Cost-effectiveness analysis of a randomized study of depression treatment options in primary care suggests stepped-care treatment may have economic benefits

In this randomised, controlled trial [20], we recruited participants from two primary care clinics in Alberta, Canada. Study groups were block randomized by day (days were randomly assigned to one of the four arms using a random number generator). We assessed participants aged 18 and above, who were able to provide informed consent. The Health Research Ethics Review Board at the University of Alberta approved the study protocol on 30th July 2013. This trial was registered with Clinical Trials database, https://​clinicaltrials.​gov/​ct2/​show/​NCT01975207 Identifier: NCT01975207. Details on methods employed in the trial have been published previously [20] and for that reason are only summarized here.

Participants were assigned to one of four arms. In Arm 1, standard care (SC), the results of PHQ-9 screening were not communicated to the patient or their physician, unless they scored positive for a risk of self-harm. In Arm 2, treatment as usual (TAU), participants who scored greater than 10 on the PHQ-9 had their scores communicated to them and to their physicians for follow up without recommendations from the study team. Arm 3 was TAU plus online cognitive behavioural therapy (iCBT). In this arm, participants who scored greater than 10 on the PHQ-9 had their scores communicated to them and to their physicians for follow-up, and patients were provided login information for a free internet-based CBT [29] and encouraged to use it. Participants who scored greater than 10 on the PHQ-9 in Arm 4, known as stepped-care pathway (SCP), had their scores communicated to them and their physicians and were offered a pre-determined treatment as indicated by the depression stepped-care pathway, described in detail elsewhere [20]. Participants in all arms had information collected on their health-related quality of life and symptoms of depression at baseline and week 12 using the EuroQol-5 dimension (EQ-5D-5 L) and the PHQ-9 measure. Those who scored over 10 on the PHQ-9 at baseline also completed the data collection at six weeks, and if at any point any participant scored “at risk” for self-harm, they were referred for urgent treatment.

Patients completed the self-reported Patient Health Questionnaire-9 (PHQ-9), an instrument that can be used to assist in screening and monitoring the severity of depressive symptoms [23], as well as the EQ-5D-5 L at baseline and 12 weeks post-randomization. The EQ-5D-5 L is a standardized measure of health status developed by the EuroQol Group that provides respondents with a descriptive system to classify their health status based on five dimensions: mobility, self-care, usual activities, pain/discomfort, and anxiety/depression [27, 30]. This measurement provides a utility score for each of the EQ-5D-5 L health states between values of 0 and 1, with 1 representing the best (perfect) health state and 0 the worst (death) state. The analysis was performed from the perspective of health care payer, and the costs and resource utilization included in the analysis were for physician, outpatient, and inpatient services. Physician services include all activities performed in primary, outpatient, and inpatient care settings. Inpatient and outpatient costs cover all activities other than physician service in these settings. Examples of these costs include salaries, drugs, medical and surgical supplies, administration, and support services. We collected data during the periods of 12 weeks pre-randomization, 12 weeks post-randomization, and from 12 weeks to 1-year post-randomization. While some data were also collected in depressed patients at week six (approximately 20% of the total sample), it is not included in the present analysis. Data on health care system usage were retrieved from health administrative databases that provide individual patient information such as age, gender, cost, diagnostics, and service date. Since there is a single health care service in the province of Alberta, this information comprised all relevant health spending for the study patients with the exception of private psychological services, which we are unable to include in the analysis. In other words, the analysis included direct medical costs only, while other societal costs, such as productivity losses and transportation, were not included.

The analysis followed the guidelines set by the International Society for Pharmacoeconomics and Outcomes Research (ISPOR) Good Research Practices Task Force [31, 32]. An intention-to-treat (ITT) approach that considered all participants allocated to each intervention arm was applied for the analysis; this approach was utilized to avoid bias [31, 33]. Missing data were handled using multiple imputation, which reflects inherent uncertainty when replacing missing data [31, 32, 34]. The data from each of the four intervention arms were further separated into a ‘depression’ subgroup of participants, who scored greater than 10 on the PHQ-9 at baseline. The analysis focuses on the depression subgroup as this is the group that may have received clinical benefit from the treatment options, since those who did not have symptoms of depression had no need for the interventions. An overall PHQ-9/EQ-5D-5 L score was estimated for each intervention arm. Imbalances in baseline PHQ-9 and EQ-5D-5 L scores were accounted for using ordinary least squares (OLS) regression in estimating their values at 12-week post-randomization. The difference between intervention arms was tested using a one-way ANOVA test, the difference between each pair of intervention arms was tested using a pairwise comparison of mean, and the difference between baseline and 12-week post-randomisation for each intervention arm was tested using paired t-test for these scores [35].

Data investigation indicated unusually higher costs for some participants, known as outliers, compared with the average cost for each study group. Outliers skew average costs and are unlikely to represent the true average expenditure for healthcare services [36, 37]. The outliers in each period and intervention arm were trimmed. We used a traditional univariate boxplot to trim physician and outpatient costs, with cost data beyond 1.5 times the interquartile range (IQR) being excluded. We did not trim outliers for inpatient costs because of the limited available data, as there were only a few participants admitted to hospital. A generalized linear model (GLM) was applied to adjust for imbalance in baseline characteristics such as age, gender, and PHQ-9 scores in estimating costs at 12-week and one-year post-randomization. The GLM extends the linear modelling approach to data that are not normally distributed. We used a gamma distribution as the family suitable when fitting skewed healthcare cost data [38]. The link function used in GLM specifies how the mean of the dependent variable depends on the predictors. In the analysis we used identity link function, implying a linear relation between the cost and predictors. Predicted cost values from the GLM were used to represent the costs for each period and each intervention arm. Discounting was not required, as the study time horizon was just one year. All costs were adjusted to a standard price year of 2017, using the Alberta Consumer Price Index (CPI).

We compared the costs and cost-effectiveness of the intervention arms at one year to capture the economic effect of the alternatives. We used quality-adjusted life-years (QALYs) as the primary outcome to measure health effectiveness over 12 months. We calculated QALYs as the area under the curve defined by the EQ-5D-5 L utility scores at baseline and one year. The RCT did not collect the EQ-5D-5 L data after 12-week post randomisation, and we therefore assumed the utility scores over the time point remained unchanged. This assumption implies that treatment effect is maximized in the first 12 weeks. We conducted a sensitivity analysis on two different scenarios to test the impact of this assumption.

Health care utilization data for participants in the RCT indicated that, during the year after randomization, a participant received either no medical services or a combination of physician, outpatient, and/or inpatient services. As shown in Table 1, approximately 1 to 2% of participants did not receive any medical services; 40 to 45% had physician visits only; 40 to 48% had a combination of physician visits plus outpatient visits; 0 to 1.7% had a combination of physician plus inpatient admission; and 9 to 17% received medical services from all the sectors (physician visits, outpatient visits, and inpatient admission).

The SCP group had a small likelihood of using expensive inpatient services in this study, which significantly impacted overall spending observed during the trial. In order to verify that the results of this analysis were not biased due to this decrease in inpatient spending, we developed a decision tree model to capture the cost of each combination of services for participants. This model splits participants into one of the five combinations reported in Table 1 and assigns the probabilities of service use to each group. We used the decision tree in the base case analysis, and because the likelihood of inpatient use is the main component driving difference in cost-effectiveness, we conducted a one-way sensitivity analysis to assess the impact of changing the likelihood.

The incremental cost-effectiveness ratio (ICER) was calculated as the ratio of differences in mean costs and mean number of QALYs [39]. The differences in means were assessed using Monte Carlo Simulation, a mathematical technique to address the uncertainty of the cost effectiveness. The input cost and effectiveness data were repeatedly modelled for 10,000 iterations on the basis of known probability distributions for generating the outcomes [31, 40‐42]. Based on results from the Monte Carlo simulations, we constructed the cost-effectiveness acceptability curves and incremental cost-effectiveness scatterplots to help the understanding of the uncertainty of the ICER [39, 43]. In addition to the analysis performed on the overall participants, the cost-effectiveness analysis was performed in a subgroup of participants who scored greater than 10 on the PHQ-9 at baseline [31, 44]. The statistical analysis was performed with the Stata software package (Release 13.1) for Windows, and decision analytic modelling analysis was performed with TreeAge Pro 2015 software (TreeAge Software, Inc).