Data collection

Data were collected at four time points. T ₀ was the baseline assessment, conducted between 17 July 2019 and 31 January 2022. T ₁, T ₂ and T ₃ represented follow-up after 1, 6 and 12 months, respectively. The study protocol provides detailed information on the data collection process. ^{Reference Kohlmann, Lehmann, Eisele, Braunschneider, Marx and Zapf19}

Depression severity was assessed using the PHQ-9, ^{Reference Kroenke, Spitzer and Williams22} and health-related quality of life (HRQL) was measured using the EuroQol-5D-5L (EQ-5D-5L). ^{Reference Herdman, Gudex, Lloyd, Janssen, Kind and Parkin23} Furthermore, depression diagnoses were ascertained using the Mini-International Neuropsychiatric Interview (MINI) at T ₁, T ₂ and T ₃. ^{Reference Sheehan, Lecrubier, Sheehan, Amorim, Janavs and Weiller24} Healthcare use, including medication intake and productivity losses, was assessed by a modified version of the Client Sociodemographic and Service Receipt Inventory. ^{Reference Roick, Kilian, Matschinger, Bernert, Mory and Angermeyer25}

EQ-5D-5L

The EQ-5D-5L is a questionnaire to assess HRQL in five dimensions: mobility, self-care, usual activities, pain and/or discomfort, and anxiety and/or depression. ^{Reference Herdman, Gudex, Lloyd, Janssen, Kind and Parkin23} Respondents rate their HRQL in each dimension on a five-point scale ranging from 1 (no problems) to 5 (extreme problems), resulting in 5⁵ = 3125 different health states.

The EQ-5D-5L is broadly accepted and applied as an HRQL measure; its psychometric properties are considered to be reliable and have been well established. ^{Reference Feng, Kohlmann, Janssen and Buchholz26} In the base case analysis, the German value set was applied to the EQ-5D-5L descriptive system, resulting in EQ-5D-5L index values ranging from −0.661 (worst imaginable health state) to 1 (best imaginable health state). ^{Reference Ludwig, Graf von der Schulenburg and Greiner27}

PHQ-9

The PHQ-9 is a measure of depressive symptom severity that is often used as a screening instrument. ^{Reference Kroenke, Spitzer and Williams22,Reference Löwe, Unützer, Callahan, Perkins and Kroenke28} It consists of nine questions regarding the frequency of typical depressive symptoms using a four-point scale (0–3). The overall PHQ-9 score is the sum of points per item. Depression severity is indicated by five levels: no/minimal (PHQ-9 score 0–4), mild (5–9), moderate (10–14), moderately severe (15–19) and severe (20–27) depression. ^{Reference Kroenke, Spitzer and Williams22} The PHQ-9 is considered to have both high specificity and sensitivity, as well as superior psychometric characteristics. ^{Reference Miller, Newby, Walkom, Schneider, Li and Evans29}

Healthcare service use and costs

Healthcare service use was estimated from participants’ responses to an adapted version of the Client Sociodemographic and Service Receipt Inventory throughout the trial. ^{Reference Roick, Kilian, Matschinger, Bernert, Mory and Angermeyer25} Data on in-patient, out-patient and non-physician services (physiotherapy, massage, etc.) and pharmaceutical intake, as well as formal and informal nursing care, were collected. At baseline, numbers of medications, GP visits, in-patient stays and sick leave days were assessed.

Direct and indirect costs were calculated on the basis of German standard unit costs. ^{Reference Muntendorf, Brettschneider, Konnopka and König30} For medication costs, a German pharmaceutical directory (Rote Liste) including pharmacy retail prices was used. ³¹ Informal care was estimated with the replacement cost approach, using gross labour costs for ‘social work (excluding homes)’ (economic sector Q88) plus non-wage labour costs as reference costs. ^32,33 Indirect costs included productivity losses due to absenteeism from paid work and were estimated using the human capital approach on the basis of gross labour costs in Germany. ³⁴ In this health economic evaluation, intervention costs were disregarded because differences in intervention costs between study arms consisted only of printing costs for one or two sheets of paper. Overall intervention costs consisting of technology and personnel costs (compared with no screening) were approximated on the basis of publicly available data. As these costs were minimal, they were disregarded in the analysis, but they are presented in Supplementary Material 3. All additional costs triggered by the intervention were covered in the costs of healthcare service use. Therefore, intervention costs were considered to be negligible and not relevant to decision-making from a societal perspective. Costs (€) were adjusted to the reference year of 2022 by applying the consumer price index. ³⁵ An overview of all considered unit costs is provided in Supplementary Material 4. Total costs were estimated by summation of costs across all available categories and time points.

QALY calculation

Quality-adjusted life years (QALYs) were estimated on the basis of EQ-5D-5L index scores. ^{Reference Whitehead and Ali36} QALYs between consecutive time points were assumed to be subject to a linear trend and calculated as follows:

\begin{align}& \,\,\,\,\,\,\,\,\, {\text{QALY}} = \left[{\text{inde}}{{\text{x}}_{{\text{early}}}} + \;\left[ {{{{\text{inde}}{{\text{x}}_{{\text{late}}}} - {\text{inde}}{{\text{x}}_{{\text{early}}}}} \over 2}} \right] \\ \times \left[ {{{{\text{reference}}\;{\text{days}}} \over {\Delta {\text{days}}\left( {{\text{early}};\;{\text{late}}} \right)}}} \right]\right] \times \left( {{1 \over 2}} \right). \end{align}

Total QALYs in the follow-up period were the sum of QALYs between consecutive follow-up surveys. Reference days were 30 for the period between T ₀ and T ₁ (1 month), 150 for the time between T ₁ and T ₂ (5 months), and 180 between T ₂ and T ₃ (6 months). Neither QALYs nor costs were discounted owing to the short follow-up period.

Base case analysis

In the base case analysis, we focused on the cost-effectiveness of the interventions from a societal perspective, considering a 1-year time horizon. Regarding costs, all available direct costs from healthcare service use, medication and informal care were included. Productivity losses due to absenteeism were incorporated as indirect costs. As effects, QALYs estimated using the EQ-5D-5L were used.

Before the analysis, we evaluated systematic differences between study arms using linear ordinary least squares regressions for continuous variables (age, baseline in-patient days, number of medications, GP visits, EQ-5D-5L index, EQ VAS, PHQ-9), binary logistic regressions for binary variables (health insurance plan, school education ≥10 years) and multinomial logistic regressions for non-binary categorical variables (gender, city size).

Cost differences during follow-up were estimated by applying generalised linear mixed models with a gamma distribution and a log-link function. Effect differences in QALYs were estimated using linear mixed-effects regression models. In both models, random effects for practices nested in study centres were assumed. As fixed effects, we considered a set of sociodemographic characteristics (age, gender, city size, living situation, education), depression risk factors (depression history, current and recent pregnancy), baseline PHQ-9 and EQ-5D-5L scores, and baseline GP visits and sick leave days as covariates. We assumed a significance level of 5%. Unadjusted incremental cost-effectiveness ratios (ICERs) between study arms were estimated to describe the additional cost per incremental unit of (health) effects due to an intervention. ICER was calculated as follows:

$${\text{ICER}} = {{\overline {{\text{cost}}{{\text{s}}_{{\text{intervention}}}}} - \overline {{\text{cost}}{{\rm{s}}_{{\text{control}}}}} } \over {\overline {{\text{effect}}{{\text{s}}_{{\text{intervention}}}}} - \;\overline {{\text{effect}}{{\text{s}}_{{\text{control}}}}} }} = {{\Delta \overline {{\text{costs}}} } \over {\Delta \overline {{\text{effects}}} }}.$$

As the ICER is a point estimate and insensitive to data uncertainty, cost-effectiveness acceptability curves (CEACs) were constructed to visualise cost-effectiveness probabilities for different WTP thresholds. ^{Reference Briggs, O’Brien and Blackhouse39} Net benefit regressions (NBRs) were conducted considering WTP margins between €0 and €160 000 per QALY with €10 000 increments. The detailed rationale behind these methods can be found elsewhere. ^{Reference Hoch, Briggs and Willan40–Reference Zethraeus, Johannesson, Jönsson, Löthgren and Tambour42} Briefly, NBRs estimate the probability of cost-effectiveness, i.e. the likelihood that the intervention has a positive net benefit. This is done using the P-value of the group or intervention estimator in the NBR, with 1/2 the P-value used when the net benefit is negative and 1 − 1/2 the P-value when the net benefit of the intervention is positive. For the NBR, we again applied linear mixed-effects regression models using the covariates described above. All analyses were conducted pairwise, comparing each intervention group separately with the no-feedback group.

Deterministic sensitivity analyses

We performed sensitivity analyses for three scenarios. In the first scenario analysis, we focused on costs from a healthcare payer perspective, including all direct costs except informal care. In this scenario, absenteeism at baseline was excluded from the list of covariates. As the second deterministic sensitivity analysis, and to test the robustness of the multiple imputation approach, we performed a complete case analysis. In this scenario, only participants for whom both societal costs and QALYs for the 12-month follow-up were available (not missing) were included.

In the third scenario, we calculated depression-free days (DFDs) on the basis of PHQ-9 scores, estimating ICERs in €/DFD. We assumed a full DFD when the PHQ-9 score was below 5 (no or minimal depression). When the PHQ-9 score was 15 or higher (indicating moderately severe depression), we assumed no DFD. For PHQ-9 values between 5 and 14, we used linear interpolation to estimate the probability that a participant had a DFD. We assumed linear changes in PHQ-9 scores, analogous to the calculation of QALYs. This reliable and well-validated approach was consistent with that used in previously published studies. ^{Reference Lave, Frank, Schulberg and Kamlet43,Reference Vannoy, Arean and Unützer44} To determine a cost-effectiveness threshold per DFD, we adapted the approach used by Unützer et al ^{Reference Unützer, Katon, Russo, Simon, Von Korff and Lin45} to the German setting by weighting their threshold with the gross household income in Germany. ⁴⁶ Furthermore, we calculated a CEAC for WTP margins between €0/DFD and €200/DFD with €20/DFD increments.

Post-hoc explorative subpopulation analyses

In addition, we conducted post hoc explorative subpopulations analyses to further disentangle possible cost and effect differences and determine subpopulations for which the interventions could possibly be cost-effective. As these analyses were not pre-specified, ^{Reference Kohlmann, Lehmann, Eisele, Braunschneider, Marx and Zapf19} they should be considered to be exploratory, and their results need to be handled with care. We focused on the categories for which the GP-targeted plus patient-targeted feedback intervention was associated with a significant (significance level: P = 0.1) decrease in the primary outcome of clinical effectiveness (depression severity), namely female gender, having had a previous depressive episode and the absence of any comorbid addiction. ^{Reference Löwe, Scherer, Braunschneider, Marx, Eisele and Mallon20} Therefore, we considered gender, depression history, and whether patients had any (behavioural or substance) addiction at baseline. We also estimated the costs, effects and cost-effectiveness of feedback interventions for the subpopulation of participants whose suspected depression was confirmed 1 month after baseline using the MINI diagnostic interview ^{Reference Sheehan, Lecrubier, Sheehan, Amorim, Janavs and Weiller24} compared with those who had not been diagnosed.

We performed subpopulation analyses by dividing the imputed data-sets with respect to the categories of interest and then conducting all analyses as previously described. Hence, we estimated cost and effect differences between study arms for subpopulations as well as cost-effectiveness. We applied the same covariates as described previously, excluding covariates with too few observations per group (pregnancy, breastfeeding and living situation), and city size was transformed to a binary variable. As randomisation was not stratified considering these categories, the results of these analyses should be interpreted cautiously and seen as inspiration for further research.