Preferences of psychotherapists for blended care in Germany: a discrete choice experiment

BC as a technology-supported therapy can be operationalized in various ways depending on the type of technology that is being integrated into the therapy process. We defined an online component as a singular online tool with a specific therapeutic function or an eMHI for a specific psychological condition. Operationalizing BC, we explicitly excluded technology that solely enables remote therapist-patient communication without offering a therapeutic functionality to make the choice alternatives more comparable. Online communication technologies and therapeutic programs or tools represent two distinct functions and therefore in our view should be studied separately. Attributes and levels for the DCE were developed using a stepwise qualitative approach. First, we conducted a literature review on BC with a focus on therapists’ perceptions of and barriers to the use of BC. The literature review was conducted in the PubMed database in July 2020 with “blended care” as main search term. We identified 226 articles on BC with five articles exploring therapist’s attitudes towards BC [8, 13‐16]. The main barriers on the side of therapists identified referred to the lack of personal communication, possible negative influence of technology on therapeutic relationship, additional workload associated with BC, uncertainty about effectiveness of technological components, concerns about legal issues/liability, and general uncertainty how to blend technology into the therapeutic process BC [8, 13‐16]. In the second step, we structured the identified factors that promote or hinder the use of blended treatment using the unified theory of acceptance and use of technology (UTAUT), formulated by Venkatesh [20]. We have chosen the original version of the UTAUT model due to its better fit for structuring our research questions We believe that the two extensions of the UTAUT2, “habit”, “experience” and “hedonic motivation”, do not (yet) play a significant role for eMHI acceptance in Germany [21]. eMHI technology is at very early stage of adoption in Germany in the moment and no sufficient time has passed to form a habit or identify fun or pleasure in eMHI usage. Venkatesh et al. defined four core determinants of users’ behavioral intention to use a technology: performance expectancy, effort expectancy, social influence, and facilitating conditions [20]. According to previous research, performance expectancy has the largest impact on the intention to use the technology [22]. Tailoring this to the context of BC, we included the attribute effectiveness of the online component used in blended treatment in our DCE design. Effort expectancy refers to the perceived difficulty of using the technology. There are different types of online-tools or eMHIs with different technological configurations as well as with a different extent of the involvement from the therapist’s side into the process, which also influences effort expectancy. As the aim of our study was not to find preferences between different types of technology that could be blended, but rather to understand therapists’ preferences for the use of technology in therapeutic process. As such we did not distinguish between different types of technology. We referred to the rather general construct as online component with a therapeutic functionality; therefore, the effort expectancy was not included as an attribute. Social influence describes the degree to which individuals perceive that individuals or organizations whose opinions and views they appreciate believe that they should use a certain technology. Anticipating peer effects, we included social influence as the recommendation attribute, relating to both recommendations by colleagues and recommendations by professional society. Last, facilitating conditions are defined as organizational or technical conditions that encourage technology use [20]. We translated this determinant as a reimbursement attribute, which includes a financial incentive for therapists that could encourage the adoption of BC [23]. Because an important adoption factor identified in previous research was the presence of a good therapeutic relationship, which depends on the concrete operationalization of the blended treatment—the number of face-to-face sessions and amount of independent work of the client with the online component—we included the attribute ratio of online and personal sessions in our design [13, 15, 16]. The levels corresponding to the attributes were chosen not only to reflect realistic scenarios but also to provide a spread enabling respondents to differentiate between levels. In the third step, we conducted semi-structured interviews with therapists with different specializations to refine our selection of attributes and levels. Our aim was to represent all four main therapeutic schools and include therapists working in both inpatient and outpatient setting. Therefore, our interview sample consisted of two CBT (inpatient care), one psychodynamic therapist (outpatient care), and two humanistically oriented psychotherapists (outpatient care). We were not able to identify a therapist with systemic background who would be available for an interview. We stopped this qualitative phase of our study after these five interviews due to diminishing further information obtained regarding our experimental set-up and difficulties identifying additional interviewees. The final experimental design included five attributes with two to three levels each (see Table 1).

We constructed the choice tasks using unlabeled, paired comparisons. We used a fractional factorial design to reduce the response burden for participants [24]. A D-efficient Bayesian design was generated using JMP software from the SAS Institute. The design was optimized for main effects and all one-way interactions based on a conditional logit model. Attributes and levels were dummy coded, and Bayesian priors were obtained from a pretest. As the number of parameters to be estimated (main effects plus interaction effects) was larger than the maximum number of choice tasks we expected to still be feasible for respondents, the final design included 32 choice tasks that were divided into two blocks to guarantee response efficiency while facilitating robust statistical analysis. Participants were randomly allocated to one of the blocks of 16 choice tasks. An opt-out option was excluded, as our research question does not aim to predict likely adoption of a concrete online-tool or an eMHI as a part of BC but rather primary aims to estimate marginal rates of substitution among attributes, compare levels and attributes of the DCE. In this case, forced choice tasks are recommended. Furthermore, the exclusion of the opt-out alternative increases in order to increase the amount of information collected and avoid decreases interpretation bias [25].

The survey itself started by informing respondents about the definition of BC and the aim of the study. In addition to collecting sociodemographic information, the first part of the survey consisted of questions on attitudes towards and previous experience with eMHIs, as well as preferences for individual components of eMHIs, such as videoconferencing or CBT-based exercises. The second part started by familiarizing respondents with the DCE elicitation format and the types of questions that would be asked, followed by an unrelated warm-up choice task. Next, all attributes and levels of the DCE were explained in a narrative fashion before respondents had to answer the 16 choice tasks. Following the DCE, participants were asked about their general views of the advantages and disadvantages regarding BC and asked to evaluate the difficulty of DCE and the survey in general.

We administered the survey online through a market research agency specializing in health care providers (DocCheck Community GmBH) to individuals who, at that time, were working as psychotherapists in Germany. Data collection occurred in August 2020. Based on a rule-of-thumb calculation proposed by Johnson and Orme [26], the minimum required sample size for identifying the main effects in the choice experiment was 47 (or 140 for all two-way interactions). A final sample of 200 respondents was targeted to also provide sufficient power to investigate preference heterogeneity and allow for sufficient variation and therefore generalizability in terms of the experiences and expectations with and towards blended care.. We collected explicit and informed consent from respondents after providing them with a detailed explanation of how their personal data would be used. Each participating therapist received € 20 as monetary compensation from the market research agency.

Data collection was paused after 30 respondents. A pretest of the DCE was conducted among 30 therapists from the same respondent pool as in the main data collection. Data from the pretest were used to refine the questionnaire part of the survey, i.e., reducing the response burden [27], and to obtain the priors for the Bayesian D-efficient design (usually calculated using 10% of the sample). Data from this pretest was also used in the main analysis.

Discrete choice data are commonly analyzed based on the random-utility framework [28] by applying different statistical models, which must be selected to fit the purpose of the study [29, 30]. In our case, attempting to elicit preferences for BC interventions among psychotherapists, who constitute a heterogeneous group concerning treatment styles or therapeutic focus, we expected a large variety of preferences. Therefore, to specifically model preference heterogeneity while also relaxing the independence of irrelevant alternative assumptions, a mixed logit model was estimated [31]. To select the utility function, the following steps were taken: First, we tested whether the inclusion of an alternative specific constant (ASC) and the inclusion of block dummy variables, indicating the survey version, would be necessary to obtain unbiased estimates based on a main effects mixed logit model [29]. Second, different functional forms of the effectiveness and face-to-face online ratio attributes were specified, namely, linear and logarithmic instead of categorical. Third, we tested several two-way interactions between attributes suspected to be correlated. Categorical variables were dummy coded throughout the analysis. In the final mixed logit model, 500 Halton draws were specified, all parameters for which we found heterogeneity were set to be random and normally distributed, and individual-level clustered standard errors were used. Marginal effects were calculated using the mixlpred command.

To investigate whether preferences specifically differed for various subgroups, interaction terms of the respective subgroup indicators (e.g., therapeutic style, clinic or outpatient, potential user of BC, previous user of BC, age, gender) and the main effects parameters were included in separate models. The significance of the difference in parameter estimates between subgroups was tested using χ² tests for joint significance. We conducted two proposed tests to assess the internal validity of the choice experiment data [32]. First, using the respdiff command, we investigated the extent to which straight lining occurred, i.e., respondents always choosing either option A or option B, indicating a lack of serious engagement with the experiment. Second, attribute dominance, relating to noncompensatory preferences, was examined by calculating lexicography scores and counting the proportion of choices based on one attribute. We assumed attribute dominance if the lexicographic score was above 90%, as suggested previously [33]. All calculations were conducted using Stata 16 (StataCorp, College Station, TX).

The survey reached 1335 psychotherapists via e-mail who were members of the DocCheck Panel. A total of 238 respondents started the survey, 38 did not complete the questionnaire, and only three dropouts occurred after starting the DCE part of the survey. Table 2 shows the respondent characteristics as well as their experiences and expectations regarding BC formats. The age and gender distribution in the sample was similar to what has been reported for the overall population of psychotherapists in Germany (49%), while our sample likely was slightly younger [34]. We observed a realistic spread of therapeutic orientations across the main therapeutic approaches, although systemic and humanistic approaches were likely underrepresented² [35]. Almost all respondents were medical psychotherapists practicing in inpatient facilities, while one-third worked predominantly in outpatient care.

Most respondents who had used BC before (26.5%) evaluated their experiences as positive (Table 2). Psychotherapists seemed willing to use BC in the future, mainly by integrating BC into regular therapy. The main reported reasons for not opting for BC in the future were the “too impersonal” character of treatment (55.6%), doubts regarding their effectiveness (11.1%), lack of compatibility with the performed therapy form (11.1%), and a lack of interest or need (5.6%). Time savings and patient empowerment were mentioned most often as potential advantages of BC, whereas lack of personal support and deterioration in the therapeutic alliance were seen as main risks. However, most respondents rated the likely impact of BC on the therapeutic alliance as positive (47.5%) or neutral (45.5%). The BC features most likely to be used in the future (Fig. 1) were psychoeducation, online exercises, online diaries, and secure video communication, while chatbots for communication were least likely to be used.

The small drop-out rate during the DCE (3 out of 203) and the type of qualitative comments collected on the overall survey indicated that the respondents were able to understand the choice tasks. Tests for the internal validity of the experimental data showed that straight lining occurred in four instances (2% of the sample), and non-compensatory behavior, with a lexicographic score of over 90%, was observed in 16% of the sample. These values lie close to the respective medians that were reported for these tests in a study summarizing 55 choice experiments [32]. In addition, 30 of these 32 observations exhibit dominant choice behavior for the ratio of online and personal sessions attribute. It is conceivable that for some psychotherapists, this attribute is indeed dominant and that non-compensatory behavior thus does not (always) imply decision heuristics.

The estimated utility function included an ASC and block-specific dummy variables to account for potential bias due to the positioning of the alternatives and the allocated survey version. Linear specifications of the effectiveness and face-to-face vs. online attributes were selected since the assumption of linear preferences regarding these two variables could not be rejected in models estimated only with categorical variables (χ²=0.02, p = 0.89 and χ²=1.46, p = 0.23, respectively). We tested the inclusion of two-way interactions between the additional reimbursement attribute and both the face-to-face vs. online attribute and the effectiveness attribute. Both were nonsignificant and therefore excluded in the final model. Therefore, for ease of interpretation, we refrained from including the interaction in the final model.

Table 3 contains the results of the mixed logit model. All main effects coefficients were significantly different from zero (columns 2 and 3), indicating their importance in the choice context. The signs of the attribute levels pointed in the expected direction, and the order of the recommendation attribute levels was logical (i.e., higher preference for recommendation by society than by colleagues), providing some confidence in the theoretical validity of the results. The recommendation by professional societies to use a certain BC intervention was the most important attribute level, while additional reimbursement played a minor role. Greater effectiveness and a higher face-to-face ratio of the BC intervention increased the likelihood of selecting this intervention. Preference heterogeneity was found for all included attributes, as indicated by the significant standard deviations of the parameters (columns 4 and 5). The marginal effects in column 6 demonstrate the change in likelihood of choosing a certain BC intervention if the attribute level changes compared to the reference category (or a unit change for linear variables), conditioning on all other attributes remaining constant. This also allows for a straightforward interpretation of the relative magnitude of the coefficients. Compared to no recommendation, a recommendation by a professional society increases the conditional likelihood of selecting a BC intervention by 25.7%. Using the effectiveness levels from the DCE, a change in effectiveness from 7 out of 10 to 9 out of 10 increased the choice likelihood by a slightly lower factor (21.0%). The smallest marginal effect was found for the reimbursement attribute (9.5%).

In terms of the subgroups, we found significantly different coefficient estimates for psychotherapists working in the inpatient vs. outpatient setting (χ²=15.70, p < 0.01), for respondents being younger or older than 50 years (χ²=17.07, p < 0.01), and for therapists predominantly practicing behavioral therapy vs. psychodynamic psychotherapy (χ²=16.62, p < 0.01). Figure 2 presents the coefficient estimates for the respective subgroups. Younger therapists put more weight on recommendations to use a BC format, the potential effectiveness of the intervention and additional reimbursement, while a higher share of face-to-face vs. online time was less important to them. Inpatient therapists are more influenced by recommendations than are outpatient therapists, who preferred a higher share of face-to-face vs. online time. Face-to-face time was significantly less important for therapists predominantly conducting behavioral therapy than for therapists conducting psychodynamic psychotherapy.

In a deviation from the main model specification we tested the inclusion of an interaction between the linear effectiveness and face-to-face vs. online attributes (see Additional file 1 Appendix Table A1). The interaction between face-to-face time and effectiveness was positive, meaning that the utility of an option with higher effectiveness and higher face-to-face time did not just increase by the additive effect of effectiveness and face-to-face time, but that the combined effect would be even larger. In other words, therapists value a higher effectiveness even more if it coincides with more f2f-time and vice versa. Finding that the coefficients of face-to-face time were not significant anymore after the inclusion of the interaction is hinting towards that face-to-face time isolated may not have much value on its own if effectiveness is low. Preference estimates for the remaining attributes were largely unaffected by the inclusion of the interaction term.

A further limitation relates to the selected characteristics in the choice experiment. While these were carefully selected using standard practices, including interviews with providers, the experimental design could have omitted important characteristics, in part because no choice experiments have been conducted in this novel but emerging research area. That BC treatment formats are also new to many psychotherapists, who often have little to no prior experience, may also be seen as a limitation. While the DCE allows us to present hypothetical scenarios, the structure, mechanism, and potential benefits/risks of BC thus remained quite abstract to respondents.

Finally, our study was conducted during the COVID-19 pandemic, where measures such as contact restrictions increased the necessity to use digital technology in mental health care. In light of this, the attitudes towards BC among therapists might have been more positive compared to pre-pandemic levels. Furthermore, it is unclear to what extent the obtained preferences will remain stable after the pandemic.