina participants			Online participants
Number of participants			403			132			131			140
Number of choice situations			6432			2112			2086			2234
Importance of attributes		Coefficient range	Attribute share	Coefficient of variation	Coefficient range	Attribute share	Coefficient of variation	Coefficient range	Attribute share	Coefficient of variation	Coefficient range	Attribute share	Coefficient of variation
Dosing		3.77	17%	1.60	3.78	17%	1.33	3.91	16%	1.85	3.29	15%	1.95
Administration		1.75	8%	4.12	1.64	7%	4.52	1.37	6%	5.93	2.34	11%	2.33
Side effect		10.00	44%	0.52	10.00	44%	0.59	10.00	42%	0.49	10.00	46%	0.70
Long-term risk		7.26	32%	0.69	7.28	32%	0.68	8.72	36%	0.61	6.04	28%	0.95
Importance of attribute levels		Estimated coefficient	Standard error	Standard deviation	Estimated coefficient	Standard error	Standard deviation	Estimated coefficient	Standard error	Standard deviation	Estimated coefficient	Standard error	Standard deviation
Dosing	1 pill once daily	2.22	(0.12)	2.58	2.35	(0.19)	3.14	2.28	(0.20)	3.46	2.05	(0.20)	4.00
	2 pills once daily	0.02	(0.09)	0.99	−0.22	(0.13)	0.46	0.27	(0.17)	1.26	−0.17	(0.15)	0.43
	3 pills once daily	−1.55	(0.11)	1.62	−1.43	(0.16)	1.25	−1.63	(0.19)	1.84	−1.24	(0.20)	1.94
	1 pill twice daily	−0.68	(0.11)	1.95	−0.70	(0.14)	1.42	−0.93	(0.21)	2.88	−0.63	(0.19)	1.63
Administration	Meal requirement	−0.79	(0.11)	2.09	−0.70	(0.16)	2.20	−0.48	(0.23)	3.02	−1.29	(0.20)	1.56
	Large pill	−0.17	(0.10)	1.86	−0.24	(0.14)	2.05	−0.41	(0.19)	2.26	0.24	(0.17)	1.44
	Neither	0.96	(0.08)	3.95	0.94	(0.13)	4.25	0.89	(0.16)	5.28	1.05	(0.16)	3.00
Side effect	None	5.00	(0.22)	2.59	5.00	(0.36)	2.97	5.00	(0.38)	2.46	5.00	(0.42)	3.50
	Moderate	−5.00	(0.22)	2.59	−5.00	(0.36)	2.97	−5.00	(0.38)	2.46	−5.00	(0.42)	3.50
Long-term risk	Low	3.63	(0.14)	2.50	3.64	(0.27)	2.47	4.36	(0.33)	2.68	3.02	(0.23)	2.86
	High	−3.63	(0.14)	2.50	−3.64	(0.27)	2.47	−4.36	(0.33)	2.68	−3.02	(0.23)	2.86

Notes to Table 3.

Respondents were asked to complete 16 choice tasks; each choice task was comprised of 3 treatment options. A total of 16 questions were skipped across the three cohorts.

The relative importance of attribute levels is described by coefficient estimates from cohort-specific mixed logit models (rescaled within each model to range from −5 for the least preferred to +5 for the most preferred level; see Appendix 1 in Supplementary materials).

The relative importance of attributes is described by the range of each attribute’s coefficient estimates as a share of the total range across attributes.

Preference heterogeneity is described by attribute-specific coefficients of variation, calculated as the sum of the standard deviations divided by the sum of the absolute values of the mean parameter estimates across all levels of an attribute

The relative importance of each attribute is described by each attribute’s coefficient range (from lowest to highest) expressed as a share of the total range across attributes. Side effects had the greatest influence on preferences (44%), followed by long-term effects (32%), dosing (17%), and administration characteristics (8%). Relative attribute importance was similar across the three cohorts. To characterize preference heterogeneity for attributes, the relative importance and standard deviations for each attribute’s levels were aggregated into attribute-specific coefficients of variation. Across cohorts, the greatest heterogeneity in preferences was observed for administration characteristics, followed by dosing, long-term effects, and side effects.

In latent class analysis, the Bayesian Information Criterion (BIC) applied to the SALC model indicated the best model fit when 5 preference classes and 3 continuous, normally distributed latent scale factors (one each for dosing and administration characteristic, side effects, and long-term effects) were used. The 5 preference “classes” represent statistical groupings of individuals with similar sets of preferences. Four of the five preference classes were consistently averse to undesirable ART characteristics, whereas one class (class 5, 4% of participants) expressed little aversion to any of the adverse ART characteristics included in the DCE (Table 4). Side effect was the most important attribute for classes 1 (40%), 2 (24%), and 3 (21%), followed by increased long-term risk. These classes differed from one another with respect to the relative importance of dosing and administration characteristics. In Class 4 (10%) dosing was most important, followed by side and long-term effects.

Table 4.

Preferences for characteristics of HIV antiretroviral therapy: Results from a scale-adjusted latent class model (N=403)

				Preference classes
		Class 1	Class 2	Class 3	Class 4	Class 5
Class size		40.2%	24.3%	20.9%	10.2%	4.5%
Coefficient estimates
Dosing	1 pill once daily	1.391	1.094	1.343	3.784	−0.190
	2 pills once daily	−0.058	0.882	−0.390	−0.544	−0.050
	3 pills once daily	−0.862	−0.387	−1.549	−3.210	−0.258
	1 pill twice daily	−0.472	−1.590	0.595	−0.030	0.498
Administration	Meal requirement	−0.527	−2.631	1.047	0.490	−0.267
	Large pill	0.074	1.203	−1.905	0.007	−0.017
	Neither	0.453	1.428	0.858	−0.497	0.284
Side effect	None	3.046	4.655	3.191	2.909	0.091
	Moderate	−3.046	−4.655	−3.191	−2.909	−0.091
Long-term risk	Low	2.663	3.465	2.244	1.744	0.593
	High	−2.663	−3.465	−2.244	−1.744	−0.593
Intercept		1.057	0.559	0.317	−0.304	−1.630
Study cohort	Duke	−0.012	−0.203	0.341	0.208	−0.334
	South Carolina	−0.110	0.482	0.576	−0.079	−0.870
	Online panel	0.122	−0.280	−0.917	−0.129	1.203
Coefficient ranges
Dosing		2.25	2.68	2.89	6.99	0.76
Administration		0.98	4.06	2.95	0.99	0.55
Side effect		6.09	9.31	6.38	5.82	0.18
Long-term risk		5.33	6.93	4.49	3.49	1.19
Importance of attributes
Dosing		15%	12%	17%	40%	28%
Administration		7%	18%	18%	6%	21%
Side effect		42%	41%	38%	34%	7%
Long-term risk		36%	30%	27%	20%	44%

Notes to Table 4.

Respondents were asked to complete 16 choice tasks; each choice task was comprised of 3 treatment options. A total of 16 questions were skipped across the three cohorts.

The relative importance of attribute levels is described by coefficient estimates from a scale-adjusted latent class model with 5 preference classes and 3 continuous normally distributed scale factors.

The relative importance of attributes is described by the range of each attribute’s coefficient estimates as a share of the total range across attributes.

Figure 2 describes visually the variation in participants’ aversion to diverse ART characteristics across preference classes; markers’ distances from the origin describe the class-specific degree of aversion to the respective ART characteristics. Most participants’ choices were influenced more by their respective most bothersome side effect and long-term risk (Panel A) than administration characteristics (Panel B) and dosing (Panels C and D); across the 5 classes, class 2 participants were most averse to both side and long-term effects. With respect to administration characteristics, class 2 participants were most averse to meal requirements, while class 3 participants were most averse to large pills (Panel B). With respect to dosing, class 2 participants were more averse to twice daily dosing than having to take two pills once a day; the opposite held true for Class 3 participants (Panel C). Class 4 participants, and less strongly Class 1 participants, were approximately equally averse to twice daily dosing as to having to take two pills once a day; neither had strong preferences regarding pill size or a meal requirement. Overall, the latent class analysis suggests a heterogeneous yet orderly distribution of preferences in the study population.

Figure 2.

Heterogenous preferences for characteristics of antiretroviral therapy: results from the latent class analysis of data from a discrete choice experiment (N=403)

There was evidence of systematic variation in preferences across the three study cohorts (p=0.0011), with increased odds of Class 3 membership among clinic participants, while online participants were more likely to be in Class 5 (Table 4). By contrast, neither the side effect nor the long-term effect selected by participants as most bothersome were significant predictors of class membership (p=0.18 for the six side effects; p=0.17 for the five long-term effects; results of this sensitivity analysis are not shown).

DISCUSSION

In this study of 403 PLWH, participants generally valued minimizing bothersome side effects and long-term toxicities more than taking a single tablet regimen or taking a smaller pill without a meal requirement. However, the data demonstrate substantial preference heterogeneity across and within the three cohorts, and preference heterogeneity was further highlighted in participants’ prioritization of desirable and undesirable ART characteristics (Figure 1). Our findings provide strong support for an individualized approach to the elicitation of patient preferences in models of shared antiretroviral decision-making.

To our knowledge this study is the largest DCE of patient preferences for ART and the first to comprehensively elicit preferences focused exclusively on patient-relevant attributes of modern ART. Mühlbacher et al. (26) used discrete choice methods to describe the preferences of PLWH in Germany in 2009–2010, focusing on the impact of therapy on life expectancy, probability of long term (hidden) side effects, flexibility of dosing, and indicators of quality of life. The most important characteristic was that the disease is not obvious to others. Another industry-sponsored study focused narrowly on probabilistic attributes, including the chances that ART would not work; would cause an allergic reaction, bone damage, or kidney damage; and that bone or kidney damage could be treated successfully (25). Participants, who were from a single demographic group, were willing to accept increased risks of adverse effects for lower risk of virologic failure, and they preferred short-term adverse effects with more certain outcomes to long-term effects with less certain outcomes. Notably absent among the attributes was consideration of cardiovascular risk associated with abacavir (1, 34–36). Based on the attributes of ‘third-agent’ antiretrovirals, which, in 2004, comprised key distinguishing features of combination ART, Beusterien et al. (23) used adaptive conjoint methods to survey 132 PLWH in the U.S. and 205 PLWH in Germany. Of the 13 attributes evaluated, developing drug resistance, the risk of lipodystrophy, the risk of gastrointestinal side effects, and regimen convenience were estimated to have the greatest impact on patient preferences for ART. Finally, Brégigeon-Ronot et al. (24) conducted a DCE of 101 PLWH in France in 2014 and identified strong preferences for a treatment with limited drug-drug interactions, diarrhea, and long-term health problems and that did not need to be taken on an empty stomach. Participants also preferred to avoid problems associated with treatment failure or ART that resulted in a higher viral load after the first weeks of treatment. Because each of the 21 DHHS recommended regimens considered for this study have well-established records of durable virologic efficacy many of the trade-offs evaluated in the above studies lack relevance in the modern antiretroviral era. The DCE presented in this study focused on characteristics of DHHS recommended regimens deemed to be particularly relevant to patients participating in shared antiretroviral decision-making, namely dosing characteristics, administration characteristics, side effects, and plausible increases in the risks of long-term complications (27).

When asked, “Who would you prefer to make the decisions about which HIV medication you will take?” 79% of patients from the clinic-based cohorts preferred shared responsibility between doctors and patients (Table 2). Although HIV treatment guidelines highlight the importance of incorporating patient preferences in selecting ART (1, 37), how to elicit patient preferences and how to engage PLWH in shared antiretroviral decision-making are largely unexplored in research to date. The findings presented here describe systematic variation in preferences among PLWH and provide a strong basis for advancing shared antiretroviral decision-making. In other settings, patients who participate in shared decision-making have better affective-cognitive outcomes, including improved satisfaction and less decisional conflict (38), and it is plausible that such patients would have increased knowledge of their medications, more realistic expectations of adverse effects, and possibly improved longer-term outcomes as a result. Models for shared treatment decisions, which have been developed for other chronic conditions such as diabetes, hypertension, and hyperlipidemia offer a template for how HIV providers could better understand patients’ ART preferences and values, and tailor therapy accordingly (39–41). Flexible, patient-centered tools that are adaptable to the rapidly changing field of HIV therapeutics could leverage technology readily available in modern exam rooms to present up-to-date risk and benefit information associated with various regimens and facilitate discussions of patients’ values and preferences. Greater engagement with patient preferences through shared decision-making may reduce preference uncertainty and improve patient activation and adherence.

We acknowledge several limitations of the study. First, HIV treatment options were described with only six characteristics presented in the form of four attributes. The number of attributes and levels presented could not cover all regimen characteristics that might be of concern to a given patient. To ensure relevance, our previous work used ranking exercises to reduce the number of dosing and administration characteristics from 12 to 4, the number of long-term effects from 7 to 5, and the number of side-effects from 18 to 6 (27). We also asked participants to select the side effect (out of 6) and long-term effect (out of 5) that they would find most bothersome, tailoring each respondent’s DCE for maximal individual relevance. Finally, to ensure treatment relevance, our selection of attribute levels was guided by actual ranges observed across the 21 recommended or alternate regimens available at the time of the study.

A second limitation relates to the use of DCE results in a shared decision-making context. For estimates to be used to facilitate shared decision-making, preferences for multiple side- and long-term effects would have to be considered jointly; side effects would need to be characterized in a probabilistic manner (rather than as the binary attribute); and clinically-relevant characteristics, such as drug interactions or treatment limiting co-morbidities would have to be considered along-side patient preference-relevant attributes. Such extensions should be considered in the development of a shared decision-making tool.

Third, while our study identified substantial preference heterogeneity, it was unable to discern the sources or consequences of this variation. Additional studies are needed to characterize the extent to which individual-level characteristics and prior experiences with ART regimens, including their dosing, administration, and risk characteristics, as well as patients’ perceptions of risk and the severity of side and long-term effects, correlate with patient preferences. Further, the extent to which patients’ preferences align with the characteristics of their current regimens, and conditions for and barriers to switching regimens, must be systematically explored.

Fourth, we acknowledge several methodological limitations. These include the lack of statistical priors and experimental design software that would have allowed us to identify an experimental design optimized for a latent class model; the inclusion of only one side effect and only one long-term effect in the decision model, the omission of interactions, especially between the two administration characteristics; and general limitations of DCEs, such as the potential for hypothetical bias (42).

Finally, study participants were recruited from diverse patient populations at two clinics in metropolitan areas in North and South Carolina as well as a national online panel. However, eligibility among online panel participants was based on a self-reported HIV diagnosis, and none of the study three study cohorts are statistically representative of people living with HIV in the United States. Results may not be representative of patient preferences in other areas.

The broad similarity in attribute importance across three cohorts of PLWH (Table 3), comparable heterogeneity in preferences for specific characteristics of ART regimens across cohorts (Figures 1 and 2), and a stated desire for involvement in shared antiretroviral decision-making by the majority of patients, underscore the need to develop flexible preference elicitation tools that can be implemented in routine clinical settings. As new ARTs emerge, such as sustained-release formulations, (43) these tools should be adaptable to new patient-relevant attributes, changing clinical trials evidence, and the ambiguity of risk information associated with newer treatments (44).

HIGHLIGHTS.

• Previous studies describing patient preferences for antiretroviral therapy evaluated attributes such as the impact of therapy on quality of life, life expectancy, the probability of certain short- and long-term side effects, flexibility of dosing, and the risk of developing drug resistance or virologic failure. Because of advances in antiretroviral drug development, many of the trade-offs evaluated in these studies lack relevance for shared decision-making in the modern antiretroviral era.

• Our discrete choice experiment, which used attributes and levels relevant to modern antiretrovirals and an innovative methodological approach to ensure relevance to each participant, was fielded to 403 treatment-experienced persons living with HIV, enrolled from two infectious diseases clinics in the Southern United States and a national online panel. Mixed and scale-adjusted latent class (SALC) logit models demonstrated that most patients valued minimizing side effects and long-term toxicities over dosing and administration characteristics, but also that individual preferences varied widely.

• Despite broad similarities in overall attribute importance across three study populations, persons living with HIV demonstrated marked heterogeneity in preferences for specific characteristics of modern antiretrovirals, suggesting the need for flexible preference elicitation tools that can be implemented in routine clinical settings. As new antiretrovirals emerge, such tools should be adaptable to new patient-relevant attributes, changing clinical trials evidence, and the ambiguity of risk information associated with newer treatments.