Bipolar disorder (BD) is a complex mental health condition with multiple and varying states ranging from elevated mood (mania or hypomania) to feelings of hopelessness and sadness (depression) [1]. It consists of several related diagnoses representing a spectrum of illness, including bipolar type I, bipolar type II, cyclothymia, and bipolar not elsewhere classified. The global prevalence of bipolar spectrum disorders is estimated at 0.741% of the adult population, and BD is associated with significant disability and costs to both health care systems and society [2-4].

The primary therapy for BD consists of mood stabilizing medications, including lithium, antipsychotics, and anticonvulsants [5-13]. Psychosocial therapies, including psychoeducation and cognitive behavioral therapy (CBT), are recommended as add-on therapy to medications to reduce relapse through improved medication adherence, identification of early warning signs, self-management, and family communication [14,15]. Psychosocial therapies delivered through traditional face-to-face methods have been shown to be effective and cost-effective adjunctive treatments to pharmacotherapy for people with BD [16-18] and other mental health diagnoses such as anxiety and depression [19].

Internet-delivered psychosocial therapies can overcome several barriers faced when seeking mental health care, such as geographic location, a limited number of service providers, and the cost of treatment. Internet-delivered psychosocial therapies have been shown to be effective and cost-effective for the treatment of depression and anxiety [20,21]. However, the evidence to support the effectiveness and cost-effectiveness of internet-based psychosocial therapies for people with BDs is limited [18,22]. To fill this gap, the MoodSwings 2.0 randomized controlled trial (RCT) was conducted to investigate the efficacy of an internet-based self-guided psychosocial intervention for people with BD [23]. This analysis reports on the within-trial economic evaluation of MoodSwings 2.0 from an Australian health sector perspective.

This economic evaluation was conducted alongside the RCT of MoodSwings 2.0 (ClinicalTrials.gov NCT02118623 [Australia] and NCT02106078 [United States]) that recruited study participants on the web from anywhere in the world. The RCT was run from 2 study sites located in Geelong, Victoria, Australia, and Palo Alto, California, United States. Details of the study conduct and analysis of the primary study outcomes have been described elsewhere [23].

In brief, adults (aged 21-65 years) with a diagnosis of BD type I, BD type II, or BD not elsewhere classified assessed by telephone using the Structured Clinical Interview for Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition, were eligible. Additional eligibility criteria included access to emergency care, visiting a health care provider at least twice per year for BD treatment, access to the internet and a computer, fluency in English, competence to provide informed consent, and willingness to provide emergency contact details. Once consent was obtained and inclusion and exclusion confirmed, we randomized participants on the secure website to 1 of 3 conditions:

Two-step block randomization was used and coded into the website during development. Research staff members were unable to view the randomization code.

For 12 months from randomization, all participants had access to the MoodSwings 2.0 website and their study arm–specific asynchronous peer discussion forum. Moderators screened discussion posts and edited or deleted those with personal contact information, profanity, or distressing content. Group 2 participants were additionally able to access 5 psychoeducational modules delivered biweekly, followed by 4 booster modules delivered at 3, 6, 9, and 12 months. The psychoeducation modules were adapted from a face-to-face clinician-facilitated manualized program, previously evaluated in randomized evaluations [24-26]. The participants randomized to group 3 were able to access the discussion forum, psychoeducation modules, and interactive CBT-based tools. This included development of a life chart, thought monitoring, simple motivational interviewing techniques, self-reflection, problem solving, identification of personal triggers, and a relapse-prevention plan [27].

This analysis was undertaken with data collected as part of the RCT approved by the institutional review board at Stanford University (Stanford, California, United States; project ID 21897) as well as the human research ethics committees at Barwon Health (Geelong; EC00208, project ID 11/73) and Deakin University (Geelong; EC00213, project ID 2021-072). The study was conducted in accordance with the ethical standards of the responsible committees.

The recommendations for economic analyses within an international trial suggest that resource use is costed with local unit costs, followed by analysis of heterogeneity [28]. This method requires country-specific unit costs from collaborators. As this study recruited participants on the web from multiple countries, local unit costs were difficult to source. The trial was managed from Australia and the United States, and an Australian health sector perspective was adopted for the economic evaluation. Health sector costs included the costs to deliver the interventions as well as the costs of other health services used by participants during the trial period (refer to Table S1 in Multimedia Appendix 1 [29]).

A microcosting approach was used to estimate the cost to deliver the 3 interventions. We estimated the personnel time required to monitor the internet-based forums as well as time for debriefing with a supervisor. Personnel costs were calculated by multiplying estimated hours by the average wage rate of a research assistant or supervisor, both with 25% added to account for employer overhead costs (eg, space and administrative overheads).

The cost of 2 desktop computers required for research assistants to monitor the internet-based forums was estimated based on an annual lease cost of Aus $800 (Aus $1=US $0.7058) per computer multiplied by the estimated time required to conduct the study (2.24 years).

The development and maintenance cost of the MoodSwings 2.0 website was provided by the study team as a single estimate. This total cost was apportioned across the 3 internet-based interventions based on complexity. The average health sector cost for each intervention group was then calculated based on the number of trial participants.

Information on participant health service use was captured through a self-report resource use questionnaire, the Cornell Service Index [30], at baseline and 3-, 6-, 9-, and 12-month follow-ups. The Cornell Service Index questionnaire asked about the number and types of medical, psychological, acute care, and support services accessed by study participants in the preceding 3 months. Standard Australian unit costs were applied. Intervention costs were added to the 3month health service costs; next, all health care service use costs over the 3- to 12-month follow-ups were summed.

All costs were presented in 2018-19 Australian dollars (Aus $). Discounting was not applied because the study time horizon was 12 months.

Self-report outcome measures were administered at baseline and 3-, 6-, 9-, and 12-month follow-ups, including the Montgomery-Åsberg Depression Rating Scale (MADRS), Young Mania Rating Scale, and the short-form health survey, 12-item version (SF-12). The MADRS score was a coprimary outcome measure that achieved statistically significant differences among the groups. It was used as an outcome measure for the cost-effectiveness analysis.

The SF-12 was used to measure participants’ health-related quality of life at each assessment time point. A preference-based scoring algorithm using British general population preference weights was applied to calculate utility values at each time point based on 6 questions from the SF-12 (short-form 6-dimension [SF-6D] instrument) [31]. Quality-adjusted life years (QALYs) were then calculated from the SF-6D utility values using the area under the curve method [32]. The use of QALYs in an economic evaluation is also referred to as a cost-utility analysis [33].

Statistical analyses were conducted using Stata software (version: 17.0; StataCorp LLC). Base case analyses were conducted on an intention-to-treat basis, including all participants with a baseline assessment. Missing cost and utility data were reported using descriptive statistics. The investigation of relationships between complete cost and outcome data with demographic and clinical variables was undertaken using logistic regression analysis.

Costs and utility values were reported at each time point by randomized group using descriptive statistics (mean and SD). The base case analysis used linear mixed effects models to evaluate between-group differences in postbaseline health sector costs, SF-6D utility values, and QALYs. Health sector costs and SF-6D utility values at each follow-up were regressed on time, baseline value, and treatment allocation with adjustment for baseline covariates specified a priori (baseline cost or utility, sex, and national origin). The model accounted for autocorrelation because of repeated measures across follow-ups and used an unstructured covariance matrix that allows all variances and covariances to be distinct.

Incremental cost-effectiveness ratios (ICERs) were calculated as the mean difference in total health sector costs between 2 randomized groups divided by the mean difference in MADRS scores. The 12-month follow-up was considered the primary time point for comparison in the main efficacy analysis, and this time point was adopted for the cost-effectiveness analysis. A nonparametric bootstrap procedure with 1000 iterations was used to calculate CIs around ICERs. Cost-effectiveness planes were constructed by plotting the 1000 bootstrapped incremental costs and incremental MADRS scores.

The incremental cost-utility ratio was calculated by dividing the mean difference in total health sector cost by the mean difference in QALYs. A nonparametric bootstrap procedure with 1000 iterations and the reordered bootstrap percentile method (1000 iterations) was used to estimate 95% CIs around each average incremental cost-utility ratio [34]. An intervention was considered cost-effective if the resulting ICER fell below the generally accepted Australian willingness-to-pay threshold of Aus $50,000 per QALY [35]. The resulting bootstrap iterations were also used to construct cost-effectiveness planes and acceptability curves to represent the uncertainty in the ICER.

Sensitivity analyses were undertaken to test the assumptions regarding missing data, including complete case analysis and multiple imputation for missing data at follow-up [36]. Missing total cost and outcomes data (utility values and MADRS scores) at each time point (3-, 6-, 9-, and 12-month follow-ups) were imputed through a resampling method using single imputation nested in bootstrapping [37]. This method generated a single call to the multiple imputation function in Stata, with chained equations and predictive mean matching, to produce a complete data set. The costs and outcomes were then analyzed with generalized linear models (GLMs) for each bootstrap resample. After the generation of 1000 bootstrap resamples, the reordered bootstrap percentile method was used to estimate 95% CIs around each average ICER [34]. In these sensitivity analyses the mean difference in total health sector costs over the 12-month follow-up among the randomized groups was estimated using GLMs [38] with the gamma family and identity link. The mean difference in QALYs among the randomized groups was estimated using GLMs with inverse gaussian family and identity link. All statistical models were estimated with adjustment for baseline covariates specified a priori (baseline cost or utility, sex, and national origin). The choice of family for each GLM was based on results from modified Park tests [38]. The link for each model was chosen based on a combination of Pearson correlation, Pregibon link, and modified Hosmer-Lemeshow tests [38].

An additional sensitivity analysis was conducted by estimating the intervention cost from population-level rollout. The average cost per study participant for variable cost items (personnel and computers) was added to the average cost of website development and maintenance per potentially eligible Australian with a diagnosis of BD. To provide a conservative estimate of potential users of the MoodSwings 2.0 program, the number of people with BD seeking care was estimated by multiplying the age- and sex-based prevalence of BD by Australian demographic statistics for the population aged 25 to 65 years in June 2018 [2,39]. The estimate was then multiplied by the percentage of people with BD using health care services for their mental health (67.7%) based on an Australian population-based mental health survey [40].

A threshold analysis was also undertaken to estimate the group 2 intervention cost required for the total cost to be the same as group 1 (control).

A total of 322 people provided consent and were screened for eligibility, with 304 (94.4%) participants randomized (refer to Figure S1 in Multimedia Appendix 1 [23]). There were no significant differences in baseline characteristics across the randomized groups (Table 1).

Self-reported resource use from the Cornell Service Index questionnaire and quality of life from the SF-12 were completed by 91.4% (278/304) of the participants at baseline, 39.5% (120/304) at 3-month, 33.9% (103/304) at 6-month, 35.5% (108/304) at 9-month, and 29.3% (89/304) at 12-month follow-ups (Table S2 in Multimedia Appendix 1). Overall, of the 304 participants, there were 84 (27.6%) with complete costs and QALYs over the 5 data collection points during the 12-month study period. Comparisons of participants with complete and incomplete data over the entire 12-month period found that sex was the only variable related to incomplete data; however, this may be due to the high percentage of female participants enrolled in the trial (228/278, 82%). It is unlikely that these data were missing not at random, given the similar patterns of missing cost and utility data that were observed across participants; as well as the qualitative differences between missing cost and utility data and their underlying values. On the basis of our exploratory analyses of missing data mechanisms, it was inferred that incomplete cost and utility data were missing at random. Multiple imputation was consequently used to account for missing data, while incorporating sex as a covariate.

Table 2 details the resources required, unit costs, and total costs for intervention delivery across the randomized groups. The average cost to deliver the control group intervention was estimated at Aus $421 per randomized participant. Group 2 and group 3 delivery costs were estimated at Aus $645 and Aus $714 per randomized participant, respectively.

The average health sector costs at each time point and totaled over the 12-month follow-up are detailed in Table S3 in Multimedia Appendix 1. The average health sector costs were not significantly different among the groups at baseline or over the 4 individual follow-up periods, except for a significant difference between group 2 and group 3 at 6-month follow-up (P=.01). The total average health sector cost, including the intervention cost, was lower for group 2 (Aus $9431) than for the control group (Aus $15,175), but this difference was not statistically significant. The total average health sector costs, including the intervention delivery costs, were comparable for group 3 and group 1 (control).

Table S4 in Multimedia Appendix 1 provides the average costs and SDs for participants who completed all Cornell Service Index questionnaires between 3 and 12 months by service use category across the randomized groups. The largest difference among the groups was noted for acute care costs between group 2 (mean Aus $1015, SD Aus $2206) and group 1 (mean Aus $6040, SD Aus $15,152).

The average MADRS scores were significantly different between group 2 and group 1 (control) at all follow-up time points (P≤.05), with a mean difference ranging between 3.6 (95% CI –0.001 to 7.2; 9-month follow-up) and 5.5 points (95% CI 1.8-9.2; 6-month follow-up; Table 3 and Table S5 in Multimedia Appendix 1 [23]). The only significant difference in MADRS scores between group 3 and group 1 (control) was at 6 months with a mean difference of 4.8 points (95% CI 1.0-8.5; P=.01).

The average SF-6D utility value was 0.63 at baseline across the randomized groups. From baseline to 12-month follow-up the average QALYs per group were not significantly different, with mean QALYs of 0.627 (SD 0.062) in group 1, 0.618 (SD 0.094) in group 2, and 0.622 (SD 0.087) in group 3 (Table S6 in Multimedia Appendix 1).

The average incremental cost per point improvement in MADRS scores for group 2 versus group 1 (control) was dominant at each follow-up time point and when summed over the study period (Table 3 and Figure 1). Dominant refers to the scenario when average incremental costs were lower and average incremental effects were higher for the intervention compared with the control group. The 95% CI ranged from dominated (higher incremental cost and negative incremental effect) to Aus $331 per point improvement in MADRS score.

The average costs per point improvement in MADRS score for group 3 versus the control group range from dominant (6- and 9-month follow-ups) to Aus $206 (12-month follow-up). Combining costs over the entire study follow-up leads to the group 3 intervention being dominant (less costly and more effective), with a wide CI from dominant to Aus $22,585 per point improvement in MADRS score (Table 3 and Figure 2).

The average costs per point improvement in MADRS score for group 2 versus group 3 range from Aus $1553 (9-month follow-up) to dominated (more costly and less effective at 3- and 6-month follow-ups). Over the entire 12-month period, the average ICER was Aus $11,140 per point change in MADRS score, with a wide spread of bootstrap iterations across all 4 quadrants of the cost-effectiveness plane making it difficult to interpret the CI (Table 3 and Figure 3).

The base case cost-utility analysis found that group 2 would be considered the dominant strategy compared with the control group based on the lower average health sector cost and an average QALY benefit of 0.012. The 95% CI for the average incremental cost-utility ratio ranged from Aus $43,000 per QALY to dominant (Table 4 and Figure 4); the lower CI was a result of lower costs and lower incremental QALYs. There was a 79% probability that the psychoeducation modules would be cost-effective at the threshold of Aus $50,000 per QALY.

The base case average incremental health sector cost for group 3 compared with the control group was estimated at –Aus $94 with an average benefit of 0.002 QALYs resulting in a dominant average ICER (95% CI dominated to –Aus $19,978; Table 4 and Figure 5). The CI is difficult to interpret because the bootstrap iterations span all 4 quadrants on the cost-effectiveness plane. The probability of the combination of psychoeducation and CBT tools being cost-effective at the threshold of Aus $50,000 per QALY was estimated at 51%.

Group 3 was dominated by group 2 in the base case because of higher average costs (Aus $7798) and fewer QALYs (–0.004; Table 4 and Figure 6). At the willingness-to-pay threshold of Aus $50,000 per QALY, the probability that the group 3 intervention would be cost-effective compared with the group 2 intervention was estimated at 7%.

The results for the comparison of group 2 with the control group were generally robust in the sensitivity analyses as shown in Table 5. The exception was multiple imputation that led to a nonsignificant negative mean difference in QALYs between group 2 and group 1 (control). The probability that the group 2 intervention would be cost-effective compared with the control group at the willingness-to-pay threshold of Aus $50,000 per QALY in the complete case analysis was estimated at 63% (Figure S2 in Multimedia Appendix 1).

The intervention cost of group 2 was also varied to assess the threshold when the mean cost difference between group 2 and group 1 would become positive. This occurred when the group 2 intervention cost was Aus $4500.

The sensitivity analyses for the comparison of group 3 with group 1 (control) were mostly consistent with the base case (Table 5). The analysis using multiple imputation led to negative incremental cost and QALY differences, both being nonsignificant, but leading to a positive incremental ICER. Across all sensitivity analyses, the probability of group 3 being cost-effective compared with group 1 (control) at the threshold of Aus $50,000 per QALY was ≤54% (Figure S3 in Multimedia Appendix 1).

The sensitivity analyses for the comparison of group 2 and group 3 were mixed. The complete case and multiple imputation analyses led to positive mean differences in QALYs and positive ICERs. The probability of group 3 being cost-effective compared with group 2 at the threshold of Aus $50,000 per QALY was ≤22% across all sensitivity analyses (Figure S4 in Multimedia Appendix 1).

Using the costs of the group 2 and group 3 interventions, if implemented across the population of people with BD in Australia (Table S7 in Multimedia Appendix 1), led to marginally lower mean differences in costs, which did not substantially change the ICERs.

To our knowledge, this is the first economic evaluation of an internet-based psychoeducation and CBT intervention specific to people with a diagnosis of BD [18,41]. The results suggest that the psychoeducation offered to group 2 through the MoodSwings 2.0 website may be cost-effective compared with an active control group of a moderated internet-based discussion board for people with a diagnosis of BD. The results also suggest that the addition of CBT tools to the psychoeducation component was not cost-effective compared with the moderated internet-based discussion board alone or the combination of psychoeducation plus the moderated internet-based discussion board.

The difference in cost between the participants randomized to the internet-based psychoeducation and the control condition (internet-based forum only), although not significantly different, showed a trend favoring internet-based psychoeducation. This was attributed to lower costs for acute care services such as hospitalizations and emergency room visits. These results are similar to those of research evaluating the costs and outcomes associated with an in-person 21-session group psychoeducation program for people with BD [42]. Over 5 years of follow-up, participants receiving the group psychoeducation had significantly fewer days hospitalized, which led to nonsignificant lower total costs for the psychoeducation group. Our results contrast with another trial-based cost-effectiveness analysis of an in-person 21-session structured group psychoeducation program that found significantly higher total costs and additional QALY gains for the participants receiving group psychoeducation compared with those receiving unstructured group peer support [43].

The MoodSwings 2.0 study group randomized to the psychoeducation modules showed statistically significant improvements in depression symptoms, as measured by MADRS scores, compared with the control group. These differences were also clinically meaningful, falling within the range of estimated minimal important difference of 3 to 6 points for the MADRS [44]. This is similar to results from the study by Lam et al [45] that found significantly improved scores on the Beck Depression Inventory at 4 and 6 months for the group receiving cognitive therapy versus a control group. The resulting average ICER for the psychoeducation intervention compared with the control group was dominant, meaning that there was improvement in MADRS scores at a cost savings.

Cost-effectiveness ratios such as cost per point change in MADRS score are difficult to interpret because of a lack of value attached to a point change in MADRS score. QALYs have inherent value-for-money connotations because of generally accepted willingness-to-pay thresholds used by health technology assessment agencies such as the United Kingdom’s National Institute for Health and Care Excellence and Australia’s Medicare Services Advisory Committee.

We did not find significant differences in utility values or QALYs among the groups over the 12-month follow-up. This contrasts with a small significant QALY gain of 0.023 (95% CI 0.001-0.56) associated with a previously evaluated in-person group psychoeducation intervention compared with in-person group peer support [43]. This may be due to the in-person mode of delivery, a longer follow-up of 96 months, use of the EQ-5D instrument, and lower rates of loss to follow-up.

Our results further suggest that the combination of psychoeducation and CBT tools (group 3) would not be considered cost-effective compared with the moderated internet-based discussion board (group 1) based on the cost-utility results. Group 3 had a trend toward lower costs and more QALYs compared with the control group, but there was a great deal of uncertainty around this dominant average cost per QALY ratio, resulting in a 51% probability of being cost-effective at the threshold of Aus $50,000 per QALY generally accepted as value for money in Australia. These results are comparable to economic evaluations of other unguided internet-based CBT interventions evaluated in people with unipolar depression [46,47].

The combination of internet-based psychoeducation and CBT tools (group 3) would not be considered good value for money compared with internet-based psychoeducation (group 2) based on the dominated average cost-utility ratio. The combination of internet-based psychoeducation and CBT tools (group 3) resulted in an average cost-effectiveness ratio of Aus $11,140 per point improvement in MADRS compared with internet-based psychoeducation (group 2). Although this seems favorable, it is harder to interpret because we do not have a willingness-to-pay threshold for a point improvement in depression symptom scores.

A prior evaluation of the MoodSwings 2.0 program found within-group improvements in depression and mania symptoms, medication adherence, and quality of life for participants receiving psychoeducation alone and psychoeducation plus CBT-based interactive elements [26]. The lack of an attention control group may explain the difference in findings compared with our evaluation.

As with all research, the results of this economic evaluation are subject to limitations. There was a high rate of loss to follow-up over the 12 months of the study period and a higher likelihood of missing cost and utility data for female participants, which may affect the validity of the results. The cost-effectiveness and cost-utility analyses would only be generalizable to the Australian context because of the exclusive use of Australian unit costs. The analytic approach followed published recommendations for the management of missing data. However, the complete case and multiple imputation results differed from the base case for the comparison of group 2 with group 3 as well as multiple imputation results differing from the base case for the comparison of group 3 with group 1 (control). There was also no treatment-as-usual control arm. The cost of programming and delivering the internet-based interventions was estimated based on the available information for this trial but may have been higher because of additional time for programming and system maintenance not captured in our projected costs. However, we found that the average total cost was lower for group 2 than for group 1 (control) until the intervention cost reached Aus $4500 per study participant, which is 7 times higher than the Aus $645 base case intervention cost for group 2.

Despite this evaluation’s limitations, it is important to report the results of economic evaluations of internet interventions aimed at supporting people with BD. Overall, there is limited literature on the cost-effectiveness of psychosocial interventions for the treatment of BD and none for BD-specific digital interventions [18,31]. People with BD are a unique population because of the symptomatology, medications, and behavioral aspects related to the diagnosis. Psychosocial interventions designed for other mental health conditions (ie, unipolar depression and anxiety) may not be appropriate to extrapolate to people with BD. It is important to tailor the information to the specific issues related to this diagnosis and evaluate program effectiveness and cost-effectiveness.

The availability of internet-based interventions is crucial, given lack of access to mental health professionals because of limited availability, geographic location, and, recently, public health measures related to COVID-19 infection control. The Australian Productivity Commission’s Inquiry Into Mental Health report recommended a national digital mental health platform with a gateway to digital and face-to-face treatment and support services. Any interventions provided through this mental health gateway should have evidence of, or at a minimum be concurrently evaluated for, their effectiveness and cost-effectiveness.

The internet-based psychoeducation provided through the MoodSwings 2.0 platform to the group 2 participants has the potential to be a cost-effective intervention for people diagnosed with BD. The group 2 psychoeducation component could be further evaluated in an implementation study for effectiveness and cost-effectiveness. Additional research is needed to understand the lack of effectiveness for the internet-based CBT tools provided as part of the group 3 intervention.