Economic evaluation of a dietary intervention for adults with major depression (the “SMILES” trial)

Australian health sector and societal perspectives were used for the analyses to reflect different decision makers and contexts as per current guidelines for reference cases in economic evaluations [9]. The health sector costs included in this evaluation were those required to deliver the dietary support intervention or the social support control condition, as well as the cost of health care resources used by participants over the trial period. Societal costs comprised the health sector costs in addition to the cost of patient transportation, food and effects on productivity. The cost and outcomes included in the analysis are reported in the Impact Inventory (Additional file 1: Table S1).

The main resource required to deliver both the dietary and the social support was the salary of trained personnel. Detailed records of the number and length of sessions for each participant were collected by dietitians and befrienders. Hourly wage rates with on-costs of 27% were applied to the total time reported to provide the support for each participant. For the dietary support group, the rate was based on a grade two, year one dietitian from the Victorian Hospitals’ Industrial Association salary circular [11]. The hourly wage rate to deliver the social support condition was based on an entry level research assistant salary (Level A, step1) at Deakin University [12]. Since the dietary support intervention aimed to modify the dietary intake of participants, the cost of food was also included in the societal perspective. Food costs were estimated at $112 per week for the dietary support group participants and $138 per week for the social support group participants. This was based on the cost of the first 20 study participants’ dietary intake at baseline compared to the cost of the prescribed healthy alternative, the modified Mediterranean diet [13]. The societal perspective also included transportation costs to attend the dietary or social support sessions. These were estimated by assuming a 30 km round trip in a private vehicle per session, and then applying the cost/km typically used for tax reimbursement in Australia [14].

Participants self-reported the use of prescription medications, over the counter medications and supplements, the number of health professional visits (GPs, psychiatrists, psychologists etc.), and hospitalisations with a resource-use questionnaire administered at baseline and the 12 week follow-up visit. The questionnaire asked for all health care use and was not limited to the use of services specifically for mental health concerns. The health-care costs included in the analyses were those paid by the government as a third-party payer and the out of pocket costs borne by participants, such as co-payments for prescription medications.

Pharmaceutical Benefits Schedule (PBS) data was used to calculate the government and patient out of pocket costs for covered medications [15]. We assumed that all participants would pay the general patient co-payment ($36.10). Online Australian retail pharmacy sites were consulted to determine patient costs for other medications and supplements not covered by the PBS [16‐18]. Health professional visits were costed using a weighted average cost paid by the government, derived from the Medicare Benefits Schedule (MBS) item reports [19]. Since a standard co-payment for health professional visits is not in place under the MBS, participants were asked to report the estimated out of pocket costs they paid for these services.

Community mental health visits were costed using data from the 2010 Australian Mental Health Report [20] and the value then inflated to 2013/2014 dollars using the total health price index calculated by the Australian Institute for Health and Welfare [21].

Hospital stays were costed using public sector average cost per separation, based on Australian Refined Diagnosis Related Groups (AR-DRGs) from the National Hospital Cost Data Collection [22]. The specific AR-DRGs were selected based on the reported reason and length of stay.

The societal perspective incorporated patient transportation costs and effects on productivity. Transportation costs for health professional visits were calculated from the distance travelled in kilometres reported in the resource use questionnaire. We assumed a private vehicle was used and therefore multiplied the distance by a cost/km typically used for tax reimbursement [14]. Transport to hospital was estimated based on the cost for a 15 km taxi ride in Melbourne [23].

The resource use questionnaire also asked participants to self-report days off from paid and unpaid work (absenteeism) as well as days worked while suffering health problems as a proxy for presenteeism. To calculate the hours of lost productivity due to presenteeism, we assumed that 1.2 h of productivity was lost per day when a person reported working while suffering health problems [24]. The human capital approach [25] was used to value lost paid productivity by using an average hourly wage rate calculated from the average weekly earnings reported by the Australian Bureau of Statistics plus 25% overhead costs [26]. Time off from unpaid activities (i.e. housework) was valued at 25% of the average wage rate plus overhead costs to represent the value of participants’ lost leisure time [27].

All costs are presented in Australian dollars (AUD) for the 2013/2014 financial reference year. Since the costs and outcomes were collected over a 12 week period, discounting was not applied.

The primary health outcome for the current study was quality adjusted life years (QALYs). QALYs are useful outcome metrics since they can be applied across different disease/disorder areas, thus allowing comparability across a range of programs when evaluating economic decisions. For example, in Australia $50,000/QALY has been used as a rule of thumb to denote value-for-money [28, 29]. QALYs are derived by “weighing” the length of life spent in a particular health state by the utility or value of that health state. Utility values (or weights) are constrained between 0 and 1 where 0 refers to death and 1 refers to perfect health and values in between denoting less than perfect health states. To assess participants’ health-related quality of life and utility, the Assessment of Quality of Life – Eight Dimension (AQoL-8D) [30] was completed at each assessment. The AQoL-8D is a multi-attribute utility instrument, with a separate utility scoring algorithm associated with the instrument allowing the calculation of utility values for each participant. The utility algorithm used in the current study was derived from the Australian general population [31]. The QALYs for the 12 week follow-up period in the current study were estimated using the area under the curve method [32].

Data management and costing were completed using Excel 2013, while statistical analyses were conducted with Stata 14 (College Station, Texas, USA). SMILES was originally powered to detect differences between groups based on the primary clinical endpoint of MADRS. Therefore, the sample sizes were not designed to test the cost-effectiveness hypotheses. The reference case analyses were undertaken as intention to treat. All enrolled participants who completed a baseline assessment were included; however, 33 % of participants did not complete all of the health care resource use questions at the 12 week follow-up. The missing at random assumption was tested through a series of logistic regression analyses comparing participant characteristics for those with and without missing endpoint data. To account for missing utility values, health care, transportation, and lost productivity costs, the ICE multiple imputation technique in Stata was used [33]. Missing data were imputed 33 times based on the percentage missing from the health care utilisation variables [34]. Generalised linear models (gamma family, log link), using the mim command to combine the estimated coefficients across the imputed datasets, were used to determine the size and significance of differences between groups for total costs (health sector and societal) and QALYs. For each outcome, unadjusted models were developed along with models incorporating the covariates of age, gender, baseline utility and baseline cost.

Incremental cost-effectiveness ratios (ICERs) were calculated as the difference in average cost between the dietary support and social support groups, divided by the difference in average QALYs. Confidence intervals (CIs) for the incremental costs per QALY gained were calculated using a nonparametric bootstrap procedure, with 1000 iterations to reflect the sampling uncertainty. The bootstrapped ICERs and the CIs were graphically represented on cost-effectiveness planes. A cost-effectiveness plane is a plot of the 1000 bootstrapped incremental costs and outcomes across four quadrants. The north-east quadrant represents the intervention costing more as well as conferring greater benefits than the comparator. The south-east quadrant shows the proportion of iterations where the intervention costs less but incurs greater benefits than the comparator (i.e. a “dominant” intervention), the north-west quadrant shows the proportion of iterations where the intervention incurs a cost but fewer benefits than the comparator (i.e. a “dominated” intervention) and, lastly, the south-west quadrant shows the proportion of iterations whereby the intervention costs less and has fewer benefits than the comparator group.

Imputation uncertainty was evaluated by comparing the reference case results to the analysis of participants with complete data. Generalised linear models (unadjusted and adjusted) were used to evaluate the size and significance of differences between groups, and non-parametric bootstrapping was also employed concordant with the reference case.

Additional sensitivity analysis was undertaken by varying the cost of the social support control condition to $0 to represent a scenario where the comparator would be a ‘do nothing’ approach. The cost of the dietary support intervention was also varied to assess the effect on the incremental cost difference between groups. Assumptions regarding the food costs were explored by varying the cost of food for the dietary intervention group.

The analysis of complete cases, defined as participants with complete cost and QALY data at the 12 week follow-up, indicated similar results to the imputed ITT analysis as shown in Table 3. The cost-effectiveness planes (Additional file 1: Figures S1 and S2) show that the dietary intervention was associated with a higher probability of reduced costs and increased QALYS (81 and 80% respectively for health sector and societal perspectives).

The results of the sensitivity analyses are presented in the Additional file 1: Table S4 and S5. When the cost to deliver the social support control condition was adjusted to $0, total health sector costs were no longer significantly different between groups for all ITT results, as well as the complete cases in the adjusted model. The costs from the societal perspective remained significantly different between the study groups, with the exception of the complete case analysis with the adjusted model, where food costs were removed and the social support costs adjusted to $0. When the cost of the dietary intervention increased by a factor of 1.5, the difference in total health sector costs between groups was no longer significant using the adjusted glm (coefficient = − 0.491; p = 0.059). As shown in Additional file 1: Table S5, the cost of food for the dietary intervention group would need to increase to $2325 for the dietary intervention to no longer be significant using the results from the base case ITT unadjusted glm model. Using the base case ITT adjusted glm model the dietary intervention was no longer significant when the cost of food increased to $2600. The sensitivity results for the complete cases showed that the results became non-significant at smaller increases in food costs.

A strength of the current study was the collection of health care resource utilisation and health preference data prospectively within the trial and the application of conservative unit cost estimates to determine the cost/QALY from a health sector perspective. This evaluation also included a societal perspective as recommended by the Second Panel on Cost-Effectiveness in Health and Medicine [9]. The study design used the social support condition as an active control, which was intended to avoid a nocebo effect that a usual care comparator may induce and to control for the non-specific benefits of face to face interactions [43].

The results of this economic evaluation should, however, be viewed cautiously due to several limitations. The sample size included in the trial was small (Dietary Support group n = 33, Social Support group n = 34) and limit feasibility and generalisability of the dietary support intervention. Despite the high prevalence of depression in Australia [44], and the presence of poor diet quality [45], few individuals who were interested in participation actually met the entry criteria for the trial. The small sample size, combined with over one-third of missing follow-up cost data and the typically skewed nature of cost data, present additional statistical challenges. While we have attempted to address these using methodologically rigorous methods and conservative assumptions, these results require replication. Moreover, while the outcome assessors were blinded to treatment allocation, trial participants were not. It is unclear whether expectation bias may have influenced the measured outcomes of this analysis in some way. However, psychological support has the greater face validity in the community as an effective intervention.

It is important to note that while there were no significant differences in resource use and costs between the groups at baseline, there were trends toward greater health sector and societal costs in the social support group. However, the adjusted analyses incorporated baseline cost values and showed that the findings were reasonably robust.

Moreover, the resource use and lost productivity data were reliant on participant recall; however, this was over a relatively brief time period (12 weeks). While recall may introduce additional bias into the results, participant-reported data is frequently used in economic evaluations conducted within clinical trials. The brief 12 week follow-up period aids participant recall, however it is unclear whether the cost savings found over the brief time period of this trial would be maintained over the longer term.

The 12 week follow –up may also be perceived as a limitation due to diminishing rates of adherence to diets over longer time periods. However, previous studies evaluating Mediterranean diets for the prevention of cardiovascular disease have shown high rates of adherence to this dietary pattern over two to five years [46‐48]. Additionally, given that the cost to deliver the intervention would be incurred up-front (in the first 12 weeks) a longer follow up period may in fact favour the dietary intervention in the cost-effectiveness analysis.

Finally, the resource use questionnaire, while capturing the bulk of service use, did not include questions on all societal costs (e.g. criminal justice costs and community care costs were not explicitly measured). This was largely because brevity of questionnaires was required so that subject recruitment would not be compromised. While it is unlikely that this intervention would affect such costs, we cannot be certain.