Towards Personalizing Treatment for Depression

Abstract

Background

While ‘personalized medicine’ commonly refers to genetic markers or profiles associated with pharmacological treatment response, tailoring treatments to patient preferences and values is equally important.

Objective

To describe and demonstrate a method to develop ‘values markers,’ or profiles based on the relative importance of attributes of depression treatment.

Study Design

Discrete choice analysis was used to assess individuals’ relative preferences for features of depression treatment. Preference profiles were developed using latent profile analysis.

Patients or Other Participants

Eighty-six adults participating in an internet-based discrete choice questionnaire.

Main Outcome Measure

Participants were presented with two depression scenarios representing mild and severe depression. For each scenario, they were asked to compare 18 choice sets based on the type of medication side effect (nausea, dizziness, and sexual dysfunction) and severity (mild, moderate, and severe); and for counseling frequency (once per week or every other week) and provider setting (the office of a mental health professional, primary care doctor, or spiritual counselor).

Results

Three profiles were identified: profile 1 was associated with a preference for counseling and an avoidance of medication side effects; profile 2 with an avoidance of strong medication side effects and for receiving counseling in medical settings; and profile 3 with a preference for medication over counseling. When presented with a severe depression scenario, there was a higher prevalence for profile 1 and patients were more likely to prefer mental health over primary care and spiritual settings.

Conclusions

Values markers may provide a foundation for personalized medicine, and reflect current initiatives emphasizing patient-centered care. Next steps should assess whether values markers are predictive of treatment initiation and adherence.

1 Introduction

The notion of ‘personalized medicine’ has been gaining increased attention in the area of mental health [1, 2]. Typically this concept refers to the field of pharmacogenomics, which can be deployed clinically to stratify patients into treatment responders and treatment non-responders based on genetic profiling [3]. However, another vision of personalized medicine is related to tailoring to the values of the patient [4]. Attention to preferences for care can have a favorable impact on treatment initiation and adherence [5] and subsequent clinical and cost-effectiveness outcomes [6, 7]. Analogous to genetic markers, profiles of genetic variation related to treatment response [8], this research seeks to identify values markers, profiles of values related to attributes of treatment that may predict treatment uptake and adherence. Although antidepressants and psychotherapy have been shown to be effective in treating major depression, many patients do not initiate or adhere to treatment [9]. In the US nationwide DECISION study, the prevalence of decisions about depression treatment was on a par with decisions for elective surgery, much lower than other equally prevalent chronic medical conditions [10]. Untreated or under-treated depression results in increased hospitalizations, healthcare costs [11], and mortality [12] and is a major public health concern [13]. Our broader line of research seeks to find specific ways to incorporate management strategies for depression tailored to what patients most value about treatment. Understanding the relative importance individuals ascribe to different features of depression treatment and whether such importance is predictive of treatment behavior can help determine how to tailor depression treatments.

Prior research assessing what is valued in depression treatment has focused primarily on patient preferences for counseling or medication treatments [14–16] and suggests that the majority of people prefer counseling over medication [17–19]. However, few patients receive their preference for counseling. In fact, older primary care patients preferring counseling are less likely to receive depression treatment altogether [20]. While a collaborative care model improves access to counseling (74 % in collaborative care vs. 33 % in usual care), many patients still don’t adhere to counseling despite a stated preference for counseling [17]. A potential explanation may relate to unmeasured values regarding particular attributes of treatment. For example, among ethnic minority patients spiritually based treatment may be particularly desirable when it comes to depression treatment [21]. If counseling is thought to be more commensurate with a spiritual model of depression than medication, patients may express a preference for counseling but be less likely to adhere to it if the therapist doesn’t share a spiritual framework for depression. Uncovering what is valued about counseling and other treatment modalities may help us understand the treatment decisions patients make.

While prior studies have reported the prevalence of patient preferences for depression treatment [18, 22], systematic approaches have not been applied for creating preference profiles, or ‘values markers.’ Values markers have the potential to tell us what constellation of valued attributes is most important to patients, thereby suggesting what might be lacking in conventional treatments. Conjoint methods, first developed in mathematical psychology [23], are intended to ‘uncover’ the underlying preference function of a product in terms of its attributes. In the arena of mental health, Dwight-Johnson and colleagues used conjoint analysis to assess the features of treatment that low-income Latinos thought would improve its acceptability [14] and Flach and Diener applied conjoint analysis in the design of an alcohol and cigarette cessation program [24].

Our study differs from prior work by determining profiles based on individual-level preference values for specific attributes of depression treatment. Our approach is similar to an approach market researchers use to understand the heterogeneity in consumer preferences in order to target their marketing strategies to particular consumer preference profiles, a notion known as ‘market segmentation.’ Profile analysis has the potential to identify patterns of treatment attributes (values markers) that patients value most—plausible targets for tailoring interventions to improve initiation, adherence, and outcomes.

This investigation had two primary goals building on our previous report [25]. First, we introduce a method to focus tailoring strategies by describing how we create values markers based on individual-level data. Second, we present how values markers derived in response to a scenario of mild depression change for severe depression. Our work moves beyond the realm of simply suggesting that we give patients the treatment they want; instead, our approach may help determine how existing treatments might be improved or tailored to better meet patients’ needs.

2 Methods

2.1 Recruitment

A panel comprising participants who have participated in studies of judgment and decision making on the World-Wide Web (http://www.psych.upenn.edu/~baron/q.htm) and responded to a request to complete a questionnaire about depression treatment preferences. Approximately 1,500 panel members have voluntarily participated over the past 10 years in numerous online studies related to decision making [25] and the panel is roughly representative of the adult US population in terms of income and education but women are over-represented [26]. For this study, 500 randomly selected members of the panel were sent an e-mail request describing the study and a URL link to access the study. Persons who responded within the first 2 weeks after the request was sent were participants for this study (n = 86). No identifying information was obtained nor was the participants’ experience with or knowledge of depression and its treatment. In this sample, 69 % were women and the mean age was 41 years. Participants received US$3 as a token of appreciation for completing the questionnaire. This study was approved by the Institutional Review Board of the University of Pennsylvania.

2.2 Conjoint Analysis Methods

We used a type of conjoint analysis known as discrete choice analysis. Discrete choice studies involve several steps. The first step involves identifying the attributes of treatment that are most salient to patients. Next, the discrete choice questionnaire is constructed based on the attributes and levels identified. Then, employing the choice data, relative preference weights for each attribute are calculated that indicate the contribution of each attribute to the choice. The resulting model can be used to estimate the change in choice expected as levels of the attributes are changed. We describe each step in the following sections.

2.2.1 Identification of Attributes and Levels

We convened three focus groups to elicit the salient attributes of treatments: two with adults from primary care settings (n = 8 and n = 6) and one with professionals who manage depression (four mental health specialists and three primary care doctors). The primary author conducted all of the focus groups. Patient participants with a history of depression and experience with depression treatment were recruited from a local primary care office. They were asked to describe what they felt were the positive and negative aspects of depression treatments in terms of making decisions to engage in a treatment. The professional group consisted of clinicians who volunteered to be part of a focus group about depression treatment. The clinicians were all part of a research center focused on depression at the University of Pennsylvania. Clinicians were asked to think about features of depression treatments that act as barriers or facilitators to patient engagement in treatment. Focus groups were audio-recorded and transcribed. Transcripts were reviewed by two of the authors, MW and JG. From the transcripts, we selected the most frequently mentioned depression treatments and attributes in order to construct the discrete choice task. See Table 1 for the attributes used for the tasks.

Table 1.

Definitions of attributes and levels provided to participants

Attribute	Level(s)	Definition
Treatment type	Medication	You take a pill every day for at least 6 months
	Counseling	You schedule appointments to talk with a professional about your life, emotions, and   depression and to learn new ways to cope and solve problems
Risk of side effects   (medication only)	Nausea	You may have an upset stomach and feel the urge to vomit
	Dizziness	You may feel unsteady on your feet when you stand up
	Sexual dysfunction	You may experience reduced sexual interest or drive. Men may experience difficulty   with achieving or maintaining an erection
Severity of side effects   (medication only)	Mild	You can easily cope with the side effect
	Moderate	You find it difficult to cope with the side effect and you may need additional   medication to treat it. The side effect interferes with some of your day-to-day   functioning
	Severe	You find it very difficult to cope with the side effect and you need additional   medication to treat it
Number of counseling   sessions (counseling only)	Every week	You attend counseling for 1 hour every week
	Every 2 weeks	You attend counseling for 1 hour every 2 weeks
Location of counseling   (counseling only)	Primary care doctor’s   office	You go to your primary care doctor’s office for counseling
	The office of a mental   health professional	You go to the office of a mental health professional for your counseling
	The office of a spiritual   counselor	You go to the office of a spiritual advisor (priest, pastor, rabbi, imam, etc.) for your   counseling

Reproduced from Wittink et al. [43], with permission from Sawtooth Software, Inc

2.2.2 Discrete Choice Questionnaire

We constructed a questionnaire that presented respondents with a series of choice sets in which attributes of medications would have to be played off against attributes of counseling in selecting preferred treatment. Discrete choice conjoint analysis simulates the selection of services or products in competitive contexts by presenting respondents with a set of products (composed of one level from each attribute) and asking which package they prefer. Because medication and counseling share few attributes, we employed an alternative-specific discrete choice design [27] that allowed us to include attributes relevant and specific to medication and counseling (separately) that might be most influential in patient decision making (see Fig. 1 for an example of a choice set that respondents were shown in the study). In addition to the questionnaire, participants were given definitions for each attribute and level (see Table 1).

Fig. 1.

Example of a choice set that respondents were shown in the study

We created our discrete choice questionnaire using SAS 9.1 conjoint analysis software. Each participant was asked to express their preference for medication or counseling based on two different scenarios: ‘mild depression’ and ‘severe depression.’ Consistent with literature on treatment effectiveness [28], participants were told that counseling and medication were equally effective for the treatment of depression. For each scenario, patients were asked to complete a series of 18 choice sets. The text preceding each choice set in the mild depression scenario was as follows: “You complain of feeling more tired than usual and you aren’t interested in doing things that you normally enjoy. Your [physician] diagnoses mild depression and recommends treatment to help. She gives you two choices for treatment. Select which option … most closely resembles the treatment you would choose.” For the severe depression scenario, the text was the same except that the depression was described as “severe depression.”

Each participant completed the two surveys that contained differing attributes of medications and counseling. For medication, nine combinations were possible; namely, three levels of severity (mild, moderate, and severe) and three side effects (nausea, dizziness, and sexual dysfunction). For counseling, six combinations were possible; namely, two schedules for frequency of counseling sessions (once per week or every other week) and three locations for the sessions (a mental health professional’s office, primary care doctor’s office, or office of a spiritual counselor). Each choice set required that the respondent choose medication or counseling. With 27 possible combinations for medicines and six for counseling, there are 54 possible choice sets. While such a design would allow us to assess interactions between attributes, we decided such a design would place an undue burden on respondents. We therefore used the experimental design routines in version 9.1 of SAS to select a fractional factorial design of 18 choice sets from the 54 possible sets that allowed us to estimate the main effects of each factor [27].

2.3 Analyses

The analysis proceeded in three stages. First, individual-level relative preference weights were estimated. Our logistic model for calculating individual-level relative preference weights produces a fixed, or overall, effect for each variable, and a random deviation from the overall effect for each respondent. The deviation from the overall effect for each individual is called the random effect because the individual effects are assumed to derive from a random distribution. An emerging standard for analyzing discrete choice data is mixed or random-parameters logit [29]. We use a Bayesian-like approach, called empirical Bayes estimation, because of the advantage of allowing for the estimation of individual-level relative preference weights functions [30]. The empirical Bayes method has the advantage of borrowing information from the entire sample in estimating the random effect for each person. We obtained an empirical Bayes estimate for the individual random effect and the corresponding standard errors. We used the SAS GLIMMIX macro, which operates by calling PROC MIXED iteratively to estimate the parameter effects. For all analyses, we used an α of 0.05 to assess statistical significance, recognizing that statistical tests are guides to interpretation and inference.

2.3.1 Latent Profile Analysis

Secondly, we clustered individual preference weights using latent profile analysis (LPA) [31]. LPA is concerned with deriving information about a categorical latent variable from the observed values of continuous manifest variables. In other words, LPA deals with fitting latent profile models to the measured data. The model groups response profiles into clusters representing preferences for treatment. The response profiles are assumed to be derived from a mixture of class-specific Normal distributions, where each class has a unique mean preference weight and corresponding variance. The number of clusters is determined using model fit criterion (e.g., Bayesian Information Criterion; BIC) and clinical relevance.

Finally, to examine the association between the mild and severe depression scenarios, we assigned each individual to the profile for which the model indicated the highest probability.

3 Results

3.1 Values Markers for Mild and Severe Depression Scenarios

The relative preference weights of depression treatment attributes were used to determine common preference profiles (values markers). The profiles representing values markers in the mild depression scenario are shown in the first three columns of Table 2. The last three columns depict profiles for the severe depression scenario. We used the Lo-Mendell-Rubin adjusted likelihood ratio test (LRT) to assess model fit. While the two-profile (null hypothesis) model yielded the best fit in comparison with the three-profile model (mild: LRT = 66.81, p = 0.09; severe: LRT = 43.26, p = 0.25), the three-profile model suggested that there may be a small but clinically important group that the two-profile model did not capture. Specifically, the two-profile model yielded two groups, one that preferred medication and one that preferred counseling, while the three-profile model revealed a group that was relatively more swayed by other attributes than treatment type. We therefore chose the three-profile model because of the implications that such a group might have for clinical interpretation described in Sects. 3.2 and 3.3.

Table 2.

Profile analysis of individual-level relative preference weights from conjoint analysis given mild and severe depression scenarios

	Mild depression			Severe depression
	Profile 1	Profile 2	Profile 3	Profile 1	Profile 2	Profile 3
Type of treatment
Medication vs. counseling	−5.68*	−0.15	5.86*	−4.79	0.81	6.20
Side effect of medication
Nausea vs. sexual dysfunction	0.17	0.20	−0.27	0.96	1.83	−1.91
Dizziness vs. sexual dysfunction	−0.01	−0.65	0.31	0.55	3.01	−1.76
Severity of medication side effect
Mild vs. severe	2.23*	6.82*	−0.93	−1.82	−2.15	3.18*
Moderate vs. severe	0.09	−4.52*	2.03*	−0.71	−0.71	1.20
Frequency of counseling
Weekly vs. every 2 weeks	−0.60	0.64	0.31	0.35	0.30	−0.58
Location of counseling
Clergy vs. mental health	0.23	0.95	−0.68	1.76*	−10.35*	1.12
Primary care vs. mental health	0.06	0.79	−0.43	−0.23*	2.34*	−0.48
Profile prevalence (%)	41	18	41	50	13	37

Reproduced from Wittink et al. [43], with permission from Sawtooth Software, Inc.

Favoring the first attribute mentioned in the row is associated with a positive mean relative preference weight for persons with that profile. A negative mean indicates that the second attribute mentioned was favored

Asterisk indicates the estimate is statistically different from zero (p < 0.05)

3.2 Profile Analysis Under the Mild Depression Scenario

Participants in profile 1 demonstrated a strong relative preference weight for the counseling attribute and for avoidance of the medication side-effects attribute. Profile 2 was associated with the avoidance of severe side-effects attribute. Profile 3 showed a strong relative preference for medication over counseling, and no strong preferences were shown for the location of counseling.

3.3 Profile Analysis Under the Severe Depression Scenario

Profiles showed some change in patterns when the depression was severe. Specifically, persons in profile 1 (strong relative preference weight for counseling) and profile 2 (no strong relative preference for counseling vs. medication) preferred mental health settings over clergy and primary care when the depression was severe. Prevalence of profile 1 (more preference for counseling) increased as well.

3.4 Changes in Values Markers for Mild Versus Severe Depression Scenarios

We found that the average probabilities for patients to be classified in a particular profile changed when the scenario changed from mild to severe depression. Table 3 provides the cross-classification of the profiles for the mild depression scenario (rows) according to the classification for the severe depression scenario (columns). Among the 35 persons whose profile indicated a preference for medicine (profile 3) in the mild depression scenario, 11 % were classified as preferring counseling (profile 1) when mild was changed to severe depression in the scenario, 20 % were in profile 2 (no preference for medication or counseling), and 69 % continued to prefer medicines (profile 3). Among the 35 persons whose profile indicated a preference for counseling in the mild depression scenario, 6 % were classified as preferring medicine when mild was changed to severe depression in the scenario, 11 % were in the no preference group, and 83 % continued to prefer counseling. For the 16 people who were classified as being in the no preference (i.e., ‘intermediate’) group in the mild depression scenario, 10 (63 %) preferred counseling and 6 (37 %) preferred medicine when mild was changed to severe depression.

Table 3.

Cross-classification of values markers for mild versus severe depression scenarios

	Severe depression scenario
	Counseling	Intermediate	Medicine	Totals
Mild depression scenario
Counseling	29 (83)	4 (11)	2 (6)	35
Intermediate	10 (63)	0	6 (37)	16
Medicine	4 (11)	7 (20)	24 (69)	35
Totals	43	11	32	86

Reproduced from Wittink et al. [43], with permission from Sawtooth Software, Inc.

Numbers in parentheses are row percentages

4 Discussion

We derived values markers, profiles of preferred attributes of treatment, based on how participants weighed various attributes of depression treatment, identifying two profiles representing a strong relative preference for counseling or medicine, and an intermediate profile associated with the location of treatment for mild depression and avoidance of side effects for severe depression. Among persons classified in the counseling profile, 83 % preferred counseling for both mild and severe depression. Among persons who preferred medicine, 69 % preferred medicine for both mild and severe depression. The intermediate profile identifies a group whose preferences are changeable and not clearly focused on counseling or medicine. The latter profile may identify a group for whom tailoring an intervention for depression may be most salient.

While accommodating patient values into healthcare constitutes a key component of the overarching principles of healthcare redesign, precisely how to assess preferences and how to utilize the information in a systematic manner that might facilitate patient-centered treatments remains an area with a need for methodological development. It is important to consider underlying reasons for preferences and to determine whether preferences are related to stable and deeply held beliefs as opposed to relative access to healthcare information, which may reinforce existing healthcare disparities [31].

Before putting our study into the context of current thinking on preferences for depression treatment, several study limitations require comment. First, the participants were recruited from an online sample of people who are part of a decision-making survey panel and therefore may react to the hypothetical scenarios very differently than would actual patients. It is also important to note that while we sent out an e-mail invitation to 500 participants, only 87 responded to the survey within the first 2 weeks. The low response rate may be due to the fact that the study focused on depression. There are many reasons people may not wish to participate in studies about depression, including the stigma associated with mental illness [33], and this may affect even confidential and online studies [34]. While it is possible that those who did participate were more likely to be familiar with depression, we do not know whether they actually had experience with depression treatment or carried a diagnosis of depression. People actually confronting these types of treatment decisions (i.e., patients suffering from depression) might respond very differently to the hypothetical scenarios. With respect to conjoint design, the number and definition of attributes and levels is of critical importance; our decisions were based on patient and expert opinions of the most important attributes and the appropriate levels to include. We note a potential limitation in only including side-effect severity and type of side effect for medicine. Had we included attributes related to side effects for counseling, we might have seen different responses. In addition, other attributes not included, such as cost, might have been highly influential. Furthermore, for the attribute of counseling frequency, we chose to include only two levels (once a week or once every other week) based on the commonly available options for counseling session frequency. We acknowledge that the differential number of levels (two for counseling frequency and three for all other attributes) may have had an effect on the importance ratings in the conjoint task [35]. However, even with these limitations, our work presents an important proof of concept for a novel method of describing patients based on their preferences for treatment attributes. While others have used individual level-preference weights and cluster analysis to determine preference patterns for treatments [36], the use of LPA provides added benefit if one is interested in determining which patients might benefit from a tailored approach to care. Specifically, LPA is driven by a statistical hypothesis that there is an underlying commonality between the treatment attribute preferences rather than simply measuring the relatedness of the treatment attribute preferences [37].

The method we used to calculate individual-level relative preference weights allowed us to learn about groups of persons within the sample who have strong preferences for specific attributes. Previous research on depression treatment preferences has focused primarily on ascertaining which individuals have strong preferences for a particular treatment type [38]. Such work can direct treatment along lines preferred by the patient. However, patients who are in a profile with no strong preference for treatment type (counseling or medication) but who are swayed by other attributes of treatment, such as side effects, might benefit from a tailored approach to discussing treatment options or to a treatment tailored to particular treatment aspects. Despite increasing access to a variety of effective treatments in primary care, engaging patients in depression management strategies remains a major hurdle. Having more knowledge about particular patient profiles for whom various attributes of treatment are most consequential could lead to more patient-centered interventions, improved initiation of treatment, and improved patient-centered outcomes.

In addition, our results shed light on the idea that certain treatment preferences may change over time. We found that presenting patients with a severe versus mild depression severity scenario had implications for both the type of treatment preferred and the relative importance of the location of treatment services. These findings suggest that preferences may need to be repeatedly assessed as health states change.

5 Conclusions

Conjoint methods sharpen the focus on ‘what it is about treatment’ that drives preferences and provides specific guideposts for how to design packages of treatments that are patient centered. Studying how preferences for attributes of treatment are related to treatment adherence, how preferences change over time as depression severity changes, and how preferences change with treatment experience are important next steps. In addition, it will be important to look at which types of attribute preferences may represent deeply held beliefs that may be based on cultural norms and traditions, and which types of attribute preferences are more transient and based on limited access to healthcare information [32]. Conjoint analysis has been successfully applied to service redesign to match with changing consumer needs [39, 40] and is increasingly being considered in medical service redesign [41, 42]. For example, conjoint analysis could be used to link patient preferences for specific attributes of both conventional treatments (e.g. medication and/or counseling) and non-conventional depression treatments (such as meditation or spiritual therapy) to observed behavior (initiation and adherence to prescribed treatment). If patients with preferences for specific levels of non-conventional treatment attributes are more likely to be non-adherent to prescribed treatments, then conventional treatments might be adapted to incorporate the desired attributes of non-conventional treatments (e.g. counseling that incorporates aspects of spirituality). Participants who fit into profile 2 identified in this study might be targeted for just such tailoring.

Key Points for Decision Makers.

• We demonstrate how discrete choice analysis can be combined with latent profile analysis to classify patients into groups according to what they value most about treatment

• Ascertaining which individuals have strong preferences for particular combinations of treatment characteristics (e.g. location of services, side-effect profile), in addition to preferences for treatment types (e.g. medication vs. counseling), can help create innovative and tailored treatment approaches. This may be of particular value to depression management in primary care given the low initiation of treatment