Cost-effectiveness of Guided Internet-Delivered Cognitive Behavioral Therapy in Comparison with Care-as-Usual for Patients with Insomnia in General Practice

ABSTRACT Study objectives Clinical guidelines recommend cognitive-behavioral therapy for insomnia (CBT-I) as first-line treatment. However, provision of CBT-I is limited due to insufficient time and expertise. Internet-delivered CBT-I might bridge this gap. This study aimed to estimate the cost-effectiveness of guided, internet-delivered CBT-I (i-Sleep) compared to care-as-usual for insomnia patients in general practice over 26 weeks from a societal perspective. Methods Primary outcomes were the Insomnia Severity Index (ISI, continuous score and clinically relevant response), and Quality-Adjusted Life Years (QALYs). Societal costs were assessed at baseline, and at 8 and 26 weeks. Missing data were imputed using multiple imputation. Statistical uncertainty around cost and effect differences was estimated using bootstrapping, and presented in cost-effectiveness planes and acceptability curves. Results The difference in societal costs between i-Sleep and care-as-usual was not statistically significant (-€318; 95% CI −1282 to 645). Cost-effectiveness analyses revealed a 95% probability of i-Sleep being cost-effective compared to care-as-usual at ceiling ratios of €450/extra point of improvement in ISI score and €7,000/additional response to treatment, respectively. Cost-utility analysis showed a 67% probability of cost-effectiveness for i-Sleep compared to care-as-usual at a ceiling ratio of 20,000 €/QALY gained. Conclusions The internet-delivered intervention may be considered cost-effective for insomnia severity in comparison with care-as-usual from the societal perspective. However, the improvement in insomnia severity symptoms did not result in similar improvements in QALYs.


Introduction
Insomnia is an important public health problem. Approximately one-third of the general population has symptoms of insomnia, while about 10% of the population meets all diagnostic criteria for an Insomnia Disorder. This makes insomnia the second most-prevalent mental disorder in the general population next to depression (Wittchen et al., 2011). The economic burden of insomnia to society is substantial, because individuals with insomnia use more healthcare services and have more functional impairments. Previous research has shown that participants with insomnia cost approximately 10 times more than good sleepers due to increased healthcare utilization, higher rates of absenteeism Amsterdam and is registered in the Dutch Trial Register (http://www.trialregister.nl/trialreg/index. asp, identifier: NTR5202). The study design has been presented in detail in the study protocol and will be briefly described here (Van der Zweerde et al., 2016).

Population
Between October 2015 and April 2018, insomnia patients were recruited through general practices. Half of the participating practices were recruited through the Amsterdam General Practitioner Network (N = 8). During the trial, an additional convenience sample was pursued due to low inclusion rates (N = 7). The 15 practices occupied a total of 50 GPs and 20 MHNPs over the 2.5 year inclusion period of the trial. Although GPs were predominately responsible for recruitment, MHNPs also participated in the process by referring patients they were already working with to the trial website or the GP.
Recruitment leaflets directing potential participants to the trial website were distributed in print in the waiting rooms of general practices, and in GPs' and MHNPs' consulting rooms. Study information was also displayed on television screens in the general practices that had such screens. GPs and MHNPs in participating practices were also asked to notify their (suspected) insomnia patients for the trial. Additionally, GPs approached patients who received consultation for insomnia within the past year by sending them the leaflet via e-mail. Individuals interested in participating were screened using an online questionnaire.
Eligible participants were 1) 18 years or older; 2) had access to the Internet; 3) had adequate proficiency in Dutch; and 4) met the DSM-5 criteria for insomnia disorder (American Psychiatric Association, 2013). Potential participants were requested to complete an online questionnaire in which the following DSM-5 criteria were assessed: a) difficulty initiating or maintaining sleep for ≥ 30 minutes per night for ≥ 3 nights per week, and for at least 3 months; and b) causing clinically significant distress and/or impairment in daily functioning. The DSM-5 diagnosis of insomnia was not validated by a clinician during an interview. Distress was measured using items of the Daytime Consequences questionnaire developed by Espie and colleagues (Espie et al., 2012). The questionnaire consists of six items which measure the effect of insomnia on energy, relationships, mood, concentration, productivity, and sleepiness. Items were rated on a 5-point scale ranging from 0 (not affected at all) to 4 (very much affected). In order to be included in the trial, a score of at least 3 on at least one domain of the questionnaire was required.
Exclusion criteria were 1) pregnancy or breast feeding during participation; 2) working in night shifts; 3) other current psychological treatment; 4) being suicidal (assessed using 5 items from the MINI diagnostic interview (Sheehan et al., 1998), and 5) schizophrenia or psychotic disorder ("Have you been diagnosed with schizophrenia or psychosis", answered with "Yes" or "No"). Other comorbid psychological and somatic disorders were not used as an exclusion criterion and neither was the use of sleep medication. Eligible participants received a study brochure along with a sleep diary, informed consent form and stamped return envelope.

Randomization and sample size
Participants that returned completed sleep diaries and informed consent were randomized on a 1:1 ratio into either the i-Sleep or the care-as-usual group. Two independent researchers conducted the randomization, using a random allocation sequence, and allocated the included patients to one of the two groups. Randomization was stratified by medication use in the past month and MHNP practice in random blocks of two, four, and six. Due to the nature of the interventions, blinding participants or MHNPs to the treatment allocation was not possible. GPs were informed when participants were randomized to ensure they would receive care-as-usual. Assuming an effect size of Cohen's d = 0.50 on the Insomnia Severity Index (ISI) and a dropout risk of 20%, a sample size of 160 participants was needed to ensure sufficient statistical power (80%) (Cohen, 1988;Horsch et al., 2015). The inclusion phase of this trial has prematurely ended at 134 out of 160 participants due to other research on insomnia planned in the participating GP practices. Based on the within-between group analysis and the sample size of n = 134 a power of 0.8 (alpha = .05) was obtained to detect an effect size of d = .24.

Internet-based intervention
The internet-delivered guided intervention, i-Sleep, is based on face-to-face cognitive-behavioral therapy for insomnia and includes the elements usually incorporated in this treatment (Edinger & Means, 2005;Edinger & Wohlgemuth, 1999;Espie, 2006;Morin & Espie, 2003;Verbeek & Klip, 2005). The i-Sleep intervention consists of five lessons with different topics: 1) psycho-education and (un)healthy sleep hygiene, 2) shortening time spent in bed and leaving the bed when awake (sleep restriction treatment and stimulus control), 3) relaxation exercises and rumination, 4) handling dysfunctional cognitive beliefs (about sleep and other) and 5) summary and relapse prevention. Participants had to fill out a sleep diary at every lesson and write down their plan for the coming period. Participants were advised to do one lesson a week and could only continue to a next lesson after finishing the previous one. The MHNP provided online written feedback every week on the homework and sleep diary. The aim of their feedback was to clarify information and motivate the participants to carry out the course and the requested behavioral changes. The structure of the intervention and content of the modules is described in more detail elsewhere (Van der Zweerde et al., 2016. Prior to the trial, MHNPs received training in using the i-Sleep program (1.5 hours) and in offering online guidance during the i-Sleep program. Further supervision from a specialized psychologist (TvdZ) was provided on an as needed basis. The time spent on giving online feedback was estimated at around 15-20 minutes per session. Participants in the i-Sleep group were also allowed to receive care-as-usual from their GP or MHNP.

Care-as-usual
The participants allocated to the control group received care-as-usual. GPs participating in the trial received a printed copy of the most recent Dutch GP insomnia guidelines which recommend cognitive-behavioral therapy for insomnia as the first line of treatment (Nederlands Huisartsen Genootschap [NHG] Richtlijnen, 2014). This ensured alignment of knowledge on the current sleeping treatments among the GPs. GPs were informed about participants who were randomized in the care-as -usual group, who were then invited to receive i-Sleep six months upon randomization. Care-as-usual provided by GPs was not restricted in any way throughout the trial; therefore, the content of the consultations in the care-as-usual group is unclear.

Clinical outcomes
Primary outcomes in this economic evaluation were the Insomnia Severity Index (ISI) (Bastien et al., 2001;Morin et al., 2011) and Quality-Adjusted Life-Years (QALYs). The continuous ISI score was assessed at baseline, post-treatment (8 weeks), and at follow-up (26 weeks). The ISI comprises seven items assessing the perceived sleep quality, insomnia nature and severity as well as its impact on physical and social functioning. Items are scored on a Likert scale from 0 (not at all) to 4 (extremely). Summed scores can range from 0 to 28, with higher scores indicating higher severity of insomnia (Bastien et al., 2001). Prior research has shown that ISI is a reliable tool with adequate internal consistency to measure perceived insomnia severity . It has also been validated for online use (Thorndike et al., 2011). The ISI was also used to calculate the percentage of participants with clinically relevant response to treatment. A clinically relevant response was defined as a change of ≥ 8 points in ISI score from baseline to follow-up (26 weeks) . The effectiveness outcomes for both i-Sleep and care-as-usual are extensively described in the clinical trial paper by van der Zweerde et al (Van Der Zweerde et al., 2020).
QALYs were calculated based on health-related quality-of-life measurements at baseline, and at 4, 8 and 26 weeks after baseline using the five-level version of the EuroQol (EQ-5D-5 L) (Herdman et al., 2011). The EQ-5D-5 L contains five dimensions (mobility, self-care, activities of daily living, pain/discomfort and depression/anxiety) with 5 answer levels (no problems to severe problems). The EQ-5D-5 L health states were converted to utility scores using the Dutch tariff, where 0 refers to death and 1 to full health (range −0.446 to 1, where negative utilities indicate that a health state is valued as worse than death) (Versteegh et al., 2016). The negative values of the tariff stem from the methods used to valuate health states; that is, in some cases, the respondents consider the health states they are presented with as "worse than death". Using the area under the curve method, QALYs were calculated by multiplying the amount of time a patient spent in a specific health state with the utility score associated with that health state. Transitions between health states were linearly interpolated.

Costs
Costs and healthcare utilization were assessed from a societal perspective using an adapted version of the Trimbos and i-MTA Questionnaire on Costs Associated with Psychiatric Illness (Tic-P) (Hakkaart-van Roijen et al., 2002). The Tic-P is a comprehensive, self-report questionnaire that focuses on estimating direct healthcare costs and productivity costs related to mental disorders. Existing evidence suggests that Tic-P is a feasible, well-understood tool that shows satisfactory reliability and construct validity in measuring visits to specialists and productivity losses (Bouwmans et al., 2013). The Tic-P questionnaire was administered at baseline, and at 8 and 26 weeks after baseline. The recall periods participants were asked to consider when reporting medication use, number of healthcare visits, and productivity losses were 1 month at baseline, 2 months at 8 weeks, and 4 months at 26 weeks after baseline.
Healthcare costs included all types of visits to primary care (e.g., general practitioner, physiotherapist) and secondary care (e.g., medical specialist, psychologist or psychiatrist, hospital admission), and medication costs. Costs of informal care consisted of patient and family costs. Informal care is defined as unpaid care provided by family members or volunteers to support patients with daily activities (e.g., eating, bathing, etc). These costs were calculated by multiplying the units of resource use by their unit price using standard costs from the Dutch costing guidelines (Hakkaart-van Roijen et al., 2015). Medication use was valued using prices from the Dutch Healthcare Institute on www.medicijnkosten.nl. Since no specific market price is available for the cost of informal care, a shadow price (i.e. proxy value) equal to the tariff of cleaning work was assumed (Hakkaart-van Roijen et al., 2015).
Lost productivity costs consisted of costs of absenteeism from paid and unpaid work, and presenteeism (i.e. an employee being present at work but not fully productive due to ongoing physical or mental conditions). Costs of absenteeism from paid work and presenteeism were valued using gender-specific incomes of the Dutch population and calculated according to the friction cost approach. This method assumes that within a specific time-span (i.e. friction period of 12 weeks), the sick employee is replaced and there are no more productivity losses thereafter (Hakkaart-van Roijen et al., 2015;Van den Hout, 2010). Absenteeism from unpaid work was calculated using the same shadow price as for informal care.
The costs of the i-Sleep intervention were calculated using a bottom-up approach. Costs that were included were 1) costs associated with using the online platform and 2) costs related to the time invested by the MHNPs in providing feedback, which depended on the number of completed lessons. Subscription costs to the online platform were assumed to be equal to €39, which is equal to the subscription costs per patient charged by a large organization hosting internet-delivered mental health interventions (Minddistrict). Costs of guidance by the MHNP were calculated using the standard reference price per consult by a MHNP (Hakkaart-van Roijen et al., 2015). All costs were in Euros and adjusted to the year 2018 using consumer price indices if necessary (Statline Netherlands, 2020). Discounting was not necessary, since the follow-up period was 26 weeks. The cost categories and prices used in this economic evaluation are presented in more detail in the Supplementary Material of this paper.

Statistical analysis
All analyses were conducted according to the intention-to-treat principle. Missing cost and effect data were imputed using the Multiple Imputation with Chained Equations (MICE) algorithm with predictive mean matching (Van Buuren & Groothuis-Oudshoorn, 2011). Besides variables included in the analysis models, the imputation model included variables that 1) differed at baseline between treatment groups; 2) were related to the missingness of data; or 3) were associated with any of the outcomes. Predictor variables that were related to missingness included age, living situation, and quality of life as measured at baseline. Therefore, it was assumed that cost and effect data were missing at random (Faria et al., 2014). By multiple imputation (MI), 10 imputed datasets were created resulting in a loss of efficiency lower than 5% (White et al., 2011). Each imputed dataset was analyzed separately and results from the 10 datasets were pooled using Rubin's rules (Rubin, 1987).
Bivariate regression was performed to estimate cost and effect differences between the treatment groups, while accounting for potential correlation between the cost and effect outcomes (Willan et al., 2004). Incremental Cost-Effectiveness Ratios (ICERs) were calculated by dividing the difference in costs by the difference in effects between the groups. Bias-corrected and accelerated bootstrapping with 5000 replications was used to estimate 95% confidence intervals around the cost and effect differences (Efron, 1994). Statistical uncertainty surrounding the ICERs was graphically presented by plotting the bootstrapped incremental cost-effect pairs on a cost-effectiveness (CE) plane (Black, 1990).
Cost-effectiveness acceptability curves (CEACs) were also estimated, showing the probability that the i-Sleep intervention is cost-effective compared to care-as-usual for a range of different ceiling ratios (Fenwick et al., 2004). Analyses were performed using IBM SPSS Statistics 24® (IBM Corp., Armonk, NY, US) and StataSE 14® (StataCorp LP, CollegeStation, TX, US).

Sensitivity analysis
To check the robustness of the results, three sensitivity analyses were performed. The first sensitivity analysis was conducted from the healthcare (government) perspective, which is used for decision-making in other countries, such as the United Kingdom. The second sensitivity analysis concerned a per protocol analysis comprising only participants who completed all five sessions of the i-Sleep intervention and all patients randomized to the care-as-usual group. In the final sensitivity analysis, lost productivity costs due to absenteeism were calculated using the human capital instead of the friction cost approach which is more common in other countries. The friction cost method assumes that lost productivity costs are confined to the period needed to replace a sick worker, whereas the human capital method estimates lost productivity costs over the whole period of absence.

Population
In total, 134 individuals who were willing to participate met the inclusion criteria and were randomized to the i-Sleep intervention group (n = 69) or care-as-usual (n = 65). In the i-Sleep group, 47 (68%) participants completed all five sessions of the i-Sleep intervention, 14 (20%) did not complete the treatment, and 8 (12%) never started the lessons. At 8 weeks, 56 (81%) and 55 (85%) participants responded to the questionnaires in the i-Sleep and care-as-usual group, respectively. At the 26 weeks follow-up, the response was 53 (77%) and 54 (83%) respondents in the i-Sleep and care-as-usual groups, respectively. Baseline characteristics are presented in Table 1. Mean scores for daytime consequences were slightly higher for participants in the care-as-usual group. The flow of participants through the trial is shown in Figure 1.

Clinical outcomes
The multiply imputed clinical outcomes at 26 weeks follow-up are presented in Table 2. The mean pooled continuous ISI score in the i-Sleep group was statistically significantly lower than in the care-as -usual group (mean difference 3.9 points; 95%CI 1.6 to 6.3). The percentage of participants experiencing a clinically relevant response according to ISI at 26 weeks as compared to baseline in the i-Sleep group was statistically significantly higher than in the care-as-usual group (mean difference 28%; 95% CI 0.13 to 0.44). The mean pooled difference in QALYs over 26 weeks between the groups was 0.01, which was not statistically significant (95%CI −0.001 to 0.02). Table 2 summarizes the pooled mean costs and the adjusted differences in costs between the two groups. Mean intervention costs were on average €99 per patient in the i-Sleep group, including subscription costs for the online platform and the costs of online consultations by the MHNP. Around one-third of participants in the i-Sleep group did not complete all sessions of the intervention. Therefore, the costs of online support provided by the MHNP were lower than anticipated at the start of the study. There was no statistical difference in total healthcare costs between the i-Sleep group and care-as-usual (mean difference €220; 95%CI −83 to 503). The largest contributor to the difference in total healthcare costs were costs of secondary care. Lost productivity costs in the two groups were relatively similar. However, informal care costs in the i-Sleep group were statistically significantly lower compared to the care-as-usual group (mean difference -€586; 95% CI −1735 to −14). The difference in total societal costs between the i-Sleep group and care-as-usual was not statistically significant (95%CI: −1282 to 645), but indicated that on average societal costs in the i-Sleep group were lower than in care-as-usual.

Cost-effectiveness and cost-utility
The results of the cost-effectiveness and cost-utility analyses are summarized in Table 3. For the continuous ISI score outcome at 26 weeks, the ICER was −81, meaning that 1 point improvement in insomnia severity score was associated with a saving of €81 for the i-Sleep as compared to the care-as-usual group. The CE plane for ISI score (Figure 2) shows that the statistical uncertainty surrounding the ICER is located in all four quadrants (Table 3) indicating considerable uncertainty. The majority of the bootstrapped cost-effect pairs (the gray dots) is gathered in the southeast quadrant, indicating that there is 59% probability for i-Sleep being more effective and less expensive compared to care-as-usual. The CEAC for ISI score (Figure 3) shows that if society is not willing to invest any money in the i-Sleep intervention, i.e. the ceiling ratio is 0 € per point of improvement in ISI score, the probability of i-Sleep being cost-effective in comparison with care-as-usual is 62%. At ceiling ratios of 250 and 450 € per point of improvement in ISI score the probability of i-Sleep being cost-effective compared to usual care is 87%, and 95%, respectively. For the response outcome, the ICER was −1122 (see Table 3), indicating that 1 respondent with a clinically relevant response extra is associated with a saving of €1,122 in the i-Sleep group in comparison with the care-as-usual group. Approximately 60% of the bootstrapped cost-effect pairs are located in the southeast quadrant, confirming that i-Sleep is more effective and less expensive. At ceiling ratios of 0, 3,000 and 7,000 € per extra clinically relevant response, the probability that the i-Sleep intervention is cost-effective compared to care-as-usual is 61%, 84% and 95%, respectively. The CUA resulted in an ICER of −28,056, meaning that to gain 1 QALY €28,056 is saved with the i-Sleep intervention as compared to care-as-usual. The probability of the i-Sleep intervention being costeffective compared to care-as-usual was 61%, 64%, and 67% at ceiling ratios of 0, 10,000 and 20,000 € per QALY gained, respectively.

Sensitivity analyses
The sensitivity analysis results are also presented in Table 3. In the analysis from the healthcare perspective (SA1), mean total costs were higher in the i-Sleep compared to the care-  as-usual group (mean difference €217, 95%CI −15 to 449). At a ceiling ratio of 0 € per extra unit of effect, the probability of i-Sleep intervention being cost-effective compared to care-asusual was 12% for the ISI score, clinically relevant response, and QALY outcomes. This increased to 95% for the ISI decrease and response outcomes at ceiling ratios of 200 and 3000 € per additional unit of effect, respectively. For the QALY outcome, the probability of i-Sleep intervention being cost-effective increased to 50% at a willingness-to-pay (WTP) ratio of 20,000 € per QALY gained. The second sensitivity analysis included only participants who completed all five lessons of the i-Sleep intervention (68% of the total sample). A larger cost difference was found between groups as compared to the main analysis (mean difference -€489, 95%CI −1442 to 464). In addition, the mean pooled difference in QALYs became larger and was statistically significant (mean difference 0.02; 95%CI 0.006 to 0.03). The CEA curves indicated that the probability of i-Sleep being cost-effective was 67% at a ceiling ratio of 0 € per extra unit of effect for all outcomes. This probability increased to 95% at a WTP ratio of 5000 € per extra unit of effect for ISI score and 75% at a ceiling ratio of 20,000 €/QALY. In the final sensitivity analysis (SA3), results were similar to the main analysis.

Main findings
This economic evaluation showed that the difference in societal costs between i-Sleep and care-asusual was not statistically significant, although on average costs in the i-Sleep group were lower than in care-as-usual. Based on the CEA curves, we conclude that i-Sleep was cost-effective in comparison with care-as-usual both for the continuous ISI scores as for the percentage of participants with a clinically relevant response from a societal perspective. Although the point estimate of the ICER in the CUA indicates that i-Sleep was dominant over care-as-usual, there was considerable statistical uncertainty. Therefore, society should be willing to pay some money to reach acceptable probabilities of cost-effectiveness. Overall, sensitivity analyses showed similar results as the main analysis.
A clinically relevant improvement in insomnia severity symptoms was observed in the trial. These positive effects were not accompanied by a comparable effect on QALYs. It is possible that the EQ-5D-5 L that was used to calculate QALYs is not sensitive enough to detect changes in quality of life as a consequence of reduced insomnia symptoms. In a recent systematic review on the economic aspects related to CBT-I, the authors also concluded that the QALY might not be the most appropriate measurement to detect significant changes in insomnia symptoms related to treatment (Wickwire & Morin, 2020).
Evidence from a previous cost-effectiveness study comparing an internet-based CBT-I intervention with a waitlist control condition suggested that internet-delivered treatments have a higher probability of being cost-effective at relatively low willingness-to-pay ratios for treatment response than in our study (Thiart et al., 2016). Participants in this study were employees with work-related insomnia. Thus, the analysis was conducted from an employer's perspective, meaning that only lost productivity costs due to presenteeism and absenteeism from paid work were taken into account, whereas the current analysis was performed from a societal perspective. However, the most probable explanation for the difference between the two studies is that in the current study care-as-usual was used as a comparator. We hypothesize that because care-as-usual is an active treatment whereas waitlist control is not, differences in effects between the treatment groups are smaller than in Thiart et al. (2016).

Strengths and limitations
An important strength of the current study is the pragmatic design of the clinical trial, which enhances the generalizability of results to the average patient consulting their GP for sleep issues. Another strength is that we used the societal perspective, as recommended by the Dutch guidelines for economic evaluations, meaning that lost productivity costs are also included. Because insomnia has significant effects on work performance, this is an important strength of the study. Also, use of the societal perspective offered insight into potential shifting of costs between sectors. That is, costs might be incurred (or saved) in sectors other than the healthcare sector. As a result, an intervention might be cost-effective from the healthcare perspective, but not for the society as a whole. However, in order to use the societal perspective, it was necessary to use self-reported outcomes to measure healthcare utilization and productivity losses. This could potentially be a limitation of this study due to the occurrence of recall bias, although previous studies have shown that self-reported measures can be reliably used for recall periods up to 6 months ( Van den Brink et al., 2005).
Another limitation was the rate of missing data (approximately 26%). Because total costs are a sum of the costs at the different time points, one missing observation already results in missing total cost estimations, which may explain the high rate of missing data. Multiple imputation was applied to account for the missing values, which is currently considered the most appropriate method to handle incomplete data in cost-effectiveness analyses (Burton et al., 2007). Additionally, the trial sample may not have been representative of the target population, with a higher percentage of respondents being white, female, and highly educated as compared to the general population. Finally, generalizability of the findings might be limited by the fact that we used a convenience sample of general practices. It is not certain whether this sample adequately represents Dutch GPs. It is possible that participating practices were more involved in research in general, more engaged with new interventions, or more interested in sleep problems than other practices.

Implications
The i-Sleep intervention was considered cost-effective for insomnia severity symptoms when considering all patients, including those who did not finish the i-Sleep intervention (intention-to-treat analyses). When looking at completers only (67% of the sample), the effects on insomnia severity symptoms and QALYs were larger than in the main analysis, which resulted in higher probabilities of cost-effectiveness for i-Sleep. Further research is needed to identify strategies to increase adherence to internet-based mental health treatments even more, as this may increase cost-effectiveness as well.
Moreover, the recruitment process was not as productive as anticipated considering the number of participating GPs and MHNPs. A possible explanation for the slow inclusion could be that GPs often identify insomnia much less than expected given the prevalence rates (Bhaskar et al., 2016;Taylor et al., 2007). Insomnia is often comorbid with other mental and physical conditions. As a result, patients tend not to report insomnia-specific symptoms and, subsequently, GPs tend to focus less on treating insomnia. It is therefore possible that participating GPs did not always mention the trial during their insomnia-related consults. However, it should be stressed that referring patients as part of a clinical trial can be much more burdensome for healthcare professionals than in real-life practice, where treatment uptake would likely be enhanced. Additionally, as seen on the patient flow chart (Figure 1), considerably more patients were referred to than included in the trial, which suggests that GPs and nurses were not mainly responsible for the low inclusion rates. Patients have been reluctant to participate as well, possible because they were overwhelmed with health complaints and, as it is always the case in a RCT, wanted to avoid the 50% chance of receiving usual care. Future research should focus on GP awareness and involvement in the recruitment process as well as real-life uptake of i-Sleep treatment.
Improvement in insomnia severity symptoms was not linearly accompanied by increases in QALYs which is the primary outcome of interest for healthcare decision makers. Therefore, future research should be conducted to assess whether the EQ-5D (−5 L) is sensitive enough to pick up changes in quality of life as a consequence of decreased insomnia severity symptoms. Finally, healthcare costs and costs related to absenteeism from paid work were similar in both treatment groups. However, costs related to presenteeism, absenteeism from unpaid work and informal care were lower in the i-Sleep group than in the care-as-usual group. Although it is reasonable to assume that the reduction in costs associated with presenteeism and absenteeism from paid work are a direct result of the improvement in insomnia severity symptoms, this is less clear for informal care costs. This should be confirmed in future studies.

Conclusions
In conclusion, based on this economic evaluation the internet-delivered i-Sleep intervention may be considered cost-effective as compared to care-as-usual for patients with insomnia treated in general practice for insomnia severity symptoms from a societal perspective. However, the improvement in insomnia severity symptoms was not accompanied by substantial improvements in QALYs and therefore the intervention cannot be considered cost-effective for QALYs.

Clinical trial registration
ID: NL5071 (NTR5202); trial name: (Cost-) effectiveness of a guided online CBT intervention in comparison to care-asusual for patients with insomnia in general practice; URL: https://www.trialregister.nl/trial/5071

Disclosure statement
No potential conflict of interest was reported by the authors.

Statement of significance
This is one of few studies to examine the cost-effectiveness of internet-delivered CBT for insomnia in the general practice. It is also the first cost-effectiveness study to take into consideration both the societal and healthcare perspectives, which yielded different results. Considering both costs and clinical effects simultaneously, internet-delivered CBT-I was found cost-effective as compared to care-as-usual with regards to clinical outcomes (insomnia severity scores). These findings are significant for decision makers given the high economic burden of insomnia and they show the potential of using internet-delivered CBT-I to address the lack of resources and expertise in treating insomnia symptoms. Further research is needed to improve adherence in internet-delivered CBT-I and better understand the relationship between insomnia and quality-of-life measurements.